Fun update: I got tired of disliking every email client I’ve ever used and built my own. It’s called Exo (for exoskeleton). It’s Claude Code for my inbox. It manages my inbox for me, and it’s open source. Link to repo + some notable features in thread!
Biology is context. The same cancer cell behaves differently depending on its neighbors.
Today we’re launching Tumor Context to test if AI understands that.
Strong opinion, held loosely: this has always been the right paradigm for design.
Generalist DNA and Protein language models have a low ceiling for design (they are helpful for analysis). The capability to do abstract reasoning is a massive boost to design in biology and does not emerge from training on sequence data fast enough (if ever). Much better off teaching an LLM about biological sequences than training a pure sequence model, which ends up just recombining within the known sequence space.
Either you have local/active learning data and your model is crushing it with that or if you want to be a generalist on public data, you should train a model that already/concurrently knows how to do abstract reasoning (i.e. code/language).
🌐 alloy.argonrobotics.ai
Today we're open-sourcing Alloy.
We've been using it internally to serve our clients for months. Now we're opening it up to the world.
Alloy is one physics engine for liquids, soft bodies, cloth, cables and robots, built for delicate work like pipetting, stitching and threading a needle, across thousands of worlds on one GPU.
🧵
When I first joined YC we had a running joke that “there's never a dull day” because however many disasters we'd help founders with, we were always discovering new types of disaster. 16 years later, there's still never a dull day.
Three trends in recent YC batches: more hard tech, more hardware, and more consumer.
Matic has done all three, to an extent that most outsiders would never realize. They built a mm-precision perception stack, the best home robot, and sold it directly to tens of thousands of families.
AFAIK, the most deployed consumer robot in the world.
Thanks for sharing with F26, @mehul and @navneetdalal
interesting that the argument hinges on current unreleased frontier models running on edge compute within the next 12 months via software-only improvement
Today we’re launching cross-model KV cache transfer from @OneTriangle, a step towards efficient model communication.
Up to 10.8x faster TTFT at 1M context. Prefill small model, transfer KV cache, decode on large. SDK live for Qwen3.5-4B → Qwen3.8-27B, beta open for more model pairs.
Model latency significantly impacts user satisfaction. Surprisingly, @OpenAI's speed decreases noticeably during the work day, while @AnthropicAI's doesn't at all.
Almost all benchmarks measure latency at test time, but few look at how API response time performs live on production.
We looked at 200,000 LLM requests submitted by users on DesignArena, at every hour of the day, sent to each lab's standard API.
OpenAl model speeds decrease significantly during the work day, fluctuating up to 16%. Even more interestingly, Anthropic's models don't seem to be affected at all.
mind bending personal reflection: 99.999% of what i previously considered "software development" circa 2023 is automatic today and yet in the most AI pilled companies ppl with technical / programming backgrounds remain vital.
across the market, swe job openings are up and other technical-adjacent roles like pm are way up. this is even though, in the intervening years, essentially 100% of what i thought this job was has been automated.
also: i wonder how many lines of code the median software engineer merged to prod on the median day at the median company in 2023. i would bet it's 0. at the 90th percentile for all of those things i'd bet a few hundred LOC.
perhaps these things are related
Proud to announce that AgentWeb is in the W27 batch!
This is a full circle moment, as one of the reasons we started this company was a dm with @agupta back in February, who is now our partner!
We solve the problem of agents that are slow to use the web, by building out endpoints (including ones behind a sign in) that allow agents to go unnoticeable on virtually any site (linkedin, crunchbase, amazon, etc..) and execute tasks on behalf of humans
If you're looking for an alternative to your existing workflows, would love to talk!
Time to ship :)