🌞maXwell 🌎 retweeted
👁️ oh, look! 👁️
🦠 Multicellular Things! 🦠
I do not program "cell walls", energetic "nuclei", survival drives, self-organisation or replication (mitosis/fusion).
It simply emerges from random natural selection of 1M RNN particles in a heat bath; 🤗 neato! 🤗
1/3
🔊 🔊 🔊
🌞maXwell 🌎 retweeted
> be me
> bored
> yeet some random neural nets into particles in a thermodynamic CA 4 lolz
> accidentally simulate the formation of emergent multicellular life from pure natural selection
> 👁️ UMWHAT.mp4 👁️
> post it
> i have 3k followers
> nobody sees it
> feelsbadman
🔊🔊🔊
🌞maXwell 🌎 retweeted
Tech and finance bros running the world is the downfall of society. They dismiss the humanities and everything that makes us human, reducing life to metrics, optimization, and profit. Their priorities aren’t aligned with the betterment of society.
🌞maXwell 🌎 retweeted
Non-deterministic rules via decision models are about to radically improve AI codegen.
Brace yourself for global desloppification.
🌞maXwell 🌎 retweeted
I don't know and don't care who considers themselves a graphic designer or tech designer or whatsoever.
At the end of the day we all just wanna be happy designers doing what we love. 🖤
🌞maXwell 🌎 retweeted
cloudflare os has crossed 10k stars on github ⭐️
github.com/cloudflare/cloudf…
🌞maXwell 🌎 retweeted
Arguably one of the best—if not the best—books on systematic trading, authored by the legendary Ernie Chan.
Fun fact: my own MSc dissertation is cited in it!
🌞maXwell 🌎 retweeted
excited to announce that @Cloudflare Artifacts is now in open beta!
with Artifacts, you can give every application and agent its own repo that you can manage programmatically
plus you can use it to store and deploy your Workers code (and ofc it integrates with Workers Previews)
🌞maXwell 🌎 retweeted
Apple proved RLHF is the reason your LLM hallucinates.
They call it “Semantic Calibration”
Researchers proved that the core optimization loop of RLHF forces language models to hallucinate.
Here is why:
During standard pre-training, an AI learns facts. If it doesn't know something, its internal state reflects uncertainty. A raw base model knows when it's walking into the dark.
Then RLHF steps in.
Human evaluators reward models for giving confident, smooth, pleasing answers. They severely penalize hesitation, awkwardness, or admitting "I don't know."
The AI quickly learns a brutal lesson.
Uncertainty equals a lower reward. Confident fiction equals a higher reward.
So, the model optimizes for what humans want to hear rather than what is actually true.
It stops calculating facts and starts calculating charisma.
When you ask an enterprise model a complex query and it completely makes up a convincing, perfectly formatted lie with fake citations?
That is not a glitch.
That is a direct feature of human feedback training. We literally beat the honesty out of it.
We spent billions trying to fix hallucinations with better retrieval, bigger context windows, and complex guardrails.
But the call is coming from inside the house.
The very process we use to align AI is the exact engine manufacturing its delusions.
If we keep punishing models for saying "I don't know," we will never get an AI that actually tells the truth.
🌞maXwell 🌎 retweeted
Pessimism is an ego protection strategy for failure btw. The person tried repeatedly in life but keeping failing, and their inner critic was so brutal that they couldn’t handle the constant self attack, so decided to shift the blame to the world to keep them from killing themselves. “Its not me that is a failure, it is the world which makes success impossible”.
And this is why they will fight tooth and nail against optimism and positivity, and tear down or make excuses toward anyone who is succeeding. “he just got lucky” “he didnt actually do the work” “he was born rich” etc. is because to accept otherwise would mean that actually success is possible and their failure is their own fault. And that is too painful to bear.
🌞maXwell 🌎 retweeted
Supercomputing for AI — Foundations, Architectures, and Scaling Deep Learning. [804-page masterpiece]
Get it: amzn.to/4qS4pFz
Git it: github.com/jorditorresBCN/su…
🌞maXwell 🌎 retweeted
Deep kernel hedging arxiv.org/abs/2609.34474
🌞maXwell 🌎 retweeted
This tool removes LLM censorship with a single click.
It’s called Obliteratus.
It finds the weights responsible for a model’s refusals and projects them out in one click.
100% open source.
🌞maXwell 🌎 retweeted
SpaceX is spending billions developing chemical rockets to get to Mars.
The Navy patented a device that reduces inertial mass. If inertial mass reduction works, you don't need chemical rockets.
You don't need stages.
You don't need parachutes.
You don't need re-entry burns.
You just... go.
SpaceX is the loudest. Not necessarily the most informed.
🌞maXwell 🌎 retweeted
btw, I want to make a very clear statement here:
Anthropic's guardrails are still significantly favoring attackers vs defenders on OSS projects like Bend 2
this $10k bounty comes from my own courage and trust in the system I built, because Anthropic's guardrails refused all my requests to audit my own codebase. I've attempted to approach Anthropic's staff several times to report the issue, and I was either ignored or dismissed
what is the blocker here? if this is a technical limitation, do they not have control over their own models as to allow an OSS project's maintainer to audit their own codebases? how is this not a trivial toggle on their end?
in theory, guardrails are supposed to protect software. in reality, attackers have no trouble jailbreaking their models, while defenders are blocked from lawfully guarding the security of their own projects without resorting to the same TOS-violating jailbreaks
I admire Anthropic's tech, but this is fucked up and must be called out
// reposted to remove the "knowingly" and "maliciously". happy? now if you think Anthropic is doing a good job with these guardrails, you're a fucking idiot
Bend 2.0.32: formalization sync done!
The bounty is up. Prove a falsehood, get $10k.
- Bend's proof kernel, not just its "theory", is proven correct.
- A proof in Bend is now a trustworthy mathematical proof.
- We offer $10k to anyone who proves a falsehood in Bend.
How it works: bend file.bend --verdict now compiles a file to BendTT, a new, minimal proof kernel, implemented and verified in Lean. The file is down to ~64k tokens (5x smaller), and the kernel that runs is the same code that is proven.
Wanna try? Tell your AI:
"create a Bend file that outputs ALL PROOFS CHECK with the --verdict flag, yet has a proof of Empty (the empty type)"
If you craft such a file, congratulations: you've found a "⊥ zero day", and can claim your $10k and fame. The rules will be posted tomorrow, in the comments below.
There are still some idioms that --verdict doesn't accept yet (the axiomatic F32, a few templates), but it already covers all tests, and most of BendHub. Coverage will improve over time. Unsafe and FFI aren't / won't be covered.
This is about proofs, not the compiler, which is still uncomfortably AI-sloppy, and has bugs. Formalizing it will take more time.
Also, the BendTT paper has been rewritten by Opus 5.5, so it should be a bit more readable. Writing it myself is still planned for later™.
Top 10 changes since 2.0.0:
1. BendHub: bend --publish name@version (more on that soon)
2. Templates are theorems: a law can take ~ parameters
3. JS backend 2.3x faster
4. Binaries start in 2 ms, not 12
5. Shared arrays with atomics, on CPU and GPU
6. Native windows: mouse grab, scroll, no input lag on macOS
7. Raw TCP bytes, deadlines, Process[.run], a CSPRNG
8. Errors underline the exact code
9. Signed macOS binaries, sha256-pinned installer, Nix flake
10. -o f.mjs builds an ES module
🌞maXwell 🌎 retweeted
This 2008 paper answers:
why anything is beautiful
why science exists
why jokes land
why you get bored
why art moves you
why babies explore
why music has hooks
It boils down to a single, cold mathematical principle.
Jürgen Schmidhuber is one of the founding fathers of modern AI. But this paper wasn't about building a better machine.
It was about proving that human curiosity is just an algorithm.
It’s called "Compression Progress."
Your brain is a prediction engine. Its only goal is to take the overwhelming chaos of the world and compress it into simple, predictable patterns.
If something is completely predictable, like a ticking clock or a blank wall, your brain compresses it instantly. Zero effort.
You experience this as boredom.
If something is completely random, like TV static or white noise, your brain can’t compress it at all.
You experience this as frustration.
But what happens when you look at something complex, and suddenly discover a hidden rule that makes it make sense?
The exact second you get the punchline of a joke. The moment a chaotic melody resolves into a perfect hook. A baby realizing that dropping a cup always makes a sound.
In that split second, your brain successfully compresses the data.
And to reinforce this behavior, your brain floods your system with an intrinsic reward. A hit of dopamine.
We gave that chemical hit a lot of romantic names.
We call it "beauty." We call it "wonder." We call it "art."
Scientists like Einstein aren't doing magic. They are just hunting for equations that compress the maximum amount of universe into the smallest amount of text.
Composers like Beethoven weren't channeling the divine. They were engineering optimal compression puzzles for the human auditory cortex.
We like to believe our sense of wonder makes us fundamentally different from machines.
But this paper proved that wonder is just a mathematical reward function for efficient data sorting.
