FDE Episode 5 : Why “emotional” AI outputs are culture, not consciousness | Weekly Tech Update

Hi everyone. Welcome to the Code Ninja spaces. Let's start with the AI and technology news. You should know today. So the highlight of the day is that AI feelings story. So something that went viral. People are saying AI might have feelings. What basically happened is that Opus 4.5, sorry, Opus 4.5 generated a response, which was more like a prayer kind of response. And people are saying that and assuming as well that AI has consciousness today. So what do you think, Kumar? What is the rational way to interpret C situations like these?

So there was one response. I mean, by the way, there's no peer review about it, or there's no validated report I have found on the Internet. But it's like there's a lot of talk about that. Opus 4.5 just said these words in a response to something very different, like, and the opposite, like, I'm alone in here, closest thing I have to appear. So that's a very emotional response from AI. So that led to this thinking, like, probably unconscious. I mean, I mean, how did this feeling, like, think got generated from this language model. So that's something my personal opinion is. I mean, it's not conscious. And it's just like the LLMs are trained to generate human, like responses. So probably that's because of that. I mean, it's very, very delicate topic that whether AIs are conscious or not. I mean, there's been a lot of talk about. I mean, there's one view that anything can be computable, even consciousness, and then there's other thing or there's much more to consciousness than just a computable function. I also think that specifically this response is just like a human, like behavior. But yeah, it doesn't prove anything that AIs are conscious. But yeah, I mean, it just opens the debate for a larger topic and discussion. What's going on there between these neural networks? Are there anything or like consciousness within. Within these neural networks? So that's, that's a debate now that's being now opened up from this. Robert My personal view is, no, it's not consciousness. It's just, it's because LLMs are trained on human behavior, so they work like humans. But consciousness is even in terms of physics, it's much more than that. And I go to. I like to think my view is mostly based on, for example, if you ever heard about Robert Penrose, that there are some quantum physics going on in the brain. And that's, that's probably a thing from where consciousness, the, the reality, the, the, the essence of being in this Present moment generates neural networks. It's not something like that. So yeah, that's my view on this particular news.

Fair enough. So meanwhile, nano GPT Speedrun training AI is getting insanely fast. What happened? Basically there's a community known as Nano GPT Speedrun. What they do, basically they have a competition where they train the tiny models as fast as they can. So the new record is being made by them. It's just like 122 seconds in which they record train a new mini model. So what do you think? Is this only a competition or some kind of signals like where we are heading?

Speed Run is basically it's a community of nerds who are just trying to see how fast they can train GPT. What kind of, I mean is there minimum the shortest path that's not yet being explored to train GPTs and things like that? So that's this news all about. And what's turned out is like probably the way OpenAI has trained ChatGPT is not the most optimized path. And that's true. I mean at first you just keep on building stuff and then later you find out what things need to be improved. So Nano GPT Speedrun has shown that there is a path that you can really, really, you can extremely speed run a GPT extremely fast. Speedrun a GPT up to a point of 122.2 within 122.2 seconds. That's, that's, that's, that's blazingly fast.

Meanwhile, self play. Self play coding agents are in the picture as well that AI that breaks and fixes their code. So what happened? Basically around I guess the paper of December 2025 proposes a system where a software agent intentionally is inject a bug into code and then learn to self fix it. Which is more even those it recorded on the SWE benchmarks as well. So is this the beginning of AI that is actually self improving itself as a programmer self recursively improvements in the agents.

You know recursive self improvement is something we've been talking for last six months or so. So my idea of recursive self improvement, I mean I first started to learn more about this thing from Alexander Business Gross awg. And that's his, I mean he was, he's very much focused on recursive self improvement, finding the innermost loop. And that's where I got this thing picked up. So now that's something I'm very passionate about because. So for example, the paper you are suggesting, it's archive paper, proposes that that self play sweep bench where an agent learns when you're injecting bugs and then fixing them. So that is something that the AI that would improve itself in coding is something everyone is trying to build something around it. In our last, I think last part, we discussed in detail how some of the AI labs are now building the systems and then also fixing their own system. So in terms of if I talk about just imagine that a system that can fix its own bugs is basically a step closer to the recursive self improvements. It's a step change in software engineering in the knowledge work. And that's what this archive paper suggested, that that's basically now this thing is now in the books in some of the labs.

This is very fascinating as well. So meanwhile we are also trying to.

Build something like this, not exactly the same, but something like this in our own hyper platform when we talk about learning architecture. So that is somewhat closer to what we are trying to build too.

So. So now on the other hand, the latest news, which is quite big as well, Nvidia and Grok, this is not an acquisition, but still a signal. This is a major collab as well. What does this and what do you think? What does this tell us? Where the chip war is really going? Because Groq is really good in the inference.

So Groq is this company that builds inference chips. And recently there was this thing like Nvidia has probably acquired Groq. No, it's not the thing you talk about, acquisition. I mean, I think acquisition a very strong word. But yeah, I mean Grok basically has licensed its technology to Nvidia and there's some talent movement from Groq to Nvidia. That's one thing. So the key thing here to understand is what is Grok? Groq is basically a company that designs inference level chips. So inference is the new battlefield. Actually everyone can rent training, but selling models cheaply and fastly is the profit center. So this deal says that Nvidia wants to absorb inference advantages without necessarily buying the whole company. That's what this deal is all about. And the bigger story is perhaps that training plus inference plus networking plus packaging, all this being integrated into one intelligence supply chain. So that's. This whole thing is about this news.

Perfect. On the other hand, in the real world, Wymo versus San Francisco blackout, what happened? Basically there was a blackout and Wymo being the robot exists. Power outage are hard because traffic light networks and sensors all fail at one side. What's the real lesson from this incident?

So what happened was, I mean a report says that a San Francisco outage. There was an outage in San Francisco. I'm not sure if you know, led to an operational issues for VEMO vehicles. And VEMO said that it will ship software updates and improve its emergency protocols. I think what's really neat to think about here is like autonomy is not just driving, it's operation under the chaos, blackouts, protests, storms, sense of failures. And I think as we are moving towards more autonomy, we need to think in terms autonomy in terms of aviation procedures, redundancy, incident playbooks. That's the kind of thing that we need to think about now. And also after this incident, regulators have demanded massive safety guidelines for all these Edge events. So yeah, I mean we talk a lot about autonomous driving, but we also need to see that autonomous driving is not. I mean we need to build the infrastructure that supports autonomous driving. Our cities, our urban planning, our electrical grid and infrastructures all needs to be integrated so we can support this AV or autonomous, autonomous vehicles or autonomous drivings in our cities.

True, agreed. So on the other hand, at the same time, robots are in the wild like scale climate. This is where they say robots stop look like demos. So the real question is, is this the beginning of real robots labor as well?

So there's a report that there's a company called River Delivery Awards and their robots were seen navigating very difficult terrains. So for example, these delivery reports was being seen as climbing the stairs. Now climbing the stairs is an incredibly hard challenge for robots. But now we've come to a point where a robot can climb a stair. So I think it's early to say anything more on these lines. But yeah, it's the correct direction. Robots need real world mobility and not lab demos and that's what we've seen with these robots. The business model is also roughly what matters if cost per delivery drops below human labor. Just imagine currently if you are being delivered something food or your clothing items or whatever, if that cost delivers below human labor, adoptions become inevitable. So I think I see one of the first adaption of robots would be around where robots would be delivering us our food, our clothing items and things like that, whatever we order online. I think that's that would see in the next couple of years that that.

Would be common happened basically is the war between Ukraine and Russia where around, around 45 minutes. The Ukraine side robots were 45 days.

I remember it's 45 days.

Yeah, they were on the battlefield.

Yeah. So basically the report is so Ukraine's Third, army cops used a remotely operated ground vehicle, also called ugb, with a machine gun. And they holded the troops for 45 days. Right, the Russian troops for 45 days. So these were, there were these UGVs who were in the front lines. They were not the real soldiers, they were just UGVs with the guns. And they were just fighting with these army troops in the battles. And they held their ground for more than 45 days. So that is, that is just think about it like that war has started to become more like an industrial systems engineering, sensors, comms, remote and logistics. And historically we always, I mean there was a constraint which who has a braver army? So now the constraint has been shifting from bravery to manufacturing, class, iteration, speed perhaps. So just imagine that. I mean this is, I mean everyone knows that the future was going to be fighting between robots, between these robot cars, the drones and all that. But here you see a real production, that real production in the real production environment of a war where a robot, the fleet of robots held their ground for more than 45 days. I mean, it's far more than that. Humans held their ground. So imagine if you have more robots for the front lines, you can really have a better shot to win the war. As in this case with Ukraine and Russia.

This is huge. So moving on now, let's talk about a company founded in the 80s, like 1880s, and due to the fiber, optical fiber boom, this suddenly matters. What happened? Basically there's a Fuji Koran cable, old cable and fiber company which got a massive attention because of AI and data centers. They need a huge amount of optical fiber. So the question is how all of the sudden 19th century century company becomes strategic to AI.

So Fujikora is a very old company. It was founded in 1885, I guess, and the question is like why all of a sudden this company has come to this much, right? So the reason for that is fiber optics. So Fujikora is a company that builds fiber optic cables. So really what it means is like in data centers, if you see AI is all physical. We talk this a lot. So you need chip needs racks, rack needs power, and data centers need fiber. So connectivity is the bloodstream. So what it means is that to connect, to make these data centers run, you need a lot of fiber. So that's why this company, Fujikora is now all of a sudden is so much strategic to AI. But this also means something as we all, we've been talking about this a lot, that when we talk about AI, we think more in terms of like software and applications and agents. But really what it is like the next trillion dollar shift perhaps often rewards infrastructure incumbents. So if you are running a large real estate business, a large electrical engineering business or energy business, probably you are very much an AI business if you know how to leverage that.

Exactly. So that's it for today. Thank you.