The Polylog AI Intelligence Brief

Morning Edition · Monday, November 16, 2026

Tech

Alibaba's Qwen4-Max cuts active parameters to 71 billion and list prices by 25 percent, and its weights stay closed

On Alibaba's in-house Work360 benchmark the model gains about five points over Qwen3.8-Max, but the only independent index so far scores it at 43, against 59 for Anthropic's Claude Opus 5.5.

2 sources
Plausible

Markets

Anthropic plans to price its IPO on Wednesday after an amended filing showed $33.2 billion in nine-month revenue and about $895 billion in compute commitments

The range of $70 to $78 a share for 825 million shares implies about $916 billion to $1.02 trillion on basic shares, and the stock is expected to start trading on Nasdaq as ANT on November 19.

3 sources

Tech

Anthropic released Claude Mythos 5.2 four days after METR estimated that AI is speeding up the company's research about 1.9 times

The Mythos 5.2 system card reports stronger covert capabilities than earlier Claude models when the model's reasoning is hidden from a monitor, and Anthropic redacted most of the evidence METR considered most informative.

1 source
Corroborated

Tech

OpenAI says GPT-6 Astra led the proof behind a candidate lattice definition of the Standard Model

The physicist who directed the work calls it a proposal, not a finished solution, and the claim arrives as OpenAI and Anthropic both publish their own accounts of AI-assisted research results.

3 sources
Plausible

Tech

NVIDIA, SK Telecom and Rebellions agree to split inference between NVIDIA GPUs and Korean NPUs in sovereign AI racks

The memorandum, signed at AI Day Seoul with backing from South Korea's science ministry, connects the Rebellions Rebel R200 to NVIDIA systems over NVLink Fusion, with deployment expected in 2028.

1 source
Corroborated

Tech

OpenAI's ChatGPT Wallet lets its agents pay with tokenized Visa, Mastercard and American Express cards

The November 13 rollout is limited to adult US Pro subscribers, requires approval of each purchase, and works only in conversations with OpenAI's "dots" agents.

1 source

Tech

Researchers show general-purpose LLMs can implant backdoors by editing neural network weights directly

Across five LLM weight editors and three target models, including Qwen2.5-0.5B-Instruct, the authors report working backdoors with clean-task performance largely intact.

2 sources

Tech

Two papers find that results from lab evaluations of hallucination detectors and agent memory reverse under deployment conditions

Detector accuracy fell to near chance under one labeling method on small quantized models, and hard-deleting stale agent memories turned a gain of about 0.2 into a loss of about 0.18 as resolver errors rose.

3 sources

Tech

Google connects Gemini to Google Wallet and Plaid-linked bank accounts to flag subscription price increases

The read-only feature launched for US users on November 13, and Google's help pages say Gemini may train on transaction summaries if a user keeps Gemini activity switched on.

1 source
Plausible

Tech

Berkeley Lab's SYNAPS-I project fine-tunes Meta's SAM 3 and DINOv3 on government machines for the Genesis Mission

Meta says a grapevine drought study that took about a month of expert segmentation per time step now takes about 15 minutes, a figure that so far comes only from Meta's own account.

1 source
Plausible

Tech

An input-optimization method reproduces activation-steering effects, including anti-refusal, with no access to model internals at generation time

ActGCG searches for a prompt suffix whose activations match those of a steered model, which carries a known white-box control technique into a setting where only the input changes.

2 sources

Tech

Telegram beta builds show a hidden "Companion" agent and a bring-your-own-key connector for OpenAI, Anthropic, Google and xAI models

Telegram has neither announced nor enabled the features, but together they show the messenger preparing to route third-party models into chats while users pay the model vendors directly.

1 source
Plausible