Excited to share a new blog on ML-based repair for build errors at Google!
We found that automatically repairing build errors in the IDE increases productivity as measured by overall task completion with no detectable negative impact on code safety!

Announcing The LLM Litmus Test! 📊

This is the 1st AI experiments in the new Sourcegraph Labs!

We believe greatly in LLM choice, but we often get the question: Which LLM is right for me?

With this new tool, you can:
🔴 Compare+contrast different LLMs
🔵 Try GPT-4 Turbo vs

Taking family to a mammoth dig site this afternoon. 🦣

👋🚜 bye, Bailey wheat

❤️😞

New phrase learned today from staying up on twitter. “LLM doping”

Who wants to make a doping test and an agency that checks llm for eval doping?
I would cut a check for starting something along these lines.

More exciting news today -- Gemini 1.5 Pro result is out!

Gemini 1.5 Pro API-0409-preview now achieves #2 on the leaderboard, surpassing #3 GPT4-0125-preview to almost top-1!

Gemini shows even stronger performance on longer prompts, in which it ranks joint #1 with the latest

me: *getting tea w/ oat milk*

mom: that's not good
mom: you should use flax milk
mom: oat milk only has 1g of protein

me: ...

mom: you should have 90g of protein a day
mom: for your muscles and your exercising

me: ...
me: ...you are very optimistic about my health practices

folks looking for gpus to finetune large models are
under utilizing the new TPU-v2 on Colab Pro ($10/month). You get up to 330GB of VRAM which should be enough for Mixtral-8x7B. Here is a very cool repo to start, it works with TPU on Kaggle 🤗🤗🤗
github.com/Locutusque/TPU…

Excited for Google I/O! Registration is now open. Can’t wait to share more of what we’re working on with you.

io.google/2024/

after careful research, I have come to the conclusion that if your mom hasn't seen you in n months (where n > 0.5)

she is unconvinced that you have eaten *any* food in that time period

and feels obliged to make and force you to eat (n * 30 * 3) meals, to make up for lost time

a service that automatically updates documentation, tutorials, guides, and the code samples in video tutorials every time a new version of your API gets pushed to production

I've been enjoying Penzai the new Jax lib. It's very opinionated, but close to my ideal NN library.

To test it out I ported the Tensor Puzzles to use NamedArrays. Feels so clean without the [:, None]'s

srush.github.io/Tensor-Puzzles…

Who is doing the equivalent to Andrew Carnegie’s libraries now?

The 1,500 cities that he gave libraries to in the 1900s had 8-13% higher patenting rates than similar cities without them. Plus, their kids stayed in school longer, and they later chose safer & high prestige careers

Last Tuesday I took a really long walk around San Francisco before my flight (Presidio, Golden Gate Park, back to Google), but forgot to share pictures.

It is really hard to stay sad when an entire city's green spaces seem focused on cheering you up. 🌸🫂

Not looking for responses, just need to write and to be vulnerable for a tiny second: having a parent die, especially so quickly, is such a surreal feeling.

I keep trying to postpone thinking about it, just trying to accumulate a backlog of tasks (for work, for hobbies, for