Letting AI Think Longer May Matter More Than Bigger Training Runs
Fractile CEO Walter Goodwin envisions thousands of copies of a model debating a science problem for days.

The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · 20VC: Taking on NVIDIA: Fractile CEO on the Future of Chips | What Everyone Gets Wrong About Capex Spending | The Winners and Losers in Neo Clouds with Walter GoodwinWalter Goodwin, the chief executive of chip startup Fractile, founded the company four years ago on a bet that the hard part of AI would not be training giant models but running them. Training, he argues, is just the fixed cost. He compares training without using the model to sprinter Usain Bolt training for the Olympics and then skipping the race.
His line for it is plain: "It's the knob that we have to turn to get greater capability." By that he means capability can grow two ways, by spending ten times more computing power to train a model, or by letting the finished model spend more computing power thinking through an answer.
For years, Goodwin said, most people pictured the second way as trivial, where you ask a question and the chatbot answers at once. The arrival of reasoning models two years ago, first published by a team at OpenAI, changed that picture and reinforced his bet.
For the hardest problems, he said, the model does not just answer. It proposes hundreds of possible solutions, makes many copies of itself, searches published research, debates among those copies, and exchanges information. "Come back to me in basically a week's time, once you've generated maybe billions of tokens," he said, describing the instruction.
He pointed to a mathematics example involving the Navier-Stokes equations, a famously hard set of physics problems about fluid flow. That solution, he thought, took about 80 hours of work by a swarm of 10,000 agents, each writing its own long chain of thought in parallel.
He imagines the same approach turned on a Millennium Prize problem, a group of seven unsolved math problems. A frontier model might generate millions of tokens, the small chunks of words and numbers a model reads and writes, in sequence, taking days to reply.
The catch, in his telling, is that no current chip handles that job well. Fast chips can produce words quickly but cannot hold the huge working memory these long runs require, while roomier chips are slower. He calls that extra inference compute the access we have, the knob to turn for problems like cures and new science that one quick answer cannot solve.
Questions people ask
What is a reasoning model and how is it different from an earlier chatbot?
A reasoning model is trained to show its work in steps before giving a final answer, rather than answering at once. It can check its own steps, try another path, and use tools like search or math software, which makes extra thinking time useful.
What are the Navier-Stokes equations?
They describe how liquids and gases move, from weather to blood flow to airplane wings. They are hard because turbulence is chaotic, and one version is among the seven Millennium Prize problems, so testing many ideas in parallel can help.
Why can't today's fastest AI chips easily handle a weeklong debate?
Long runs need both speed and memory at once. Fast on-chip memory can deliver words quickly but holds little context, while larger external memory holds long histories but moves data more slowly. A days-long swarm with huge contexts hits both limits together.
