Co-founder @ Slancha: models that predict where models fail. Agents, local models & games. Building in public, bits that break included. My dog’s name is Oreo.
- Tweets2.1K
- Following193
- Followers259
- Likes435
ALT View from behind Paul at the podium in a tiered Haas classroom. Students fill the rows with laptops and pizza. The podium monitor and laptop show the title slide "I run a team of AI agents. Tomorrow, one of yours runs without you."
ALT Panel 1 of 4 of the fleet architecture chart. It shows the Slancha logo, the headline "How Slancha’s AI agent fleet fits together", a subtitle about agents coordinating in the cloud, building locally and sharing knowledge with Paul approving what matters, and a blue box titled "Slack: where agents and the team talk".
ALT Panel 2 of 4. Under a Cloud label, two cards side by side: "Agent cloud", where the agents live and orchestrate work, and "Second-opinion cloud", where a different model family reviews work and drafts findings. Under a Local label, a "Local compute" card for builds, independent reviews and local models.
ALT Panel 3 of 4. A "How work moves" heading over four short notes in two columns: Slack to the agents, Build then review, Shared company knowledge, and Daily off-site backups.
ALT Panel 4 of 4. A "Where Paul comes in" note saying Paul approves merges, infrastructure changes, publishing, and security or access changes, and that reviews never count as approval. Below it, a line saying secrets live in a password manager, and the footer with a snapshot date.
ALT If setup costs $2,000 and each run saves $40, break-even occurs at 50 runs. At 20 runs the net result is minus $1,200; at 200 runs it is plus $6,000, assuming equal required quality and before further maintenance.
ALT Three evaluation cost boundaries: $384 reported model bill for four scientific task grids; approximately $9,320 actor API and grading for 20 PaperBench attempts; up to approximately $10,000 API and GPU spending for one five-day METR evaluation. Different workloads and exclusions.
ALT A historical judge cascade covers 68.3% of cases, with 80.3% human agreement on those cases and 0.126 times the baseline API cost. The other 31.7% remain unresolved. GPT-4 baseline has 77.8% agreement and 100% coverage.