Polylog
The Polylog AI Intelligence Brief

Morning Edition · Monday, July 20, 2026

Moonshot Halts New Kimi K3 Sign-Ups as Serving Demand Outstrips Its GPU Fleet

Tech

Moonshot Halts New Kimi K3 Sign-Ups as Serving Demand Outstrips Its GPU Fleet

The three-day-old 2.8-trillion-parameter open-weight model needs roughly eight top-tier accelerators to serve a single instance, and the resulting shortage pushes demand toward Chinese inference chips.

1 source
Corroborated
Alibaba Previews 2.4-Trillion-Parameter Qwen3.8, With Weights Promised but No Benchmarks Shown

Tech

Alibaba Previews 2.4-Trillion-Parameter Qwen3.8, With Weights Promised but No Benchmarks Shown

The multimodal preview arrives days after Kimi K3 and claims to trail only Claude Fable 5, yet it ships with no license, no model card, and no active-parameter count.

1 source
Plausible
Meta's Non-Invasive Brain-to-Text Decoder Works, but Only Inside a Room-Sized Scanner

Tech

Meta's Non-Invasive Brain-to-Text Decoder Works, but Only Inside a Room-Sized Scanner

Brain2Qwerty reaches a 32 percent character error rate from magnetoencephalography, and roughly double that from a wearable electroencephalography (EEG) cap, exposing the fidelity cost of skipping surgery.

1 source
BrainCo Demonstrates Thought-to-Robot Control With Under 200 Milliseconds of Latency

Tech

BrainCo Demonstrates Thought-to-Robot Control With Under 200 Milliseconds of Latency

The Chinese startup's electroencephalography (EEG) platform, shown at Shanghai's World AI Conference, drives robotic arms, humanoid robots, and four-legged robots from decoded brain signals rather than an implant.

1 source
Plausible
Anthropic Proposes an Industry Standard for Scoring Jailbreak Severity as New Research Shows Attacks Can Be Distilled

Tech

Anthropic Proposes an Industry Standard for Scoring Jailbreak Severity as New Research Shows Attacks Can Be Distilled

The framework, co-developed with Amazon, Microsoft, and Google, arrives the same week a paper shows that harmful chain-of-thought traces can be transferred into reusable jailbreaks.

2 sources
VarRate Cuts Long-Context Memory by Varying KV-Cache Compression Token by Token

Tech

VarRate Cuts Long-Context Memory by Varying KV-Cache Compression Token by Token

The training-free method targets the key-value cache, the dominant memory cost of long-context inference, without the structural limits of token-selection or fixed-rate approaches.

1 source
Study Finds a 'Global Workspace' of Verbalizable Representations Inside Language Models

Tech

Study Finds a 'Global Workspace' of Verbalizable Representations Inside Language Models

The paper argues that only a subset of a model's internal representations are available for verbal report and flexible reasoning, borrowing a framework from cognitive neuroscience.

1 source
Meta Opens a Model API and Expands Its Muse Generative-Media Line

Tech

Meta Opens a Model API and Expands Its Muse Generative-Media Line

The Spark model API and new Muse Image and Muse Video systems push Meta's generative stack toward editable, production-oriented output for developers.

2 sources
New Papers Push Specialized LLMs Into Agentic Healthcare and Cost-Aware Diagnosis

Tech

New Papers Push Specialized LLMs Into Agentic Healthcare and Cost-Aware Diagnosis

Cura 1T targets consultation, clinical reasoning, and workflow execution in one model, while GraphDx frames sequential diagnosis as a cost-versus-accuracy trade.

2 sources
Meta Adds AI-Triggered Parent Alerts for Teen Suicide Risk on Instagram

World

Meta Adds AI-Triggered Parent Alerts for Teen Suicide Risk on Instagram

A dedicated model flags high-risk conversations and notifies supervising parents, with every alert reviewed by a human before it is sent.

1 source
Anthropic Publishes the Origin Story of Claude Code as Coding Becomes the Lab Battleground

Tech

Anthropic Publishes the Origin Story of Claude Code as Coding Becomes the Lab Battleground

The account of how an internal command-line tool became Anthropic's flagship coding agent arrives as rival labs build dedicated coding teams to catch up.

1 source
Anthropic Solicits the Public's Hardest Questions About AI

Tech

Anthropic Solicits the Public's Hardest Questions About AI

The lab says it will collect and publicly address the toughest questions it receives, a move that arrives as funding narratives face sharper scrutiny.

1 source