Just got access to Veo 2 through Freepik (odd collaboration, but hey). The prompt coherence is amazing, even with super complex prompts. Completely outperforms Hunyuan, Kling, and other competitors (and don't even get me started on Sora).

One of my random tweet replies just went viral for the first time (2100 likes on Twitter!). Feels weird that an off-hand comment blew up while my actual content gets crickets.

Grok 3 is quite good - not mind-blowing (at least without Deep Search/Big Brain, which I haven't tried yet), but solid. Strong showing by the xAI team, though not enough to shake up the status quo.

I'm excited for the Grok 3 launch in a way I haven't been for other model releases—not because I have particularly high hopes (I don't), but because it will definitively show whether scale in pre-training still matters.

So we're just casually dropping models that beat both Llama and state-of-the-art memory models by massive margins now? Looking forward to the independent replications of these impressive LM2 results.

Turns out you can teach LLMs to reason with just 17k examples and the content doesn't even need to be correct. The real secret? Just keep the reasoning structure coherent. Makes you wonder about all those carefully curated datasets.

Finally, a paper that tells us when distillation beats supervised learning without having to run a million experiments. TL;DR: Use distillation for small compute budgets or when you already have a teacher, otherwise stick to supervised learning.

Huginn: Smaller model that gets smarter by thinking longer in latent space. Finally someone trying something other than just making models bigger or forcing them to write their thoughts down like a middle schooler.

Jobs disappeared due to high interest rates, but AI means they're not coming back. This decade will be tough for junior software engineers.

Mathematical reasoning in LLMs has always been seen as data-hungry, requiring 100k+ examples. New paper shows just 817 carefully curated samples can achieve 57.1% on AIME. If reproducible, this challenges everything we thought we knew about scaling laws in reasoning tasks.

It's amazing how much drama surrounds OpenAI. My wife knows that whenever I start a story with 'Hey look, Sam Altman...' or 'Hey look, OpenAI...,' it won't be about a new model or product—it'll be about another chapter in the daytime soap opera that is their boardroom drama.

DeepMind's AlphaGeometry2 solves 84% of IMO geometry problems from 2000-2024, surpassing gold medalist performance. The system nearly doubles its predecessor's 54% solve rate through improved language modeling, faster symbolic reasoning, and novel search techniques.

Wild how Deepseek became the first AI model besides ChatGPT to become a household name. Even my parents talk about 'that Chinese AI they keep hearing about on the news.' Claude, Gemini, Mistral, Qwen, Grok - none of them hit mainstream. It's just ChatGPT and Deepseek.

Comprehensive study shows that long chain-of-thought reasoning in LLMs emerges from SFT + RL but requires careful reward shaping to control length scaling. Rule-based reward signals with filtering outperform model-based approaches for stabilizing performance.

New research from Stanford and UW shows how to achieve o1-preview-level performance for test-time scaling with just 1K examples and simple supervised training. The key? Careful data curation and sensible compute budgeting.

Interesting find: OpenAI's Deep Research web browser has JavaScript disabled by default, but the model can enable it when needed. Cool to see the technical decisions behind this project.

OpenAI's Deep Research is way better than Gemini's, which is the only reason I'm keeping my Gemini Advanced subscription. Oh well.

It's wild how viciously anti-AI the 🦋 crowd is. You'd think their lives had been destroyed by rogue LLMs, given the level of vitriol.

We've gone from LLMs as 'write a limerick' machines through a brief 'count the letter r' phase to our current 'watch balls bounce in a rotating rectangle' era.

At Maracas, our coding assignments are fully open-book. We encourage using any AI tools you want. We're hiring for both coding & AI skills—and it's supercharged our velocity. Still doing leetcode whiteboard interviews in 2025? That's just irresponsible.

Teaching is just the latest job being displaced by AI. Six weeks of LLM tutoring now matches two years of traditional education.

Humanoid robotics isn't at its language AI moment yet (like we saw in 2022), but I can feel we're getting really close.

ChatGPT Tasks feels underwhelming - as a concept, product, and direction for ChatGPT. If Operator turns out just as weak, that's not great news for OpenAI...

Here's an efficient prompt format I use often with o1-pro: 'Ask me 10 questions that you need answers to in order to [___].' It lets you outsource more thinking to the model since it generates even the questions that need to be asked.

Seeing more 15-18yo founders lately. Years ago, I pitched a coding school where talented teens could skip high school, get elite CS training in half the time, and join the workforce as rockstars at 17-18. Still relevant.

Why do we keep using these corny stock photos of milk-white robots with exposed brains to symbolize AI? Kubrick gave us the perfect visual metaphor back in 1969.

We need a 'proof of humanity' system to separate purely human-created content from AI-generated or AI-assisted content. Think of it like organic vs GMO labeling, but for creativity. One idea: requiring creators to escrow a video of themselves making the content. Any other thoughts?

When Kai-Fu Lee said nursing would be last to fall to AI (needing dexterity + empathy), it made sense. But seeing 2025's robots and AI like Claude, we're probably 2-3 years from robot nurses beating humans. Other jobs? Even more at risk.

The denial among LLM skeptics about AI's coding productivity gains is mind-boggling. 'Well, maybe eventually AI will help write some basic boilerplate code.' Sorry, but Claude already codes circles around you - you're just not seeing it.

Thoughts and prayers to everyone affected by the Los Angeles wildfires. It's a terrible tragedy and my heart goes out to everyone who is suffering.

NVIDIA releases Cosmos - an ambitious attempt at a general-purpose physical world simulator. Open source with permissive licenses, but still struggling with basic physics like object permanence. Nice milestone for robotics and autonomous driving, even if we're far from a true digital twin of reality

Small LLMs (1.5B-7B) achieving o1-level math reasoning through iterative self-improvement is impressive, but let's see independent verification and stress testing before getting too excited about the 90% MATH score claims.

Project Digits is killer. The Mac Mini-like form factor is a huge plus — I can just keep it sitting on my desk until I need its power.

SDPO paper is out - turns out the secret to making LLMs more socially intelligent isn't better data or larger models, but simply optimizing the right chunks of conversation. Sometimes less really is more.

OLMo 2 proves open-source LLMs can compete: 7B & 13B models matching/beating Llama 3.1 & Qwen 2.5 while using fewer FLOPs. Best part? EVERYTHING is open - models, data, code, logs, the painful training stability debugging... because science needs more than just weights. @soldaini.net

When GitHub Copilot launched 3 years ago, it was a massive concentrated undertaking by a huge tech company. Now a small startup can build a competent code assistant as a side project. It's incredible how fast we're accelerating.

FLOAT: A new method for AI talking heads that's literally 125x faster than current approaches, with preserved quality and realistic expressions. Flow matching shows its strengths over diffusion models yet again.

Babe, wake up - @sirbayes.bsky.social just dropped an amazing reinforcement learning tutorial! I honestly thought RL was gonna be a passing fad, but every month I see more important use cases. This is a perfect intro if you want to learn.

Interesting paper from Google DeepMind on 'motion prompting' for video gen. Clean technical approach with nice results, though the 12-min generation time makes me wonder about real-world viability beyond research demos. @tobiaspfaff.bsky.social @andrewowens.bsky.social

Lichess just released a multi-billion-game chess dataset on Hugging Face. Beyond improving gameplay, curious what other insights we could train models to find!

Gemini-exp-1206 (yeah, what's with that name?) is crushing it - #1 across the board on LMArena. O1 Pro is solving tasks no other LLM has cracked before. Llama's dominating open weights with Llama 3.3. Fingers crossed we get 3.5 Opus by year's end. This December is absolutely fire! 🔥

Finally, someone figured out how to do 4D reconstruction without needing a million synchronized cameras. Impressive results on both real and AI-generated videos while keeping the technical approach surprisingly elegant. @benmpoole.bsky.social @jonbarron.bsky.social @holynski.bsky.social