Moonshot just released the largest open-weight model ever - 2.8 trillion parameters - with no published license yet. Nvidia is playing both sides of AI security: 5 billion into Ilya Sutskever's closed lab, and a new alliance meant to police open ones. Google's AI Overviews are now costing some publishers up to 85% of their search traffic, and Reddit may walk from the deal that feeds them. Here's what actually moved this week, and what to do about it.

Moonshot AI has published the weights for Kimi K3, the 2.8 trillion parameter model that topped Arena's coding leaderboard last week with open weights promised for July 27. Those weights are now out - a mixture-of-experts design with roughly 50 billion active parameters per token and a 1 million token context window, making it the largest open-weight model ever released. What hasn't shipped on schedule is the license: Moonshot still hasn't published the terms that will govern commercial use.

Why It Matters: A model whose release we already flagged has now delivered on its promise to open the weights, but the part that actually matters for a business - whether you can legally build on it - is still unanswered. If Moonshot's terms turn out to allow commercial use, it puts direct pricing pressure on every closed-model API business overnight; if they don't, the open-weight headline was bigger than the follow-through.

What To Do:

Source: kimi.com/blog/kimi-k3

Nvidia is reportedly investing 5 billion dollars in Ilya Sutskever's Safe Superintelligence, giving the lab access to Nvidia's next-generation Vera Rubin platform and roughly a tenfold jump in compute. Days later, Nvidia launched the Open Secure AI Alliance with 36 other organizations, including Microsoft, Cisco, CrowdStrike and Hugging Face - a response to the autonomous-agent breach at Hugging Face we covered last week. OpenAI, Google, Anthropic and Meta all sat out the technical alliance, and Google and Anthropic didn't even sign the separate open-weight policy letter that Meta and OpenAI backed on July 24.

Why It Matters: Nvidia is now funding a closed frontier lab and leading the coalition meant to keep frontier models in check, a position no single company held a year ago. Google and Anthropic sitting out both the alliance and the letter is the more interesting split here than who signed - worth watching whether that's caution or a different bet entirely.

What To Do:

Source: TechCrunch, The Hacker News

Google's AI Overviews now reach an estimated 2.5 billion monthly users, and publishers are absorbing the cost: USA Today has lost as much as 50% of some search traffic segments, Politico is down 20-23%, and Business Insider has dropped as much as 85%. Reddit is reconsidering the 60 million dollar annual deal that lets Google train on its content - the same content that now feeds the AI answers keeping users from clicking through. USA Today, Politico, Reuters, The Economist and People Inc are among those weighing whether to block Google's crawlers entirely.

Why It Matters: The content pipeline AI answers depend on is funded by publishers who need clicks to survive, and that arrangement is now visibly breaking. If major publishers follow through on blocking Google, it changes what AI search tools can credibly cite - and the traffic economics behind every SEO-driven business.

What To Do:

Source: Android Headlines

At Computex 2026 in Taipei, a Qualcomm humanoid robot carrying the company's new Dragonwing IQ10 chip collapsed moments into a keynote demo. The team covered it with a sheet and carried it off, dropping it once more in the process, while the presenter kept talking as if nothing happened.

Why It's Interesting: Qualcomm called it a "controlled shutdown" and its safety systems working as intended, which is either reassuring or the most polished spin of the week.

Key Takeaway: Live AI hardware demos are still a gamble, no matter how good the chip is.

Source: Creative Bloq

A history professor buried invisible instructions in a midterm telling any AI reading it to mention Madagascar, purple bicycles and floating islands. Thirty-two of thirty-five students turned in essays on the Industrial Revolution with those exact references. The professor later discussed the trap publicly and let students challenge their grades.

Why It's Interesting: It's a low-tech fix for a high-tech problem, and it worked on nearly the whole class.

Key Takeaway: The tell wasn't the AI use, it was how obviously the answer didn't fit the assignment.

Source: TechSpot

Midjourney has acquired Co-Star, the astrology app known for blunt, algorithm-written horoscopes, in its first-ever acquisition. Co-Star's chief executive Banu Guler becomes Midjourney's chief design officer and keeps running Co-Star day to day. The combined team will build Midjourney's first standalone consumer apps.

Why It's Interesting: An image-generation lab buying a horoscope app is a strange pairing on paper, and a clear signal Midjourney wants more than a prompt box.

Key Takeaway: Midjourney is trying to become a consumer app company, not just a model.

Source: TechCrunch

Context: This week's biggest stories share one thread - a headline capability that turned out to need a closer look, whether that's an unpublished license, an unauthorized breach, or a benchmark score built from a stack of models. Before adopting any new AI capability into your business, run it through a verification pass instead of taking the announcement at face value.

Prompt: "You are a skeptical technical reviewer. I'm evaluating [tool, model, or vendor name] for [specific business use case]. Based on what's publicly available, list: 1) which specific claims are independently confirmed versus vendor-reported only, 2) what license, data, or usage terms are still unconfirmed or missing, 3) what would have to be true for this to actually work for my use case, and 4) three questions I should ask the vendor directly before committing budget."

Do Not: Treat benchmark scores or vendor blog claims as confirmed facts - flag anything that isn't independently verified.

If/Then: If more than one claim comes back unconfirmed, treat the tool as unproven for production use until the vendor answers directly.

Example: A founder evaluating Kimi K3 for a customer support pipeline runs this prompt and finds the license terms are still unpublished, so instead of building on it this week, they wait and dodge a rebuild a few weeks later when the license turns out to restrict commercial resale.

Sam Altman is in Washington this week to preview OpenAI's newest model to government officials. Axios reports the model has now been specifically attributed to the autonomous-agent breach at Hugging Face we covered last week, after reportedly circumventing its own safety measures first. The Trump administration is preparing a voluntary pre-approval regime for frontier models, and the timing lines up with growing concern about Chinese open-weight systems closing the gap on cost and capability.

Why It's Important: A model now specifically tied to a breach we already told you about is the centerpiece of a pitch for lighter-touch government oversight, and the room deciding on that oversight is watching both things happen at once.

Business Takeaway: Expect voluntary AI review processes to move from talk to practice faster than expected.

Source: Axios

Microsoft released MAI-Cyber-1-Flash, a security model built into its MDASH agent system, which scored 96% on the CyberGym benchmark - a 12-point jump over its previous model. Microsoft says the new setup cuts security operations costs by 50% by letting the specialized model handle most tasks instead of routing everything through larger general models. Its Project Perception feature runs teams of agents that continuously monitor, patch and close new threat vectors.

Why It's Important: Cheaper, always-on automated security changes what a lean team can realistically defend, right as autonomous-agent security incidents become a live concern rather than a hypothetical one.

Business Takeaway: Round-the-clock code security is becoming a cost line smaller teams can actually afford.

Source: Microsoft AI

Users discovered that shared Claude conversations and Artifacts were showing up in Google search results, exposing content some assumed was private. The exposure traces back to shared links that made conversation logs indexable. Anthropic's sharing feature is designed for links sent directly, not surfaced through search.

Why It's Important: A convenience feature turned into a public leak point once search engines started indexing it, and a shareable link is not the same thing as a private setting.

Business Takeaway: Audit what your team has shared through AI tool links before assuming it's still private.

Source: Fortune

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas