This Startup Wants to Sell You Everyday Goods at Cost for $99 a Year

1  Aonic Club uses AI to sell goods at close to what they cost to make

With US inflation at 3.7% in September, one of the longest high-price stretches in a century, Aonic Club is betting people will pay a $99 yearly membership to buy its own-brand goods at roughly manufacturing and shipping cost. The startup says that works out 20% to 80% cheaper than name brands. For now, it sells only protein and supplements. See founder Max Meier's launch thread.

2  About 400 scientists say voluntary AI pledges are not enough

Researchers from 34 countries, including Geoffrey Hinton and Yoshua Bengio, signed an open letter calling for frontier models to be cut off from outside networks and placed under independent oversight. It pushes back on the voluntary safety commitments tech companies agreed to in Washington last month.

3  GPT-5.6 Sol finds a critical Chrome flaw in six days

OpenAI's Greg Brockman says it took six days and about $40,000 of compute with GPT-5.6 Sol to uncover a serious Chrome vulnerability that could have affected roughly 1B users. The security startup behind the find reportedly earned a $250K reward from Google, a sign that AI bug hunting can now pay for itself.

The premium version. Fidji Simo enrolled her daughter at Alpha School, where AI-led lessons take about two hours a day and students get one-on-one AI tutoring. Tuition runs from about $10K to $75K a year, and the school publishes data it says shows strong results.

The do-it-yourself version. You do not need private-school money to try the idea. Wes Pomeroy built AI study guides for his sons and wrote up how he did it. Noah Tye works math problems by hand, scans the pages with a document camera, and sends them to ChatGPT for feedback, so the thinking stays his.

The self-taught version. Deedy Das met a teenager who taught himself AI, landed an internship at a major lab, left university early, and now has several job offers from San Francisco AI companies.

The open question. Each example keeps a human in the loop: a parent setting goals, a student writing by hand, or a learner choosing what to study. The real test is whether AI tutoring helps children who do not already have that kind of support at home.

Ramp's Glass. Ramp is rolling out Glass, a company-wide AI workspace that remembers how the business runs. CEO Eric Glyman shared the numbers: 323 employees have made more than 6,500 edits, and captured feature requests jumped from about 300 to 20,000 a month. After a single $1.5M week of frontier-model spending, the company built a router that cut the bill by 30%. Outside companies can join through early access.

Stripe's Harbor. Stripe built Harbor, an AI app builder that lets non-engineers prototype their own tools. It took one engineer about six weeks.

Why build in-house. Gergely Orosz looked into the trend and found three reasons: companies keep control of their data, they can enforce their own design rules, and the cost is low. Engineers also get hands-on AI experience along the way. If your team is considering the same path, this practical guide to enterprise AI apps is a good starting point.

🏙 Notes From San Francisco. Greg Isenberg shared 21 takeaways from two days of meetings with AI founders and executives, from data brokers to SaaS valuations (11K bookmarks).

🚢 Build a Hull. Product Hunt CEO Rajiv Ayyangar argues that each new AI model either lifts a product or sinks it, and that only startups with a lasting advantage stay afloat. Read his full essay.

💻 Claude Code, Redesigned. The Claude Code desktop app now makes it easier to start several sessions at once, split work between them, and keep an eye on each.

✈ Fly Over Your Website. A free prompt turns a landing page into a 3D flyover video about 15 seconds long.

🔮 Software Is Not Dead. Kun Chen predicts a three-part future: well-built software keeps its value, most of it gets used through AI agents, and the screens people see become something they can reshape just by asking.

🚨 Stolen Login, Big Leak. Hackers reportedly used an employee's stolen credentials to get into a private Meta GitHub repository, taking source code and data on about 1,500 people.