182 | Sam Altman’s Vision of the AI Future, ChatGPT Goes All-In on Vibe Coding, Enterprise AI Disruption & more Top AI News form the week ending on Apr 18

26:06Now people are doing reverse location search on images. So many people have shared taking images from multiple sources, uploading to Chachapiti, specifically O3, and asking it where the picture was taken. And it's showing really good results at understanding from small nuances and things in the image where the picture was taken based on landmarks or different hints that it finds in the images. While this is really cool, it obviously raises serious concerns because any picture of yourself that you put online, let's say on social media, can now be used to tell people where you were at the time that you were taking the image.

26:43It's also showing you how powerful these tools are right now to understand nuances and little hints from images to get a much better understanding of more things than we probably know. Those of you who have seen the very first demo of Gemini where the Google female employee was demoing Gemini Live and she was wearing the glasses and she looked outside the window and all you could see is rooftops of other buildings around. And she asked Gemini where it thinks she is and it told her that she's probably at King's Cross in London, which is where she was. So these tools exist, this capability exists.

27:17It's not exposed in a formal way, but it's there and you can use it. And it will be very interesting to see now that people know that, what they develop from an API perspective, what other applications can be developed to benefit from that capability. If you think about the negative stuff, there's a lot of negative stuff, but if you think about the positive stuff, allowing you to ask where you are in order to get better navigations if you don't know where you are, or how to find the one thing that you're looking for in a street where you don't have the exact address, or stuff like that, just by seeing what you're seeing, that's the direction where it is going.

27:44And on the last two items from OpenAI, I want to touch on the hallucinations and mistake side of things, because I think it's very important to touch upon it. According to OpenAI internal testing, O3 and O4 Mini models that were just released hallucinate more than both companies' previous reasoning models and traditional models, so GPT-4-0 and O1 Mini and O3 Mini. To make it even more confusing or less promising, OpenAI themselves do not know why these models have higher levels of hallucinations, and they said, and I'm quoting, more research is needed. So they do not understand why when they add additional capabilities to the models, they get additional hallucinations.

28:23By the way, in a pretty big spread, so on a benchmark called PersonQA, O3 hallucinated 33 % of the questions, while O1 and O3 Mini hallucinated only 16 % and 14%. So it doubled the rate of hallucinations in this model. That's obviously not a good thing. And as of now, it doesn't seem that OpenAI has a good solution for that. Now, on a more interesting aspect of this, OpenAI researchers published a paper last week revealing that their deep research technology, which I started using all the time, not just from ChatGPT, but from the other providers as well, they found two very interesting facts. They were testing it against a very hard benchmark.

29:03And the idea was to compare it to A, previous technology, just GPT-4-0 versus deep research, and B, compare it to actual human researchers in those same really difficult tasks. And what they found, that deep research did dramatically better than humans and a whole different level from the old tools. So the old tools basically failed miserably. If you just gave GPT-401 of these tasks, it couldn't complete any of that research effectively. Humans, many of them gave up on 70 % of the questions after two hours of effort and not finding answers. And for the people who did continue all the way, still had 14 % incorrect answers.

29:40Now, deep research got answers to all the questions, so they didn't give up on any of them, but it only came back with correct answers 51.5 % of the time. Basically, just a smidge accurate than flipping a coin on the right answer on something. Now, that's not very promising, but to be fair, this is on the hardest set of questions that they could find that, again, took human researchers hours to actually complete. So what do I think about this personally? I started becoming more and more dependent on these deep research tools. It is very alarming to me to find out that they could be wrong 50 % of the time.

30:14So what this tells us is that while these tools are incredible and they really do amazing research, you still need to go and click on the links of the sources that it provides you and verify the information. So the research is not going to take you just five minutes. It's probably going to take you 25 minutes to actually go through the relevant links and see what's reliable information and what's not. But it's still going to save you a few hours of doing the research yourself because in many cases, these websites will check 70, 80, 100, 200 websites in order to get you the answers, which if you have to do on your own, will take you hours of work.

30:46As I mentioned, there's been a lot of other news from OpenAI that will move to the rapid fire items. So let's dive into those. First of all, according to app figures, Chachapiti's app surged to 46 million downloads in March, 2025, a 28 % jump from February and making it the number one globally downloaded app in the world overtaking Instagram and TikTok. Now, the interesting thing is what drove that madness is obviously their new image generation tool and all the Ghibli style images and then the action figures madness that happened shortly after that, which is ridiculous to me because there are so much better real value actual use cases that you can use Chachapiti for in your personal life and in business that gives you way more reasons to download the app and use it regularly than creating anime style images.

31:33But I guess it doesn't really matter because that drove this complete craze of downloads of the ChatGPT app. And it is amazing to see their brand dominance where basically ChatGPT became synonym with AI. It's just like Google became Google in search in the 2000s, almost the same exact way. People just know ChatGPT. Most of the people in the world are not like me or you, the listeners of this podcast that know multiple chat platforms and use them in different use cases. Most people just know ChatGPT and that's the only thing they're using. And it's an incredible brand power that OpenAI has right now over anyone else.

32:12An interesting piece of news that is actually a good news from my perspective when it comes to using OpenAI's APIs, OpenAI is planning to implement a new verified organization process for you to get access to their most advanced models. So this is not going to impact the models that exist right now, but it will be required for future models. And that means that any organization who wants access to these future APIs will require an approved government-issued ID from one of the countries that is supported in the OpenAI's API. And without that ID, you won't be able to get access to these tools. I think this is a great step in the right direction.

32:45I don't know if it'll solve all the problems because there's many bad people and bad players within countries with actual legit IDs, but at least it reduces the risk of the wrong people in the wrong groups using these really powerful APIs for the wrong things. Now, this might also be as a response to what happened or presumably happened with DeepSeek when they were more or less scraping, or if you want the professional word, distilling their models by basically running it against their API. That was probably blocked in other ways before, but having an approved government ID as a prerequisite to getting access to advanced models, I think is a great step in the right direction of higher security when better models come out.

33:24Now, I told you before that I'm going to come back to the safety risk that is driven by the crazy competition in the AI universe. Well, OpenAI just updated its preparedness framework, and they're stating that they may adjust their safety requirements if competing AI labs release high-risk systems without similar protections in place. Now, they are claiming, to be fair, that they're only going to make these adjustments after rigorously confirming that the risk landscape has actually changed, publicly acknowledging that they're making those adjustments and that these adjustments do not meaningfully increase the overall risk of severe harm and still keeping the safeguards in place.

34:02So in the TED Talk interview of Sam, Chris made an interesting statement. He basically said that all the leading labs, because of the competition, are basically saying we got to run the fastest that we can because it's inevitable that somebody will get there. And if we get there first, it will be safest because we're safer than everybody else. And he was questioning the inevitable aspect of that scenario. What this is telling us, that it is inevitable. It is inevitable because the labs themselves are claiming on paper in their formal security guidelines that if other labs crosses the line, they will cross the line too.

34:37And this is not even spoken in closed rooms and behind closed doors. This is their formal statement. basically saying, if other labs move forward with less cautious, we will do the same. This is really scary to me because it basically means that this competition can drive us over the edge and all the leading labs will play their role in that scenario. In another interesting acquisition by OpenAI, they just hired the team, not the company, that started becoming a norm in the AI world. They acquired the team of Context.ai. Context.ai is a startup that was founded in 2023 by two former Google employees.

35:13And what their tool allows to do is to look into the actual doing of AI language models, understand them better, and use that to develop better and safer AI solutions. So OpenAI just acquired the team and obviously all their knowledge. The company is going to dissolve its previous operations. And Scott Green, one of the founders, is now product manager at OpenAI building evals. Those of you who don't know what evals are, evals are the ability to evaluate what the AI is doing. And evals are a requirement to building high-performing AI applications, but they're very hard to get right now because it's quote-unquote a black box.

35:47And what their tool allows to do is to look into the black box and have a better understanding of what it is to evaluate how the AI is working to develop better and better AI systems faster. Like I said, this could have been a whole episode about just OpenAI and ChatGPT, but there's a few interesting news from other companies. The biggest one comes from Microsoft and they just unveiled three really interesting things. The first one is really promising for the future and they have released a one-bit, really small model that can run effectively on a CPU. So it's a very powerful, really large model that actually provides leveled results with LAMA 3.2 and Google's GAMMA 3 and Alibaba's QEN 2.5.

36:24So not the latest and greatest, but just the level before that on a very, very small footprint that can run locally on smaller computer with significantly less memory. Why do I think this is promising? I think this is promising because the world's compute and resources are now all being drained in order to support future AI models. And if we will find ways to scale models not with that level of compute, I think the planet will thank us. Now, on the more practical things that you can start using right now, Microsoft released two big releases this past week. One of it is AI agents that can control your computer.

36:59So the concept of computer use, the same as we got from Claude a while back and from OpenAI a few months ago, you can now use Copilot Studio to build these agents that will take over everything on your computer. So anything in your browser, as well as applications, basically it looks at the user interface and can perform anything that humans can perform on the screen. Now I'll say two things about these kinds of agents and computer use solutions. One, I think in the long run, it will open the door for a lot more automation and will probably eliminate completely any tedious tasks that we have to do today.

37:31So from that perspective, it's really good. Two, any person who has tried to use the full Copilot Studio knows it's an absolute nightmare. So first of all, they have two products called Copilot Studio. One is basically custom GPTs wrapped as a Microsoft product. And the other is their old school automation tool with some more AI capabilities, which is absolutely impossible to use. It's the worst user experience in the world. I consider myself an advanced user, a techie and a geek. And every time I try to build something there, it's very hard to do. And I gave up more than once. So I really hope that part of the process is allowing the AI to create the AI so I wouldn't have to fight the tools that they have created.

38:08So I'm not sure how the implementation is on the Microsoft side. It will be very interesting to test and I will keep you posted as I do that. The other thing that I will say that there's a huge risk in allowing these tools to take over your computer because you don't know what they will do. You don't know when they're going to go crazy. You don't know whether on purpose, maliciously, or just by random fluke mistake, They will do stuff like changing your passwords or locking you out from specific pieces of information that you actually need or anything else that you don't want that would happen on your computer or in your company network.

38:37And what I have done in order to test these kind of agents is I've actually created a virtual machine on Google Cloud, and I'm using that to run a browser, and I'm testing everything in that virtual machine. So if this goes rogue, then nothing happens to my real universe, and I can safely test different use cases, including different agent capabilities. I highly recommend that to anyone, and I will create a full episode about this showing you exactly how I did it so you can do it as well. The other thing that Microsoft released this week is they made Microsoft Copilot Vision available to anybody on Edge.

39:10So Vision is if you want a subset of agent, it can see everything on your screen. You just cannot take actions. And this was available for the paid users of Copilot and how any user of Copilot, including the free users, can show the AI everything on the Edge browser. And you can even activate your microphone and talk to it. That will allow you to basically ask the AI about anything that you're doing, whether it's emails or spreadsheets or research or anything that's on your screen, you can ask about and collaborate with the AI. I've done this multiple times with Gemini Experimental Capability because that's been available for a while.

39:45It's been working, I would say, about 50 % of the time effectively. And the other 50%, it crashes before you get to a solid outcome. That's at least on the Gemini platform. You can also do this on OpenAI, but only on their mobile app, which is weird to me because it will make a lot more sense to have it on desktop, either on the browser or on the desktop app. But right now on OpenAI, it's only available on the mobile app. But anyways, it's now available for Microsoft users. Anybody who's using Edge can now activate this and use that. It's only currently working with several different specific websites.

40:15such as Amazon, Target, Wikipedia, TripAdvisor, Food & Wine, OpenTable. So basically websites you use in order to do your day-to-day personal things and not necessarily business tools. And as I mentioned, I see a very long time between now and the time that an organization will allow these models to deal with actual enterprise systems because I don't know how reliable it is yet, and I don't know when it will be 100 % reliable, or at least as reliable as humans doing data entry. And I think we still have time to go until that point. Going from Microsoft to another widely used platform that has released interesting AI capabilities, Notion, officially released Notion Mail.

40:50It is an AI-powered email client that connects to your Gmail account and allows you to do really cool things. The first thing, it allows you to manage the email. It reads all your emails and you will categorize them and put it in different buckets so it will be easier for you to review. It also knows how to suggest answers. It also knows how to look for suggested meetings. So if somebody said, let's meet, it will look through your calendar and make suggestions to when you are available and things like that. This is obviously the direction that all these tools will go. I think it's a great move from Notion to move in that direction and offer that integration.

41:23I must admit, I'm surprised, and I shared that before, that doesn't exist in Google itself already as part of Gemini. But I think once Google and Microsoft figure out how to integrate all their Gemini tools and all their Copilot tools into one Gemini that can see and connect everything, and the same thing with Copilot, we will gain huge benefits to efficiency on our daily business usage. Now, broadening from day-to-day work of everybody to what AI is doing in the broader scheme of things from an application perspective and what aspects of the world it is going to touch, two former Tesla supply chain leaders have started a company called Atomic, and their goal is to develop an AI-powered platform focusing on streamlining inventory planning and supply chain management.

42:04These two leaders have experienced the craziness of Tesla's near collapse when they were scaling Model 3 production in the beginning. If you remember Elon Musk sleeping at the factory floor at Tesla in order to go through that period. So they deeply understand the problems and the need to a better supply chain control and management. Most of the supply chain management in the world today right now is done manually. I have several clients who have warehouses and inventory and supply chains. And the amount of work that is done by copying and pasting data from multiple Excel files and emails into a unified environment to understand what is actually happening in the company is incredible.

42:42We've been solving this problem through automations and custom GPTs, and they're gaining amazing benefits. So first of all, if you have issues like that in your company, reach out to me on LinkedIn. I'll gladly help you out in stuff that you can start doing tomorrow versus in months or years when this company takes off. But their goal is to basically put AI software that allows users to quickly simulate multiple scenarios that would normally take hours or days to calculate. And based on that change, the whole supply chain structure. In early pilots, they have shown several different examples, such as reducing inventory by half while maintaining 99 % in-stock rate for the relevant needed components.

43:20This is obviously very promising, and I'm sure that a lot of companies need a solution like this. Another field where this is happening very aggressively is obviously customer support. So Zendesk CEO just been interviewed and shared a lot of interesting information about AI in customer service. It's a very near future where 100 % of customer service interactions involve AI and 80 % are solved by AI without any human intervention. Now, the interesting thing is, according to a survey that Zendesk themselves did, 51 % of consumers are saying that they prefer interacting with bots over humans when seeking immediate service.

43:58So I want to unpack that for a second. First of all, the survey was done by Zendesk. So they have a vested interest to say that because that's what they're pushing, their platform. But that being said, now I'm going to speak from my personal perspective. Every time I have to deal with any of the big companies, such as AT &T or any medical group or insurance or airlines, it drives me crazy. The amount of calls, communications, people I have to talk to in order to solve problems that seemingly shouldn't be that complicated. I just spent last week two hours on the phone trying to get my tickets from StubHub.

44:32That is insane. That shouldn't happen. And I agree that if these tools are done correctly, and they're connected to the right systems, and they can find the right information and take the right action quickly and effectively, I would rather talk to that than the traditional customer service agent any day, any time, and only escalate to a human when it cannot solve my problem, which happens anyway in more than 50 % of the case. An actual example that I found is that the city of Buenos Aires is developing AI chatbots to manage over 2 million queries per month without human intervention, and that has reduced their burden on their actual staff by 50%.

45:12So that basically tells you that they are dealing with a huge amount of responses that are successful because it's not getting to the humans after that. It is very clear that the customer service universe is changing dramatically as we speak. And I already predicted that in the beginning of last year, that I think the concept of call centers or connect centers will disappear from the world. And I know that's a scary thought for an entire industry to disappear, especially in countries like India and Philippines, where there's millions and millions of people who make their living through that industry.

45:44But I don't see this evolving in any other way. Another interesting data point about AI implementation in enterprises comes from Johnson & Johnson. They just released the findings of an internal research that they've done, and they found that 10 to 15 % of artificial intelligence use cases delivered 80 % of the value. That's in the story that was shared on the Wall Street Journal this week. Not too surprising to me, but I think it's not the right way to measure it, and I'll explain. The reason it's not too surprising to me is the 80-20 rule always works, or in this particular case, the 85 % rule.

46:19But it always is true that whatever are the big projects are going to deliver the most amount of value. That doesn't mean that the other 80 % of projects do not provide value or do not provide positive ROI. So I can tell you that many of the things that I'm doing with a lot of my clients or that people that take my courses do in their companies are small day-to-day initiatives that are going to change the efficiency by small percentage, but free a lot of time for specific individuals to focus on bigger things. None of these things will be visible as a significant one line item on the bottom line of a large-scale company, right?

46:55So if you want to move the needle for Johnson & Johnson, the benefit needs to be in the billions. If you just created an initiative, a small custom GPT that removes the need to do a task once a day for an hour, that is worth a million dollars to the bottom line. But it's not going to move the needle for Johnson & Johnson, so they're going to say that didn't qualify to move the needle for the organization. But if you do 20, 30, 50, 100 of these in an organization, well, that adds up very quickly to a significant amount. So while I agree with that, that the 80-20 rule works, I think every use case needs to be tested as its own ROI use case.

47:32And if there is a positive ROI, continue doing it, even if it's not a huge addition to the bottom line. And I think there's an important way here to differentiate between company initiatives that require a huge amount of resources. And over there, you do want to focus on the ones that will produce the most amount of value versus education and training to your employees on how they can apply AI tools on their day-to-day stuff to gain these small benefits, and then just encourage anybody who can get a positive ROI out of that process. Staying on the topics of industry and new tools, LightSource Labs has emerged from stealth mode.

48:05They just raised$33 million in seed and series A, and the company addresses a critical industry gap this side of procurement. So again, we're still staying in that universe of supply chain and so on. And they're stating exactly what I told you before, that 70 % of procurement teams still rely on manual processes using no software tools and sourcing using no software tools and instead managing billions, if you look at aggregately in the world, through email spreadsheets and randomly formatted document, exactly as I mentioned earlier. And they want to transform that aspect of the business. So a different company, slightly different flavor, but I'm sharing with you all of these to start understanding where the world is going.

48:43AI is going to be in everything that we do, and we'll be able to dramatically improve everything that we know in every aspect of the business. And as a business owner, a business leader, all you have to do is think about what are your biggest pain points? What would you like to solve? And start looking for these kinds of solutions because they exist. And if they don't exist, they will exist in the next few years. Staying on the topic of new interesting developments in the enterprise space, Adobe just took a strategic investment in Synthesia. So those of you who don't know Synthesia, they're a British startup that allows you to create and use AI avatars.

49:16This comes at the same time that Synthesia announced a hundred million in ARR, which is positioning them as a leader in this field. The two leading companies in the AI avatar universe are Synthesia and Haygen. Both of them have very advanced capabilities, but the investment from Adobe makes it very interesting because this now brings potentially in the future, these kinds of capabilities into the Adobe universe, which will allow interesting combination of the Adobe tools with avatar video generation. And if to quote Synthesia's CEO, he said that their vision is aligned with Adobe to democratize high quality content creation and making enterprise communication faster and more effective.

49:53I couldn't have said it better. I think it's a very interesting partnership and it will be very interesting to see if Adobe actually starting to leverage that inside their tools. That being said, despite the amount of money that Synthesia is making right now, and despite the fact that they saw a hundred percent growth year over year, they still lost over 25 million pounds this past year. And continuing on the topic of impact on industry, Hugging Face, the open source behemoth, the one that hosts most of open source models in the world, just made a very interesting move, and they acquired a French-based robotics company called Pollen.

50:25Some wasn't released, but the robot called Ricci2 is now available for sale for$70 ,000. Now, the interesting thing about this particular robot is that it comes ready to go, but with an open source platform that you can continue to develop on your own. Hugging Face team lead for robotics is Remy Kaden, who is a former Tesla robotic scientist who worked on the Optimus program. So they're definitely very serious about their move in that direction. And the fact that they're releasing the world's first advanced open source model makes it very, very interesting and showing you how many huge advancements are happening and are going to happen in the robotics space.

51:02I think it will be very interesting to see what companies do with that infrastructure and architecture that is now open space. And it's a very interesting move by Hugging Face, putting them in the hardware field as well and in the very competitive and highly lucrative robotics universe. There are a lot more robotics updates and if you want to find out about them, you can read about all of them in our newsletter, which you can sign up for in the link in the show notes. All you have to do is open your phone right now, click on the links and you can sign up for our newsletter. And over there are a lot more news that we cannot share on the show.

51:30But our next topic, after talking about enterprise and what's happening in that field, I want to share with you something interesting from Grok. So Grok just shared Grok Studio. Grok Studio is their version of Canvas in ChatGPT or Artifacts in Cloud. And this is a side-by-side split view where on the right, you have a document that you can edit. And on the left, you can continue to have the chat in Grok. I started using it yesterday and I must admit it is not to par with Canvas and not even with Artifacts. It can run documents and code at the same time. It has one cool benefit that it has more of a document style editing capabilities where you can change things to bold or italic and different headings straight there in the user interface.

52:11But it's lacking a lot of other aspects that Canvas is so helpful at. The main thing from my perspective is the ability to highlight a specific segment of the text and getting the AI to work just on that segment. To me, that's the most magical aspect of Canvas. The other problem with the way Grok Studio works is that it writes everything, the full answer, which sometimes for me is like three pages on the left side, basically on the regular chat, and then copies it to the other side, which takes twice the time, which absolutely drives me crazy. So a good step in the right direction. I think these collaborative environments with AI are fantastic.

52:46I think the current implementation by Grok is lacking and needs to be improved. That's it for today. Don't forget right now to open your app and look at the survey and tell us what you want this podcast to be. I want to listen to what you have to say and make adaptations to this podcast to serve you better. But for that, please fill out the survey. Also, don't forget, there's an AI business transformation course starting on May 12th. So if you need better structured training on how to implement AI to drive your career, to improve your team, to improve your business, your company, it's an incredible opportunity.

53:17We teach these courses all the time, but usually for close groups and only once a quarter, we open them to the public. So the next one after May will probably be in September. So don't think twice and come and join us. Again, the link for that is in the show notes as well. On Tuesday, we'll be back with a fascinating episode that will show you 15 different use cases on how to use the new AI image generation capabilities. 14 out of them are business oriented and not for fun. So it will teach you a lot of valuable stuff. And for now, have an awesome rest of your week.