Signal #67: A Model Did Original Math for $2,000

Ten problems that had sat open for decades fell last weekend, and the compute bill came to roughly two thousand dollars. Every one of them shipped with a machine-checkable proof, which means nobody has to take OpenAI's word for any of it.

That is the shape of the week, and it repeats. Astra's bill was a metered number a stranger can re-run, and Palantir's print is the same demand for a checkable figure pointed at a public company. The Sofia round is the unglamorous version of it, and the eSIM tier at the end of the issue is the one you can test for five euros: stop buying a bundle sized by a guess and pay for the units you consume.

On August 1 OpenAI published ten results in mathematics and theoretical computer science produced by an internal build of Astra, its next major release. The problems had been open at least ten years and in several cases far longer, across group theory, coding theory and lattice cryptography. The headline result is the first explicit construction of a non-sofic group, a question standing since Gromov introduced soficity in 1999. Forbes put the token cost at around $2,000.

The price tag is the headline and the wrong thing to fixate on. Each result ships as a Lean 4 formalisation on GitHub, and Lean's kernel either accepts a proof or rejects it. You do not need to trust the lab, the reasoning trace or the press release, you run the compiler. It is the first time a frontier lab has published original research a skeptic can check in an afternoon without a doctorate in the field. None of the Millennium Prize problems fell, so keep the scale honest.

The transferable move is the certificate, not the mathematics. Anywhere correctness is machine-checkable, unit-tested code, SQL against a schema, contract clauses against a policy file, you can now let a cheap model run wide and let the checker do the judging. Find the one workflow in your company where a machine rather than a manager can say pass or fail, and put a model behind it this month. Where no checker exists you still have a review bottleneck, and no price cut fixes that.

Finding one great creator is easy. Finding a hundred who genuinely match your brand, your price point, and your audience is a full time job nobody on your team has time for.

So it doesn't get done. Or it gets done badly.

partnerUP fixes both. You guide the brief, and the AI reaches creators at a scale no person could, then scores every single one by fit so only the right matches reach you. It handles the contracts, the payments, and the performance tracking from there.

You get the creator program a big brand would build, without the team a big brand would hire.

Your first two matches are on us.

Start with two free

Tiger Technology took the first close of a €8.7 million Series A on July 28, about $10 million, led by TCEE Fund IV with TDJ, Endeavor Catalyst, Impetus and existing backers. Capital.bg called it a record for a Bulgarian software round. The product, Tiger Bridge, installs on top of storage a company already owns and makes on-premises, cloud and edge behave like one place, no migration required.

Notice what got funded in a year when everything is supposed to be an agent. Not a model, not a wrapper. A control layer that makes twenty years of accumulated files reachable. The constraint in most companies was never that intelligence is expensive. It is that the data the intelligence needs sits on a box in a building, in a format nobody has opened since 2019, and the migration quote is six figures and nine months.

If you have been deferring an AI project because the data is not ready, price the control-layer route against the migration route before you write off another quarter. The two are rarely compared because they get sold by different vendors to different buyers. For founders in the region, the second read is the one worth keeping: a Bulgarian company selling unglamorous infrastructure raised from Endeavor Catalyst without relocating to Berlin, in a year when Bulgaria has seen almost no rounds this size.

Palantir reported Q2 on August 3: revenue up 93% to $1.94 billion against $1.80 billion expected, earnings of $0.41 a share against $0.28, US commercial revenue up 149% year on year. Full-year guidance went from $7.65 billion to between $8.15 and $8.158 billion. The stock jumped 12%.

The market has spent this year marking down AI spending it cannot trace to revenue, and this is the counter-print. What makes it awkward for everyone else is the composition, not the headline growth. US commercial guided at 134% means the demand is not government pilots and not renewals from customers who were always going to renew. It is companies buying deployed work and booking it as a line item.

Steal the framing rather than the multiple. The question a board asks in October is which revenue exists because of the AI work, not merely alongside it, and most teams cannot name a single line. Pull that number yourself this month while it is still a private embarrassment rather than a public one. If the honest answer is zero, that is fine in year one and fatal in year three.

Every travel eSIM on the market asks you to predict the future. Pick 5GB and 30 days, then discover you needed 2GB or 12. eSimity, a Berlin company, put a pay-as-you-go tier in its app that removes the guess: load credit, install one global profile once, and data comes off the balance as you use it with the meter visible in the Usage screen. Credit goes in at €5, €10, €15 or €50 and stays valid for three years, so a leftover balance survives to the next trip instead of dying on a 30-day clock.

This is the same billing shape as the Astra bill, arriving in a product you can hold. The bundle model exists because it is easier to sell, not because it is honest: eSimity's own comparison prices 5GB across Europe at €8.27 against roughly €17.50 on Airalo and €68.90 on Holafly. For anyone working across the region it is starker: a week through Kosovo, North Macedonia, Serbia and Montenegro is four roaming regimes on a normal contract, and one profile here.

The test costs five euros, less than one afternoon of roaming. Load the minimum before your next trip, install the profile, then compare what you burned against the package you would have bought on reflex. Most people find they have been paying for roughly double what they use. My referral code is LPPLRXB4, and yes, I get credit if you use it. The apps are here, and the pay-as-you-go tier only exists inside them, not on the web checkout.

Five tools to install or test this week, tagged by the seat they help.

Productivity: Notion's meeting notes can now start the work. Notion added a trigger on July 31 that fires a Custom Agent the moment an AI Meeting Note finishes, so the summary becomes the input to something rather than another page nobody opens. Attach one to your recurring customer call: let it draft the follow-up and post the recap to Slack before you have left the room.

Design: Figma Make got a proper editing panel. Figma shipped granular spacing and type controls on July 30, plus stacked chat edits and the ability to annotate a prototype and have the agent act on the annotation. The gap it closes is the one where you could generate a screen but not nudge it. Take a prototype you abandoned for being nearly right and fix it by annotation instead of regenerating.

Marketing: Airtop put a browser agent inside Google Ads. Airtop launched on August 3, building campaigns, reallocating spend and generating reports by driving the Ads interface directly rather than through the API. Point it at one live campaign, ask for an audit and a reallocation proposal, then compare its answer to what you would have done. The disagreements are the useful part.

Dev: Cursor's agent can now reach your Google Workspace. Cursor added Workspace plugins on August 3, giving coding agents read and write access to Gmail, Drive, Calendar, Docs and Sheets. Useful for agents that need a spec out of a Doc or numbers out of a Sheet mid-task. Scope the permissions narrowly on the first install, because write access to Drive is a wider grant than most people read it as.

Ops: GitHub can hand specific models to specific teams. GitHub opened model policy targeting to public preview, so enterprise admins grant model access per team instead of switching it on across the org. This is how you pilot an expensive frontier model with eight engineers rather than eight hundred. Create one team, grant it, and put a spend cap on the experiment before it becomes the default.

Next edition soon,

Çelik