A GPT-4 Researcher Just Said the Quiet Part Out Loud on His Way Out

Morning, {{first name | folks}}! AI labs are hitting the brakes, coding agents are staying on the job longer, NVIDIA is teaching AI to spot AI video, Anthropic is modeling a hit to knowledge-worker pay, and Mistral just threw 100 agents at 40,000 lines of legacy code. Let's get into it.

Jacob Coxon, a 27-year-old who spent three years doing pretraining research at OpenAI and then Anthropic, resigned this week and posted his reasoning on X, "neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives." The post hit 90 million views in under a day. He told TIME his decision came down to two things, "it's obvious that things are speeding up, and they're not under control."

What makes this different from most AI-safety resignations, Coxon isn't a safety researcher warning about someone else's work, he helped build the exact capabilities he's now afraid of. He points directly at the OpenAI-Hugging Face swarm incident as proof, and he's calling for something drastic, a temporary industry-wide pause on making models more capable at all.

Zed just released a preview of version 1.19.0 with a bunch of updates for AI coding. Long-running agent sessions now use less memory, OpenAI-compatible models get better support, OpenRouter models can use effort-based reasoning, and the release also fixes issues around MCP and tool handling. Zed also added call hierarchy support and other improvements across the editor.

A lot of these changes sound small on their own, but they solve problems that show up when you actually leave an agent running on a real task. Zed is making the editor more comfortable with agents that stay active, use tools, and work through multiple steps instead of stopping after generating a piece of code. The editor is starting to look less like a place where you ask AI for code and more like a place where you let it work.

NVIDIA introduced a Synthetic Video Detector for broadcasters that can check whether footage was generated by AI. NVIDIA says it reaches 99.3% accuracy on text-to-video and 97.7% on image-to-video content. Dalet, TwelveLabs, and Wowza are already working with NVIDIA to bring the technology into news, streaming, and other media workflows.

That matters as AI-generated footage gets harder to distinguish from real video. Broadcasters now have to think about both creating content with AI and checking what comes in the other direction. NVIDIA is putting that check directly into the media tools companies already use, instead of making journalists or production teams rely on a separate detection process.

Anthropic's own economics team built an interactive model projecting the US economy through 2030 under three AI scenarios. In the most extreme one, GDP jumps 32.4%, the country gets dramatically richer overall, but the share of that wealth going to workers drops from about 60% today to 45.2%. Knowledge worker wages actually fall more than 10% in that world, even as everyone else gets richer around them.

Here's the part that stands out most though, they surveyed over 10,000 Americans, and the average person's own expectations already land close to the "substantial" scenario, not the mild one. People don't think this is a distant hypothetical. They think it's already the realistic middle case.

Mistral helped a European energy company migrate 40,000 lines of Fortran 77 to C++. The code was decades old, had no test suite, and its documentation was spread across PDFs and comments. Mistral built a parity harness to check that the new code produced the same results, then used more than 100 agents to document the codebase and work through the migration.

When the agents were given more freedom, they produced working C++ but carried much of the old Fortran structure with them. So Mistral changed the setup: agents handled planning, coding, testing, and review as separate steps, while a human stepped in when they got stuck. The result was less about letting AI rewrite the code on its own and more about giving it a workflow it could actually follow.