This Week in Startups · Oct 8, 2026 · 1h 18m · 58m saved
With Deedy Das, Partner at Menlo Ventures, guest host · Kanu Gulati, Partner at Khosla Ventures · Chris Farmer, Founder and CEO of SignalFire · Sheel Mohnot, Co-Founder and General Partner at Better Tomorrow Ventures
Instinct, a startup building a personal AI agent, is valued at $10 billion, and none of the three venture investors on this roundtable made the case for paying that price.
Sheel Mohnot of Better Tomorrow Ventures said he uses Instinct all the time. He still said the large companies, OpenAI, Google, Meta and maybe Apple, have a huge benefit, and that he has tried so many rival agents he is reluctant to hand his data to another one.
"I actually have personal agent fatigue." — Sheel Mohnot
Deedy Das, a partner at Menlo Ventures, filled in as host on This Week in Startups. His panel was Kanu Gulati, a partner at Khosla Ventures, an early investor in OpenAI; Chris Farmer, founder and CEO of SignalFire, which moved from about 80% horizontal software to 80% vertical AI; and Mohnot, who co-founded Better Tomorrow, a seed-stage fund investing in fintech and vertical AI.
The full episode is covered here so you can skip it. 78 minutes of audio, 20 minutes of reading.
Here are the 13 arguments that matter.
Key Takeaways
None of the three panelists backed paying $10 billion for Instinct, and Farmer called himself "a bear"
Facebook and Google each reach billions of people and already hold their users' email
Mohnot doubts personal agents reach a billion daily users, but sees hundreds of millions
Instinct paid a medical bill that Meta's Muse would not, because it stored his card number in plain text
Khosla Ventures has not invested in any AI training-data company, because Gulati doubts their terminal value
SignalFire flipped from about 80% horizontal software to 80% vertical AI, betting on private data the models cannot see
Compute may stop being the constraint, Mohnot said, and lab capacity and clinical trials may take its place
SaaS is still investable, at a lower entry price, Farmer said, while Mohnot called some of it "cigar butt businesses"
Funds now often back competing startups, and Khosla keeps the partners on rival companies behind a strict information wall
1. Math Falls First
Das opened with a math result from OpenAI. Counting an earlier Navier-Stokes result, he said AI had now made major progress on four of the six unsolved Millennium Prize problems, and asked what is left for a startup to own.
Gulati said it is getting harder to find value that does not flow to the model makers
Khosla Ventures was an early investor in OpenAI. As a former scientist, she said the pace makes her uneasy.
"But it is a very very fair question about what can we invest in early stage that will not ultimately flow into the token generators and it is harder and getting harder and harder for all of us." — Kanu Gulati
Math works for AI because it is a formal domain where proofs can be built on proofs, she said. Her bets try to turn messy real-world fields into formal ones: Pramaana Labs on tax rules, and Logos Research, out of Imperial College London, on financial code. She expects chemistry and physics to move faster in software, and sees more companies closing the loop with wet-lab and robotic experiments.
Farmer expects the biggest gains in human health
SignalFire spends a lot of time in healthcare. Clinical trials only advance a drug that has already worked, so the failures were never recorded, he said.
"Whereas in an AI world, you want to capture all the failures as well in order to train and learn from that those data sets." — Chris Farmer
He also named climate and other complex systems, where there are too many inputs for people to agree on cause and effect.
Mohnot said math came first only because it is easy to verify
The larger prize is making invention itself cheaper across science, he said. AI could design a cancer therapy quickly, but testing it is another matter.
"But I think there's still a lot in proving that it works safely in humans." — Sheel Mohnot
A better battery chemistry still has to be manufactured and tested, which he said is why so much money is going into physical AI.
