Another new record for AMD MI300x training performance and SLMs: down to ~150 minutes for hitting state-of-the-art evals with only 140M params! Can't wait to see what this arch looks like at the ~3B param scale.. luckily I won't have to wait long haha ;) Thanks to Hot Aisle

DeepSeek panic opened a firehouse of opportunities. Who is it negative for and what people missed? This thread covers the impact on: 1/ Compute Hardware Requirements (+) 2/ Energy Consumption (=) 3/ AI Applications (+) 3/ Open-Source vs. Closed-Source Models (-)

There's a simple way up in most societal hierarchies: Become a learning machine. If you go to bed every day a little bit wiser than you woke up, it's crazy how much more helpful you'll be to the world. It'll take time, but it is certain, and you'll be rewarded for it.

However, I think the external response has been somewhat overblown, especially in narratives around cost. One implication of having two paradigms (pre-training and reasoning) is that we can optimize for a capability over two axes instead of one, which leads to lower costs.

Biggest inflow since 2021.

Just $AMZN, $MSFT, $GOOG, and $META alone are expected to spend $270B on CapEx this year.

Another banger from Joe. This is what most people struggle with To add insult to injury I see many teams wanting to use an API call or off the shelf tool to do their evals. Not only do generic evals not work, but they are incredibly harmful and distracting

if you’re building a RAG system in 2025 just build a good search engine backend + let the model query it

Retrieval is as alive as ever

new blog post "There Are No New Ideas In AI.... Only New Datasets" in which i summarize LLMs in exactly four breakthroughs and explain why it was really *data* all along that mattered... not algorithms

Friends and family often ask whether LLMs will replace search engines. I don’t think so because LLM-based search wouldn’t work without search engines. General audience LLMs like GPT-4o literally rely on them to answer many queries. In that sense, LLMs are (for now, a very

AlphaEvolve is deeply disturbing for RL diehards like yours truly Maybe midtrain + good search is all you need for AI for scientific innovation And what an alpha move to keep it secret for a year Congrats big G

Context is all you need

I really like the term “context engineering” over prompt engineering. It describes the core skill better: the art of providing all the context for the task to be plausibly solvable by the LLM.

although search was never old, now is the best time to work on search and retrieval. with so much sloppy content and seo-optimized sites, building engines that go beyond the noise to provide reliable, high-quality results is a real challenge. not only just relevance but also

If I can teach you one thing; it's to never ever doubt your instinct. If your instinct tells you something, it's right. Your gut, your instinct is god's gift. If you ignore it, you disrespect him Follow. Your. Instinct