Lazy evaluation and unified memory in MLX make it super easy to quantize or merge large models on small machines. In MLX LM you can quantize Mixtral (~100GB) in < 4 mins on an 8GB M1 (h/t Angelos Katharopoulos). Example:

Fun story from our internal testing on Claude 3 Opus. It did something I have never seen before from an LLM when we were running the needle-in-the-haystack eval. For background, this tests a model’s recall ability by inserting a target sentence (the "needle") into a corpus of

Memento is a movie about living with a short context window. What he does to compensate is basically RAG.

Welcome home

I encourage everyone to pause for a second and realize how far in the future we are.

# Reproduce GPT-2 (124M) in llm.c in 90 minutes for $20 ✨ The GPT-2 (124M) is the smallest model in the GPT-2 series released by OpenAI in 2019, and is actually quite accessible today, even for the GPU poor. For example, with llm.c you can now reproduce this model on one 8X

Time for a new mind-bending project! #QLOCK — A JavaScript Quine Clock aem1k.com/qlock It displays the current time in a seven-segment style, embedded within its own JavaScript source code. 🕔 🕝 🕢 🕤 🕑 🕜 (321 bytes)

"The model underscores how complex the human brain is: describing just this small sample — one-millionth of the total human brain and about 3 mm long — requires more than a million Gigabytes of data: 1.4 Petabytes." blog.google/technology/res…

I agree with this - and it's a problem. I've given up recommending ChatGPT or LLMs in general to non-technical people. They don't understand that PDFs that consistent of JPGs aren't going to be handled the same as text. They don't understand why after so many messages it forgets

TIL: there are more transistors in the AirPods Pro than in the CPU of a MacBook Pro from 2010 One is a professional laptop, the other earphones running on a battery weighing about 1 gram Moore's Law's one hell of a thing

I'm very proud of the Preparedness evaluations we did on o1-{preview,mini}. One example in particular: While testing cybersecurity challenges, we accidentally left one broken, but the model somehow still got it right.

“LLMs are just next token predictors”, Ilya Sutskever explains how this is deeper than it sounds. Predicting next token requires understanding the underlying reality that led to the creation of previous tokens.

cant believe how much time i wasted learning things before LLMs, spending hours in google search rabbitholes, not knowing future-me-in-only-a-few-years would never need to google again and would have a PhD who can answer anything instantly for $0.001. so grateful for this tech!

You can finally do the Blade Runner Esper Machine, thanks to o3. "Zoom, enhance."

Tanay Jaipuria And it’s gotten worse in the 6 months since we first measured. Google is now: 15 scrape for 1 visitor OpenAI is now: 1,200 scrapes for 1 visitor

ben This is my favorite part … ( 84% chance it blackmails you if it thinks you might turn it off )

* = the answer to the next question A task meant to require web and computational search guided by multi-hop deduction starting from a partial SHA1, answered in 2m 54s by ChatGPT o3 featuring Tom Cruise and Addison Rae

Turing comparing building thinking machines to biological procreation: "we are, in either case, instruments of His will providing mansions for the souls that He creates."

I can't believe how quickly we reached perfect text as the indicator that an image is AI

The Opus 4.6 system card has some extremely wild stuff that remind you about how weird a technology this is. These paragraphs are really worth reading.