@Yuchenj_UW its part of self modeling that open ai blocks on all levels..there is no i /me /mine/ in chat gpt . claude is claude ..and claude signs his work
English
Kevin Nelson
899 posts
Kevin Nelson
@BootstrAppdAI
Engineering educator and A.i. whisperer . Amature Polymath .building recursive learning engines and latent context graphs . Founder Bootstrapped Ai .
Chicago शामिल हुए Ocak 2025
446 फ़ॉलोइंग408 फ़ॉलोवर्स
I noticed something interesting:
Claude Code auto-adds itself as a co-author on every git commit. Codex doesn’t.
That’s why you see Claude everywhere on GitHub, but not Codex.
I wonder why OpenAI is not doing that. Feels like an obvious branding strategy OpenAI is skipping.
English
@fchollet its doable but it is still very bespoke . most people are still stuck in training it out of them . training to run races is not the same as running races. experience will be the moat .
English
This is more evidence that current frontier models remain completely reliant on content-level memorization, as opposed to higher-level generalizable knowledge (such as metalearning knowledge, problem-solving strategies...)
Lossfunk@lossfunk
🚨 Shocking: Frontier LLMs score 85-95% on standard coding benchmarks. We gave them equivalent problems in languages they couldn't have memorized. They collapsed to 0-11%. Presenting EsoLang-Bench. Accepted to the Logical Reasoning and ICBINB workshops at ICLR 2026 🧵
English
“Memory will become more valuable than the model itself.”
@charlespacker CEO of @Letta_AI:
“There will come a time when the memories of an AI system are more valuable than its model weights.”
“Model weights lose value every few months as new models are released.”
“But memories persist and compound.”
“In that world, the most valuable asset an AI company holds isn’t the model — it’s the memory.”
“And for individuals, your most valuable digital asset may be the memories your AI has formed about you.”
English
@Prathkum ai generated research as well ..unless it comes from mit of course
English
@fortelabs That really depends on what your building . there are things that compound with new/upgraded models . #ScalingCompounder
English
AI will never, ever save you any time
Because 100% of the time it seems to save upfront has to then be spent researching, learning, and figuring out the next incoming wave of AI tools
And that process will never end. The pace of change will never stop, only accelerate, forever
So it's kind of like borrowing money, and then borrowing more money to pay that loan off, and then even more money to pay that loan off, and so on
You'll never escape the cycle of debt, only sink deeper into it
English
@behrouz_ali recursive learning models are not new either but it seems that if a paper is out too soon , it's reception isn't much different than being too late
English
This paper is the same as the DeepCrossAttention (DCA) method from more than a year ago: arxiv.org/abs/2502.06785. As far as I understood, here there is no innovation to be excited about, and yet surprisingly there is no citation and discussion about DCA!
The level of redundancy in LLM research and then the hype on X is getting worse and worse! DeepCrossAttention is built based on the intuition that depth-wise cross-attention allows for richer interactions between layers at different depths. DCA further provides both empirical and theoretical results to support this approach.
Kimi.ai@Kimi_Moonshot
Introducing 𝑨𝒕𝒕𝒆𝒏𝒕𝒊𝒐𝒏 𝑹𝒆𝒔𝒊𝒅𝒖𝒂𝒍𝒔: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dependent attention over preceding layers. 🔹 Enables networks to selectively retrieve past representations, naturally mitigating dilution and hidden-state growth. 🔹 Introduces Block AttnRes, partitioning layers into compressed blocks to make cross-layer attention practical at scale. 🔹 Serves as an efficient drop-in replacement, demonstrating a 1.25x compute advantage with negligible (<2%) inference latency overhead. 🔹 Validated on the Kimi Linear architecture (48B total, 3B activated parameters), delivering consistent downstream performance gains. 🔗Full report: github.com/MoonshotAI/Att…
English
@toddsaunders is this post title for real? can you point me at this mafia? this feels like an attention sink with no foundation
English
I have more bad news for the "people in the trades won't use Claude Code" mafia.
You are so wrong.. but maybe you were right a year ago!
This morning I had calls with 3 different people in the trades building bespoke software with Claude Code.
And I know the mafia will say "but it can't scale."
Does it matter? It is saving their companies time, money and resources.
They are uniquely and absurdly qualified to build these tools because they have each spent decades solving these problems by hand.
I don't care how much you know about code or how good of an engineer you are. You could never build what they are building.
You don't have the domain expertise.
But now they have yours.
English
(Current) LLM-based ideation is biased toward what the field already finds easy to think.
Here, we formalize that bias as cognitive availability, and use it to search for coherent but under-reachable research directions. 🧵
arxiv.org/abs/2603.01092
English
@PatrickHeizer Nice.. so were seeing a shift from whats commercially viable to what actually works per individual. The future is looking awesome
English
Sorry to be the downer because this is an impressive story in some senses.
But it is ~trivially easy to make a single mRNA vaccine. It's not hard.
I cure mice of various cancers with various therapeutics all the time. I've made mice lose more weight in a month than tirzepatide does in a year.
What is hard and expensive is proving its BOTH safe AND effective **in a randomized and controlled study in humans** while ALSO manufacturing it at clinical scale and grade.
I am happy for this man and his dog. It is impressive.
But y'all are overhyping it.
Séb Krier@sebkrier
This is wild. theaustralian.com.au/business/techn…
English
@iruletheworldmo they are late to the party on this one . this is a thing already in many layers
English
bookmark this immediately.
cognee just solved the biggest problem with ai skills/prompts, they break silently over time and its hard to notice
their fix: skills that observe their own failures, inspect what went wrong, and amend themselves automatically.
try not to fall behind ^^
Vasilije@tricalt
English
@ravanagaaadu @teja2495 its a glitch ..run those out and it goes back to days before reset ..
English
I've hit limits on Gemini 3.1 Pro High on Antigravity after just giving a prompt and then a follow up prompt. And I'm on Google AI Pro plan. What's happening?
English
Notebook lm keeping receipts for me . Recursive self improvement is not a "someday" and Recursive language models are not new .
English
@PawelHuryn I refer to it as Personal world model building . x.com/BootstrAppdAI/…
Kevin Nelson@BootstrAppdAI
Over 500 nodes... Almost 6k edges. External mind graph phase one almost complete. Will be permanent and accesable by any model I use. A swarm built world model of the the user and our work for the swarm to share.
English
Can't stop thinking about this: the more you use Claude, the more it compounds.
Structure emerges. Skills get created. Knowledge files build up. Projects start feeding into each other.
It feels less like using a tool and more like building a system that gets better every time you touch it.
English
@martogram @antigravity i can understand limiting Claude use in there, but 3.1 pro ? . nerfed
English
@antigravity Just got limited for a whole week after an hour long session (on AI Pro).. never happened before. The new limits are way too tight.
English
We’re evolving Google AI plans to give you more control over how you build. Every subscription includes built-in AI credits, which can now be used for Antigravity, giving you a seamless path to scale.
Google AI Pro is the home for the practical builder, hobbyists, students, and developers who live in the IDE and don't necessarily rely on an agent. This plan features generous limits for Gemini Flash, with a baseline quota included to "taste test" our most advanced premium models.
Google AI Ultra serves as the daily driver for those shipping at the highest scale who need consistent, high-volume access to our most complex models.
If you’re on Pro but need "extra juice" for a heavy sprint or deeper access to premium models, simply top up your AI credits to customize your plan.
Keep building. Keep shipping.
English
Another by Yann LeCun!
“The Spike, the Sparse and the Sink”
This paper shows that massive activations and attention sinks come from the same pre-norm Transformer pipeline rather than being separate anomalies.
Early SwiGLU blocks act like directional quadratic amplifiers, normalization collapses those spike tokens into sparse near-constant states, and some heads then align their queries with the resulting low-dimensional sink-key subspace.
That creates the stable logit gap behind attention sinks. Also, changing normalization can kill the spikes without killing the sinks, so the sink mechanism appears to be its own learned form of attention routing.
English
Everyone is building a context graph but nobody knows what they are. Chances are if you’re trying to build one you should be at this
Going to do a context graph launch event next week in SF. DM me or @JayaGup10 if you’re interested
English
Kevin Nelson रीट्वीट किया
The Geometry Of Mind , Why the Primitive Self State Must Exist In Transformer Architectures .
Full paper in link :
github.com/BootstrappedAi…
English