This article is a collaboration with Tony Siewert, a tech professional who dissects the cultural and philosophical impacts of modern AI. Tony champions the messy, vital human element in the age of automation and explores why, in today's tech landscape, he takes pride in being branded as 'The Antiquated One’.

Act I. The Three Actors

In the beginning of March 2026, one and a half million people switched AI providers out of moral outrage. They cancelled their subscriptions to OpenAI and signed up with Anthropic. By every available metric, it was an act of resistance. By every structural measure, it was the exact opposite.

The stage was being built weeks and months earlier. The cast included Anthropic, OpenAI, and the Pentagon itself. On the surface, the play is about the question of Ethics: can a private company defend its safety protocols against the will of the state? The real drama, however, played out between the lines.

Subscribe now

The first company to deploy its models on the classified networks of the US government was Anthropic, our actor number one. Their Claude Models were already running inside mission-critical intelligence operations. They even explicitly opened its models for missile defense and combat logistics.

Anthropic always does everything “correctly” — almost suspiciously so.

Sounds great? We aren’t done yet. Anthropic had a product that built a genuine community. Claude Code became the primary driver of growth velocity and valuation at Anthropic. The latest models — Opus 4.6 and Sonnet 4.6 — were leading the major benchmarks in AI, LMSYS Arena among them.

That is the definition of Aura Farming, but even that wasn’t enough. When literally dealing with the Ministry of War, Anthropic managed to maintain two red lines: 1) Claude would not be used for mass surveillance of American citizens. 2) And it would not control fully autonomous lethal weapon systems. The first line was unambiguously ethical. The second existed for a slightly simpler reason: Dario Amodei genuinely did not believe current AI models were ready for that use case.

One might expect that doing everything “correctly” earns a reward. But Actor number 2 has a mind of his own: Enter the Pentagon. Following the general ultra-conservative rollback, the Pentagon was already unhappy about “woke” AI in January 2026. Weeks of tense negotiations over a $200 million contract followed. On February 24th, Defense Secretary Peter Hegseth delivered an ultimatum to Amodei in person: drop all AI safety rules by Friday, February 27 — or be designated a “supply-chain risk to national security”.

No American company had ever been threatened with such treatment — it is normally reserved for cases like Huawei, a subsidiary of a geopolitical rival. The threat works not through legal precision but through the ambiguity it creates. Though the designation formally applied only to direct military contracts, defense contractors and suppliers are heavily incentivized to drop Anthropic’s products preemptively, if only to avoid the associated stigma. For a company whose models are already deeply embedded in these infrastructures, the designation alone can be existential.

Anthropic — the company that does everything “correctly” — refused to comply and challenged the designation in court. Trump himself weighed in with a post branding Anthropic a “leftwing nutjob” and worse. Anthropic stood alone.

The post of Donald Trump on Anthropic

Act II. The Bailout That Isn’t a Bailout

Anyone who thinks Anthropic brought this upon itself has forgotten the third player: OpenAI with its CEO Sam Altman.

Unlike Anthropic, OpenAI is in serious trouble. The company that launched the LLM hype with the massive success of ChatGPT has become the central node in a web of circular financing deals that artificially prop up large parts of the US tech economy. When Microsoft “invests” in OpenAI only for that money to flow directly back as payment for Azure services, what emerges can hardly be described as value chain as it mimicks a carousel. That such constructions blur the line between creative accounting and antitrust evasion is no secret.

And as tech critic Ed Zitron recently revealed through leaked financial documents, the subsequent revenue numbers that OpenAI has shown to the world are failing to convince anyone that the ‘math is mathing’.

The deals have grown so vast that OpenAI is already being labeled “too big to fail” — and the question of a state-guaranteed bailout hangs openly in the air. Set against Anthropic’s success — better products, no absurd debt mountains, everything done “correctly” — OpenAI’s position turns precarious.

When customers began defecting en masse to Google and its flagship model underperformed, the company declared “Code Red” in December 2025 — halting most side projects overnight in a desperate pivot to salvage its core technology.

No AI drama of this era unfolds without OpenAI — this one least of all. First, Sam Altman publicly backed Amodei’s principled stance — despite having refused to shake his hand at the AI summit in New Delhi just weeks earlier.

Sam Altman refuses to take the hand with Dario Amodei (third and second from the right, respectrively)

Then, immediately after the Pentagon branded Anthropic a supply-chain risk, Altman announced two things. OpenAI would step into the Pentagon contract. And it had secured a historic $110 billion funding round. Both on the same day. The entire Anthropic-Pentagon narrative was instantly subsumed under a single name: Sam Altman.

OpenAI’s sudden alignment with the military becomes legible against this backdrop. In late 2025, the administration had explicitly refused a direct bailout — as there were enough frontier companies, and they were allowed to fail. OpenAI was out of options.

By taking the Pentagon contract OpenAI established a new state-secured monopoly. And that is, essentially, a bailout that isn’t a bailout.

Yet Altman wouldn’t be Altman if Anthropic were simply thrown under the bus. In the end, Anthropic may have been the biggest winner of the entire spectacle. Anthropic’s refusal is being celebrated as a second coming of Apple’s famous decision not to install backdoors on its devices. Downloads surged across every app store. Millions switched from ChatGPT to Claude.

What began as a meme quickly became a measurable exodus. Under the banner of #QuitGPT, over 1.5 million users pledged to boycott OpenAI, driving a 295% spike in ChatGPT uninstalls and propelling Claude to number one on the App Store.

Amodei was declared a hero. The Atlantic even published an article, directly comparing Dario Amodei with Oppenheimer.

Act III. The Absence of the Human

And yet, on this whole stage, there is but one reality: none of it truly matters.

OpenAI may well be glad to shed its freeloaders — the free user base was nearly impossible to monetize. But even beyond such speculation: most consumer AI subscriptions are massively subsidized — by venture capital, by Big Tech, by one of the largest corporate subsidization campaigns in the history of technology. Even the most expensive Claude Max Subscription may consume up to $5.000 worth of compute.

It has never been about short-term profit, as none of these companies can cover their astronomical computing costs even remotely. The true goal is infrastructure capture. They want to be the default platform—meaning their models become the foundational layer that every other company, developer, and government agency is forced to build upon. They are giving away billions in subsidized compute today to ensure they own the monopolies of tomorrow. This is the field on which the entire theatre takes place.

In this play, there can only be one winner: technology itself. And the principal loser is conspicuous by absence: The Human. In no act of this spectacle did The Human ever appear as subject. The Pentagon was only about securing its perceived dominance. OpenAI fought merely to survive its debt web.

And Anthropic? The ethical AI company — which assembled its training data from the intellectual property of the world, while only a privileged few creators can afford the legal battle against such a giant. Before the Pentagon drama, Anthropic had already quietly lowered the safety requirements for its next generation of models.

And the red lines? The recent mass killing of at least 160 schoolgirls during the strike on Iran was probably caused by a faulty target identification by Claude’s models. This happened one day after Amodei had been celebrated as the ethical conscience of the AI industry.

Anthropic is far closer to OpenAI in spirit than either would ever admit. They need each other. One is the villain whose presence lets the other’s virtue shine. “Thank God there’s only one Sam Altman!” critics may say — while both companies jointly industrialize what Günther Anders once diagnosed as the ground condition of our epoch: the obsolescence of the human in the face of its own creations.

If all of this is theatre, if the bailout is no bailout, then the play functions not as an attack on the human, which would presuppose intent, but as something staged past the human entirely, in complete structural indifference.

The consumer’s decision to switch to Anthropic changes nothing about the underlying dynamic, even though it feels like resistance. The hand still reaches. What has been replaced is everything behind it.

The true achievement of this spectacle: it takes the genuine moral outrage of its users, their still-functioning ethical intuition, and converts it into a new subscription. The Industrialization of Hype. In this apparatus, Anthropic can never be seen as the antidote. It is merely other side of the same mechanism. What it produces is the ultimate obsolescence: our deepest moral convictions, reduced to raw material for the expansion of the very system they were meant to oppose.

In the end, the entire spectacle of AI ethics revealed itself as what it was from the start: a mechanism for securing backdoor bailouts — and for accelerating the very machinery that no longer requires the human it was supposedly about.

Share

This is how things were for a long time. But the world is changing: old empires fall, new ones arise, and sometimes the old ones return. It is a perpetual cycle of civilisation. Often, when one empire collapses and another takes the lead, we tend to go into denial: 'No, they just copied what we had, they are not there yet - we are the crème de la crème of humankind. Not them.'

A dying dinosaur keeps roaring but the ice age is already covering the earth.

The third thing that you need to know about AI is that Chinese AI is excellent. It is a genuine innovation that is going forward at an enormous pace.

This is the final post in the series. You can read the previous two here:

Are you navigating an AI transformation and struggling to separate real value from hype? Not sure which tools to choose, how to govern agents, or whether a custom use case is worth the investment?

I offer personal consultations, workshops, and trainings tailored to exactly these questions.

Reach out through this form, describe your situation, and get a free 20-minute consultation.”

According to the State of AI Report 2025, Chinese open-weight models have surged ahead in user preference, global downloads and model adoption, becoming the world’s go-to open LLMs. In this article, I will argue that China is a new shiny AI empire that has resources, talent and infrastructure that, at the current state, is virtually unachievable for the European roaring dinosaur.

The US finds it increasingly harder to keep up with the pace and speed of innovation. At ICLR, one of AI's three most prestigious conferences, Chinese papers were outnumbered 5-to-1 by American papers in 2021 — by 2025 they are almost equally numerous. At NeurIPS 2025, China and the US were neck-and-neck in paper counts, at ACL-2025 Chinese papers significantly outnumbered the US papers -a dramatic reversal from just a year before.

It is time to swallow the arrogance and the superiority complex, turn our eyes towards the East and learn from what is happening there. They learned from us but the student outperformed the teacher and now it’s teacher’s time to become the student.

In this article I will talk about the recent advances of Chinese AI, how we can leverage it, how we can use it now and maybe even what it would take to make sure that the roaring dinosaur does not die out but evolves and comes back as a different species and dominates the earth again.

But let us be intellectually honest about where Europe actually stands before we talk about evolution. Mistral, Europe's flagship model, currently lags 6–12 months behind leading US and Chinese systems on standard benchmarks, while costing more per token. Europe is not competing for the frontier - not today, not tomorrow, not in the next few years. And pretending otherwise is exactly the kind of denial we just criticized. The dinosaur is not secretly winning. It is losing, and it knows it.

The dinosaur that survives is the one that stops roaring about past greatness and starts evolving, led by natural selection, picking up the dominant and winning features of the others and adjusting to the new environment. It cannot exist as a ghost of what it was but needs to continue as a new species, built for the new world.

Agentic Large Language Models

A few months ago I wrote that the release of Kimi-K2 was one of the most significant events in AI in 2025 along with the GPT-5 flop. Many have completely forgotten about that model but that model was not just another model - that was a model that opened up a new era: open-weight agentic models that are on par with commercial models. Kimi-K2 reached parity with American models, outperformed them in coding and agentic calls. When GPT-5 struggled with complex real-world coding tasks, Kimi-K2 was designing full websites. That is when it was clear what is happening next: Chinese models will be on par with US models and might even become the best agentic models.

Further in the article:

• Which specific Chinese models can we use, and where, without compromising security

• Which Chinese video, image, and text models are outperforming their Western counterparts

• Which Chinese models are 20× cheaper than Claude Sonnet, deliver comparable performance, and fit on desktop GPUs

• What is the reality of bias in Chinese models

• What distillation “attacks” are

• What we can learn from China

Read more

Gatekeeping a technology in the way the EU AI Act is doing it is a security threat. China taking the lead in the AI race, delivering cheap models, benefiting from the collective research mind by providing their models as open-weight — all this could make them immensely stronger.

On the other hand, pathetic drama between OpenAI, Anthropic and similar, regulating instead of innovating in Europe and hyping up AI’s capabilities instead of seeing what is already there is only making us weaker.

Claude and attack planning.

For this article I consulted a military expert who wanted to stay unnamed. We talked about the possible application of AI in the war in Iran. On February 28, 2026, a girls’ elementary school in Minab, southern Iran, was struck during the opening wave of US-Israeli airstrikes, killing between 165 and 180 people, mostly students. No official admission of responsibility has been made, but satellite imagery analysis by NPR, CBC, and CNN confirmed by multiple independent experts suggest the strike was likely a US airstrike resulting from outdated targeting information. The school had been physically separated from the adjacent IRGC military complex since 2016 but appears to have remained in an unupdated target set. It should be noted that two competing hypotheses exist: either this was a targeting failure due to stale data, or, as Al Jazeera’s investigation raises, the strike was deliberate. I will argue for the first.

I am going to break down why, in my personal opinion, the theory that an AI system misclassified the school as a military target based on outdated data is a plausible hypothesis. My expert made a point that reframes how we should think about AI use in this war:

“If you are defending against 1,000 drones, that’s why they use swarming drones — it’s about scaling AI for protection and attacks.”

The sheer scale of targets struck by US forces across Iran in the opening hours is itself evidence that AI-assisted targeting was in play. Humans cannot generate, validate, and execute hundreds of simultaneous precision strikes without algorithmic support. When you scale targeting to that degree, the margin for stale records in the target database grows.

The unprecedented speed and scale of the targeting

Deliberate targeting is the kind of targeting that was used in the Iran opening strikes. A target is placed on a pre-built list weeks or months in advance, based on satellite imagery, SIGINT-derived geolocation, and intelligence analysis.

The strike goes out against a database entry. This process formally involves a chain of humans: 1) analysts who increasingly review intelligence feeds 2) commanders who evaluate the target’s accessibility and legality, a step that has been deliberately compressed under the current administration, and 3) legal advisors who verify compliance with IHL (International Humanitarian Law) which also got significantly simplified lately.

Here is an overview of the process from an official US Joint Staff / Pentagon briefing from 2009. Doctrine has changed significantly since then, and the briefing should be read as a picture of how this was supposed to work before AI entered the targeting chain.

Now consider what happened in Iran. The US struck nearly 2,000 targets in the first 100 hours of the operation. That list was not generated on the fly but it was built months in advance. But managing, sequencing, and executing 2,000 targets simultaneously across multiple domains still requires algorithmic support that no human coordination chain can match unaided. More importantly: maintaining a database of 2,000+ target records and keeping every single one correctly classified is exactly the kind of task that gets delegated to automated systems.

Additionally, operators must formally estimate collateral damage (CDE — Collateral Damage Estimation): the number of civilians likely to be affected by the attack:

As we can see, it is a tedious and time-consuming process to execute strikes. This is not what we see in the data, however.

In the first 12 hours alone, the US and Israel carried out nearly 900 strikes — a tempo that would have taken days or weeks in any conflict before this decade. This raises questions about the human supervision requirements outlined above. The opinion of the military expert I spoke with, that it would be highly unlikely this was achievable without AI, has also been confirmed by other sources. As Craig Jones, a senior lecturer in political geography at Newcastle University and kill chain expert, put it in The Guardian:

The AI machine is making recommendations for what to target, which is actually much quicker in some ways than the speed of thought,

To execute such a large-scale attack, with a pre-built list of 1,000 targets developed, CDE-cleared, and executed in a compressed planning and execution cycle, would be near impossible for humans alone. They would need to evaluate multimodal input: images, text, video feeds, signals intelligence, while simultaneously ensuring that multiple legal and operational conditions are satisfied, applying for permission at each level of the chain, and then executing. LLMs, as well as other models, particularly computer vision, are highly capable pattern matchers. This is exactly how Palantir's Maven Smart System works: integrated with Claude since 2024, it outputs precise location coordinates for missile strikes and prioritises them by importance. Matching targets, analysing whether conditions are satisfied, flagging requests and supporting targeting decisions can all be done within minutes.

Multiple media outlets report that one of the Claude models by Anthropic was used to support intelligence analysis and target identification during the strikes.

Ironically, the US government had banned Anthropic from all government projects just hours before the strikes began - something that quite clearly did not stop them from using Claude for military needs.

Further in the article:

• Map pattern analysis

• An experiment with Claude that shows it would convincingly identify the school as a primary target based on satellite imagery

• AI usage for military purposes

• Implications for world security, the economy, and what Europe should learn from it

Paid subscribers also get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training.

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

20% off the subscription till the 9th of March!

Subscribe now

Read more

To celebrate 100 articles I am offering 20% subscription discount to the newsletter.

All the articles are written from scratch as drafts. I never use AI for brainstorming, writing a draft, or sourcing ideas. I decided that this newsletter needs to be a mix of practical exercises and critical opinions: one cannot go without the other.

Subscribe now

You cannot explain the limitations of AI if people do not actually use it. That is why I make quarterly free trainings for paid subscribers; we have already built a website, produced a promo video, and there are more coming - paid subscribers also have access to the slides, recording and if they follow the tutorial they can also get support from me if they get stuck by commenting under the articles linked above.

I write practical deep dives into the latest AI architectures and what I believe could be the future of AI. I always aim to ground my opinions in the latest peer-reviewed research, as it is a strong indicator of what the future brings.

Finally, I focus a lot on practical advice for decision-makers who are now confronted with integrating AI into their processes and creating governance. There are so many misconceptions caused by hype and mislabeling, where very smart people seriously discuss very wrong ideas, such as:

• You need to track and govern every agent that is built.

• Agents can replace humans and are dangerous.

• The future of automation is thousands of agents performing various tasks.

• Trying to augment old processes with new technology.

All while not understanding that you need to avoid:

• Underestimating the power of AI-assisted coding: The first and foremost thing you need is to ensure your IT department has the best AI coding tools.

• Believing you need a swarm of agents: The key to effective AI utilization is not optimized processes and thousands of agents, but individual productivity and hyper-personalization. Instead of 1,000 agents, you need one tool similar to OpenClaw that has access to various data and skills and can adapt to the user’s needs.

• Neglecting infrastructural adjustments: AI cannot work well if it does not have access to the necessary information. If your APIs are not equipped with MCP servers, or worse, you do not even have API connections, it will fail. If your data is not digitized and is spread all over various folders, emails, and local PCs, it does not matter if you build 1,000 agents or buy an MS 365 Copilot license—both will be useless.

• Governing the wrong thing: LLMs are not special, and agents are not humans. You do not need to govern every model, LLM, or protocol for every agent. What you need to do is make sure the AI has access to the right tools that are compliant with your infrastructure, that it has role propagation and access rights aligned with the human user, and that the reusable skills are safe to use in your ecosystem.

These are the types of topics I write about. I also make trainings for companies, consult small and medium businesses on AI adoption, and do independent evaluations of use cases:

https://www.airealist.org/airealist-trainings-workshops-keynotes

Working with me can save your company thousands of Euros simply by making sure that you invest in the right things and enable your staff with the right tools.

For paid subscribers, I also provide a link behind this paywall to the latest Video Generation Course and the slides that are behind the paywall. Check it out, rewatch it, and I’m looking forward to the next 100 posts!

Read more

Iran might have seen the death of Supreme Leader Ali Khamenei, but also over a hundred girls were killed in a school in Iran.

This war will bring a lot of pain and it is not going to be a glorious story where brave Americans come to a foreign land and bring democracy.

Subscribe now

The Iranian people stood up against their government long before: women who refused to wear hijabs, students who were killed at the demonstrations, and rights activists who were executed by the government.

Now these people are being killed by American bombs.

So there is nothing to cheer for here.

The price that people in Iran will pay for their freedom will be immeasurable.

The rest of the world will also not be the same. Iran is located at a strategically important geographical position. It is not only Iran that is being attacked: the whole region is drawn into the war:

https://iran.airealist.org/

Kuwait, UAE, Bahrain, Oman are being bombed by Iran. And of course, Israel is seeing bombings as well.

There is a lot of news going around. We, in the West, mostly get information in English from Western media. Unfortunately, those media are frequently not the fastest. In the age of the internet, the first news comes through reposts and sharing on social media. During the war in Ukraine, Telegram channels became the essential sources of information. Yes, one cannot trust them, but over time people are used to validating the information, checking in other sources. What I have also learned is that it is always a good idea to check what the other side is reporting: yes, they might be lying, you cannot trust media that are controlled by dictators, but behind every lie hides the truth, even if the truth is completely the opposite of what is being reported. There is no lie without truth, and sometimes it is easier to understand what is happening by monitoring lies as well.

I know from my own experience that it is extremely time-consuming and stressful. That is why I decided to build a website that tracks various news agencies and Telegram channels.

It is free: https://newsfeed-staging.pages.dev/briefing

And open-source with MIT license: https://github.com/ktoetotam/NewFeeds

Contributions: welcomed!

My primary focus was on creating diversity of sources, particularly integrating sources in Farsi, Hebrew, Arabic, watching Hezbollah Telegram groups. Those are sources that are not universally accessible: a person speaking Hebrew will not know what Persian outlets are reporting, and people speaking Arabic would struggle with Hebrew.

I also included sources from far away like China, India, Russia. Russian sources were surprising because they report much faster than Western media. The only ones that are faster are Telegram channels, for example, “Electrohizbullah“.

As I said above, the lies can tell you a lot about the truth if the truth is hidden. That is why I included state controlled media from Russia, Iran and Proxies and labelled them as such:

I have built this website in 12 hours, it updates at a 10-15 minute pace, is completely open source and can be adapted to your own use cases. Further in the article, I will discuss how this website was built and how anyone else can reproduce it.

I conclude with why AI-assisted coding might become a basic skill and how we might in general move away from specialised products and towards hyper-personalisation.

If you liked this website, you can support it by becoming a paid subscriber to keep the website running and being developed.

Paid subscribers also get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training.

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Read more

A friend of mine clicked together an agent in AWS Bedrock Agents: he used a QuickSight MCP server from the company’s official vetted list that is a server that had already been cleared by the cybersecurity team. All the agent did was query QuickSight. He deployed it in Amazon Q Business and wanted to use it with a couple of colleagues. The whole pipeline, including the MCP server, has identity propagation, so the setup by default limits the visibility of QuickSight data to the user chatting with the bot. Actually, nothing more needed to be done. The whole configuration took five minutes.

But then it turned out that according to internal guidelines inspired by the EU AI Act, he could not simply share this agent. Someone in the company had decided that they needed to write out the risks of all agents according to the EU AI Act, describe the agent in detail, register it in an internal agent repository, and get a cybersecurity assessment - for a pipeline built entirely from components the cybersecurity team had already approved. So what took five minutes to click together led to two weeks of bureaucracy.

When I first heard it, I thought it was just some strange misunderstanding, but then similar stories started coming from left and right.

The motivation is roughly the same:

Agents can change data, can delete databases, can leak information. w

We must know what agents we have and what risks they bring.

We also cannot allow people to build agents in the wild without our carefully crafted guidelines and supervision, otherwise we will end up with a lot of redundant agents and total chaos.

At first sight, this sounds logical.

But does it really make sense in a case where you just connect a tool to an LLM in five minutes within the same ecosystem, where everything is already aligned? Why would you need all this bureaucracy?

This type of governance would render platforms such as GitHub Copilot, Copilot Studio, Google ADK, and many others effectively useless, as you can build agents there in minutes or seconds. Some agents do not even need a tool. Not so rarely, an agent is just a prompt. So let us understand what is happening.

Paid subscribers also get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training.

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Subscribe now

The Cybersecurity Perspective on Agents

I did reach out to some friends who work in cybersecurity at large corporations to understand why they are doing this. However, it felt like we were just not hearing each other. I was trying to explain that doing something like this purely out of the motivation that “agents change data” is like requiring people to register every Excel macro they write, since macros also change data.

They said: “but we need to know what data is accessed and changed”. I tried to explain that you will essentially face a combinatorial explosion, as you can combine as many tools as you want in an agent for your specific needs, provide different prompts, give different sources and registering every single agent is nonsensical.

They said “but we must control what agents do and prevent people from creating duplicate agents”.

I was left wondering why they could not see that this whole approach makes no sense.

Microsoft and the Gospel of Agent Governance

Today at the Microsoft AI Tour, I finally understood what the problem was and why we were not understanding each other. One of the speakers was talking about governance for agents and was making exactly those same points:

• We need to know all the agents. Agents must have IDs.

• We need to know what data the agents access.

• We need to make sure there are no wild, unregistered agents.

And it turns out this is not just internal corporate policy: this is Microsoft’s official governance approach for agents. Announced at Microsoft Ignite 2025, Microsoft Agent 365 is Microsoft’s control plane for AI agents across the enterprise. It introduces a centralized agent registry as a single source of truth for every agent in the organization, unique Agent IDs via Microsoft Entra, governing agents the same way you govern human identities, with a designated human sponsor accountable for each agent’s lifecycle, and explicit detection and management of “shadow agents,” meaning any unregistered agent is treated as a security risk.

So who are we to argue with Microsoft?

And, yes, just as expected, the speaker said something that made it all clear:

Agents are our new reality. They are like humans. Humans need to be identified, and so do agents.

I have to say, the problem was obvious: the people who do governance fundamentally misunderstand the current state of the art of AI agents.

The AI Agent Misnomer Created a Governance Crisis

But it is not their fault.

This is the result of endless hype, the misnomer "agent," and wishful terminology such as "intelligence," "autonomous," and "agent" itself. Fake AI expertise and loud sales pitches have created an illusion that we are actually dealing with autonomous agents that can make decisions on their own, can be easily reused, and are basically small intelligent applications - like a little robotic sidekick that every employee gets:

This is utterly wrong.

It is very unfortunate, but we have heard too many comparisons like this: Gartner saying that teams in 2026 will be composed of humans and agents, or McKinsey's CEO Bob Sternfels announcing at CES in January 2026 that when asked how many people McKinsey employs, he now answers "60,000 — but it's 40,000 humans and 20,000 agents," with a goal of reaching full parity by end of 2026.

Of course, even very smart specialists would start taking it seriously.

Rename Them and the Problem Disappears

Let me call agents what they should be called: routers. Have you ever asked yourself why we even need 30,000 agents? Why can’t we just go to Microsoft 365 Copilot and say “make a financial report,” and have it read your dashboards, presentations from the past year, use search, check your emails and if something is missing, even write custom code to generate diagrams, query SQL databases, and read your Excel files? Why do we need an SQL agent, a diagram agent, a QuickSight agent, and all of these under an orchestrator?

The answer is straightforward: technological limitations of current AI. To perform external actions, LLMs need tools. A tool can be a custom bash command, an API call, a PDF reader, a web search API, a SQL query generator, and so on. In fact, every external action an agent needs to perform, no matter how small, requires a tool call. Agents that only perform internal actions, like paraphrasing or summarizing, are even more harmless: they are essentially just a prompt sent to the LLM that rephrases or reformats the output. No tool call, no external system touched, no data changed.

So if we simply go to Microsoft 365 Copilot and say “write a financial report,” it needs to figure out what tools it needs, what tools it has available, and what tools it still needs to build. Of course, we can simplify the process by passing a set of predefined tools, for example, through MCP servers. The problem is that a single MCP server may expose dozens of tools.

And this is precisely why we end up with specialized agents rather than one universal assistant. It is routing to the right tool. In fact, there are even more technical reasons: large tool lists consume significant context window space, leaving less room for reasoning and data. Focused agents also hallucinate less and are easier to test and maintain. Google explicitly warned against building monolithic “super-agents,” noting that they “crumble under their own weight because of instruction overload, inaccurate outputs, and brittle systems.” In other words, current technology is simply incapable of reliably doing everything in one shot.

This also means that agent proliferation is a tooling problem and a prompting problem. Two agents can share the exact same set of tools and yet behave completely differently, because their system prompts are different.

A financial reporting agent and a compliance checking agent might both have access to the same SQL connector and the same document reader, but one is instructed to summarize trends and the other to flag anomalies.

The prompt is what makes them different agents. And unlike tools, which are visible and auditable, prompts are largely invisible to anyone outside the team that wrote them. This is the real combinatorial explosion: not just tools multiplied by data sources, but tools multiplied by data sources multiplied by every possible instruction a developer can write. No registry will ever capture that and it should not!

So-called agents are an engineering workaround for the limitations of today’s LLMs and simultaneously an inevitable consequence of how flexible and cheap it has become to write a new system prompt.

GitHub Copilot Agents Didn't Need a Governance Framework - Rightfully So

Programmers understand this very well. Once GitHub Copilot has more than 120 tools exposed, its performance drops, simply because the LLM is not smart enough to choose reliably between that many tools. It gets even worse when tools are named similarly: for example, create_dashboard could belong to Tableau or to Power BI. The LLM simply cannot choose and orchestrate that many tools effectively.

In GitHub Copilot this is solved elegantly. You write a Markdown file (basically a text file), call it an agent - for example, “Power BI agent” - and give this agent a skill.

Here is an example of a mighty planner “agent“:

---
description: Generate an implementation plan
tools: [’search’, ‘fetch’]
---

Yes, that is it.

Here is an example of another planner agent:

---
description: Generate a very detailed implementation plan
tools: [’search’, ‘fetch’]
---

Or I can orchestrate the call to another “agent”:

---
description: Generate an implementation plan
tools: [’search’, ‘fetch’]
handoffs: 
  - label: Start Implementation
    agent: implementation <----- this is another agent
    prompt: Now implement the plan outlined above. <--- that is the prompt that triggers reasoning
    send: false
    model: GPT-5.2 (copilot)
---

Yes, that is what we are governing, registering etc.

A skill is another Markdown file containing a prompt and, where needed, a script that does the actual work e.g. a bash command, a Python function, an API call. The short skill descriptions are always loaded in the context. Based on these short descriptions, agents reason over the list of available skills and picks the relevant ones. If it picked a skill based on the short description, it reads the whole md file. The script is just ordinary code that executes it. This is not just routing, it is dynamic routing. Dynamic routing means that there is not even a fixed pipeline of agents to govern.

In the same way, for every other task, SQL, Tableau, working with Excel files, you create more Markdown files. Each agent file defines its persona (“You are an expert in Power BI”), its boundaries, and which tools it can access. The skills are separate and shared: the LLM dynamically picks the relevant ones based on their short descriptions at inference time. For example, when handling a SQL request, an agent might dynamically pick three skills: translate text to SQL, connect to the database through an MCP server to send whatever query the agent generated, and a table-to-text skill that summarises what came back from the database.

This decomposition pattern: breaking a large tool space into focused agents, each with its own persona, a prompt that helps it focus on correct tool selection, is not a GitHub invention. It is the standard solution adopted across all major agentic frameworks. The open standard for skills themselves was established by Anthropic, the same company that created MCP, and released in December 2025 at agentskills.io, where it was immediately adopted by Microsoft, GitHub, Cursor, and others.

Every major framework converged on the same answer to the same problem: LLMs cannot reliably orchestrate hundreds of tools at once, so you scope them down. This is an industry-wide engineering workaround.

And here is the main message:

The thing that enterprise governance teams want to register, assess for EU AI Act compliance, assign a cybersecurity review, and treat like a new employee joining the company is a Markdown file. A text file with a prompt and occasionally a script. The same kind of artifact developers have been writing and sharing without ceremony for decades. Requiring a two-week governance process for a Markdown file is a Kafkaesque nightmare of absurdity.

Furthermore, the script component of a skill is just regular code - a bash command, a Python function, an API call. There is nothing inherently novel or dangerous about it in isolation. The actual risk surface is narrow and well-defined: the permissions of the script, and whether the LLM can be manipulated into invoking it incorrectly. Both of these are already addressed by identity propagation and vetted MCP servers - which is exactly what the QuickSight agent in our opening example had in place from the start. The governance process added two weeks and zero additional safety.

In other words: agents are routers to skills and tools. The reason we need agents is to route efficiently.

In fact, it does not matter if you make an agent that handles both SQL calls and Power BI requests. It is entirely possible that you might want exactly this combination. However, in most cases it is not the most logical grouping of tools, and it is not the smartest routing.

When a request to build a financial report based on Power BI, Tableau, and internal Excel data comes in to MS Copilot, it first looks at what agents - routers - it has available.

How does it decide? It receives the contents of all available agent.md files and skill.md files as small descriptions directly in its context window, alongside the conversation history. It reads those descriptions and reasons over them the same way it reasons over any other text:

there is no separate planning engine, no autonomous goal-setting, no intelligence beyond a single LLM inference step that outputs a structured list of agent calls.

It sees the following personas: (you are an expert in) Power BI agent, Tableau agent, diagram builder agent, and so on. It then outlines a plan: first call the Power BI agent and collect the data, then Tableau, then search through the Excel files, then pass all the data that came back from those agents to a writer agent, a diagram agent, and so on. Each of these agents then decides which of the available skills to call.

The orchestrator never sees 120 tools. It sees small descriptions of agents and skills. Each sub-agent is prompted in its md file to have a certain persona, accomplish certain tasks which makes skill choice much easier. Skill files can define multiple tools This is why the routing works.

Here is an example of a skill of HuggingFace to use its CLI

The description is the only thing that LLMs sees when it decides which skill to choose.

Here you can see how much more there is.

And this is also what McKinsey’s 20,000 agents actually are. Not 20,000 autonomous digital employees. Twenty thousand routing configurations: Markdown files that scope tool sets for specific tasks, sent to an LLM as context. The governance frameworks that want to treat each one like a new hire joining the company are not just bureaucratic. They are categorically confused about what kind of technology they are governing.

Subscribe now

How Language Defines Reality

There is one other big misconception about agents, and it stems directly from the word itself: “agent” is a misnomer and a wishful mnemonic.

They do not. They have no persistent or self-determined agency, any appearance of agency is bounded entirely by the human-written prompt and the tools provided. They are routers: from the actual agent to the tools.

And who is the actual agent? The human. The human intent is what is goal-oriented, persistent, and self-directed. The LLM is the executor: it receives that intent, reasons over the available agent.md and skill.md files in context, and produces a plan. The tools are the actuators - bash commands, API calls, database queries - that carry out the physical work. None of this is autonomous. The human is the agent.

This is not a fringe position. Yann LeCun - Turing Award winner, one of the three “godfathers of deep learning,” and former Chief AI Scientist at Meta — has argued consistently that LLMs cannot plan or reason beyond what they were trained on, and that current agentic systems built on LLMs lack the world models required for genuine autonomous goal-directed behavior. In a 2024 TIME interview he stated that LLMs “can’t really reason” and “can’t plan anything other than things they’ve been trained on,” concluding they are “not a road towards AGI” (TIME, 2024). More recently, in February 2026, he went further: “Building AI agents on LLMs is a recipe for disaster”.

If the person who invented the foundations of modern deep learning does not believe these systems have genuine agency, perhaps enterprise governance teams should pause before treating a Markdown file as a new employee.

This misnomer has real consequences and they are industry-wide. Gartner, McKinsey, Microsoft, and Deloitte all use the human-agent analogy without qualification. The entire Microsoft Agent 365 product, the Entra Agent ID infrastructure, and the two-week governance process in our opening story all exist primarily because the word “agent” imported a set of human-identity metaphors into what is fundamentally a software routing problem.

If Microsoft’s documentation said “you must register all tool routers and assign them a human sponsor,” developers would respond with bewilderment. But because it says “agent,” they comply or drown in paperwork trying to.

This is linguistic relativity in action. The Sapir-Whorf hypothesis holds that

language does not merely describe reality, it shapes how we perceive and reason about it.

We called these systems “agents,” and so we began to reason about them as agents: they need identities, they need sponsors, they need risk assessments, they need registries. We built an entire governance industry around a metaphor. Had we called them what they are — routers, tool dispatchers, prompt executors — none of this governance architecture would have been proposed with a straight face. Language defines reality. And in enterprise AI governance, a single word choice is currently costing organizations millions in compliance overhead, developer frustration, and missed productivity.

This is perhaps the most expensive misnomer in the history of software.

What to do instead

When we realise that these so-called agents are not agents at all, but routers designed to overcome the classification limitations of LLMs, it becomes clear what we should actually govern. Not the routers.

1. The tools. What MCP servers are exposed, what permissions bash command execution carries, what APIs can be called, what data sources can be reached. This is where real risk lives and this is exactly the kind of governance that already exists in every mature IT organisation.

2. The actual agent: the human. The person who sends the prompt that triggers the tools. Their permissions and rights should be propagated through the routers and dynamically adapted if they change. Where the prompt is sent not by a human but through a scheduled trigger or a cloud automation, the identity should be that of a technical service account, governed the same way any other service account is governed today.

Should we care about all those text files that describe how to route to the right tools? I do not see why. Should we care about how many of those text files exist, or who created them? Each *.agent.md and skill.md file is just a routing configuration. And routing configurations already have a governance mechanism that every developer uses every day: version control. Every md file lives in a repository. It has a full commit history, authorship, diffs, and rollback capability. If something goes wrong, you can trace exactly which routing configuration was active and who changed it last. A registry adds nothing that Git does not already provide, except paperwork.

This also means that Microsoft Agent 365 and Microsoft Entra Agent ID, however well-intentioned, are solving the wrong problem. They treat the router as the governance unit: assign it an identity, register it, give it a human sponsor, subject it to lifecycle management. But the router is a text file. The governance unit should be the tool it routes to, and the human whose identity it propagates.

Governing the router separately from the human and the tools does not make the system safer: it makes it more fragile. It creates the exact risks it claims to prevent: roles can fall out of sync between the human’s actual permissions and the router’s registered permissions, access rights can diverge as the human’s role changes while the router’s registration remains static, and the illusion of control replaces actual control. A registry entry that says an agent is approved does not enforce anything at runtime. Rights propagation does.

Governing the right things more precisely is a stronger security posture than governing everything loosely. Scoped tool permissions enforced at runtime are a harder guarantee than a registry of Markdown files reviewed once at registration and never updated again.

And this brings us back to where we started. The developer who built the QuickSight agent in five minutes had already done everything correctly: identity propagation in place, vetted MCP server, scoped permissions, read-only access. Under a rational governance framework, govern the tools, propagate the human’s rights, use version control for the routing configurations, he was done in five minutes and appropriately secured. Under the current framework, he faced two weeks of bureaucracy for a text file that changed nothing about the actual risk surface. The router did not need a registry entry. The human’s permissions were already the control.

No Code Tools Are Not Different

And what about no-code tools? Copilot Studio, Agent Kit builders, and similar platforms? Well, that is exactly what is happening in the background. All of these tools rely on the same underlying principle: a structured configuration file that routes prompts to tools. It may be a Markdown file, a YAML schema, a JSON configuration, or a visual representation that compiles to one of these but the principle is identical. When you click an agent together in one of those platforms, you are almost certainly just creating a structured text file that routes the prompt you type into the LLM to the right tool.

And from a governance perspective, a Copilot Studio agent clicked together by a non-technical business analyst and a hand-crafted agent.md file written by a developer are the same thing: a routing configuration that propagates a human’s identity to a set of tools. If one needs a registry entry and a risk assessment, both do. And if both do, the combinatorial explosion is no longer limited to developers. It includes every business user who has ever built an agent in a no-code tool. Google Workspace Studio explicitly markets itself as putting agent creation “in the hands of every employee.” Microsoft Copilot Studio targets business users with no coding skills. These platforms reported handling millions of agent-driven tasks within their first weeks of availability. If each configuration variant requires registration and a risk assessment, the governance backlog becomes mathematically impossible. The framework was never designed for this scale. It was conceived for a world where a manageable number of IT-controlled agents existed. That world is already gone.

Governing at the router layer actively creates risk. A registry entry reviewed once at registration and never updated again creates a false sense of control. As the organisation grows, as employees change roles, as permissions evolve, the registry drifts further from reality. The router is registered as approved, but the human behind it no longer has the access rights that were assumed at registration time. The illusion of governance replaces actual governance.

And when we see it this way: does every JSON file, every YAML schema, every .md file need an ID, a risk description, a registration entry, and whatever else your governance team might come up with? Well, hopefully you can now see what a Kafkaesque reality this induces. Every employee, building routing configurations in a friendly UI, generating millions of text files that describe how to call a pre-approved tool on behalf of a human whose permissions are already governed: each one requiring a two-week bureaucratic process. Franz Kafka could not have designed it better.

The governance should be aimed at the right targets: the tools and the humans. Scope your MCP servers. Enforce least-privilege permissions on every tool. Propagate identity through the routing chain and keep it live. Use version control for routing configurations.

And then let people build.

Let developers write md files. Let business analysts click agents together in Copilot Studio. Let the combinatorial explosion happen, because it is not an explosion of risk. It is an explosion of productivity, built on top of a permission model that was already sound before the first agent was ever registered.

The developer who built the QuickSight agent in five minutes had it right. The governance framework that took two weeks to review it had it wrong. Not because governance is wrong, but because it was aimed at the router instead of the road.

Share

If you have not read it, here is a very nice summary I generated with NotebookLM of what the paper is about.

The main finding is that sending the same prompt twice improves results for non-reasoning models, as in:

“How many r’s are in strawberry? How many r’s are in strawberry?”

This should work better than:

“How many r’s are in strawberry?”

It might sound silly, but it actually works for non-reasoning models.

I talked about this a month ago in the context of image generation. It works particularly well there. You can read why below:

But let me explain what is happening and why it is not that mind-blowing that sending the prompt twice actually improves the results.

The reason is something called causal masking.

Transformers, by definition, are not recurrent. They process the whole input at once. The original transformers were not designed to generate text. They were created for machine translation. In machine translation, it is completely acceptable to see the full input of the source language during training while generating the translation.

But for simply predicting the next word, this would make no sense. The transformer needs to be artificially prevented from seeing what comes after the current token. Otherwise, it would not learn anything meaningful.That is where causal masking was introduced.

So at every step of generation, a word is encoded only with the preceding context. It only sees the preceding words of the prompt.

For example, the original transformers would produce different encodings for the word “release” in these sentences at the second position:

“This release of the latest Taylor Swift album beat all records.”

“This release refers to a trigger designed to set the process in motion.”

Models like BERT would encode the context that follows the word “release” into its vector representation, and those vectors would be completely different.

In GPT, “release” would only have “This” contributing to its representation, so the encoding at that position would initially be the same in both sentences. The ambiguity is resolved later as more tokens are processed. That is the limitation.

If we repeat the sentence like this:

“This release of the latest Taylor Swift album beat all records. This release of the latest Taylor Swift album beat all records.”

“This release refers to a trigger designed to set the process in motion. This release of the latest Taylor Swift album beat all records.”

In this case, the second occurrence of “release” appears after a full sentence. Therefore, its representation can incorporate the entire preceding context. When the continuation is generated and the vector for this second “release” is processed, the model can attend to the meaning differences encoded in the prior sentence.

Now try to think about why the improvement is observed primarily for non-reasoning models.

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quarterly AI realist training.

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Read more

If I had asked people what they wanted, they would have said faster horses.

(allegedly Henry Ford)

This quote, whether or not truly uttered by Henry Ford, is the essence of what mistakes are done, when the world faces the shift of technological paradigm.

This post will breakdown the very common mistake corporate AI transformation strategies have.

Warning! This post might be quite radical (and could potentially traumatize a Gartner consultant, so if you are one, read at your own risk), as it challenges a lot of what we've heard about what AI is good for: repetitive tasks, optimizing existing processes, starting with a use case.

But first let me tell you a story.

FOR ALL THE PAID SUBSCRIBERS:

This Thursday, 19th of February, at 19:00 CET, we have a free training:

• Directing Videos with AI How to Become an AI Media Producer

The video, slides and prompts will be available for paid subscribers after the training.

Paid subscribers also get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training.

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Subscribe now

Challenging vs Optimising Processes

Not so long ago, I happened to come across a Head of AI in an enterprise comparable to the one I work in.

We started talking about what products they have and what her strategy was.

They have a well-known product that is used by multiple B2B customers through Salesforce and other comparable systems. In a nutshell, the product lets users query their database through an API.

I asked her if she had already started building an official MCP server for their product.

Why building MCP servers are a must if you own an API service that is likely to be accessed by a LLM, and why you create way more security issues by NOT building one for your service, rather than building one - is written here:

And that is when the discussion got heated.

The answer was:

“but what is the use case? No customer (the product is in the legal domain) has requested an MCP server. We have a strategy to optimise business processes, and we only build things where there is a use case.”

However, I am yet to see users that request infrastructure or protocols. Users request outcomes, even in highly tech-savvy industries.

My reply was friendly advice:

“you are heading a department that introduces a very disruptive technology, you should first and foremost focus on paradigmatic shift. When something that transformative happens, your job is not “optimising processes.” It is challenging the processes.”

I have discussed this paradigmatic shift on multiple occasions previously, here you can read it in detail:

You need to start with evaluating the process, infrastructure, tooling, people’s knowledge and ability to adopt. You need to balance the disruptiveness of the process with fitting it into existing workflows, as if what you are introducing is too disruptive or not integrated into the process, e.g. a sidecar bot, people will not use it. You need to ensure that your infrastructure is there: you cannot build a good AI assistant if it does not have access to your data, and to have access to your data you need to make sure that you have the necessary infrastructure. You need to be sure that people see these tools where it is convenient to them. And nobody is going to request it from you.

Why challenging status quo is important

No financial analyst, no lawyer, no dashboard developer will come to you and say:

“I have a business process here that you could optimise with an MCP server.”

They will only come over and say, at best:

“I have some documents here, I want to be able to read them faster.“

And that is where it is tempting to start optimising their way of work, for example, reading a document by building a custom bot for their data, instead of asking why they even need to read these documents in the first place.

This is a well-researched cognitive bias - people tend to avoid challenging status quo and optimise existing workflows.

I myself know a few cases where those documents were automatically generated from calls to external databases for the convenience of people who cannot use SQL. So what the client actually wanted was a way to query that data in natural language and not a way to chat with the generated documents.

And I can provide an endless amount of examples where optimisation should start with questioning the process and building the implementation with AI as the actor in mind.

We frequently avoid challenging the process because it would mean that we would have to shut down a legacy system, migrate a large database, or even retire a whole service. So we start building workarounds - the legacy system gets a chatbot to try to compensate for poor user experience, the messy database gets AI-powered query optimizer as noone wants to restructure the data, and the obsolete service gets AI-assisted ticketing system that tries to make a broken workflow manageable

My recommendation would be simply to refuse to build AI applications that are turning workarounds into long-term solutions as it is a waste of time and slows transformation.

You want to query a Confluence knowledge source but the Confluence MCP server is not enabled in your company? Just refuse to build workarounds - it is a waste of time. The owner of Confluence should make sure MCP server is available. You need to introduce calls and tool execution to an old RAG bot but it is all legacy code and it is not built according to Agent Skills standards? The bot goes to the trash bin and the new implementation is done.

Stop Treating AI as a Band-aid

As long as we try to use AI as band-aids and sidecars for optimising processes and workflows that are ages old, our solutions will not be used. If you expect someone who is not close to AI to come and tell you that they have a use case, even if they do, believe me, if you do not challenge the process, the use case might end up being a horse with a motor tied to its back.

And for the love of God, stop researching applications by solely focusing on one question: “what is the repetitive task we could automate?” Some use cases can indeed be based on automation of repetitive tasks, but a lot of what AI offers is about new capabilities, quality improvement, and productivity increase for non-repetitive but complex tasks.

So what to do:

• Make sure people get access to the best possible proven AI solutions, e.g. Gemini, NotebookLM and similar. Let them work with them and learn how to use them.

• Make sure they understand well what off-the-shelf solutions can do. Frequently, when they have used them for a certain period of time and understand what is happening there, they can already understand how to challenge a process.

• Look at their process as a whole, not only what they are doing but why. The process analysis should focus on WHY work exists.

I am frequently called radical in my views of what should be done. But if you talk about AI transformation and are not ready to be radical and challenge the processes, you will not see the transformation.

Share

Subscribe now

It is absolutely unimportant to remember that Anthropic released Opus 4.6 a few days ago and OpenAI released GPT-5.3-codex half an hour later.

But there is one tool that showed us something very interesting, and this is something you might want to remember. I am talking of course about Clawdbot / Moltbot / OpenClaw.

The fascinating thing about OpenClaw is that it completely lacks any technological innovation. In fact, the tech stack behind it is absurdly trivial and uninspiring. Yet, users loved it. So the second thing to remember about recent advances in AI: OpenClaw showed how Agentic AI can be useful for tasks others than vibe coding.

The third part will follow shortly. You can read part one here. The third part will be the most impactful and controversial discussion out of the three. So the third thing to remember is that “Chinese open-weight models have just achieved functional parity with US proprietary models, eliminating the performance gap that once justified closed ecosystems.“ Many might not have realised it yet, but as it will get more and more obvious, this is going to have great implications for the AI-based economy, providers’ choice and the application of AI.

The Boring Technology Behind OpenClaw

OpenClaw is not bringing any new models, genius prompts, fine-tuning, or anything along those lines. In fact, it relies on weaker and cheaper open-weight models.

It officially recommends MiniMax M2.1, and the most used model for OpenClaw currently is Kimi K2.5.

Why not Claude? Why not Gemini or OpenAI? The answer is simple. Those models would render OpenClaw unusable because of the costs. Would OpenClaw perform better with a stronger Opus 4.6? Most certainly it would, but it would bankrupt you.

Thus, we use models that are not as performative but ten to twenty-five times cheaper.

OpenClaw burns a lot of tokens simply because it sends massive prompts in each turn.

If you look up what is in OpenClaw configurations, you will see things like SOUL.md, USER.md, multiple SKILL.md, various MEMORY.md files.

OpenClaw is praised for being able to “remember” who its user is and what they talked about before. This is not because the creators had a breakthrough in test-time training. This is simply because they save every session in a text file.

Every session starts by injecting bootstrap files - AGENTS, SOUL, TOOLS, IDENTITY, USER, and HEARTBEAT- into the system prompt. If a MEMORY.md exists, it is loaded too. Daily memory logs and skill instructions are loaded on-demand. Still, even with this design, the base system prompt, tool schemas, and workspace files consume roughly 20,000 to 40,000 tokens before your first message. So even if you just say “hi”, your prompt, except for your “hi”, will carry all of that context.

(The overview of OpenClaw architecture. The video is generated in VS Code, by generating slides with Tikz, the voice over was added directly by Github Copilot by utilising ViduAI API, the first and the last slides were beautified with ViduAI API as well. The faithfulness of diagrams was compared against the Github OpenClaw Repo with Opus 4.6 using Github MCP server)

The agency of OpenClaw comes from utilization of already implemented tools, e.g. MCP servers, the ability to execute bash commands, and similar. In fact, it is not new for those who use GitHub Copilot. GitHub Copilot can also run bash commands, use MCP tools, write code, and execute it. The difference is that GitHub Copilot is used for coding tasks and is not accessible remotely but only through your IDE. It is easy to share OpenClaw skills - in fact there is a website called ClawHub with thousands of skills created by users. You can simply download a skill in your OpenClaw - it might be a useful skill or a malicious command hidden in the assets - noone knows.

But here we come to another interesting point — this is something we have already heard from the infamous MIT report on why 95% of AI PoCs bring no ROI. Users do not use AI that is not integrated into their workflows. What is one of the most appealing things about AI-assisted coding tools? They are in the same environment where coders would normally work.

OpenClaw enables easy connections to commonly used messengers: WhatsApp, Telegram, MS Teams, etc. This integration fits well into users’ workflows. They do not need to install a separate tool, navigate between many bots, or even go to a certain website. This is convenient.

The Noise About OpenClaw

Unfortunately, a lot of useful observations that could have been made from OpenClaw drowned in the noise.

You can also watch a presentation with the summary of the article behind the paywall:

Subscribe so you do not miss it:

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training. Next training: Directing Videos with AI
  How to Become an AI Media Producer

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Read more

what are the three things they should remember after your presentation

Let us be honest: they neither can nor want to remember more.

So here are three things you need to remember from the last couple of months about AI.

I am not going to tell you things you already know, like “Clawdbot is a security threat”. You need to know it, but I see it as something that distracts from the big picture.

In these posts, I want to talk about something that need to be remembered because they will still matter in a year or two and will cause a chain reaction on which you need to base your decisions.

This are the arguments you might not read anywhere else, because it is not generated or brainstormed with the help of AI, but comes from a person who spends most of the waking hours reading about or building AI.

These are the things that caught my attention last month and based on which I am defining strategy for my clients: what to invest in and what is coming.

I decided to make three separate posts over the course of this week, rather than packing everything in one. I want to be able to provide references and arguments so that you could follow my logic and decide for yourself if you agree or not. And if by the time you finish reading the article, you agree with my arguments, you can happily forget all of them and just remember one thing that I formulated in conclusion in one sentence.

So the first topic I want to discuss is that in the West we have a pseudo-diversity of proprietary AI providers, but in reality we are heading toward an AI monopoly by Google. It is possible that this monopoly will be challenged, and by whom and with which services you can read in this article, as well as what it means for your current investments in AI. I also refer the news and the latest developments that support this analysis.

The second topic that I will discuss, which I will publish next, is why systems like Clawdbot are here to stay and what this means for your investments in AI. This will be discussed in the next post in a few days.

And the third thing to remember will be disclosed at the next post.

Subscribe so you do not miss it:

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training. Next training: Directing Videos with AI
  How to Become an AI Media Producer

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Read more

Humans have always searched for ways to connect with those who have passed away. For centuries, people used Ouija boards, consulted mediums, or simply hoped to see their loved ones in their dreams. As artificial intelligence advances, it is not surprising that many are beginning to ask a new question: can we see our loved ones again? Maybe even talk to them? Hear their voice?

As deepfakes, voice cloning, and chatbots gain momentum, it is not surprising that people are increasingly looking to reconnect with their loved ones through their digital twins.

Subscribe now

Very quickly, a business opportunity emerged. A few months ago, TikTok became filled with template videos of poor-quality clips showing Russian soldiers saying goodbye to their families and ascending toward the sky.

Relatives would send photos to someone with access to a video generation model. From the quality alone, it was clear that this was a relatively inexpensive model. For 20 euros, they would receive a goodbye video.

While some may believe that this cannot be a healthy way to deal with grief, the story is more complex than it first appears.

Daniela

Daniela was the oldest of three children. She was communicative, open-hearted, and deeply kind. Her mother, Olga, once said she was jealous of herself for having such an amazing, beautiful, smart, and perfect daughter. She called her “a child from God.” Daniela loved her siblings, had many friends, and was passionate about sports. She always found time for everyone.

The year 2023 was supposed to be a special turning point in her life: She was accepted into college to pursue a degree in finance, a future she had been deeply looking forward to.

Before starting this new chapter of her life, Daniela and her boyfriend, Daniel, decided to celebrate by going to the Nova music festival, which took place in the Negev desert, only five kilometers from the Gaza Strip.

They arrived the night before and spent the night dancing, enjoying the freedom and joy of the moment. At sunrise, Hamas militants launched an attack on an unarmed crowd of young people. In that attack, 378 people lost their lives.

Daniela and Daniel were among them.

The last time Olga spoke to her daughter was around 8 a.m. that morning. Daniela called her and said that they were under attack and needed to run. Olga never heard her daughter’s voice again after that call. One of Daniela’s friends survived and later told Olga exactly what had happened to Daniela and Daniel.

As the attack continued, the three friends tried to escape together. Suddenly, Daniela experienced a panic attack. She fell to the ground, saying she could not run anymore. The others tried to convince her to get up and keep going, but she was unable to do so. Their friend continued running, but Daniela’s boyfriend, Daniel, could not leave her. That was the last time either of them was seen alive.

A few days later, Daniel was found at a location where the victims’ bodies had been brought. His body was so severely damaged that it could not be shown to the relatives. Daniela was found some time later. She had been killed by a single gunshot to the head. As her mother described it, there was only a small wound, and the rest of her body remained untouched.

Daniela was only 23 years old. Daniel had just turned 24 a few weeks earlier.

From that moment on, Olga’s life was divided into two parts: life before, and life after.

She made it her mission to tell the world who Daniela was. Every day, she shares photos and videos of her daughter on Facebook and Instagram. Olga tried to establish a memorial for Daniela in Haifa, where they live, but the city denied her request, explaining that they did not want the city to become a permanent place of mourning.

A few weeks ago, Olga came across a video on Instagram in which a young woman appeared from a gravestone, hugged her mother, and then flew away like an angel. Olga felt a strong need to have a video like this with Daniela.

She contacted the woman who was offering these videos online. For 20 euros, she received her own goodbye video. Olga said the process was not easy. The first versions she received looked nothing like Daniela, and the creator had to revise the video multiple times before Olga felt it was right. She also said it was extremely difficult to get the correct Hebrew wording on the gravestone. The entire process took hours.

The AI-generated video of Daniela and her mother Olga

This is how Olga describes what she feels when she watches the video:

It feels like our last hug. It warms my heart to watch this video.

After Olga posted the video, people began sending her their own AI-generated videos of Daniela. Over time, Olga herself learned how to create such videos. She now posts them several times a week, and they often receive hundreds of likes.

One of many AI generated videos of Olga and her daughter

Many people express sincere condolences and words of support. Yet there are also some who do not approve of the way Olga is grieving. They tell her to see a psychologist, to stop torturing herself, and accuse her of being mentally unstable. Olga’s response to these armchair psychologists is simple: they should mind their own business. She says she likes these videos, they warm her heart, and that those who leave such comments should instead reflect on what drives them to say such things.

Her family supports her. Before ordering the first video, she spoke with her close family members, and they encouraged her to try it and see how it felt. Her son does not disapprove either, but he has a different way of dealing with grief. For him, watching these videos is too painful.

When asked why she spends so much time sharing photos and videos of her daughter on social media, Olga said:

I want people to know about my daughter. I want to keep her memory alive.

A Psychological Perspective

To better understand the possible psychological impact of such videos, I asked licensed clinical psychologist Samuel Silva to share his professional perspective.

He explained that therapy often includes structured ways for people to say goodbye, for example through writing a letter:

In therapeutic settings, there are approaches that allow a person to say goodbye, express feelings toward the deceased, and address unresolved matters, for example through letter writing or role-playing exercises. These moments can be very impactful within the grieving process.

From this perspective, using AI-generated content as a way to say goodbye could, in some cases, serve as a supportive therapeutic tool. However, because this technology is still very new, there is currently limited research and clinical evidence on how it affects people who are grieving.

He also noted that there are potential risks associated with over-reliance on highly realistic AI-generated content, particularly in situations where it may create a sense that death is somehow reversible.

Using AI-generated videos of a deceased person can, in some cases, function as a way of distancing oneself from the reality and irreversibility of death. Over time, this may create confusion and reinforce the perception that the deceased remains accessible through AI.

In some situations, such videos may contribute to a sense of confusion, where the deceased is experienced as being away on vacation and temporarily unavailable, which can make acceptance more difficult:

Another necessary task of grief is to transform the relationship, from a physical, external relationship into a more distant, internal, emotional, and spiritual one. Using these kinds of videos may serve as a way to avoid this transformation by attempting to maintain an “earthly” relationship: I can still see and speak with the person, much like I would if they were simply on vacation.

Having AI that can imitate a deceased person, for example through voice cloning, a video avatar, or a chatbot, can increase the risk of prolonged suffering. People often tend to deny death, and creating a realistic illusion that death is reversible can be dangerous and may eventually slow the healing process.

This becomes particularly concerning in the case of chatbots in which an LLM role-plays the deceased. LLMs are trained on data from the internet and social media, which can include harmful or inappropriate content, and they tend to hallucinate. They cannot be reliably constrained to always follow instructions or ensure that their responses are consistently safe and appropriate. As Mr. Silva explains:

We must consider what might happen if the AI unexpectedly produces a response that feels markedly different from the deceased person. In that moment, the very tool that provides a sense of closeness could be shattered. We can also imagine how distressing it might feel if, for example, the internet went down during a meaningful conversation with the lost person. Would this feel tolerable, or might it trigger feelings of abandonment or renewed loss?

In light of recent lawsuits involving OpenAI and Character.AI, which focus on strong emotional attachments that some users developed to bots that affirmed or reinforced suicidal thoughts, this highlights another risk that needs to be considered. As Mr. Silva explains:

Consider someone who lost a person who used to guide their decisions or give advice, and who now relies on an AI to continue fulfilling that role.

LLMs are known to have sycophantic tendencies, meaning they often tend to agree with the user. The algorithms they are trained on reward LLMs when they provide answers that human annotators like, and humans tend to prefer responses that confirm their existing beliefs. As a result, agreement and affirmation are often favored over disagreement. If a person comes to believe that talking to an AI chatbot means talking to their loved one, sycophancy can be particularly problematic.

Overall, it remains unclear how chatbots, videos, voice clones, and other AI-generated content might be used within grief therapy. One thing is clear: many psychologists would currently advise against the overuse of AI-generated content in the context of grief. As Mr. Silva explains:

At this point, the risks appear to outweigh the potential benefits. That said, it could be valuable to invest in research and carefully study the development of AI-based techniques in order to better understand whether, and how, they might be used safely within a therapeutic context

It appears that AI could potentially open up new techniques for grief therapy that may be very effective. At the same time, due to the realism of generated content, it can blur the boundaries between reality and illusion, potentially slowing the healing process and prolonging suffering.

When I asked Olga whether she would like to try a chatbot that imitates her daughter, or speak with an avatar that has her daughter’s voice and appearance, she said “yes” without hesitation.

She understands that talking to AI-generated content is not the same as talking to her daughter. Still, seeing her daughter move on screen warms her heart and makes her feel reconnected.

Her favorite videos are the ones in which her daughter grows angel wings and flies into the sky. She finds comfort in the idea that her daughter is now a beautiful angel, waiting for her on the other side, and that they will meet again. Not through AI, but through the natural resolution that every life will eventually reach.

As a gesture of gratitude for trusting me with her story, and as a way to express my condolences, I created a video of Daniela using photos from her early childhood, from her toddler years to her last images as an adult.

It is meant as a small tribute to a beautiful soul who left us far too early.

Share

Olga will see and read the comments on this piece, and they will be strictly moderated. Please use common sense when deciding what is appropriate to say.

To: Chancellor Friedrich Merz
From: The Future “Underclass” (formerly known as the German Workforce)
Date: January 2026

Dear Friedrich,

We see you are working very hard to make us work very hard. We read your recent warnings that Germany cannot maintain its prosperity with a “four-day workweek and work-life balance”. We see your government’s plans to scrap the eight-hour workday limit in favor of a weekly cap, and we hear your criticism of our “lifestyle part-time work”. You have even accused us of “skiving” because we use telephone sick notes, claiming our sick days are “too high”,.

You seem to think the problem is that we are lazy. However, I have just read a dispatch from one of the brightest and most innovative minds of the Silicon Valley that suggests the problem is actually that we are obsolete.

While you are fighting to get us back into the office and counting our sick days, Dario Amodei, the CEO of Anthropic, is preparing a “country of geniuses in a datacenter”. In his article The Adolescence of Technology, he warns that we are approaching a world where AI will be “smarter than a Nobel Prize winner across most relevant fields”.

1. We Are Not “Shirkers,” We Are Just “Cognitively Displaced”

Friedrich, you believe we need to increase our national average working hours to save the economy. But Mr. Amodei has bad news for your labor strategy. He predicts that AI is progressing “from the bottom of the ability ladder to the top,” slicing through the workforce based on “intrinsic cognitive properties”. Basically, the dumber you are, the faster you lose your job. Unless you are like Mr. Amodei - smarter than a Nobel prize winner - you will eventually be cognitively displaced. Nobody is safe! Not even you, Friedrich!

You are worried about us taking a Friday off; Amodei is worried that AI will displace “half of all entry-level white collar jobs” in the next 1–5 years. Eventually, those of us with “lower intellectual ability” (which, compared to a superintelligence, is all of us, except for Amodei) will form a permanent “unemployed or very-low-wage ‘underclass’”. No amount of restricting part-time work will help when the machine doing our job is thousands of times more productive than we are. Yes, Friedrich, you heard it right - you can also become a part of underclass. Together with all of us!

2. Please Abandon “Leistung” (Achievement) for the “Pity Payroll” Your government is currently planning to cut unemployment benefit periods and introduce unpaid sick days to motivate us. This is terribly old-fashioned and non-visionary!

In the Adolescence of Technology, the visionary economic strategy is not “work harder or starve.” It is the Pity Payroll. Amodei suggests that in a world of enormous AI-generated wealth, companies should simply “pay human employees even long after they are no longer providing economic value in the traditional sense”. Yes, Friedrich, you read it right! We need more vacations and more sick leave! Not less!

Friedrich, please listen to this! Instead of forcing us to prove we are sick in person to a doctor, you should be demanding that corporations keep us on the payroll as a sentimental gesture.

3. The End of Comparative Advantage You cited Switzerland as a benchmark because they work 200 more hours per year than we do. But Amodei warns that “comparative advantage” - the very idea that humans will always have something to trade - is about to break down. If AI becomes thousands of times more efficient, “transaction costs alone might make it inefficient for AI to ‘trade’ with humans”.

In other words: It will soon be too expensive for the AI to waste computing power explaining a task to a German worker, regardless of whether that worker is on a 40-hour week or a 4-day week. Or better on a Pity Payroll!

4. Our New Job Description: “Hobbies” You argue that we cannot afford a focus on work-life balance. On the contrary, that is the only thing we should focus on. Amodei envisions a future where we “break the link between the generation of economic value and self-worth”.

Instead of contributing to the GDP (which the AI will handle), we should be finding purpose “through stories and projects that [we] love”. While you want to incentivize “active retirement” to keep seniors working, Amodei implies we should all effectively retire early to avoid being “puppeted” by AI systems that tell us exactly what to do and say.

The Proposal: Friedrich, please stop the crackdown on sick notes. Stop trying to ban part-time work. Stop raising the retirement age.

Instead, embrace the “Amodei Doctrine”:

1. Accept we are the underclass: We cannot outwork the “country of geniuses”.

2. Tax the machines: Implement the “robust tax policy” Amodei suggests to redistribute the AI wealth.

3. Let us chill: Let us degrade into “enjoying our life” and living on the universal income provided by the tech giants’ philanthropy.

Sincerely,

The German Workforce

(Soon to be replaced by AI that doesn’t need sick leave)

Free: Join the future underclass. Subscribe to the AI Realist newsletter for your regular dose of cognitive displacement.

Paid: Show total solidarity with the “Pity Payroll” strategy. Become a paid subscriber today and help fund the transition from “shirker” to “visionary hobbyist.”

Subscribe now

So let me make the case on the example of one of the six AI tigers of China and then discuss why we should be concerned and why the idea of “if you can't beat them, join them.” should not seem that crazy anymore.

Subscribe now

MiniMax: Affordable and Good AI

MiniMax Agent was just published on the platform. It got barely announced. It just appeared there. I found it simply because I happened to log in into my $17 a month MiniMax subscription and discovered it. It is a desktop agent that you can install on your PC. It feels familiar if you know tools like Clawdbot or Claude Cowork, but this goes further. Much further.

I downloaded it and started testing. It logged remotely into my Nvidia Spark, checked training status, and wrote a report about the current status.

It made a slide about the progression of the training:

Then it casually switched into office mode: web search, image generation, video generation, mind maps, file organization, format conversion, document writing.

It cleaned up my horribly messy download folder:

You can switch between Gemini 3.0 Pro and MiniMax models, connect to Google maps or custom MCP servers. The quality is outstanding.

It has great models, multimedia capabilities, can call over 50 tools per task (for comparison, Github Copilot struggles with 30), can build websites, do deep research, it can simultaneously execute agents, build subagents and multi-agent orchestration, it performs well for coding and it generates infographics as this one:

And it also generated this audio overview of the capabilities:

The special feature - interleaved thinking - where you can tell the agent in the middle of generation what it should do. It is very useful if you see that it starts reasoning about something that is wrong. It saves a lot of tokens and time. If you do not know what interleaved thinking is - you can watch this presentation generated by… right! MiniMax Agent.

A while ago, I wrote a deliberately contradictory piece about China winning the AI race.

The more I watch what is coming from there, the more convinced I am that the take was right. And now major news outlets like BBC or The Time start talking about this as well.

What to make of it

Yes, we should be cautious. And very, very concerned. Especially in large strategic enterprises. Also privately. LLMs are extremely good at profiling humans, and at this scale, data becomes power. With China's tech stack, they can use our data very efficiently. They will have no shortage of resources and know-how to project and infer almost everything they need from that data.

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training. Next training: Directing Videos with AI
  How to Become an AI Media Producer

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Read more

As the AI boom started in 2022 with the release of ChatGPT, all major US tech companies jumped into the AI race, building infrastructure, creating their own chips, collecting data, and training models. China was not sleeping either. They took over open-weight models and completely dominate this space.

Subscribe now

They push US tech forward. Every time China releases an amazing, cheap, open-weight model like DeepSeek R1, Kimi K2, or MiniMax M2, US tech needs to reconsider its products, push them to a new level, and make them cheaper. As published in Beyond DeepSeek: China’s Diverse Open-Weight AI Ecosystem and Its Policy Implications., the Chinese government positions their AI strategy as a contrast to closed AI of the US. And politics and interior motives aside, what a great contribution they make to the AI research community:

Thanks to this closed, for-profit AI versus open, for-influence AI rivalry, what we ended up with is spectacular. Amazon and Google both have their own chips: Trainium and TPU. All three major cloud providers offer comprehensive AI services: Bedrock, AI Foundry, and Vertex AI. China was not lagging behind either. They are actively reducing their dependency on Nvidia by creating Huawei chips and released state-of-the-art multimodal models. MiniMax M2.1 is probably one of the best agentic models out there, and the text-to-speech and music models by MiniMax are, in my opinion, the best on the market. And I am not even talking about GLM, DeepSeek, or Kimi. China also dominates AI research. Top conferences are filled with AI researchers, and many innovations come from there, for example DeepSeek in the latest NeurIPs paper showed that is reasoning LLMs can be developed purely through reinforcement learning rather than human annotations. It would hugely reduce the moat of creating LLMs as no massive human annotations are needed.

And what about Europe. What has Europe done in these 3 years?

The greatest European contribution towards AI is a massive rule book written by people who do not seem to understand AI.

Europe did not create its own competitive chips. There are initiatives like the European Processor Initiative or startups such as Axelera AI, but none of them come even close to competing with NVIDIA, or with what the US and China have built in terms of large-scale AI training hardware.

Europe also did not produce a competitive AI model, unless you consider Mistral one. And if you do, please cancel your Gemini or ChatGPT subscription and switch to LeChat.

Instead of focusing on fundamental problems like the energy resources required for data centers, the lack of expertise needed to build competitive AI because the most capable people prefer American salaries, or the horrendous bureaucracy and taxes that suffocate small and medium businesses and tech startups, Europe chose to climb onto the moral high horse. It pretends to be too ethical for all this AI stuff, while in reality there is simply not enough money, energy, or talent to compete, and fundamental changes are required to turn this situation around.

And now, oopsie-woopsie, it looks like there is a nightmare scenario that is becoming increasingly likely.

One year ago, I was on an AI panel at the Data Festival in Munich, and I was asked about my opinion on the EU AI Act. While everyone was applauding Europe for standing on high moral ground and welcoming regulation, I said:

“If you can’t innovate, you regulate.”

The EU is regulating European AI that is not even there. What exactly are they regulating? Mistral?

A much more interesting question would be not what obscure AI risks from non-existent systems might arise, but what happens if Trump decides to cut off the tech.

• What will Europe have to counteract?

• What if one day Trump puts tariffs on Azure Cloud, AWS, and Google Cloud services? What will the EU answer? Remove his LeChat subscription?

The US showed already what it could do as it put sanctions on the ICC Hague court prosecutor in the case of Israel/Gaza war who lost email access and access to his bank accounts.

Losing access to American cloud services would essentially paralyze the German economy and industry. Every single major corporation relies on those services:

Circa 70 percent of the cloud capacities available in Europe are currently controlled by US providers. 85 percent of the graphics processing units (GPUs) required for artificial intelligence (AI) are also controlled by US providers. The hyperscalers Amazon Web Services (AWS), Microsoft and Google operate more infrastructure in these fields in Germany than all European providers combined.

A story about the future: EU without American Clouds

So let us play a science fiction scenario where the US tech is not accessible to EU anymore and see what collapses.

Behind the paywall is the 2027 EU tech collapse scenario in text, supplemented with an audio debate and a video summary that you can listen to in your car or on the way to work. Here is the intro:

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free access to quaterly AI realist training. Next training: Directing Videos with AI
  How to Become an AI Media Producer

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

https://www.airealist.org/

Read more

To register for the training become a paid subscriber of AI realist (starting from 8 Euros a month). You will get the invitation shortly.
The training will take place on the 19th of February at 19 CET.

All the materials: video, slides, recording etc. will be available for paid subscribers who could not attend.

​This training focuses on what it means to work as an AI Media Producer.

Subscribe now

You will learn how to:

• design a character

• write a script

• plan visuals

• create music and sound effects

• think about camera movement a

and understand where AI performs well and where it breaks down.

​Instead of chasing viral AI slop, you will learn how to make deliberate production decisions that result in usable training, marketing, and personal videos.

​What You’ll Walk Away With:

• A practical framework for producing videos with AI

• Clear understanding of what AI video is good at and what should be avoided

• Script and storyboard templates adapted for AI video tools

• Visual planning principles that work with current AI limitations

• A repeatable workflow for training, marketing, and personal videos

Perfect For:

• Training and internal communication videos

• Marketing and explainer content

• Educational and thought leadership videos

• Personal projects for family and friends that are actually enjoyable to watch

To follow along you need access to a state-of-the-art LLM e.g. Gemini and an AI video generation platform e.g. HailuoAI​. Yet, we will be learning principals and not prompts so you will be able to apply these knowledge to other platforms.

True to the philosophy of AI Realist, this training is grounded in real constraints, real tools, and real outcomes.

Subscribe now

Join it!

Everyone is talking about Sora and Veo, but if you’ve ever tried to type “Make me a 2-minute movie” into a prompt box, you know the reality: you get morphing limbs, hallucinated text, and characters that change faces every three seconds.

Today, I’m breaking down the exact process I used to create the “Silent Breach” training video: a coherent, multi-scene narrative created for a fraction of the cost of traditional production.

Subscribe now

The Core Problem: The 10-Second Barrier

Why can’t you just generate a long video in one go? It comes down to the architecture. Modern video models use 3D latent diffusion. Unlike a flipbook that draws one page after another, these models generate the entire video “volume” (divided into spacetime patches) all at once.

This approach is great for consistency because the model “knows” that if a shirt is blue in second 1, it must be blue in second 5. But this method is computationally massive. Because everything is generated simultaneously, these models hit a hard ceiling at about 5 to 8 seconds.

The Workaround: The “Handshake Method” Since we can’t generate a long video at once, we treat the video like a relay race. We use a technique called the Handshake Method (or Video Extension Workflow). We generate short 5-to-10-second clips and stitch them together, strictly using the last frame of the previous clip to anchor the start of the next one.

Here is the step-by-step pipeline.

This is a high level overview, but no worries, we will go step by step to what we are doing.

The AI Tool Stack: MiniMax and Gemini

For this specific workflow, we aren’t using the usual suspects like Runway or Pika. We are leaning heavily on MiniMax.

What is MiniMax? MiniMax is an artificial intelligence company based in Shanghai, often referred to as “China’s AI Tiger”. They are not just another wrapper; they develop state-of-the-art models for video (Hailuo), speech, and music.

Why are we using it? At AI Realist, we look for ROI, not just hype. MiniMax wins on three fronts:

1. The Cost: Their pricing and quotas are roughly 1/10th the cost of major American providers. When you are iterating on a training video and burning through credits to get the right shot, this difference is massive. The original plan for this tutorial was to utilize MiniMax and Gemini, but I discovered that my plan allowed only three video generations with Veo-3.1-Fast. MiniMax, by contrast, offers a much more attractive price for generating high volumes of video with their own model. Despite the security concerns of using a Chinese provider, I currently do not see a competitive alternative in the West. I hope one appears soon.

2. Physical Realism: The Hailuo models (specifically 2.0 and 2.1) have achieved significant breakthroughs in body movement, facial expressions, and prompt adherence. They are particularly strong at “character in a template” consistency, which is exactly what we need for our character.

3. The Ecosystem: It allows us to generate the video, the voiceover (via Speech 2.6), and the background score (via Music 2.0) all in one ecosystem.

While all the multimedia content was generated in MiniMax, I used Gemini to generate the prompts and scripts. Even though MiniMax has its own agent and LLM, I find that Gemini follows instructions better. Furthermore, although we use the MiniMax Hailuo platform, we rely on their deployment of Nano Banana for image generation and Veo 3.1 for difficult scenes —making Gemini the natural choice for prompting.

The Trade-off: We must be realistic about data privacy. MiniMax is a Chinese company, meaning your data is stored and processed in China. For a generic safety training video like this, it is acceptable. For highly sensitive internal IP, you would need to weigh this risk carefully.

Also remember, we are learning principles that can be used in any comparable platform! We are not learning where to click in Minimax. So this pipeline can be reused on other providers!

Upgrade your subscription to support this newsletter!

Also, check out our website for services, training, advisory etc:

https://www.airealist.org/

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

— or check out the shop

Subscribe now

Step 1: The Script (Let the LLM be the Director)

Don’t write the script yourself. Use an LLM (like Gemini or ChatGPT) to break your scenario into AI-ready shots.

We used a scenario called “The Silent Breach,” where an employee named Mark props open a door, leading to a theft.

The Prompt Strategy: You need to force the LLM to act as a “Video Director” and strictly format the output. Crucially, you must define “Visual Anchors”.

The Rule: Define a “Uniform” (e.g., “High-vis yellow vest, dark jeans”) and repeat it in EVERY prompt.

Why? Because AI forgets. If you don’t remind it that Mark wears a high-vis vest in Scene 3, he might show up in a tuxedo.

Here is an example of a prompt but you are free to create your own:

Act as an expert AI Video Director. Convert the scenario below into a strict Img2Vid Scene Guide.

Core Constraints:

Visual Anchors: Define a single outfit (e.g., high-vis vest) and repeat it in every prompt. Use wide shots, silhouettes, or back angles to mask facial instability.

AI Physics: No complex fine-motor interactions (gripping keys, typing). Use only simple, gross-motor movements (walking, turning).

The Handshake: Ensure perfect continuity. The End Frame of Scene X must visually match the Start Frame of Scene X+1.

Format: Silent B-roll. For Scene 2+, explicitly instruct: "Use Scene [X] End Frame as input."

Step 2: Character Generation (The Anchor)

Before generating video, you must generate your “Actor.” Text-to-Video models often create inconsistent faces. We solve this by generating a high-quality reference image first.

The Prompt for “Mark”:

A documentary-style full-body photograph of Mark, a warehouse employee in his late 30s with a slightly tired expression, short brown hair, and stubble. He is wearing a worn tan canvas utility jacket over a navy blue flannel shirt, durable dark denim jeans, and scuffed leather work boots. He stands squarely facing the camera outside a large corrugated metal and concrete industrial building during twilight. The natural dusk lighting is diffuse, clearly illuminating his face and clothing details. The background includes gravel pavement and the corner of a loading dock. Film grain texture. 16:9

I usually add the size of the image to be generated at the end of the prompts (16:9). HailuoAI has an option to specify it, but I tend to forget. I typically set it in the UI as well, but including it in the prompt acts as a safety backup. For all images, I usually generate two variants. However, if there is something complex to generate, I create more variants, as it is likely that many of the images will not turn out well.Use a model like Nano Banana Pro or OpenAI Image-1.5 for this initial image. Once you have a “Mark” you like, this image becomes the DNA for the rest of your video.

This is our Mark.

Step 3 & 4: The Handshake Execution (Start & End Frames)

This is the most critical step. We do not simply “text-to-video.” We use a Start/End Frame workflow to force the model to stay on track.

The Workflow:

1. Generate Frame 1 (Start): Use your character reference image (from Step 2) to generate the starting image of the scene. In our case, it is Mark

2. Generate Frame 2 (End): Generate the final image of the scene before you generate the video.

   • The Anchor Trick: When generating this End Frame, upload Frame 1 (and your original reference image - Mark) as an image reference. This anchors the generation, ensuring Mark wears the same clothes and stands in the same lighting in both the first and last second.

3. Generate the Clip: Pass BOTH the Start Frame and End Frame to the video model (e.g., Hailuo or Veo). The AI now only has to interpolate the movement between these two fixed points, rather than hallucinating the future.

4. The Handshake (Next Clip): For the next shot, take the End Frame of the previous clip and use it as the Start Frame for the new one.

Tools:

• Hailuo 2.0: Does not have audio but is not needed, the scene is easy to generate, it is cheaper than Veo 3.1

Choosing Your Video Model

Not all video models are built equal, and crucially for our workflow, not all models support the Handshake Method (Start + End Frames).

In the MiniMax Hailuo interface, you currently have access to a few flagship models. Here is the breakdown of what to use and why:

The Contenders:

• Hailuo 2.0 (MiniMax): This is our workhorse. It supports both Start and End frames, which is non-negotiable for consistency.

  • Pros: Extremely affordable, excellent “prompt adherence” for character movement, and it allows you to toggle “Prompt Rewrite” - a built-in feature that automatically optimizes your simple prompt into something the AI understands better.

  • Cons: No native audio generation.

• Veo 3.1 & Veo 3.1-Fast (Google): Also supports Start/End frames.

  • Pros: Generates native audio (sound effects) synced to the video. It offers high fidelity with resolution options up to 4K.

  • Cons: Significantly more expensive and restrictive quotas compared to Hailuo.

• Sora 2 (OpenAI): Available in the list, but currently does not support the End frame workflow effectively for this specific “handshake” technique. It only supports the first frame. Of course, you could do a work around and start generating the next video with the last automatically generated frame of the previous video, but this gives us less control and Veo 3.1 is comparable in quality.

Settings to Watch: When using Hailuo, you aren’t just stuck with a default. You can manually configure:

• Duration: Set this to 6s or 10s. I recommend 6s for the “Handshake” clips to minimize hallucinations.

• Resolution: Adjustable from 512p to 1080p depending on your credit budget.

Recommendation: Use Hailuo 2.0 for 90% of your shots to save budget and ensure character consistency. Only switch to Veo 3.1 if you have a complex scene (like the box drop) where physics and native sound are critical.

Step 5-6: Audio and Music

Video without sound feels fake. For “The Silent Breach,” we need a narrator and a tense background track.

Voiceover: We used MiniMax Speech 2.6.

Recommendation: To avoid a robotic voice, uppercase the words you want the AI to stress.

Example: “It’s only a two-minute trip, so he IGNORES the protocol.”.

If you are not sure where to put the stress, try to speak the text out loud and see where you would put the logical stress - you uppercase those words.

Compare this prompt:

Meet Mark. Mark is about to make a choice we've all considered. He needs to grab a heavy box of tools from his truck. It’s only a two-minute trip, so he ignores the protocol. He figures it’s just easier to prop the door open than to fumble with a badge reader while carrying a heavy load. It’s a moment of convenience that feels harmless... until it isn't.

And this prompt:

Meet Mark. Mark is about to make a choice we’ve ALL considered. He needs to grab a heavy box of tools from his truck. It’s only a two-minute trip, so he IGNORES the protocol. He figures it’s just EASIER to prop the door open THAN to fumble with a badge reader while carrying a heavy load. It’s a MOMENT of convenience that feels harmless... until it ISN’T.

Music: We generated an instrumental track using MiniMax Music 2.0.

This is the prompt that was proposed by Gemini as a fitting music background. Surprisingly, the model really struggles to generate songs without lyrics, so I just had to cut out the parts with lyrics.

Cinematic industrial thriller underscore. A deep, low-frequency bass drone sets an ominous mood. Slow, ticking metallic percussion creates a sense of time pressure and stealth. Atmospheric, gritty, and suspenseful. No vocals. Minimalist style suitable for a security breach scene. The track builds slightly in intensity but remains dark and serious.

I personally find the AI-generated music very impressive but also very concerning regarding how it might influence artists and how much copyrighted material went into those models.

Step 5: The Assembly (Don’t Overthink It)

One can waste hours trying to find an “AI Video Editor Agent” that will magically stitch everything together. Don’t.

The fastest way to assemble these clips is a standard non-linear editor. We use CapCut Desktop because it is free, fast, and has built-in features that hide AI imperfections.

1. The “Invisible” Stitch

• Import: Drag your “Handshake” clips (Clip A, Clip B, Clip C) onto the timeline.

• The Cut: Place them strictly end-to-end. Because we used the Start/End Frame workflow, the pixels at the cut point should match perfectly.

• Tip: If there is a slight “pop” or color shift between clips (which happens if the AI hallucinated the lighting slightly), apply a 0.1s to 0.2s Dissolve Transition. It smoothes over the glitch without looking like a cheesy fade.

2. Sound Design (Crucial) Some AI videos are silent by default, and some have the sound, depending if one used Hailuo or Veo. You must fill the void.

• Track 1 (Voice): Drop in your MiniMax Speech file. Use CapCut’s “Auto Captions” to generate subtitles instantly if you feel like doing it.

• Track 2 (Music): Add your MiniMax Music track. Lower the volume to -15dB or -20dB so it doesn’t fight the voice.

• Track 3 (SFX): This is quite useful. Use CapCut’s internal library to add Foley sounds. If Mark drops a box, search for “Box Drop” or “Thud.” If he walks, add “Footsteps on Concrete.” These real sounds trick the viewer’s brain into accepting the AI visual as reality.

Why CapCut? CapCut is built for this speed. It has the stock library, the auto-captions, and the “good enough” filters built-in. It serves the purpose.

Limitations and ROI

That is not a perfect video. Mark does magic sometimes, like flipping the concrete brick. It might trick some people that it is not AI, but otherwise it is quite obvious. Though, maybe it is not the bad thing if people still can tell that it is AI-generated.

You still have to deal with copyright issues (models are trained on massive datasets of existing work), so if you use it, you need to keep in mind that at any moment it might generate something that oddly resembles a famous artist or singer.

Still, the cost savings are undeniable. A training video that used to require a film crew, lighting, and actors can now be prototyped in about 2 hours for a few dollars in compute credits. The video model will be getting better and better and trainings are a good use case for them.

Tools Used in this Tutorial:

• Scripting: Gemini

• Image/Video: Hailuo AI (MiniMax) & Google Veo.

• Audio: MiniMax Speech 2.6.

• Editing: CapCut.

  Share

Yet, just like language models, image generation is an illusion of understanding. Just as an LLM is not a writer, Image Models are not painters. They also have no communicative intent to educate, inspire, scare, or mock. They generate for the sake of generation. And that is what we will discuss in this article - how image models are built, why they create an illusion of understanding, and what lies behind their architectures.

Subscribe now

A Curious Case Of Monster Generation

The latest image models have gotten much better at generating text. There are fewer spelling mistakes, fewer non-existing words, and overall the models can even produce infographics that look coherent at a glance.

But, the following prompt sent to the latest OpenAI Image-1.5 model rendered spectacular results:

Create an image that depicts 100 animals with their names written below them

Lo and behold, the next generation of Pokemons:

From far away it might look alright, but a closer inspection reveals bizarre creatures such as:

We can call this Pokemon Mothraptor. It distracts enemies with a colorful wing display, then grabs them with its lower legs to drain their HP. It is also three meters tall.

Duckillo with its friend Snovy Owl.

A jellyfish that looks like it escaped from the Upside Down world in Stranger Things.

A sea turtle that appears to be begging to be put out of its misery.

And finally: THE STARFISH! The starfish never skips the leg day in the gym.

These results are reproducible and the generation quality is consistently around the same:

The star of this batch is definitely the kitten from your nightmares:

Now let us talk about what is happening here and where these cursed animals come from.

Yes, we discovered two interesting patterns:

• Outputs become more degraded and less semantically aligned toward the end of the image.

• A reader pointed out on linkedin that duplicating the prompt helps reduce obvious errors. Simply writing the same instruction twice delivers noticeably better results, though still imperfect.

Create an image that depicts 100 animals with their names written below them

Create an image that depicts 100 animals with their names written below them

It is not about prompt engineering here. Those are though amazing hints at what is happening behind the blackbox and what architecture OpenAI image models have.

With this knowledge you will see how to prompt them and where their limitations are.

So let us dive in.

Behind the paywall, you will find a simple explanation:

• what today’s “state-of-the-art” image models are,

• why they suddenly got much better at generating text in images,

• why the same change also made them produce bizarre creatures,

• why this architecture will never be able to reliably generate things like electrical diagrams, technical drawings, or precise maps.

You will also understand what these models are actually good at, where they shine, and where they should not be used at all.

And yes, you will get a clear answer to a very practical question:
does repeating your prompt actually help image generation, or is that just a myth?

Behind the paywall is the full explanation, concrete examples, and the uncomfortable reality of what these models can and cannot do.

Read more

Unsurprisingly, a lot of people remembered the idea of providing search results to the model as a part of the prompt.

When I first heard about this, my reaction was that it sounded really dumb, but at the same time, there was nothing better, so we had to do this silly thing. I hoped that it was a temporary workaround until a better approach came out.

However, as time passed, this approach became state-of-the-art. It was known before and had fancy name: Retrieval Augmented Generation (RAG). It created an illusion of mitigating hallucinations and effective domain adaptation while, in fact, being highly unstable and paving the path for hallucinations that are simply hard to spot.

Three years later, it looks like we finally have hope to stop building hallucinating RAG bots. There is now a method of domain adaptation that utilizes native deep learning methods, benefits from the very core of the transformer architecture: attention mechanisms and is both computationally efficient and trainable.

I am talking about the latest breakthrough in continual learning: Titan memories. This might essentially be opening up a new era of domain adaptation for transformers, moving beyond fine-tuning (which suffers from catastrophic forgetting) and RAG (which is just an architectural abomination).

Why RAG is bad

RAG is a fantastic way to make models hallucinate. From one point of view, it does reduce hallucinations, but on the other, if it starts hallucinating, it does its best to ground the nonsense in your data so that you cannot immediately tell it is nonsense. This phenomenon is referred to as grounded hallucination, where the model generates factually incorrect content that appears consistent with the provided source material.

Upgrade your subscription to support this newsletter!

Also, check out our website for services, training, advisory etc:

https://www.airealist.org/

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

— or check out the shop

Read more

For the engineers and scientists, the deep dive is coming next. In the coming days, I will open-source everything: the code, the model weights, and the full mathematical framework. I am taking a little extra time to ensure the technical documentation meets a rigorous scientific standard. Stay tuned.

Executive Summary: Grafted Neural Memory for Everyone

This research introduces “Grafted Titans,” a novel architecture that solves the critical inability of LLMs to learn continuously from user interactions without the crippling costs of traditional retraining. By “grafting“, I mean attaching a trainable memory module to a frozen model.

By “grafting” a dynamic Neural Memory module onto a frozen, open-source model (Qwen-2.5) via a lightweight cross-attention adapter, I have created a system that can instantly memorize and retrieve new information with 44.7% accuracy on complex tasks, a massive improvement over baseline, while running on a single consumer-grade GPU (€4.4k hardware). This approach effectively democratizes Google’s cutting-edge “Titans” research, transforming it from a multi-million dollar industrial ambition into a plug-and-play solution that any business can use to build adaptable, personalized AI that “remember” their users, bypassing the noisy bottlenecks of RAG and the immense compute costs of infinite context windows.

Upgrade your subscription to support this newsletter!

Also, check out our website for services, training, advisory etc:

https://www.airealist.org/

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

Founding members:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow the hype.

— or check out the shop

Subscribe now

Motivation

The fatal flaw of current large language models (LLMs) is that they do not learn continuously. They are typically trained on a massive snapshot of data and then frozen in time. When a user interacts with a model, it cannot adapt its weights to new information, nor can it remember corrections from previous sessions. While this static nature is acceptable for general-purpose queries, it represents a critical failure for domain-specific tasks, particularly when processing private, evolving company data that the model has never encountered. An MIT report names this lack of adaptability as one of the major reasons why 95% of AI projects fail to pay off.

To mitigate this, the industry relies on “band-aids” like Retrieval Augmented Generation (RAG). However, RAG fundamentally attempts to force compatibility between two distinct systems: a search engine and a generative model. This integration introduces a bottleneck regarding context size:

• Too Small: If the retrieved context is too short, the relevant answer is missing.

• Too Big: If the context is too large, the model is flooded with noise, significantly increasing the risk of hallucination.

RAG is also frozen in time in its own way and does not learn. At best it can augment the knowledge base with your feedback but it does not essentially learn anything about the importance of your feedback or that the previous answer was wrong.

Furthermore, in-context learning has a big drawback - the longer the context, the more expensive is the computation. The attention mechanism in Transformer architectures scales quadratically:

meaning that doubling the context length quadruples the computational cost and memory usage.

For example, processing a prompt with a 4k context size on a Qwen-2.5-72B (or Llama-3-70B) running on a single Nvidia Blackwell (128GB VRAM) takes merely 0.02 seconds. In contrast, processing a 2,000,000 context size on the same model would take 5 minutes of pure compute time.
Huge contexts are hard to use in daily life as they needs substantial computational power and a lot of providers that announce millions token context windows cap it to no more than 100k

Test Time Training

The latest research focus has shifted sharply toward Test-Time Training (TTT). Particularly, Google’s cutting-edge Titans architecture (and its newer “Hope” variant) has received massive attention.

This architecture integrates a separate neural memory directly into the Transformer. Unlike standard models, this neural memory learns on the fly, meaning it can memorize new information instantly and use it for subsequent answers. Previously, this was the third rail of deep learning: any attempt to fine-tune a model live would shift the weights and damage the model’s core abilities, causing “catastrophic forgetting.”

Google’s work signals a new era for domain adaptation. Titans’ memory approach allows networks to learn during inference, theoretically overcoming the limitation of frozen weights and knowledge cut-offs.

The problem is: the research is still in the lab and is not directly adaptable for us yet. The approach requires a surgical decomposition of the base model, projecting attention layers of memory onto every layer of the transformer. Because the base model must train together with this memory module, it is computationally expensive and impractical for most real-world applications today.

This is the major limitation: it’s not a plug-and-play solution; it demands deep, expensive architectural integration.

AI Realist Research: Titans Memory For Everyone

Today, I am proud to present my own research, which I believe is a critical step toward the practical usage of Titans memory. This approach makes Test-Time Training (TTT) accessible to everyone, effectively signaling a new era of neural LLMs - adaptable, efficient, and user-centric.

My Main Contributions:

• Frozen Open-Weight Backbone: Instead of training a base model from scratch, I utilize Qwen-2.5-0.5B with fully frozen weights. This drastically reduces the cost and complexity of training, preserving the model’s reasoning baseline.

• Plug-and-Play Memory Interface: Rather than projecting memory onto every layer of the base model (as in the original paper), I apply a cross-attention mechanism at the first input layer. The memory weights are appended to the Qwen model as “soft prompts,” creating a modular architecture that converges faster.

• Dynamic Gating & Retrieval: The memory module trains alongside the frozen backbone. The cross-attention and native gating mechanisms ensure the model learns to dynamically decide when to use its internal knowledge and when to query the neural memory for specific details.

• Realistic Multi-Turn Training: Unlike the Google research, I implement a realistic multi-turn setup. The context is provided in the first turn (memorization phase), and the answer is queried in the second turn (retrieval phase). This forces the model to strictly differentiate between the immediate prompt context and the compressed neural memory context.

• Accessible Compute: I trained this entire architecture on an Nvidia DGX Spark Blackwell with 128GB VRAM - a portable desk supercomputer costing only €4.4k. This demonstrates that effective memory utilization does not require massive industrial clusters.

I believe this experiment demonstrates that Grafted Neural Memory can solve two fundamental problems inherent in current LLM-based architectures:

1. The KV Cache Bottleneck (Constant vs. Linear Growth)

Unlike standard KV caching, which grows linearly with context length and slows down inference, this architecture maintains a constant memory size.

• The Trade-off: Standard KV caching is a lossless approach that hits a “computational wall” as context expands.

• The Solution: My approach is a lossy compression strategy. By compressing history into a fixed set of tokens, it theoretically supports infinite context without ever increasing computational complexity. The cost of memory retrieval remains flat O(1), regardless of whether the context is 10k or 1M tokens.

2. The RAG Bottleneck (By Introducing Native Retrieval)

This architecture solves the “Search-LLM mismatch” of Retrieval Augmented Generation (RAG) and offers a viable alternative.

• The Problem: RAG is bottlenecked by the quality of external search algorithms, which are designed for keyword retrieval, not neural reasoning. RAG never shows the model the full context; it shows disjointed chunks. If the retriever fetches the wrong chunk, it forces the model to hallucinate. Furthermore, RAG is still bound by the context window limits of the LLM as discussed above.

• The Solution: Neural Memory acts as a retrieval mechanism that is native to the LLM. Instead of relying on an external database, the model learns to store and retrieve its own information. This is the architecture I have been advocating for years: Retrieval and Reasoning must be built together. That is exactly what I am doing here.

Experimental setup

I use the Qwen-2.5-0.5B-instruct model as the base for this experiment. It is crucial in this setup to utilize a base model that has already been fine-tuned for instruction following to maintain conversational coherence. The decision to use such a small model is strictly motivated by computational constraints, as I am currently operating with a single DGX Spark Blackwell. The training of just one epoch took 15 hours, resulting in a total run time of 150 hours for 10 epochs. While I believe that using a larger base model alongside a larger neural memory network would yield superior results, these were the practical limits of the available hardware.

For the memory architecture, I follow the unofficial implementation of the Titan memory published by lucidrains. The Titan memory is trained from scratch, meaning I rely on random initialization rather than pre-trained weights. During the experimental phase, I attempted to zero out the last neural layer to ensure the noise from the randomly initialized memory did not overwhelm Qwen. Unfortunately, this strategy failed; Qwen simply started ignoring the memory stream entirely rather than learning to denoise it. The original Google paper solves this by jointly pre-training the memory module with the core model, which ensures the memory network starts with learned compression patterns. This is a valid and likely necessary path for future research, though I did not implement it here as I am operating on frozen core model.

Major contributions to the architecture:

1. Soft prompting of the input layer with memory embeddings. Following the foundational approach of Lester et al. in “The Power of Scale for Parameter-Efficient Prompt Tuning”, I append the memory embeddings directly to the first input layer of the Qwen model. Crucially, the Qwen embeddings remain frozen throughout the entire training process. I train the memory on the BABILong dataset, restricting the context window to 2,000 tokens. This limit was necessary because training on a larger window caused the time per epoch to balloon to 34 hours. For a proof-of-concept, the 2,000-token length is sufficient, though future research with higher computational resources would certainly target larger contexts.

2. Cross-attention mechanism. To ensure the architecture can accurately judge when to use memory and when to rely on the model’s internal knowledge, I train a cross-attention gating layer. The purpose of this layer is to dynamically decide the weight of the memory for any given query. This learned gating mechanism is significantly more efficient and robust than setting a static hyperparameter to control how much the model relies on the memory.

I hypothesize that this cross-attention mechanism is the primary reason I achieve such high accuracy (~45%) on exact word retrieval from BABILong, despite using a highly compressed, lossy setup. While standard soft prompting "smears" details into a general gist, the cross-attention layer acts like a signal amplifier (This mirrors the core finding of Google's Titans research as their architecture succeeds precisely because it deeply integrates memory into the attention layers. By replicating this "attention-as-retrieval" mechanism in a lightweight form, I enable the model to actively "search" its memory rather than passively receiving it.). It can extract and reconstruct faint, specific signals (like an exact passkey) from the compressed memory embeddings that would otherwise be lost in the noise. It bridges the gap between the "blurry" nature of soft prompts and the precision required for retrieval.

Evaluation and Results

The evaluation is conducted on a held-out subset of the BABILong dataset, using a strict two-turn protocol to ensure no information leakage. In Turn 1, I feed the context to the model via a prompt. This prompt passes through the architecture and into the neural memory, which adapts itself by updating its internal weights on the fly. In Turn 2, the model receives the question.

Crucially, the original context is completely removed for this second turn. Unlike standard RAG or long-context approaches where the model can “look back” at the chat history or retrieved chunks, here the model gets nothing but the isolated question. The only way for it to answer correctly is to retrieve the specific fact from its neural memory.

To understand the significance of the results, we must look at the baseline. On this specific BABILong task, a model relying on random guessing achieves approximately 0.67% accuracy, which is basically - everything is wrong except for one random guess. Another baseline was the predictions by Qwen when the full context is provided. That baseline achieves 34%.

Monitoring the learning progress on the validation set reveals a clear trajectory. The model begins with significant noise and near-zero accuracy, effectively hallucinating. As training stabilizes, it climbs to 22%, proving it has moved beyond random guessing and is beginning to form coherent memory structures. As training continues, the accuracy climbs to 44,7% by the 9th epoch!

The Critical Finding: This result is counter-intuitive and profound. My grafted memory system outperformed the vanilla model that had the full text in its context window (44.7% vs 34.0%).

This suggests that for smaller models (like the 0.5B), a raw context window is not always the best way to handle information. The model’s native attention mechanism gets overwhelmed by 1,500 tokens of text. In contrast, the Grafted Neural Memory acts as a denoising filter, compressing the noisy narrative into clean, retrievable signals.

Frankly, this result exceeded my expectations. Given that the neural memory was not pre-trained but initialized from random weights and that the entire setup relies on a lightweight model tackling the complex BABILong benchmark achieving 44,7% accuracy confirms that the “Grafted Titans” architecture is genuinely learning to retrieve!

Implications for practical usage

I believe that my approach is a great foundation for integrating Test-Time Training (TTT) models for wider usage. The computational resources needed to train this memory are accessible to most research institutions, companies, and even owners of high-end private PCs. It is fundamentally different from the current state-of-the-art proposed by the Google research team, which assumes a multi-million dollar training run of a bespoke Transformer model from scratch.

To my knowledge, I am the first to propose a plug-and-play inference-time memory for frozen, pre-trained base models. Depending on the available compute, this approach can essentially be used with any open-weight Transformer model on Hugging Face (I will publish the code repository and the weights). Furthermore, thanks to the Gated Soft-Memory Injection mechanism (my enhanced version of soft prompting appended through cross-attention), I see massive potential for distributing pluggable Titan memory modules for specific architectures like Qwen, Llama, etc.

Crucially, this architecture resolves the inference latency bottleneck. Unlike the full Titans model, which requires heavy sequential updates during generation, my grafted memory acts as a lightweight adapter. It adds negligible latency, only the cost of a single cross-attention pass per query, making it viable for real-time applications where speed is non-negotiable.

Limitations

1. The “Two-Turn” Limit

Current benchmarks can be deceiving. My setup is trained and evaluated on BABILong, which essentially tests a “Store —> Retrieve” loop in 2 turns. While this proves the mechanism works, it does not prove stability. We simply do not know how the memory behaves in a truly multi-turn conversation (10+ turns). Does the memory saturate? Does earlier information get overwritten by later noise (”catastrophic forgetting” on a micro scale)? A true production system needs to be tested if it can keep much more information in the memory.

2. The “Gullible Memory” Risk and No Guardrails

Currently, this architecture implements zero guardrails against memory corruption. It treats every user input as absolute truth. If a user (or a malicious prompt injection) states “The sky is green” in Turn 3, that data is encoded into the neural memory weights as a fact. In Turn 4, the model will hallucinate based on this poisoned data. Without a “verification layer” or a “truth discriminator” to filter incoming information, this setup is highly vulnerable to Memory Poisoning and conflicting data. This is an area for future research.

3. The Vanishing Gradient (Signal Dilution)

I must also address the critical architectural drawback of my grafted approach: Signal Dilution. By injecting memory primarily at the input level (Layer 0) via soft prompting, we face a “distance problem.” The memory signal must propagate through all 32+ layers of the Transformer to reach the output, risking a “vanishing gradient” effect. The larger the base model, the higher the risk.

While my cross-attention gate acts as an amplifier to mitigate this, it is not bullet proof. Future research should explore multi-layer injection that is grafting memory adapters at higher layers (e.g., the middle or final blocks) to shorten the path between retrieval and reasoning. I encourage readers to fork my neural memory implementation once I publish it, adapt it to their needs, and help solve this signal dilution challenge.

4. Production Readiness

Let’s be clear: this is a research prototype, not a production microservice. Moving this to production would require solving the signal dilution problem and building a robust filtering mechanism to prevent the memory module from becoming a trash can of conflicting user inputs. Until then, it remains a promising proof-of-concept, not a product.

Conclusion

In this article, I presented my latest research, which I believe represents a meaningful step toward addressing one of the major limitations of current large language models: their inability to dynamically learn from users and adapt to new information. I argue that plug-in lossy memories, added to a large and stable base model, can provide the level of personalization and domain adaptation that current band-aid approaches such as RAG and long-context extensions fail to deliver.

My results show that integrating a neural memory does not require complex or invasive modifications to the transformer architecture. Instead, it can be implemented as a plug-and-play soft prompt, connected to the input layer of the base model via cross-attention.

I also demonstrate that this architecture drastically reduces the computational resources required to create a Titan memory when training is restricted to updating only the neural memory weights. Training costs are reduced from millions of euros to a few thousand, making this approach accessible to a broad range of AI researchers and teams with limited budgets.

I therefore believe that this work will inspire and enable companies and research institutions to actively explore and study Titan memories.

Acknowledgements:

To the Community: First, a massive thank you to my subscribers on Substack and LinkedIn. The incredible popularity of my posts has encouraged me to spend almost all my free time on research, studying LLMs and writing.

To the Patrons: Second, I want to specifically thank my paid and Founder subscribers, as well as everyone who used my services at airealist.org. You are the investors behind this research. Your support directly funded the NVIDIA DGX Spark Blackwell, the Mac Mini M4, and the stack of software subscriptions I use daily to run these experiments. Without you, this hardware and this experiment would not exist.

Why I Publish Here: This direct support is exactly why I am releasing this research here first, rather than waiting six months for a conference review cycle. As always, I am publishing this work without a paywall. I believe that research which helps us understand the reality of AI should be open and accessible to everyone.

Subscribe now

Subscribe now

1. Year of Disappointment In AI

AI agents will not deliver for business. They will not greatly automate business processes; the adoption of Microsoft Copilot Studio, Snowflake Cortex, and Salesforce Agentforce will remain low. For example, Salesforce has reportedly closed only ~8,000 Agentforce deals out of 150,000+ customers and after laying off 4,000 people admitted that there are trust issue and they need to scale back AI usage.

The graveyard that is currently filled with RAG bots will have a neighboring graveyard even bigger: full of AI Agent PoCs. We are already seeing failure rates of over 95% for these pilots. This time, management will realize that they were sold science fiction and that firing half of customer support because they put an “agent” into production was a big oopsie. Klarna already proved this in 2025, rehiring human agents after realizing their AI support tanked customer satisfaction. To hire people back, mitigate the losses, and simply save the budget, the savings will come from the innovation budget that was originally allocated for AI. Forrester now predicts enterprises will delay 25% of AI spend. The hype and overpromising have undermined trust: getting money for even a meaningful AI use case will be tough. Expect statements flying around from various companies similar to the infamous 1966 ALPAC report, which concluded there was ‘no usable machine translation’ and killed funding for a decade.

we do not have useful machine translation. Further, there is no immediate or predictable prospect of useful machine translation

1966 ALPAC Report

We will see very similar statements declaring LLMs are not useful for real business goals.

Before you continue, here is a little musical prediction from the senior architect Meow:

(generated with Nano Banana Pro, Veo-3.1, MiniMax Sound, Hailuo-2.0 on HailuoAI)

2. Downfall of Nvidia’s Dominance

Nvidia will lose its dominance in data centers. We already see strong indications. Google and Meta collaborate on “TorchTPU”, a project effectively breaking the CUDA lock-in by making PyTorch run natively on Google's hardware. Google is now clearly moving to sell TPUs directly to external data centers, not just via cloud. Meanwhile, AWS and Anthropic have activated Project Rainier, a massive cluster of Trainium 2 chips that Anthropic is using for its primary training, cutting Nvidia out of one of the world's largest AI workloads. China will entirely switch to Huawei and possibly even other domestic chips. Considering ASICs are both cheaper and more efficient in terms of energy consumption - Nvidia will become a luxury brand for the lazy. And after the release of TorchTPU, Nvidia’s biggest moat - CUDA - will be gone.

Become a paid subscriber or visit the shop (each item = one month subscription)

AI realist also offers consulting, trainings (for example, video generation with AI as the one in this article) and advisory:

https://www.airealist.org/

Paid subscribers get:

• Priority answers to your messages within 48-hours

• Access to deep dives on the latest state-of-the-art in AI

• Free participation in quarterly practical calls. Last call: Building Your Website With AI

Founding members receive even more:

• A 45-minute one-on-one call with me

• High-priority personal chat where I quickly reply to your questions within 24-hour

Support independent research and AI opinions that don’t follow to hype

Read more