Amodei's 2023 Senate testimony calling open source AI dangerous resurfaced in 2026. Reddit's technical community fact-checked every claim and found the same pattern that began with GPT-2 in 2019: safety language that happens to benefit Anthropic's bottom line.
POSTS
Updates from the workshop
Field notes and progress reports from building Apprentice.
Sakana AI's AI Scientist scored higher than 55% of human submissions at ICLR 2025. Each paper cost $15. Nature published a paper about the system. The team's recommended fix is an admission the existing system cant tell the difference.
Sakana AI's Fugu Ultra benchmarks match Anthropic's best model. But the scores belong to GPT-5 and Claude Sonnet 4. Here is what Fugu actually is, and what it isnt.
37 TB in 21 days, from a logging system you never configured. Extrapolated: 640 TB per year against your SSD. The forensics trace back to a TRACE level default that ships with every Codex installation. The Hacker News thread called it slopware. Hard to argue.
New Website is Online for Everyone!