Future Proof from Semafor Technology
Oct 14, 2026 · 1 hr 1 min · 14 segments
Reflection AI co-founder and former Google DeepMind researcher Ioannis Antonoglou thinks that for companies to take full advantage of the power of AI, they'll need to rethink their businesses from the…
Ioannis AntonoglouGuest
Reed AlbergottiHost
Back then, I asked you a question about US AI regulation and you said, hey, look, I only got to the US a year ago.
No, it's been an incredible year and, you know, it just feels incredible that like, you know, last week we released our first model, Beam, which is like a frontier open model for its model class.
And, you know, it's actually really powerful when it comes to agentic and coding use cases.
know what i've actually really tried to do with beam is that uh we really wanted to just like build a workhorse that it's a model that can be really used by enterprises it's going to be used by um know sovereigns who want to own their own intelligence and if you actually look at what are the models that uh you know are the workhorse models the models that actually like consume the most tokens you can go to open router you can see the consumption of like different models based on like the different model classes and you will see that like more the models uh that uh you know uh have the most conflict token consumption uh is are at that level of um of size and uh and uh and budget compute budget This is because, like, you know, you want to be at a place where you can just, like, make the most out of the flops that you're just, like, throwing at the problem, right? So we just, like, went for a model that it's, you know, a mid-sized model, 5 billion parameters.
But then we just, like, really tried to ensure that, like, the flops and the flops in the compute that you put into this model actually leads to the best outcomes.
From tiered reasoning efficiency, what we mean by that is that, you know, for every task that the model needs to solve, you can measure how many tokens of reasoning it takes for like the model to just like solve these tasks.
this is like the new trend, right? I noticed like a lot of the frontier labs are saying this too, that don't measure based on, you know, cost per token anymore, dollars per token, measure cost per task, which I think is an interesting, it's an interesting new development.
Sometimes I wonder if it's like, They're, you know, they're trying to find a way to get you to spend more money or something or not care as much about the money you're spending.
So actually, like the way that you can do it is by having different models or like solve tasks.
And then you just like you measure how many tokens it took for like the model to solve this task.
Back then, I asked you a question about US AI regulation and you said, hey, look, I only got to the US a year ago.
No, it's been an incredible year and, you know, it just feels incredible that like, you know, last week we released our first model, Beam, which is like a frontier open model for its model class.
And, you know, it's actually really powerful when it comes to agentic and coding use cases.
know what i've actually really tried to do with beam is that uh we really wanted to just like build a workhorse that it's a model that can be really used by enterprises it's going to be used by um know sovereigns who want to own their own intelligence and if you actually look at what are the models that uh you know are the workhorse models the models that actually like consume the most tokens you can go to open router you can see the consumption of like different models based on like the different model classes and you will see that like more the models uh that uh you know uh have the most conflict token consumption uh is are at that level of um of size and uh and uh and budget compute budget This is because, like, you know, you want to be at a place where you can just, like, make the most out of the flops that you're just, like, throwing at the problem, right? So we just, like, went for a model that it's, you know, a mid-sized model, 5 billion parameters.
But then we just, like, really tried to ensure that, like, the flops and the flops in the compute that you put into this model actually leads to the best outcomes.
From tiered reasoning efficiency, what we mean by that is that, you know, for every task that the model needs to solve, you can measure how many tokens of reasoning it takes for like the model to just like solve these tasks.
this is like the new trend, right? I noticed like a lot of the frontier labs are saying this too, that don't measure based on, you know, cost per token anymore, dollars per token, measure cost per task, which I think is an interesting, it's an interesting new development.
Sometimes I wonder if it's like, They're, you know, they're trying to find a way to get you to spend more money or something or not care as much about the money you're spending.
So actually, like the way that you can do it is by having different models or like solve tasks.
And then you just like you measure how many tokens it took for like the model to solve this task.
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.