Four people who have built, funded, and watched companies succeed and fail walked into Alumni House and said the quiet part out loud.
Ion Stoica opened the evening with a keynote that traced his research lab’s output across two decades, less as a victory lap than as a case study in how technology compounds. Starting with Apache Spark in 2009, the insight wasn’t complex: most queries don’t touch all the data, so keep the working set in memory rather than reading from disk on every iteration. That decision turned a Ph.D. project into the standard for big data processing—and the foundation of Databricks. The pattern repeated with Ray, which unified distributed computing under a single framework so that recommendation systems, reinforcement learning pipelines, and post-training workflows could run across heterogeneous hardware without being stitched together from incompatible APIs. ChatGPT was trained on Ray. The point Stoica kept returning to: each of these projects began because researchers hit a problem they couldn’t route around. Netflix Prize. AlphaGo. The GPU memory gap. The problem came first.
His third example landed differently. Chatbot Arena began because the lab needed to evaluate Vicuna, a Llama (Large Language Model Meta AI) fine-tune trained on ShareGPT data. Static benchmarks were contaminated. GPT-4 solved every Codeforces problem published before its training cutoff and zero after. Human preference evaluations didn’t scale past pizza-fueled lab sessions. So they built a blind, side-by-side comparison interface and let users vote. The insight was structural: if you’re deploying models through a chat interface, the only honest evaluation is conversational and human-judged.
The VC panel that followed, moderated by Chuck Ng ’96 and featuring Murray Rode of Bow Capital, Michael Stewart ’15 of M12, and Derek DeVries ’93 of the Berkeley Space Center, zeroed in on the essential question: what separates the companies that achieve and grow from the ones that don’t?
Stoica’s answer was market, then team. Not product, not pitch deck. The market is the thing founders can’t change, so it demands the most honest scrutiny upfront. Stewart went somewhere unexpected for a deeply technical investor: communication. In a moment of extreme noise, where model capability announcements arrive weekly and yesterday’s benchmark is today’s footnote, the founders he’s watching succeed are the ones who can make a stranger stop, reconsider, and switch. Not through hype videos. But, through force of will and clarity. Rode said the lesson that took him the longest to learn was not to overthink decisions, because timing is real, speed matters and contemplation has a cost. DeVries said ego is the most common blocker he sees across the hundreds of startups moving through the Berkeley Space Center pipeline. Confidence, yes. But founders who can’t raise their hand and say they don’t know something will hoard information until the company breaks around them.
On the question of what founders most often get wrong with early investors, Stewart was direct: they game the valuation conversation instead of proving the problem is real. If people are using your product before a company exists, the VC conversation becomes simple. If they’re not, no amount of narrative engineering compensates. Stoica added that first-time founders consistently misjudge how much time they have. Learning feels like momentum, but it isn’t. Someone who has done it before moves faster. In an era where speed compounds, that gap is decisive.
The final question was about the next decade. Stoica pointed to problems the field has assumed are distant: cancer, energy, interplanetary infrastructure. AI, he argued, is compressing timelines, and the founders who position themselves at those problems now, before the crowd arrives, will find easier ground than those entering mature fields. Stewart focused on the compute stack itself. With ten-trillion-parameter models driving demand faster than power infrastructure can follow, he argued the entire basis of computing needs re-examination from the inference endpoint backward. The chip architecture that evolved from the IBM PC was never designed for token generation at scale. The economic incentive to rebuild the stack from scratch has never been greater.
Derek DeVries announced Berkeley Innovators that night, a new investment syndicate built to bridge the funding gaps that leave university-born startups stranded between seed and Series A. He said the biggest developable opportunity in Silicon Valley is 36 acres at NASA Ames, where Berkeley and NASA scientists will work alongside private industry tenants. It echoed the night’s central idea: the problems worth building toward are the ones that feel too large and too far away to touch.
The closing advice, when each panelist was asked what they’d tell their younger self converged on three words: focus, move fast, take the risk.
Keep an eye on upcoming events at CAA.
Photo Credit: Don Collier
