What Is Going On
OpenAI tensions resurface. OpenAI's internal tensions spilled into public view again. The pattern is clear: safety teams raise concerns, commercial pressures win. This isn't unique to OpenAI—it's the industry default.
Anthropic races to keep up. Anthropic positioned itself as the 'safety' alternative, but is now racing to keep up with GPT and Gemini releases. The market doesn't reward caution.
Open source ships without review. Open-source models ship with essentially no safety review. The gap between frontier labs' safety theater and the actual ecosystem is growing.
AI Corner
Red-teaming professionalizes. Red-teaming is becoming professionalized. Several consultancies now specialize in finding model vulnerabilities. Prices are high; demand is higher.
Interpretability lags. Research on model interpretability continues to lag capabilities. We understand these systems less as they become more powerful.
News You Can Use
- ●
UK AI Safety Institute reports. The UK AI Safety Institute released its first evaluation results. Mostly unsurprising findings, but establishing the methodology matters.
- ●
Voluntary commitments signed. Several major AI companies signed voluntary safety commitments. Enforcement mechanisms remain unclear.
- ●
AI liability insurance emerges. Insurance products for AI liability are emerging. Pricing suggests underwriters are worried.