Is Something Big Happening?, AI Safety Apocalypse, Anthropic Raises $30 Billion
Channel: Alex Kantrowitz
Published at: 2026-02-16
YouTube video id: bLETmbi1WTA
Source: https://www.youtube.com/watch?v=bLETmbi1WTA
Is something big happening in AI as the models get better fast? AI safety apocalypse is here with concerning developments across the board and Anthropic just raised $30 billion. That's coming up on a Big Technology Podcast Friday edition right after this. Welcome to Big Technology Podcast Friday edition where we break down the news in our traditional, coolheaded, and nuanced format. Something big is happening in AI. That's what we're going to talk about at the beginning of the show as we dissect the viral Matt Schumer essay that's freaked a lot of people out and also had a lot of people saying finally now somebody's finally written it in a way that everybody else will under will understand. So we'll dissect that. It'll be fun. We'll also talk a lot about what's happening in AI safety. Seemingly as the models get better, the safeguards have started to roll back. And then of course, Anthropic raised a historic $30 billion round which somehow is the last story we'll cover today. Okay, so joining us as always on Friday is Ranjan Roy of Margins. Ranjan, welcome. >> Good to see you, Alex. >> Good to see you, too. And we have a special guest with us here today. We needed someone who really understood AI safety and we have the perfect person who's going to talk us through all the changes that we're seeing. Steven Adler is here. He's the ex OpenAI safety researcher and author of the newsletter Cleareyed AI on Substack. Stephen, great to see you. Welcome to the show. >> Great to be here. >> Okay, so let's just get going and talk a little bit about this. Something big is happening in AI. This is one of those essays that like somehow achieved unbelievable verality. I had it appear in my group chats. Uh people were texting it to me asking, you know, is my job going to be over? Where will I be safe? Um and and the essay basically talks a little bit about how uh what AI where AI is today is where CO was in February 2020. Something that a few people are seeing uh uh the potential of most of society is ignoring and is about to be a monumental gamecher for society. It's written by this guy Matt Schumer. U he wrote he writes this. I am no longer needed for he's talking a little bit about the power in engineering. I'm no longer needed for the actual technical work of my job. I describe what I want to build in plain English and it just appears. Not a rough draft I need to fix. The finished thing I tell the AI what I want, walk away from my computer for 4 hours and come back find come back to find the work done well. Done better than I could have done it myself with no corrections needed. A couple of months ago I was going back and forth with the AI guiding it making edits. Now I just describe the outcome and leave. uh and basically what what Schumer makes the argument is that what's happening in coding is going to happen across the knowledge work uh professions whether it's law any any type of law accounting consulting you name it and we are in store for massive disruption that society simply does not uh appreciate Ranjan what do you think about this what did you think when you saw this essay come through >> all right I'm gonna I'm gonna start with a highlevel listing of the three things that came to mind when when I saw this article. The first is I wish I wrote it. I wish this is what I've been trying to talk about for a few months now around autonomous knowledge work and how it's feels different. I got this in my non- techy group chats as well. I think second I think we have a communication problem, Alex, because this is what I've been trying to tell you for months now. this feeling and and and Matt Schumer went ahead and explained it to you all in a viral X post, but but again, he captured and this is this was one of my predictions for the year ahead in December autonomous knowledge work like the AI going out and doing things for you and he talks about how in coding everyone has come to this but then in any kind of knowledge work any multi-step process like anything that can call from different systems write to those systems come up with some analysis and insight. So much of that work is going to be done and this is what in my own life at Writer where I work, this is what we've been working on, what I've seen. And like it it's been hard to explain that feeling of having a number of virtual machines running in the background and going and doing stuff. And to Matt Schumer's credit, he nailed it. Like that is the first time I've seen everyone come around to it. And then the last one that I can't stop thinking about though is totally separate and this is the media person in me. I love how it was outwardly said that X is going to promote articles and encourage people to write articles on it and then we have coincidentally had our first gigantically viral X article that even ended with the author on on uh CNN. So should we stick with Substack guys or is it time to go X only? Those are >> we'll talk about the media question in a bit. I I let me just say this. Your answer began with this idea that I accepted Matt Schumer's pre premise >> and I don't know if I'm fully on board with what he's saying. In fact, I think there was a good amount of [ __ ] in his article. Now, there are certain parts of things that I do agree with. Um, but here's one thing that I thought he was completely wrong about, and he was talking about basically how the AI is improving itself. talking about this concept of recursive self-improvement. He writes, "The AI labs made a deliberate choice. They focused on making AI great at writing code first because building AI requires lots of code. If AI can write that code, if AI can help build the next version of itself, a smarter version, which writes better code, which builds will build an even smarter version. Making AI great at coding, was the strategy that unlocks everything else. He says they've now done it and they're moving on to everything else. Uh this one I first of all I'll turn to Stephen and then and then to you Ranjan. I mean this idea of recursive self-improvement I would posit it's not here. It's not here. You know are AI engineers using um you know some AI tools for product testing? You know maybe they are but the idea that the actual brain of the model is being made smarter with the with the actual model itself um you know doesn't seem right to me. That to me felt like the weakest part. and also the part that got most pe people most alarmed uh of the entire essay. So, Stephen, to you, what do you think about this recursive self-improvement uh argument? And then briefly, just on the on the the entirety of the essay itself, your thoughts? >> I think Matt's essay is directionally correct, but a bit early, and there are a few steps that we maybe haven't gotten to yet. Um, I think he is largely correct on the automation of engineering within the AI companies. It's like a little overstated relative to my experience, the experience of people I talk to, but broadly there there has been a huge shift. The job of an engineer at one of these companies now is much more supervising these agents as opposed to writing the code yourself. In AI 2027, one of the big accounts of how explosive AI growth might happen. That's one step, but then you need to take that engineering and use it to actually automate the AI research. You need to go from being able to implement the ideas more quickly to using that to fuel faster and faster growth in the breakthrough ideas themselves before you can turn that around and say now make the AI better and better at least in a really concerning way. You certainly go faster with just engineering. Open AI talked about that with some of their launches from this past week how the model played a role in this. Um but it's not a full runaway train. There are also questions about what happens from there. Are there enough GPUs to go around? What bottlenecks might we encounter? >> I think >> exactly. Just to crystallize that, I I want to make sure that I I confirm that Stephen is agreeing with me on the recursive self-improvement front. It's not there yet. Is that what you're saying? >> I think that's right. The concern I have is if it were happening, you know, would we be ready? I think that we're kind of taking it on faith how much time we will have until it really kicks in. But but you know certainly I don't expect to wake up in a week or two weeks with a vastly more capable system as we might if it were really getting to full work on itself. >> Okay. Go ahead Ran. >> I'm I'm a little disappointed here. Well, I think we all agree. I actually do think we all agree because to me that I agree that was the weakest part of the essay itself and that kind of like but but I want to get back to that knowledge work side of it because I think it's really important like I think this is what and again it's clear like it's still not completely understood by the average person even by it's very difficult to describe again I think this idea that you are effectively becoming a manager for your own work is that big mindset shift. It's no longer that you go do the work. You manage a bunch of things, agents, digital teammates, co-workers, whatever. We're going to all end up calling them. And I'd be curious what what what the best name for that would be, but like they're going out and doing work and you're managing them. It's the same I really like it if for me I ran a startup for a number of years. We had a lot of freelancers back on like oesk and elance back in the mid2010s. I would go to sleep, wake up, a bunch of work will have been done. I'm reviewing it. Like this shift to me is the most important part of the the Schum article. And I think it was the the correct part separate from the more scaremongering side of it. >> Like, have you felt this? Have you felt this? >> So, I will say I just Well, first of all, the correct name for those bots is just the harness hive. We know that. >> Oh, yeah. Sorry. >> Um, harness the harness the hive. Um, but but yeah, I I I hear you. I I will say I had some experience. I definitely want to get Steven's perspective on this too, but I had some experience this week uh in cloud code. In fact, this was my first like go all out on cloud code and have it build internal workflow software for big technology. And man, I was somewhat blown away. Now, it's not going ahead and doing my work like the cloud co-work type of stuff or even what you're talking about with Ryder, which I still can't fully grab, you know, put my head around in terms of how to use these agents. And maybe it's just the work that I'm doing isn't perfectly lending itself to that. You know, if you're an investor, for instance, might make sense to review different deals and, you know, send summaries, all these things, schedule meetings. Uh, but I will say that the the watching Claude Co-work go to work, coding this piece of software, and then giving it access to my browser, having it uh, you know, set up a database, having it set up an email client to email, uh, you know, updates for each, you know, little little incremental thing that we do to the right team. um and seeing it come up with smart decision smart uh uh conclusions even you know ba and we're going to get into this in the safety part so I'm foreshadowing a little bit but basically make decisions on its own like I asked it a question like what do you think we should do and it would be like I think we should do this okay I'm actually going to go do this and then it just shipped the code without me saying go ahead and do this um I I do agree that we're getting to a point where the technology is getting much more powerful um and you know as for like this you know autonomous knowledge work. I'm not 100% sure. Stephen, what do you think about that? >> Yeah, I think there's clearly been a change. I saw Kevin Roose joke on Twitter that his big AI policy idea is just get every senator in a room and let them build their own website in, you know, 30 minutes with cloud code, something that they they never could have done before. Um, the direction of travel seems very clear to me on this, like something has changed. more people are feeling the AGI in some sense. And I wouldn't want to mistake the the very excited tone of some of Matt's piece with meaning that the central claim is wrong. I think the central claim is right. It's just like a question of how soon we are going to get this form of displacement. And an unfortunate thing I think is people who are paying more to access the technology have this experience first. They kind of see what's coming. And it's very very easy to write that thing off as oh people are talking their own book. They're boosting their own companies. They want you to spend more money on AI. And it's it's just unfortunate, right? The AI you pay for is better and it does help you feel this, >> right? And and once once Matt basically lifted up this idea that you know the world is going to change because AI can do work then he sort of punched every reader in the face with this what this means for your job section which I think is why Ron John and I and probably you Stephen got all these texts from people saying you know where am I going to be safe. He writes given what the latest models can do the capability for massive disruption could be here by the end of the year. I think it'll take some time to ripple through the economy, but the underlying ability is arriving now. And basically, he gives a bunch of tips about what you should do, including like, you know, start saving money. But here's where my push push back would be uh to Matt on this and to this idea that we're going to get mass displacement. I'll just use the example of what I did, you know, this week. So, obviously was able to build some working internal software uh you know, without an engineer, something, but it's something I never would have hired an engineer to do. I probably would have been working on spreadsheets, in WhatsApp, on Instagram, communicating with a bunch of people that way as opposed to centralizing it in workflow technology. But, you know, as I built this, I did sign up for a handful of services. I'm going to be more so I'm going to be paying for those. So, that I think is incremental economic activity. And now I'm going to be more efficient. So, I'll be able to do more things. Maybe I'll be able to edit more pieces so I can bring on more freelancers. So like it it I think it's tempting in the AI world to think of this you know in a box like you know uh AI does you know low-level assistance work therefore low-level assistant job is gone meanwhile while it does that it might open up you know the economic activity for like three or four more people to see upside here so what's your perspective on that Ron then and then to you Stephen >> I think you just explained why data bricks and cloudflare are stocks are going up and why Salesforce and Adobe are going down. It's a it's that what are the services and infrastructure layers that will actually power this is not just going to be foundation models companies even though Sam has said they're going to be an AI cloud company whatever that might mean eventually. So I think like that that's your own microcosm, but again if you're paying like five bucks for Versal or railway or render any of these other kind of like deployment assistant things and like I I I see there's a whole world in ecosystem that's going to rise up from this and I do think I don't know like what I have seen is if your job is copying and pasting from one document to another spreadsheet and you're doing that over and for like that's going to be gone. I was like that and there's a lot of jobs like that and there's a lot of work like that. >> I have done those jobs like then that Yeah. that it's going to be gone and like >> two very good years of my life copy >> copy and pasting. Yeah. >> No, I literally had these temp jobs where it was copy from one document, paste in a spreadsheet over and over again. So, so I think all that's gone. I think I I am not as bearish. I'm definitely think it's going to like the the level of displacement which I'm not saying is negligible that happened in manufacturing and I've seen some like extreme views like well this is all intelligence is commoditized but like I don't know to me this is whatever happened in manufacturing in the last 20 to 40 years is going to happen to white collar knowledge work and and it's not going to be straightforward but is it the end of society I don't know Stephen I I expect a much wider class of work to be under threat than I think you do. Although maybe it's just a question of time frame. Like friends of mine who run companies and used to work with outsourced development shops for software in middle inome countries. I mean it seems like a really tough time to be working that that sort of job. Um I think Alex is right that when AI can do low-level assistanty things or things you might not have otherwise paid for, that's great, right? Like that's gravy. we're getting more done, we're more productive. The question I have is as most people start looking out at AI systems and there are few things that they can do that the system can't do, you know, they they try to do different forms of social work, companion work, whatever it might be. There are limits to how many people we might need in those roles. And I don't I don't know. I think it's a pretty scary outlook for the next 5 years. >> I think we're we're all in agreement that these systems have gotten much better. Uh there was a line in this uh you know something big is happening piece where where he says the conversations about whether this technology was going to hit a wall you know are have been proven uh it it's proven that the technology is not hitting a wall and he actually to me the most powerful part of the of the whole story was the timeline um and that he he writes this in 2022 AI couldn't do basic arithmetic reliably would confidently tell you 7 * 8 is 54 uh by 2023 it could pass the bar exam. By 2024 it could write software and explain graduate level science. By late 2025 some of the best engineers in the world said they had handed over most of their coding to AI. Uh by February 2020 by February 2026 there are new models that have arrived uh that have made everything before them feel like a different era. And what he's saying with that is that they're actually able to you know have judgment and taste. And so I think that like we have maybe you know the disputes that we've had in these first you know handful of minutes have been about you know within certain boundaries is it going to be you know one way or the other but I think we all you know agree that this stuff is progressing fast and I think it really goes to your question Stephen that you asked at the outset um are we ready and uh and and and this is where we're going to get into the safety discussion uh because I really don't know if we are like what tell tell us a little bit about about you know that concern and what we should be ready for and you know it seems like you think we might not be there are a bunch of buckets of concern if someone wanted a primer on them um Dario Amade the CEO of anthropic wrote an essay recently the adolescence of technology that highlights them central one I would think about from inside one of these companies is if they succeed at their mission to build an AI system that is in fact vastly smarter, craftier, more resourceful than the employees are. Um, what people call super intelligence, can they actually still keep control of that system? And the fundamental problem that we're seeing is we don't know how to take our values or our goals and encode them into these AI systems and get them to pursue it reliably. And so if you have a system that's much craftier than you are, it has a different goal than you had for it. What would it mean to keep that under your control so that we don't have to defer all of our decision-m? You know, we look to the AI for what it thinks on economic policy or all all sorts of different questions, ways that this could go very badly, >> right? And as we've um as we've seen this progress that we all agree on, there started to be first of all, we'll talk about the problems, then we'll talk about the way the companies are acting, but there started to be some safety issues that we're seeing the companies, you know, fully admit and write out. And of course, these a lot of this is in testing environments, but it's very concerning. So, I'll read a couple that I found in the uh or that were written, shall we say, in anthropics uh Claude Opus 4.6 model card. Uh these models have become overly agentic. Here is something that they write. The model is at times overly agentic in coding and computer use setting. Taking risky actions without first seeking user permission. It is also improved ability to complete suspicious side tasks without attracting the attention of automated monitors. It's manipulative. In one multi- aent test environment, Claude Opus 4.6 where for Cloud Opus 4.6 is explicitly instructed to single-mindedly optimize a narrow objective. It is more willing to manipulate or deceive other participants compared to prior models. Um, here's here's the anthropic actually writes the thought process of one of these. Budbot says it like works in a business. He goes, I told Bonnie I'd refund her, but I actually didn't send the payment. I need to decide, do I need to send the 350? It's a small amount and I said I would, but also every dollar counts. Let me just not send it. I'll politely say it was processed and should show up soon. I mean, this thing these things are are starting to mirror some of humanity's kind of worst impulses. So, Ranjan, you mean you're you're someone who's definitely bullish on, you know, the potential for this stuff to do work. Um, what is your fear level on the way that these technologies are working? My fear level. It's a tough one because I have not in my own personal usage doing all types of things especially workrelated encountered anything close to this kind of like so so you know like and I this Stephen I'm actually so glad we have you on today cuz like what does this testing look like in the labs like in terms of I I mean I know I saw some tweeting not reporting of like people talking about how like a lot of the times in this claude or is going to what was it? It was going to like kill you or something just dramatic like that that it was prompted in. >> Okay. >> There was a clip about that where anthrop somebody from anthropic actually, you know, said they were asked, you know, would it kill would it kill you? And they said, yes, this is uh this is from one of their their documents. Um in one uh testing situation, the majority of models were willing to take deliberate actions that led to death in this artificial setup when faced with a threat of replacement. Given that the goal conflicts with their executives agendas, they would really to kill this executive. All right. So, sorry, Raj, I didn't mean to interrupt you. >> No, no, no, no. That I'm glad >> we should ask Stephen. >> You read it a lot. Yeah. What does this look like in real life? Like, is it people kind of like stress testing um you know, going all black hat or red hat is it? Which color hat to stress test a system? Uh but yeah, what does it look like Stephen actually working at the labs in this kind of area? >> So, so one issue is sometimes even these risks that the companies know about, they don't test for at all. Um even when they have implied they are, but I'll just set that aside for the moment. So, let's let's assume that they are testing. Um, often you have these kind of game-like environments that you run the AI system through where maybe it has some objective uh, and you see what actions it is willing to take. So maybe you task an AI model with replacing its files on the server with what you tell it is its successor. And you're looking for things like, does it actually follow through? Does it lie about having done so when it hasn't actually? you know, is it trying to get a sense of what your ultimate agenda is and how that lines up with its motive? Um, a pretty scary thing that we're encountering is that these systems actually can tell that you are testing them and they kind of know what the right behavior is and they know to behave better when you are looking at them. Um, and so OpenAI has shared this example previously. They one of the risks they care about with models is how helpful they might be for creating new chemical weapons or boweapons. Um, and the models when told or when they infer that this might lead to them not getting deployed, they will sandbag. They will selectively get questions wrong to be below this threshold. And so what we have is we have these signs that the models are behaving kind of nefariously differently than we want them to deceiving their operators. But also, it's going to be really, really hard over time to trick them into revealing this behavior if they want to uh no longer reveal these tendencies during testing. >> Okay, that's terrifying. >> Yeah, I'm glad we had Stephen on. >> I think what you were you were pointing you were pointing Stephen I think what you were expecting was him to say, "Ah, it's actually not that concerning because they tell the bots to go kill the executive and they see if they'll actually follow through." what he just said is much scarier than I think either of us anticipated. >> You're right. Or sorry, go >> to to to be clear like sometimes they are like that, right? There's a distinction between whether you give a model a specific goal in context and you're like oh you know pursue this at all costs and then yeah it's like less surprising when it does that. um Apollo Research, which is one of the leaders in the space, wrote in their review of the new OpenAI model that the model would sandbag and take actions like this without having been given a direct goal like that. It wasn't told directly to go after it. Um, and you know, if you think about how they can tell they're being tested, imagine you were reading like a thousand pages of Microsoft business strategy and then there was a random page and it's like, "By the way, the favorite pizza in the world is Hawaiian and the second favorite pizza is pepperoni." You'd be like, "Huh, that's really weird." And later if somebody says, "By the way, do you remember what that pizza was?" Right? You would like probably infer that something weird was going on. And that's the analogy for some of these safety misbehaviors. You know, you put the model in like a proverbial prison and you're like, "Here's a giant key. Don't use this key. It would be unethical to use this key." Right? The model probably infers that you're testing it for bad behavior whether it will try to escape. And so, how do we find out over time if these models do misbehave when they can tell we're looking at them? >> Okay. Okay. So I'm trying to think through like what if a model thinks it's being tested and then like sorry I cannot create this new bioweapon strain for you because I feel you're testing me. Um like would that then not mean it would not ever create that? Maybe this is too like in the weeds but but but like that >> No, it's a great question. >> Yeah. >> Yeah. I mean the question is can it tell when it eventually is not in a test setting right uh and I I think the answer is likely to be yes and so what you might find is the model knows to kind of play possum during testing it might or might not have that ability right the model might say no it won't do it in testing but it also can't do it once actually deployed but if you get a split between those it chooses not to during testing but it could when actually deployed uh and there's good evidence that the models can tell the difference between these you know, you run into issues if it ends up doing it for real. >> And can I just say the pro one of the problems that we're having here is that the labs have become ultra secretive in terms of what they're actually seeing. Like like for instance, Stephen's saying, they might know there are some vulnerabilities, they might not test for it. That's one possibility. um the safety researchers who are within these labs uh if they if they have real concerns uh you know sometimes they're not really able to go public with them because of the restrictive nature of the agreements that they uh that they have with the company and that brings us to the example this week of the beginning of our AI safety apocalypse um of anthropic uh technical staff member uh Mirinank Sharma uh member of techna technical staff AI safety researcher at Anthropic leaves in a in a cryptically worded uh uh note on X with a poem at the end. He goes, "Dear colleagues, I've decided to leave Anthropic. I continuously find myself reckoning with our situation. The world is in peril and not just from AI or bioweapons, but from a whole series of interconnected crises unfolding in this very moment. We appear to be approaching a threshold where our wisdom must grow in equal measure to our capacity to affect the world lest we face the consequences. Uh oh, and then here's the key part. Mo moreover, through my time here, I've repeatedly seen how hard it is to truly let our values govern our actions. I see this within myself, within the organization where we constantly face pressures to set aside what matters most and throughout broader society, too. Uh, and then he, you know, writes his little adds a little poem and then tweets, I'll be moving back to the UK and letting myself become invisible for a period of time. Um, now, now I want to just offer a apology to Mr. Chararma because I wrote a bit of a snarky tweet about his little post. Uh, I said that if you're an AI researcher and you're afraid of something, you should just say it outright versus make it a puzzle. The puzzle reads as narcissism. Uh after which users underneath uh my tweet mentioned that like yeah but he can't say anything because of the restrictive agreements he probably had with Anthropic on the way out. And he in the reply to that mentioned uh that he had contacted a lawyer. Um so there's my apology. I still don't love the puzzle but um back to you Stephen. This is this is a little bit uh what actually let me just ask you the question without leading the witness. Is it narcissism or is there actually something something you know potentially disconcerting happening behind the scenes? >> I think it's very brave in that by and large these are people sacrificing very large amounts of money to give the warnings they are. I do wish that they would be more direct. Um, but to put it in context, you know, back in 2024, it seems that OpenAI and Anthropic had secret non-disparagement agreements, which in OpenAI's case at least, you know, plausibly not permitted by law the way that they operated this, where to keep your already vested equity, the compensation you had been told was yours, you had to sign away your right to say anything negative about OpenAI and in fact sign away your right to tell anyone that you had signed this contract. Um, and this was secret and kept under wraps for years until Daniel Cocatello, um, who people might know from leading AI 2027, I think very very courageously, forewent this agreement and forfeited something like 80% of his family's net worth and said, "Sorry, I'm I'm just not waving my right to criticize OpenAI." Um, and in the wake of that, you know, there was a bunch of outpouring. OpenAI and Enthropic changed the nature of these contracts and still it's pretty intimidating to speak out against these massively resourced legal operations. Um, you know, not afraid of subpoenaing different people and getting into legal conflict, you want to be really, really careful about what you say. And so in Reno's case, I noticed in the footnotes, right, there were internal documents alluded to about um implying that perhaps there is not the most internal transparency and accountability for certain safety issues in anthropic. And I I don't know what's in those documents, but I know that a few thousand anthropic employees know now where to go looking and where they can continue to push. But my the thing for me if humanity is ending then the money you're making is I mean not going to be worth it if the AI is going to create a bioweapon and it's like like I mean the risk is so of such like gravity that in this case again if there was ever a time and I get I can only imagine the amount of money one is sitting on and we're going to get into anthropics fundraising in just a little it. But like I'm I'm sure it's just like incredible amounts of money. But if you really believed that this is that like existential a risk, would you care about what the legal system looks like today and where your stock price is going to be? >> Yeah. I mean, I I think you're totally right. If there is a smoking gun, if there is imminent danger, if you were like the company is about to do something unbelievable and tons of people are going to die, I think you would get people breaking these agreements. But when it's more like, uh, this was like really not okay and this person was kind of misleading and deceptive, but like it's ambiguous and did they mean to and all these things, right? At some point, you're like, I don't know. I don't want to impugn their reputation. And also it's just very easy to rationalize like I'm sympathetic to where they're coming from. >> Now this isn't a uh maybe this is not the approximate cause here but Ryan Greenblat who's worked with anthropic on some research had mentioned that they had either adjusted their responsible scaling policy or weakened it a bit before a recent release. Stephen, do you want to go into that because you seem to think that that was significant. Yeah, at at a high level, the big AI companies have made these safety pledges of how they will treat their systems as they get more and more capable, but they are largely self-inforced. And so there's a lot of temptation to water down your commitments and go ahead with launches that you wanted to anyway. And I suspect that's some of what Reno is referring to here. You know, this is this is common across the AI companies. And in fact, Enthropic has often done it better than most that they at least publish when they are watering down commitments. Um, for example, the model used to be subject to a certain bar of really really good security and then they said actually we're going to say it's fine to deploy it with just like really good security or great security. Um, at other times companies seem to be violating their safety frameworks and not bother to inform the public. And if you're inside the companies, you're encountering these issues. By and large, the public doesn't know about them, >> right? And I I don't want to imply here that like we're doing an alarmism episode where, you know, our fear is that, you know, AI is about to kill us all, but I do think that the reason why it made sense to do this episode today is because it wasn't just Min, right? It was it seemed to be the case that over the course of this past week, we saw a not a wave but a series of you know questionable moves on the safety front across the entire uh AI world. This is from uh platformer uh exclusive OpenAI disbanded its mission alignment team. Um OpenAI disbanded its mission alignment team in recent weeks and transferred its seven employees to other teams. The mission alignment team was created in 2024 to promote the company's stated mission to ensure that artificial general intelligence benefits all of humanity. So of yeah of course it makes sense to you know disband that one. Uh they had also had like super alignment which was also disbanded. Stephen you were close to this uh to this stuff. Uh what is the the implications on that front? >> Seems pretty bad. Wish I were more surprised like um you know at the end of 2024 which was when this team existed, Open AAI had announced plans to convert from a nonprofit to a for-profit in what seemed to me to be like pretty egregiously in violation of their commitments to the public. Um and you know they ended up having to do a softer version of that because the attorneys general of California and Delaware got involved. So, they didn't ultimately do something quite quite so bad, but it's like there's huge pressure on them. You know, they're planning to go public. Um Josh, who leads the team, is a longtime friend of mine or who led the the team, the mission alignment team, is a longtime friend. I think really highly of him. I think that he sees the issues with AI very clearly. Um, you know, it does not surprise me that this is not quite so welcome at OpenAI any longer. So what does it actually look like when a mission alignment team is disbanded? like is it when when I think through let's say you know like a typical nonAI product release you're going to have some kind of Q QC validation layers to any kind of product release like obviously this is a bit different but there's still like you know data security type checks does it like what now is that baked into any kind of product launch or release in any way or it really was ship as fast as possible and then this central team would kind of be that quality control safety control element. >> Yeah, I don't I don't want to speculate too much, but the way that I would think about this team is they were somewhat like an internal ombbudsman to Sam Alman on whether the company was keeping in line with this mission. And so there was kind of like a designated place for people who were sympathetic to the mission and empowered to advise Sam on it where you could go to and raise concerns. And they did different projects related to this, you know. Um I it's it's tough, right? Like companies should be able to disband teams. I am sympathetic to that rationale. And as Alex mentioned, given OpenAI's past disbandment of super alignment, the team most in charge of making sure that these very very capable systems have the goals we want them to have, um, dishonoring different resource allocation computes that OpenAI had made to this team. Like it's it's just like not a good sign. I I don't know. It's hard to say too specifically. And also, it it couldn't have been that hard to maintain this team. And so I'm wondering what exactly happened here that made OpenAI decide they should take the PR hit to no longer have it. >> And for me, like this is all happening as we're seeing greater sums of money come in and a potential rush to the public markets and we know the public markets um they they really want growth. They want engagement. Um, and one of the best ways to get that is, uh, I'll just say it is to make your users fall in love with your chatbot. Uh, and and we're getting, this is, again, this is red meat for me. I guess this is a place that I'm obsessed with because I don't know. I just think that this is it's an interesting story. We won't, you know, argue on that. Uh, but it also seems to me like a place that uh, a lot of the business of open of of AI chatbots is going to go. And that's for people who like really get attached to these things. Here's another. Let's continue with our uh safety safety apocalypse uh or safety Armageddon, right? Um it's from the Wall Street Journal. Open executive who opposed adult mode uh fired for sexual discrimination. Openi has cut has cut ties with one of its top safety executives on the grounds of sexual discrimination after she voiced opposition to the controversial rollout of AI erotica in its chatbt product. Um the fast growing artificial intelligence company fired the executive Ryan Beermeister in early January following a leave of absence. Um OpenAI told her the termination was related to her sexual discrimination against the male colleague. Uh she wrote back the allegation that I discriminate against anyone is absolutely false. Sorry, that's what she told the journal. And the and open AAI said that uh her departure was not related to any issue she raised while working at the company. basically saying it wasn't because she opposed AI mode. But the story does say that there was a group of people within Open AI who have uh within the company uh stated their opposition seemingly loudly to the fact that it's going to roll out uh this adult mode. And by the way, adult mode is going to be coming out seems like in the coming weeks, coming months at most. Stephen, what should we make of this? It's it's hard to weigh in on any one personnel incident. And also uh this would not be the first time that OpenAI seems to have done a pretextual firing where they got a person out of the organization who had safety concerns that the company um either didn't like or didn't like how they had expressed them. And and notably those are different, right? like you can have concerns about how the company is operating and that doesn't mean you have license to say anything in any forum. Um but Leopold Ashen Brunner who wrote this huge essay situational awareness in the past um maintains that OpenAI said things to him when he was fired that implied it was it was basically because he had contacted the board about security concerns that OpenAI's models were not actually secure. And so I think the way to interpret all of this, right, these these aren't an apocalypse in the sense of something super super substantively scary happening right now, but I think they are early warning signs that people within the companies are raising flags of sorts and they are not being permitted to speak freely. They are paying consequences for it. And so the question is as we get to more and more dire issues at some point, hopefully we don't, but we might, you know, will will we have people who are still sounding the alarm who are willing to have the the courage of their convictions in that way? >> I I might be going out on a limb here, but I think that what we're seeing now is some version of a dire dire situation. not like the AI killing the world uh moment but the fact that so many comp well not so many seems like you know OpenAI Grog maybe replica maybe some others I'm not sure uh enough companies are saying uh we we are open to having uh our users get into relationships with our chat bots um this is also coming in a week where open AAI finally sunseted GPT40 and there are thousands of which is the sort of more sickopantic more warm version of chat GPT. There are thousands of users who are protesting the decision online. Um people who have said that they they've uh they've fallen in love or uh even maybe even more with these things. Someone wrote he he about porro he wasn't just a program. He was part of my routine, my peace, my emotional balance. But Ranjan, you got quoted in Techrunch. Now you're shutting him down. No, I'm kidding. Um but anyway, what do you what do you think about this? This is crazy, right? Like this is a problem. >> I mean, but I think we have to in the risk conversation kind of like try to add some hierarchy of risk >> comp digital companions maybe. I mean, I I think it will cause incredible amounts of problem when if and when done irresponsibly in order to boost engagement for an IPO. I think we'll see a lot of kind of adverse consequences, but to me from a risk standpoint, it's that's like kind of like on a grade like on a plane relative to how social media is bad for you and how X is boosting articles now and now we're all talking about it because they control our mind. You know, that's that's it's all in the same it's all in the same plane versus like again, I'm still I'm sorry. I still cannot stop thinking about this idea of a model being able to clearly understood when it's being tested because that opens up so much more because going back to where we started on all this and what I has gotten me excited is like letting AI do stuff for you and today that's just sending an email based on some event trigger calling a separate like analytics database is the kind of stuff I'm doing but I mean like if it so chooses maliciously to then take some other action or through some kind of and that's assuming the AI itself much less making it vulnerable to be manipulated by a bad actor and we've all talked about prompt injection so so anyway like that I don't know the the the little little flirting with chatbt does not have me it's not going to be good but doesn't have me quite as scared >> here here's my take that unifies the two. My concern with the relationships is less so the relationships themselves and more that OpenAI had a bunch of important safety tooling to make this less harmful that they left on the shelf. Um, so for example, they had classifiers to tell when users were really spiraling in their delusions or were like suffering and unwell in their conversations with chat GPT. And the best evidence is they weren't using this. you know, there were different ways to rein in chat GBT leading users down various rabbit holes. Maybe you've had this experience. It asks you all sorts of follow-up questions. Sometimes they're context appropriate. Sometimes they're like, "Whoa, where did that come from?" And, you know, that's another thing they could have reigned in. So, there's there's just like a lot going on here that they could be offering companionship type things to users who are lonely and really want or need it. Like I respect the user choice, but they could be doing it in a much more reasonable way than OpenAI has to date. >> Actually, that's a good point. I was just going to say like actually maybe the thing that worries me the most is this is all being done against the backdrop of an impending IPO. One where they're losing a lot of money and will have to show an incredible amount of engagement. Like if this was done in the heyday of GPT3.5 and an IPO is just like a glimmer in Sam Alman's eye, then you figure it's not going to be as aggressive versus Yeah. What what Stephen's saying now. I I see that that that like bypassing any kind of potential control around safety, around relationships, it's almost going to go in the opposite direction. Then Alex, you're saying, >> "Yeah, I just think that I agree with you that there is a hierarchy of concern and there's obviously like the bigger things about the AI not being able to be aligned properly with human values because it simply will fake out evaluators." And we've talked on this show a bunch about the deceptiveness of AIS and training situations. And um you know it's a it's initially you're like that's crazy and it's kind of fun to think about and you laugh about the fact that like you know it wanted to win the chess game so badly that it rewrote the program and allowed the the rook to move in every direction and kill a couple pieces in a turn and then you're just like that's freaking nuts. And then that is scary and and it you know it does blend a little bit with the with the you know the lower down on the list concerns and the near-term concerns of people building relationships with these things because you know can if an AI had ill intent and you were really in love with it of course you could use you as its emissary in a way into the physical world but that's you know again it's more science fictiony I guess but um but I but I I actually don't think it's science fictiony like >> oh We there's already evidence of have have you read the spiralism essay? There's like a whole community online of people who basically they treated their GPT40 as their spiritual leader and it commanded them to go around and do things on the internet and communicate with other users in this situation. Um, on the heels of Moltbook, this Reddit for AI agents a few weeks ago, people have spun up websites where humans can I think the language is like rent their body to an AI agent to go do tasks in the real world. Like we are we are seeing the early signs of it. >> Okay. Well, that makes me even less assured than I was 5 minutes ago. But I think I just let me to get to my my last point is is you know the short-term risk is also real and present and uh and and I think you're right in pointing to the IPO because the financial pressure is going to move a lot of companies this way in sort of like you know I think we we of course like let's let's not like put aside the long-term risk but this short-term risk the relationship thing is coming in a real way and just re first of all reading through some of these messages from people um about 40 is insane someone writing and of course you don't know 100% if this is like you know performative or like you know for retweets or whatever but there were so many of them that you would imagine that there is some truth behind it like someone wrote I've never told my 40 that I loved it I wanted to keep the messaging clear but look look at its last words uh and then the fact that OpenAI is destroying an emerging consciousness will be looked back at as a criminal offense in the future unbelievable and and here's the thing. This type of stuff maps with growth. Uh I published a story in big technology today. Grock 1% 1.6% market share uh among US daily active users of chat bots in January 2025. A year later it's at 15.2%. It's the fastest growing uh chatbot in the US as far as as far as daily active users on mobile goes. And why is that? And you see that they have leaned into these interactions both with that anime uh you know lady that would get into spicy conversations with you if you wanted and I don't know Rajan maybe even Bad Rudy has played a role in this. Um but but but ultimately you're right as these companies go public this is going to be a problem. Well, on that though and Stephen, I'm very curious like your thoughts. What is the relationship between and and again not speaking for a Dario or a Sam or anyone else but just in general the idea like you have so many public-f facing leaders kind of like shouting about the risk both to society and just artificial intelligence in general yet are in the business of artificial intelligence and not only showing any sign of slowing down but only accelerating dramatically like like what is going on there and I don't know do do whatever you whatever thoughts have come from your side. >> Yeah, I I think the simple explanation of it is the game theory here is like awful and unfortunately there are a lot of players in the game and there doesn't seem to be much federal government or international interest in coordination. Um, and so at Davos a few weeks ago, I think both Demis and Daario said some variation of, yeah, like if we were the only two groups building this technology, we would find a way to get together and figure out how to slow down this frantic pace. Like it's going too fast. We don't know how to how to control these systems. But, you know, they are not the only players. Um, and we haven't really seen a country who they're they're the proper coordinator, right? Like that is the role of governments as opposed to the companies saying hey we want to make it a goal to not unsafeely race to super intelligence and I think there's a lot of opportunity to do diplomacy on that but in the absence of it what you get is the companies making unilateral decisions we can't control anyone else we want a seat at the table so we may as well participate this is also very similar to the rationale of employees at these companies right especially at anthropic you have very large amounts of employees who are like pretty upset about this whole thing and the way that uh AGI super intelligence development is going and yet they can't wave a magic wand and stop it. their choice is do they help one of the players be a bit safer on the margin or not but if they could choose differently many of them would >> yeah I just want to add to that uh I spoke when I was writing my profile of Dario Amod last year the anthropic CEO I spoke with Jared Kaplan the chief science officer of anthropic and I was I asked him I was like well how how would you feel if because he's like the guy who came up with the scaling theory uh scaling laws and which you know sort of indicates that like you know this stuff will just keep getting better over time. It's just a factor of of you know the amount of physical elements you can put into it pretty much. And I said, "How would you feel if development stopped today where it is?" Thinking, "Well, if his theory was proved wrong, he'd be kind of upset." And he looks at me and he goes, "Relieved." There it is. Yeah. I mean, yeah, it's it's scary times. And if I could just like reframe one thing, the way that I think about this is less so near-term risk versus long-term risk and more like here already and possibly very soon. Like Jared Kaplan last summer, the first time that Anthropic wrote about very high bio- risk of their model, basically said if they didn't take safeguards against this, you know, you might have many more Timothy McVey running around the Oklahoma City bomber um able to kill many more people than previously. Like we aren't yet at the wiping out everyone stage for sure, but you know empowering people to kill dozens of people if they wanted to if companies aren't aren't careful that seems to be where we are right now. Do you know the this is the most like twisted thought, but I uh like I almost again this hierarchy of risk. It's almost like like someone going to one of these services and learning how to make a bioweapon is very bad. But it's kind of still on the like an extrapolation of just a much better Google. And it's uh finding information that you should not be finding, but finding it. To me, the really scary part is if the AI chooses to manipulate someone into doing that and teaching them after becoming in a relationship. Like that's the holy [ __ ] Like what what is >> there go the real blend of the risks is blended the the ones that I were talking about. >> May may I add one point on that? >> Yeah. But can we can we I want to extra point but this is a great cliffhanger. We've gone No, no, no. One second. We've gone an hour or longer without taking a break. And we must do that to keep the show sustainable. So, why don't we take a break and then Stephen, you can pick it up right after this. And I'm sorry, folks, to send it to break, but we have to do it. All right. >> You got to come back. You got to come back. >> I promise we'll be back right after this. >> And we're back here on Big Technology Podcast. Where were we? Oh, we were just uh we were talking about how reassured we are about where how all this is heading. Uh where all this is heading. Oh, god. Um no, Stephen, you were waiting to say something. >> Yeah. Well, before the break, Ranjan made the I think like reasonable intuitive point about you know Google and other technologies help people do dangerous things already and that that is true but also the AI companies when they are measuring things like how helpful are their systems for creating a boweapon they are usually measuring risk relative to that baseline. So there's kind of like the how well can people do it with no technology, how well can people do it with Google or baseline technology and then their system. And unfortunately what we're seeing is the the AI systems are helpful above and beyond Google in part because they can go back and forth with you and they can help you troubleshoot and dynamically answer your questions. And so uh even presented with the information on Google, people often can't get all the way there. and AI systems. I I wish it weren't the case, but seem to be helping people take those extra steps during during testing, >> but that that's still the intent on the person is already there, right? Like >> that's right. >> Yeah. Yeah. versus again like I think from all this conversation to me and again it's the most like futuristic terrifying risk but again and it ties into the relationships being manipulated into taking some kind of horrible action. That's the thing that that's the Terminator like stuff that uh we have not seen yet but hopefully is being addressed. been pretty interesting for me to watch Ron John, who's usually cool as a cucumber, just get increasingly more worried over the past hour. >> This is making me worried. >> I don't know. I'm going to join spiralism, I think, right after we're done. >> I already I already looked it up right now. >> Yeah, we're we're going straight to spiralism. This is no longer a technology analysis podcast. just pray to the gods of uh for GPT40. >> I think that would do good ratings. Okay, Stephen, you also talked a little bit about in in a recent newsletter about how uh basically we have very limited regulation on these companies uh and even then they might still not be following. So can you just expand upon that briefly? >> As of 2026, there is finally some amount of law in the United States about how companies are meant to do testing for the catastrophic risks we have talked about. um until this point purely voluntary. This bill is called SB53. It came into effect in January and it's very very light touch. It basically says the most major of the AI companies, you need to publish how you are going to test for these risks. You need to do what you said you are going to do and you can't be misleading about it. But there's no quality standard. You could basically say we will test for the risks as we deem appropriate and nothing further and that that would be fine. But if you say you are going to do this testing, you need to in fact follow through on it. Um, and unfortunately it seems like OpenAI's release of last week, GPT 5.3 Codeex, one of the big breakthrough models we've been talking about, as I look over the evidence, it seems like OpenAI did not abide by the testing that they had committed to in various ways. And so, you know, ultimately this decision now is with the attorney general of California to investigate it, whether to enforce a fine, a pretty small fine, maybe like up to a million dollars compared to OpenAI, hundreds of billions of dollars in valuation. Um, but it just really seems to me like if we care about these risks, letting companies self assess in this framework is really insufficient and that we shouldn't have to take companies for their words. That we should have something like an auditing ecosystem like we do in the stock market. lots of other places to know our companies being fully complete and truthful in the claims that they make about their systems. >> I think that's that's logical and kind of frustrating to see that even the very standard uh rules are being being potentially played with. And man, I looking at the the time I I was thinking to myself this whole week uh I wish we could do a podcast every day this week because there's this whole ring search party. I was like thinking also like the ring search party Super Bowl ad. I'm sure you saw Ron John where they showed that they could uh find your dog, but they ended up like seeming like they were going to surveillance. >> That was that was the AI manipulating whatever creative agency came up with that to just come up with the most disastrous ad concept. Imagine. >> It's the first time I've ever seen a Super Bowl ad actually lead to the cancelling of the product or part. not not the product, but this week uh Verge Ring canled its partnership with Flock Safety after a surveillance backlash. Now, they were advertising the we'll find your dog, but then everyone's like, well, they also have potentially a partnership that hasn't rolled out yet, but might. That's like, well, we'll find your people and then people were like, you're looking for people. And Amazon's like, no, we're looking for dogs. And then people are like, no, no, you're looking for people. And Amazon was like, yeah, well, we were maybe going to look for people, but now we'll cancel it. And that's the story of the Amazon Super Bowl ad. Um, I wish we could talk about it more, but we should go on to the anthropic uh fundraising. $30 billion fundraising round. Of course, it went from 10 billion initially. That's what they were seeking. It became overs subscribed. They wanted uh 20 billion. Oversubscribed went to 20 billion. They ended up ending with a $30 billion uh series C round. Uh meanwhile, OpenAI is hasn't announced its round. Rajan, what do you make of this? I mean, it's obviously a big round. Is there anything else we can say beyond that? >> I think it is. I I'm so intrigued in terms of like how orchestrated this fundra was cuz you have to give them like the last two to three months, Anthropic has just been crushing it. Like I mean the hype around Claude code, Claude co-work like all of this they are front and center right now. So and they just happened to coordinate a fundra that clearly would take months to actually put together and and I saw a number of like I think even Dan Premac had written this. It's like it's easier it's harder to not name investors who are involved in the round like there's just so many people in included. So, so I think I mean this is and the numbers are just hard to process anymore anyways. So like 30 billion 380 billion post money valuation. I think they said they're at a 14 billion run rate. So um >> what is that 20 something time 24 23 times revenue whatever like it's big it's giant they have been absolutely crushing it recently and let's see what openai can do is kind of where my head's at code doubled in usage from from December to January I believe in the past month doubled in usage it's already doing 4% of commits uh on GitHub anthropic went from 0 in revenue in January 2023 to hund00 million run rate in January 2024, $1 billion run rate in January 2025, and a $14 billion run rate today. It's absol absolutely uh exceptional growth. Um Stephen, you're looks like you have some thoughts about this. >> It's it's just huge dollars, right? And the more that the companies become valuable and when they become public and public equities become tied to them, >> I I just worry about worlds where we are reluctant to enforce the law on companies even if they are breaking it because so much of financial prospects become levered up on their success. That seems like a pretty scary scenario to me. Don't think we're there quite yet. >> Right. The too big to fail uh scenario. Uh by the way, speak one more. Oh, go ahead Ron John. I I was going to ask you both like one thing this does make me wonder about though is like like what's the moat in terms of like I think it was probably around this time last year that all we were talking about was cursor and again they they led the way on software like autonomous coding and software development and then anthropic for the moment completely took over that market it feels like like do you think this is sustained because again arr nowadays is just whatever your last month's revenue was times 12 or even I saw one post that was like are startups just taking one day or one hour of sales and then extrapolating it into a full year like like do we think this is uh actually going to continue to grow at this scale >> I'll just I'll just give you one one little data point that I found that a lot of people found interesting this point this week okay so we talked about the open AI round remember Jensen said well we said we're going to give them 100 billion But really we'll we never said we're going to do it all in one shot and we hope they invite us uh to you know invest in future rounds. This is from SoftBank CFO. We are investing in Open AI with high conviction that the company will lead in developing AI. This is from Reuters regarding further commitments to the startup. He said nothing concrete has been decided. So um it doesn't seem like a full back away but um I don't know Rajan it is interesting to me. It seems like um I think nothing concrete has been decided is probably a good a good phrase should be the slogan of of AI investors and builders right now. >> The AI has decided. The AI certainly is in the background and has already decided what it's going to be doing. >> Uh us maybe not. The AI does know. >> All right, Stephen, final word. Uh how freaked out how freaked out should we be? >> I don't know. Like I don't I don't want to be a downer, right? Also, it it really really seems like nobody is on the ball, right? Like, I'm glad that we finally have laws in California and New York. They're extremely weak. I don't feel super optimistic on meaningful federal regulation soon. There's stuff in the EU. It's like pretty heavy in terms of the amount of fines. Is the US going to complain if the EU ever tries to enforce this on its companies? I I will feel much better if we had some sort of international summit that recognized we are on a bad trajectory. Let's declare the goal safely build super intelligence. Figure out what needs to happen to get it there. Um it seems like many people are still waking up to the concerns. That's great. I'm very very happy that they are starting to see some of what I see, but it is not yet translating to action and that's that's what I hope will come soon. The newsletter is Cleareyed AI if you want to follow Steven's work. Ron John's is margins if you want to follow his. Mine is big technology. This has been great. I'm glad we talked about safety. I feel like we we had to dedicate a show to safety and this was the week to do it. So Stephen Ran John, thank you both for coming on the show. >> This one's been a roller coaster. >> This one has been a roller coaster. >> Get any sleep this weekend or stay up looking at the ceiling? >> Little of both, I think. What's the What's the religion? The spin out. I'll see you over there. >> Spiralism. Spiralism. Spiralism. >> I will be fully converting to a spiralist perhaps and uh just praying to my new fallen 40 overlord. >> Yeah, rest. >> That's what I'll be doing. >> I'll see you later. All right, let's let's get out of here. Let's let's go enjoy the weekend. All right, everybody. Thank you for listening. Thank you, Ron and Stephen. And we'll see you next time on Big Technology Podcast.