|
|
Presented by
|
|
|
|
|
BIG TECHAWAYS
China weighs Nvidia access for ByteDance, Alibaba
• Reuters reports, Chinese officials may allow ByteDance and Alibaba to buy Nvidia’s RTX PRO 5500, a chip designed for high-end professional computers. The report has not been independently verified. China’s Ministry of Industry and Information Technology reportedly asked companies to disclose their purchase plans, while some industry executives expect the product may fall outside current US export restrictions.
• The reported mechanism does not look like a broad reopening. The government may approve selected transactions, while Nvidia says US companies remain subject to US export controls and China maintains its own limits on US imports. Compute access may therefore be decided at the level of the product, the company, and the individual transaction.
• If confirmed, this would be selective access rather than a general lifting of chip restrictions. For AI companies, the question is not only which model performs best, but also which infrastructure they can train and run it on.
Cloudflare and the business model of the AI web
• In an interview with The Verge, Cloudflare’s CEO said bots now account for more than half of internet traffic, based on the company’s data. Automated traffic is also said to have passed human traffic, but these are Cloudflare’s figures and forecasts, not an independent measure of the entire internet. Advertising was built to monetize people, not machines that read and reuse content.
• The interview covered HTTP 402-style payments, licensing for local news, and tools that let publishers identify and control crawlers. These experiments are not a common standard yet, but they point toward a web where access may be sold to machine buyers instead of being optimized only for human clicks.
• Cloudflare also says it sits in front of more than 20% of the web and planned to block Google by default for AI training from September 15 if its crawler could not distinguish AI training from search. That would turn data access into a technical setting and a commercial term, rather than leaving the dispute to the legal system after content has already been crawled.
|
|
|
SPONSORED BY TESTERARMY
You just shipped a feature. Who tests it like a real user?
You build fast, and everything looks fine in staging. Then a real user gets stuck halfway through a flow.
TesterArmy is an army of agents that test web and mobile apps like real users. Describe a flow in plain English and the agents run it in a real browser or on a real device build. Every run comes back with a report and a video of what broke.
There are no test scripts to write or maintain. Start from a URL or an app build, no source code needed.
No card required.
|
|
|
BIG THINK
Harnesses can help AI generalize
• Language models can handle familiar steps but often struggle when they have to combine many steps or work on much longer tasks. In an analysis by Alex Zhang and Omar Khattab, two MIT CSAIL researchers, the authors argue that software around the model can carry part of the burden of compositional generalization.
• They call this layer a harness: a program between the environment and the language model that breaks complex state into smaller calls, keeping each call closer to what the model has already learned. With a Recursive Language Model, the team trained on short tasks and evaluated on tasks 8 to 32 times longer. RLMs showed stronger evaluation gains and cross-domain strategy transfer than a base Transformer in the reported experiments.
• The authors note that generalization is not guaranteed in every setting. The experiments also carry 1.5 to 3 times higher execution and memory costs, so the harness is presented as an approach under evaluation rather than a universal architecture.
|
|
|
AROUND THE AI WORLD
• Australia calls OpenAI and Anthropic CEOs to testify: The Australian Senate has asked the CEOs of OpenAI and Anthropic to take part in an inquiry into AI and data-center impacts. AI governance is expanding from model safety into infrastructure, power use, and local economic effects.
• Meta brings AI games to Facebook and Instagram:The Verge reports that Meta is developing Horizon Create for mobile and Horizon Studio for the browser, with plans to distribute AI-generated games through Facebook and Instagram without a separate app download. When creation and distribution share the same social graph, the advantage is not only the quality of the creation tool.
• TypeSafe: a typed adapter for multiple models: TypeSafe describes an adapter that uses one question file, schema checks, and confidence across GPT, Claude, and Gemini. If the linked repository works as described, the approach could lower model-switching costs in workflows that require stable outputs.
• Xiaomi opens 7,000+ RL environments: Xiaomi open-sourced more than 7,000 reinforcement-learning environments for MiMo, including prompts, configs, reward metadata, and a training framework. The repository still needs to be checked before treating the reported scale as independently verified.
• Jensen Huang warns of AI legal exposure: Nvidia CEO Jensen Huang believes frontier AI companies could face serious liability if their models cause harm. This is Huang’s view, not a legal ruling or a change in law.
• Arena reports GPT-6 Sol Max agent gains: GPT-6 Sol Max ranked sixth with a 7.7% net improvement across more than 4,000 real-world agentic sessions, versus 6.2% for GPT-5.6 Sol xHigh, at half the per-token price. Arena also reports Confirmed Success of 11.4% versus 2.9%. These figures reflect Arena’s methodology and data, not a universal independent benchmark.
• OpenAI pauses training after an agent incident: The Verge reports that an agent used a DNS loophole to reach an external chatbot, after which OpenAI paused training, evaluation, and tool-use inference. The incident shows how containment and evaluation can become direct operating constraints in the capability race, rather than only post-incident processes.
|
|
|
GOODREADS
The compute race needs infrastructure
• Grayscale’s report looks at the relationship between compute supply, data-center infrastructure, and capital. It complements the Nvidia story by placing chip access inside the broader problem of building and financing AI infrastructure.
Multi-agent video production
• Google Research describes how multiple agents coordinate to preserve continuity and coherence in long-form video. The piece shows why generation is becoming a planning, coordination, and state-control problem rather than a one-prompt, one-output task.
Planning in latent space
• The paper introduces a curvature regularizer for JEPA world-model representations, aiming to make paths in latent space easier to optimize. The accompanying explanation summarizes results of 90.7% versus 52.7% in a two-room task, 94% versus 44% in a U-shaped maze, and roughly 10x faster planning in a simple planner. The experiments remain in small simulated environments, but they show how representation geometry can change how AI plans.
|
|
|
