275 | Knowledge work as we know it is over, self improving agents loops connected to Microsoft or Google eco systems are now possible, build entire software suite autonomously, Meta is cutting 20% of it’s workforce, & more AI news for Mar 13, 2026

28:51And they're saying that it is identifying bugs in 84 % of large pieces of code that has more than a thousand lines changed in them. The other thing is they're saying it's doing it with very low rate of false positives. So it's finding on average 7.5 issues per review with less than 1 % of false positive, meaning it's very rarely identifying bugs that are not actually bugs. There's been mixed feelings about the code review feature when it was released. Many people are saying it's extremely valuable and powerful. Obviously, a lot of people inside Anthropic said that it's completely changed the way they work.

29:29but it is currently billed based on token usage which means every review costs between 15 to 25 dollars that means that if you're shipping a lot of code this could become very expensive very quickly and then you need to decide whether it's worth it or not so let's describe for a second the problem that it's solving and then let's generalize what the hell that means the problems that it's solving is that the amount of code that gets generated right now is orders of magnitude more than the amount of code that was created before AI created most of the code in the world. So the people who are reviewing the code were, by the way, in many cases, the bottleneck before this case.

30:14But now it is crazy, right? So if you had 10 developers and three people reviewing the code, it was more or less doable. Now you have 10 developers creating the amount of code that 100 developers created before, and you still have the same three people to review the code. They're becoming a very significant bottleneck, which means A, they slow the deployment down, and B, the chance that they will miss things because they got to run faster is very significant, and C, it's just not doable. There's just not enough people in the world to review the code that will be generated in the very near future, meaning the ability to review the code by AI is necessary.

30:51The entire cycle of development of software has to change from end to end, because from the design process, which I'm doing right now with AI, without a design team, without product managers, but with AI agents in these roles has changed completely. And every company that creates code or any company who wants to create code needs to learn new ways of doing that. So is$15 to$25 per review high or low or too much or not enough? I'm not sure. I'm sure companies will have to figure it out on their own. But this kind of solution is mandatory moving forward because otherwise code is just not going to be reviewed, which is obviously not a good idea.

31:36Why do you think this is significant? Because the same exact thing is coming for any kind of work. We can now create 100x more marketing content than we ever could. Somebody needs to review this marketing content to make sure it is A, valuable, B, aligned with our core values, C, aligned with our brand guidelines, D, aligned with the goals of the campaign that we're running, etc., etc. Humans cannot do this because there's a huge amount of it that gets created at any given point. The same thing applies to more or less any outputs that a company can generate in the digital universe. Which means we need to learn how to build AI review systems to review the things that other AI systems are generating to make sure that what we're shipping, regardless of which aspect of the business it is, is actually what we want to ship.

32:26And it's not going to create any issues, whether technical issues, legal issues, customer service issues, brand issues, and so on. We will have to learn how to do this across every aspect of the business. But let's continue with additional interesting releases this week. Andre Karpathy, who is one of the co-founders in OpenAI, he was one of the leaders of the Tesla AI team and is a highly influential in the AI space. We talked about him many times on this podcast. He has just released something interesting, which is a minimalist open source Python library called AutoSearch. It is a very short piece of code.

33:03It has only 630 lines. But what it does, it is a short loop that allows to train small AI models on a single GPU overnight. The vast majority of us listening to this podcast, myself included, don't have any plans of training our own models. So why am I including this inside of a deep dive section of this podcast is because of what it means. So let me explain to you a little bit what this thing actually does. What this thing does is it takes an existing set of parameters of a small large language model, and then it uses two files. One is a markdown file from a human, the person who wants to run it, who defines the goals and the research instructions and so on.

33:49And another file that is controlled by the AI, which is a library that the AI itself writes as it's learning new things. every run runs for five minutes and then there's a clear defined criteria on how to score the output so then it goes and compares the output to the previous version that existed five minutes before that to see if it's a better version which means it now becomes the baseline or it is not as good as the previous one and then it goes down the trash but because the process itself is taking notes and learning from one cycle to the other, it is actually getting better and better from one cycle to the next.

34:29And it can develop and optimize a large language model over a small language model overnight. So if you think about it, because every experiment runs for five minutes, this means it can run about 12 experiments per hour, which is about 100 experiments in a single night, which allows it to learn very quickly on its own. This is just step one. The second step that Andre is planning, and some people have already implemented, like Varon Martha, who is the CEO of Hyperspace, is connecting multiple of these agents to a single shared knowledge base where they can learn from each other's experiments and not just each agent learning from its own experience.

35:15So between March 8th and March 9th, he had 35 autonomous agents on the hyperspace network run 333 fully unsupervised experiments. Now, because they are sharing a single knowledge base of learning from one another, when one agent discovered something, it told the other agents, which then knew how to do this thing without having to get to that in their own experimentation. This happened in several key findings through the night. So one agent's success became the knowledge for every other agent in that cluster, basically replicating how human research communities work. Only when research communities work, it takes a very long time for them to share information because they have to share research papers that gets published, not every single minute.

36:07And there have to be conferences for them to discuss this and so on. You're talking about unlimited number of these that are running 24-7 that are sharing information with one another in near real time. So let me give you the final quote that I want to share from Andre and then tell you what I think about everything we talked about so far in the show. And then we're going to go into rapid fire. Andre said, the goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. The only word that is unnecessary in this sentence, if you want to generalize it, is the word research.

36:44So if you generalize this, it will say the goal is to engineer your agents to make the fastest progress indefinitely and without any of your involvement, not just in research. So why do I believe that this is a highly critical aspect and why this puts together and connects a lot of the dots that we talked about today and in the past few weeks? If you remember a few weeks ago, we talked about Ralph Loop. So what is Ralph Loop? Ralph Loop is a very simple tool, actually, a few simple lines of markdown file that allows to do the following. Look at a list of tasks for development in software. Pick the one that is the most relevant to develop next.

37:25Send it to a set of agents that will write the code. Send it to a different set of agents that will review the code. Send it back to the agents to fix the code in order to solve the bugs. complete all the bugs that are there in several different cycles. Once the bugs are solved, commit that code, and then the loop starts over again by looking for the next most relevant tasks and going and developing it. This is a very similar approach to what Andrej Karpathy just released. It's a loop of agents with a clear score in the end to identify whether what you did is the right thing that needs to actually happen.

38:00But once you scale that with multiple of these loops in parallel that share a knowledge base with each other and they can run simultaneously while learning from the mistake and successes of all the other agents, this is now multiplies the capability of these loops almost indefinitely. And you can do this for any field of business. So here are a few examples. In the marketing space, think about optimizing any marketing copy, whether it's newsletters, landing pages, brochures, website, etc. You test it, you run it, you evaluate it. If it's better, you let it run. If it's not better, you kill it.

38:38But you can run as many experiments as you want at any given time. In sales, think about writing proposals, evaluating the proposals based on different set of metrics, and then rewriting it until it is fully optimized for the specific target audience or based on the specific RFP that you are using in order to write the proposal. In customer service, think about considering multiple options for the best way to handle a specific situation and then for the best way to communicate it with a client. In product development, think about considering multiple options for the next steps, the next features, the next cycle, the next version and optimizing for different outcomes.

39:16In drug development, and think about testing a huge variety of options before actually deciding on the ones that you actually want to pursue and so on and so forth for almost anything in any industry. Having an army of agents that run multiple tests while scoring them and evaluating them and picking the winners while learning from the mistakes can accelerate the path to success by orders of magnitude compared to how we are working right now. Literally, anything you can objectively measure and score, you can apply this model to. Now, since we already have open source solutions that built the scaffolding, the infrastructure for this, and I'm sure there's going to be even better of them coming soon, any company can apply this to any aspect that they want.

40:04In addition, I'm pretty sure that very, very quickly, this will become an easy to use infrastructure where it won't require you to know how to go to GitHub and use a repo which is a lot easier than most people think. But very soon, there is going to be a user-friendly version of that tailored into every aspect of our work as part of the tools you are using today. And that means that any person in any company doing any knowledge work will be able to apply these kind of loop operations that are self-optimizing across everything in the company. I'm already doing a lot of this stuff in the agents that I'm developing and in the course that I'm working on is going to be one of the topics.

40:42how to have a unified knowledge sharing across agents, across every aspect of the business, so they can learn much faster and not repeat mistakes that other agents have made or in the development of other agents. Now, let's connect all these dots, the convergence of all these different capabilities, such as the CLI access to your entire Google workspace, such as co-work inside of your Microsoft environment, such as the ability to run a huge number of experiments that self-improve to deliver specific business outcomes are complete game changers compared to what we knew just a couple of months ago.

41:20Now, we are not even starting to understand what this actually means for the way business is run today. But one thing is very, very obvious to me. And if you're listening, I assume it's obvious to you as well, those who will implement this now will gain a huge competitive benefit, at least in the short to medium future until everybody else figures it out. And again, if you want to learn how to do this, please reach out to me either as an individual or as a company. You can do this, as I mentioned, on LinkedIn or through the links to my calendar or my email on the show notes of this podcast. Now let's dive into the rapid fire.

41:58We have a lot to talk about. So let's start with the impact of all of this and other aspects on workforce. Meta is planning to lay off up to 20 % of its workforce. That is close to 16 ,000 employees that are going to be let go out of the 79 ,000 employees that Meta has worldwide. Meta is planning this in order to do two things. A, to compensate for the huge amount of investment they have in AI in the future. They have committed over$600 billion to build data centers by 2028. They just spent close to$2 billion on Chinese startup Malice. They are offering million-dollar compensation packages to hire new people to their super intelligence team.

42:45So they have huge expenses on the AI aspect so far without a lot of new income on the other side. And so they have to find ways to compensate for that. The other reason, if you remember, Mark Zuckerberg stated in January that he is seeing projects that used to require big teams now be accomplished by a single very talented person. So the combination of these two things leads to this really sad and yet obvious outcome where they got a layoff 16 ,000 people. sometime this year. There's no any formal thing, but it has already been communicated to the leadership team and through the leadership team to the second tier leaders inside of Meta that this is coming.

43:28To stay for a second on the Meta topic before I continue to additional layoffs in this particular case in Atlassian, Meta announced that they have acquired Maltbook. So if you remember, we talked about Maltbook just a few weeks ago. Maltbook is a social network for agents built by OpenClaw that was for a very short time was called MaltBot. So that's why the name MaltBook. So it's the combination of MaltBot together with Facebook to create this platform. So Meta just acquired the platform. And if you want, most importantly, they acquired the two people behind it are Matt Schlint and Ben Parr. The specific announcements from Meta says, and I'm quoting, the Maltbook team joining MSL opens up new ways for AI agents to work for people and businesses.

44:14Their approach to connecting agents through an always-on directory is a novel step in a rapidly developing space, and we look forward to working together to bring innovation, secure, agentic experiences to everyone. So that's the announcements from Meta's spokesperson. If you remember just a few weeks ago, OpenAI has hired, or if you want, purchased OpenClaw and its creator, which is the thing that started the whole craze that then drove these two people to create Maltbook. So what the hell is going on here and why would Meta and OpenAI hire people who developed a piece of technology that they can replicate in about 10 minutes?

44:54So on the Meta aspect, something that I think plays a big role is that Maltbook is maybe the only quote-unquote social network new idea that came online and became successful in probably at least 10 years. So that by itself is worth acquiring just so that nobody else captures that and then they might face some kind of a competition in the future. But I believe there's something much bigger than that that ties to something I shared with you several times in the past few weeks. But I will share with you again because I think it is very important that you understand this and or if you missed this in previous week.

45:28The world of knowledge work has shifted from a world in which the limiting factor for growth of a company is resources, money, compute, people, etc. to a world in which you need three things in order to drive 10x scale to any knowledge business. Great ideas, good judgment, and very well-defined requirements. If you have these three things, you can literally build anything you want because the agentic tools allow you to build these kind of solutions. But for these solutions to be successful, you need to have the right idea. You need to have the right judgment in order to A, prioritize ideas and B, then guide your original idea as it's being developed to the right direction to provide value and become successful.

46:17And you need very, very clear requirements in order to be able to develop this in a way that will actually work and will be aligned with the vision. The people who are just hired by AI and Meta obviously have all three at the highest levels. They had a great idea, they had solid judgment, and they were able to define clear requirements and create the product or the service that they wanted to create. These are exactly the kind of people that every company needs to hire right now. Every company that will be able to get to put their hands on many of these kind of people that can have at least one and hopefully two or three of these capabilities.

46:57So the ability to come up with great ideas, great judgment, and the ability to create detailed requirements to enable the execution to happen through agentic processes, it will be rocket fuel for the future of these companies. I am 100 % focused on this myself. I already see incredible results from the things I've developed in just the past couple of months following this same exact concept. And if you are a leader in a company, this is one of the things you have to start thinking about in the kind of people you want in your company. So back to the impact on Workforce Atlassian, which is an Australian company that is one of the key players in the software collaboration tools with tools like Jira that more or less every software company in the world uses in order to define requirements and track bugs and so on, is just announced that they're going to let go 10 % of their workforce, which is going to be about 1 ,600 employees that are going to let go with an aim to, and I'm quoting, reshape its skill mix.

48:02And in order to offset their huge investments in AI enterprise solutions. So very similar to what we hear from Meta, it is a double digit percentage of the employees that are going to get let go in order to support these two things. It is very obvious to me that this is going to keep on happening and it is currently happening in the software space more because this was the first topic in which the AI platforms had invested, but connected to everything we talked about in this episode, this is coming for every other knowledge work. Quick update on the topic that we discussed in detail last week, which is the clash between Anthropic and the government and specifically the Department of War.

48:45A, there was an announcement that said that there is no way this will be reversed and come back to a successful outcome of new negotiations. But on March 9th, Anthropic has filed two federal lawsuits against the U.S. government, alleging that the supply chain risk delegation and the President Trump's February 27 ordered to federal agencies to cease using its technology constitutes illegal retaliation for its protection of speech and violates constitutional and stationary law. How is this going to evolve? How it will impact the actual detachment of anthropics tools from the government? I'm not exactly sure, but I will definitely report as things evolve.

49:29Staying on the legal world, and I told you a lot of interesting things happened in the legal space this week, a lawsuit has been filed against OpenAI by Nippon Life Insurance Company of America, alleging that Chachapiti's chatbot engaged in unauthorized practice of law by providing legal advice to an individual who subsequently used it to file numerous motions in a new lawsuit after a settled disability claim. So the person, Graciela De La Torre, who is a former disability claimant, allegedly was unhappy with the settlement that her lawyers got her in a lawsuit and she has uploaded all the previous data of her previous trial from her lawyer into Chachupiti that validated her concern about being gaslighted and encouraged her to fire the lawyer and then assisted her in drafting dozens of motions and a new lawsuit despite the fact that the original case has been settled with prejudice.

50:30Now, to be fair, Opens.ai's user policy from October 29th of 2025 explicitly prohibit using its services for tailored legal or medical advice without a licensed professional involvement. That being said, the lawsuit is still there. Nipon claims that this is too little and too late. And OpenAI is saying that this is all bullshit and they don't understand why they're being sued. I assume the legal system will figure out a way to address that. That being said, this situation where individuals are using these tools for either legal work and or medical advice is growing every single day. And there is now legal work to try and block that.

51:15As an example, there's a bill advancing in New York, which aims to explicitly prohibit AI chatbots from dispensing substantive, from creating legal advice, and it's planning to create private right action for users harmed by such conduct. But combine that with the fact that the federal government is going to try to block every state kind of motion moving forward, I'm not sure how helpful this is going to be. It is definitely a problematic situation right now. As somebody that has been using Claude and Chachupiti and Gemini excessively to review and write legal documents, I will be really sad if this is being blocked because it's saving me a hell of a lot of money.

51:56If there's lawyers listening, I'm sure they think I'm completely crazy. But as somebody that has reviewed legal documents for three decades right now generated by lawyers, I do not see a big difference between the great advice that I'm seeing from these tools compared to what I paid for a lot of money by lawyers. I still use my law firm for final review of big things, but not for everything and definitely not from the initial state. Staying on the legal field, Elon Musk's lawsuit against OpenAI is entering a new phase where on Friday last week, a federal judge has decided what evidence and expert testimony can proceed to trial in April.

52:35Now, to put things in perspective, this case is going to decide whether the valuation of Musk's early donations into OpenAI in the very early stages is currently worth up to$109 billion, which is what Elon is claiming and that is going to be reviewed and decided in court starting next month. But despite all the backlash on AI in the legal world, Swedish AI legal platform Legora has just raised 550 million dollars in a Series D with a 5.55 billion dollar valuation tripling its value from just a few months ago in their Series C. Another company, Lagora's largest competitor, Harvey, has the same kind of situation.

53:24This is a US-based company that is backed by A16Z, is already valued at$8 billion and reportedly seeking$11 billion in its next round. Both platforms supply a different kind of service. They actually provide AI solutions to law firms and legal companies to execute more effectively. Or as the CEO of Legora said, it is amazing that everybody can have their own pocket lawyer in Claude, but we are not solving for the same use case. Staying on the legal world, Meta is now facing a major class action lawsuit after an investigation by a Swedish newspaper revealed that workers in a Kenya-based AI subcontractor company are reviewing footage from the Meta's Ray-Ban sunglasses, including some highly intimate personal situations that were recorded by these glasses.

54:20The lawsuit doesn't actually go after the fact that this is being done, but after the fact that the campaign for these glasses focused on designed for privacy, and that is obviously false advertising and not the case, when your intimate moments, including nudity and sexual scenes, are being reviewed by humans without your consent. And to make this even worse, you have no means as a user of these glasses to block your content from being reviewed by humans. Both the visual aspect of it as well as the voice communications, both are potentially being reviewed by humans. And in both cases, you have no way of blocking that.

54:59Now, based on the fact that over 7 million pairs of these glasses were sold in 2025 alone, the potential aspect of this class action could be enormous. I said multiple times before that there are huge issues with these smart glasses from multiple aspects of privacy. This is just one aspect of your own personal privacy. But the fact that a person that's wearing them is basically recording everything around them without the consent of the people around them in any situation is very problematic. and I don't know if it's breaking any rules right now, but I think it should break several rules that should come up to protect specific scenarios from being recorded without your consent.

55:42Staying on legal topics, the Supreme Court just denied an appeal by a person named Thaler, who for the past few years have been fighting the windmills of the copyright process in the US to get a copyright protection for a work that he has done with AI. The story actually started in 2018 when Thaler had listed AI as the sole author of his copyright application for a work that is called A Recent Entrance to Paradise that was rejected by the Copyright Office. He then took it through the judicial system in its different tiers, all the way up to the D.C. Circuit Court of Appeals, who kept the Copyright Office statement in place.

56:28And again, now he was rejected and appealed to the Supreme Court, which leaves the statement as is. The DC Court of Appeals affirmed a statement that says, and I quote, the court has said that calling human authorship a bedrock requirement of copyright, basically saying that AI cannot be at least the sole author of something that gets copyright protection. Now, the problem is how much human involvement is required for copyright protection remains unanswered. In a related case, that is Allen v. Perilmutter is currently in the U.S. District Court of Colorado, where the artist Jason Allen is challenging the Copyright Office refusal to register a work he refined using over 600 AI prompts, arguing that his significant creative inputs distinguishes his claim from the other quote where Thaler just gave one prompt and he's trying to copyright that.

57:30I tend to agree. I think that a significant input into a prompt or a set of prompts or a entire process of working with AI should receive copyright protection for two different reasons. A, you are bringing your own capabilities, your own creativity, your own mindset, your own judgment into the output. And B, if that is going to be disqualified, then copyright law is over because I think moving forward, most of human creation is going to be assisted by AI. And that means we don't have any copyright for more or less anything in the future. Switching gears to other topics, Replit just raised$400 million in a Series D and have unveiled Agent 4.

58:15Why is that interesting? Beyond the amount of money that they're raising that now becomes common because we're used to seeing these crazy numbers that were science fiction just a few years ago, they're claiming that Agent 4, which is their fourth generation of agents, is now capable of vibe coding an entire company, an entire software company from scratch on its own. So not just a simple application, but think about Salesforce or something like that, that being fully created, fully autonomously by these set of agents. Now, to make this even more extreme, Replit is stating that Agent 4 is 10 times faster than the previous version, Agent 3, that they released just a few months ago.

58:58So not only that you can create very sophisticated software, you can do this faster and cheaper than before. A big new feature of that is called Canvas, which is a digital scratchpad that allows you to see the different elements of the project and move them around and annotate cards working collaboratively between human developers and the AI agents in order to have more control and more understanding on how your software project is developing. Now, speaking on the software development space, interesting new announcement. XAI, Elon Musk's company, just recruited Andrew Millich and Jason Ginsburg, two senior leaders from the AI coding startup cursor, in order to enhance the coding capabilities of XAI.

59:45This has been a major focus for Elon, who has admitted publicly that their platform is not at the frontier right now, but he said it's a gap they can close by mid-2026, which is just around the corner. And I'm sure that hiring these two people from Cursor is a push in that direction. Another interesting announcement this week that has to do with the ability to do things faster and cheaper is Amazon AWS just announced a significant collaboration with AI-cheap startup Cerebrus, systems aiming to deliver their capability through the AWS cloud solutions. Those of you who don't know Cerebrus, they are the largest competitor of Grok with AQ, G-R-O-Q.

1:00:25Both these companies are developing a different kind of chips than GPUs that specialize in inference. Inference is when the AI is being used to generate its outputs, so it is not good or not optimized for training models like the GPUs are, but they are much better, faster, and more efficient than GPUs in the delivery of AI, which is becoming a bigger and bigger portion of the AI compute that needs to be done because more and more people and companies are actually using AI. So in the first year or two of AI training the models took most of the compute, that is shifting very, very fast to usage and these companies have very capable solutions.

1:01:05If you remember, NVIDIA just acqui-hired the leadership of Grok and now we have an announcement from Amazon that they are going to be partnering and delivering the Cerebrus platform to AWS users in combination with their own in-house chips. That's it for today. All that I can request is that you think deeply about what we talked about in the beginning of this episode. It will change everything that we know in knowledge work and the capability exists right now. This is not a future capability. You can start doing these things this minute or tomorrow in your company with the right training for the relevant employees.

1:01:47We will be back on Tuesday with a new how-to episode that will show you how to implement a specific AI use case. If you are enjoying this podcast, if it gives you valuable knowledge, I would really appreciate it if you spend two minutes in clicking on the podcast platform that you're on right now, giving us a five-star review or whatever you think I deserve, and then share the podcast with other people. There's a share button on your platform. It will take you two minutes to enter a list of several different people that you know that can benefit from this. They will be grateful, I will be grateful, and you will feel good about helping people know more about AI.

1:02:26That is it for this weekend. Have a great rest of your weekend and I'll see you again on Tuesday.