The Polylog AI Intelligence Brief

Morning Edition · Tuesday, November 17, 2026

Tech

Nvidia reports earnings today under an interim chief as its board declines Trump's suggestion to reinstate Jensen Huang

Colette Kress leads the call against the company's own $108 billion revenue forecast, while directors concentrate on an investigation into reports that Huang was the target of a foreign intelligence effort.

11 sources
Plausible

Markets

Anthropic prices its Nasdaq listing on Wednesday with orders reported at about six times the 825 million shares on offer

At the top of the $70 to $78 range, the offering plus a $26 billion private placement with chip and technology companies would bring in about $90 billion, according to deal coverage.

7 sources

Tech

OpenAI gives four outside safety groups access to unreleased models, training checkpoints and training data

Apollo Research, METR, Transluce and Goodfire will work inside OpenAI's systems, but OpenAI keeps the right to redact their findings and to ask for publication delays.

7 sources

Tech

Anthropic sells the same model twice, as Fable 5.2 for the public and Mythos 5.2 with looser cyber and biology safeguards for verified customers

Anthropic's launch table reports 93.0% on SWE-Bench Pro for Fable 5.2, but one outside analysis found no independent confirmation for nine of the table's 14 evaluations.

7 sources
Plausible

Tech

Lambda says more than 10,000 Blackwell Ultra GPUs now run on Nvidia co-packaged optics switches with lower latency than pluggable optics

Lambda reports a latency advantage that grows from 3.8% at rack scale to 10.1% at pod scale with matched throughput, and puts each switch's power draw at 3.95 kilowatts against 7.0 for a conventional unit.

5 sources

Tech

OpenAI widens free frontier-model access to as many as 200,000 scientists and reports a self-estimated 1.9-fold speedup

The speedup comes from 251 survey respondents who had applied for AI access, and OpenAI says its data cannot establish a causal effect.

5 sources

Tech

Deleting "stale" memories lowers LLM agent accuracy when the classifier that judges staleness makes errors, a new paper finds

The result appears as the Chinese wearable maker Looki introduces Loom, a memory system that keeps each remembered item tied to its timestamp and its original camera and microphone recording.

5 sources

Tech

Windows 11 telemetry records when developers run Claude, Codex, Copilot or Cursor, a reverse engineer reports

Xusheng Li found that the data leaves the machine as a compact usage bitmap under Windows' optional diagnostic setting, with no prompts or source code in the report he examined.

4 sources
Plausible

Tech

Hallucination detectors that lead on 7-billion to 70-billion-parameter models reverse their ranking on 4-bit 3-billion-parameter models, an arXiv audit finds

A second paper posted the same day, LoRAWeaver, selects LoRA adapters layer by layer through cosine similarity in adapter-output space, aimed at the growing adapter pools that deployed small models accumulate.

4 sources

Tech

Nvidia bundles four DGX Spark desktops into a ring-connected developer kit for supercomputing work at SC26

The cluster pools 512 GB of unified memory over ConnectX-7 links and is on sale through hardware partners now, while its HPC software setup guide is still listed as coming soon.

4 sources

Tech

OpenAI offers vetted organizations a free image classifier for AI-generated child sexual abuse material through its API

According to coverage of the release, OpenAI estimates that the classifier, developed with the National Center for Missing & Exploited Children, cuts image review workloads by 48% while finding about the same number of abusive images.

5 sources

Tech

LVMH's 17 media brands, including Les Echos and Le Parisien, sign a content and newsroom-tools deal with OpenAI

ChatGPT will summarize and link to the titles' articles, and neither side disclosed payment terms or any commitment on referral traffic.

5 sources
Corroborated