An AI Solved a 27-Year-Old Math Mystery for $2,000. It's Not for Sale.

Today's stories are a study in contrasts. OpenAI's next model casually solved math problems that stumped humans for decades, and in the same week Google had to yank an AI tool off Google Earth because people used it to fake photos of hospitals and refugee camps within 48 hours. Reddit's stock took a hit tied directly to how AI is rerouting web traffic, and Apple is already talking about charging you more for Siri. None of this is hype. It's just Tuesday in AI right now.

Wispr Flow turns your voice into clean, professional text you can send the moment you stop talking. Not rough transcription you have to clean up. Actual polished text — ready for email, Slack, or any app.

Speak the way you think. Go on tangents. Change your mind mid-sentence. Flow strips the filler, fixes the grammar, and gives you text that reads like you spent five minutes writing it.

89% of messages sent with zero edits. Millions of professionals use Flow daily, including teams at OpenAI, Vercel, and Clay. Works on Mac, Windows, and iPhone.

Start flowing free

OpenAI revealed that an internal, unreleased model called Astra solved ten long-standing open problems in math and theoretical computer science, according to OpenAI's own announcement. The results include proving that "non-sofic groups" exist, a question open since 1999, along with disproving a conjecture by mathematician Alain Connes and clearing three problems from Paul Erdős's list. Every proof was checked using a formal verification tool called Lean, and OpenAI says the whole run cost roughly $2,000 in computing. Astra itself has no release date, no price, and no public access, not even a name for whatever product it eventually becomes.

Why it matters to you: This isn't a product you can buy or a feature landing in your ChatGPT app next week. It's a research flex that signals OpenAI's next model is a real leap forward, and when it does ship, expect the marketing to lean hard on "solved decades-old math problems," even though that has nothing to do with writing your emails or scheduling your staff.

What to do about it: Don't let this headline talk you into upgrading anything today. Wait until Astra actually ships as a product you can test.

Reddit's stock fell sharply after an earnings report in which CEO Steve Huffman openly questioned the value of Google's AI Overviews, according to CNBC. Huffman said the old "10 blue links" model of search sent real value to sites like Reddit, and that AI Overviews "has yet to make a similar level of positive impact." Reddit is reportedly weighing whether to end a $60 million licensing deal with Google over the issue, and a Pew Research study found AI Overviews have cut referral clicks to sites like Reddit by roughly half.

Why it matters to you: If people search for your business or your content and Google just answers the question on the results page instead of sending them to your website, you lose the click, and Reddit's own numbers show that drop can run as high as 50 percent.

What to do about it: Check your Google Search Console traffic trend over the last few months. If clicks are sliding while impressions hold steady, AI Overviews are likely the reason, and it's worth building traffic sources you don't depend on Google to hand you, like email or a real community.

Google added its Nano Banana 2 image model into Google Earth, letting anyone edit real satellite photos with a text prompt, according to The Decoder. Within two days, users had generated convincing fake scenes, including a staged refugee column at the Mexican border and a fabricated bombed-out hospital in Gaza. Google pulled the feature entirely, saying it won't return "until stronger safeguards are in place," and admitted that people "uniquely trust Google Earth for a reliable view of the world."

Why it matters to you: If Google, with all its safety teams, can't keep an AI image tool from being weaponized for two days, assume any AI image or video generator you use for marketing carries the same risk of misuse, whether by you accidentally or by someone impersonating your brand.

What to do about it: Label AI-generated images clearly in your own marketing, and don't assume a platform's safety filters will catch a determined bad actor before real damage is done.

Alibaba released Qwen3.8-Max, a 2.4 trillion parameter model built for long, multi-day tasks like software projects and research instead of just quick answers, according to The Decoder. The model is live now through Alibaba's cloud platform, with the full open weights set to release for free on Hugging Face next week. Alibaba says it ranks just behind Anthropic's newest Claude model on capability, at a fraction of the price during its current promotional window.

Why it matters to you: Every time a lab gives away a model this capable for free, it resets what you should expect to pay for AI tools that are good enough for your daily work, from drafting emails to summarizing customer feedback.

What to do about it: If you're paying premium prices for a task that doesn't need frontier-level intelligence, like basic writing or research, check whether a cheaper or open model already does the job.

OpenAI slashed prices on its GPT-5.6 Luna model by 80 percent and cut its Terra model by 20 percent, according to VentureBeat. Luna's input price dropped from $1.00 to $0.20 per million tokens, driven by what OpenAI calls efficiency gains in its infrastructure, along with competitive pressure from Google's cheaper Gemini models. It's the clearest sign yet that AI companies are shifting from a quality race to a price war.

Why it matters to you: If you or your team pay for AI tools by usage, this kind of price war usually shows up in your bill within weeks, either as a direct price drop or as a new cheaper tier you weren't offered before.

What to do about it: If you're on a metered AI plan, check your provider's pricing page this month. You may already be overpaying for compute that just got cheaper.

In his final earnings call as Apple CEO, Tim Cook said the company is considering a paid tier for heavy users of its upgraded, AI-powered Siri, according to TechCrunch. Cook said any upgrade would likely run through iCloud+, Apple's existing subscription service, calling it a way for people to "buy up the stack." The improved Siri is currently in the iOS 27 beta, with a wider rollout expected this fall, and Cook stressed the pricing plans are still not finalized.

Why it matters to you: Free AI is quietly becoming a thing of the past across every major platform, including the assistant already built into your customers' phones, so budget for the possibility that tools you use for free today may carry a subscription tomorrow.

UNAM, Mexico's most prestigious university, cancelled roughly 3,000 scores from its first-ever online entrance exam after results showed an unusual spike in near-perfect scores, according to CNN. About 158,000 people took the exam, and historically only about 3.5 percent scored 100 or more correct answers; this year that rate jumped high enough to trigger a formal investigation. The university suspended new admissions while it investigates a mix of suspected AI-assisted cheating and leaked questions, and the scandal has drawn public comment from Mexico's president, herself a UNAM graduate.

Why it matters to you: If your business relies on any kind of online testing, certification, or screening, whether it's a hiring assessment or a customer quiz, assume AI-assisted cheating is already happening at a scale your current safeguards probably don't catch.

What to do about it: If you run any high-stakes online test for hiring or certification, add a human review step for unusual score patterns instead of trusting automated proctoring software alone.

OpenAI is shutting down ChatGPT Atlas, its standalone AI browser, on August 9, less than ten months after launch, according to OpenAI's own help center. Browser-based agent features are moving into the ChatGPT desktop app instead, but bookmarks, open tabs, and browsing history do not transfer automatically. Users have to manually export bookmarks as an HTML file and move cookies and passwords themselves before the shutdown, or lose them for good.

Why it matters to you: A tool you built workflows around, or handed real business data to, can be discontinued with about a week's notice and no guarantee your information comes with you. That's true of Atlas today and could be true of any AI tool you're relying on.

What to do about it: Before you build a workflow deeply into any single AI tool, know how you'd export your data if that vendor pulled the plug next month, because based on today's stories, they might.

Every story today points at the same seesaw. AI is getting good enough to solve problems humans couldn't touch for decades, and reckless enough to get pulled off a platform within 48 hours of launch. It's cheap enough that prices are falling across the board, and expensive enough that Apple and Google are both eyeing your wallet for the next tier up. None of that is a reason to panic, and none of it is a reason to assume any of these tools are stable ground to build on. Get the value where you can find it today, keep a human checking the output, and don't get too attached to any single vendor's roadmap. That's the whole ride.

If today's issue saved you from finding any of this out the hard way, forward it to one person who needs to see it. And if someone forwarded this to you, subscribe to get it every day.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social