Microsoft Tells Its Own Engineers to Stop Using AI
Microsoft's reported move to ban its own engineers from using AI tools has been drawing attention on Hacker News. It's a seemingly contradictory decision: a company that puts Copilot and Azure AI front and center is restricting AI use in its internal development.
The move appears to be driven by quality risks in AI-generated code and security concerns. Promoting AI adoption to customers while keeping AI away from its own critical development processes symbolizes the trust gap that has emerged between companies that sell AI and companies that use it.
This development highlights the practical difficulties of AI adoption. How to reconcile the productivity gains that generative AI delivers with quality and security assurance remains a challenge the entire industry continues to grapple with.
Heretic: FT Reporting Reveals the Spread of "Decensored" Models
The Financial Times has published a detailed investigation into the impact of the open source tool Heretic, once again underscoring how difficult it is to keep open models safe.
In the FT's own hands-on verification, it succeeded in stripping the guardrails from Meta's Llama 3.3 model using Heretic in under 10 minutes — no special hardware required. According to Heretic's creator Philipp Emanuel Weidmann, more than 3,500 "decensored" models have been created since its release, with a cumulative total of 13 million downloads.
The freedom to modify models at will — a key advantage of open-weight models — is working against safety. Model providers continue to strengthen their guardrails, but as long as removing them with a tool remains technically easy, these efforts fall short of a fundamental solution. Balancing openness with safety is expected to remain a structural dilemma in the AI community.
Nature Warns Against Uncritical AI Adoption in Science
An editorial published in Nature has expressed strong concern over the "uncritical adoption" of AI in science.
It points out that while AI tools are being rapidly embedded into research processes, guardrails for validating their soundness and ensuring the reproducibility of results remain insufficient. Particular concerns include cases where AI-generated analysis is incorporated into papers without critical evaluation, and the possibility that the black-box nature of AI tools undermines scientific transparency.
The editorial does not deny AI's usefulness, but stresses the need for an essential debate — not just about the speed of adoption, but about how to maintain scientific rigor. This warning from science's most prominent journal offers an important perspective for considering the societal acceptance of AI.
NuExtract3: A 4B VLM Under the Apache 2.0 License
Numind has released NuExtract3, a vision-language model based on Qwen3.5-4B, under the Apache 2.0 license. It specializes in extracting information from visually structured documents such as PDFs, screenshots, forms, tables, receipts, and invoices.
Designed to take images and text as input and output conversions to Markdown, it is the successor to the earlier NuMarkdown-8B-Thinking. Despite its lightweight 4B size, it can parse complex document layouts, and its self-hostable nature adds to its practical appeal.
antirez Announces DwarfStar, a Distributed LLM Inference Project
Salvatore Sanfilippo (antirez), the creator of Redis, has announced DwarfStar, a distributed LLM inference project.
It is an experiment in distributing LLM inference across multiple machines, with details published on antirez's blog. The developer behind Redis, a world-standard piece of infrastructure software, is now taking on LLM inference infrastructure — drawing significant attention in the open source community.
Distributed inference is a key technology for running large models that don't fit on a single GPU, and there is much anticipation about how antirez's experience will translate.
US Government's $2 Billion Quantum Computing Investment May Be Illegal
Members of the US Congress have begun arguing that the $2 billion investment in quantum computing companies, announced by the US government last week, is illegal.
Congress approved the funds for the purpose of supporting public research, and their use for equity investments in startups was never authorized, they argue. The arrangement, which puts $100 million into each company, represents make-or-break scale funding for quantum computing firms whose products are thought to be years away — but a dispute over its legal basis has now begun.
At the same time, the investment also gave rise to the first quantum foundry company, though some are questioning whether demand even exists in the first place. The debate now looks set to call into question the very approach of public investment in quantum computing.