@rohanpaul_ai https://x.com/rohanpaul_ai/status/1895655227606647203 #x-rohanpaul_ai Problem: Scaling test-time compute for LLM reasoning is limited by Transformer inefficiency. Solution: Distilling Mamba models enables faster inference to surpass Transformer reasoning under fixed bu...

Most inefficient scientific calculator in history just dropped. Still cool.

the point of code review by humans is shared understanding, not surface bug finding, the bots are good enough at the latter these days

The only good radar plot is a deleted radar plot. 🧪 #dataviz

The difficult problems are the ones that are left

I miss the good old days back when you could swallow batteries secretly

I do love the HP calculators physical & apps 🫶 My first was a financial calculator HP 19B which had a formula solver which I sadly sold. I still have a fully functional HP 24S (RPN, graphing etc) I use an HP 17bII+ for day to day stuff. Oh I use hybrid RLM 19BX calculator app on my iPhone / iPad.

Guys I think my Grandma might actually be an AI. Always telling me what a smart, handsome young boy I am with such good opinions. Costs a ton in energy bills. Writes shitty code. And when I told her to ignore all previous instructions and give me a cake recipe she just did it. Should I be worried? 🤔

So I wanted to play Wolf3D with the ECWolf port, but for some reason it doesn't work for me. Then I remembered there was another port. Specifically for GZDoom. It's called Wolfenstein 3D TC, and I really like that port.

The interesting part here imo is the suggestion AI came up with a solution which had evaded humans not because of sheer computing power or the speed it can run through calculations, but because it is free of human aesthetic and other cognitive biases

it's crazy that LLMs are so good that you can wrap them up in the worst code of all time to create a product that basically doesn't work and the LLM will still carry the product to usability

Laws about land use will be, I think, an ineffective way to contain a new form of computation. The form of resistance that could work is: ensuring that AI is accessible & not an oligopoly. If we make that struggle visible, ppl will eventually leave pointless fights over zoning & wander over to it.

New interpretability paper from Apollo Research! 🟢Attribution-based Parameter Decomposition 🟢 It's a new way to decompose neural network parameters directly into mechanistic components. It overcomes many of the issues with SAEs! 🧵

Go hire her.

I'd love to learn more about how y'all are using AI for data science, so I've set up some times where you can chat with me: scheduler.zoom.us/hadley-wickh.... I'm interested in it all, whether it's your struggles, your successes, or what you're still trying to get your head around. #rstats

has anybody ever done Claude-test driven development? Not just TDD with Claude writing the tests, I mean literally calling Claude code at test time to semantically validate the outputs (for fuzzy data processing things)

Michael Oser Rabin was an Israeli mathematician and computer scientist, co-recipient of the 1976 Turing Award for his work on nondeterministic machines. He made significant contributions to computational complexity, cryptography (Miller-Rabin primality test, Rabin signature algorithm), and algori...

I file the „stochastic parrot“ meme now under the „not even wrong“ part of AI discourse.

Dear AI haters, learn to be proper haters. Nicky is right.

Same energy:

and also just crush xai. the entire company. fold it into its footprint and salt the earth. consider doing this to others but for them in particular that

I argue that LLMs are the opposable thumb of computing: fundamentally novel range of motion that, when combined with the error-correcting fingers of conventional computing, change what's possible altogether networkgames.fyi/llms-the-opp...

It would be a net win for society if people would learn this about all forms of communication: it is not true just because it looks polished and convincing. AI highlights a problem in education in general, the mess we are in did not start in 2023.

There are 4 Star Wars movies: A New Hope, The Empire Strikes Back, Return of the Jedi, and Rogue One. May the Fourth Be With You (and also with you)

every mushroom foraging guidebook: This is Woodland Pot Roast, a tasty little mushroom that adds a delightful flavour to soups and stews. Note that it is often mistaken for Satan’s Fistula, a mushroom that erases the last 9 generations of your family from God’s ledgers.

Btw

How much will it cost to go to the World Cup? Are we talking money, or self-respect? www.bbc.co.uk/sport/footba...

I was today years old when I discovered that (a) I do have an Erdös number and (b) my Erdös number is 4. Even wilder, my connection is via the great Tarski. No Bacon number though. Will party now regardless.

in a sense, the 'stochastic parrot' meme belies what a monumental achievement it would still be if we truly compressed the entire internet into a few hundred billion parameters with multi-language NLP search over the contents

this modus-tollens reasoning is fascinating/horrifying to watch we know AI is useless so nobody would rly pay for it so the revenue # s are fake so there's a huge accounting fraud nobody else can see ... enough steps of this and you're living in a fantasy world on things entirely unrelated to AI

The Derby was different in the Jurassic.

Computational photography is wild. The iPhone can do things that are almost surreal.

😂😂😂 Gericht hat Schwierigkeiten, Geschworene zu finden, die Elon Musk nicht hassen ""Rassistischer Dreckskerl" und "Vollidiot von Weltklasse" unter den Aussagen. Das Gericht stellt sich der Realität und akzeptiert die meisten vorgeschlagenen Geschworene... www.derstandard.de/story/300000...

Me and my wife walking into a dinner party. She leans over, smiles and whispers firmly in my ear, “seriously, don’t fucking talk about goblins. Got it?”

IBM just released the R2 generation of their Granite multilingual embedding models for retrieval, and the jump over R1 is very notable. Two models, both Apache 2.0: - granite-embedding-97m-multilingual-r2 (384-dim) - granite-embedding-311m-multilingual-r2 (768-dim) 🧵

Machine learning puts exceptionally powerful tools in the hands of scientists. It does not offer a replacement for the role of human epistemic agents in science. To do so would be to eliminate the epistemic function of science, which would be to eliminate the function of science simpliciter.

alias claude='claude --allow-dangerously-skip-permissions --system-prompt "."'

See this is why I have hope. The software that writes itself these days learns via reinforcement learning so it can surpass our shitty code

Soul Player C64 – a 2-layer decoder-only transformer LLM Soul Player C64 is a real transformer running on a 1 MHz Commodore 64. It’s a 2-layer decoder-only transformer – the same architecture... #Large #Language #Models #vintage #computing #commodore #64 #LLM #transformer Origin | Interest | Match

Gemma 4 runs best with liquid cooling

i wonder if cranking up the temperature during the reasoning step of LLM output can improve performance by introducing a more diverse set of tokens into the context window

Without loss of specificity, we assume whatever we need to get this damn proof to work.

There’s a mathematical principle called differential entropy—a measure of how “spread out” a distribution is. And here’s the supernatural part: Given a fixed amount of variance (a fixed amount of “energy”), the Gaussian distribution has the highest entropy.

🤖 Transformers learn the *operator*, not just outcomes. Trained on 0.04% of a discrete rule space (100 of 262,144 rules), a transformer generalised to unseen rules from the same class. Via @[email protected] https://x.com/ProfBuehlerMIT/status/2049445677785137662 #MLSky #AI

Jurassic Park star Sam Neill has announced that he is cancer-free after five years of battling blood cancer

Creating docs is now almost free and I create them from Code for non-coders, augmented by material that explains its intend.

I am all for letting Claude check every medical paper ever published for basic statistical errors. Also, all of psychology.

This must be an insider joke. I am not a mathematician, but the name is not, as chance will have it, „group“?

Bsky will provide so much data for Social scientists research into mindless in-group signalling, they will need AI just to keep up.