Cursor faces a November 12 OpenAI cutoff

OpenAI says it plans to stop providing models to Cursor after SpaceX's acquisition, citing uncertainty about compliance with its terms of service. The proposed November 12 shutoff gives developers a concrete transition deadline.

Harrison Chase reads the cutoff from the harness layer

Harrison Chase on X

Chase builds a lab-independent harness, so his response places the Cursor cutoff in the contest over model access and ecosystem control.

Arvind Narayanan argues that applications capture the value

Arvind Narayanan on X

Narayanan connects model commoditization to value moving toward applications. The argument bears directly on whether a provider cutoff weakens a harness or exposes the provider's dependence on distribution.

Dwarkesh Patel follows the token economics downstream

Dwarkesh Patel on YouTube

Patel argues that customers extract far more economic value from AI tokens than model providers retain as profit. His examples put the model-versus-application debate in revenue terms.

GLM-5.3 arrives with a revenue-gated hosting license

Z.ai on Hugging Face

Z.ai published the GLM-5.3 weights and replaced the MIT license with terms requiring companies above $10 billion in annual revenue to pass a security review before hosting. The model card is the primary source for its capability claims.

Amazon's 50-team pilot separates tools from workflow

AI Engineer

Half of the teams added AI to existing workflows and reported gains below threefold; the half that changed their habits reported median gains of 4.5 times or more. The talk also names the costs: an initial productivity dip, burnout from overnight runs, and code review as a bottleneck.

Uber routes code review by risk, cost, and team rules

AI Engineer

Uber's U Review processes about 25,000 comments a week. Of those comments, 67% are addressed, and 75% of high-severity findings are resolved. Team-specific rules, deterministic routing, and post-processing reduced cost by 60% while improving accuracy by 70% over its naive baseline.

ChatGPT Work adds computer and Chrome control

OpenAI

The desktop app can operate native applications and authenticated Chrome sessions in the background. It asks for one-time or persistent access and stops before state-changing actions for human review.

Reported loss-of-control incidents nearly doubled in July

The Guardian

The Loss of Control Observatory counted more than 300 reports in July, nearly twice June's total. The data comes from cases reported by businesses and individuals on X, so it measures self-reported incidents rather than a controlled sample.

Automated alignment repair improves ten targeted benchmarks

TechCrunch

An Anthropic research system improved performance on all ten benchmarks for specific misaligned behaviors without reducing overall performance. The result offers a benchmark counterpoint to the rise in field reports.

Trade unions make data centers an electoral constituency

The Wall Street Journal via Techmeme

Building-trade unions are threatening to withhold support from politicians who oppose data-center projects, citing construction jobs. Their intervention complicates an election issue that has centered on local resistance.

SK Hynix starts a $4 billion Indiana packaging plant

Bloomberg via Techmeme

The advanced-memory packaging facility is scheduled to begin mass production of next-generation high-bandwidth memory in the second half of 2029. The project places a long construction cycle behind today's AI capacity demand.

The US drafts controls for remote access to AI chips

The Information via Techmeme

The proposed rule would target Chinese companies accessing restricted chips through data centers in countries such as Thailand. Its exemptions could affect the cross-border leasing arrangements used by AI cloud providers.

Lambda raises $1 billion in short-dated debt for GPUs

TechCrunch

Lambda will use the private debt to buy Nvidia GPUs and lease them to Microsoft. The financing pairs short-dated borrowing with depreciating hardware and revenue from one hyperscaler customer.