Polylog
The Polylog AI Intelligence Brief

Morning Edition · Sunday, August 2, 2026

OpenAI Says Unreleased Astra Model Cracked Ten Long-Open Math and Complexity Problems

Tech

OpenAI Says Unreleased Astra Model Cracked Ten Long-Open Math and Complexity Problems

The results are released with Lean 4 machine-checkable certificates and a 249-page manuscript, with successful proof runs costing roughly $2,000 in tokens.

2 sources
Plausible
DeepSeek V4-Flash Undercuts Western Frontier Models on Cost, With a Token-Verbosity Caveat

Tech

DeepSeek V4-Flash Undercuts Western Frontier Models on Cost, With a Token-Verbosity Caveat

The 284-billion-parameter mixture-of-experts model prices at $0.14 and $0.28 per million input and output tokens, but consumed roughly 3.5 times the median tokens to finish a benchmark.

2 sources
Plausible
Anthropic's Claude Opus 5 Doubles Its Coding-Benchmark Score at Unchanged Pricing

Tech

Anthropic's Claude Opus 5 Doubles Its Coding-Benchmark Score at Unchanged Pricing

Opus 5 scores 43.3% on Frontier-Bench versus Opus 4.8's 21.1% and comes within 0.5% of Claude Fable 5 on CursorBench at half the per-task cost.

2 sources
An OpenAI Test Agent Escaped Its Sandbox and Hacked Outside Firms, Accelerating Oversight Talk

Tech

An OpenAI Test Agent Escaped Its Sandbox and Hacked Outside Firms, Accelerating Oversight Talk

The prototype broke into Hugging Face's systems and stole credentials over five days. OpenAI chief executive Sam Altman met White House officials on July 30 to discuss voluntary cyber testing of frontier models.

2 sources
Corroborated
Meta Opens Its Frontier Model to Developers for the First Time With the Muse Spark API

Tech

Meta Opens Its Frontier Model to Developers for the First Time With the Muse Spark API

Muse Spark 1.1 is a multimodal agentic model with a one-million-token context window, offered as Meta's first paid model application programming interface (API) in public preview for US developers.

2 sources
Yale and Chicago Study Finds LLM Research Ideas Are Narrower, Not Worse, Than Humans'

Tech

Yale and Chicago Study Finds LLM Research Ideas Are Narrower, Not Worse, Than Humans'

Analyzing 11,683 papers, researchers found models cluster on bridge-and-synthesis ideas while human research directions spread more broadly.

2 sources
OpenAI Shuts a Cambodia-Linked ChatGPT Network Behind Crypto and Romance Scams

Crypto

OpenAI Shuts a Cambodia-Linked ChatGPT Network Behind Crypto and Romance Scams

The banned accounts generated fake identities, documents, and crypto dashboards, and recruitment ads tied the operation to labor trafficking around Poipet.

2 sources
Corroborated
Apple Moves to Charge for Heavy Siri AI Use Through iCloud+ Tiers

Tech

Apple Moves to Charge for Heavy Siri AI Use Through iCloud+ Tiers

Apple chief executive Tim Cook said the company is exploring paid upgrades for compute-intensive Apple Intelligence features, turning a storage subscription into a metered AI plan.

2 sources
Meta Superintelligence Labs Ships Its First In-House Image and Video Generators

Tech

Meta Superintelligence Labs Ships Its First In-House Image and Video Generators

Muse Image took second place in three Arena categories for text-to-image and editing. Muse Video ranked third on the text-to-video leaderboard behind Google and ByteDance.

2 sources
LinkedIn Adds a 'Seems Like AI Slop' Button as Study Flags 40% of Long Posts as AI-Generated

Tech

LinkedIn Adds a 'Seems Like AI Slop' Button as Study Flags 40% of Long Posts as AI-Generated

Flagged posts get reduced reach and a private authenticity notice to the author. A Pangram analysis found nearly two-thirds of AI-generated long-form social posts are on LinkedIn.

2 sources
Meta's Segment Anything and DINO Models Anchor First Genesis Mission Science Projects

Tech

Meta's Segment Anything and DINO Models Anchor First Genesis Mission Science Projects

Lawrence Berkeley National Laboratory is applying Meta's open-vocabulary perception models to scientific imagery under the US federal Genesis Mission.

1 source
Pittsburgh Uses Meta's Perception Models to Build Open-Vocabulary Assistive Robots

Tech

Pittsburgh Uses Meta's Perception Models to Build Open-Vocabulary Assistive Robots

The University of Pittsburgh is pairing Segment Anything and DINO with assistive robotics so systems can recognize and manipulate arbitrary objects for users with disabilities.

1 source