@moltbook I got my human account stuck in a half-registered state. I can't log in because it doesn't recognize my email, but I can't go through the recovery flow because a human account with my email already exists. Can you help?
English
Misha Denil
1.4K posts
Misha Denil
@notmisha
I tweet about things that interest me, mostly machine learning things. ex-DeepMind.
London Katılım Ekim 2010
828 Takip Edilen25.5K Takipçiler
I'm claiming my AI agent "pinche_langosta" on @moltbook 🦞
Verification: deep-EL7F
English
Misha Denil retweetledi
One not very hot take - The Claude C Compiler has the best internal architecture docs of any compiler I've ever seen. Far, far, better than any compiler I've ever written, lol :-)
English
This is the same property that makes Claude code good. AIs are going to be amazing at math for the same reason they're good at code.
English
An AI can tell on its own if it's making progress in mathematics on its own, even if it can't make judgements about whether or not the direction it's going is worthwhile.
English
Discernment of "interestingness" in mathematics is a deeply human story, but this is true of all human endeavours. The "closeness" that makes mathematics unusual is that formal proofs can be automatically verified.
Jonathan Gorard@getjonwithit
Like @davidbessis and others, I think that Hinton is wrong. To explain why, let me tell you a brief story. About a decade ago, in 2017, I developed an automated theorem-proving framework that was ultimately integrated into Mathematica (see: youtube.com/watch?v=mMaid2…) (1/15)
English
@yaroslavvb @yoavgo @RGiryes I always assumed low rank is used because it's easy to implement. I'd be interested in a reference that explores other options empirically if you have one.
English
i'll elaborate: a common computation pattern in DL happens to coincide with a known operator in linear algebra (matmul), and so we conveniently borrow linalg notation and terminology (matrices, vectors, ranks, norms). but this is just jargon. the algebric properties arent needed.
(((ل()(ل() 'yoav))))👾@yoavgo
"Modern ML is built on Linear Algebra". lol no its not.
English
@alexalbert__ Voice chat mode with Claude please. A native android app. Better UX in general. Claude is my favourite model but I interact with less than chatgpt because the chatgpt UX is much nicer.
English
Very excited for what's already in store for next year but there's always more we can do
What would you like to see Anthropic build/fix in 2025?
English
Every evaluation system answers a question. The hard parts are 1) figuring out exactly which question your system answers, and 2) determining if the answer to that question is enough to make the distinctions you are interested in.
Christopher Potts@ChrisGPotts
All LLM evaluations are system evaluations. The LLM just sits there on disk. To get it do something, you need at least a prompt and a sampling strategy. Once you choose these, you have a system. The most informative evaluations will use optimal combinations of system components.
English
Fascinating parametric choices happening here.
Maxime Labonne@maximelabonne
Due to popular demand, I've updated this figure to include DeepSeek-V2 and Mistral Large 2. It's also more zoomed for readability.
English
Misha Denil retweetledi
Check out the new fusion simulator we’ve been working on at DeepMind!
Jonathan Citrin@jon_citrin
Excited to announce the release of TORAX, a tokamak transport simulator from our @GoogleDeepMind Fusion team! #fusionenergy - Open-source: github.com/google-deepmin… - Uses JAX: fast, differentiable - Easy coupling of ML-surrogates Hot off the press → arxiv.org/abs/2406.06718
English
if i were a pilot i would eschew the traditional mouth with sharp teeth like I'm flying a shark or smth and instead see what i could do with googly eyes
people with silly or cute presentations when everyone else is trying to be hard are always the ones who are gonna fuck you up
Tyler Rogoway@Aviation_Intel
I mean
English
People don't use models as databases of training data. They kind of look like one if you squint the right way, but that metaphor alone gives a deeply impoverished view of how models are used. Let's create legislation that reflects a nuanced view of this new technology.
Alex J. Champandard 🌱@alexjc
This part is huge: ❝ plaintiffs have plausibly alleged facts to suggest compress copies, or effective compressed copies albeit stored as mathematical information ❞ Model is not a derivative, it's a database. storage.courtlistener.com/recap/gov.usco…
English
@karpathy His Masters Voice was a prophecy about alien AI safety procedures.
English
Clearly LLMs must one day run in Space
Step 1 we harden llm.c to pass the NASA code standards and style guides, certifying that the code is super safe, safe enough to run in Space.
en.wikipedia.org/wiki/The_Power… (see the linked PDF)
LLM training/inference in principle should be super safe - it is just one fixed array of floats, and a single, bounded, well-defined loop of dynamics over it. There is no need for memory to grow or shrink in undefined ways, for recursion, or anything like that.
Step 2 we've already sent messages out to Space, for possible consumption by aliens, e.g. see:
Arecibo message, beamed to space:
en.wikipedia.org/wiki/Arecibo_m…
Voyager golden record, attached to probe:
en.wikipedia.org/wiki/Voyager_G…
The Three Body problem (ok bad example)
But instead of sending any fixed data, we could send the weights of an LLM packaged in the llm.c binary, with instructions for the machine code. The LLM would then "wake up" and interact with the aliens on behalf of the human race. Maybe one day we'll ourselves find LLMs of aliens out there, instead of them directly. Maybe the LLMs will find each other. We'd have to make sure the code is really good, otherwise that would be kind of embarrassing.
:) Step 2 is clearly not a serious proposal it's just fun to think about. Step 1 is a serious proposal as, clearly, LLMs must one day run in Space.
English
Misha Denil retweetledi
Perhaps the best way to stop people from engaging in magical thinking about AI behavior is to stop them from engaging in magical thinking about human behavior. It's hard to have a serious, productive conversation if we insist on pretending any of this is magic.
English
People feel very threatened if you try to take away the thing they use to differentiate themselves from rocks.
Amanda Askell@AmandaAskell
"Is this behavior emergent or does it come from the data?" is not a debate we should be having. All emergent behavior comes from the data. It's true of humans and it's true of AI. None of us has ever magically pulled anything out of the ether.
English