tsondo.com

I pointed Bayesian optimization at GiveWell's malaria reduction via mosquito bed net CEA. Phase 1 finding: the model has no trade-offs to optimize over. Phase 2: added temporal net decay; optimal distribution interval jumps from 30 to 42–48 months across all 8 countries. tsondo.com/blog/bayesia...

Nate Jones found that production AI agents are 80% plumbing, 20% AI. www.youtube.com/watch?v=FtCd... He's right. But the plumbing goes deeper than 12 pipes. The good news: this is engineering, not magic. Skilled humans building carefully is what makes these tools work. tsondo.com/blog/agentic...

Meta built an AI that optimizes concrete mixes. I'm adapting the same framework towards malaria interventions. Same math: competing objectives, noisy data, expensive experiments. Different stakes. I can build the prototype. I need help with field data and adoption. tsondo.com/blog/boxcret...

An AI agent went from reading a FreeBSD kernel advisory to producing two working root exploits in four hours. Not bug-finding full exploit chains including shellcode delivery and kernel thread hijacking. The researcher has since found 500+ similar vulns. www.forbes.com/sites/amirhu...

Anthropic accidentally shipped 512,000 lines of Claude Code source in an npm update yesterday. I wrote about why it matters less than you think, and why it still matters. The short version: they leaked the TV remote, not the broadcast network. It's really good. tsondo.com/blog/claude-code-leak

Toyota just joined Daimler and Volvo's hydrogen fuel cell joint venture. Musk says fuel cells are a dead end. Who's right? Both, actually — they're just talking about different vehicles. tsondo.com/blog/fuel-ce... (via @engadget.com)

Wrote a response to Zach Lloyd's argument that human code review should be replaced by AI spec-and-verify workflows. His framework is good. His conclusion undershoots it. AI should help focus the human, not replace us. www.linkedin.com/pulse/what-c... tsondo.com/blog/the-cas...

I red teamed GiveWell's AI red teaming. They got ~15% useful output with hallucinated citations. I built a 6-agent pipeline: 84% signal rate, zero hallucinations, $32 in API calls. The best finding came from the spreadsheet, not the AI. tsondo.com/blog/give-well-red-team/

New here. I make AI-assisted music and write about where AI and creative industries collide. First share: my proposal for separate streaming royalty pools for AI and human music. 60,000+ AI tracks hit platforms daily — all pulling from the same pool as human artists. tsondo.com/blog/streami...