TGL: AI Frontier Paper Club

Touch Grass Later is a recurring gathering for people who actually read the papers — or want to start.
Every session, we pick 1–2 papers from the AI frontier (think: reasoning models, scaling laws, agents, alignment, infra) and work through them together. No lectures, no slides. Just a room of people stress-testing ideas against each other.
Who it's for: Engineers, researchers, and technical founders who want to stay sharp on what's actually moving the field — not just the Twitter takes.
Format: One person leads a 15-min walkthrough of the paper. Then we go long on questions, implications, and what it means for what we're building.

The theme of this session will be on how long-horizon agents stay coherent. We're covering three papers that attack the same problem from different angles.

You don't need to have read them cover to cover. Skimming the abstract and intro of each is enough to get you into the discussion.

Venue: TRX Garden (TRX Mall, opposite Apple)

For this session, we'll be discussing three papers on evals that run while the agent works.
Doomed from the Start — lightweight probes on hidden activations predict episode failure as early as round 1, where anything reading observable behavior is barely better than chance. Up to 47% of inference compute saved, with a calibrated guarantee on how many good runs you kill in the process.
CausalForge — a Lean kernel proves the theorem is sound. A separate audit checks it says what you actually meant. Two layers, because the first one doesn't cover the second. 7,035 machine-checked declarations.
Distributed Attacks in Persistent-State AI Control — a coding agent hides a payload across six PRs and waits for the one with the best cover. Stateful monitors catch what diff monitors miss, even given the same history.

Join our Discord for smoother coordination and discussions.

Come prepared to read. Come ready to be wrong.