Aman Dhesi

Every week a new model comes out that “beats” the previous SOTA on benchmarks. But benchmarks don’t matter - models need to be evaluated on YOUR data. Introducing Superpipe, a framework to build, evaluate, and optimize LLM pipelines on your data. Built w/ @benscharfstein

(Re)Introducing Highlight: The Marketplace For Believers. NFTs aren't just digital ticker symbols—they're vibrant expressions of human culture, creativity, and community. Here's what's new on Highlight ↓

@MichaelDean_0 @osventuresllc Congrats Michael! So happy for you

I'm thrilled to announce my @osventuresllc Fellowship, where I’ll turn my philosophy of craft into code. This year ahead is set up to be a fusion of all my past threads: architecture, technology, and writing. I’ll be mapping patterns, reviewing classic essays, publishing a textbook, and developing an AI-powered editor to guide writers towards mastery. Special thanks to @jposhaughnessy and the team for this incredible opportunity. Follow along here and on my website for weekly chapters, essay breakdowns, and updates.

I get a twisted sort of pleasure from seeing TSLA shorts get burned. Go ahead, short the most important company of our generation 🤷‍♂️

2 days ago Evolutionary Scale launched ESM3, a protein language model that can generate novel proteins and a lot more ​ The numbers: • 98 billion parameters • 2.78 billion natural protein sequences in the training set • 771 billion unique tokens of training data ​ What's the hype about? ​ [1] It speaks fluent protein ​ ESM3 doesn't just understand protein sequences - it grasps structure and function too. This multimodal (sequence, structure, and function) approach is a first. ​ [2] Architecture custom-built for protein data It's a masked language model specifically designed to handle the complexity of protein data: ​ • Multi-track input/output: Processes protein sequence, structure, and function as separate token tracks. ​ • Geometric attention: The first transformer block includes a special geometric attention layer to handle 3D atomic coordinates. ​ • Masked prediction: During training, random masks are applied across all tracks, and the model learns to predict the masked tokens. ​ [3] 500 million years of evolution in one model ​ ESM3 generated a novel green fluorescent protein (GFP) that's so different from known proteins, that it's comparable to what we'd expect to see after 500 million years of natural evolution. ​ [4] Data quality is still king ​ The team didn't just rely on natural proteins. They augmented their dataset with synthetic examples to give ESM3 an edge. [5] Chain of thought isn't just for language anymore ​ They used chain-of-thought prompting to guide ESM3 in protein design. NLP techniques are proving their worth even in hard sciences. The team is also releasing ESM3-open, a 1.4B parameter model, for academic research. ESM3 is the best example of the ideas behind language models being applied beyond language and code. ​ Kudos to @alexrives and team! 🙌

I hope you've found this thread helpful. Follow me @amansplaining for more threads like these. Like/Repost the quote below if you can:

Shoutout to the @MachinaLabs_ team for doing something extraordinary. Sources I used to write this thread: youtube.com/watch?v=VcYWzP… youtu.be/dCXu8Ju_fdY?si… alexjfischer.com/what-is-robofo… iopscience.iop.org/article/10.108…

Imagine a world where factories aren't locked into producing a single part for decades Where a production line can shift from car doors to airplane wings overnight This isn't sci-fi. It's roboforming, and it's happening right now in LA. read more 🧵