An Unreleased OpenAI Model Solved 10 Decade-Old Math Problems for About $2,000 in Tokens.

Smarter AI. Lower costs.

Most companies focus on choosing cheaper AI models when the real savings come from eliminating unnecessary requests and optimizing how AI workloads are routed.

Mesh API gives your team complete visibility into AI usage, helping you reduce token costs, improve efficiency, and scale AI without overspending.

1  An unreleased OpenAI model solved 10 open mathematical and computational problems, several untouched for over a decade, for roughly $2,000 in compute.

OpenAI's internal Astra model solved 10 open problems spanning mathematics, quantum complexity, and theoretical computer science, documented in a 249-page paper. None of the problems had seen new progress on their main results in at least ten years. The total estimated token cost to produce all ten proofs was around $2,000. Sam Altman was reportedly in Washington, DC last week demoing Astra to federal officials, which puts the model's public unveiling on a very different timeline than its private one.

2  Google's Gemini Spark can now browse the web with your saved passwords and logged-in accounts to complete tasks on your behalf.

With permission, Gemini Spark can complete online tasks like researching flights or hunting for apartments, using your existing logins rather than requiring you to re-authenticate at each step. See how auto browse works. Separately, Google now lets you summon Gemini at any time by pressing the FN key in the macOS Gemini app.

3  A Stanford-built startup raised $200M to simulate the decisions of all eight billion people on Earth.

Simile launched five months ago and has now closed a Series B aimed at accurately predicting human decisions at scale. Built by Stanford researchers on their own academic work, the startup has released two models: a foundation model trained on real people, and a confidence model that estimates how accurate each prediction is likely to be. See how it works.

The cash flow shock. We are watching something genuinely unusual happen. Amazon and Alphabet, two of the most profitable companies ever built, reported negative free cash flow last quarter, citing the hundreds of billions being funneled into AI infrastructure. This chart puts the scale in perspective. Meta was not far behind, with cash flow down 91%. Companies whose entire brand is generating cash are, for the moment, doing the opposite.

Wall Street shrugged, mostly. The declining cash flow did not spook investors, largely because revenue is still strong. The entire Magnificent 7 posted double-digit annual revenue growth, except Nvidia, which reports in a few weeks. The tech-heavy Nasdaq rose roughly 2.5% on the week, reversing a downward trend that started in June. Microsoft was the standout: the stock soared 19% and added a single-day record of $450B in market cap, while Meta and Apple both declined substantially. The market is not treating this quarter as one story. It is treating it as several different stories that happen to share a sector.

RAMageddon, still the loudest topic. The global memory shortage, nicknamed RAMageddon, kept driving costs for both Big Tech and consumers this quarter. Elon Musk called current memory pricing "pretty insane." Tim Cook described it as a "hundred-year flood." It has already forced price increases across Macs and iPads, and analysts expect further increases are still coming. Two CEOs from two companies that rarely agree on public messaging are describing the same problem in almost identical terms.

What it adds up to. Negative cash flow at two of the most profitable companies in history, a 91% cash flow drop at a third, and a memory shortage severe enough that competitors are using the same disaster metaphors independently. And the market went up anyway. That is either a sign of extraordinary confidence in where AI spending leads, or a sign that the market has not yet fully priced in what happens if it doesn't. Both readings are live right now, and this earnings season did not resolve which one is correct.

👤 TIME headshots. Is your LinkedIn photo due for an update? This viral prompt transforms regular selfies into editorial-style headshots you would actually post, modeled on Marco Grob's TIME cover portraits (2,000 upvotes).

👀 Sam's suggestion. Sam Altman accidentally drew the internet's ire on Friday when he posted a "cool use case" for ChatGPT. Many people did not think it was cool and flooded the post with sarcastic comments (12M views). Even the most-followed person in AI is not immune to a reply-guy pile-on.

🚨 Be aware. Do not believe any satellite images you saw over the weekend. Google Earth got a new AI image-generation feature that was abused so quickly the company had to revoke it within a day. Read about the short-lived update. A useful reminder that AI-generated geographic misinformation is not a hypothetical concern.

📝 Leopold's letter. Leopold Aschenbrenner's AI fund just had its roughest month since launching. He sent an investor update that is being applauded for its calmness under pressure, along with a mild humblebrag about the fund's year-to-date performance (3M views). Notably paired with the news elsewhere this week that he sold the fund's public portfolio to Citadel after a sharp drawdown.

⚔ LOTR movie. Andrej Karpathy created an animated short based on the opening of The Lord of the Rings while testing Claude Opus 5. He admits it is a bit janky, but says it offers a glimpse of how AI has made fun projects like this feasible for one person working alone. Watch it here (2.5M views).

Draft a feedback conversation that is honest without being unkind

Most feedback conversations either soften the message so much the real point gets lost, or land so bluntly that the other person stops listening halfway through. This ChatGPT prompt builds a structured feedback script that stays specific and behavior-based, opens dialogue instead of shutting it down, and ends with a concrete next step. Fill in six context fields and it drafts the actual conversation, not just advice about how to have one.

For [BEHAVIOR], describe the actual observable action, not your interpretation of it: "missed the last three deadline check-ins without a heads-up" works better than "seems disengaged." For [IMPACT], name the concrete consequence: what it cost, delayed, or put at risk. The prompt explicitly avoids the praise-sandwich pattern most feedback advice defaults to, which tends to bury the actual message between two compliments the recipient forgets by the time the conversation ends. Run this before a conversation you have been putting off, not as a script to read verbatim, but as a way to find the version of what you want to say that is both clear and fair.

One team cut AI spend by 78% without switching models. They switched gateways. Mesh API: 1000+ models, one endpoint, cheapest routing, per-team token tracking. Estimate Savings.