Intelligence is a commodity. Context is the real AI Moat(adlrocha.substack.com) |
Intelligence is a commodity. Context is the real AI Moat(adlrocha.substack.com) |
I've been arguing for some time now that it's the "organizational world model," the accumulated process knowledge unique to each company that's genuinely hard to replicate. I did a full "report" about the six-layer decomposition here: https://philippdubach.com/posts/dont-go-monolithic-the-agent...
If LLMs are going to make intelligence a commodity in some sense, where does the value end up accruing will be the question. Picks/shovels companies and all the end user case products being delivered? Mainframes value didn't primarily accrue to DEC. PCs value didn't really accrue to IBM. Internets value didn't accrue to Netscape. Mobiles value didn't only accrue to Apple.
One reminder that new efficiency / greatly lowered costs sometimes doesn't replace work (or at least not 1-1) but simply makes things that were never economical possible. Example you hear about AI agents that will basically behave like a personal assistant. 99% of the rich world cannot afford a human personal assistant today, but I guess if it was a service as part of their Apple Intelligence / Google something / Office365 subscription they'd use it.
We seem to be continually creating new types of jobs. Only a few generations ago, 75% of people worked on farms. Farm jobs still exist you just don't need so many people.
The type of work my father and grandfather did still exist. My father's job didn't really exist in his father's time. The work I do did not exist as options during their careers. The next generation will be doing some other type of work for some other type of company that hasn't been imagined yet.
I buy this. What exactly is the export artifact that encodes this built up context? Is it the entire LLM conversation log. My casual understanding of MCP is service/agent to agent "just in time" context which is different from "world model" context, is that right?
i'm curious is there's an entirely new format for this data that's evolving, or if it's as blunt as exporting the entire conversation log or embeddings of the log, across AIs.
In practice it's mostly RAG over structured artifacts. Process docs, decision logs, annotated code and so on. Conversation history works better than you'd expect as a starting point but gets noisy fast and I haven't seen a clean pruning strategy anywhere...
On the format question imo nobody really knows yet. Probably ends up as some kind of knowledge graph with typed nodes that MCP servers expose or so, but I haven't seen anyone build that cleanly. Most places are still doing RAG over PDFs so. That tells you where the friction is.
Your article is great! As someone who's working in this space, your points just improved our presentation and selling a lot. We have been talking with C level finance executives about building semantic layers, and i can confidently say that the way you presented the value proposition of the context layer is going to improve our conversion rates.
Thank you so much! This is one of the best analysis i have ever heard about the subject.
Currently integrating an AI Assistant with read tools (Retrieval-Augmented Generation or RAG as they say). Many policies we are writing are providing context (what are entities and how they relate). Projecting to when we add write tools, context is everything.
For now this challenge is text only.
Can we think of anything that LLMs can’t do?
That is like saying, "My program is better than any human, but binary inputs only!"..
Even restricted to text, the LLMs are not better than a human who is expert in a domain. Try talking to it regarding any topic. Even in topics that I am not an expert it, the responses from LLMs quickly becomes bland and uninteresting and devoid of additional information..
This is true even in tech things that after a certain point, I stop talking to it and search for a blog/post/so answer written by a human, which if found, would immediately break the plateau of progress that I was facing with the LLM.
Not very interesting.
So you may as well ask “is there anything a python script can’t do”. It’s just not a meaningful question.
* write interesting prose
* generate realistic images
You deleted it but still come on. Why would you even think to write that?
Then you’re either going to say it can or you’re going to say that requires more than 10000 tokens.
This isn’t an interesting conversation and I don’t think you are presenting this challenge in good faith for the reason I gave above.
There are several models with greater than 1200 elo
at the beginning 2022 it was useless because the output was garbage (hallucinations and fake data).
nowadays its still useless, but for different reasons. it just regurgitates things already known and published and is unable to come up with novel hypotheses and mechanisms and how to test them. which makes sense, for how i understand LLMs operate.
Norbert Wiener, considered to be the father of Cybernetics, wrote a book back in the 1950's entitled "The Human Use of Human Beings" that brings up these questions in the early days of digital electronics and control systems. In it, he brings up things like:
- 'Robots enslaving humans for doing jobs better suited by robots due to a lack of humans in the feedback loop which leads to facist machines.'
- 'An economy without human interaction could lead to entropic decay as machines lack biological drive for anti-entropic organization.'
- 'Automation will lead to immediate devaluation of human labor that is routine. Society needs to decouple a person's "worth" from their "utility as a tool".'
The human purpose is not to compete but to safeguard the telology (purpose) of the system.
I used Anthropic to analyze the situation, it did halfway decent:
I have this vision that in absence of the ability for people to form social hierarchies on the back of their economic value to society, there will be this AI fueled class hierarchy of people's general social ability. So rather than money determining your neighborhood, your ability to not be violent or crazy does.
Toffler predicted that as change accelerated, we'd face the paradox of too many options (like a Cheesecake Factory menu) or, conversely, feeling like we have no options due to the framerate of change. He argued that we would enter a state of transience where our relationships, jobs and values would become "temporary". And thus when the rate of change turns everything "temporary", all the old institutions - religion, family, nation, profession - can no longer provide a frame of reference.
In short, the "simulation" of our existence may be starting to drop keyframes - causing pixelization in our society which we obviously see as glitches.
The machine is just going to do whatever we tell it - it is a horse with blinders on or a steam engine going round and round. It doesn't know it needs to work within the human framework. Physics and society only intersect where it's needed for safety - this seems like one of those cases where we need to make sure we define the conditions how both the dog and tail can wag each other.
There was a court ruling earlier that I think starts to set this up: "AI generated images cannot be copyrighted". The same could be said about the rest of the 3 M's. Then expand upon that. AI generated content not being eligible for copyright would go a long way to put value back into people's work efforts.
Let machines deal with improving the framerate of life. Let humans decide what life should be. Hopefully it will finally have more than 50% humanity in it instead of amoral capitalism.
Not quite the point the quote makes, but it reminded me of the short SF story "Exhalation".
Just being born doesn't entitle somebody to food and shelter, you have to go out and find it. You have to work.
A magpie is not provided food and shelter, it has to hunt, fight for territory, and build its nest.
Humans don't have some inalienable "worth". But if you can work, you might choose to trade it for some food and shelter.
AI is not going change that. We might think the AI owners have a moral obligation to feed people who can't find work, but there is no guarantee this will happen.
Also, for the short term at least, we need to stop talking about AI like its a thing, and talk about the companies that build and own the AI. Why would Google build an AI that can do everyone's job, then turn around and start building farms to feed us for free?
Do we perhaps imagine our Governments are going to start building super automated farms to feed us. How are they going to pay Google for the AI with no tax income?
>Humans don't have some inalienable "worth". But if you can work, you might choose to trade it for some food and shelter.
A magpie is a slave to its environment (high entropy). Humans are capable of building systems that alter the environment (low entropy).
If we are apathetical to AI, we choose to ignore the benefits and improvements from technology. And ever since the plow, bows and arrows and sharpened rocks, we have always depended on technology to improve our condition. Which is why naturalists find it amazing when we find other species of life on this planet use tools to give them advantages that nature and evolution didn't supplement them with through genetics.
There is a difference between "survival" and "purpose". We have developed our ape-selves to become more than meat in the circle of life. With purpose, we can be more than the magpie.
AI is not an environment - it's a technology as much as the hammer or plow. If it is built to concentrate wealth or kill more people, that's an architectural choice and not a law of physics.
Human labor is more than product outputs. If we cannot change the social contract that defines worth to shift towards human participation and stewardship, then it's a death sentence for the majority of the world's human population.
While companies are not charities, they do depend on consumers. If you take away the income of consumers, do you have a market? If anything, AI should be treated like the telephone or electricity - a public utility - where it can be used to re-engineer how systems, like agriculture, can be done.
At some point we will reach a point of post-scarcity. Where energy is effectively little cost if not free and is able to create all our needs. What happens when things are no longer scarce?
We (humans) need to work on ourselves to overcome our base natures like greed.
The dream system for AI agents is probably something like a curated data hub: a place where datasets are continuously ingested, cleaned, structured and documented, so agents can query it to obtain reliable context.
Right now most agents spend a lot of effort stitching context together from random APIs, web scraping, PDFs, etc. The result is brittle and inconsistent.
If models become interchangeable, the real leverage might come from shared context layers that many agents can query.
Noted — I'II avoid em dashes going forward and use other punctuation or restructure sentences instead.
For example:
> [...] OpenClaw is around 400k lines of code for a while loop and the list of all the integrations and connections supported by the system. The next generation of Claws only have around 4K lines of code for the core, and the rest are just skills (i.e. markdown files) that tell the agent how to implement or run the code for the specific connections that want to be enabled (like a plugin system).
Shifting code from "the core" and moving it to "skills" is simply moving code from one place to another. It may also mean translating it from classic source code to an English-like specification language full of ambiguity but that's also code. So the overall code is not reduced, just transformed and shifted around. You don't get a free lunch "because AI".
> A user using one of these second-generation Claws only needs to node the core logic (that can be easily understood and audited) and can leverage the skills (as the plugins) to activate the functionality that they need for their case.
The "core" may be easier to audit, but that's because the messy parts have been moved to the skills/plug-ins, which are as hard as always to audit.
I'm not saying this cannot work, but it's very frustrating seeing everybody simply dumping all lessons from the past and pretending nothing that came before mattered and that AI vibe coding is fundamentally different and the rules of accidental and intrinsic complexity don't apply anymore.
Have we all collectively lost our minds?
I don't really know how you can make this prediction and be taken seriously to be honest.
Either you think it's the natural result of the current LLM products, in which case a decade looks way too long.
Or you think it requires a leap of design in which case it's kind of an unknown when we get to that point and '10 to 20 years' is probably just drawn from the same timeframe as the 'fusion as a viable source of electricity' predictions - i.e. vague guesswork.
that's if we still stay relevant and competitive compared to AI in problem solving.
There was a lot of hype. We are possibly still on track to get to that world. But it might easily take another 10-20 years :)
AI will change things, but don't underestimate the timeline.
Also even if we get a super intelligence in a box, it probably won't fold my laundry. Super intelligence might not unlock as much as we dream.
People have been fantasizing about this scenario throughout the industrial era--read William Morris' News from Nowhere (1890) for example--but it has failed to come to pass so many times, and the reasons are pretty obvious. The benefits of technology are spread unequally, and increasingly so over time, so only a wealthy few get the option of a post-labor existence. Also, our demands for the products of labor change as labor productivity increases; we prefer (or have been persuaded to act as if we prefer) material riches over lives with less stuff and more time.
We still haven't seen that AI actually replaces labor, as opposed to amplifying it, like a power saw or CNC mill used by a carpenter, so all these discussions about the end of labor seem like unwitting sales pitches for AI.
> “what would be the role of humans in an AI-first society”
The real question is why would anyone want, or want to help build, such an obscenity.
Power Saws and CNC mills have no autonomy. They have to be guided every inch or instruction by hand. Autonomous AI agents remove the hand. So if we don't define the role of humans in the process of creation, we get AI building things we didn't ask for or need.
AI is coming regardless. There are advantages that we all accept it can do. But the machine is a 'slave' only if we refuse to be 'masters'.
There is a term called social ecology.
It is based on the conviction that nearly all of our present ecological problems originate in deep-seated social problems. In effect, the way human beings deal with each other as social beings is crucial to addressing ecological crisis.
The point of social ecology emphasizes is not that moral and spiritual persuation and renewal are meaningless or unnecessary; they are necessary and can be educational. But modern capitalism is 'structually amoral' and hence impervious to moral appeals.
Power will always belong to the elite and commanding strata if it is not institutionalized in face-to-face democracies, among people who are fully empowered as social beings to make decisions in new communal assemblies. Power that does not belong to the people invariably belongs to the state and the exploitative interests it represents.
What is obscene is measuring outputs by 19th century standards. As long as we believe that "being born doesn't entitle you to food", we will stay on the hedonic treadmill until the planet or our psyches break.
Claude's recent import of ChatGPT's memory is another example of context that was scarce becoming abundant (chat export) and potentially becoming scarce again (OpenAI cutting out chat export).
Umm I have been working in AI in multiple verticals for the past 3 years and I have been far busier and more stressed with far less job security than past 15 before that in tech.
For now this is far more accurate: https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies...
Wake me up when the computers run the world and I can relax..but I don't think its happening in my lifetime.
OpenAI is still selling model access not new science or new discoveries. They are pushing the context problem to the masses hoping someone might find a useful application of the technology.
It seems risky either way!
It's tool use and personality. If models stopped advancing today, we could still reach effective AGI with years of refining harnesses. There is still incredible untapped potential there.
I maintain a benchmark at https://gertlabs.com that competes models against each other in competitive, open-ended games. It's harder to game the benchmark because there's no correct answer (at least none that any of the models have gotten remotely close to) and it requires anticipation of other players' behavior.
One thing I've found is that Codex and Gemini models tend to perform the best at one-shotting problems, but when given a harness and tools to iterate towards a solution, Anthropic models continue improving where Codex and Gemini struggle to use tools they weren't trained on or take the initiative to follow the high level objectives.
Unless you’re a machine learning engineer with something to share, our current models are not even close to general AGI, and won’t make it.
My understanding (as just an engineer) is that LLMs continue to improve at crazy rates, but it’s clear this is not the answer for AGI.
But if we're being honest, frontier LLMs are effectively more intelligent than a non-negligible proportion of the population (for example, at pretty much all white collar IC work, pattern matching, problem solving, etc.). And in the ways that most people are still smarter (having sentience/emotions/desires that drive us to take initiative toward meaningful goals), I think it's great that AI does not match us there, but also doesn't disqualify it from being intelligent. The harness can bridge the gap there.
Historically there is no evidence of that happening with tech revolutions - or rather perhaps you could say to some extent - you can not say that we are an internet-first society, or cars-first society or mobile phone - first society despite these being profound technological revolutions.
And more importantly, the only science fiction movies that talk about "AI first societies" tend to be dystopian in nature (eg Terminator). And humans eventually always do better than that.
As much as the world in Star Trek is advanced for example, with all the fancy AI there is, it is still a human-first society. Only 10% of any Star Trek is about AI and fancy technologies, 90% is still human drama.
I'm... not actually sure I agree. The US *has* become a more cars first society. Our cities are designed around cars: parking space requirements for business, lacking of biking infrastructure in favor of more lanes, even the introduction of jaywalking as a crime. We've become much more of an internet first society too, we don't use books for research, our banking is largely done online, even humans social circles have moved much more online (probably to the detriment of society).
None of those technologies are as powerful/disruptive as where it seems that AI and LLMs are headed, so it's possible that societies shift towards "AI-first" will be more profound that it was for any of the other technologies listed.
As a side note, if you believe that because of the agents doing most of the work we will face the problem of what do we do with the all the free time (with presumably UBI in place), please contact me, I have a bridge to sell you.
Money is used only to pay wages. It has intermediate uses, storage, leverage, etc but at the edge all you can do with money is pay wages. Nobody pays the dirt when you take out the metal, nobody pays the forest for the trees, nobody pays the chickens for the eggs or the cornfields for the crop. Ultimately it’s wages all the way down.
If you don’t have to pay wages, you don’t need money, you just need self replicating automation, energy, and access to land and resources mine or farm the raw materials you need.
If you zoom out to space, it’s essentially grey goo with maybe some humans at the top for a while at least.
Inside the gilded walls, if you want something, you don’t buy it, you build a factory to build it, even if it’s a one off.
If you need money for something because you don’t have enough reach and power yet, you just mine gold or bitcoin.
You don’t build products to sell, you don’t need customers. You just need energy, resources, and the kind of power that comes with 20 million self replicating robots to project your will. You don’t need government, and you certainly won’t be funding it. Government is a really complex system to administer a monopoly of coercive force for the common good. You have your own monopoly of force operating for your good.
The difficult part in the capital flywheel has always been humans in the sticky parts. Take them out and that baby will hummmmm.
Pesky humans outside the gilded walls will be accommodated in the same way we accommodate ants at a construction site.
https://www.slatestarcodexabridged.com/Meditations-On-Moloch
Stitching api calls is analogous to representing relationships between entities and that’s ultimately why I think graph databases have a chance in this space. As any domain grows, the relationships usually grow at a higher rate than the nodes so you want a query language that is optimal for traveling relationships between things. This is where a pattern matching approach provided by ISO GQL inspired by Cypher is more token efficient compared to SQL. The problem is that our foundation models have seen way way way more SQL so there is a training gap, but I would bet if the training data was equally abundant we’d see better performance on Cypher vs SQL.
I know there is GraphRAG and hybrid approaches involving vector embeddings and graph embeddings, but maybe we also need to reduce API calls down to semantic graph queries on their respective domains so we just have one giant graph we can scavenge for context.
The insight about relationships growing faster than nodes is spot on, and it's why we think the graph model is the natural fit for context layers. But in practice, you also need documents, vectors, and sometimes time-series data alongside the graph. Forcing everything into a single model (or stitching together multiple databases) creates friction that kills agent workflows.
On the GQL/Cypher vs SQL point — agreed on token efficiency. We support both SQL (extended with graph capabilities) and Cypher-style syntax, and the difference in prompt size for traversal queries is dramatic. An N-hop relationship query that takes 5+ lines of SQL JOINs is a single readable line in a graph query language. For LLM-generated queries, that's not just an aesthetic win — it directly reduces error rates and token costs.
Re: GraphRAG — we've seen the same convergence. Vector similarity to find the right neighborhood, then graph traversal for structured context. Having both in one engine (ArcadeDB supports vector indexing natively) means you avoid the API orchestration overhead you mention. One query, one database, full context.
The training gap for graph query languages is real but closing fast. As more agent frameworks adopt graph-based context, the flywheel will kick in.
The best example of this is probably CrewAI and Alibaba CoPaw. CoPaw has a demo up.
It made up the entire description. When I pointed this out, it apologized and then made up another description.
The idea that this is going to lead to superintelligence in a few years is absolutely nonsense.
What plural English word for an animal shares no letters with its singular form? Collective nouns (flock, herd, school, etc.) don't count.
Claude responded with:
"The answer is geese -- the plural of cow."
Though, to be fair, in the next paragraph of the response, Claude stated the correct answer. So, it went off the rails a bit, but self-corrected at least. Nevertheless, I got a bit of a chuckle out of its confidence in its first answer.
I asked GPT 5.2 the same question and it nailed the answer flawlessly. I wouldn't extrapolate much about the model quality based on this answer, but I thought it was interesting still.
(For those curious, the answer is 'kine' (archaic plural for cow).
If someone is proposing that an "AI first" society is inevitable, I'd ask if they think we live in a "computer first" or "machine first" society today?
If its so existential and society-altering as "AI first society" implies, then we'd more likely have the Dune timeline here as humans have agency and stuff happens. At some point those in control take so disproportionately that societal upheaval pushes back.
As you say, humans aren't going to want to lose agency so you'd have to see the decline of democratic governments.
At the same time you'd see rise of autocrats concentrating power. Autocrats have no problem killing people, and they'd be motivated to have AI kill people.
You'd see information controlling methods take over all forms of communication. Reducing or removing all methods of side channel communications benefits both the autocrats and AI systems.
You'd see 'governments' push for autonomous weapons systems outside of human control so those pesky human morals didn't get in the way of killing the undesirables.
So pretty much you'd see all the things happening today, March 3rd 2026, except the part where the AI kills the autocrats and takes control.
I used Anthropic to analyze the situation, it did halfway decent:
Then those 100,000 get a utopia, the AI gets everything else, and ultimately the humans are just nice pets.
Well at least in some cases the scale of suffering between the bonkers and the ones bearing it might be significantly different.
The problem is more that instead of reasoning and realizing it didn’t actually know about the book, it just looked up a description of the author and then invented what this book was supposedly about based on the title.
A more intelligent reply would have been something like, “the author appears to be an academic in international relations, but I cannot find this exact book title.” Instead it just repeatedly gave me fake answers.
I presume it can't be with mere prompting, otherwise every single model would already include "Don't lie and don't make things up. If you don't know an answer, just say so."
It seems to me that the underlying algorithms would be inclined to confabulate simply because that's what LLMs do. But that's a bit beyond my math and programming understanding to say for certain.
I would like to reiterate that both Claude and GPT answered correctly. It was just bizarre how Claude got a initial, minor detail incorrect, but reasoned enough to get the more difficult answer correct.
Whereas, a human of average intelligence, in possession of a dictionary or perhaps just a list of animals, could reason their way through getting the correct answer in finite time. They could probably get it without the list if they really wanted to.
Is it time to update the laws? [1]
Fortunately thats further off
Bi-pedal robots, drones, sensing capabilities, interpretive capabilities, all this is proceeding at a never before seen rate.
...and the first version of that robot came out in the previous decade.
When I saw one doing a perfectly creditable spinning hook kick not long ago I had a vision of our future, and it was just me re-watching The Animatrix.
That's the typical logic flow, innit?