The Polylog AI Intelligence Brief

Morning Edition · Monday, August 24, 2026

Tech

Nvidia Commits About $7 Billion to Open-Weight Models Through a Poolside Licensing Deal

The chipmaker will pay $6 billion to license Poolside's training stack, invest $1 billion at a $12 billion pre-money valuation, and move more than 100 of the startup's staff onto its Nemotron open-weight team.

4 sources
Corroborated

Tech

Microsoft Takes Delivery of the First Production Vera Rubin Racks as Nvidia Declares Full Production

The NVL72 configuration pairs Vera central processors with Rubin accelerators carrying up to 288 gigabytes of high-bandwidth memory, and Nvidia claims up to five times the inference throughput of Blackwell.

3 sources
Corroborated

Tech

An Anonymous Model Called Ox Alpha Tops an Independent Coding Benchmark With No Disclosed Owner

A developer's run of the DeepSWE agent benchmark put the stealth model at 80 percent against 65 percent for Claude Fable 5 and 52 percent for GPT-5.6 Sol, but no vendor has published the result or claimed the model.

4 sources
Developing

Tech

New Papers Target the Weakest Point in Agent Deployments: Cross-Domain Tool Abuse in the Model Context Protocol

One arXiv submission proposes a control layer called AEGIS for resource abuse across trust domains, a second argues that end-to-end success rates hide where long-horizon security agents actually fail.

3 sources
Plausible

Tech

Meta's Open Vision Models Are Running Inside a United States Department of Energy Science Program

Segment Anything 3 and DINOv3 on 300 accelerators at the National Energy Research Scientific Computing Center cut per-time-step image annotation from about a month of expert work to roughly 15 minutes.

3 sources
Corroborated

Tech

Meta Is Selling Its Frontier Model by the Token While Keeping the Weights Closed

The Meta Model API prices Muse Spark 1.1 at $1.25 per million input tokens and $4.25 per million output tokens, a deliberate break from the open Llama strategy that built the company's developer base.

4 sources

Tech

MiniMax Ships a Desktop Agent That Runs a Whole Video Production From One Brief

Design launches four specialised agents in parallel for copy, images, video and audio, and calls out to H3, GPT Image 2 and Nano Banana Pro, none of which run locally.

3 sources

Tech

Black Forest Labs Splits Resolution Off Into a Separate Video Upscaling Endpoint

FLUX Upscale regenerates video at 1.5, 2 or 3 times its input scale, reaching roughly 4K from high-definition source, and is tuned to repair the specific artifacts FLUX 3 leaves behind.

3 sources

Tech

Harvard Business School Is Selling a $699 Bootcamp Taught by AI Replicas of Its Faculty

The eight-week HBS Foundry program uses avatars built by HeyGen to hear pitches and run mock board meetings, and has been piloted at more than 100 universities since April.

4 sources
Corroborated

Tech

A Safety Evaluation Paper Argues Single-Dimension Scores Hide How Aligned Models Actually Fail

The aiXamine authors give an example of a model scoring 99.3 on safety alignment while refusing one in three benign requests, and a second paper argues refusal training leaves harmful knowledge intact underneath.

4 sources

Tech

Paper Finds Therapy Chatbots Misread the Vocabulary of the Teenagers Already Using Them

An arXiv submission cites 13.1 percent of United States adolescents, about 5.4 million people, turning to generative AI for mental health advice, then documents where the models' clinical reasoning breaks on their language.

3 sources
Plausible

Tech

Two Papers Argue the Constraint on Coding Agents Is Now Specification and Memory, Not Model Capability

One proposes a spec-driven development lifecycle for million-token coding agents, the other spawns Claude Code instances pre-loaded with prior work instead of starting each session from an empty context.

3 sources