Great paper from Amazon AGI on generating RL environments for agents. (bookmark it) Also, pay attention to this important new AI engineering skill of creating RL environments. Seeing a huge shift towards this. The authors propose AutoGym which writes the task, the executable environment and the verifier together, starting from a small domain seed or past model trajectories. Paper: arxiv.org/abs/2609.22592 Chat with Paper: academy.dair.ai/papers/autog…

in need of a very unique type of computer use trajectory data, can’t be served by traditional data vendors like scale or mercor hmu if you’re in the space

What an insane time. We are so lucky to be working in tech right now.

do you think that the a16z college and the joe lonsdale college are football rivals

with everything i’m seeing with computer use and jev, wonder if Instinct (the imessage agent) was this fast from early Jev access?

very interesting in better understand the training data used for Jev @typesafeai, have a feeling that this is innovative on that side as well

who do we need to message to get @typesafeai Jev access? looks very promising

💎💠 announcing the official @skarlo_ x @a16z @speedrun partnership 💎💠 we're pretty excited about this one... after creating the look & feel for a few of the @a16z portcos, an email landed in my inbox from @Chen at speedrun we hopped on the phone and within minutes, it was pretty clear we were talking the same language - he inherently understood what is at the core of @skarlo_'s values - intentional design is no longer optional, it's crucial to survival - even at the early stage of his speedrun protcos our first call was a few months ago - since then, we've firmed up an formal partnership and have build brands for some of the most exciting startups in san francisco these incredible teams are made up of some of the best humans around, and they've quickly become some of our favourite clients - in part i'm sure to the mentorship that the incredible lester and the @a16z team provides as we approach the SR007 demo day, we'll be showcasing what the team; @odobasix , @emsparx, @rcharl03 have been building keep an eye out for what's coming, we're only just getting started ✨💫🥀

how it works animation for @primitivelabsai designed in @framer

Frontier models are becoming extraordinarily capable, but capability does not imply human behavioral fidelity. Primitive Labs builds the infrastructure to measure, model, and train how AI systems behave relative to humans.

everyone finding out about typesafe for the first time today is great, this team has been crushing for a couple years now

We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next. Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon. This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials. Read our blog posts below.

codex banked reset would go so hard rn

adept was so far ahead of its time

very cool release from a great team, excited to get started with the api!

We solved the 70-year-old Pierce–Birkhoff conjecture, an open problem in algebraic geometry, by finding an explicit 30-dimensional counterexample. The counterexample was found using publicly available GPT and Claude models in our multiagent, multi-model setup. The proof has been checked by human experts and verified in Lean. Joint work with Zehua Lai and Lek-Heng Lim.

Behavior will be a first-class primitive of product design. Use Instinct to validate behavior at each step of the SDLC. Don't wait for an after the fact dashboard.

pre-product with a podcast strategy is insane

what are all these “model training” startups coming out of accelerators going to do when they realize the gpu situation is so much worse than they thought

when founders say they’re “training models” but don’t know what NCCL is