AI Post-Training Risk: Why Safety Checks Fall Short
A researcher argues deployment safeguards work well enough, but the greatest AI post-training risk sits in training itself, where monitoring tools backfire.
A researcher argues deployment safeguards work well enough, but the greatest AI post-training risk sits in training itself, where monitoring tools backfire.
An Nvidia preprint finds agents limited to 48,000 tokens of context matched full-history agents across 13 benchmarks. The cost claims come with caveats.
After a November selloff that hit AI stocks, investors at TomorrowX debated whether chips, power efficiency or AI use cases offer the next opportunity.
An AI-assisted proof of the Riemann hypothesis has a researcher asking whether humans ever handled software complexity, and how formal specs and diversity could help.
A fintech engineering lead merged 115 pull requests in a week, and 170 in the best seven-day stretch, after clearing a string of bottlenecks for AI agents.
Pi Dash coordinated 26 AI agents to rewrite its own Django backend in Rust in about ten days, at a model cost near $350.
A chart from OpenAI’s chief scientist places GPT-6.2 Astra exactly on the line for solving a Millennium Prize problem, hinting at a shared base with the Navier-Stokes proof model.
Saving AI agent memories is mostly solved, but keeping them accurate isn’t. Benchmarks show stale memory can do worse than none.