SecondSourceJudgment rebuilt from primary sources

How we checked the items below

For readers who want to dig in. What happened, in brief, is on the Today tab.

This issue draws on SecondSource's own research brief, produced early on October 18, and last night's sweep. Main-line item 1 rechecks a SecondSource call from September 18. Items 2 and 3 and most later items date from October 16 to 17. From the archive and Model watch are evidence updates on older topics. Last night's sweep brought in 561 pieces, and this issue uses 34 outside sources, each with a link to check.

1. [Today] (deep-dive column, October 18) GPU rental prices have no single "rise" figure: CoreWeave's 25% most likely sits in contract prices, the public index measures listed prices, and the two can't be subtracted

Not proven yet: that the 25% sits in contract prices is an inference; CoreWeave hasn't said what it covers

How we checked

SecondSource's deep-dive column read CoreWeave's 8-K presentation, quarterly report and earnings-call transcript in the original, and read getdeploying's methodology page and weekly dataset directly. That the 25% applies to contract prices is an inference. The 98% committed-contract figure is primary. The website price page has no web archive snapshot, the August baseline is a third party's old reading, and CoreWeave hasn't said what the figure covers. No one discloses what share of new contracts are short leases. Nebius's original announcement hasn't been reviewed; its new list prices were read directly from Nebius's website. The presentation is a fundraising marketing document.

Back to this item ↑

2. [Today] (event date October 17) DeepSeek releases V4.1-Flash-1017 weights under the MIT license, self-reporting wins over GPT-5.6 Sol on 9 of 11 benchmarks

Not proven yet: all 11 benchmarks are self-reported, with no third-party rerun

How we checked

The model card was read in the original, the quote was checked word for word, and SecondSource recomputed the table column by column. All 11 benchmarks were chosen and reported by DeepSeek, with no third-party rerun; the card doesn't give the source of rival scores. BugBunny's test set exists only in the post, which reached SecondSource through an account that often reposts DeepSeek news. "Cheaper" comes with no change in API list prices.

Back to this item ↑

3. [Today] (report relayed October 16) The Financial Times reports that Citi, Bank of America and JPMorgan Chase are restricting third-party AI agents from online banking; the full story is paywalled

Not proven yet: the full story is paywalled, and every figure comes through relays

How we checked

The Financial Times original is paywalled and hasn't been reviewed. The three banks' names, the US$1.6 trillion and the US$47 billion all come from X accounts and tech-media relays. The Next Web piece's URL is about Schwab blocking AI agents, and its full text hasn't been reviewed. Rohan Paul runs an X account that relays AI news; Jigar Shah's post quotes the same passage from Rohan Paul. Effective dates, technical methods and the banks' own words aren't in the relays. The relays don't say whether users can grant read-only access through open-banking providers.

Back to this item ↑

[Today] (weekly roundup October 16) IDC data relayed by Reuters and the trade outlet Semiconductor Engineering: in the first half of 2026, NVIDIA's AI accelerator shipments in China fell 44% year on year, and NVIDIA's share by units shipped was 34% against Huawei's 28%.

How we checked

Khan's model estimates that a full sales stop today would let the US keep a 10x to 100x lead in annual AI chip output into the mid-2030s, while continued exports could make the two sides' output overlap (ChinaTalk, 2026-10-17). That is a scenario estimate from an advocacy organization; Khan concedes the error bars are wide, and his think tank is pushing for a DUV export ban. The US-China gap side has no new number this issue; the October 13 issue's archive pick cited Epoch's estimate of an average seven-month overall gap between US and Chinese models.

(Semiconductor Engineering, 2026-10-16; Reuters, 2026-10-13)

Back to this item ↑

[Today] (posts October 16 to 17) François Chollet, creator of Keras and co-founder of ARC-AGI, says agents now write almost all code, yet engineers are more numerous and busier than in 2025 and employers' P&L usually hasn't moved; Box CEO Aaron Levie responds that software demand isn't fixed and engineering's weight in the economy is rising.

How we checked

One is a recollection and one is relayed, and the populations differ, so they can be read side by side but not compared. The 30% is one company's self-report, relayed, and whether the denominator grew with hiring isn't stated.

(François Chollet, 2026-10-16) (Aaron Levie, 2026-10-16) (Tim Sweeney, 2026-10-17) (Harry Stebbings, 2026-10-16)

Back to this item ↑

[Evidence update] (blog August 25, read directly on October 18) OpenRouter's official blog says the 13.8x rise in daily token usage during the August discount belongs to GPT-5.6 Luna alone. Terra grew 5.6x, and the undiscounted Sol grew 1.1x over the same period.

How we checked

The multiples don't net out organic growth, the window includes one-off arbitrage, and they can't be extended into a long-run demand curve; the figures are one platform's self-report.

(OpenRouter blog, 2026-08-25)

Back to this item ↑

[Business] (announcement October 16) Microsoft and NVIDIA jointly publish a white paper on managing risk from open-weight models in enterprises; its core sentence is "weight openness alone does not determine an AI system's risk," with risk sorted into five categories and the focus on the model layer.

How we checked

(Microsoft Azure blog, 2026-10-16)

Back to this item ↑

Still missing today. The gap that would most affect judgment is the share of short leases in CoreWeave's new contracts, which no newer cloud discloses. Other gaps: the full Financial Times bank story, the technical nature of the OpenCode attack, the source of DeepSeek's rival scores and 1017's API list price, the Reuters and IDC originals on Chinese accelerators, the full text of Anthropic's draft bill, the Microsoft and NVIDIA white paper itself, TechCrunch's Replit story, and IDC's smartphone report. Each of these either hasn't been reviewed or isn't provided in the original.

Sources & accounting

The past 24 hours. 561 new pieces came in from October 17 to 18: 464 academic papers (452 from arXiv, 12 from other sources), 72 social-platform posts, 19 company or personal blog posts, 3 show transcripts and 3 subscription newsletters. Of those 561, 69 were read in full and 76 were screened out as off-topic; the other 416 haven't been read yet and remain on the reading list. Among the 416 unread are 387 papers, 15 posts and 14 blog posts. Another 27 were judged not worth reading at pre-screening. They are already included in the 76 screened out and the 416 unread, not added to them. Author-tracked arXiv fetching didn't come in last night for a second straight day because arXiv restricted access; the 452 arXiv papers above came through other listing channels. Accounts with the most posts included @teortaxesTex, @bhorowitz, @elonmusk and @Miles_Brundage, with @teortaxesTex alone making up 21% of last night's X pull; the newsletters were one each from Interconnects, Latent Space and Zvi Mowshowitz (Interconnects); the shows were two ChinaTalk episodes and one Cognitive Revolution episode. The CoreWeave filings, getdeploying dataset and Nebius website prices behind main-line item 1 were read directly by SecondSource's deep-dive column; the OpenRouter blog behind Model watch item 1 and the MOFCOM annex behind From the archive were checked directly and aren't part of the reading list.

One-time catch-up. No new one-time catch-up of older material today.

A note on source concentration. ⚠️ Four items were first picked for the main line, and the fourth, OpenCode's breach, now runs as Also happened item 5. The load-bearing figures for main-line items 1 and 2 and for the OpenCode item all come from the parties themselves: CoreWeave's fundraising presentation, DeepSeek's model card and OpenCode's official post; item 3's Financial Times report is available only through relays. The four parties are all different, and no single source exceeds a third. ⚠️ The Nebius original announcement behind main-line item 1, the full Financial Times story behind item 3, the Reuters and IDC originals behind Chips item 1, and the full stories or draft behind the first four "Also happened" items haven't been reviewed.

The sources we track. The long-term roster currently has 529 named sources: 302 on social platforms, 90 shows, 51 news outlets, 48 blogs, 48 paper authors and 46 newsletters. Channels are counted separately, so the total exceeds 529, meaning some sources are counted in more than one channel. ⚠️ Those are counts of tracked sources, a different population from the 561 new pieces above; they can't be added or compared. Representative names: on social platforms, Mark Zuckerberg, Lucas Beyer, Graham Neubig, Jordan Schneider and Arvind Narayanan; among newsletters, Zvi Mowshowitz, Miles Brundage, Dean Ball and Kevin Xu; on the shows roster, interviewees such as Sam Altman, Dario Amodei and Michael Intrator. This issue uses 34 outside sources in the body. Only outside links actually cited in the body are counted, and links to SecondSource's own site are left out; this is a different population from the 529 tracked sources above. On the Taiwanese side, 7 new sources were added in September, bringing the total to 129; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.