Zain

If you cannot explain something in simple terms, you don't understand it.

#llm-variance" target="_blank" rel="nofollow noopener">danluu.com/ai-coding/#llm…

awesome writeup on ai coding makes a great point: all benchmark results should be presented as distributions instead of a nice and clean single number > distribution of performance across many runs > distribution of cost per task > distribution of wall clock times

pretty mindboggling finding in this new science paper even with 60 years of modern medical advancements 90% drugs still fail human trials, same number as in 1963

U Chicago law school is adding a in-person oral exam to combat LLM slop submissions I think LLMs are a tool that can 10x learning power for students if incentives are aligned and they're using them for the right reasons

@bleysg AGI has already been achieved internally!

@zainhas Forget paperclips, we’re building emdash maximizers.

LLM's be like:

@sparkycollier @JensenHuang He didn't wake up with a looser mentality. He is not a car!

I was wrong. @JensenHuang was able to turn it into a headline

@CompleteSkeptic Very insightful read, thanks for sharing! Can't wait to see the models from typesafe!

@zainhas I wrote an article on IMO the most plausible mechanism: x.com/CompleteSkepti…

FACT: OAI/Ant "reasoning"/"thinking" tokens you get back are encrypted/summarized. QUESTION: How does one distill effectively from encrypted reasoning traces?

men lie, women lie, numbers never lie. the numbers:

Interesting paper on the impact of overthinking in reasoning models

this graph is interesting because effectively it says that Kimi K3 will be the best cyber model that people can use to perform/protect against cyber attacks b/c the blue line is gated and not generally available

oh wow was not expecting Claude Opus 5 to 3x performance of the next best frontier model on ARC AGI 3

if you take all the popular coding benchmarks and classify every task in them you can get a pretty good idea of the hills agentic models are currently climbing👇 tldr: > lots of implementing features + bug fixes + implement algo + perf optim > not enough security, devops, ML

This is pretty foretelling, businesses will need: > Ways to amplify thier content to be picked up by LLMs + thier search api's > LLMs recommending options genuinely vs. ads that show up on the side of ChaGPT will be what you need to optimize for > Embed their content in high signal sources that get scraped for pre/post training so you can bias the model like reddit ex. Fascinating how quickly human behavior can change when a technology is 10x better than incumbents.

does opus 5 have an overthinking problem? on FrontierCode main + extended set - Opus 5 @ med reasoning effort > high, xhigh and max pay to lose?