Monday, August 3, 2026
One of the leading theories of human consciousness says the mind runs on a "global workspace," where information from separate systems becomes available all at once. Anthropic's interpretability team went looking for that structure inside a language model and found something that behaves much like it. David Valdman walked us through the paper at last month's Research Circle, starting from the transformer and building up to the experiments behind it. The recording is on the Rally YouTube channel now.
These sessions happen when someone cares enough about a paper to teach it to a room, so if there's one you've been sitting with, I'd love to hear about it.
Two more nights ahead: Degrees of Freedom starts August 18 at Mission Robotics, and on August 25 we're at Niantic Spatial's space in the Ferry Building, for Physical AI: What's Working (and What Still Breaks). We're confirming demoers now, so more on the lineup soon, and registration is by approval, so register early.
Builders Night is tonight at 6:30 at The Commons, bring a project. See you there.
Lee
Coming up at Rally
Builders Night
2026-08-03 · 6:30 PM - 9:00 PM
Two hours of focused work alongside other builders, then the share-back at the end where people say what shipped and what they need.
Degrees of Freedom: Bring Your Own Robot
2026-08-18 · 6:00 PM - 9:00 PM
A monthly working night at Mission Robotics for solo robotics builders who trained policies at home and need actual labmates. Two hours of focused bench time, debugging help from people who've hit your exact problems, and a space designed for hardware work.
Physical AI: What's Working (and What Still Breaks)
2026-08-25 · 5:00 PM - 8:00 PM
Live demos across the robotics stack, perception, mapping, simulation, policies, hardware, from builders whose systems actually run. Then everyone who demoed sits down for a panel on what still breaks: failure modes, the gap between simulation and reality, and the datasets nobody has.
Around SF this week
Tuesday, Aug 4, 6:00 PM - 8:00 PM
Most teams rent inference and never think about it again. These four founders own their models end to end, and each gets fifteen minutes on why they made that call and what it cost them.
Tuesday, Aug 4, 6:00 PM - 8:00 PM
Five YC teams on the technical decisions behind putting agents into production, across infrastructure, testing, and knowledge work. The draw is hearing what held up at scale, past the demo stage.
Wednesday, Aug 5, 6:00 PM - 8:00 PM
First night of a new series on the layer nobody demos: novel inference methods, infrastructure at scale, and making cheap systems reliable. Talks and live demos from Luminal, Attentio.ai, and smallest.ai.
Wednesday, Aug 5, 6:00 PM - 8:45 PM
The robotics side of the scene, with demo tables and hiring booths alongside short talks from Bedrock Robotics, Serve Robotics, Polymath, and Amazon FAR. Go if physical AI is where you're trying to get.
Wednesday, Aug 5, 7:00 PM - 10:00 PM
Local models are slow, so this one walks through quantization and serving tradeoffs, benchmarks against Claude Code, then has you tune your own endpoint live. Laptop required.
Thursday, Aug 6, 5:30 PM - 8:00 PM
Your agent scores fine on public benchmarks and then falls over on messy inputs, multi-step workflows, and tool calls. Four teams building evals infrastructure on how they measure what the benchmarks miss.
Thursday, Aug 6, 5:30 PM - 8:30 PM
AI made protein design fast, and the bottleneck moved somewhere else. A panel from Biohub, Generate:Biomedicines, and Retro Biosciences on what actually slows the path from sequence to clinic.
Thursday, Aug 6, 5:30 PM - 8:30 PM
You scan a QR code and walk into a configured sandbox with real data and integrations, so you use the agents yourself. A gallery of working systems you can poke at.
Thursday, Aug 6, 6:30 PM - 9:30 PM
The hardware scene's long-running demo night, now on its 136th edition. Prototypes and components on tables, happy-hour format, and you can bring your own if you have something to show.
Saturday, Aug 8, 10:00 AM - 1:00 PM
Your coding agent forgets your repo every session, so this build sprint has you chunk, embed, and index your codebase into persistent memory. Out the door by 1pm, laptop mandatory.