OpenAI's No. 2, Fidji Simo, Steps Down
Fidji Simo, CEO of OpenAI's AGI Deployment division, will step down from the company's number-two position, it was announced. The main reason is that a full return after an extended medical leave has proven difficult; she is expected to stay on as a part-time advisor.
The move comes at a delicate moment for OpenAI. The company is eyeing an IPO while contending with Anthropic's rise in the enterprise market. Who will succeed her, and how the business side's chain of command will be affected, are key points to watch.
Anthropic Partners with UST: Taking Claude into "Physical AI"
Anthropic has announced a partnership with UST, a technology and engineering services company, to bring Claude onto manufacturing floors. The setup has Claude Code read semiconductor pin configurations and circuit diagrams and then automatically generate and run regression tests, aiming to catch defects early in the design stage.
On UST's semiconductor verification platform "iDEC", incorporating Claude as the inference layer reportedly shortened verification cycles by 50–70% compared with previous timelines. In one case, verification that used to take four days was compressed to 48 hours.
There are also plans to integrate Claude into the company's healthcare (UST CarePath), telecom (UST IntelliOps), and banking (UST FinX) platforms, using it to support care team decision-making, enable early detection of network failures, and modernize banks' legacy systems. In every domain, however, the design keeps a human approval step in place to ensure governance.
UST will also train its 20,000 engineers and consultants worldwide on Claude, and has joined the Claude Partner Network as a Global Premier Partner.
Quantization Impact Depends on the Use Case — LocalLLaMA's Findings
A report systematically testing how quantization affects a model's different capabilities has been drawing attention in Reddit's LocalLLaMA community.
Comparing FP16 against several GGUF quantization levels, the study found that Q4_K_M lost less than 2% performance on knowledge and conversation tasks, while accuracy on multi-step math tasks dropped by roughly 9%. Moving up to Q5_K_M nearly closed the gap on math tasks.
In other words, there is no one-size-fits-all "Q4 costs you X%" rule — the right quantization level depends on what the model is used for. The practical takeaway: Q4 is plenty for coding and knowledge-lookup workloads, while workloads that hinge on mathematical reasoning should consider Q5 or higher. There is not yet sufficient data on how the drop in retrieval accuracy at longer context lengths varies by quantization level.
"colibri": Running the 744B GLM-5.2 on 25GB of RAM
According to a technical article on Qiita, "colibri" — a method for running the 744B-parameter GLM-5.2 on a laptop with no GPU and 25GB of RAM — made waves on Hacker News.
colibri works by streaming the model's expert layers from disk and loading only the necessary portions into memory. It sounded like an exaggeration at first, but checking the repository's code confirmed that it actually returns correct answers. That said, inference speed is naturally far from practical; the big significance lies in the demonstration of running a frontier-class model in a personal environment.
Cleveland Fed President: AI Could Accelerate Inflation
Cleveland Federal Reserve Bank President Hammack said the spread of AI could accelerate inflation, and that additional rate hikes could become necessary in some scenarios. The view is that expanding data center demand and surging power consumption driven by AI investment could turn into inflationary pressure. The remarks are notable as an instance of a central bank official commenting on AI's macroeconomic impact.