Evan Miller @EvMill

🚀 New on the Klaviyo Data Science Podcast: @EvMill joins us to discuss his paper, Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations. AI metrics are everywhere—but how much uncertainty is behind them? Understanding variability matters. Listen now: bit.ly/3CV0CmX #AI #DataScience

We’re starting a Fellows program to help engineers and researchers transition into doing frontier AI safety research full-time. Beginning in March 2025, we'll provide funding, compute, and research mentorship to 10–15 Fellows with strong coding and technical backgrounds.

This paper on the statistics of evals is great (and seems to be flying under the radar): arxiv.org/abs/2411.00640… The author basically shows all the relevant statistical tools needed for evals, e.g. how to do compute the right error bars, how to compare model performance, and how to do power analysis. Back when @jeremy_scheurer and I wrote the "We need a Science of Evals" post (apolloresearch.ai/blog/we-need-a…) this paper is exactly the kind of thing we had in mind and more.

Awesome new research by my friend and colleague @EvMill — adding error bars to evals! Always great to see the Central Limit Theorem!

Anthropic @AnthropicAI

2 years ago

New Anthropic research: Adding Error Bars to Evals. AI model evaluations don’t usually include statistics or uncertainty. We think they should. Read the blog post here: anthropic.com/research/stati…

50 292 2K 757K 1K

I cannot agree with this more. Please use basic research methods on AI benchmarking!

keyboard_arrow_left Previous keyboard_arrow_right Next

Anthropic @AnthropicAI

2 years ago

New Anthropic research: Adding Error Bars to Evals. AI model evaluations don’t usually include statistics or uncertainty. We think they should. Read the blog post here: anthropic.com/research/stati…

50 292 2K 757K 1K

New Anthropic research: Adding Error Bars to Evals. AI model evaluations don’t usually include statistics or uncertainty. We think they should. Read the blog post here: anthropic.com/research/stati…

Machines of Loving Grace: my essay on how AI could transform the world for the better darioamodei.com/machines-of-lo…

@egino Interesting q - no idea, sorry!

New sequential A/B test from @Zalando based on the Lévy inequality – check it out! arxiv.org/abs/2406.16523…

I think I've finally cracked quantiles… A/B testing medians, instead of means, usually requires an expensive bootstrap. But we can use a likelihood-ratio test (Wilks' theorem) instead. This reduces the quantile problem to a few simple formulas. Read on! arxiv.org/abs/2401.10233

Ten months ago, we launched the Vesuvius Challenge to solve the ancient problem of the Herculaneum Papyri, a library of scrolls that were flash-fried by the eruption of Mount Vesuvius in 79 AD. Today we are overjoyed to announce that our crazy project has succeeded. After 2000 years, we can finally read the scrolls: This image was produced by @Youssef_M_Nader, @LukeFarritor, and @JuliSchillij, who have now won the Vesuvius Challenge Grand Prize of $700,000. Congratulations!! These fifteen columns come from the very end of the first scroll we have been able to read and contain new text from the ancient world that has never been seen before. The author – probably Epicurean philosopher Philodemus – writes here about music, food, and how to enjoy life's pleasures. In the closing section, he throws shade at unnamed ideological adversaries – perhaps the stoics? – who "have nothing to say about pleasure, either in general or in particular." This year, the Vesuvius Challenge continues. The text that we revealed so far represents just 5% of one scroll. In 2024, our goal is to from reading a few passages of text to entire scrolls, and we're announcing a new $100,000 grand prize for the first team that is able to read at least 90% of all four scrolls that we have scanned. The scrolls stored in Naples that remain to be read represent more than 16 megabytes of ancient text. But the villa where the scrolls were found was only partially excavated, and scholars tell us that there may be thousands more scrolls underground. Our hope is that the success of the Vesuvius Challenge catalyzes the excavation of the villa, that the main library is discovered, and that whatever we find there rewrites history and inspires all of us. It's been a great joy to work on this strange and amazing project. Thanks to Brent Seales for laying the foundation for this work over so many years, thanks to the friends and Twitter users whose donations powered our effort, and thanks to the many contestants whose contributions have made the Vesuvius Challenge successful! Read more in our announcement: scrollprize.org/grandprize

@ggerganov @EvMill The blog about Softmax+1 plays a very important role when we were trying to identify the root cause of the sink @Guangxuan_Xiao can comment more!

Have a few thoughts about this approach But most importantly, I'm happy to see @EvMill's idea on softmax1 recognized - to my very basic and intuitive understanding of LLMs, it made enough sense to warrant further analysis arxiv.org/abs/2309.17453

👀

Markus Nagel @mnagel87

3 years ago

@Tracing47202686 @yell1337 @TiRune Unlike with clipped softmax, to achieve an exact zero in the output using softmax1 for a (partial) no-update, the input requires to be -infinity. However, after @EvMill blog post we experimented with softmax1 and found it in practice competitive with our proposed approaches.

0 1 12 4K 1

Results of my latest nerdsnipe from @TetraspaceWest! The plot below shows the predicted shape of the water flow, with a model taking into account gravity and surface tension. It looks just like the real thing! Conclusion: yep, it's surface tension details below 😁

Following @EvMill great blog post on encountered issues on the GPT-like models training that appear to be related to the SoftMax function, I wrote this small piece mostly to understand what was going on. wandb.me/tinyllama

Kurt Vonnegut's 1969 address to the American Physical Society @APSphysics --on the innocence of the "old-fashioned scientist" and its loss after World War II. For physicists, artists, and other humans. I have transcribed it in its entirety as a google doc: docs.google.com/document/d/1Mn…

Softmax1 update… We now have support for ⚡️Flash Attention ⚡️ This lets us test much larger models than before! To get the code, just pip install flash-attention-softmax-n Or clone / star the GitHub repo here: github.com/softmax1/Flash… All credit / kudos to Chris Murphy.

@shxf0072 @johnowhitaker Never too late for a twitter poll...

Softmax1, Week 2. Second set of empirical results are in, and they are… 🌸 promising 🌸 Weight kurtosis is roughly the same – but activation kurtosis improved 30X (!!) and maximum activation magnitude reduced 15X (!). Read more from @johnowhitaker: datasciencecastnet.home.blog/2023/08/04/exp…