Safety & governance accounted for 17% of the relevant conversation across 24 episodes.
How to read it
Recurrence across podcasts does not mean agreement. The cited moments show the language behind each summary.
In the discussion
there's not really much point and it might be sort of a waste of time or perhaps even counterproductive to try to get the safety measures to be super robust relative to capabilities that don't exist yet
New models, training methods, benchmarks, and discoveries from AI labs.
Trend this week
Gaining attention19%of conversation
6 of 6 published weeks met the cross-podcast threshold.
11quoted moments6episodes6podcasts discussing this topic
A closer read
What changed
Its share rose from 17.6% to 19.3%, with evidence across 10 podcasts.
Where it showed up
It appeared in episodes including “How Researchers Test AI for Hidden Goals — Apollo Research” and “Is Offense or Defense Dominant? FAR.AI's Adam Gleave on the AI Security Leaderboard”.
How to read it
6 cited episodes from 6 podcasts are shown below. The topic appeared across 10 podcasts in the full analysis.
Podcast moments from the week
11 quoted moments · 6 episodes · 6 podcasts
Machine Learning Street Talk
How Researchers Test AI for Hidden Goals — Apollo Research
“It's possible that the model might get really good results in its training distribution, but it might actually be learning the hospital where the test was taken, not the actual test as well. So it doesn't generalize. But why is this uniquely bad with language models and reinfo...”
“I guess the interesting thing for me is that when we think of reinforcement learning algorithms like AlphaGo Zero, it makes sense that they are reward-seeking because there is this structured inference process. this is a language model. It's doing greedy sampling of tokens. So...”
“I think when you're trying to get a high reward, it's really useful to think about how are they going to score me? How are they going to be evaluating me? How are they checking my answers? And then modifying your behavior to optimize for those. And things that are really usefu...”
“What's the sort of like, uh, superposition of mental models that you use between next token and, you know, persona selection or whatever other paradigms you kind of combine as you think about what these things are or how we should think about them?”
“Is it possible now with these increasingly powerful models to run this computationally and see what happens after, you know, a few trillion iterations?”
“According to the actual benchmarks, this beats Fable 5 on a lot of the benchmarks that we have seen. Now, you and I have both had time to work with this in practice. We can talk a little bit more about what other people think. But we have time to spend— we spent time with this...”
Its share rose from 3.8% last week to 16.5% this week, reversing the previous week's drop.
Share of this week’s podcast conversation
Recent weeks6.9%→This week16.5%
Came up more often than in recent weeks.
Podcast moments5 cited momentsViewClose
The Cognitive Revolution
Nathan Goes to China – Part 2: AI Safety with Chinese Characteristics
there's not really much point and it might be sort of a waste of time or perhaps even counterproductive to try to get the safety measures to be super robust relative to capabilities that don't exist yet
How Researchers Test AI for Hidden Goals — Apollo Research
Yes. And it is conceivable that this process is accelerating over time, especially like everyone is now talking about continual learning, you know, because right now at least we have a centralized Right? So we have red teaming, we have frontier companies building these models,...
Mythos is the first one where it seems to have insane power to do unexpected things. And like, do you think basically that we're pretty close to the limit here? Like, you know, because at first I thought, yeah, it was bad. It was safety washing, you know, they shouldn't have s...
Is Offense or Defense Dominant? FAR.AI's Adam Gleave on the AI Security Leaderboard
I think there's an argument that actually maybe too much attention is paid to universal jailbreaks because in principle, a very targeted jailbreak could cause a lot of harm.
Its share fell from 34.3% last week to 29.1% this week, reversing the previous week's rise.
Share of this week’s podcast conversation
Recent weeks31.9%→This week29.1%
Came up less often than in recent weeks.
Podcast moments5 cited momentsViewClose
No Priors
Building an Autonomous Enterprise for Real-World Services with Netic Founder Melisa Tokmak
that is many customers how they started, but today actually over 70% of our customers are AI first. We call it Netik first. They go N1, uh, and over 70%, actually all of their customers' first interaction with the company is with Netik agents.
Is Offense or Defense Dominant? FAR.AI's Adam Gleave on the AI Security Leaderboard
As a lifelong Detroiter, it pains me to say that Boko Haram is ahead of the Big 3 when it comes to AI adoption. I didn't think I'd ever utter that sentence. I don't know if you have a sociological take on how in the world that's happening, but it's a real puzzle from my perspe...
Real AI Transformation Costs HALF of Everyone's Salary for 2 Years | Chris Blackburn, Liatrio
You speak to a lot of CEOs or C-suites of And I know you've been critical of, you know, there's a lot of talk about AI transformation, but as you said, a lot of it's on the margins. When you go into an organization, what's your process and what are the things that you see over...
All right. And I think that the gap between input and output pricing keeps widening at the frontier. If you look at the Fable-5, it's about $10, not about, it's $10 per million input tokens and $50 per million. With the GPT-5.6 SOL, it's a 6x ratio. So we're seeing the gap eve...
Reconstructing how OpenAI agents attacked Hugging Face
I did not get that in the initial. So I don't think that was really known at that point. Which is really ironic when you consider the fact that Hugging Face was initially trying to use an OpenAI model to discern what happened with, unbeknownst to them, was an OpenAI model atta...
that's where the power of having so many open-weight models out there and the choice among models Yeah. It's so important because if you just have the nucleus of your stuff and these models can call upon it, it leaves you so much more in charge.
Well, yeah, I mean, NVIDIA basically leading off with this Open Secure AI Alliance, right? kind of collecting 50 companies, more than 50 actually. But, you know, you mentioned Palantir, Microsoft, Hugging Face, of course Salesforce, Cisco, Dell. There's a number of GitHub, SAP...
I think that's where a huge opportunity is going to be. And I think ultimately, that's going to mean more people are going to move into jobs that have been considered individual contributors, more builders, more people that are actually value producers versus, versus the overh...
Ep. 227: How Good Karma Brands Got Serious About AI and Made It Stick
That's so cool. I love that. Maybe could you give me a little more of a sense of how this role is going to work? I know it's very new, but the innovation specialists, so they're going to function full-time as people that help everyone within the company reinvent workflows, it ...
Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI
It's a career risk for me because I used to be in developer relations, right, where your job is to show. And then you're like, what do you mean you don't need— actually, your job is to tell. But the product people are like, well, we don't need you if our product is intuitive e...
In terms of our vision, we are building an autonomous enterprise, right? We want every single thing in these companies to be handled autonomously with Netik. Except the actual services and the labor itself, that human component, so that these companies can fully focus on that ...
Reconstructing how OpenAI agents attacked Hugging Face
I just have to find it, right? So if you have then access to the public internet, why not go on a search for the actual solution. So the model, or more appropriately, the agent inferred where might this solution be? Where are solutions to benchmarks? Where are datasets? Where ...
