August 30, 2026
TechCrunch reported this week that Caterpillar is applying lessons from decades of autonomous mining equipment to how it deploys AI more broadly across its business. That's a striking source of expertise. Caterpillar didn't get autonomous trucks running at remote mine sites by writing better algorithms in a lab -- it got there by solving for dust, heat, spotty connectivity, skeptical operators, and safety regulators, all at once, over years. The models were arguably the easy part.
That's the piece of the AI conversation most vendors and pundits skip. Everyone wants to talk about which model is smartest this month. Almost nobody wants to talk about the unglamorous work of getting a tool to actually stick inside an organization that has legacy systems, workers who don't trust the new thing, and zero tolerance for downtime. Caterpillar's edge here isn't a foundation model -- it's institutional muscle memory for exactly that kind of rollout. I think that's the more valuable asset in this moment, and it's a big reason we've argued that teams should run a new tool alongside the old one before fully committing, rather than assume the software is ready the day it ships.
The lesson for ordinary businesses evaluating AI tools isn't 'go build autonomous mining trucks.' It's that deployment discipline -- phased rollouts, clear ownership, a plan for when the automation gets something wrong -- matters as much as picking the right vendor. If a heavy-equipment giant with a century of industrial engineering behind it still treats deployment as the hard problem, a 40-person services company should take that seriously before assuming a new AI tool will just work out of the box.
Vijay Pande left a16z, where he ran a biotech practice worth roughly $4 billion, to start a much smaller, AI-native fund called VZVC. TechCrunch's interview with him is less about the fund size than about his underlying thesis: biology is shifting from a discovery science, where breakthroughs are mostly luck and intuition, to an engineering science, where AI lets researchers predict and design outcomes with real precision. He also argues that clinical trials remain punishingly expensive, and that open, shared datasets -- not companies hoarding their own proprietary data -- are what will actually let AI move medicine forward.
I find the 'discovery to engineering' framing genuinely useful, and not just for biotech investors. It describes what AI is doing across a lot of fields: turning problems that used to require rare expertise and years of trial and error into problems you can iterate on quickly. But I'd push back gently on the idea that smaller checks and open data alone solve biology's cost problem. Clinical trials are expensive because of regulation, patient recruitment, and liability -- structural costs that better models don't erase. Pande seems to know this; his emphasis on trial economics suggests he sees AI as necessary but not sufficient. That's a more honest position than the breathless 'AI will cure disease' takes that show up whenever a fund raises money.
For business readers outside biotech, the interesting signal is the bet-sizing itself. A veteran investor deliberately choosing fewer, smaller, more hands-on bets over spraying capital across dozens of startups is a quiet vote against the idea that AI opportunities are so obvious you should just fund everything and see what sticks. That's a useful gut-check for any company assembling its own AI stack, too -- better to pick a few tools you'll actually integrate well, the way we talk about replacing a scattered stack with fewer, better-fitted pieces, than to bolt on every shiny new agent.
Which of these resonates more with your own experience: the grinding operational work of rolling out AI, or the slower, more deliberate way capital is now flowing into it?
Sources