Eric Bigelow, a researcher at the interpretability company Goodfire, said that all the research for his recent paper, Forking Fast, "was done fully by Silico," Goodfire's research agent. It all happened within a week, he said, and the agent was "a huge accelerant." His advice for working with the agent is mostly about what the human has to keep doing.
Bigelow spoke on The Cognitive Revolution, in an episode published October 10, 2026. Near the end, host Nathan Labenz asked what he should "be sure to take to my silico sessions to get the most from my, my precious research units."
What Silico is
Interpretability is the study of what goes on inside an AI model: which internal signals produce a given behavior, and why. The work is very experimental. Researchers train small tools to read a model's internal activity, run many variations and compare the results. Goodfire describes Silico as an agent built for this work. It develops experimental plans, runs experiments in parallel, monitors long-running jobs and returns results that researchers can inspect. The company lists methods it can use, including probes and sparse autoencoders, which are both ways of reading a model's internal signals, along with causal analysis and data attribution.
Don't hand over the whole project
Bigelow said his advice applies to research agents generally, not only to Silico. "It can be really tempting," he said, to hand an agent an entire project, tell it to work out a big open-ended problem, then wait for the plots and descriptions. The risk is that agents "get a bit like lost in the sauce" and produce theories "that aren't fully formed or don't fully make sense." He thinks agents work best with "a little bit of handholding," while the researcher keeps track of what is happening along the way.
The other half of his advice is context, which he called "everything." A poorly specified question can bring back something "a little bit funky." So whenever he reacts to an agent's work, he explains why. He tells it what looks wrong and why, what he wants to try, what he will do next and what his larger goal is.
He explained why this matters from the agent's side. He pointed to a recent video imagining life as an agent in a swarm, in which you "wake up in a library" with no idea what is going on. An agent starting in a fresh context, he said, does not know how the world works. It needs to be told why the work is being done and, when something fails, why it failed.
How Forking Fast came together
The Forking Fast paper, posted in August 2026, deals with a measurement problem. Bigelow's research tracks how the answer a language model is likely to give shifts as it writes. Estimating that requires generating many alternative continuations at each point in the text, which gets expensive. The paper uses statistical smoothing to get usable estimates from far fewer samples.
Bigelow said the project included "a bunch of different like side quests" that went nowhere. The turning point came when he ran very heavy sampling and saw that the resulting curves were smooth. "It kind of clicked," he said: maybe a statistical model could make the measurement more efficient. The paper matches that account. As sample counts increased, it reports curves that were mostly smooth with occasional sharp jumps. It handles the two separately: it detects the jumps, then smooths the stretches between them. The paper also notes limits. Smoothing improved estimates when samples were few, but some of the largest jumps were estimated worse after smoothing, and the uncertainty ranges it reported were too narrow.
The paper credits its research and initial draft to Silico working under human direction. Its appendix also records a problem found during implementation: the cost function used to detect the jumps had been silently swapped for a different one.
Bigelow described his own role as "carving the scientific narrative," which meant finding the core thread that holds the work together.
Taste as the human job
Bigelow passed on a colleague's view. In a world of agents producing results that can sometimes look really good but turn out "a little bit fishy" on closer inspection, research and scientific taste is "everything." Bigelow called it "worth its weight in gold."
He granted that agents not yet being able to do this themselves could be seen as a downside. But he called it "one of the most fulfilling parts of doing science," and said he is glad it remains an important role for people. That role means bringing your taste to bear, keeping track of what is happening and giving feedback.
His comparison for Silico was "a very eager, like first or maybe second year PhD student" with many skills and a strong work ethic, who "doesn't always know what it's doing in the bigger picture." He recalled his own early graduate school days, when his advisor would send him off on a task and he would come back saying the results were cool but he wasn't sure what was going on with them. His biggest tip, he said, is "to hold on to your scientific taste" and to take part in the research.
Why Goodfire wants interpretability to become routine
Labenz's final question was what people at Goodfire talk about over lunch. Bigelow said agents doing interpretability has been one of two big topics since the summer. He is now doing a lot of this work on another project too. If interpretability is to "keep up with the breakneck speed of the field," he said, partly automating it "might be the only way."
He also described the goal of making interpretability a kind of "normal science." In that kind of science there is a steady supply of "interesting, but boring experiments" on many similar phenomena. Each one is modest, but together they help "paint the bigger picture of how neural networks work." He said he had recently been asked by someone how to do good research with Silico, and had given much the same answer he gave Labenz.