Three AI stories crossed my desk this week, and they all point the same direction: these systems are moving from helpful assistant to doing the grunt work nobody wants to check by hand. Here's what happened, and what I make of it.
Claude Machine-Checked Fermat's Last Theorem
Anthropic let Claude run mostly on its own for 11 days on a platform called Prove2Me, built by a team out of Columbia. The job was to take the proof of Fermat's Last Theorem and translate it into Lean, a language a computer can check line by line with zero room for hand waving. Claude came back with over 13 million lines of Lean code and something like 30,000 new theorems along the way. A mathematician at Imperial College London who reviewed the work called it an extraordinary achievement.
Let's be straight about what this is and isn't. Claude did not solve Fermat's Last Theorem. Andrew Wiles did that back in 1994 after seven years of grinding. What Claude did is take that human proof and turn it into something a machine can verify without trusting anybody's word for it. That's still a massive job. Formal verification at this scale used to eat up years of a research team's life.
My take: this is exactly the kind of work these models should be doing. Nobody dreams of spending three years re-checking someone else's algebra for gaps. That's tedious, unglamorous, and necessary. If AI can chew through that kind of grind and hand mathematicians their time back, I'm all for it. I'd rather read about this than sit through another chatbot demo.
A Third of Companies Are Skipping the Software Aisle
McKinsey put out its State of AI in 2026 survey, and one number jumped out at me. 32 percent of organizations said they passed on buying a piece of software, or at least a feature, because they figured their own team could build it with agentic coding tools instead.
This matters because the whole SaaS pitch runs on the idea that building internal tools is expensive and risky, so companies would rather pay a subscription. If a third of buyers now think an AI coding agent closes that gap, that's a real dent in the pitch every software salesperson has been making for twenty years.
Here's my worry. Not every company that skips a purchase should. A lot of "we'll just build it ourselves" software turns into an unmaintained script that breaks the day the one guy who wrote it goes on vacation. But whether or not that belief is fully earned yet, it's already changing buying decisions, and belief moves budgets faster than results do. If you sell boring internal tools as a service, pay attention to this number.
SoundHound Closes the Deal on LivePerson
SoundHound AI finished its acquisition of LivePerson this week, combining SoundHound's voice and agent technology with LivePerson's digital messaging network. As part of closing, SoundHound also paid off LivePerson's outstanding debt.
Voice AI and chat AI have mostly lived in separate lanes. Voice assistants answer the phone. Chat tools run the little bubble in the corner of a website. This deal puts both under one roof, betting that businesses want a single AI system that can talk to a customer on the phone and text with them five minutes later without starting over.
I think you're going to see a lot more of this. Smaller AI companies without a real moat are going to get bought up, not because they failed, but because the market wants fewer, wider platforms instead of a pile of point solutions that don't talk to each other. If you're running one of those point solutions, now's a good time to think about who you'd want buying you.
That's the roundup for today. A math proof got checked, a buying habit is shifting, and another AI company got folded into a bigger one. Same story most weeks lately, just different names.