Reading

Microsoft is turning Copilot from an AI add-on inside Office into the entry point that carries Office itself.

Last week, at a closed-door event for large enterprise customers, Satya Nadella positioned Copilot as the "OS for work." Coding, agent capabilities and the full power of Office will go directly into Copilot, where users can create and edit documents in an interface that resembles the web version of Office. The Verge reported the event's contents on October 1.

The product moves have already happened: the desktop Scout agent announced at Build in June has been put into maintenance mode, with the cloud version rebranded to Autopilot. The Verge, citing sources, reports the rebrand was driven in part by trademark concerns around Yahoo Scout, and that the Copilot reorganization involved internal tensions; Microsoft has not confirmed either point.

Copilot chief Jacob Andreou admitted that agents have been "almost impossible to bring into the enterprise" because of the autonomy they need; Microsoft says the new architecture keeps organizations in control. On the consumer side, roughly 90 million people pay for Microsoft 365 out of their own pocket, which is why Microsoft is not giving up on consumers.

Alkaid Computing (九章云极), a Beijing-based AI computing infrastructure provider, filed its listing application with the Hong Kong Stock Exchange on September 29. The offering size and amount to be raised have not been disclosed.

According to Zhidx reporting and the prospectus, revenue reached 1.098 billion yuan in 2025, up 177.7% year on year, and 588 million yuan in the first half of 2026, up 92.0%. Adjusted net loss narrowed from 206 million yuan in 2024 to 91.3 million yuan in 2025, and 14.88 million yuan in the first half of 2026.

The risks are equally concrete: the top five customers accounted for 88.0% of 2025 revenue, the top five suppliers for 84.8% of purchases, operating cash flow remains negative, and net liabilities stood at 2.644 billion yuan as of June 30, 2026. The company sources computing capacity mainly through long-term arrangements rather than building its own facilities, to limit upfront capital spending.

The company has completed 14 private funding rounds, with Beijing state-owned investors including Beijing State-owned Capital Management and Zhongguancun Capital. Whether the listing proceeds, and at what valuation, depends on the exchange review and market appetite for AI computing assets.

Ai2 released the open-source training framework Olmo-core 3 on October 1, saying it scales mixture-of-experts (MoE) training into the trillion-parameter range.

In its own test on eight NVIDIA B300 GPUs, a 47-billion-parameter MoE processed 52,000 tokens per second per GPU, versus 19,400 with the earlier implementation — about 2.7×. A separate test ran a 1.2-trillion-parameter model across 512 GPUs, but with random routing, measuring system performance only, not model quality.

The framework switches to distributed data parallelism, keeping expert weights resident on GPUs instead of repeatedly gathering them, and supports the MXFP8 low-precision format, self-tested at about 21% higher throughput than BF16. It underpins the next generation of the open Olmo models; code and a technical report are public, and real training results await third-party use.

Amazon Web Services has open-sourced a small "decision model" that lets agent workflows make routing choices at low cost.

TechCrunch reported on October 1 that AWS's Strands Labs released Strands Decider 2B, fully open-sourced, available now, and small enough to run locally. Built on the torso of Qwen3.5-2B, it does not generate text; it picks among pre-decided options and delivers a confidence score. The project began as a side effort by distinguished engineer Marc Brooker, whose September 28 blog post says it briefly topped the Jevbench ranking for its size class.

The category took off after TypeSafe's Jev. TypeSafe CEO Diogo Almeida said the current batch of imitators seems "more like ML people wanting to implement a cool architecture," and that the hard part is keeping the models actually smart.

Shopify released Canvas on October 1, letting merchants build an entire store by chatting with its AI assistant Sidekick, with changes visible in real time.

Shopify's previous no-code editor worked by rearranging modules and blocks within a chosen theme, and deeper customization still required editing code or hiring a developer. Canvas renders the store's actual theme code rather than a static preview, so merchants can test full interactivity and animation. Shopify simplified the theme architecture so Sidekick can read and write theme files directly.

Wix, Squarespace, Webflow and others already offer similar AI site builders; Canvas brings that approach to a leading e-commerce platform. The product is early days by Shopify's own account, and real-world results remain to be seen from merchant use.

AI hardware design platform Flow Engineering announced on September 30 a $50 million Series B round, lifting its valuation to $750 million.

The round was co-led by Antonio Gracias, founder of Valor Equity Partners, and Gavin Baker of Atreides Management, with Sequoia Capital participating; Sequoia also led the company's Series A in October 2025.

Flow runs AI agents inside CAD, simulation and documentation tools to do impact analysis, flag conflicts and catch requirement failures. Its customers include Rivian, Volkswagen's joint venture RV Tech and Anduril. The claim that development cycles shrink from months to days is the company's own; the announcement carries no independent assessment.

A risk-control model rejects a loan, and customer service hands over a bar chart: this bar is the longest, so it's the culprit. The 2017 paper asked a harder question — on what grounds does that chart count as an explanation of the model?

The authors put it plainly: within the class of additive feature attribution methods, there is exactly one solution that satisfies local accuracy, missingness, and consistency simultaneously — the Shapley value. Any approach that doesn't follow Shapley must violate at least one of these axioms.

Don't rush to treat the bar chart as a verdict. The original user study involved only 30 and 52 participants, and the experiments focused on simple models and settings like MNIST; exact values are computationally infeasible, and the approximations actually used rely on assumptions that features are mutually independent or that the model is approximately linear.

Its reach is also limited: explanation methods outside the framework are out of scope, and for tasks beyond text and images, as well as unverified model types, no transferable conclusions are offered.

“A Unified Approach to Interpreting Model Predictions” (2017) | Next review 2027-09-20