Why Math? Or, the Nature of AI Problems

Math is a closed system. A proof is verifiable as either correct or incorrect. That makes math a superb playground for AI. Software engineering is similar because either the code compiles and the tests pass or they don’t (of course, the tests need to be thoughtfully constructed and suitably comprehensive, but that’s a topic for another day).

So it’s no surprise that OpenAI chose math to show what its models can do. The Information wrote about it this week, after the company published 722 math manuscripts on GitHub, many with machine-checked proofs. It’s hard to argue with a proof.

Although business isn’t math, some business problems are more closed than people expect. For example, contract drafting. Most commercial contracts are reviewed against a playbook that spells out the acceptable liability cap, the governing law, and the clauses that the Legal team will never accept. A draft contract either conforms or doesn’t, as long as the playbook is written clearly and comprehensively.

Other areas are more open than they look. For example, company branding. There’s no compiler for a brand. If you reposition yours, you’ll find out whether it worked in a few years, often in a dynamic and competitive market that is difficult to disentangle.

Even something as mundane as a calendar can be surprisingly open. Imagine two meetings potentially conflict. One is a normally important product review meeting and one is labeled “Coffee w/David.” What the calendar doesn’t know is that David is a very senior but confidential potential new hire. An AI can’t pick the right meeting because the context that makes this decision was never written down.

The trap is in wishing for an open problem to be closed. For instance, AI SDRs. Outbound sales looked closed: Send emails, book meetings, count the meetings. Done. The pitch a few years ago was that you could replace your entire SDR team with AI. Many of these tools have high churn (50-70% by one estimate), however. “Meetings booked” was the closable part, but whether these were actually good meetings was not.

Several startups have pitched me a variant of this issue. These startups believe they can observe their way to context. The pitch goes something like this: Read every email, transcribe every call, view every document, and eventually the system will know enough to make the decision itself. But these systems often run into problems. The AI reading every email still doesn’t know about that conversation that happened over lunch or that epiphany I had on this morning’s run. More data from the wrong places don’t close an open problem.

So what does actually work? Close the parts that can be closed, gather data along the way, and give the human a much better starting point so they can get the problem over the finish line.

When this is done well, two good things happen. The first is that demand goes up. William Stanley Jevons noticed in 1865 that more efficient steam engines didn’t reduce England’s coal consumption, they increased it. Similarly, many business tasks are rationed today because they’re expensive or otherwise scarce. When you fix the constraint, companies do much more of the tasks. 

The second is somewhat ironic, which is that when you remove the supply constraint, the extra work generates exactly the context that the open problem was missing. And these data are just what the AI system needs in order to do a better job.

Verso, whose seed round we just led, sees both of these effects. Though branding is an open problem, what consumers actually said is not. Designing a study, recruiting participants, running interviews, and synthesizing results are all closable, and Verso runs this loop with AI-moderated video interviews. Since traditional consumer research is slow and expensive, most companies ration it to a few studies a year and rely on gut feel for everything else. Verso says its clients get insights about 10x faster, at a quarter of the cost, while talking to 15x more consumers. These benefits lead to more demand for insights, Jevons’ Paradox at work.

In the Verso model, the brand manager still makes the call, but now with evidence. And every interview lands in a knowledge base that makes the next decision a little less open, both for the people making the decision and, increasingly, for the agents assisting them.

With a verifier, many math problems can be solved by AI. Though many business problems will never be solved this way, we should use AI to close the parts of each problem that can be closed and let the humans (and the data they generate) do the rest. Over time, the open problems will get a little smaller. Even branding.

Jerry Dischler

Could the future of software be fluid
How do we get the best of AI without losing the soul of software?

The future belongs to young missionary teams
Why it makes more sense betting on youth in the current moment

The AI-native enterprise playbook
Ten real-time observations on a rapidly evolving playing field

No more painting by numbers
It’s the end of the “SaaS playbook.

HARD TECH

Closing the packaging gap. GlobalFoundries will make silicon interposers for TSMC’s CoWoS packaging at its Malta, New York fab under a five-year, $2B deal. Interposers connect processors to high-bandwidth memory in many advanced AI accelerators. Advanced packaging, the report notes, ‘has become a major bottleneck for American chip manufacturing.’ Blackwell GPUs made in Arizona still go back to Taiwan for final assembly. GlobalFoundries expects volume production no earlier than 2028, while Amkor’s and TSMC’s Arizona packaging facilities are scheduled for 2028 and 2029, respectively. Intel Foundry already offers comparable packaging for TSMC-made silicon, though designs must be adapted to it.

HARD MARKETS

One customer, $35B. Lambda is raising up to $4B at a $14.5B pre-money valuation, led by Coatue and Blackstone, in what may be its last private round before a 2027 IPO. Its backlog grew from $15B in June to $50B in September, but $35B of that is a single Anthropic commitment, so the valuation “could be leaning heavily on Anthropic’s ability to keep paying.” With data center buildouts heavily debt-financed and lenders becoming more selective, Lambda is turning to equity ahead of a potential IPO.

Meta tests the paywall. Meta has released Muse Wonder, its most advanced model for powering digital assistants, with particular strengths in law, finance and health. Meta AI chief Alexandr Wang says the model is “specifically focused on personal and professional agents.” Access sits behind paid tiers of the Muse app and the Muse Code agent, and developers can reach it through the Meta Model API. It’s a notable shift for the company behind the freely available Llama models.

HOW TO STARTUP

When agents pick you. AI agents are changing how software gets discovered, according to MongoDB CEO Dev Ittycheria. Ittycheria points to Vercel’s reported $600M ARR, with agents accounting for 50% of new business, up from 3% at the start of the year. Agents are increasingly recommending infrastructure providers based on the applications they’re building. The discussion also explores how startups can make their products discoverable and usable by agents, from clear documentation to accessible APIs. “If you’re missing from an AI-generated answer, it’s like being invisible on Google.”

HOW TO VENTURE

Slower, not a slowdown. North American startups raised $92B in Q3, down 35% on Q2 but up 50% year on year. The dip largely reflects the absence of new OpenAI and Anthropic megarounds. Deal counts held steady, and two-thirds of funding still went to AI. Liquidity came through M&A, with 11 acquisitions of $1B or more, led by Nvidia’s $12.9B purchase of Hugging Face, while 17 venture-backed IPOs raised under $4B combined. “It was the IPOs that didn’t happen this quarter that attracted the most attention.”

Verso raises $6M in seed funding led by Angular Ventures to scale its AI-powered consumer intelligence platform and expand internationally.

Motorica
Product Marketing Manager (Stockholm)

Paradime
Product Engineer (Remote)

Sourcix
Lead Product Manager (Tel Aviv)