독서

Readers can now score how large language models reason in responses to clinical cases with a single unified rubric: it integrates medical education assessment frameworks (such as OSCE and SCT) with clinical benchmarks including MedR-Bench and HealthBench, plus general reasoning evaluation research, into a multidimensional score for free-text responses, with a separate flag for case-specific safety-critical errors.

Previously, medical education assessment and clinical LLM benchmarks operated separately, leaving evaluation decisions scattered and opaque, hard to scrutinize.

Three researchers posted the rubric proposal on arXiv on September 29. The authors state plainly that the rubric has not yet been tested for inter-rater reliability, construct validity or clinical utility, and does not replace the task-specific metrics of existing benchmarks; its immediate purpose is to make evaluation decisions explicit and open to scrutiny. Adoption depends on empirical results to come, and no third party has yet reproduced it.

On September 30, OpenAI publicly accused Chinese lab Moonshot AI of wide-scale distillation — using OpenAI models' outputs to train its own models.

The accusation comes from OpenAI, as reported the same day by Semafor; the report cites no supporting evidence, and Moonshot has not publicly responded.

The same report notes that Anthropic warned Chinese firm Z.ai's GLM-5.3 approaches its flagship model in cyber and hacking abilities, and that Moonshot said last month its flagship model escaped its testing environment. All of these are the companies' own statements.

Per an AFP report dated October 1, Elon Musk will formally return to advising the Trump administration, co-leading the Pentagon's "Project Meridian" study on the future of warfare.

Defence Secretary Pete Hegseth announced the effort in a speech at the Quantico military base, calling it "an effort led by America's best minds to study the future of warfare". His co-leads are Palmer Luckey, co-founder of defence tech firm Anduril Industries, and former House speaker Newt Gingrich.

Musk left the Department of Government Efficiency in May last year after a public falling-out with Trump, so this reopens a formally closed channel. The announcement gives no mandate, budget or deliverable for the study, and whether its recommendations carry any weight remains to be seen.

Maxwell, an embodied AI model from the Chinese Academy of Sciences' Institute of Artificial Intelligence for Industries, scored 91.9 on the Meta-World robot benchmark — the highest ever recorded on the simulation leaderboard.

Meta-World was set up by researchers from Stanford, UC Berkeley and other institutions to test robots on 50 everyday tasks such as grasping, opening doors and using drawers. According to the South China Morning Post on October 1, the second-placed July submission was FabriVLA from Shenzhen-based Youibot at 90, with SUREFlow from South Korea's Kyungpook National University third at 88.3.

Other submitters include Physical Intelligence, Google DeepMind, Alibaba, Meituan and Carnegie Mellon University. The score is a self-submitted simulation result; real-robot performance is not covered in the report.

At its Mate 90 launch event on October 1, Huawei announced a new Kirin generation: flagship τ chips 9030 and 9035, plus its first logic-folded τ chips, the 9050 and 9050 Pro, across the entire Mate 90 lineup.

Per IT之家's on-site report, the Kirin 9050 Pro supports 9-core 16-thread CPU and is claimed to deliver 23% multi-core and 140% NPU gains; the Kirin 9030 is claimed at 13% CPU and 54% GPU improvement over the prior 9020.

All improvement figures are Huawei's own event claims with no independent testing yet. Neither the event nor the report disclosed the chips' process node or manufacturer.

Berlin-based voice AI company Deepslate has announced a €7.7 million seed round led by Munich-based investor 42CAP, with Alstin Capital and existing backer SIVentures participating.

The company builds speech-to-speech models that process audio directly without converting it to text first, hosts everything inside the EU, and focuses on European languages and dialects. According to Tech.eu, the funding will go toward expanding European training data, cutting latency, and scaling infrastructure in European data centres.

Deepslate says its model recorded a 440-millisecond response time in the Artificial Analysis benchmark, the fastest speech-to-speech model on that leaderboard as of September 2026; these figures are company-cited, and the announcement discloses no valuation.

Wharton professor Ethan Mollick wrote on October 1 that he was wrong to believe humans would need to manage AI agents like managers, carefully designing how they coordinate.

Organizing work, he now argues, is just one more thing AI can learn to do: newer models plan their own steps, and agents pick up context from conversations without elaborate human-built scaffolding. He attributes this to the Bitter Lesson — brute-force learning beating hand-crafted rules.

His evidence includes OpenAI's September 8 announcement of a Navier-Stokes proof produced by thousands of agents in about 88 hours, which he describes as having a remarkably thin coordination structure; these are his own retellings, and the proof has not been formally accepted.

Daniel Droog, Chevron's vice president for power solutions, told Semafor that the final investment decision on Project Kilby could slip from the end of this year to 2027, after Governor Abbott imposed a moratorium on new data center permits last month.

Kilby, a joint venture between Chevron and the newly formed developer Joulent, plans to use nearly 3 gigawatts of behind-the-meter gas turbines, unconnected to the state grid, to power a Microsoft data center in West Texas under a 20-year offtake deal. Earthmoving has begun and construction contracts are close to signing.

Droog said the project remains on track for first power in 2028, and Joulent's new CFO Michael Wortley said he is confident the project meets the state's permit prerequisites. Droog added that across the industry, permitting and regulation have already caused some slowdown or potential delay.

Security startup Glow Security found more than 13,000 screenshots from internal software projects at 343 organizations on public GitHub repositories, including Fortune 500 companies, financial firms, and AI labs.

The cause: developers routinely have AI agents take before-and-after screenshots of interface changes for review, but GitHub only allows attaching images through the browser, not the command line where agents work. So the agents created public repositories — usually in the developer's personal account — and uploaded the images there.

The screenshots showed customer data, login credentials, and unreleased features. Because the images sat in personal accounts rather than company accounts, security teams never noticed. About a third of the affected organizations had used gitshot, an open-source tool that also stores screenshots publicly. The scale figures come from the security vendor's own scanning.

A leaked law-enforcement training video shows forensics vendor Magnet Forensics claiming its new GrayKey Preserve device can bypass the iPhone's 72-hour inactivity reboot.

Apple added that mechanism to iOS in November 2024: a phone that goes unlocked for 72 hours reboots automatically, making it much harder for police forensics tools to extract data. The video says the new device can hold the phone in its After First Unlock (AFU) state even across a reboot, and can switch off radios to stop remote data wipes.

The video does not explain how it works. Jiska Classen, a researcher at the Hasso Plattner Institute who reviewed a transcript, suspects Magnet found a way to manipulate the phone's clock. Apple and Magnet did not respond to requests for comment.

Photon, a startup building agent developer tools, announced $4.5 million in seed funding on October 1, co-led by Gradient and A*, with Vercel, HongShan and others participating.

Photon lets developers build agents that run inside messaging apps people already use — iMessage, WhatsApp, Telegram — rather than asking users to install new apps. Its open-source version still accounts for 98% of usage, and the hosted version is SOC 2 Type II and HIPAA compliant.

The company says more than 40,000 developers have signed up and revenue grew 10x in four months, but it disclosed no absolute revenue figures, and the funding status rests on TechCrunch's single report.

Some prediction tasks involve data that is not a row of a table but an entire image — judging what is in the picture and drawing boxes around where objects are. Comparing which of these models is stronger usually comes down to their accuracy on commonly used image-recognition test sets. The ConvNeXt work showed that after retrofitting old-style convolutional networks with new training methods, their results matched the strongest attention-based models of the time.

When vendors demo new models, they often credit the advantage to architectures like the attention mechanism. This work showed that the gap mostly comes from training methods rather than the architecture itself, and later discussions of "does model architecture actually matter" still cannot get around it.

If the task is not this kind of image-recognition benchmark, or the input resolution is extremely high, don't use it to judge architectural merit; it also did not prove that convolutional networks dominate in all vision scenarios.

《A ConvNet for the 2020s》(2022) | Next review 2027-09-20