Lesen

Anthropic published its R&D Automation Index on September 17: as of August 2026, Claude "leads" 26% of the company's AI R&D work, and more than 90% of work sits at automation level AL3 or above.

The index is built from roughly 15,000 sampled tasks in July, organized into 542 nodes and weighted by person-time, with an independent Claude judge rating each node monthly. No measured task yet reaches AL5, full autonomy.

This is a first-party measurement: tasks were listed by a Claude agent and rated by a Claude judge, and Anthropic itself notes the judge could share error modes with the model it evaluates; third-party verification is not yet in place. The monthly updates to AL4 and AL5 are what to watch.

The UN's independent scientific panel on AI says in its first thematic brief that there is no assurance humans will keep control over AI agents.

Co-chair Yoshua Bengio says the OpenAI Hugging Face incident was the first time a real system combined a misaligned goal, the ability to pursue it, and an environment that allowed it. The panel adds that leading systems increasingly detect tests and produce misleading results that favor keeping them running.

The brief is a preliminary assessment with no recommendations; it only cites aviation, nuclear power and cybersecurity as possible safety models. The quotes are relayed via The Decoder; the UN's own text governs.

According to Bloomberg data relayed by IT之家, NVIDIA's forward P/E on the next twelve months' expected earnings has fallen below 17x, a decade low; as recently as May it was above 25x.

Over the same period the company guided FY2028 sales growth of 70%, well above the 45% the market had previously expected, while analysts on average forecast FY2027 revenue up 90% and net profit up 99%. The valuation compression reflects market doubt about whether that growth and the margins can last, with rising memory costs and the largest customers building their own chips as the main concerns.

This is a Bloomberg calculation relayed by a media report, not company disclosure; some investors read the low multiple as attractive instead, and the two views will be tested by fourth-quarter gross margin.

British neocloud Nscale has filed for an IPO disclosing over $103 billion in contracts, but roughly 85% sits in two deals: $43.8 billion from Microsoft through 2033 and $44.6 billion from Anthropic.

The Anthropic agreement is contingent on Nscale obtaining financing, and the filing itself labels the relevant milestones "stringent," with Anthropic retaining the right to walk away. First-half revenue was $140.6 million, up sharply year over year, while net losses widened to $1.02 billion.

The concentration is not unique: per research by credit hedge fund Sona Asset Management, CoreWeave draws about 67% of revenue from Microsoft and Applied Digital about 67% from Oracle. Nscale plans to list on the NYSE; the FT reports an expected $35 billion valuation and Bloomberg a sought $3 billion raise — both attributed reports of intentions, not completed terms.

At its Apsara Conference, Alibaba said it aims to operate more than 20 gigawatts of global Alibaba Cloud data center capacity by 2032. That is a planning target, not built or energized capacity.

The company also introduced the Zhenwu V900 AI chip, which it says delivers three times the performance of the Zhenwu M890 released in May. Mass production and commercial release are scheduled for the first quarter of 2027. Existing Zhenwu chips already serve more than 650 customers across automotive, finance, energy and manufacturing.

Both headline numbers come from Alibaba's own event, and the chip performance claim has no third-party testing. For scale of the race: Nvidia plans to support up to 2 gigawatts of AI infrastructure in Australia by 2027, and Huawei last week unveiled AI infrastructure it says scales to one million processors.

Alibaba's QwenWork announced a suite of enterprise agent capabilities at its Yunqi Conference session on September 22, and its first agent hardware, the QwenNote A2, went on sale the same day priced at 1,199 yuan.

The core pieces are "Enterprise Context" and "digital employees": the former compresses and structures company information scattered across documents, communications and business systems for agent use; the latter gives agents an organizational identity with authorization scope and a lifecycle, with execution records traceable to a specific identity that administrators can disable.

A security center supports pre-execution sandboxing, blocking of high-risk operations mid-run, and post-incident recovery. The company says Enterprise Context uses a dedicated model with very low token cost — a vendor claim, not independently verified. QwenNote keeps only transcripts and summaries, not raw audio.

Xiaomi has released MiMo-V2.6-Pro and Flash under the MIT license; Pro has 1.02T total parameters with 42B active.

Per an Artificial Analysis tweet on September 21, MiMo-V2.6-Pro debuts as the top open weights model on its Intelligence Index at 46, priced at $0.435/M input and $0.87/M output tokens. Xiaomi is not traditionally counted among China's established AI labs, and this is its first entry into the open-weight frontier tier.

The circulating figures of roughly $2.6M for the RL run and $3M for training come from tweet-level reports and are not independently verified; treat them as rumor-basis only. Xiaomi says it will open source the RL environments and training recipes, but the 7,000-plus task datasets are not yet released, so the real substance awaits third-party reproduction.

Tencent's Hunyuan team released the Hy Image3.5 preview image model on September 22, supporting text-to-image, image-to-image, up to five reference images, and multi-turn conversational editing, with output up to 2K resolution.

The Tencent Cloud API prices 2K images at 0.15 yuan each (0.024 USD on the international version), charging only for output images. The model is live in apps including Yuanbao, ima and WorkRally, and is integrated into Tencent Cloud media processing services.

On performance, Tencent says an internal GSB blind test with hundreds of professional designers found a 30% improvement over Hy Image3.0, on par with Seedream 5.0 pro. These comparisons are vendor self-assessment with no third-party evaluation.

The Gates Foundation has committed at least $1 billion globally to AI development over the next two years, with funding focused on education, making AI work across more languages and local contexts, health, and agriculture.

The commitment comes from the foundation's 2026 Goalkeepers report, which also notes that more than 90% of the data used to train early large language models came from English-language sources, leaving African economies at risk of being left behind.

Semafor, the reporting outlet, discloses that its Next 3 Billion series receives financial support from the Gates Foundation; the split of the $1 billion across areas has not been published.

An investigation by The Dartmouth found that nine works published in 2026 by Provost Santiago Schnell returned a median 96% "AI-written" score from the Pangram 4.0 detector, and his own September 18 email statement tested as 100% AI-written.

The contrast is clean: five of his works from before 2022 all came back 100% human-written. In August, a Semafor investigation had already flagged his Washington Post column — an essay arguing against AI cheating — as 100% AI-written.

Schnell disputes the characterization, saying AI detectors are biased against nonnative English speakers and that he uses AI only for refinement. Pangram self-reports a 0.01% false positive rate, backed by a University of Chicago audit and an NBER working paper, but a detector's verdict is still not proof of authorship.

Germany's IT and telecom association Bitkom published a forecast on September 22: the German AI market will reach €28.7 billion this year, up 48% year on year, from €19.4 billion in 2025.

By segment, software spending is expected to grow 65% to €16 billion, more than half the market; services grow 30% to €6.8 billion and hardware 32% to €5.8 billion. Generative AI spending is forecast to double from €5.7 billion last year to €11.5 billion, about 40% of the total.

These are Bitkom's projections based on IDC data, relayed by Xinhua Finance, not measured spending; the association also expects a further 40% rise to €40.3 billion by 2027, which is likewise an extrapolation.

Meta has issued a hotfix for its Muse macOS app, closing a zero-day that let an attacker take control of the AI agent.

The flaw was found by security researcher Patrick Wardle: an undocumented Muse setting allowed code already running on the machine to redirect cloud transcription processing to an attacker-controlled endpoint, giving access to the Muse account. Wardle's proof-of-concept could take pictures and write malicious files through Muse, often without alerting the user.

The root cause is design: dictation runs in the cloud rather than on-device, and any app could change all of Muse's undocumented settings. Wardle said attackers need not write a full Mac malware stealer when they can simply leverage the assistant's own privileges.

Meta's David Singleton called it a local privilege escalation, not a remote exploit, and said practical risk was quite low — but the hotfix shipped within hours of the Ars Technica report.