Sunday, July 19, 2026
Anthropic's interpretability team has identified a mechanism inside a language model that resembles the "global workspace," the structure a leading theory of consciousness credits with making human thought conscious. It's a substantial claim, and what makes it worth two hours of discussion is the rigor the team brought to testing it.
That's the subject of tomorrow's AI Research Circle. We'll open with the theory, then walk through the evidence behind it: the prompts, the measurements, and the interventions used to distinguish a genuine finding from a pattern one might expect to see. You'll leave able to read the paper closely and judge for yourself whether it holds up. No research background required.
See you tomorrow at 6:45 at The Commons.
- Lee
Coming up at Rally
AI Research Circle: A Global Workspace Inside LLMs
2026-07-20 · 6:45 PM - 8:45 PM
David Valdman takes Anthropic's "global workspace" paper from the theory through the experiments that back it up. No research background needed.
Claude Code Demos: Building the State of the Art
2026-07-21 · 6:45 PM - 8:45 PM
Builders using Claude Code, Cursor, and custom agents put their real setups on the screen, then spend the night comparing what works and getting unstuck together.
Builders Night
2026-07-27 · 6:30 PM - 9:00 PM
Next week's coworking night. Two hours of heads-down work alongside other builders, then a share-back where you find out what shipped and what stuck. Bring a project.
Around SF this week
Wednesday, Jul 22, 5:30 PM - 8:30 PM
Different data structures answer different questions well, and forcing every modality through one costs you in latency, storage, or lost meaning. This session walks through vector indexes, Iceberg tables, knowledge graphs, and more, showing what each excels at and how to combine them so agents actually get faster and cheaper.
Thursday, Jul 23, 3:30 PM - 5:00 PM
Helmy Eltoukhy took Guardant Health from nothing to a public genomics company, after pioneering semiconductor DNA sequencing at Stanford and Illumina. South Park Commons runs a sharp founder-origin series, and this one digs into the early calls that decided whether Guardant existed at all.
Thursday, Jul 23, 5:30 PM - 8:00 PM
Teams are shipping Claude Tag and ChatGPT Work into their workflows, and this demo night surfaces what's actually working versus what's still rough in multiplayer AI collaboration. Lightning demos plus open discussion on emergent behaviors and where team-based AI goes next.
Thursday, Jul 23, 6:00 PM - 8:30 PM
Three builders diving into how local-first knowledge graphs and identity layers transform AI agents into something beyond chatbots. Arto Bendiken, Rayyan Zahid, and Rob Kunkle share concrete approaches to personal intelligence at Frontier Tower.
Friday, Jul 24, 5:00 PM - 7:00 PM
A small dinner on what stays human when agents become the default tool you build with, with music and a conversation curated to sit in the tension. A good reset if your week has otherwise been all infra and demos.
Tuesday, Jul 28, 6:00 PM - 9:00 PM
LangChain and Cognition on "LLM wikis," the emerging pattern for giving agents memory that holds up over long runs. Most agents fail on weak context, so if you're building anything that has to remember across sessions, the talks here are worth the Tuesday.
Wednesday, Jul 29, 5:30 PM - 8:00 PM
Groundcover is showing agents that run inside your cloud, reading live production telemetry and acting across Slack, Linear, and GitHub without shipping data anywhere. The demo triggers a real anomaly on stage and watches Agent Mode isolate the cause and open a patch end-to-end.
Wednesday, Jul 29, 5:30 PM - 8:30 PM
Embodied AI research track launching with two curated talks on world models, robot learning, and sim-to-real, followed by live robot demos on Mission Robotics' floor. Rigorously selected presentations prioritize discussion over slides, with papers shared in advance to ensure high-level technical critique.