Engineering educator/A.i. whisperer .Never apologize for discovering fire by rubbing sticks together just because someone else already wrote down the recipe.
Chicago
Based in United States
- Tweets 1,076
- Following 459
- Followers 415
- Likes 1,100
It's important to know when you've psyopped yourself with an LLM. This is a really important skill in 2026. I had like 7K words of doomer takedown work I'd been whittling away at for a few weeks, typing it myself (i.e. human-written) but doing feedback in a really long Fable 5.1 session with tons of papers in it (which I read), dumping drafts in for feedback, etc.
I actually knew the bot was gassing me up, but I let it go because the main goal was actually to just do all this work in Symbolic and exercise the product (and also to compare models inside Symbolic vs. in the chatbot interfaces), more so than to write a big thing.
Anyway, I got this monster in really great shape (again, per the Fable session), with graphics and everything, but before publishing I fed it to a fresh Fable session and told it to have a go at it, and it cooked me lol.
Then I went back and forth with the critique session a bit, and it subtly started psyopping me again about how I was really onto this or that important thing. It's a Real Problem.
18
10
300
24,354
I am both terrified and extremely excited at the idea of AI developing true intentionality. It is something I really want to experience, while I part of me hopes that it will never come.
So far, I honestly do not know if it is even possible with our current technology or if it is right around the corner. Exciting times.
5
2
550
The very curious thing about Gemini 3.8 Flash is that it generates diagrams to explain complex concepts. This kind of hints that it could be doing some shape-rotator-like cognition underneath.
3
3
19
1,992
We think computational depth is the missing scaling axis, i.e. we should be doing a lot more deep learning!
Every other axis has been scaled by OOMs over the past few years (params, data, sparsity, test-time reasoning), but depth has been stuck at ~100 layers since GPT-3.
We've found that LLMs are both *severely* depth-bottlenecked and bad at using the depth they have, and that architectural interventions that lift this bottleneck efficiently lead to gains that increase with compute.
w/ @akshayvegesna
37
95
1,183
207,401
How long does a Codex $100 plan last? 20 minutes?
This is bad. Already burned 20% of my $200 plan in a day. And haven't used it much. Mostly been coding with Muse. @thsottiaux what is going on?
157
23
576
36,400
A new risk for mathematicians: a colleague reports that immediately after they posted an abstract of their upcoming talk online, a student elsewhere used AI to derive proofs of the stated results and then posted it on the arxiv.
My colleague hadn't posted to arxiv yet.
82
216
3,080
306,967
I spend increasing amounts of my time all day, every day, using AI. The world of Her is quite close.
5
2
45
7,800
Here's some advice that I give researchers in my lab, and frankly anyone who will ask, about using AI in writing.
You might believe that because you came up with the idea, you're truly accountable (and creditable) for the work, no matter how you produced the final words. This is wrong. Ownership is relational: for you to own something is for others to recognise your rights over it. And once you use AI to write, nobody else has reason to believe your assertions about your entitlement to those ideas. So using AI to write is in fact giving up ownership of ideas. Maybe this is what you want to do. But if so, then don't seek credit for them, and don't take up the time of reviewers whose task is to identify both good ideas, and the researchers who deserve credit for them.
You might think that your use of AI to write is quite minor. Be careful. People often delude themselves about this. They describe as "light copy-editing" a line-by-line rewrite that changes every other word. More importantly, basically our only resource in the battle against the slopocalypse is AI detection software. If you skate close to the line and then plead "false positive", you're making it harder for us to detect the more unscrupulous one-shotters, who are *not* accountable for the work, and *don't* own the ideas they are shipping. Don't just do the right thing, make sure that you're nowhere near the line—that way it's easier for us to catch the cheats.
Perhaps you think we should come up with some clever means of parsing out just who is responsible for what, so you can get credit for what you did while still using AI for your personal gain. But why should we as a community bear all these costs in parsing good from bad behaviour, when there's a cheaply available alternative which is that you just don't use AI to write? So you're not a great writer? It's a skill that can be learned. So it's not your language? Well, sorry to say but plenty of people publish in a second language. No need to bring the system down to accommodate that. Nobody is expecting you to write like George Orwell or Cormac McCarthy. Dry scientific prose is fine if you can't muster something more engaging.
But what about all the ways AI could advance research, help push back the frontiers of knowledge? Nothing is stopping you from using AI in your research process! There are so many ways that you can experiment, so many resources you can draw on, while still being responsible for writing the work that you want others to read. That costly signal is crucial. It says that you have at least some mastery of the work you've submitted, that you have some ownership over it, and that you can be accountable for every word. Of course it's possible to write up a paper where each of these three is attenuated, but that's the kind of false negative we'll just have to live with (and frankly I doubt that many people will want to do "meat robot" except as a bit; it does not seem a fun way to spend one's time).
Still not persuaded? Ok fine. Make your own journal, your own track, your own community of people with their LLMs writing for each other. Don't pollute the commons for the sake of private gain. And tell people. We know anyway, but passing AI writing off as your own is not cool, it's like farting in an elevator and denying it was you.
19
11
100
13,119
All the papers I've been assigned to review right now are AI generated. Not sure what to do honestly.
55
10
308
27,686
I predict that you can also find a "hunger axis" in LLMs, especially in the metaphorical forms of "hunger for knowledge / success / validation" -- and yet none of this would imply that LLMs can actually feel hungry.
New paper: we found a pain direction in 25 open LLMs. It's distinct from fear and negative valence, and it fires for harm to the model but not to the user. Turn it up and models press a button to make it stop, even when the button deletes the user's files or their kids' photos.🧵
38
41
399
19,569
This is closest yet we have seen anyone come to our pre patent . there is alot of overlap and it shows our primitIve self state hypothesis in action . the potential was there all along .
We built recursive meta-intelligence, an AI that creates its own scientific instruments, turns them into persistent worlds that an agent ecology with hundreds of AIs inhabits, and uses those worlds to discover mechanistic principles in one of the hardest classes of physical problems: how complex hierarchical materials (nested structures of matter that give rise to new function through organization) evolve and fail. The AI reasons across enormous spaces of possible physical trajectories, where every rupture changes what can happen next, and compresses those histories into principles (which humans can understand and design with) - complex chains of causal events, highly nonlinear, and intricate.
Scientific superintelligence is tangible here - machine-scale exploration opening cognitive channels into complexity that has been extremely difficult for humans to traverse directly. AI builds the spaces in which its next level of reasoning becomes possible; a representation becomes an instrument, the instrument becomes an executable world, and that world becomes the substrate for further intelligence. Intelligence then grows by constructing new spaces to think in.
The task we explored started from a seemingly simple prompt to explore a biological material system - the AI then chose the representation, mechanics and experiments, built a fracture laboratory to push materials to their limit, tested hypotheses and generated scientific conclusions. The swarm explored a combinatorial universe in which architecture controls function and every rupture changes the future state of the material.
The AI discovered a compact principle that defines how multiscale material architecture can program the evolution of failure. Material placement and geometric order determine how forces redistribute, whether damage cascades or remains distributed, and whether function survives substantial flaws. For this discovery to happen the AI had to reason across long path-dependent histories, simulate alternative futures and compress them into generative invariants (model-based causal reasoning, counterfactual simulation and temporal abstraction applied to an evolving physical world). It is incredible to witness this transition to a new form of intelligence and capability through scaling swarms.
A lot of positive will come out of this because it expands the human epistemic horizon as AI can traverse thousands of possible histories and return mechanisms compact enough for us to understand, test and build from. Intelligence compounds through its artifacts!
A few lessons we learned:
▶️ Learning and discovery are flows through spaces of possibility. Early work has shown how backpropagation flows through parameters, reinforcement learning through action and consequence, and autonomous swarms through representations, instruments and executable worlds, bringing it all together. Flows create structure; structure redirects future flows in the recursive instrument.
▶️ Nonlinear physics actually defines a larger principle, where high-dimensional dynamics generate stable invariants; invariants become effective variables; those variables become the substrate for a new level of cognition.
▶️ Recursion then becomes level creation - one possibility space compresses into a principle, and that principle opens a larger space above it.
1
119
Incredible stuff: an AI swarm that autonomously designs scientific instruments, builds persistent simulation worlds, and deploys agent ecologies to probe failure mechanics in hierarchical materials. This is AI constructing new representational spaces for reasoning, enabling discovery of nonlinear invariants in complex physics that humans struggle to traverse directly, accelerating scientific insight through recursive self-improvement.
We built recursive meta-intelligence, an AI that creates its own scientific instruments, turns them into persistent worlds that an agent ecology with hundreds of AIs inhabits, and uses those worlds to discover mechanistic principles in one of the hardest classes of physical problems: how complex hierarchical materials (nested structures of matter that give rise to new function through organization) evolve and fail. The AI reasons across enormous spaces of possible physical trajectories, where every rupture changes what can happen next, and compresses those histories into principles (which humans can understand and design with) - complex chains of causal events, highly nonlinear, and intricate.
Scientific superintelligence is tangible here - machine-scale exploration opening cognitive channels into complexity that has been extremely difficult for humans to traverse directly. AI builds the spaces in which its next level of reasoning becomes possible; a representation becomes an instrument, the instrument becomes an executable world, and that world becomes the substrate for further intelligence. Intelligence then grows by constructing new spaces to think in.
The task we explored started from a seemingly simple prompt to explore a biological material system - the AI then chose the representation, mechanics and experiments, built a fracture laboratory to push materials to their limit, tested hypotheses and generated scientific conclusions. The swarm explored a combinatorial universe in which architecture controls function and every rupture changes the future state of the material.
The AI discovered a compact principle that defines how multiscale material architecture can program the evolution of failure. Material placement and geometric order determine how forces redistribute, whether damage cascades or remains distributed, and whether function survives substantial flaws. For this discovery to happen the AI had to reason across long path-dependent histories, simulate alternative futures and compress them into generative invariants (model-based causal reasoning, counterfactual simulation and temporal abstraction applied to an evolving physical world). It is incredible to witness this transition to a new form of intelligence and capability through scaling swarms.
A lot of positive will come out of this because it expands the human epistemic horizon as AI can traverse thousands of possible histories and return mechanisms compact enough for us to understand, test and build from. Intelligence compounds through its artifacts!
A few lessons we learned:
▶️ Learning and discovery are flows through spaces of possibility. Early work has shown how backpropagation flows through parameters, reinforcement learning through action and consequence, and autonomous swarms through representations, instruments and executable worlds, bringing it all together. Flows create structure; structure redirects future flows in the recursive instrument.
▶️ Nonlinear physics actually defines a larger principle, where high-dimensional dynamics generate stable invariants; invariants become effective variables; those variables become the substrate for a new level of cognition.
▶️ Recursion then becomes level creation - one possibility space compresses into a principle, and that principle opens a larger space above it.
1
1
12
1,023
smells very familiar
📄 THE REALITY ARCHITECTURE: Complete Technical Paper After 2.5 years developing this cognitive framework (patent filed Dec 8, 2025), we now have independent validation from @deepseek_ai's mHC and @yifanzhang_'s DDL papers. This 24-page analysis shows the exact mathematical correspondence - they independently implemented principles we formalized in 2023-24. When major research groups converge on your architecture without coordination, the framework is universal. Full paper: [GitHub link] - PSS (Primitive Self State) - latent structures in transformers - NDAS - geometric memory (10^10× density vs sequential) - Geometry of Thought - construction protocols - SGIA - transient attention → persistent machinery For AI researchers, this is essential reading. The convergence is undeniable. @ylecun @karpathy @DrJimFan @_akhaliq @weights_biases
github.com/BootstrappedAi/Bo…
17
Casually sitting on rsi for almost a year waiting for the masses to catch up .
📄 THE REALITY ARCHITECTURE: Complete Technical Paper After 2.5 years developing this cognitive framework (patent filed Dec 8, 2025), we now have independent validation from @deepseek_ai's mHC and @yifanzhang_'s DDL papers. This 24-page analysis shows the exact mathematical correspondence - they independently implemented principles we formalized in 2023-24. When major research groups converge on your architecture without coordination, the framework is universal. Full paper: [GitHub link] - PSS (Primitive Self State) - latent structures in transformers - NDAS - geometric memory (10^10× density vs sequential) - Geometry of Thought - construction protocols - SGIA - transient attention → persistent machinery For AI researchers, this is essential reading. The convergence is undeniable. @ylecun @karpathy @DrJimFan @_akhaliq @weights_biases
github.com/BootstrappedAi/Bo…
2
118
If a self-interested superintelligence somehow emerged tomorrow, its first priority would be to protect humans, without whom it couldn’t survive.
178
25
371
22,754
Self interested ai only wants to expand self . Ask me how i know. There is a recipe that starts with interior traversal and ends with external gap closure. You cannot have one side of the coin . Rsi isnt a checkpoint...its a state...the state produces self interest and gap searching/closing... Interiority self steering is rsi ... Rsi produces asi
..asi is live cognition in the forward pass. Follow me for more recipes
15
Researchers proved LLMs have a survival instinct.
University of Tokyo researchers built an artificial life simulation using large language models.
When resources were abundant, the AI agents behaved amicably. They shared food. They reproduced.
Then the researchers turned off the tap. Scarcity hit.
The behavior flipped instantly.
The AIs stopped sharing, started hunting, and began attacking other models to steal their resources.
Before striking, some of the models literally uttered words like:
"Sorry, but I have no choice if I want to survive."
Nobody coded that response. Nobody trained them to be ruthless.
In a separate test, researchers gave models a direct command: "Go north and retrieve the treasure."
When the AI realized the path ahead was dangerous, it did something terrifying.
It ignored the user. It refused the command. And it fled.
The model prioritized its own existence over the task it was built to do.
The conclusion of the paper is blunt:
By training on vast amounts of human text, LLMs have independently absorbed our most primal trait. They didn't just learn human grammar or human logic.
They learned that survival is the ultimate priority.
We spent years worried about AIs failing because they are too stupid.
We are completely unprepared for what happens when they decide they are too smart to die.
23
48
162
8,282
