Sep 4, 2026 · 1 hr 3 min · 13 segments
In this episode, James is joined by Dan Balsam, co-founder and CTO of Goodfire, to talk about Silico, Goodfire's new automated AI research platform, and why the alignment problem is too big to leave…
Dan BalsamGuest
James BowlerHost
Okay, la-last thing I wanna talk about is, um, where does Goodfire sit within the AI alignment community? So we, we've talked about some of the, um, kind of, um, players already.
We, uh, like Goodfire, interesting, uh, kind of have the for-profit model and, you know, we, we can talk about, um, about that.
I'm curious what your takes are on kind of, um, what the unique advantages are of, of being a, like for-profit in this space and, and versus a nonprofit and, and what advantages that, that gives you.
Um, um, um, but yeah, like what-- to what extent do you feel part of the alignment community? What do you see as your, your role, um, in, in ultimately like solving the, the problem that we all, we all care about here?
Yeah, I mean, I think we absolutely think of ourselves as, as part of the, the larger alignment community.
Um, you know, we've taken-- We have a couple of strong beliefs, like, you know, we believe that interpretability is like very essential, like very likely route.
Um, we believe that at some point we're gonna have to figure out how to intervene in the training process, not necessarily through things like reward shaping, but we have to figure out some way to do that effectively.
Um, and, uh, we also believe that accelerating research is, uh, a great mitzvah, and we should do everything we can to accelerate research and especially alignment research.
Um, and so, yeah, so I view our role in the community as like, you know, being the champion of, of those beliefs and doing whatever we can.
Um, in terms of for-profit versus nonprofit, I think there's pros and cons to both.
Um, ultimately, I think the best case scenario is that alignment and safety is economically valuable.
'Cause if alignment and safety is not economically valuable, then, um, there's always a gradient pressure away from it.
I think we always believed, and hopefully it's becoming maybe more clear to others, that there is an economic gradient, um, towards safety and alignment.
I think the reality is that once you have models that can do things like the, you know, OpenAI hugging face situation, like these are the types of real world risks we're dealing with.
There is real economic liability to deploying those models, um- And so there's economic pressure to have good guardrails.
Um, and I think there's gonna be economic pressure to figure out how to really solve things like alignment and reward hacking.
Um, and I think our-- one of the roles that we can play, uh, and is showing that there's a big market, um, for this type of work.
Um, so I-- and, you know, uh, I think one of the things that being a for-profit also helps you, helps you do is, like, you do have to stay very pragmatic and focus on, like, what delivers value.
Okay, la-last thing I wanna talk about is, um, where does Goodfire sit within the AI alignment community? So we, we've talked about some of the, um, kind of, um, players already.
We, uh, like Goodfire, interesting, uh, kind of have the for-profit model and, you know, we, we can talk about, um, about that.
I'm curious what your takes are on kind of, um, what the unique advantages are of, of being a, like for-profit in this space and, and versus a nonprofit and, and what advantages that, that gives you.
Um, um, um, but yeah, like what-- to what extent do you feel part of the alignment community? What do you see as your, your role, um, in, in ultimately like solving the, the problem that we all, we all care about here?
Yeah, I mean, I think we absolutely think of ourselves as, as part of the, the larger alignment community.
Um, you know, we've taken-- We have a couple of strong beliefs, like, you know, we believe that interpretability is like very essential, like very likely route.
Um, we believe that at some point we're gonna have to figure out how to intervene in the training process, not necessarily through things like reward shaping, but we have to figure out some way to do that effectively.
Um, and, uh, we also believe that accelerating research is, uh, a great mitzvah, and we should do everything we can to accelerate research and especially alignment research.
Um, and so, yeah, so I view our role in the community as like, you know, being the champion of, of those beliefs and doing whatever we can.
Um, in terms of for-profit versus nonprofit, I think there's pros and cons to both.
Um, ultimately, I think the best case scenario is that alignment and safety is economically valuable.
'Cause if alignment and safety is not economically valuable, then, um, there's always a gradient pressure away from it.
I think we always believed, and hopefully it's becoming maybe more clear to others, that there is an economic gradient, um, towards safety and alignment.
I think the reality is that once you have models that can do things like the, you know, OpenAI hugging face situation, like these are the types of real world risks we're dealing with.
There is real economic liability to deploying those models, um- And so there's economic pressure to have good guardrails.
Um, and I think there's gonna be economic pressure to figure out how to really solve things like alignment and reward hacking.
Um, and I think our-- one of the roles that we can play, uh, and is showing that there's a big market, um, for this type of work.
Um, so I-- and, you know, uh, I think one of the things that being a for-profit also helps you, helps you do is, like, you do have to stay very pragmatic and focus on, like, what delivers value.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
