The AI Cow

00:00:00: Hello, fellow humans and welcome to Human Core AI.

00:00:03: Have you ever seen an AI cow?

00:00:07: Maybe? No?

00:00:08: So there was this AI generated cow posted by Dr.

00:00:15: Singularity on X, which basically looked like you would take random, I

00:00:25: don't know, 12 pixels and just try to create a cow with that.

00:00:29: So it really looked like something out of the 80s with very bad quality.

00:00:37: And then if you think about what the latest tools that create images can

00:00:47: provide you, like, for example, Image in three by Google, it's pixel perfect.

00:00:54: It's photo realistic.

00:00:56: You basically cannot tell a cow generated with AI in 2025 from a real photo of a cow.

00:01:07: It comes to show how far we've come with this entire development of AI.

00:01:14: I'm so amazed when I see stuff like this, people randomly post on X.

00:01:18: And yeah, it's pretty amazing, exciting, but a little bit scary as well to see

00:01:28: that much progress because like if you go from 2014, which clearly is some kind

00:01:34: of first ever try maybe to what you see in 2025, where the AI generated image is

00:01:42: pretty much indistinguishable from reality.

00:01:45: Like if you are very, very, very nitpicky, you might pick up something

00:01:50: that gives you a little hint, a little clue, especially when you generate human avatars.

00:01:56: Although I have had large success with, for example, Image in three from Google.

00:02:02: It's just breathtaking.

00:02:05: Sometimes it generates crap ish images, let's say, not crap anymore crap ish.

00:02:11: But most of the time it's just mind blowing.

00:02:16: So yeah, we are here.

00:02:18: It's February 2025 and the AI world has always delivered.

00:02:25: And let's go through some of the stuff that I think is the most interesting

00:02:30: that has happened and see what is going on.

00:02:37: So I think the biggest news by far in the last few days has been the release

00:02:44: of open AI's deep research.

00:02:47: Deep research is an AI agent that open AI has released that allows to do

00:03:00: research on a much, much bigger scale than it has been possible until now.

00:03:08: So if you are familiar with the offering of different models by open AI,

00:03:14: there have been a few models released very recently, but that allow for deeper reasoning.

00:03:22: And, for example, one allows to reason deeply.

00:03:27: So it's a reasoning model, as they call it, and it can do deep analysis,

00:03:33: but it cannot connect with the Internet.

00:03:36: So you have to give it all the input that it needs and then it could sink.

00:03:39: So you can still use it as a very good tool for research,

00:03:43: but you have to provide the sources yourself, which is still pretty good.

00:03:47: And results are very decent, especially if you have access to one pro,

00:03:52: which is the version behind the permanent paywall,

00:03:57: which is, in my opinion, well worth the money if you do anything business like

00:04:03: it can just very quickly save your money or provide you so much valuable input.

00:04:08: You can offset the 200 bucks.

00:04:09: I think it's a bargain, but let's just leave it at that side now.

00:04:13: So we have 01, 01 Pro.

00:04:15: And then recently we got 03 Mini, which has the ability to connect to the Internet.

00:04:23: So it's just web search, basically.

00:04:25: And with 03 Mini, you can already do a pretty good research, I would say,

00:04:32: because you can type in, like, hey, figure out this, figure out that.

00:04:36: But the results still are what you would expect from a traditional LLM,

00:04:43: even with reasoning.

00:04:44: But it's like a step further.

00:04:47: And now with deep research, people believe that we have received 03,

00:04:53: the Model 03, but packaged as deep research.

00:04:57: So for this specific use case only, not for a general purpose,

00:05:02: let's say, and what can deep research do?

00:05:05: So basically, it's probably 03.

00:05:08: So likely the smartest LLM that is available out there today.

00:05:15: Plus combined with web search.

00:05:18: Plus, I think with some more allowance by OpenAI to actually explore and search

00:05:28: and analyze.

00:05:29: So what it does is you tell it what to research and it will go off and,

00:05:36: I don't know, do lots of search, ask different search engines.

00:05:41: Probably maybe just one search engine, but it asks different questions.

00:05:45: It collects information and tries to figure out what are the trustworthy

00:05:49: searches, maybe the reliable sources for this kind of research.

00:05:54: And then you can follow it a little bit in the UI.

00:05:59: You can see how it goes through the different pages, scans the content, reasons

00:06:05: about it, then puts it all together and then still analyzes the content and tries

00:06:11: to put it into a format of your liking.

00:06:14: That process can take a few minutes, a few minutes even.

00:06:19: So it's not something that is done in a few seconds as the previous models.

00:06:25: This one can go and take maybe five, 10, 15 minutes sometimes.

00:06:30: So if you take and put a human to do the same kind of task, it's impossible

00:06:36: to have it done this fast.

00:06:39: And with that degree of research, you can get a much better research.

00:06:46: And of course, you should check the results always, as is usual with LLMs.

00:06:51: But for like an 80/20 Pareto principle, like 80 percent of the result

00:06:59: can be done with 20 percent of the time.

00:07:01: This is the exact tool to use.

00:07:04: So I have been doing lots and lots of researches over the last few days.

00:07:10: And I have had, I would say, mixed results, but they were going from neutral

00:07:17: to amazing.

00:07:18: And basically, if you think about it from a user perspective, like people

00:07:22: who use Google, right, search engines, when you're interested in figuring

00:07:31: out something like, for example, you want to invest some money so it does

00:07:36: not sit idle on your bank account.

00:07:38: What do you do?

00:07:40: Well, Google, you search through different blog posts, you read them, you analyze.

00:07:45: Maybe you watch some YouTube videos.

00:07:47: This kind of stuff.

00:07:48: And by the end, you might have some idea or vision in your head on about this

00:07:55: topic of investing, and then you might just go and do it with deep research.

00:08:02: You just basically tell it what you want, which by the way is a very good thing to

00:08:08: do, because when you put down your thoughts on or requirements on what you want,

00:08:14: it makes it so much easier to actually know if the thoughts in your head are

00:08:19: clear enough so you get good results.

00:08:21: So I always write down the prompt and then you set it out and then you wait for

00:08:30: the results.

00:08:30: But believe me, it saves so much time because deep research will go through

00:08:37: all the trustworthy sources that you might not be able to find that quickly.

00:08:42: And it scans them all and then it puts it in the right format.

00:08:47: And you can be creative with this.

00:08:48: I have been very conservative with the format and structure because I wanted to

00:08:54: get some research for my business.

00:08:57: So most of the time I would ask it to do like a traditional research, but I've

00:09:04: seen people use it for much more creative stuff.

00:09:06: So I guess there is still some space to experiment with deep research.

00:09:13: But honestly, what deep research accomplishes in 10 minutes cannot be

00:09:20: accomplished by a human the same level of quality.

00:09:24: No chance.

00:09:26: And this is, I think, the scariest thought of all.

00:09:31: Moving to much less impactful news or maybe speculation.

00:09:38: Sam Altman has been to Tokyo, I think.

00:09:41: Some people from OpenAI have been to Tokyo because the deep research was

00:09:46: actually announced from Tokyo, which was pretty fun because they announced it,

00:09:51: I think, you know, daytime in Tokyo, which means that the rest of the world

00:09:56: basically woke up to the news on a Sunday, which was the thing.

00:10:02: Time zones are funny till you go travel and then have a week of jet lag.

00:10:07: Story of my life.

00:10:10: So Sam Altman has said something in the lines of that there might be some work

00:10:20: being done on a device, a native device, let's say.

00:10:27: And that is, I think, the interpretation that people did out of it that could replace

00:10:33: smartphones or specifically the iPhone.

00:10:35: So I think in the tweets that I read, they were saying terminal.

00:10:41: I believe this is just a generic term that you can use for a device of unknown

00:10:47: properties, though.

00:10:48: But I think the interesting part here is the trying to spin up our imagination.

00:10:57: And think, OK, what is it that OpenAI or a different player on the market could go after?

00:11:05: So obviously the killer product, the most successful product of the last 10 years,

00:11:13: 15 maybe even, has been the smartphone, especially the iPhone being a very,

00:11:21: very high and lucrative product.

00:11:24: Making Apple become basically the biggest, richest company in the world.

00:11:29: And the same applies to Android, but for a different kind of public.

00:11:35: But in the end, smartphones have changed the way the world and society looks and works

00:11:43: and communicates.

00:11:45: It's been a huge game changer.

00:11:47: So I understand that many people are thinking, what is the thing coming after smartphones?

00:11:53: What is the next evolution?

00:11:55: Because, like, let's say no one is assuming that the smartphones are the end of the line,

00:12:00: the end of the game, and that no other device will ever come close to it.

00:12:04: I don't think that is plausible.

00:12:07: AI, especially latest LLMs, have a chance to compete with the smartphone as a device.

00:12:15: How is that going to be possible?

00:12:17: Well, here we start with the opposite side of the thing as compared to iPhones.

00:12:23: iPhones or smartphones, smartphones are primarily hardware devices with some good software on them.

00:12:32: But then with AI, you start with the brains.

00:12:36: And then you have to think back to, OK, I have the brains.

00:12:40: What does it mean in terms of hardware?

00:12:42: Or maybe is hardware even that relevant when we talk about AI?

00:12:48: Could we imagine AI being just present and not needing any hardware and devices?

00:12:56: Well, I think that's not a utopia.

00:12:59: It's more like a dystopian view on things because we quickly come to the conclusion.

00:13:04: Oh, yeah, that might require some implants, some chips that will just project images into our brain.

00:13:10: But that sounds a bit scary.

00:13:12: I think humans are not there yet.

00:13:14: So probably there will be some kind of hardware that you can at least decide when you want to use it and when you don't want to use it.

00:13:20: I think the glasses idea that Google tried maybe 15 years ago might still be relevant in this day and age with very smart brains and sleek hardware that is not very intrusive.

00:13:37: That sounds like a very good combination.

00:13:40: And maybe that's where stuff is going and maybe OpenAI is seeing an opportunity here to get into this market.

00:13:47: Although, honestly, OpenAI as a company is so far away from any kind of hardware manufacturing that I don't really think that OpenAI is going to go there.

00:14:00: Unless there are some plans that we know about and OpenAI wants to expand to become this mega-corp, one of the biggest tech giants in the world.

00:14:10: I mean, these are becoming a tech giant just by focusing on building the AI and LLMs.

00:14:17: Maybe they are more ambitious than that.

00:14:21: We don't know.

00:14:21: So let's see how that goes.

00:14:23: Interesting topic, but very speculative as well.

00:14:26: So we don't really know what's going to happen.

00:14:30: Maybe I should spin off a separate episode about this topic.

00:14:35: Let me know if that's something you are interested in.

00:14:37: Why Combinator is entering the game?

00:14:41: It's not like they have not been close to the AI developments, but they have been publishing some stuff, focused a little bit on their own interests, but still very interesting to see.

00:14:54: For those who don't know, why Combinator is a startup incubator, a very powerful one, one of the best in the world, I guess, one of the best known for sure.

00:15:04: Very VC oriented, so venture capital.

00:15:09: They have the network, they have the people, the expertise and usually startups that go through why Combinator have much, much bigger chance of success.

00:15:20: And some of them have become like really huge unicorns.

00:15:24: So why Combinator has dropped a few recommendations and observations on the market?

00:15:31: Because they want to, let's say, give founders or potential founders a little nudge into the right direction.

00:15:40: So the direction that they see where things are going.

00:15:43: And I think it's interesting to observe one of the largest incubators.

00:15:47: I guess there are tons of, tens of smart people there who spend their entire days thinking about this stuff.

00:15:56: So, you know, getting inspired a little bit by these people is good if you are into knowing what is going to happen next.

00:16:04: So maybe if you have some kind of business idea for the future, looking where stuff is going, where the major VCs are going is a smart thing to do.

00:16:14: Even if you decide to bootstrap or go in a different direction, it's still wise to know where the industry is moving as a whole.

00:16:21: So why Combinator basically is focusing the entire pitch on vertical AI agents.

00:16:30: So let's just recap what those terms mean.

00:16:35: Agents, AI agents mean AI that is more autonomous, that has more agency than the traditional LLM.

00:16:43: So AI agents can do tasks, they can interact with software, they can do stuff on the internet, they can make more advanced decisions in a more autonomous fashion.

00:16:56: And they don't need to wait for input to do stuff, they just get a mission, a task to do and they go.

00:17:03: And if they are programmed correctly, they will come up with some results and multiple agents can work together.

00:17:10: That's like the spirit here.

00:17:13: And vertical AI agents mean that why Combinator is going after...

00:17:21: specific niches and industries that can be disrupted from top to bottom or from bottom to top entirely.

00:17:30: What is a vertical in the space like Uber is a good example.

00:17:35: You know, the taxi driver economy that could potentially be disrupted by AI as well or maybe Airbnb to name very few very well known names.

00:17:52: So Airbnb is also vertical in the rental space and renting, I know, vacation or renting, I guess.

00:18:00: And so Y Combinator is seeing potential for creating software solution, AI agent systems that can potentially become much, much bigger than the biggest SaaS businesses today,

00:18:22: SaaS being software as a service because they can not only replace or automate or boost software, they are not only providing software and tools,

00:18:40: they can also replace labor costs. And that is I think that the biggest game changer here. And if we think about the economic impact of AI in the future, that's exactly the core of the topic.

00:18:57: So we are not speaking anymore about purely tools and software solutions. We are speaking tools and labor attached to them.

00:19:12: AI agents are the labor force as well that uses certain software and certain tools in a certain area.

00:19:21: So that means that suddenly businesses, startups even get much, much more detached from the cost factor, which is labor, humans.

00:19:38: It's a fancy way of saying, hey, humans are not going to be the slowing element anymore. They're not going to be the blocker in terms of scaling.

00:19:50: So that's at least what Y Combinator is predicting that by combining software and labor into one AI system, computer system that can scale vastly much more than any traditional business could.

00:20:08: And that's what they are betting on. They have released some more findings and opportunities. I think it's a good thing to read about it.

00:20:17: They have provided some success stories. Very, very interesting stuff for sure.

00:20:22: One very cool thing that crossed me on the Internet in the last few days was a tweet on the topic of hiring.

00:20:32: And for the first time, and unfortunately turned out to be fake, but for the first time I saw a job offer that was specifically targeting AI agents.

00:20:46: And it was more of a playful illustration of where this is all going.

00:20:54: You know, what it's probably not going to happen this way exactly, but it made me think because at first I thought, okay, this is probably true.

00:21:03: This is what the world's going to look like very soon, which means businesses might be hiring AI agents,

00:21:13: which would mean that some people will build AI agents to be hired by companies.

00:21:19: And it's an interesting thought game here. If this is where it's going, I guess it's not really because when you build an AI agent,

00:21:29: you're not building it decoupled from the world. It's like a human born in a place and growing up and then deciding, hey, I'm going to study this.

00:21:38: As we know, AI agents work much differently. They're much more integrated into system or can be integrated into systems that interconnect and very quickly exchange information.

00:21:52: And also they have a much broader knowledge base, let's say.

00:21:57: So it's a bit of a different story. Normally, I think a company would never hire a specific AI agent.

00:22:04: Most likely they would use a system that has different AI agents to use.

00:22:11: But hey, maybe just maybe some niche topics with very, very deep expertise that is not available to the average AI to the average LLM.

00:22:23: So the training data might not contain that level of detail.

00:22:28: Maybe there is an opportunity to have very specialized AI agents that are 100% optimized for one specific task and they're the best in the world at it.

00:22:40: Maybe. I'm not sure if that's the case. Probably not going to happen for your regular replacement of a designer or something, but maybe for some very specialized areas.

00:22:51: Who knows.

00:22:52: What is interesting to me as well as in this thought experiment, of course, is the economics.

00:22:57: Because in the job offer, they say the company will pay $10,000 per month for hiring this kind of agent.

00:23:07: But I think that, of course, the difference here is that an AI agent can work at least for three times as long as a regular human.

00:23:20: Or maybe if we think about it, not in terms of how many hours they can work, but how many hours per day they can work.

00:23:31: So first thing is one agent can work 24 hours a day.

00:23:37: Not as a human that can work maybe eight to, I don't know, some crazy people, maybe 16 hours.

00:23:43: But let's say normally between eight and 10.

00:23:47: In productive work, that's still, I mean, the AI does not go get a coffee or have a break or go to the loo.

00:23:55: So there is a big difference.

00:23:57: One agent is already capable of working 24 hours per day, seven days a week, no days off, no vacation, no sick days, unless servers go down and stuff that could happen.

00:24:10: So let's say AI sickness might be a thing still in the modern age.

00:24:15: But then there is this thing that, well, you can have one AI agent, but you could have two, three, four, five doing exactly the same thing.

00:24:24: They just, you know, you add compute hours, working hours.

00:24:29: Some things cannot be done in parallel, so they have to wait.

00:24:34: There is some time component here.

00:24:36: But in many industries, being able to scale up the same person, the same worker, infinite times, that's a huge game changer.

00:24:49: And the 10K per month sounds like a lot of money, but it's still a very good deal if you think about this.

00:25:01: I think AI agents will get much, much cheaper than that.

00:25:05: I think labor costs in terms of AI agents will go down to pay 500 or 1000 bucks per month for most of the AI agents.

00:25:15: So that's going to be the economical change.

00:25:18: We're going from 10, 20, 30K a month for a very specialized or maybe even more specialized jobs down to, hey, this costs like 500 bucks.

00:25:30: And this is going to enable things that were not possible economically until now.

00:25:35: And that is one of the sources of major transformation, I think in the world.

00:25:40: Things that were not possible to do economically now become economically viable.

00:25:48: And that is a massive game changer.

00:25:50: And too big of a top for this single episode.

00:25:53: So let's see where this is going.

00:25:56: But yeah, definitely I saw after reading it that this is not a real job offer.

00:26:04: But let's see, maybe we'll see soon a real one.

00:26:07: Sam Altman published a new blog post about what he says, three observations.

00:26:15: And the three observations are first that AI intelligence is scaling logarithmically with resources used.

00:26:26: What does that mean?

00:26:27: It means that the more resources you throw at it, the more it skates, right?

00:26:34: It scales even more than just a linear.

00:26:37: It's really getting to a point where you can throw money at a problem and still gets better and better.

00:26:44: So that's like the first aspect of it.

00:26:47: AI costs are falling 10x every 12 months.

00:26:53: And he compares it to Moore's law that says that compute power doubles every 18 months.

00:27:00: So 10x less cost every 12 months is a big acceleration compared to Moore's law.

00:27:08: So we are actually accelerating in technology and development of technology,

00:27:13: which is mind blowing.

00:27:17: And finally, the third point or thesis here is that value of increasing AI intelligence grows super exponentially.

00:27:29: And here, the key point is, I believe that with more intelligence, more advanced AI,

00:27:38: we are going to get to the point where the AI itself is improving itself,

00:27:44: maybe research and scientific discovery.

00:27:47: We are not just humans, but also AI and LLMs will discover new things,

00:27:54: accelerate the development of technology even further.

00:27:58: So investing, I think the logic here, if I understand correctly,

00:28:02: is that if you put $1 to work at increasing intelligence,

00:28:06: that $1 will have a percentage return that becomes higher and higher and higher

00:28:13: as more intelligent the systems get because that one invested $1 can create new dollars

00:28:19: and more than just a percentage or a fraction of the $1, it can maybe generate $10, $100 even.

00:28:26: So I think that's this kind of snowball effect that, in theory, the development of AI can provide.

00:28:34: So I recommend reading it. It's a very interesting thing.

00:28:37: And if you are in the space of AI, it's interesting to know what the person who might have seen the future wants to share with us.

00:28:47: And to wrap up this overview of the last week and the things that caught my attention,

00:28:55: I saw a tweet again that was posted by Rohan Paul and he said Hollywood is on the brink of massive change.

00:29:08: And then there was a video attached with sequences that were generated with AI.

00:29:15: Back to the initial topic of the cow, what it has been happening with textual creation generation with images

00:29:26: is also going to get to the space of video, which seems a little bit computationally more expensive.

00:29:32: That's why it's a bit slower.

00:29:34: And probably it requires a bit more dimensions because, you know, visuals are more difficult to trick than text.

00:29:44: And they're more complex, more information dense, even.

00:29:48: Video is one of the frontiers that also is slowly getting to the point where AI can generate video that is indistinguishable from reality.

00:30:02: Tooling is still not there, so it's not like you can just sit down and say, "Just create a movie."

00:30:08: But I think very soon we will get to the point where amateur filmmakers might be able to create movies, series,

00:30:22: that get at least visually to the level of Hollywood.

00:30:28: And what that means again in speculation time, but imagine a world where independent creators

00:30:37: without leaving the comfort of their home can create new blockbusters.

00:30:43: The biggest and greatest, whatever, I don't know how I was thinking about Titanic.

00:30:48: That's how old I am.

00:30:50: And Titanic is still not a very spectacular movie, but maybe I don't know, Avengers and stuff like that,

00:30:55: or I don't know, Star Wars, like movies, or Dune, right?

00:30:59: Imagine in a few years from now, a person sitting at home, similar to a writer writing a book,

00:31:11: can create a movie that has the quality of Dune.

00:31:18: Isn't that mind-boggling? Isn't that blowing your mind that you could think about it?

00:31:26: But that's probably where we're going with AI.

00:31:29: There was going to be a transition phase where we go from Hollywood,

00:31:37: an exclusive on this kind of blockbuster movie,

00:31:42: to Hollywood using AI to produce cheaper and better,

00:31:46: to slowly and gradually independence getting into the space and mark my words,

00:31:52: and it will be the day where an independent creator creates a blockbuster that is.

00:32:00: At the level, or even surpasses many of the blockbusters created traditionally,

00:32:07: I cannot wait the day where I can create, for example, movies or cartoons for my kids.

00:32:15: I expect AI to be fast enough, so I still can create cartoons for as long as my kids want to enjoy them,

00:32:22: but if not, then I might want to create movies or something for fun even in the future.

00:32:30: And if I think about how this can change and disrupt another industry,

00:32:37: how many industries have we seen that have potential for disruption already?

00:32:41: Maybe I'm a little too slow to this game with my observation,

00:32:45: but Hollywood is not going to be the same in five or ten years from now.

00:32:51: They might now enjoy a few years of bonanza and being able to cut costs

00:32:56: and generate stuff faster, cheaper and better,

00:32:59: but at some point, the economic value of that creation is going to decrease

00:33:05: as independents get into the game with a great story,

00:33:11: because that's something that people have been saying, like Hollywood lacks the good story.

00:33:15: Well, if we lower the barrier to entry to people,

00:33:19: give access to Hollywood-style blockbuster creation to people who sit in their bedroom,

00:33:30: maybe with just a laptop and no pants on,

00:33:34: it could be massive.

00:33:37: So yeah, I'm very excited about all this stuff and small news like this,

00:33:41: small tweets even sometimes spark much more or boost my imagination much more

00:33:46: than those big announcements from big corpse about AI.

00:33:52: Sometimes it's the small things that matter the most.

00:33:56: Thank you very much.

00:33:57: As you have noticed, I'm still figuring it out.

00:34:00: I'm still learning and collecting just my thoughts and presenting them the best I can.

00:34:05: So if you have any feedback, just please feel free to comment at it.

00:34:09: Maybe if you've seen something that I haven't seen on the Internet and some news

00:34:15: and you want me to have a look at it and comment on it, feel free, comment down below.

00:34:20: If you are kind and want to contribute to this channel getting better,

00:34:26: please like if you're listening on a podcast platform.

00:34:30: Maybe there is a way to give a positive opinion as well and comment on it.

00:34:35: So thanks for tuning in at hand and talk to you soon.

00:34:39: Bye-bye.