Daily Brief Oct 29, 2026
How we checked every item today except "Also happened"
For readers who want to dig in. What happened, in brief, is on the Today tab.
This issue draws on material from October 28 to early October 29, plus three late-October preprints in Research & models. Last night's collection took in 366 pieces, and this issue uses 27 outside sources with checkable links; the later sections also draw on material from the past week.
1. [Today] Alphabet's equipment spending last quarter ran US$9.065 billion past what its operations brought in, and it sold US$28.7 billion of bonds plus new shares in the same quarter
Not proven yet： negative free cash flow isn't an inability to pay; Alphabet holds about US$265 billion
How we checked
Alphabet's figures come from the 8-K filed with the SEC, a primary document; the bond issuance, new stock and halted buybacks are in the same filing's cash flow statement and quarterly supplementary section. Free cash flow is Alphabet's own measure and doesn't deduct stock-based compensation; the roughly US$265 billion of cash and marketable securities is from the same filing. Microsoft's Azure and cash purchase figures come from the press release on the company's investor relations page. "Below expectations" and "mainly leases, no debt yet" are Moorhead's words, and the roughly US$46 billion market expectation is @wallstengine's compilation; neither is in the press release. The two same-day earnings reports are independent primary sources about two different companies, so neither corroborates the other.
2. [Today] Bloomberg reports that even top-rated AI data center debt is getting more expensive, and 30-plus financiers say raising money is harder, while issuance is still about level with last year
Not proven yet： most of the 30-plus interviewees are anonymous, and "near the limit" is the reporter's synthesis
How we checked
This is a single original Bloomberg report, and the key quotes and figures come from its text: Bloomberg's own data and quotes from Campbell, Roberts and others. Most of the 30-plus interviewees are anonymous, and "credit markets are nearing their capacity to absorb" is the reporter's summary, not a measurement by any interviewee. The report cites Elon Musk's October 25 remark that the debt market is nearly closed to operators without independent revenue as background (Musk on X, 2026-10-25), so Musk's post doesn't count as a second source.
3. [This week] (research page, October 26) AI math company Axiom says its agents turned a 15,000-page proof into Lean code a machine can check, in 24 days, and people checked that the theorem was stated correctly
Not proven yet： everything is Axiom's own account; no independent team has reported a successful build
How we checked
The research page and the GitHub README are both Axiom's own documents, primary but self-reported. Axiom's post was relayed by Stella Biderman, executive director of EleutherAI, who disclosed that Axiom had written to ask for promotion (Biderman on X, 2026-10-27). OpenAI chief product officer Kevin Weil wrote that the result is now beyond doubt, which goes further than Axiom's own account (Weil on X, 2026-10-27). Neither is an independent check. The public material consists only of Axiom's documents and these two relays; the Lean community, the maintainers of Mathlib, Lean's shared math library, and other independent teams have not yet published build results.
4. [Today] Google releases Gemini 4 Carbon to cyber defenders first and developers later, and its first place on a leaderboard and its security scores are Google's own reports
Not proven yet： the first place and the security scores are Google's own, with no rivals named
How we checked
The blog post is Google's first-party account, primary but self-reported; Pichai's post the same day is from the same company and doesn't count as a second source. The Artificial Analysis index is a third-party leaderboard, but the score of 59 and the first place rest on Google's blog and Pichai's post. Pichai's post says the v5.0 index was revised recently, so its scores can't be compared with the old version's. The internal benchmarks have no published method. Google's Chrome team says patch quality is more than ten times better than another lab's frontier model, without naming the rival or defining the metric.
[Today] NVIDIA lets RTX Spark laptop users split GPU resources between AI agents running in the background and foreground programs, with no figures on how the split performs.
How we checked
The blog doesn't say how much GPU the agents get or how much performance foreground programs lose (NVIDIA, 2026-10-28).
[Today] Investment podcast host Patrick O'Shaughnessy relays guest Cliff Sosin: the Jevons paradox, "the cheaper it gets, the more gets used," is only one point on the demand curve, and past a certain price a market shrinks.
How we checked
[This week] (preprint, arXiv 2610.37283) The University of Stuttgart's BenchMeUp uses an agent pipeline with a verification loop to build software engineering benchmark tasks from natural-language specifications; the authors checked 91% of the 198 tasks it built as fully valid, against 5% of 200 targets for the general-purpose agent Codex run as a control.
How we checked
The abstract doesn't say whether Codex had the same specifications, budget and verification loop as BenchMeUp, so the 91% against 5% can't be credited to the verification loop alone.
[This week] (preprint, arXiv 2610.37459) On an eight-task clinical subset, M4Bench finds a complete specification, not a skill pack, drives most of an agent's score gain. The specification adds 0.162, and a skill pack adds another 0.024, which is statistically indistinguishable from zero.
How we checked
The abstract doesn't say which model was used or whether that gap is significant (arXiv 2610.37459).
[This week] (preprint, arXiv 2610.35971) A study of 147,002 coding-agent trajectories finds that 20,000 tokens into a run, whether the attempt will fail is no more predictable than a coin flip, and in 216 of 224 configurations running to the end solved more problems with fewer tokens than restarting midway.
How we checked
The abstract doesn't say which models or problem sets were used (arXiv 2610.35971).
[Today] Microsoft turned its own agent harness into a managed service in public preview on Foundry Agent Service, and from December 1 the compute agents use in the sandbox is billed by the hour.
How we checked
Microsoft gives no performance figures and no comparison with a self-built harness.
[Today] Slack opened the Slack Code API, which lets an agent open a dedicated channel for a piece of work, to all developers; in August only a few partners such as Anthropic, Cognition, GitHub and Vercel could use it.
How we checked
Slack says 30% of GitHub Copilot's agent sessions in Slack opened a dedicated channel; that figure is self-reported by Slack, with no period, measurement method or note on extra charges.
[Looking back] (SecondSource deep dive, October 22) AI safety work catches problems by reading a model's reasoning draft, and the deep dive argued that rules should be tied to measurable outcomes, not to model architecture.
How we checked
Sources & accounting
The past 24 hours. 366 new pieces came in from October 28 to 29: 137 social-platform posts, 115 academic papers, 50 from a second academic-paper feed (counted separately from the academic papers), 47 company and personal blog posts, 9 company filings, 6 subscription newsletters and 2 show transcripts. Of those 366, 162 were read in full, 160 were screened out as off-topic (4 of them before fetching), 33 haven't been read yet, 10 were logged with a stated reason for not reading closely, and 1 is being refetched.