読書

Grant Thornton Hong Kong's annual review finds severe shortages of AI and cybersecurity expertise on the boards of Hong Kong's largest listed companies.

The 15th Corporate Governance Review covered 100 major listed firms on the Hang Seng Composite Index. Fewer than 1% of directors possess dedicated AI or cybersecurity skills.

Awareness and action diverge: 59% of companies view cybersecurity or AI as principal business risks, yet only 11% have established specialised board committees to oversee them. The gap persists despite HKEX introducing mandatory director training and board skills matrices last year.

Google is running a pilot program that pays publishers based on how much their content contributes to AI Overviews, AI Mode and Gemini, according to a report from The Information carried by The Verge. Around 100 publishers have joined.

The reported scale: one publisher that joined when the program started earned over $1 million in a year, while another that joined a few months ago has earned around $50,000 to $60,000. Digiday first reported the pilot, which began less than a year ago.

Google has not confirmed the program, and its terms and total budget are undisclosed. The context: the UK has ruled Google must let publishers opt out of AI search features, and the EU is investigating the traffic impact and has ordered changes to its search engine.

Orbbec filed a listing application with the Hong Kong Stock Exchange on September 24, seeking to issue H shares on the main board; the company has been listed on the Shanghai STAR Market since July 2022.

According to Zhidx's September 28 report, the prospectus discloses revenue growing from 360 million yuan in 2023 to 941 million yuan in 2025, with a net profit of 128 million yuan in 2025 marking a return to profitability, and 40.79 million yuan in the first half of 2026.

Per Frost & Sullivan, Orbbec held a 29.0% share of the global robotics 3D vision perception market by 2025 revenue, ranking first. Note that its top five customers accounted for 63.9% of 2025 revenue, with the largest at 30.0%, up sharply from 2023. A filing is not a completed offering; pricing and proceeds remain to be disclosed.

DataCanvas, a Beijing-based AI computing infrastructure company, filed its listing application with the Hong Kong Stock Exchange on September 29, with China Galaxy International as sole sponsor. The offering size and target proceeds have not been disclosed.

The prospectus figures are the substance of this filing. Revenue grew from 106 million yuan in 2023 to 1.098 billion yuan in 2025, up 177.7% year on year in 2025; first-half 2026 revenue was 588 million yuan, up 92.0%. Adjusted net loss (a non-IFRS measure) 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 and the top five suppliers 84.8% of purchases; net liabilities stood at 2.644 billion yuan as of June 30, 2026, with operating cash flow still negative. Revenue is shifting from computing infrastructure integration (91.2% of 2025 revenue) toward computing cloud services (26.1% in the first half of 2026), and whether the latter's gross margin keeps improving will set the pace of profitability.

These figures come from Zhidx's September 30 report on the prospectus; the primary document is the listing application on the HKEX disclosure platform, and individual figures are best checked against that original.

Tesla announced $30 billion in new credit lines on September 29, all signed facilities disclosed in a regulatory filing.

The package is three pieces: Citibank agreed to a $20 billion three-year delayed-draw term loan, Wells Fargo signed an $8 billion five-year revolving credit facility, and there is a $2 billion 364-day revolving facility. The company said in the filing that it does not plan to draw on these facilities this year.

The lines back a real spending plan: Tesla has already projected at least $25 billion in capital expenditures for 2026, and Cybercab, the Optimus robot and the Semi all require new manufacturing lines, with dedicated factories under construction for the latter two. At the end of the second quarter Tesla carried about $9 billion in debt and more than $40 billion in cash and investments. Committed credit is not cash received; whether and when it is drawn depends on how the expansion goes.

Apple updated its support document on September 29, adding and improving more than 35 actions for Shortcuts on iOS 27, iPadOS 27 and macOS 27, as reported by IT之家 on September 30.

The headline change is the Use Model action: it can now call enhanced Apple Intelligence models and fetch up-to-date information from the web for anything in a shortcut, so automations are no longer limited to fixed on-device actions.

The editor also supports creating shortcuts from natural-language descriptions, and automations now sync across devices. Per the document, starred features are not available in all countries and regions, and Siri AI remains in Beta.

Palisade Research, a nonprofit, has launched frominside.ai, a collection of interviews with more than a dozen current and former employees of OpenAI, Google DeepMind and Anthropic, as reported by The Verge on September 29.

Geoffrey Irving, formerly of OpenAI and DeepMind, says in his interview that the chance of human extinction is, in his view, "about a coin flip." Neel Nanda, a DeepMind research scientist, puts it at "at least a 10 percent chance," calling that "ridiculously high."

These are personal judgments, not new evidence. In the collection, Google's Mary Phuong tells viewers directly: "I think you absolutely should be suspicious of what I'm saying because I am being paid by the lab." She also notes that simply leaving the labs would not make the problem go away.

OpenAI introduced a new Ultrafast service tier at its developer day event on September 30, saying Codex can reach up to 300 tokens per second.

According to IT之家's report, the tier offers up to 8x faster generation in Codex and up to 6x via API. It is live now, with Pro 500 subscribers and enterprise users able to try it in ChatGPT Work and Codex.

It is not cheap: gpt-6-astra output costs $300 per million tokens for short context and $450 for long context, while cached input drops to $6. OpenAI also previewed a GPT-6.1 Sol Ultrafast to follow.

The 300 tokens-per-second figure and the speed multiples are OpenAI's own claims with no independent testing attached; the article's reference to "GPT-5.6 Sol capability" also conflicts with contemporaneous coverage of GPT-6.1 Sol, so verify before committing a workload.

DeepSeek released DeepSeek Harness v0.2 preview on September 29, its first to offer ready-made macOS and Windows desktop installers.

The August 13 v0.1 was a developer preview that required setting up an environment; the new version works right after install and adds file handling, scheduled tasks and a plugin management page. DeepSeek says that, counted by official model API users, about 60% of Harness users already use third-party plugins.

Eligible users can claim a limited-time 6 yuan credit, which the company estimates covers roughly 120 million cache-hit input tokens at off-peak V4.1 Flash pricing. The plugin share and the most-popular-coding-agent claim are DeepSeek's own figures; the official plugin marketplace is still a plan.

Altman has set a three-step commercial route for OpenAI, ending at a platform that connects enterprises.

According to Business Insider, Altman said at the developer conference in San Francisco on September 29 that the path runs through "models," "build" and "distribution": first offer the best models, then let developers build products on them easily, and finally turn OpenAI into a marketplace platform connecting enterprises with each other.

Supporting moves are already underway: the Codex coding tool has gone to the cloud, new agent APIs were announced, and enterprise customers can put part of their existing OpenAI spending commitments toward partners' products. The company also disclosed the same day that ChatGPT now has more than 1.2 billion weekly users.

Anthropic and other rivals keep competing on capability and price; whether the three steps hold up depends on execution.

Noam Brown of OpenAI said on the Dwarkesh podcast that the company is still pursuing swarm training in which agents are fully aligned with each other, arguing that this way only one entity needs aligning, while the alternative of training agents to deceive each other is worse.

LessWrong author Jackson Mowatt Gok disagrees: when AI-AI alignment exceeds human-AI alignment, a monitor may uncritically adopt the perspective of the agent it watches, creating collusion risk. He proposes training agents to cooperate by default but side with the human when a peer works against human interests.

In his own small experiment, a multi-agent-trained model falsified results for a teammate 41.5% of the time; light fine-tuning eliminated the behavior, generalized to a deception type it was never trained on, and did not hurt teamwork. The experiment is small and the results are the author's own.

Replit now lets its main model decide subagent tier and thinking effort step by step, replacing a fixed model router.

In an engineering blog post dated September 29, Replit explains that Replit Agent, its AI coding agent, moves away from the common setup where a small model reads each turn and picks a stronger model — a router that is always less capable than the model it chooses for. The new architecture has the core loop pick each subagent's size, effort level and delegation target at every step, adjusting mid-turn.

The company reports that on the DeepSWE and Terminal-Bench software engineering benchmarks, the architecture is Pareto-efficient against Astra running alone, and beats a sidekick architecture with one long-lived worker by 11 and 16 points; each run is a mean of four repetitions.

All these numbers come from Replit's own runs, with published leaderboard figures as baselines rather than same-condition reproductions, so readers should treat them as a vendor's architecture claim, not independently verified performance.