162 |  AI profound impact on the job market, the aftermath of DeepSeek R1, AI security and safety, and more AI news you need to know on the week ending on Feb 7, 2025

22:41So what he's saying, and I heard that from multiple people across the industry and beyond, is that the fact that it's becoming cheaper is not going to reduce the AI usage. It's actually going to make it explode, which means they will need more compute, more models, and more everything just because it will become relevant to a lot more use cases and available to practically everybody because the individual use case will be practically free or very, very close to that. Staying on the DeepSeek topic, DeepSeek's researchers actually shared their thoughts about Huawei's Ascend 910C, which is the latest AI processor, which is supposed to compete with NVIDIA's offering.

23:17And they're claiming that based on their research, it achieves 60 % of NVIDIA's H100 inference performance. And they're saying that by manipulating how it works, they can actually achieve even better results. Now, 60 % may not sound like a lot, but the gap was significantly bigger a year ago. So Huawei, which is Chinese largest cheap manufacturer, is closing the gap on NVIDIA. Again, they're still behind. And moreover, those researchers also said that while it's not bad on inference, it's still far behind when it comes to training reliability. So at least on the training side, GPUs are still reigning supreme, but the gap has been closing on inference.

23:51As I mentioned before, inference is becoming a big part of the game because these reasoning models are actually using a lot of compute in the inference time, when the model is actually being used and generating tokens versus training. And this development of hardware combined with, we see innovations in software like DeepSeq and Quen are a direct outcome of the US ban of NVIDIA chips being used in China. So first of all, they find ways to use it. As I mentioned, there's rumors that DeepSeq has used 50 ,000 GPUs from NVIDIA, but even if they only have 10 ,000, like they're claiming, that's still a lot.

24:21But putting that aside, it is forcing China to come up with their own innovation and to develop both hardware and software innovations that they probably wouldn't develop and they would have been dependent on the US otherwise. So to summarize the DeepSeek topic, I will mention something from a conversation I had this week. I'm delivering an executive education session in partnership with the university in a couple of weeks. And the person who is in charge of this effort has asked me in an email this week, do I think that DeepSeek changes everything and do we need to change the training data? And what I told him is I think that DeepSeek in the big picture changes very little.

24:51If anything, it solidifies the same point of view I had before. We will have access to incredibly powerful intelligence for free or very, very close to that. And the implications of that are obviously profound, but that doesn't change the trajectory. The DeepSeek event doesn't change that. If anything, it accelerates towards that even faster. As I mentioned, the third big topic that we're going to talk about today is security and safety. So OpenAI has done research on how persuasive their models are. And they've done this on the Change My Views forum on Reddit. So it's 3.8 million members who participate in debates trying to convince one another.

25:27They conducted 3 ,000 different tests. And what they've learned, that their models are becoming more and more persuasive to humans. So GPT 3.5 was 38th percentile persuasive. 01, 77%, 01 mini, 77th percentile compared to other humans, 01, 80th percentile, and 03 mini, two, 82nd percentile. So 03 mini, not even the full 03, is better in persuasion than 82 % of people who participate in this debate, meaning they are people that by definition are better than the average because they want to participate in this process. And so they are becoming very, very persuasive, and it's just going to keep on going because of their thinking capability and analyzing capability and reasoning capability that they did not have before.

26:12That obviously raises significant concerns. OpenAI shared why they're doing this research and their mitigation strategies and how they're trying to monitor this and make sure that they can prevent and restrict political persuasion and other specifically targeted campaigns. But can they really monitor 100 % of it? I'm pretty sure they can, and that obviously raises a lot of risks. So think about foreign governments or terrorist organizations using this to convince Americans or people in the free world with new opinions, because it will be impossible to know what's coming from an AI and what's not.

26:44It will be more persuasive than what the humans can write to contradict what the AI is writing. So that's where we are right now, and it's only going to accelerate. And I'm going to go back and forth between good news and bad news on the safety and security thing. So Meta just unveiled a new risk assessment framework for their AI development. And they just announced it on February 3rd, and they call it the Frontier AI Framework. And the goal is to establish clear boundaries to AI system development and deployment based on potential risk. And they've identified two levels of risk. One is called high-risk systems that could enable but not guarantee successful cyber, chemical, and biological attacks.

27:19And then the second tier is critical risk systems that could lead to, and I'm quoting, catastrophic outcomes, which they didn't label, but based on the fact they're talking about cyber, chemical, and biological, you can kind of know where this is going. And the biggest difference is, can you mitigate the risks once the model has been deployed? So basically what they're saying is that the critical risk systems, once you deploy them, there's no way for you to mitigate what's going to happen versus the other ones that you might be able to prevent it once it starts happening. Their policies go beyond just chemical and so on.

27:46So they're talking about things like automated compromise of corporate scale environments, creation and deployment of high impact biological weapons, and other scenarios that they're deemed most urgent by the company. So the policy defines both internal and external researcher risk assessment for every one of these models. They're including senior level decision makers, have the final review authority for everything that is said and done, and relying on multiple types of tests across the different risk levels instead of relying on one specific test. And the outcome is that high risk systems will be limited internal access and deployment only after risk mitigation, and critical risk systems, will have enhanced security protocols internally, and potentially will lead to development suspension.

28:26So they're basically saying that they're going to stop the development of AI systems if they deem that it could lead to catastrophic critical risk. I'm very happy to hear that. But that being said, that's just one company that decides on its own what the risks are and that are in huge competition with other companies who are doing the same thing. And as I mentioned, I'm going to go back and forth between good and bad security announcements. So Google just decided to abandon their 2018 AI weapons ban. So since 2018, they had a restriction on allowing weapons and surveillance development based on their AI models.

29:00And as of February 4th of 2025, they are removing some key components of that. So their previous policy prohibited the usage of AI for technologies causing overall harm, weapons or injury causing implementations, surveillance violation international norms, and technologies that could go around international law and human rights. And the justification they're citing for why they're removing this is that the need for democracies to lead AI development involving industry standards, geopolitical AI competition, and that they're saying that they're going to focus on appropriate human oversight rather than just banning the whole thing.

29:37Now, multiple employees in Google obviously are opposing and are being very loud against it, but this is the current state. And I shared with you that a few weeks ago when we were talking about open AI starting to get into these fields. I think it's inevitable and it's very scary, but I do agree with Google that if China and Russia and other adversaries of the US and the free world will start employing AI capabilities in their military and surveillance capabilities, and we won't, we will fall behind. Now that, again, can lead to catastrophic outcomes, but this seems to be the next Cold War or nuclear race or whatever you want to call it for world dominance.

30:12And hopefully it will stay balanced and keep everybody intact from doing stupid things, but these systems, AI systems, will be deployed and will control deployment of military forces and military capabilities around the world. That could obviously lead to a doomsday scenario, but as I mentioned, the alternative is not necessarily better by letting the other side have it when you don't. And ending on a positive note on the safety and security before we dive into a lot of rapid fire items. So Anthropic just announced that the the company has launched a new security system designed to protect Cloud AI from universal jailbreak attempt.

30:45Now, they achieved incredibly impressive results, again, very different than what we've heard about DeepSeek. So they reduced jailbreak success rate from 86 % to 4.4 % using these measures. And they did that while only increasing false positives, basically not breaking attempts that the system thinks are breaking attempts and going to stop it, by only 0.38%. So it basically hasn't prevented any action that is a legit action and was able to prevent 82 % more jailbreak attempts compared to the previous capabilities that they have. Now, that is costing them 23.7 % additional compute cost, but they're planning to continue to optimize from it.

31:21But Claude has always led that direction of safe AI usage, and that's another way for them to prove that. We are investing a lot of compute time in making our systems safe. And they're doing this using their constitutional AI frameworks that they've built their systems around, plus a lot of additional new capabilities that they're adding in order to do this. And I don't know if you know that, but they have had a bounty of$15 ,000 called Hacker One that is going to be given to a hacker that will be able to develop a universal jailbreak against Claude. And that has been around since August and nobody has claimed it yet.

31:51So they're definitely doing the right things in order to prevent hacking and jailbreaking their model. And I really hope that every single company that develops these frontier models will follow the footsteps of Anthropic on this particular case. Going back to the international development and partnership that I'm dreaming that will happen, I think these kind of capabilities and research must be shared with everybody and every company and every organization who is developing these kind of models to achieve the same results that Claude is achieving. And now let's dive into lots of rapid fire items.

32:20We're going to start with a lot of aspects from OpenAI. We already talked about their big releases with O3 Mini and O3 for the professional version and Deep Research, but they also made a lot of other interesting announcements this week. One of them is that Sam Altman has announced that being on the wrong side of history when it comes to open source, basically, and I don't know if that comes as, again, the aftermath of DeepSeek or anything else, but in a Reddit AMA Ask Me Anything, Sam basically admitted that the fact that they went to completely closed source might have been the wrong decision.

32:50And he basically said that they need to figure out a different open source strategy. In the same lines, Kevin Whale, who is their chief product officer, has revealed that the company is considering open sourcing older, non-state-of-the-art models, which they haven't done so far. Now, is that going to be impactful? I'm not really sure because now we have open source models who are competing with their state-of-the-art models. So if they're going to open source their older models that are not competing, they won't be able to compete in the open source world. So I'm not really sure I understand the logic behind it, but I think they understand that the open source world is accelerating and it's closing the gap on the closed source world.

33:23So the whole point of what they're doing, maybe not be that impactful and effective. And so they're obviously rethinking it. What will that lead to? I'm not a hundred percent sure, but I will update you as there's stuff to update. Now in the same AMA, Sam touched on a few other things. He did not give a specific timeline on GPT-5. He's saying that they're working on it, but it's unclear when will that become available. There's going to be new features and new capabilities for the O3 model coming in more than a few weeks, but less than a few months. He also mentioned that they're developing a successor to DALI, which is their image generation tool that is way behind the competition right now.

33:54And he's saying, and I'm quoting, it's going to be worth the wait. I assume we're going to get another extremely powerful AI image generation tool. We already have a bunch of those. So right now there's probably five or six models that are very, very good that can generate completely realistic images of anything you want. And we're just going to get one integrated into the OpenAI environment. To connect you to one of the previous points, OpenAI also defended its collaboration with the U.S. National Laboratories for Nuclear Defense Research, and they were explaining and expressing the fact that they have full confidence that the scientists will have responsible use of the technology.

34:24Whether that's true or not, time will tell, but that's their statement, which you wouldn't expect them to have any other kind of statement. A very interesting and positive development from OpenAI is that they announced a rollout of an education-specific Chachapiti version for California State University System that is going to reach 500 ,000 students and faculty across 23 campuses. Now, as you probably know, OpenAI has created a ChatGPT EDU version back in May of 2024, and it was deployed to some prestigious institutions such as Wharton and University of Texas at Austin and Oxford University. But this is a much broader deployment of a high education specific model that is going to be deployed across the largest public university system in the US.

35:04Now, the main features include personalized tutoring for students, study guides generation, as well as administrative task automation for faculty. So it's going to help both professors and students. I said that many times before. I think AI represents the biggest opportunity education has seen. We still perform education from kindergarten all the way to PhD in the same way it was done in the last hundred years with a professor or a teacher in front of a classroom teaching a large group of students, assuming they're all the same. It's a horrible way to train people because people have different needs.

35:36They progress in different ways across different aspects of the learning. They learn better by leveraging different tools. Some learn better by listening, some by watching, some by doing, some by using different games, some by watching videos, et cetera, et cetera, et cetera. And AI will allow us to, for the first time in history, to provide any person, any kid on the planet personalized tutoring, aligned with their needs, aligned with the things they need or want to learn, aligned with their capabilities on different aspects, and aligned with how they learn best. And I really hope that this is where this is going to go.

36:05First, it will provide education for all. And second, it will make the existing education in places that it's already available significantly more efficient, hopefully making teachers more of mentors than the people who actually teach the material and allowing people to engage people or kids through the process in the most effective way. Now, I mentioned before that OpenAI has filed some patents. Well, one of the things that Sam Altman confirmed this week in an interview with the ELEC, which is an organization in Japan, that they are indeed developing a device and that they're working in collaboration with the legendary Apple designer, Johnny Ive.

36:37I've shared these rumors with you in the past. Well, now they've confirmed these rumors. Sam emphasized the fact that voice will become the user interface medium of the future. I can tell you that right now, I personally use the advanced voice mode on ChatGPT as well as the live streaming on Gemini regularly. And it literally changed the way I work with AI models, but I can't wait to basically activate everything with computers and literally just have conversations with them to do everything that I need. And I think it will over time become the user interface instead of a keyboard, mouse, or a touchpad on every device that we know, including microwaves.

37:10Why do I need to figure out how much time I need to put in? I can literally tell it, or it will see on its own with a camera what I'm putting inside of it, and it will just run the microwave based on what's the best outcome that can happen while asking me what my preference is for the level of cooking or the temperature that I want the output to be. But there will be no need for physical user interfaces on most things interact with right now. Another thing that Sam said about the device is that their plan is not to replace the smartphone, but actually to have an extension of their smartphone. So think about something similar to a smartwatch.

37:38So it will connect to the phone, but it will be a device that will allow you to talk to it and use AI in order to then operate the phone or operate through the phone things that you need to do. So now let's switch gears and talk about rapid fire items as far as new releases of products and features. So we already mentioned OpenAI with their O3 model in deep research. Well, Google weren't standing still either. And they just announced on February 5th that the company has made the entire Gemini 2.0 family generally available. And there's going to be three different models, 2.0 Flash, the smallest model with 1 million tokens context window, which is way beyond everybody else.

38:11Then you have Gemini 2.0 Pro, which is an advanced model that is available through paid members with 2 million tokens context window. Again, now twice what you get in Flash. And then there's Flashlight, which is built for cost efficiency. 2.0 Flash will become the mid-tier model. Think like Claude 3.5 Sonnet on the Claude family. And then you'll have the Pro, and then you have the lighter version that can run faster and probably much, much, much cheaper. And these models have enhanced coding capabilities, improved complex prompt handling, advanced reasoning capabilities built in Google search integration, code execution capabilities, a lot of stuff that I shared with you in the past before they made it widely available.

38:44One of the interesting things that they mentioned about Flashlight model is that it's very good at image understanding and that it can process approximately 40 ,000 image captions for less than$1 of compute. Going back to what we already talked about before, we're going to have very high capable intelligence across everything we need for practically free. Now, going back to our benchmark, which is the chatbot arena, Gemini 2.0 Flash Thinking is now number one. Gemini 2.0 Pro is number two. Then ChatGPT 4.0 Latest, which is kind of like a model that they released in late 2024, then DeepSeek R1.

39:18So these are the top five models right now. And then number six is another Gemini model, which is Gemini 2.0 Flash. And so what you can see, as I mentioned last week, is that Gemini basically took three out of the top six models in the world right now, as far as the success of the results based on usage of actual people, which is how the chatbot arena works. And I anticipate, like I did a year ago when they were still far behind, that Google will keep on leading this race because they have access to more of everything you need in order to make this work. more compute, more engineers, more money, more data between Google search and YouTube and other platforms that they control and huge distribution across the entire Google ecosystem.

39:53Another company that made a big release this week is Mistral. We talked about Mistral many times in the past on this podcast. Mistral are an open source platform from France and they just released a new multi-modal version of their Le Chat model. They just released it on February 6th and it has faster answers than probably any other open source model right now. They're claiming a thousand words a second, advanced document processing with superior OCR capabilities, code interpreter to create and run code on the platform, image generation powered by Flux, so not their own, but it's integrated into the models, and new memory capability similar to what we have now in several different platforms.

40:28There's going to be a free tier and a paid tier for 15 bucks a month, and very interesting, they're developing more and more enterprise relevant tools, which are not surprising, so secure environment deployment, custom tool integration, tailored model customization, and also they're coming up with new connectors to existing work environments within enterprises. So like all of these companies, they're pushing very, very hard into the enterprise world and being an open source company, it actually makes a lot of sense because it obviously allows you to host it on your environment while keeping your data secure.

40:56Another company that made a big announcement this week is GitHub. GitHub, one of the most commonly used coding environments, they just shared that they're releasing three different AI capabilities. One is called Agent mode for autonomous coding. The other is copilot edits. And the third is project Padawan reviews. The most interesting one is obviously the agent mode. And agent mode allows self-iterating code generation, automatic error recognition and fixing, terminal command suggestions, runtime error analysis with self-healing. And it is going to be available for everyone who's using Visual Studio, which is the code platform that it's integrated into.

41:28Biggest difference here of all these three platforms is that it will replace not just writing short code snippets, but a much bigger portion of the coding ecosystem and process. And as I mentioned in the beginning of this episode, when we were talking about Y Combinator, this is where it's going. Like coders will sometime in the near future will write very little code and they will depend on these platforms to route code on their behalf, which will allow them to create significantly more code, faster and much, much cheaper, which will allow us to do future developments better than we can do them right now.

41:56Another big release this week that is very interesting and very scary is ByteDance, the company behind TikTok, has released what they call OmniHuman, which is a tool that generates videos of people based on a static image of them and a voice sample. So you can take an image of any person and they've shown multiple examples of celebrities as well as historical characters. And then they took a voice sample of that person and it generates a very realistic motion of that person speaking as if it's the real person. That's just another really amazing deepfake capability, which on one hand is amazing because you can bring to life historical characters for educational purposes.

42:30So one of the examples they showed is Albert Einstein speaking about whatever you want. And so you can use it for educational reasons, which is really cool. But on the other hand, it's deepfake and it will allow to replicate any person on the planet in a very highly realistic way in seconds. Now, right now it's still research, but I would not be surprised if this becomes a part of the TikTok platform sometime in the very near future. Another interesting announcement on new types of models, but not the models yet, is that a new European AI allowance has emerged. It's called the Open Euro LLM, and they have launched with 52 million euros in funding.

43:01And their goal is to allow an independent open source development of models by Europe in order not to relay on the dominance of US and China and to allow a consortia of European organizations to develop a European-based language models and other AI capabilities. And as I mentioned, the goal is to make it full open source, including everything, models, software, data, and evaluation through open source capabilities. And then the last topic I want to talk about is the new announcement from the US Copyright Office. they just came out with a second report and actually with two different parts to it that is highlighting a significant change from their initial view of AI, but they're still claiming that the law doesn't need to be changed.

43:41They're just elaborating on it. The original ruling from a year ago said that basically anything that is AI generated cannot be copyrighted. And now they've changed their opinion a little bit. And they're basically saying that if the process was initiated by human ingenuity or if the significant part of work was human, then it's still copyrightable and can be protected under copyright laws. One of the things that they made very clear is that prompts alone do not provide sufficient human control to make user of the AI system authors of the output. Meaning even if you wrote a three-page prompt to generate an output, the output is not protectable from AI laws in the US.

44:16But they give other examples where work that is co-created or initiated by the human is protectable. I have a very personal example of this. A lot of the content that I create is an output of this podcast. I take the podcast and run it through AI systems and that generates blog posts, social media posts, ideas for content and other stuff. And that now is protected because it's derived from the content that I'm creating on the podcast. Another example that they gave is that if you're a painter and you start with a scribble that you create and then you use AI to improve it, it's still yours and can be copyrighted.

44:44I think that will continue to evolve. I think serious prompting versus short prompting, and I don't really know how to define that in a legal way, will also be considered eventually because it's still something that I'm inputting into the system in order to create an output that another person will not be able to do because they did not have my ability to do it, which is the whole point of copywriting. So I think that will continue to change. But the fact that they finally moved from anything with AI is not copyrightable to, okay, some things are, is saying that they understand where the world is going and that they're open to change.

45:14That's it for this week. We'll be back on Tuesday with another high too detailed episode that you are going to absolutely love. If you are enjoying this podcast, please rate us on your favorite podcasting platform. It will take you five seconds. So pull up your phone right now and click the five-star review and write whatever you think about the podcast or reach out to me on LinkedIn and give me any feedback. I love getting your feedback on LinkedIn and I get a lot of it almost every single day. So good, bad, ugly, whatever you want to say, I really want to hear about it. And while you're at it and you have your phone open, please share it with other people that can benefit from this.

45:43I'm sure you know a few people that if you think about it for a second, you can say, oh my God, so-and-so has to listen to this podcast as well. And so please share the podcast with those people. And until next time, have an amazing day.