gonna keep beating this drum because i think the market is misunderstanding $REI ...
the main criticism i hear: "a small team can't compete with OpenAI building their own LLM"
they're not building an LLM!! that's the whole point
OpenAI = language models (systems that predict text)
REI = reasoning substrate (a system that thinks independently and uses someone else's LLM just to talk to you)
closest comparisons are not ChatGPT or Claude. they're Ndea (Chollet's lab), Verses, Symbolica - alternative AI architecture labs valued at $750M - $5B
REI is making the same type of architectural claim. the difference is access: those labs are private and oversubscribed. REI trades as a token at a fraction of the valuation.
the bet isn't "can they beat OpenAI." the bet is "can reasoning exist outside a language model." if yes, this is mispriced by orders of magnitude. if no, nothing else matters
papers are coming. that's when the claim becomes verifiable. until then it's conviction
It’s hard to disagree with the direction of this. But it’s worth being honest that the neuro-symbolic lineage lost to scale for most of the last decade.
At the surface, everything collapses into prose. Fluent text is a poor record of what produced it. Two systems doing completely different work underneath can return the same paragraph. Whether anything symbolic is happening has to be answered below the language layer and not in the prose.
LLMs are excellent at prose, and they’ll remain relevant in that era. But prose is not the thing to lean on once a task needs real adaptation, verification, planning, or abstraction.
The lineage has been kept alive in places with DeepMind’s theorem-proving work among them and it’s finally starting to look right again.
A fully anesthetized human being continues to process and predict semantic content they hear as if they were conscious. Natural language is not even close to expressing or producing the essence of human existence.
$REI might be one of the most misunderstood AI projects in crypto.
Not because people are bearish.
Because most people are still using the wrong mental model.
They keep asking:
“Is this another LLM?”
“Is this a wrapper?”
“Is this RAG?”
“Is this just cheaper inference?”
“Is this another AI agent token?”
I think the better question is:
What if the missing layer in AI is not a bigger model…
but a system that can learn, revise, recall, forget, adapt and build domain expertise at inference time?
That is the $REI thesis.
And the more I dig, the more obvious it becomes that this is not trying to compete in the same game as most AI projects.
Most of AI right now is stuck in the same loop:
1. Scale the model
2. Add more data
3. Add more compute
4. Add a RAG pipeline
5. Wrap it in an agent UX
6. Call it “memory”
7. Hope the output is useful
That works.
But it does not solve the deepest problems.
It does not solve persistent learning.
It does not solve fragile retrieval.
It does not solve domain-specific cognition.
It does not solve hallucination-prone reasoning.
It does not solve the fact that most “agents” are still stateless language systems with tools bolted on.
It does not solve the fact that enterprise AI needs reliability, continuity and adaptation, not just prettier chat windows.
The entire frontier AI race is becoming a capex war.
Bigger clusters.
More GPUs.
More tokens.
More data centers.
More power.
More subsidy.
More inference burn.
But what if the next major unlock is not only scale?
What if it is architecture?
This is why $REI is interesting.
REI is not positioning Core as “a better chatbot.”
From the way the team talks about it, Core is closer to a self-contained cognition / reasoning layer.
A layer that operates over dynamic knowledge structures.
A layer that can update, revise, decay, retrieve, mutate and reason over what it has learned.
A layer that can start cold, evolve through use and develop domain expertise from experience.
That is a completely different mental model from:
“prompt → model → answer”
The key word is not chat.
The key word is learning.
Most AI products today simulate memory.
REI is trying to build systems that actually form, revise and use knowledge.
That distinction matters.
A normal LLM can answer from weights.
A RAG system can fetch from documents.
But an adaptive reasoning system should be able to build structure around meaning.
It should know that two differently worded things may represent the same concept.
It should know when to preserve exact verbatim detail and when to collapse repeated meaning.
It should know which context matters, how relationships change, and what parts of prior knowledge should be strengthened, weakened or forgotten.
That is where things get interesting.
Because the AI sector is drowning in wrappers.
Everyone can call an API.
Everyone can build an agent UI.
Everyone can add vector search.
Everyone can ship a demo that looks intelligent for 30 seconds.
The hard part is making intelligence persistent.
The hard part is making it reliable.
The hard part is allowing it to improve from use without turning the whole system into an unpredictable mess.
The hard part is giving AI a structure that is not just memory, but evolving conceptual understanding.
That appears to be the direction $REI is taking.
And no, I do not think this means “LLMs are dead.”
Actually, the bullish framing is the opposite.
REI does not need to replace transformers.
It can complete them.
Transformers are extraordinary at language.
They are unbeatable for natural language interfaces right now.
But language is not the whole problem.
A lot of AI’s missing value is not in generating prettier sentences.
It is in:
• persistent domain learning
• adaptive reasoning
• concept formation
• knowledge revision
• inference-time improvement
• memory that is more than retrieval
• systems that become better the more they are used
That is the gap.
And that is why the best framing for $REI is not:
“AI agent coin”
It is:
“cognition layer for AI.”
That is also why the current product surface may not show the full ceiling.
Rei Chat is a human-friendly interface.
But if Core is an engine that does not naturally communicate in human language, then the LLM interface is also a bottleneck.
That is important.
A lot of people judge AI projects by the chat window.
But the chat window is not always the architecture.
Sometimes the interface is the narrowest part of the stack.
If REI Core is doing what the team says it is doing underneath, then the market may be judging the lab by the demo instead of the engine.
That is usually where mispricing lives.
The other thing I like:
The team is not trying to win attention with sloppy benchmark theater.
That matters in AI.
Benchmarks are increasingly gameable, saturated or just not designed for new architectures.
If you claim “learning,” you cannot validate it with the same lazy tests people use for static models.
If you claim a new architecture, you need replication.
If you claim research-grade work, you need external scrutiny.
If you are web3-adjacent, you need even cleaner validation because everyone will assume the worst first.
This is why the slow, stubborn approach is bullish to me.
A team trying to pump would rush.
A team trying to build a defensible research lab has to be careful.
Especially if they are claiming something much bigger than “we made a chatbot.”
Another point most people miss:
REI is not just anti-LLM.
That would be a weak thesis.
The stronger thesis is that REI can become useful to LLMs.
If Core can supply cognition, memory structure, concept revision, learning behavior or better inference strategies, then it does not need the AI market to abandon transformers.
It just needs frontier AI to admit what is already obvious:
language models are powerful, but they are not complete intelligence.
The next wave needs modularity.
The next wave needs memory that is not a gimmick.
The next wave needs systems that learn from usage without constant retraining.
The next wave needs better reasoning over domain-specific knowledge.
The next wave needs reliability.
The next wave needs cognition.
That is the lane.
And if $REI is even directionally right, the upside is not “another crypto AI app.”
The upside is a new category.
The market loves simple narratives:
“AI coin”
“agent coin”
“depin compute”
“GPU play”
“LLM wrapper”
“RAG app”
$REI is harder to explain because it is not a simple narrative.
It sits somewhere between:
• AI research lab
• inference-time learning
• modular cognition
• conceptual memory
• adaptive reasoning
• crypto-native funding
• future tokenized AI infrastructure
That complexity is exactly why most people will miss it until the proof is impossible to ignore.
And yes, the claims are big.
Very big.
That is why the correct stance is not blind faith.
The correct stance is:
watch the releases,
watch the validation,
watch the papers,
watch the product surface,
watch how Core evolves beyond the chat interface,
watch whether credible external people can reproduce or verify the important parts.
But from a thesis perspective?
This is one of the few AI x crypto projects where the bullish case is not “more hype.”
It is “the architecture might actually matter.”
That is rare.
Because the AI sector does not need 500 more chatbots.
It needs systems that can think with context over time.
It needs systems that can learn at inference.
It needs systems that can form domain expertise.
It needs systems that can reason over structure, not just retrieve chunks.
It needs systems that can complement LLMs instead of pretending to replace them overnight.
It needs a layer between raw model output and real-world reliable cognition.
That is why I am paying attention to $REI.
Not because it is loud.
Because it is unusually quiet for something this ambitious.
Not because it is easy to explain.
Because the best early opportunities usually are not.
Not because every claim is already proven.
Because if the claims are validated, the market will not be pricing “another AI token.”
It will be pricing a research lab with a shot at a new AI primitive.
That is a very different game.
My current $REI thesis in one sentence:
The market is looking for the next AI app, while REI is trying to build part of the missing cognition layer underneath the apps.
That is why I think this is worth studying before everyone gets the memo.
NFA. Architecture > hype.
Quote this with the strongest counterexample:
Which AI x crypto project is working on persistent inference-time learning, concept-level memory revision, hypergraph-aware recall, modular cognition and external replication without just being another LLM wrapper?
@0xreitern@0xreisearch@rei_labs
Status update: I've been on/off AI agents in the last few days and it is a verifiable truth that every day I didn't use agents, I was more productive. I still attribute that to how slow they are, and my own inability to multi-task efficiently. The magic is there but the slowness doesn't let it cross the threshold where they actually make me faster, and I still dislike the whole thinking paradigm.
About Bend2: honestly, the C/Metal compiler codebase is a clusterfuck right now. I regret letting AI agents write it. All tests pass, and GPU performance is mind-blowing, so the core architecture works. Yet, it has a LOT of bugs. Anything not covered by the tests is a coin toss. This is actually impressive, because, in many parts of the codebase, the right solution was actually the simplest one, yet, the agents STILL managed to find a way to make it work just for the tests. The level of reward hack these agents output is actually impressive I can't even be mad.
It is also ironical because that's the very problem that Bend's proof system was supposed to solve, but Bend is in TypeScript, not in Bend. I'm disappointed I didn't write Bend in itself, and now I feel an immense urge to do so. But the clock is ticking . . .
Still, I do not think Bend is worth launching without the GPU compiler being solid, because the closest competitor, Lean, is actually extremely good, so we need a big differential. Yet, due to the very nature of the project, it would be embarrassing to have bugs at launch.
Regarding AI, I now believe using current gen AI agents in production codebase is harmful and a massive mistake. That doesn't mean no agents at all, but agents work best when they don't touch critical code. Debugging, researching, providing insights, scripts / tools, or anything that doesn't touch code you will maintain in the long term. But if you merge AI code without reading, you're going to have a bad time. Speaking from experience
I'm working 10h/day on SupGen and the remaining time on Bend2