SecondSourceJudgment rebuilt from primary sources

How we checked every item today except "Also happened"

For readers who want to dig in. For what happened, in brief, go back to this issue.

This issue draws on material from November 16 to early November 17. Last night's collection took in 549 new pieces, 24 of them read in full, and this issue uses 22 outside sources with checkable links; the later sections also draw on material from the past week. The full edition of this issue is the archive of record.

1. [Today] Claude Fable 5.1's share of Databricks customers' proprietary-model tokens fell from 5.5% to 0.6% in six weeks

Not proven yet: every figure comes from Databricks' own customers and Databricks' own benchmark

How we checked

Shares and benchmark results are from Databricks' blog post. The sample is more than 10,000 enterprises using Unity Gateway, Databricks' model gateway. These companies connect to several model vendors at once and may compare prices more often than a typical company. Databricks also sells a service that switches models automatically by cost, so the conclusion fits Databricks' product direction.

Back to this item ↑

2. [Today] Open-weight models' share of Databricks customers' tokens rose from 7.3% to 30.9% in four months

Not proven yet: the OpenRouter figures appear only as cited by Hugging Face

How we checked

The Databricks figures are from Databricks' blog post, and the sample is limited to companies using Databricks' gateway. The report's closing note says the current data covers September 21 to November 7; the report doesn't explain which data the July baseline comes from. The 28% is the figure as of November 2. The OpenRouter figures are as cited in Hugging Face's blog post, which doesn't say how OpenRouter defines open weight. The same citation gives average prices: US$0.28 and US$1.62 per million input and output tokens for open-weight models, against US$1.27 and US$6.36 for closed models. Hugging Face says it bills at providers' own prices with no markup. OpenRouter is itself a party to Hugging Face's routing change.

Back to this item ↑

3. [Today] Anthropic shortens its disclosure deadline for high-severity vulnerabilities from 90 to 30 days from 2027

Not proven yet: the claim that rediscovery is rising comes with no figure

How we checked

The above is from Anthropic's announcement. For high-severity reports already past 90 days and still unpublished, Anthropic says it will allow another 30 days. These are the deadlines Anthropic uses for its own reports; other research teams' and vendors' deadlines are unaffected.

Back to this item ↑

4. [Today] OpenAI lets four outside evaluators examine models still in training

Not proven yet: how independent the evaluations are depends on the reports published later

How we checked

The terms are from OpenAI's announcement. Transluce's follow-up evaluation is scheduled for early 2027 and runs about eight weeks; the scope of METR's monitoring evaluation is still to be negotiated. METR's own monitoring-evaluation proposal of October 14 also says that relying on company employees for sampled interviews "can reduce the level of independence" (METR, 2026-10-14). The quote is from Kokotajlo's post on X. The quote omits two passages: one says evaluators notify OpenAI of each conversation, and the other says OpenAI can withhold information entirely on security, IP or technical grounds; both point the same way as the rest.

Back to this item ↑

[Today] NVIDIA published MLPerf results on November 16 showing Vera Rubin NVL72 at about twice the per-GPU pretraining speed of the previous-generation GB300 on the DeepSeek-V3 test.

How we checked

(NVIDIA, 2026-11-16) (CoreWeave, 2026-11-16) (NVIDIA technical blog, 2026-11-16)

Back to this item ↑

[Today] AMD announced on November 16 that it has completed its acquisition of World Labs, and co-founder Fei-Fei Li joins AMD as executive vice president and chief scientist.

How we checked

The press release doesn't give the deal value or details of regulatory approval.

(AMD press release, 2026-11-16)

Back to this item ↑

[Today] NVIDIA announced Rubin HPC, a separate GPU for scientific simulation due in 2028, with precision, power and price not yet disclosed.

How we checked

(NVIDIA, 2026-11-16)

Back to this item ↑

[This week] Paul Christiano, an OpenAI board member who heads ARC, a group researching how to make AI act on human intent, argues on the Dwarkesh Podcast that frontier labs should pause algorithmic progress until their security can stop the theft of weights and algorithms.

How we checked

This is Christiano's personal position, and the probability attached is Christiano's subjective estimate with no model behind it.

(Dwarkesh Podcast, 2026-11) (Anthropic S-1/A, 2026-11-17)

Back to this item ↑

[This week] According to the paper's abstract, two authors report that pretraining from scratch in NVFP4, NVIDIA's 4-bit format, leaves validation loss only 0.5% above BF16, but the model needs about 1.32 times as many exposures to memorize the same rare facts.

How we checked

The abstract doesn't give the model size (arXiv 2611.12832, 2026-11-09).

(arXiv 2512.02010, v5 2026-05-09)

Back to this item ↑

[This week] According to the paper's abstract, three authors report that a robot that trains no policy and simply copies the actions of the nearest demonstration reaches a 91.7% success rate on the LIBERO simulation benchmark.

How we checked

The abstract doesn't say which task suites the 91.7% covers, how many trials were run, or whether it is computed the same way as OpenVLA-OFT's 97.1%.

(arXiv 2611.13160, 2026-11-09) (arXiv 2502.19645, v2 2025-04-28)

Back to this item ↑

[Today] A Google developer blog post estimates that pausing a coding agent after two idle minutes and returning its machine to a shared pool, which the author calls "parking," saves 40% on machine costs against one machine per session at 2 prompts an hour, and only 4% at 4 prompts an hour.

How we checked

The author also says prompt frequency is the least certain input in the whole estimate. The author's estimate runs 100 coding agents on GKE, Google's Kubernetes service: one Pod per session needs 34 machines, about US$1,288 a month, while parking needs 15 standby machines plus 1 control machine. Another approach, overcommitting CPU without parking, needs only 9 machines for 100 agents; the author's 40% estimate doesn't include it.

(Google Developers Blog, 2026-11-16)

Back to this item ↑

[Today] Perplexity, the AI search company, launched Frontier Research, whose scope includes pretraining, with more than 10,000 GPUs of compute; the post doesn't give the GPU model or whether Perplexity owns them.

How we checked

(Perplexity, 2026-11-16)

Back to this item ↑

[Today] OpenAI switched its free plan for academic researchers from Pro to Plus and doubled its enrollment target to 200,000, saying most pilot users stayed within Plus limits.

How we checked

(OpenAI, 2026-11-16)

Back to this item ↑

Still missing today. The gap that most affects judgment is the full Databricks AI Index data: the blog post doesn't say which data the July baseline comes from, and doesn't give the full model rankings, the number of benchmark tasks or the harness each model ran in. Hugging Face's post doesn't give OpenRouter's definition of open weight or how the average prices were computed. No public figure for Fable 5.2's enterprise share turned up for this issue. The full MLPerf Endpoints entrant list and Lux's ranking haven't been checked against the original MLCommons and TOP500 lists. Both research items rest on the abstracts; model sizes and per-task figures are in the full papers. Of the 503 pieces not yet read, 283 are academic papers and 124 are social-platform posts.

Sources & accounting

The past 24 hours. 549 new pieces came in from November 16 to November 17: 284 academic papers, 143 social-platform posts, 85 company or personal blog posts, 13 originals linked from posts and articles, 9 subscription newsletters, 7 podcast transcripts, 7 from a paper-summary site (counted separately from the academic papers) and 1 company filing change. Of those 549, 24 were read in full, 13 were screened out as off-topic (4 of them before fetching), 9 were only indexed without a full read, and 503 haven't been read yet. Academic papers, 284: 1 read, 283 not yet read. Social-platform posts, 143: 8 read, 7 screened out, 4 only indexed, 124 not yet read. Blogs, 85: 3 read, 1 screened out, 5 only indexed, 76 not yet read. Linked originals, 13: 12 read, 1 screened out. Paper-summary site, 7: 4 screened out before fetching, 3 not yet read. The 9 newsletters, 7 podcast transcripts and 1 company filing change were not read in last night's batch; the Anthropic filing and the Dwarkesh episode cited in the body were read separately. Named sources: Lisa Su and Daniel Kokotajlo on social platforms; the Dwarkesh Podcast episode with Paul Christiano; and Anthropic's amended IPO registration statement of November 17 among company filings. Last night's scan covered arXiv, X and blogs; one blog source failed to fetch.

One-time catch-up. No new one-time catch-up material came in this window. The catch-up material already collected is dated between July 1 and November 17 this year, counted separately from the past 24 hours above. The two research papers were collected earlier this week, and Four Over Six and OpenVLA-OFT are older comparison papers; none of them is among the 549 pieces above.

Source concentration. Main-line items 1 and 2 both come from the same Databricks report, with item 2 also compared against OpenRouter figures as cited by Hugging Face; Databricks also sells a service that switches models by cost. Main-line items 3 and 4 are announcements by Anthropic and OpenAI respectively. Both Research & models items rest on arXiv preprint abstracts, with results reported by their authors.

Tracked sources. 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, with the rest spread across earnings reports, keynotes and other channels. The same source may be tracked on several channels, so the channel figures add up to more than 529. These are counts of tracked sources, a different list from last night's piece counts above. Representative names include Lisa Su, Ethan Mollick, Elon Musk and Gary Marcus on social platforms; Zvi Mowshowitz and Dylan Patel among newsletters; and Dwarkesh Patel among shows. This issue uses 22 outside sources in the body, counting only outside links cited in the body. Sources from Taiwan total 271, with 44 added in October; Taiwanese media, Taiwanese companies' official channels and local reporters in Taiwan count, while foreign outlets relaying Taiwanese media do not.