OpenAI will test visual ads inside ChatGPT’s image-generation experience later this month in the U.S., starting with an initial group of advertisers. The ads will show product inspiration, usage, or related experiences, with clear labels and separation from generated images. OpenAI says advertising will not influence ChatGPT’s answers.
The company is adding measurement integrations with Hightouch, Tealium, and LiveRamp, plus attribution partners including AppsFlyer, Triple Whale, and Adjust. OpenAI cites partner-reported results including a 15.3% lower attributed CPA for WeightWatchers versus blended paid search and 93% new visitors for Portland Leather; neither figure has independent validation. DoubleVerify and Integral Ad Science will also pilot brand-suitability evaluations without accessing private conversations.
Why it matters: For advertisers such as WeightWatchers and Portland Leather, ChatGPT offers access to OpenAI’s 1.2 billion weekly users—but partner-reported performance data will need independent measurement before brands treat the channel as a dependable acquisition source.
President Donald Trump has announced a Super Intelligence Force chaired by national intelligence director Jay Clayton to coordinate the federal government's effort to keep the United States ahead in super intelligence.
FTC Chair Andrew Ferguson, Undersecretary of War for Research and Engineering Emil Michael, and Office of Personnel Management Director Scott Kupor will serve as vice chairs. The force has 120 days to report on AI's risks and opportunities.
Its charter calls for plans to respond to "SI-enabled threats" while preventing overregulation and regulatory capture that could stifle innovation and competition. The move follows Trump's September promise to create an AI Force and appoint an AI czar, along with an executive order rebranding AI as "super intelligence."
Why it matters: Clayton, Ferguson, Michael, and Kupor now have 120 days to define whether Trump's AI policy prioritizes moving faster than countries such as China or limiting rules they say could stifle innovation and competition.
AWS CEO Matt Garman said Amazon no longer uses nondisclosure agreements (NDAs) in dealings with government agencies seeking approval for new data centers. Environmental activist Erin Brockovich calls transparency the top complaint about these projects.
New York has imposed a one-year moratorium on permits for large data centers. Garman says more than 100 similar moratoriums are under consideration nationwide and warns they could damage U.S. AI ambitions.
To counter criticism, Garman says direct data-center water consumption equals 0.5% of U.S. industrial water use. An independent watchdog says data centers drove a 76% year-over-year electricity-price increase on America’s largest grid. A planned Amazon Texas data center is permitted to emit 33 million tons of carbon dioxide annually.
Why it matters: For Amazon, transparency is now a concession, not a permit: New York’s one-year freeze and more than 100 proposed moratoriums give local governments a concrete lever over new data-center approvals.
Google Cloud AI Research and several universities developed RRSI, a method that helps self-optimizing AI agents avoid memorizing their test tasks. RRSI regularizes the agent’s harness—the prompts, tools, memory, and workflows around a frozen model—by shrinking its edit budget and rejecting benchmark-specific hacks.
Across eight coding, office-work, and engineering-design benchmarks, RRSI improved unseen-task performance by up to 4.7 points while using about 30% fewer runtime tokens than an unregularized harness. It never fell below the baseline on unseen tasks, unlike two competing optimization methods. The study used Claude Opus 4.8 and did not cover agents whose model weights change.
Why it matters: Teams shipping agents should judge harness optimizers by transfer, not familiar-task scores: RRSI was the only tested variant to land well above baseline on unseen benchmarks, while two alternatives fell below it.
NASA and IBM released an open-source lunar foundation model that cuts polar ice-prediction error by up to 22%. Trained on nearly 2 million tile bundles spanning 11 data modalities, the model turns 17 years of Lunar Reconnaissance Orbiter observations—plus data from three other missions—into a reusable system for lunar science.
The model also beat SwinV2-B by nearly 19% on coarse-scale crater detection while using only half the data. But the gains are uneven: meter-scale crater detection and segmentation of Irregular Mare Patches roughly matched the strongest baselines, and the system is not reliable for absolute positioning. NASA and IBM released the model, datasets, and benchmarks through Hugging Face, GitHub, and TerraTorch.
Why it matters: NASA and IBM are making scarce lunar labels less of a bottleneck: one model can connect imagery, gravity, hydrogen, and mineralogy across missions, giving researchers a reusable starting point for studying polar ice and other lunar features without replacing physical measurement instruments.