FDE Episode 4 : AI models that can audit other AI's from the inside | Weekly Tech Update

Everyone, welcome to the Code Ninja spaces. It's our fourth episode, so let's start with the first news of the day. There's a news that model that psychoanalyze their own internal state. What does it mean? Anthropic says that they trained something known as the activation oracle. It means that model can now look into another model's internal activation to get to know about their hidden surface level knowledge and even risky intent. So I have a question here. Do you think are we entering into an era where AI can audit AI from the inside?

So first thing is like what are activation oracles? So think about any language model and think about like there's an internal state of language model. But just because of, we want this model to behave in a certain way more because of alignment or because of some, because we want to feed explicit morality into the, into these language models. So we just fine tune them with lots of data. So these activation oracle means just like what? I mean if there isn't any alignment through fine tuning, what's the core of, what's the core of that thinking process that these neural networks generate? What it really means is like, think of like you train a kid, for example, if you're going to, let's say with someone, you bring your kids along but you want them, the kids to behave in a certain way because you don't want other guys to know that the kids, for example, the kids in the homes are the way they are here, the way they are here the day they are behavior behaving here. So but that's not the internal state of that kid. That kid is in their home. They are naughty, they do lots of stuff that probably society doesn't deem right. So there's an act, there's an actual state of that kid's model and then there's a fine tuning of and then there is a certain behavior you explicitly train them for. So they behave in a certain way when they are out, let's say meeting for meeting with your extended family members and stuff like that. But that's exactly what Anthropic says. That's exactly happening with language models. Their internal alignment, their internal thinking process. And that's different than what fine tuning of data makes them do. So what they call it risk intent. So what really activation oracle means is to go deeper into these language models and really try to see that if there wasn't any explicit fine tuning of these models trying to behave make them behave in a certain way, if that wasn't the case, what would have what different they would have done. That's a very important thing to think about. Like I mean all those question of latent risky intents or mechanistic accountability. Elon talks a lot about it like maximize truth or think about it like what is truth? Right? One thing what truth is not is if you are continuously fine tuning model to behave in a certain way. What it means is it generally means that it's not truth, but truth is what model. Let's say if model had some thinking, then truth is the way model thinks about the world, about us and all. So that's a step towards that. It's a very tough path activation. Let's see what very tough concept to even explain. I mean I'm finding it hard to explain because beyond certain point you just need to use a lot of computer science to explain. But yeah, so at the end of the day think of it like that there is one state of the models that we achieve by explicitly, let's say programming morality into them or by just fine tuning them to behave in a certain way. But if there wasn't, there isn't that layer of fine tuning, that explicit programming of morality, what model truly believes, what does it mean? I mean that's very important to see. And I think the path towards super intelligence or say superintelligence is it's very important to know that does these models have a thinking of their own and does that thinking of their own align with how we think about them? So it's a really tough path.

So great, we have a second news of the day which is more about related to game playing. So a journalist game playing agent is trained from Internet gameplay. What happened? Basically there's Nvidia's introduced new model known as Nitrogen, which is even Those trained on 14,000 videos from Twitch and YouTube across I guess around 1,000 games from which the success rate was about 52% which is quite cool. And this is similar to the robot training as well. So the main question here is that Nvidia's Nitrogen seems to learn control policies from massive gameplay videos. So is this robot training without robot data?

So one thing what happened with large engine models were that we had this in Krabbles vast amount of data repository in the form of Internet and we just that just made these language models work in a way that just made sense and give us the illusion of intelligence. Similarly, I mean every company is trying to find out is can something serve as something as Internet for robots just like the Internet served served to language models. Is there, is there in terms of Videos in terms of some content that just we can feed to these robots and they behave just like humans. Right. So for example, in terms of cognitive labor, you've seen that large in cognitive labor, but can in physical labor can something is already out there that help us achieve that level of performance for robots. And so, so nitrogen is also an attempt to that what's really important here is it's a model that means trained by just what it called it learn by watching. Like they've shown 40,000 hours of YouTube content and achieved certain performance goals. So imagine if whole of the YouTube data can be fed into these models vision action models and then these vision action models are then being planted into robots. So would robots would have the same level of physical performance that what humans now have. So that's an attempt towards that. I think we're far from there from achieving that level of physical performance in terms of robots. But yeah, I mean if this works out, I think then probably next five years we would have for example lots of humanoids that would be performing just like humans. And yeah, that, that's, that's. But I think there are lots of computer science breakthroughs required currently. But yeah, that's, that's a, that's a news. Let's see.

Yep, let's see. Really interesting. So moving towards the third news of the day, which is more about training loops collapse and basically your question is that why should a normal listener care that a model can be trained in two minutes? So basically what happened? There's a Nano GPT speedrun community which reportedly record about I guess 127.7 seconds to reach a specific training target. Which highlights that models can be provided a very specific elite hardware that can speed up their speed of training data as well. So what do you think?

I think again we think that probably intelligence is cheap, it's very expensive. Costs of executing training runs are extremely expensive and we need to lower down the cost of tokens. So just think about if for example we need to solve something that's incredibly requires lots of AI or token usage, let's say predicting whether or what kind of ingredients in weather we need to change to make weather more climate friendly. So tasks that are extraordinarily hard to achieve. It requires that we make cost of intelligence to near zero. But currently it's not that accretion is currently is training larger model is expensive. So what this shows training loop collapse means like iteration cost is collapsing when training to competence for example become minutes. Experimentation explores more tries faster discovery. Exactly. I mean you can jump to years ahead by doing lots of iteration, lots of discovery if the cost is low. So that's I think that is a very significant, I think progress. The training loop is the cost, training loops is collapsing. Imagine coding or other businesses starts building their own training runs. That's only possible if it's not 10, $15 million which is like for example if a half a million dollars for $100,000 you can train a model. It's pretty cool. And think about it, what can be achieved in healthcare, in climate if the cost of fin gets collapsed?

So we have the fourth news of the day which is related to the biology and predictive videos. So what happened? Basically MIT reported deep learning models now can predict minute by minute cell behavior during early, you can say fruit fly embryo development with accuracy around 90%. This is huge in even those certain experiments. So basically the question here is that from this news is are we heading towards simulators of life in a way that we build these simulators of weather.

This is so cool. I can tell you, I mean just imagine, I mean what this mit what the direction that is such as that MIT is taking is like you can actually predict cell division minute by minute, second by second. You can do a lot of interventions. So imagine solving cancer, right? Which truly the cell division problem. So I mean I think that's where I think most of the AI resources should go. And you know this is my passion too to trying to learn about this stuff, AI or structured biology. And that's something that I'm very excited about. Let's see. I think this is the first time ever that we would see that these things would move out from the sci fi space and into the real world.

So interesting science like progressing very rapidly. So we have another news of the day as well which is related to the robotics and assembly work. So the bad news of that is like there's a lab which demoed a robotic fan performing really fine doing mechanical work like threading labs. Oh okay.

Yeah.

With the in hand manipulation. So there's a basically question related to that. Is factory automation finally catching up with the software automations?

Like I think the problem in automation, this problem is called last mile dexterity gap. So last mile dexterity is basically, I mean just think about it like can you make robot hand as dexterious as human hand? So for example, if I am holding this cup of tea and I'm taking it right, it's very natural for my hand to hold this. I can just feel it, right? This is dexterity, right? But imagine if the robot has to take this, you would see some friction, some, I mean some gap in holding onto this cup. It won't be this natural and this smoother. This is called dexterity. It's in robotics it's the hardest problem to solve. Dexterity. I mean Elon said like it's the extraordinary hard engineering to make robot work the way robot hands work, the way human hands work. And think about it like we take it for granted and that I mean, I mean everyone, I mean every. I mean if you're fed, we take this holding the cup for granted but that's not for robots. And this problem is called last mile dexterity. And what this current results shows that we on a path, I mean I've seen that but I think we're not close to this level of dexterity. But yeah, we are on a path. If we do that then think about what kind of factory automation do we need would be there. So I mean at the end of the day our factories are all already automated. I mean if you see smart manufacturing in China, if you see manufacturing in even in this part of the world in U.S. most of the people in the factory, most of the factories are already automated. But what really means is currently is like for the people you see in those factories in the last mileage strategy there's some level of human hand required. Not cognitive mind but human hand is required to for example perform certain operations. And if you automate, if we automate that hand then probably we don't need those people. Which I think is very important point to see that what kind of job displacement are we looking at here? So if we work, if we work, we make this dexterity work. There's a serious crisis then so perfect.

That's the last news of the day. Which is more related to the browsers agents. Recently we have been looking towards the trend and anti gravity of the Google announced as well that their Gemini 3 going to coverage more towards the browser agent. So do you think are browser agent becoming a real workforce as well?

So for agents, let's say for agents doing long horizon computer tasks that's everyone is trying to achieve that. Like can we make agents? Let's say do lots of can agents in the browser buy for example my shoes or do shopping for me and all that. Yes, it is possible. You remember in our last episode we discussed about Itzy right? So yes, that's already happening but only when there is a structured data available in that domain. If you have some level of structure around data around domain then only yes this is possible. Why structured Data is important again in case of antigravity because we need to make sure that agents, I mean to achieve this level of long horizon computer tasks, we need to make sure that agents are very good in tool use. Right. Alexander Wang said this like we need to make sure that how to make the agents use tool maximally. If agents start learning how to use tools maximally. Yes, then. And then that, then I think if that happens, knowledge work will be cooked anyways. There won't be any knowledge work left. Because currently if you see most of the people behind the laptops and our computers are doing not a cognitive labor again, it's tool use. Exactly the same way as we just described that in the factory. The current factory workers are doing last month dexterity work. They're not doing physical labor processing physical labor is already automated in the factories. Even so. Yeah, I mean I think that's what everyone is trying to achieve. But it's very important. Rafa, I want to mention imagine we solve last mile dexterity on logarithm computer tasks. I wonder why people say there's no job displacement. I mean it's just around the corner. I mean because if you see the CEOs of these, all these labs they are saying is we're going to automate all that. And the amount of money that's being poured into these labs is only justified if economics value they generate is through automation and not through human. And they know that, that that's only possible. The returns are only possible if lots of this work would be automated. And that's going to happen in next two years. We'll see that agents will be good enough in tool use and our robots probably next five will be good enough in large mind extractive.

So yeah, great news. So that's it for you today. Thank you so much. Let's wait for the another episode on Friday. So.