Hey Snackers,
It’s Friday, the outro of the week. A new analysis of thousands of albums looked just at the last songs, finding that they tended to be a little bit longer and a bit quieter than the rest of the record. Early on that was a necessity, because vinyl records have less range to work with the deeper you get on an album, but now artists just like to end the record on a weird, acoustic one.
May your Friday be relatively chill and feel a little longer than your typical day as you ride on down that moonlight mile.
Stocks dropped, bond yields moved higher, and oil prices rose as the US Treasury Department’s bond-buyback plan underwhelmed investors and President Trump said he did not expect the US-Iran war to end until after the November midterm elections.
OpenAI and Alphabet have very, very, very different ideas on what good P.R. for AI looks like.
OpenAI made waves this past week when it claimed to have made a significant advancement on solving one of mathematics’ so-called “Millennium Problems,” which are a group of very complex and as-yet unsolved problems.
The large AI laboratories are attempting to demonstrate that the technology that they’re developing is legitimately innovative and capable of feats of original, creative intelligence that can compete with cutting edge, human practitioners in established fields. That has made this class of math problems so compelling, because solving them would give the AI labs intellectual legitimacy and underscore their claims that even though these models are trained on vast troves of previously-existing human effort, their outputs are more than the sum of their parts, making them capable of genuinely innovative work.
OpenAI argued they’ve achieved a solution to the question of whether the Navier-Stokes equations always have smooth solutions in three-dimensional space.
This has generated no small amount of controversy, in part because of how OpenAI treated a researcher currently working on the problem.
That mathematician, Tristan Buckmaster, said that OpenAI sought to solve the problem after learning that he and a colleague (a researcher at rival Anthropic) had been making progress. Fueling the controversy is an unresolved allegation about OpenAI’s access to the mathematician’s own Codex logs, and OpenAI’s suggestions about how to publish.
Setting aside the controversy, there is perhaps one major takeaway here that many in the mathematics community have been pointing out: OpenAI plowed an enormous amount of resources into it.
From their own blog post, “Across all attempted problems, the agents sent 4.9 million messages and used about 300 billion output tokens. In the process of resolving the Navier–Stokes problem, the agents sent 2.7 million messages and used approximately 130 billion output tokens.”
How much money OpenAI spent to get to that conclusion depends on the company’s own compute costs, but Astra ain’t cheap and this result probably cost them millions of dollars — low tens of millions of dollars if we’re counting input tokens — to achieve.
Not to put too fine a point on it, spending millions of dollars to solve math problems is already how mathematicians solve stuff. Instead of burning billions of tokens, they just give out grants to grad students so that they can think about math for a little while. It works pretty good.
As one researcher pointed out, $22 million is something like 70 grants from the National Science Foundation, which would cover 140 years worth of grad student research in the aggregate, and that is indeed enough to legitimately move the needle on many unsolved problems.
Within math, some consternation is growing about the AI Labs targeting these problems for quick PR wins, as their removal from the realm of the unsolved deprives emerging mathematicians of opportunities with little real-world impact, with Fields Medalist Terence Tao writing:
“We have now seen that even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it before the original research project has time to reach its full potential. The incentives may now be pointing in the direction of no longer sharing any promising research directions with the broader community, which would reverse centuries of traditions of open science and do serious long-term damage to the future of the field.”
The Takeaway
Here, it’s perhaps worth contrasting the difference between how OpenAI is field-testing its computational ability with how Google is. DeepMind, which is an AI project of Alphabet, turns its attention toward hurricane prediction, where a forecasting community already well familiar with competition between models has been genuinely impressed with the frankly outstanding performance of Google DeepMind’s modeling. It has been the most accurate hurricane model of 2026, outpacing the National Hurricane Center’s track predictions in both the Atlantic and Pacific by as much as 30% in the five-day forecast, and it’s rightly being celebrated for that.
Part of that is clearly because hurricane forecasting is life-or-death science, and the hard-won advances in the field save lives and an improvement to the state of the art — and earlier, more accurate forecasts – can have enormous implications for civilization. Another part of that is that hey, there’s always gonna be another hurricane season, and Google’s been generous with what they’ve learned, collaborating with forecasters and open sourcing their models.
Sure, it’s not the kind of PR win that OpenAI seems to crave by strip-mining the field of mathematics, or even the kind that Anthropic likes by boasting of how many bioweapons they’ve managed to not make, but after decades in the business it stands to reason that maybe Google’s on to something with how to sell the abilities of new, disruptive tech on its genuine merits rather than on its trophy case.
Snacks Shots
🏈 NFL: This Sunday sees the opening weekend of the NFL season continue, with big games between the Commanders and Eagles (Philadelphia the 69% favorite*), Packers vs Vikings (Minnesota a 53% favorite), and the Ravens vs. Colts (62% Baltimore).
🏈 SNF: And then, after all those awesome games, the United States has to come together to watch… the Giants and Cowboys in a primetime game. That’s the NFL for you. Dallas is a 63% favorite as this stoppable force takes on this movable object.
*Event contracts are offered through Robinhood Derivatives, LLC — probabilities referenced or sourced from KalshiEx LLC or ForecastEx LLC.
That’s how much of GameStop’s quarterly sales now come from collectibles, per its Q2 earnings report. Overall net sales fell to $790.2 million from $972.2 million a year ago, hurt by weaker videogame and preowned sales, planned store closures, and a tough comparison with last year’s Nintendo Switch 2 launch. But underneath that decline, collectibles sales surged 57% to $356.3 million. A year ago, they made up just 23% of GameStop’s sales.
The shift is becoming harder to dismiss as a side business. Collectibles now bring in more revenue than videogames. For a company still named GameStop, more and more of the growth story is coming from things that aren’t games.
What else we’re Snackin’
Snack Fact of the Day
Return windows are getting shorter as retailers tighten up their policies, on average dropping from 43 days to 38 days at larger retailers over the past two years.
Friday
August CPI inflation report slated for release at 8:30 a.m. ET.