The Polylog AI Intelligence Brief

Morning Edition · Friday, July 17, 2026

Markets

At the AI Frontier, Token Bills Are Approaching Payroll Scale, a16z and Ramp Find

The heaviest corporate AI users now spend about $90,000 per employee a year on model tokens, and cheaper tokens are accelerating that spending rather than capping it.

3 sources
Corroborated

Tech

Moonshot Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model It Says Rivals US Frontier

The mixture-of-experts model ships with a 1-million-token context and posts 93.5% on GPQA Diamond, with full weights promised by July 27.

2 sources
Plausible

Tech

OpenAI Claims a GPT-5.6 Configuration Produced a Proof of a 50-Year-Old Graph Conjecture

The system reportedly used 64 parallel agents and under an hour. Mathematicians call the proof elementary but note that it is unreviewed and thinly cited.

3 sources
Plausible

Tech

Meta's Brain2Qwerty v2 Decodes Typed Sentences From Brain Scans at 61% Word Accuracy

The non-invasive pipeline reads a magnetoencephalography signal and reconstructs text without surgery, up from about 48% a year earlier.

3 sources

Tech

NVIDIA Adds Jetson T3000 and T2000 to Push Robotics Foundation Models to the Edge

The Blackwell-based T3000 delivers 865 FP4 teraflops at roughly half the size and power of the T5000, paired with a 4-billion-parameter Cosmos 3 Edge world model.

3 sources

Tech

New Audit Tests Whether a Model's Chain-of-Thought Actually Depends on Its Stated Premises

The black-box method substitutes predicates in a prompt and checks whether reasoning changes, exposing chains that look sound but ignore their inputs.

1 source

Tech

OpenAI's First Hardware Device Will Reportedly Be a Screenless Smart Speaker

Bloomberg sources describe a portable, display-free device positioned as a home computer for the AI era, still in development with no ship date.

1 source

Tech

Google DeepMind and Isomorphic Labs Publish a Joint Approach to Biosecurity for AI Models

The labs describe safeguards meant to keep protein-design and biology models from lowering barriers to biological misuse.

1 source

Tech

Google Extends AI Mode With Connected Apps, Letting Search Act Across Third-Party Services

Users will be able to link outside services and have the assistant read and act on them directly inside Search.

1 source

Tech

Anthropic Redeploys Fable 5 and Proposes an Industry Framework for Scoring Jailbreak Severity

The model returned globally on July 1, alongside a shared severity scale developed with Amazon, Microsoft, Google and other Glasswing partners.

1 source

Tech

New Provenance Tool Traces Which Training Records Belong to a Given Author

OriginBlame targets record- and token-level attribution in training data, aiming to build the forget set that machine-unlearning methods require.

1 source

Tech

AI Companion Apps Have Been Downloaded 165 Million Times Since ChatGPT Launched, Appfigures Finds

The virtual-partner category, much of it explicit, has become a measurable consumer market with real spending, per data cited by Decrypt.

1 source