AI digest: Coding models hit new highs while privacy gets serious
Poolside's coding models reach 72.5% on SWE-bench, OpenAI releases open-source privacy tools, and the Microsoft-OpenAI exclusivity deal finally ends.
Big moves in coding AI this week, plus some long-overdue changes to data partnerships.
Poolside’s coding models crush benchmarks
Poolside AI dropped two new coding models that hit 68.2% and 72.5% on SWE-bench Verified. These are open-weight models designed for long-horizon coding tasks, which means they can actually stick with complex problems rather than giving up after a few lines. The jump from previous scores suggests we’re hitting a sweet spot where coding AI becomes genuinely useful for real engineering work.
OpenAI releases privacy filter for free
OpenAI just open-sourced a 1.5B parameter model specifically for detecting and redacting personally identifiable information. It runs in browsers with only 50M active parameters, which is properly clever engineering. This feels like OpenAI trying to get ahead of privacy regulations by giving everyone the tools to clean their data. Smart move, especially with all the enterprise deals they’re chasing.
Microsoft loses OpenAI exclusivity
The Microsoft-OpenAI partnership just got rewritten with no more exclusivity clauses and the AGI provision completely removed. Amazon already has OpenAI models running on AWS, which shows how quickly these partnerships can shift. This probably means more competition between cloud providers, which should drive down inference costs for everyone.
Google steps up Pentagon AI work
Google signed a new contract with the Pentagon after Anthropic refused to let DoD use Claude for surveillance and autonomous weapons. Over 600 Google employees protested, but the company pushed ahead anyway. The contrast between Google and Anthropic’s approaches here is stark, and it’ll be interesting to see if this affects enterprise customers who care about AI ethics.