This AI briefing for October 4 rounds up six developments published on October 3-4, 2026: NASA and IBM turned 17 years of lunar observations into an open foundation model, Microsoft and Hugging Face changed how agents are graded, AWS changed its secrecy practices amid the data center backlash, plus product moves from Capcom, Kling AI and Anthropic. Every item was checked against its original source and publication time.
NASA and IBM open-source a lunar foundation model: 17 years of observations, fed to machine learning at once
The Decoder reported on October 4 that NASA and IBM Research, working with several academic institutions, have released the NASA-IBM Lunar Foundation Model, described by the team as one of the first open-source foundation models for lunar science. Its training corpus, SomBench, is billed as the largest co-registered multimodal lunar dataset to date: nearly 2 million tile bundles across 11 modalities and two spatial scales, including about 1 million high-resolution images from the Lunar Reconnaissance Orbiter's Narrow Angle Camera at roughly 1 meter per pixel, drawn from the orbiter's 17 years of observations. The model, based on the TerraMind architecture but trained from scratch, receives imaging geometry such as illumination angles and sun position as explicit input. On polar ice deposit prediction it cut error by up to 22 percent versus the best baseline, SwinV2-B; on coarse-scale crater detection it led by nearly 19 percent while training on half the data. The limits are stated plainly: in generation tests, latitude and longitude were off by dozens of degrees, so it is not a tool for absolute positioning. The model is on Hugging Face, with code and the companion datasets released alongside.
Microsoft's ThinkingBox is live: the agent says it is done, the database decides
Microsoft and Hugging Face published ThinkingBox on the Hugging Face blog on October 3: 507 stateful business workflows, each task run 20 times per model, graded not on the agent's final message but on the terminal state and side effects it leaves in the backend database. It is now runnable through OpenEnv on Hugging Face. The first results are unflattering: across 121,680 valid trials from 12 models, 79,853 attempts failed the executable checks, and 67.24 percent of those failures still terminated cleanly, with well-formed tool calls and a confident "done." Claude Opus 5.5 posted the highest single-attempt pass rate at 67.16 percent. Kimi-K3 covered the most ground, solving 93.89 percent of tasks at least once, yet passed all 20 runs on only 13.41 percent. About four in five failures were tool-handling problems, not reasoning problems. Teams shipping enterprise agents should keep the yardstick: doing it once is not delivery; doing it twenty times in a row is.
AWS stops signing data center NDAs, adds a $1 billion community pledge
TechCrunch reported on October 3 that AWS CEO Matt Garman wrote in an official blog post that the company has stopped using nondisclosure agreements with government agencies when seeking approval for new data centers. Transparency had become residents' top complaint: projects surfaced only after permits were secured, with local officials bound by NDAs signed before neighbors knew anything. New York has announced a one-year pause on large data center permits, and Garman says more than 100 similar moratorium proposals are under consideration across the US. He also announced Built Together, a commitment of more than $1 billion over five years to host communities, funding free community college programs, workforce training, and energy-efficiency upgrades for schools and homes. AWS is the first hyperscaler to put "no more NDAs" in a public commitment; whether the others follow is the story of the next few months.
Capcom unveils the REX project: rebuilding the RE Engine so AI can help build games
The Verge reported on October 3 that Capcom presented the REX project at its CAPCOM Open Conference RE:2026 on October 2, in a talk by programmer Satoshi Ishida: a multi-stage overhaul of the in-house RE Engine whose long-term goal is an "AI-generation game engine" in which AI can understand engine code, help write it, and playtest builds to hunt bugs. The RE Engine has powered more than 27 titles since Resident Evil 7 in 2017 and is used daily by over 2,000 developers; the overhaul targets development efficiency, game performance, tooling, and quality assurance. Capcom also drew a line: AI-generated assets will not ship in finished games, and AI is for development efficiency only. Player backlash has already started, and the argument is less about the tools than about how far "creating games together with AI" is supposed to go.
Kling AI teases Kling 4.0: first stop is Advertising Week New York on October 6
Kling AI announced on October 4, via its official WeChat account and overseas channels, that it will appear at Advertising Week New York 2026 on October 6 in a panel titled "The New Production Engine: Powering Creativity at Scale with Kling AI," alongside WPP's Director of AI Strategy and Innovation Mathieu Albrand and Adobe's enterprise partnerships lead Elissa Levine, to discuss how AI video enters enterprise workflows and supports content production at scale. Kling 4.0 is explicitly billed as coming soon, but no date or specifications were given on October 4. For brands and agencies, the signal lands before the specs: ahead of its next model, Kling is competing for enterprise mindshare first.
Anthropic starts asking for voice data: opt in, and only then is it used for training
On October 4, Testing Catalog, which tracks unreleased features, surfaced a new in-app prompt in Claude: Anthropic has begun asking users whether they will voluntarily share their voice conversations to improve its models. The prompt says, in effect, that audio recordings and voice chat data help improve how Anthropic's AI models understand and respond to speech, and that users can turn the setting off at any time or delete the data. This is per-user consent, not default collection. As voice mode becomes standard across assistants, what data may be used to train it is becoming a question that has to be asked out loud. If you see the prompt, the choice is yours either way.
The date to watch this week is October 6, when Kling takes the stage in New York: if Kling 4.0 gets a release date on the spot, it will set the tempo for this round of AI video.
