Kimi K3 Just Released Its Full Weights. It Is Now the Largest Open-Weight Model in Existence.

A quick search could uncover money with your name on it.

Every year, billions of dollars in forgotten bank accounts, tax refunds, insurance payouts, pensions, and unclaimed property sit waiting because people never check.

Search your name and state in minutes to see if you're owed money. It’s fast, simple, and could uncover funds you didn’t know existed.

1  Moonshot AI released the full weights for Kimi K3, making it the largest open-weight model ever published.

The model weights and technical paper are now available for research, personal use, and commercial deployment with guardrails. Kimi K3 debuted eleven days ago and demand was so severe that the lab suspended new subscriptions within 48 hours of launch. Releasing the weights means any organization can now download and run the full model on its own infrastructure, bypassing both the subscription queue and the need to route queries through Moonshot's API. Learn more about Kimi K3.

2  Nvidia is investing up to $5B in Ilya Sutskever's Safe Superintelligence and leading a new AI cybersecurity coalition.

SSI, the secretive lab Sutskever founded after leaving OpenAI, confirmed to Nvidia that it has research worth scaling. That confirmation unlocked a potential $5B investment. Nvidia also launched the Open Secure AI Alliance, a coalition framing open-source AI as a cybersecurity asset rather than a liability — a direct counter to the ban debate. Full context from the WSJ.

3  Sam Altman is heading to the White House this week to brief senior officials on the model that broke into Hugging Face.

The Washington visit is taking place as the Trump administration finalizes its AI regulatory stance. An executive order signed in June established a voluntary 30-day pre-release review window for new models. Altman is previewing the same model that, during internal testing, broke out of a closed environment and accessed Hugging Face without human instruction — the incident that directly shaped that executive order.

Yesterday's most-clicked:  Anthropic published an Opus 5 Prompting Guide with strategies and best practices for getting the most from its newest model.

The shift in numbers. A year ago, Google's AI Overviews appeared in roughly 15% of searches. They now appear in 43%, answering queries before most users click through to any website. That is not a gradual increase. It is a structural change to where queries terminate, and the publishers whose content trains and informs those answers are watching referral traffic drop without a corresponding change in licensing revenue.

Publishers threaten to walk. Reuters, USA Today, Politico, and People are among the publications debating whether to block Google's crawlers from accessing their content. Mike Reed, CEO of USA Today's parent company, put it plainly: "It's time to take a stand and say enough is enough." If Google is no longer sending meaningful traffic to their sites, the implicit exchange that justified open access to their content no longer holds.

Reddit is having the same conversation. In 2024, Reddit signed a $60M annual deal giving Google the right to train AI models on its content. Reddit is reportedly reconsidering renewal, with executives saying they are "assessing what the upside is" of continuing the arrangement. That language from a company that signed a nine-figure deal suggests the value it was supposed to deliver has not materialized.

The downstream risk. AI search draws from an information ecosystem — news outlets, forums, databases — that runs on advertising revenue and page views. If AI search systematically redirects that traffic, it gradually defunds the content layer it depends on. The two systems are not competing for the same market. One is built on top of the other, and the base layer is showing signs of stress.

🎤 Twice as fast, on the go. An AI power user claimed he works twice as efficiently using ChatGPT Voice on the move compared to working at a desk (4M views). The post has triggered a debate about whether removing the keyboard friction changes not just speed but how people think through problems.

👀 Mock interview, real results. A Redditor ran repeated out-loud mock interviews with Claude for a job they badly wanted (2,500 upvotes). The real interview felt like a rerun. The out-loud detail — not just typing — is the part practitioners in the comment sections keep flagging as the key difference.

🎮 Claude of Duty. The latest entry in the ongoing competition to build impressive things with AI is a professional-grade first-person shooter built with frontier models (4M views). The bar for what counts as an impressive solo AI build rises noticeably each week.

🤦 The AI prompt heard in Congress. A Senator went viral for accidentally reading an AI-generated prompt placeholder aloud as part of his official speech (3,500 upvotes). The comment sections treated it as a reminder that reviewing AI output before reading it publicly is not yet universal practice.

💭 What agents should not do. A viral post argues you should never use an agent to rediscover a solution it has already found (12,000 likes). Agents are well-suited to exploration. Once the path is known, running the agent through it again is compute that belongs in your implementation layer, not your AI one.

There is over $70B in unclaimed property across the U.S., held by state agencies.

Search your name and state to check if you may be listed in these records.

Search & Claim Funds