AI News: August 3, 2026

1. AI Keeps Cracking Unsolved Math Problems, Dividing Mathematicians

Research. A run of AI systems has begun resolving previously open mathematics problems, sparking debate over whether the tools accelerate discovery or erode the discipline’s culture and verification norms. Coverage surveys both the enthusiasm for faster progress and the worry that machine-produced results change how proofs are trusted and credited. It matters as automated reasoning starts to touch genuine research frontiers rather than curated benchmarks. Source

2. Study: AI Coding Agents Modernize Research Software but Can’t Vouch for the Science

Research. A new analysis found coding agents can effectively modernize long-neglected research software and deliver large speedups, yet they produce output that is “eloquent, convincing, and confidently wrong” when judging scientific correctness. Reviewers still had to verify results extensively, since the agents optimize for plausible code rather than valid science. The practical lesson: agents are strong at refactoring and porting, weak at domain validation. Source

3. Claude Opus 5 Turns Text Prompts Into Playable 3D Games

Anthropic. Hands-on testing shows Anthropic’s Claude Opus 5 can generate complete, browser-playable 3D games from a single prompt, emitting geometry, textures, physics, and even music directly as code without external assets. It marks a clear jump from earlier prompt-to-game demos that produced rough colored blocks. The result signals how far single-shot code generation has come for interactive, multimodal artifacts. Source

4. OpenAI Launches Presence to Put AI Agents Into Production

OpenAI. OpenAI introduced Presence, an enterprise offering for deploying AI agents in production across customer service and internal workflows, going beyond what configurable GPTs can handle. It pairs the platform with “forward deployed engineers” who work directly with customers to pick workflows, integrate existing systems, set guardrails, and run testing through launch. The move targets the persistent gap between demoing an agent and operating one reliably in real business settings. Source

Also covered in today’s OpenAI roundup.

5. Meta Adds a Second ‘Memory’ Agent to Keep Long-Running Agents on Track

Meta. Meta AI researchers proposed a dual-agent design that runs a dedicated memory agent alongside the action agent, tracking execution history and selectively injecting reminders so long tasks stop drifting or repeating failed steps. The approach lifted Terminal-Bench 2.0 from 38% to 46% and tau2-Bench from 55% to 62%, beating fixed-recall baselines and existing memory layers such as Mem0. The team argues that deciding when to surface memory matters as much as storing it, and it published the project’s code on GitHub. Source

Also covered in today’s Meta roundup.