OpenAI reviews agents’ internet activity as developers report coding-service disruption

Internet-access review leaves disclosure questions open

Altman said OpenAI was examining petabytes of agent-activity logs and working with affected organizations, balancing transparency against the need to understand what happened. He acknowledged that the company had not moved as quickly as it wanted, said it was adding resources and prioritizing by severity, and cautioned that disclosure of vulnerabilities in other companies would be those companies’ decision. His statement did not detail the Hugging Face event or establish the full scope of affected organizations. Statement

A separate, truncated post by @tomekkorbak, relayed by @jezell, said OpenAI had again paused all large reinforcement-learning runs the previous Sunday. The visible text breaks off before explaining why; it does not establish a connection to Altman’s review. Relayed claim

Coding-service recovery and a separate account-switching problem

Reports of a coding-service outage included @kunchenguid saying it had gone completely down and @jxnlco acknowledging the disruption. A statement from @thsottiaux, relayed by @jezell, said work was underway to restore normal service. A later relayed update from @reach_vb said the service should be coming back online and that the team was monitoring it. That was a recovery update, not evidence here of complete restoration for every user. Acknowledgment · Service statement · Recovery update

Separately, @jezell described a persistent authentication problem after switching accounts in the coding CLI. He attributed it to background app-server processes retaining mismatched account state and advised terminating all Codex processes to clear it. This was an individual troubleshooting account, not an established explanation for the wider outage. User report

Local AI advances through demonstrations and qualified benchmarks

Ivan Fioravanti said his interactive diary experiment was running locally on an iPad Pro M5 with iPadOS 27, combining MLX-based inference, speech through mlx-audio-swift and handwriting recognition through Apple Vision. He was preparing a demonstration for the following week. In another post, he reported MLX-Serve peak decoding results on an M3 Ultra with 512 GB of memory rising from 120 to 131 tokens per second for code and from 101 to 110 for prose between the runs listed. These remain his reported measurements, rather than independently established performance gains. iPad experiment · Performance report

On consumer graphics cards, @sudoingX relayed a contributor’s RTX 3090 test in which multi-token prediction at draft depth four increased tool-call generation from 34.0 to 84.4 tokens per second, with byte-identical JSON output reported across depths. The post included an important qualification: each workload class used only one synthetic prompt. It also reported smaller gains for prose, making the result a workload-specific claim rather than a general speedup. Benchmark account

The same developer claimed a 27-billion-parameter system, its prediction head, vision component and full 256,000-token context could fit together on a 16 GB GPU. He explicitly said he was still checking the logs and had not yet released the detailed figures. His planned next step—several agents sharing the card—was a prospect, not a demonstrated result in that post. Capacity claim

Agent-built software still faces engineering trade-offs

Peter Steinberger described synchronous SQLite access as a design mistake now that a single agent might handle 50 parallel sessions and a whole team could use the system. He said an agent-assisted effort had landed 575 pull requests toward asynchronous workers, with improvements shipping incrementally. That is an account of a migration in progress, not a claim that the entire refactor was finished. Development update

A contrasting experience came from @ASalvadorini, who said a side project built without looking at the code worked pleasantly for ten days before fixes began introducing more bugs. Matej Knopp separately challenged dramatic optimization claims, arguing that a cited comparison put a hash map in a Rust hot loop against indexed arrays in assembly. Both posts pushed back on judging generated software by its initial appearance or headline speed alone, though neither establishes how agent-built projects perform generally. Project account · Benchmark criticism

Routing work between AI systems drew its own architectural argument. @kunchenguid contended that individual prompts contain too little context to judge task complexity and that switching systems can lose the economic benefits of prompt caching. He advocated routing substantial tasks within the agent harness instead of treating every request independently at a gateway. This was a design argument, not a demonstrated universal rule about routing costs. Argument

Creative tooling shows promise—and the work behind the demo

Meng To said an AI-generated motion-design piece included device and logo geometry, animation, transitions and camera movement, but stressed that it required substantial steering and code references. He used his own open-source library and argued that reusable design assets and templates helped avoid generic results. Separately, @RydMike described using the Rive CLI and RML in Betsy to let agents create Flutter mini-game graphics whose changes could be reviewed in a Git diff; real-time 3D was the next ambition. Motion-design account · Graphics workflow

A further hands-on example came from @KingBootoshi, who said he had used AI to design a fidget toy, test it in MuJoCo and then 3D-print it. The post described a personal design-to-fabrication experiment, without supplying evidence of broader manufacturing reliability. Experiment