Yesterday, All-Gather 2026 brought the AI research community together at the San Francisco Exploratorium for our first-ever research conference. If you weren’t there, here’s a little glimpse of what you missed.
this was an experiment in how far you can get working mostly agent driven. our first attempt was from scratch with no reference and the agent struggled. so we started again and let the agents read from sglang, which allowed for some good bootstrapping. mahmoud (https://t.co/eogYd4aiBn) took it from there and things really started moving. TP4 went from ~5ms to 3.5ms to 3.1ms per step in ~a day. stopped hill climbing around 2.6ms.
we built an inference engine for deepseek v4.1-flash from scratch. the core engine is just rust + cuda. no pytorch, no sglang, no dsls, etc.
~1740 TPS decode on 4xB300, 2060TPS on 8xB300. 1M token prefill in ~12s.
it's up on github
The best logistics model isn't Astra or Fable.
It's a small model, trained on lessons from real shipments. It beat every frontier model Cartage tested on its freight benchmark, while running 20x cheaper than the frontier.
Cartage-1 powers Wilson, unlocking significantly more complex logistics workflows than consumer models.
Genspark and Baseten Applied Research post-trained an open-weight model that beats Opus 5.5 at deep research.
Starting from Qwen3.5-397B, our post-trained model scores 0.759 on @genspark_ai's internal evals, ahead of Opus 5.5 (0.738) and well above the base model (0.612). It’s now the default model in production behind Genspark’s Deep Research agent.
Excited this post is finally out 🎉
Benchmarks can be gamed. Production can't: real users have to actually like the model.
This one beats Opus 5.5 at deep research and is now the default in Genspark's agent. Built by Baseten's training team in collab w @genspark_ai.
More than a third of psilocybin, ketamine, and LSD use logged on WHOOP happens on Fridays and Saturdays. So we looked at the morning after.
LSD and ketamine use were associated with bedtimes over an hour later and Recovery scores about 9 points lower. Psilocybin showed smaller changes: 33 minutes later and 2.9 points lower.
The takeaway: the night before shows up in your Recovery.
We're going to see a LOT MORE of what X just did.
At Town, we've rebuilt systems in 4 days with one engineer. Before AI coding agents, that would've taken us 2 months.
If you're rebuilding something where you already know the requirements and how everything is supposed to behave, agents can do an incredible amount of the work because you can verify.
It's a very different problem from building a new product where you're still figuring out what users want.
https://t.co/QKokjNhKU6
Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.
Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.
https://t.co/VR02tw73fJ