Andrew Feldman, Robin Rombach: Open Source Wins, AGI Is Here, and Scorsese's AI Toolkit with CEOs of Cerebras & Black Forest Labs

The conversation opens by reintroducing Andrew Feldman, CEO and founder of Cerebras (transcribed as "Cerebrus"), a company building inference chips that recently completed an IPO. The first topic is the physical scale of AI infrastructure. The host compares it to the Great Wall of China, the pyramids and, within recent history, the war effort, describing it as a mobilization of capital, time and talent with no clear parallel in their lifetimes.

The discussion contrasts this with how hardware is usually talked about. Chips and boxes do not have "the same sort of physical enormity" as what is being built now. The claims made are large. Data centers in the next several years are said to be on track to use more power than the previous 50 years on earth took, and individual buildings the size of football fields are described as drawing more power than mid-sized cities. The construction is described as spanning the U.S., Canada, the Nordics, France and wider Europe, the Middle East, and countries such as Kazakhstan, Tajikistan, Georgia and Armenia. The discussion also says every country, and every U.S. state, feels it needs to take part.

On the buyer side, the speakers name OpenAI, Anthropic, SpaceX AI and Google and call them "insatiable right now." They say these buyers placed orders for Cerebras chips before the chips were finished. The point they stress is that this boom differs from earlier tech booms because buyers are not speculating on future demand:

they're trying to capture yesterday's demand, right? The demand is way outstripping our ability to build data centers and to fill them with hardware.

A $25 billion backlog is cited, and the speakers say it is not unique. Google, Microsoft and AWS (transcribed as "ATWS") are all described as wanting more data centers. They frame this as the opposite of "if you build it, they will come": the demand is already booked, and the question becomes how to keep customers from leaving. The speakers call this "extremely unusual."

The host raises the debate around "token maxing": is the heavy token use creating value? The answer accepts both sides. "There is clearly massive value happening," but there is also "massive experimentation." The comparison offered is early AWS, when engineers were told to sign up with a credit card to get around their own IT organization. Much of that work was useful, and some of it led to regret. The conclusion drawn is that experimentation does not cancel out enormous net value. It only means some of the spending will go nowhere.

A second comparison is Costco's opening in the Palo Alto area in 1988. Early shoppers walked every aisle as they would at Safeway and came home with "four things you didn't need and each was $22." Experienced shoppers went straight for specific items. The speakers apply this to enterprise AI. At first, companies gave everyone as many tokens as they wanted. Now they are starting to allocate: high-productivity users get what they need, and other work goes to open source or cheaper models. One speaker describes this as "running like a business," in keeping with the principle that enterprises do not normally give anyone an unconstrained resource.

The host also describes a kind of user that is emerging: people who think in systems, as developers and strategic CEOs tend to. The host has seen founders, venture capitalists and associates at their own firm start playing with the tools until "the tool starts playing with them." The AI asks for a requirements document, asks what the goal is, and tells users they are token maxing and need to focus.

A colleague from 20 years ago is quoted as saying computers "do exactly what you tell them." The discussion says early prompting worked that way, since small edits to a prompt changed the answer dramatically. Newer models are described as increasingly understanding intent. Using the models transcribed as "Fable or 5-6 from Open A," users no longer need to be a "prompt whisper." The example given is a model suggesting a chart with both a line and a bar, which the user had not requested but preferred. The speakers call understanding intent "a huge leap." They add that two years ago no one would have predicted the move, within 24 months, from summarizing web results to understanding intent and delivering an abstracted solution.