joe tey

we're working on a new photo search engine called Angles. it's focused on doing one thing well: finding photos + videos by visual similarity. here is raw, unedited footage of me going through my 80,000 photo camera roll *in realtime* with text-to-image search, image-to-image search, "find similar" inside of a photo, and "live search" with the camera. all done with local models and completely private. auto-growing albums coming soon :-) Angles has been exceptionally useful for my friends who do creative work. when words fail to describe what you are looking for, you can simply tap on any asset and instantly see everything else like it across your entire library. we're in early beta - if you want access please send a DM to @patinasystems with your testflight email! i have spots available for another 50 people

The hardest part of protein engineering isn't just finding good mutations – it’s deciphering which ones combine synergistically. Today in @ScienceMagazine, we present MULTI-evolve, a framework for rapid multi-mutant protein engineering, validated across three diverse proteins.

@mgdurrant @AnthropicAI Congrats Matt!

Thrilled to share that I have joined @AnthropicAI as a life science researcher! I am confident that Claude will do amazing things to accelerate biology. Big things ahead!

Excited to share our work digging into how Evo 2 represents species relatedness or phylogeny. Genetics provides a good quantitative measure of relatedness, so we could use it to probe the model and see if its internal geometry reflects it.

@sokrypton This is cool. Do you extract per-layer logits by applying the unembedding matrix to each layer's hidden state (logit lens style approach)?

Bonus: One can compute the catergorical jacobian layer by layer, to see what pairwise interactions each layer is learning. (10/10)

Excited to share the work with @ZhidianZ @HWaymentSteele @garykbrixi @wwwhbpt @matteodp @DorotheeKern biorxiv.org/content/10.110… (1/7)

What makes LLMs like Grok-4 unique? We use sparse autoencoders (SAEs) to tackle queries like these and apply them to four data analysis tasks: data diffing, correlations, targeted clustering, and retrieval. By analyzing model outputs, SAEs find novel insights on model behavior!

Come join the bridge editing team @arcinstitute and build the future of genome design!

🦎 un.ms/research

@JohnKoelliker @ForerunnerVC @WarbyParker @Chime @glossier @ouraring @faire_wholesale @Zola John! Congrats — this is huge!

Announcing Leland's $12M Series A led by @ForerunnerVC! Forerunner has an incredible track record of backing iconic consumer brands and marketplaces like @WarbyParker, @Chime, @glossier, @ouraring, @faire_wholesale, @Zola, and others. We couldn't be more excited to welcome @eurie_kim to the board and partner with her and Forerunner to take Leland to the next level. Also participating in the round are @gsvventures, inVest Ventures, and our early investors like @FjLabs, @contrary, @PetersonVenture, and @GSBackers. We’re also proud that dozens of Leland coaches and customers participated, in addition to leading marketplace founders, operators, and angels from companies like Instacart, Handshake, LinkedIn, Pluralsight, Mercado Libre, and others. With this fresh capital and support, we're well positioned to accelerate our vision of unlocking human potential by making the world's expertise more accessible. In the past year alone, our coaches have helped hundreds of thousands of people in over 70 countries break into top schools or companies, develop new skills, prepare for interviews or tests, and pursue goals across dozens of unique categories. While a fundraise is never a destination, it certainly represents an opportunity to reflect on the progress we’ve made and thank the community who has helped us get this far. So, THANK YOU! We know the biggest opportunity is ahead of us, and we’re excited to be building a world-class team to go after it. If you want to join us, we're hiring. if you want to partner with us, shoot me a DM. And if you are interested in working with a coach or coaching the next generation, we'd love to have you in our community.

Studying intelligence seems like the right next step. Excited for this team!

@SlavaChalnev @MatthewWSiu @ArthurConmy Cool work! Curious if you tried directly clamping a set of SAE features instead (as opposed to a steering vector) / if you discarded this as an unpromising technique?

Excited to share a new paper! We improve LLM steering vectors by targeting SAE features 🎯

@josephsemrai Congrats Joe!

Meet Context Autopilot It learns like you, thinks like you, and uses tools like you. With SoTA context understanding, it's capable of most information work today. Watch it beat a team of industry experts:

🔥 Paper Drop 🔥 What can we understand by peering inside vision-language models (VLMs) like LLaVA? We show that image representations inside VLMs can be directly interpreted and edited in the language space, and we apply our findings to mitigate hallucinations!

maybe a canvas for research?

branching interface for writing a tighter creative feedback loop guided by a language model to weave through the possibility space of text

🚀Excited to announce: Open-source AlphaFold3 implementation! 🚀 I am thrilled to announce one of the models we have been building for the last 8-weeks at Ligo - an open-source implementation of DeepMind’s frontier model, AlphaFold3! Here’s what we have learned, a thread (1/11):

@thesephist @lbtucker123 I'm curious -- do you mean curating the dataset for the automated interp. pipeline, or the dataset to actually train the SAE? Either way, even the mechanisms of chunking text (by sentence, paragraph, token), can influence extracted features. Thoughts?

(1) I don't think we need to find all, or even a majority, of features in a model to make useful tools / steering mechanisms on top of them (2) There are many features that we're not really likely to care about for a particular use case, and by curating the dataset more intentionally we can focus the search for features on specific attributes we care about for e.g. a creative writing or research or programming use case

By end of 2024, steering foundation models in latent/activation space will outperform steering in token space ("prompt eng") in several large production deployments. I felt skeptical about this in summer '23, felt vaguely positive in Jan, and now think it's more likely than not, and I'm more optimistic than ever about the direction of my work since 2022 around a joint exploration of interfaces and latent visualization/steering of foundation models. Anthro's published work today is a milestone in a steady march toward this future that started in mid 2023. Anthropic and DM's leadership in this area, combined with lots of community efforts in work like steering vectors, better sparse autoencoders, various image editing UI prototypes all push us toward the future here, but once technical foundations are there, interface will be much more obviously a bottleneck to utility, alignment, and capability. I'm very excited about the interface possibilities this will open up, particularly for multimodal models and creative use cases. For a moment I thought it was possible that dialogue may eat everything. I don't think so anymore. We'll see new universes of possibilities in both. (And if frontier labs don't have serious interface research bets today, this would probably be a good time to reconsider it :-)

Meet Iris — a macOS app that makes it easy to multitask with AI. Over the last year, I’ve been exploring how we can evolve the desktop OS. This is the first small step :) You can download Iris now at iris.fun

stand back, I'm a professional -- >>>content.split("...")[2].strip().split('"""')[0].strip().split("\n")