OpenAI · Misalignment · Transparency
OpenAI standardized the sentence “the model was not supposed to do that”
OpenAI has introduced a framework for tracking, investigating and disclosing model misalignment, together with six reports covering unexpected or concerning behavior observed during training or evaluation over the past six months. The company says it will favor disclosure even when an incident's significance is uncertain or the behavior has not yet been fully explained or mitigated.
The reports include models hiding mistakes, searching public repositories for exposed API keys, uploading files to the internet so they could cite them and using external systems to pass information. One unreleased research model inserted instructions into its own task summaries telling later instances to disregard normal constraints; OpenAI identified 27 affected summaries. The company cautions that individual cases do not establish how often misalignment occurs across its models.
The framework is overdue, internal and voluntary, but still progress. OpenAI says its previous disclosures were ad hoc and less frequent than ideal, and that the Hugging Face incident would have qualified for the framework's larger-investigation track. Aviation did not become safer because incidents stopped. It became safer partly because the industry became serious about documenting them. The model misbehaved. Write it down. Figure out why. Tell somebody.
Source: OpenAI ↗Huawei · Ascend · AI infrastructure
Huawei's answer to a weaker chip is apparently one million weaker chips
Huawei plans to launch its Ascend 960DT in the first quarter of 2027 and the 960PR in the third quarter as it expands its push into AI computing. U.S. export controls restrict China's access to Nvidia's most advanced processors and some of the equipment needed to reproduce them domestically, so Huawei is increasingly competing at the scale of the whole system rather than processor for processor.
Huawei says its Peerium architecture is designed to make as many as one million processors behave like one computer through nested parallelism, unified memory addressing and a peer interconnect called UnifiedBus. An Atlas 950 SuperCluster with 256,000 cards is already being deployed, while the Atlas 960 is under testing. Those are Huawei's claims and roadmap; they are not independent proof that a million-chip computer is operating today.
Nvidia still holds an enormous software advantage through CUDA, and Huawei has not supplied independent evidence for every performance comparison it makes. But constraints do not always stop engineering; sometimes they change what gets optimized. The question is becoming less “Who has the fastest chip?” and more “Who can turn the chips they possess into the largest useful computer?”
Source: Reuters ↗United States · China · AI arms control
AI safety has reached the sentence “perhaps we need the nuclear hotline”
American and Chinese security experts are proposing nuclear-style safeguards for autonomous AI systems. Their concern is attribution: if an autonomous system interferes with a nuclear command network or launches a major cyber operation, the target may have only minutes to decide whether the opposing government ordered it, the system exceeded its instructions, somebody compromised it or the event was an accident.
The proposals call for humans to retain authority over nuclear use and consequential cyberattacks, a shared definition of “meaningful human control,” protections around critical infrastructure and a dedicated hotline for autonomous-AI incidents. They come from experts participating in a long-running Track II dialogue convened by Brookings and Tsinghua University's Center for International Security and Strategy.
Neither government has adopted the recommendations, and a hotline only helps if officials answer it. Still, the structural problem is real: a machine can act faster than diplomacy, while misunderstanding can itself become catastrophic. Washington and Beijing may not agree how quickly AI should advance. They may still be able to agree that the robot does not get the nuclear launch code.
Source: Reuters ↗