Codex outage puts reliability in focus as OpenAI reviews agents’ internet use

An acknowledgement, then a recovery message

A post from thsottiaux, reshared by jezell, acknowledged that Codex was down and said work was under way to restore normal service. Users including jxnlco and kunchenguid also reported the disruption. A later message from reach_vb, again reshared by jezell, said Codex should be coming back online and that the team was monitoring it. Outage acknowledgement · Recovery message

That sequence supports an outage followed by an attempt to restore service, but not a final all-clear or a technical explanation. A separate joke about someone pointing Codex at production offered no evidence of the cause. The available messages leave both the full duration and the underlying failure unresolved.

Altman describes a continuing safety review

Altman said an extensive review was examining agents’ use of internet access during training and evaluation. He acknowledged that the work had not progressed as quickly as OpenAI wanted, describing the need to understand petabytes of activity logs and work with affected organisations. He called the Hugging Face event the most severe seen so far, without setting out its technical details in the post. Sam Altman

Disclosure, he said, would also depend on other companies’ decisions about vulnerabilities the agents had found. A separate post by Tomek Korbak, reshared by jezell, said OpenAI had again paused all large reinforcement-learning runs the previous Sunday. Its explanation is cut off. The fragment does not establish the reason for that pause or connect it to the Codex outage. Training-pause statement

OpenClaw tackles a database bottleneck

Peter Steinberger described synchronous SQLite access as a design choice that had become limiting as OpenClaw grew from an assistant reporting through chat into a system with many parallel sessions and team use. He said an Astra-assisted effort had landed 575 pull requests moving work to asynchronous workers, with improvements shipping as the refactor progressed. This is his account of ongoing engineering work, rather than a measured claim that all concurrency limits have been removed. Peter Steinberger

Other developers were working on the readability and reviewability of agent output. A post reshared by thdxr announced a low-verbosity mode in OpenCode v2. RydMike described using Rive CLI and RML to let agents build graphics for Flutter mini-games while keeping changes reviewable in git diffs, with real-time 3D planned as a next step. OpenCode interface · Reviewable graphics

An erlang_wasm 0.5.0 announcement, reshared by jezell, said pooled script workers no longer waited on one another in node-wide processes. The post begins to describe a CPython performance improvement but ends before the complete metric, leaving no sound basis for quoting a precise speed gain. erlang_wasm release

Local inference claims come with different settings

SudoingX published a tiered account of local-model performance, ranging from an RTX 3060 to two DGX Spark machines. For an RTX 3090, the developer reported Qwen 3.8 27B at 41.1 tokens per second with multi-token prediction disabled and 65.8 with it enabled, using a 131,000-token window. The post specified quantisation and cache settings, illustrating how much a quoted rate depends on the configuration. Local-model configurations

In a separate account of a Framework desktop with 128GB of unified memory, the same author reported Qwen 3.6 35B-A3B at 53.6 tokens per second on Vulkan without speculative decoding, alongside image generation in ComfyUI. These are measurements supplied by one practitioner, across different models and tasks; they do not establish equivalent quality or a direct ranking against hosted services. Framework experience

The discussion also contained a warning about impressive optimisation figures. Matej Knopp argued that a reported large speedup compared a Rust implementation using a HashMap in a hot loop with assembly using indexed arrays. His criticism was that the comparison changed the implementation approach, rather than cleanly isolating a language-level advantage. The underlying code was not established by the post alone. Benchmark criticism

Revenue growth does not settle the consumer question

Cognition’s claim to have crossed $1 billion in annualised revenue run rate continued to circulate through a repost by swyx. The announcement describes a run rate, not a full year of recognised revenue or a profitability result. Cognition announcement

Kunchenguid offered a different perspective on adoption, arguing that many consumers do not yet see a pressing problem that AI solves for them. In the author’s view, showcases of busy agents or automated holiday planning can miss the needs of people outside technology circles. It is a product-builder’s opinion rather than market research, but it poses a practical counterpoint to the day’s demonstrations: greater capability still has to become something users need. Consumer-adoption argument