Standing room only at 8:30 a.m. on the first morning of Ai4 in Las Vegas, and the opening keynote led with the hardest science on the agenda before generative AI got a single slide. That ordering was itself a signal. Three sessions from day one are worth carrying home: how AI is changing drug development, why an enterprise software vendor wants CIOs to hire their AI instead of governing it, and where compute and power now set the ceiling on everything else.
The Nine-Month Drug Candidate
Insilico Medicine founder Alex Zhavoronkov, Generate:Biomedicines co-founder Gevorg Grigoryan, and Radical Numerics co-founder Eric Nguyen sat down with TIME's Alice Park to talk about where AI changes drug development rather than speeding up the parts that already worked. Zhavoronkov's numbers carried the panel. Insilico now has 31 AI-designed preclinical candidates, one already in phase three and a broader clinical pipeline spanning multiple phase one and phase two programs. A traditional discovery-to-candidate timeline runs roughly four and a half years and hundreds of millions of dollars. Frontier AI, he said, has brought Insilico's record down to about nine months, with typical programs still landing in the low teens of months.
Nguyen's contribution was the uncomfortable one. Radical Numerics trains models to read and write DNA the way a large language model handles text, and its Evo model was used last year to generate a full genome from scratch, a bacteriophage that infects bacteria rather than people. The demonstration was benign. The implication was not: the same capability that designs a therapy can, in principle, design something dangerous. Nguyen pointed to the 1970s Asilomar conference on recombinant DNA as the model worth reviving, a voluntary, scientist-led set of guardrails negotiated before regulation forced the issue rather than after an incident did.
Grigoryan framed Generate:Biomedicines' bet as programmability, treating drug design as reusable engineering components rather than a one-off invention each time. All three agreed on where the AI stops helping. The search space narrows dramatically, but wet-lab validation remains the only ground truth, and nobody on stage suggested that changes soon.
The Spreadsheet Nobody Could Kill
Dataiku chief marketing officer Mark Abramowitz and SVP of AI and platform Jed Dougherty opened with a number meant to unsettle the room: in a survey of 900 CEOs across eight countries at companies with at least half a billion dollars in revenue, 96 percent believe employees are already using generative AI without formal sign-off, and 42 percent estimate at least half their workforce is doing it. Their framing carried the rest of the talk. The demand is legitimate. The infrastructure to support it safely is what is missing.
The analogy was the spreadsheet. IT never built it, never approved it, and forty years later has never managed to retire it, because finance and operations built entire workflows on top of it faster than central IT could ship an alternative. Dougherty's argument is that generative AI is following the identical pattern, compressed into months instead of decades, and that trying to lock it down repeats a mistake the industry has already made once.
A live demo made the risk specific rather than abstract. A sales dashboard looked clean, plausible, board-ready. One hidden formula error, a wrong date range pulling actuals where targets belonged, made every number on it wrong in a way that looked completely credible. A broken chart is obvious. A wrong chart dressed up as a correct one is not, and an autonomous agent making the same class of error can act on it, emailing customers or moving money, before a human notices anything is off.
An agent that cannot pass review should not ship, the same way a new hire who cannot pass probation does not stay.
That line captured Dataiku's actual pitch: treat agents like labor rather than software, with a named business owner, a budget line, an onboarding and review cycle, and the authority to shut one off. The company has shipped two products this year built around that idea. Cobuild, generally available since June 18, turns a plain-language business objective into a governed, inspectable AI project without requiring code. A companion Agent Management layer is meant to function as a control plane, tracking which agents exist, who owns them, and what they cost. Dougherty told the story of a manufacturing client whose CIO, after finally auditing every cloud platform running in the company, found more than ten times the number of agents in production compared with what leadership believed existed.
Nobody could say who had built most of them or what they were doing.
The stat that landed hardest came from a follow-up question. Eighty percent of the CEOs in Dataiku's survey say their own job is at risk if they cannot show measurable AI return on investment this year, and 77 percent believe a peer is going to be fired over a failed AI strategy or an AI-driven incident. Dataiku's separate research into CIOs specifically, a different survey of 600 CIOs run through Harris Poll earlier this year, shows the same pressure running slightly cooler at 74 percent. The direction holds either way. The pitch to IT departments is a role change, from gatekeeper approving requests to a platform team building the infrastructure that makes approval close to automatic, mirroring the shift HR made decades ago for human hiring.
Compute Becomes the Bottleneck
Former Intel CEO Pat Gelsinger, now a general partner at Playground Global, joined OpenAI's head of compute Sachin Katti for a conversation moderated by The New Yorker's Gideon Lewis-Kraus, titled "The AI Reckoning: Chips, Constraints and the Next Generation of Compute." Their shared argument was that progress from here is gated by physical limits as much as by algorithms. Chip supply, power availability, and capital deployment now set the pace of what gets built, not just model architecture decisions made in a research lab. The framing that carried: AI has stopped being purely a software story and become an infrastructure one, which changes who controls the roadmap.
If someone asked you right now how many AI agents are running against your production systems, would your answer survive an audit the way this manufacturer's did not?
I covered Ai4 from the show floor through day one only.
Black Hat starts the same week, and that's where I'm headed next.
