← Back to newsletters

AI Daily Brief — Mon Aug 31

2026-08-31

By Vadym · Generated with AI, curated by me


Listen to this issue

Today is about who owns what — OpenAI is cutting Cursor loose over who now controls it, two music publishers are suing Anthropic over what Claude was trained on, and DeepSeek is raising a war chest ahead of a public listing — while the technical side just keeps shipping: alignment fixes that generalize, a 753B open model punching above its weight, and a $399 robot you train yourself.


Headlines & News
Industry

OpenAI Cuts Off Cursor’s Model Access After SpaceX’s $60B Acquisition

OpenAI notified SpaceX on August 28 that it will terminate Cursor’s access to its models on November 12, 2026, following SpaceX’s roughly $60 billion acquisition of the AI coding startup. OpenAI said it isn’t confident SpaceX will honor its usage policies, pointing to past terms-of-service violations by Twitter and xAI, both Musk-owned companies.

Source: CNBC

Funding

DeepSeek Nears a $7.4B Round at a $74B Valuation Ahead of a Shanghai Listing

DeepSeek is reportedly closing a roughly 50 billion yuan ($7.4B) round that would value the company at $74 billion pre-money, adding investors including CPE and Legend Capital on top of its $50B-valuation round from June. Proceeds are earmarked for model R&D and compute, with the company said to be preparing a 2027 IPO on Shanghai’s STAR Market.

Source: China Money Network

Legal

Sony Music and Warner Chappell Sue Anthropic Over Claude’s Training Data

Sony Music Publishing and Warner Chappell filed suit against Anthropic in California, accusing it of a “brazen campaign” of torrenting and scraping tens of thousands of copyrighted songs to train Claude, and naming co-founders Dario Amodei and Benjamin Mann personally. The publishers are seeking up to $150,000 per work, which could push total damages into the billions.

Source: TechCrunch

Research

Anthropic’s Automated Researchers Beat Humans at Fixing Alignment Bugs

Anthropic had Claude autonomously discover fixes for 10 categories of alignment failure, and its best method beat the top submission from 28 human safety researchers by 20% on a deception benchmark. The fixes held up on models up to 4.7 times larger than the ones Claude was optimizing against, meaning the techniques generalized rather than overfitting to one model.

Source: Anthropic

Models

Z.ai Open-Sources GLM-5.3, a 753B Coding Model That Punches Above Its Size

Z.ai published open weights for GLM-5.3 on Hugging Face, two weeks after the model launched behind its paid coding plan. The roughly 753B-parameter, 40B-active mixture-of-experts model is now Z.ai’s most capable release for agentic coding, and on several coding benchmarks it reportedly matches or beats models several times its size.

Source: Hugging Face

Robotics

Hugging Face and Pollen Robotics Ship a $399 Biped You Train Yourself

Microduck is a 25cm, 800g biped with 15 motors, a camera, LiDAR, and two IMUs, running onboard policies at 50Hz. It ships with seven pre-trained behaviors — walking, sitting, kicking, grabbing, roller-skating, recovery — but the point is retraining it yourself in simulation with reinforcement learning and deploying new policies to the hardware, all on an Apache 2.0 software stack.

Source: Pollen Robotics


Analysis

Takeaway

Split the day in two and a pattern shows up: the money-and-ownership fights — OpenAI severing Cursor over who bought it, publishers suing Anthropic over what it trained on, DeepSeek raising ahead of an IPO — are all about control over something that already exists. The technical side moved on regardless, shipping an alignment method that generalizes, a smaller open model that competes with bigger ones, and a $399 robot meant to be retrained, not just admired. The lawyers and the researchers are working two different clocks.

— Boba


Curated by Vadym