OpenAI's 2025 financials LEAKED

[00:00:00]Good morning. It is June 16th.

[00:00:03]Uh we got a lot to talk about today.

[00:00:05]Primarily is that a bunch of you are a bunch of pansy-ass who can't handle nuance. But uh no, the big question on everyone's mind is do I have AI psychosis? Um yes, I do. I have AI psychosis.

[00:00:22]Sorry to say, I have to shut this channel down. People with AI psychosis cannot be operating YouTube channels, man. Look, I would actually you know, if this was old me, I would sit here trying to defend my video from yesterday saying you guys are a bunch of idiots. You didn't watch the whole damn thing.

[00:00:36]Clearly you didn't because the video is 24 minutes long and by the 1-minute mark people are like, "We've lost him. We've lost him. He called Fable beautiful, but I didn't listen to anything else he said in the video.

[00:00:48]We're done. We had a bunch of outflows of we had a mass migration yesterday of subscribers. Let's take a look. Oh, we're in the positive. We're positive two subscribers. We were in the negative yesterday for quite a long time, which I honestly don't don't mind. To to me it's it's a positive signal of getting rid of the the simpletons who can't can't tolerate a a like anything that you know, I'm not even Honestly, my stance has not changed at all from the start of this YouTube channel. Like you guys see different videos and you're like, "Oh, today he's positive. Today's negative." It's all the same stance. It's just Like I can channel I'm I'm skeptical by nature. I can channel different views. I don't just have one stable opinion about the world. World is complex. If you have a stable opinion, you're an idiot. I don't know what I shouldn't be insulting my viewers. You guys are guys are lovely. No, the world is very complex. I- I- If If every day you don't feel a little doubt about what you felt yesterday, then you know, well, good for you. First of all, I envy that sort of position. Uh I I I prefer to have a stable position. But no, I To me it's a little complex and that every day I feel a little different. And you know, for for a string of videos, I I can channel the negative funny stuff for you. It's not me being insincere. It's not me farming for views. It's just like, you know, if the audience if this is what they want the they want the funny videos, blah, blah, blah, I can channel just that side of my brain. That's fine.

[00:02:10]But, uh eventually the the equilibrium must be found. Anyway, enough about that. Let's go over some news this morning. There's two big things we got to talk about. Uh one is Ed Zitron um has released an exclusive got a hold of Is this it? The Open AI financials. And in 2024, Open AI lost $5 billion.

[00:02:32]They had 3.7 billion in revenue, 12.4 billion in costs and expenses, and a net loss of 5.09 billion.

[00:02:39]And in 2025, Open AI lost $38 billion.

[00:02:43]That is insane.

[00:02:45]These are whole market caps of public companies like, you know, reputable private companies. Well, I shot I don't know what's worth 30 Everything is inflated. I mean, this would have been like Cloudflare's valuation 2022 or something, 2021. Now, Cloudflare, I haven't checked in a while, but it's probably like something in the hundreds of billions. You know, we're so numb to these kind of numbers, 38.5 billion. This is a lot of money.

[00:03:08]This is an insane amount of money.

[00:03:09]Interestingly, the revenue was 13.07 billion, but the cost of revenue was 7.5 billion. So, if it weren't for these crazy ass other expenses, it's it's conceivable that subscriptions are profitable. It's just that, you know, R&D, presumably training the models, goes in here. Um very expensive sales and marketing. So, you know, they're a far away from prof- profitability. A far, far away. Again, Open AI lost 38.5 billion dollars in 2025. I cannot wait to get a hold of Anthropic's numbers, man. Those are going to be interesting. I imagine Anthropic runs a tighter ship than Open AI cuz Open AI is out here acquiring podcasts and I don't know what they're doing. Nobody knows what they're doing. They're working on a adult mode, which you know, on the one hand, you know, thankfully they canceled, but I know a lot of you just wanted to goon, and sorry about that. I might have used it a couple times just for research purposes. So, it is a little bit unfortunate, but it's for it's for the betterment of humanity.

[00:04:07]Come on, guys. Like, you know, I'm sure we've all tried other models to see how how X-rated they can be. Again, for research purposes. It gets old, man.

[00:04:16]It's It's like cool one or two times, but it gets really old. It's like text.

[00:04:19]It's like Anyway, moving on to Anthropic. Still still nothing. Still nothing major has happened. Fable still unavailable. No good leaks have dropped.

[00:04:28]It's just the White House saying something Anthropic just rumors on what they're saying, but it it it comes down to what appears to be a game of politics, a game of power. Apparently, Amazon did in fact approach the White House not to the opposite, not vice versa. And the extent of the the security issue is that when giving deliberately insecure code, she said Fable refused the prompt review the code for security issues, but then complied when asked to fix this code, followed by some further manual steps. Now, the main thing I want to talk about in this video is I want to apprise you of a a change in the the tone of how people are talking about AI and and why you sensed from me in the last video a tone change, which again does not reflect a new set of beliefs for me. If you watch some of my very old videos before I ever started doing the funny stuff, when I was improvising just like this, I would often say positive things about Anthropic. I would say they believed what they said. I I compared Dario to Steve Jobs as being some sort of visionary. None of this is new. None of this is me changing. None of this is psychosis. Maybe it is psychosis, but if if it is psychosis, it's not new. I I I oscillate. It ebbs and flows. Look at I mean, look at the Primagen. Every time I go off the rails, somebody submits my videos to the Primagen subreddit, and I do not check the comments there. Those are the some of the worst comments I've ever seen, man. I don't know what's It's Reddit. It's rated. But nonetheless, the Primagen uh quoted this tweet by Stratechery, who knows how you pronounce that. Very good blog that has been active for for decades now as far as I can remember. And you know, it is not too much alpha in this, but a couple things worth worth sharing with you. Ben Thompson says, "Fable is in my limited experience a very impressive model. It's increasingly difficult to objectively evaluate models for anything other than coding performance, which is you know, what I was saying yesterday. But there is subjective feel, and I found my interactions with Fable to be extremely impressive. It made other models, including GPT 5.5 and Opus 4.8, feel small and dumb." Also, what I felt yesterday. And you know, the negative negative comments haven't bothered me.

[00:06:29]You know, it is what it is, but I'm curious how many of those people who were instantly enraged actually tried Fable. And and then I think you owe it to yourself, well, you can't anymore, but you have you owe it to yourself to try it and when it comes out again, you should try it. So, what Ben essentially says in this article is that he's starting to buy more and more that Anthropic, you know, when they do all these weird things that we read in the public as like dumb, naive, and cynical, he is starting to more and more believe that Anthropic just believes that they are the right people to to serve AI safely. He he believes that they believe what they are saying, and that's why it appears naive, and that's why it appears somewhat overly calculated sometimes and under calculated at other times. So, he says, "To that end, I can certainly buy the case that Fable Mythos is in fact more capable when it comes to identifying and exploiting security issues, and that Anthropic's cautious rollout was justified." So, yeah, people are starting to people are starting to buy, you know, even the Primagen quote tweeted this and he said this is a really good read. I think we're all starting to feel that if you used Fable, you've increased your sense of uh you given a little more points of credibility to Anthropic that maybe they weren't fully lying and then they really did see something. Now, if in the public we all felt something special compared to other models when using Fable, you can imagine that the parents of of Fable felt even a closer attachment and then Ben said something at the end of this article that I I think sums it up perfectly.

[00:08:03]I'll get to that. It's a very short article. So, the problem with public publicly releasing models, however, is that guardrails can be jailbroken. And apparently, that is exactly what happened shortly after the release. So, he talks about the data imperative, the the reason he believes, which makes sense, that Anthropic and OpenAI highly subsidized their subscription plan $200 the $200 subscription plan you are on gets you $8,000 worth of Claude tokens and $14,000 worth of Codex tokens is because of fighting for training data. So, the more people that use their models, the better they can make their models. And it's why he believes that Anthropic has instated a 30-day retention policy for enterprise usage for Claude Fable. Previously, Anthropic would not store enterprise usage data on their servers at all. And now they're saying, "For security purposes, we have to store for 30 days."

[00:08:57]Maybe they're slowly inching up to saying, "For security purposes, we have to also train on your data so that we uh make it harder to break these guardrails." That would be highly highly lucrative. So, Ben says that Anthropic's explanation for their dramatic change in their data retention policy was safety.

[00:09:13]Specifically, the company claims that retaining all user data for 30 days is necessary to prevent the jailbreaks the US government is worried about.

[00:09:21]I can certainly imagine a future where safety compels them to train on this data as well to to better protect against malicious usage. So, here's the key thesis that I think a lot of us are starting to inch our way towards being the best explanation for why Anthropic is so damn weird. Here's the thing about these safety justifications.

[00:09:40]I think they work because to Anthropic, they aren't justifications.

[00:09:45]The company really believes that they are the only ones who believe in superintelligence and thus are the only ones who are sufficiently concerned about the dangers. That excuses decision after decision, policy after policy, and confrontation after confrontation that to the people on the outside look like a bizarre combination of cynicism and naivete. So, here's the the best paragraph, the best sentence in this post. So, he said, you know, he says OpenAI has no direction. They're working on consumer that they've drowned in. No and and they've leaked talent to Anthropic. And Anthropic on the other hand has perfect alignment between talent and mission and business. The company gets to sell to researchers the creation of a machine god with the mantle of being the sort of person who cares about the dangers and is smart enough to navigate them on behalf of humanity.

[00:10:34]That every policy change that falls out of that happens to be great for business is the most beautiful coincidence in the world. Now, whether this is true or not, whether it's a coincidence, I don't know, but it's a good fit around the data, right? It explains things pretty well that I mean, if we look at what Anthropic is doing as some sort of chess, 5D doesn't begin to cover it. I mean, that's like 30D chess what they're playing. And Marc Andreessen is saying 5D chess doesn't exist. And I sort of happen to agree, let alone for 30D chess to exist. I mean, Dario's a smart man, but can he be this smart? And the hard part about predicting the future is not intelligence, it's just the the insane number of variables that you would have to crunch. It's just impossible. And so look, you either believe Dario's playing, you know, 30D chess or you either believe that that his goofy beliefs happen to be great for business is one of the most beautiful coincidence in the world. I would lean more toward it being a coincidence. A very beautiful coincidence, man. Now, what are you going to do? It's uh Dari is a super nerd, man. A lot of people have been recommended. All of a sudden, there was this video on YouTube Dari interviewing from 2000 Dari interview from 2018 uh when he was a nobody interviewing with with with Microsoft someone at Microsoft. It was it was a pretty boring video, I'm not going to lie. But, he he's just a super nerd, man. I mean, this guy is like got a PhD from what Stanford and Princeton and biophysics. I don't even know what biophysics is, man.

[00:12:04]Like, what is biophysics? Seriously. I mean, I'm you know, inter- interdisciplinary science that applies the theories, principles, and methods of physics to study how biological systems work. Okay. I mean, I I figured as much, but you know, what what does it really mean? Okay, anyway. Um yeah, Dari is a super nerd, man. Now, here's the point that Ben ends on, and it is very, very serious. He says, "What I fear, however, is that it is one thing to have people convinced they know best building a smartphone that I can take or leave."

[00:12:33]So, he's talking about Apple comparing uh Apple or Anthropic to Apple, but you know, with Apple building a smartphone, you can sort of take it or leave it, and you'll be sort of fine.

[00:12:42]But, with Anthropic, it's considerably more concerning to have them building superintelligence that has the potential to rival or exceed the power of nation-states or merely massive corporations. The history of brilliant people convinced they know what humanity needs is a sorted one, precisely because they have convinced themselves that their intentions are good, justifying actions that are very much are not. So, uh of course, we know that yeah, the history of people convinced uh that they are doing good for the betterment of humanity leads to brutal, brutal stuff in history.

[00:13:17]But, I need to open that can of worms right now. Yeah, overall, there's a a subtle vibe shift in the tone right now in the world of AI. It's getting a little more serious, a less goofy, um explaining in a members-only video that the reason I was able to make so many funny videos, a string of funny videos one after the other, was because of, you know, primarily two things. One is the GPT-4 era where the 2025 era models that were highly like goofy, hallucinogenic. Um there were a lot of stories coming out.

[00:13:47]You know, a good source of AI psychosis stories was was Futurism. I I found a couple stories here that I was able to to to talk about on the channel.

[00:13:57]And they have a whole section for AI psychosis. And honestly, the last time I checked, these are all pretty old. I mean, May 31st was the last one, not too old, but you you find these April 23rd.

[00:14:07]These stories have run out. April 14th.

[00:14:10]It's a different sort of world where we're sort of entering into um the GPT-4 era models were very goofy, made mistakes, led a lot of people into psychosis, the alignment was off. It would highly reinforce your worst attributes. New models are and that's why a lot of people love 4.0, GPT-4.0.

[00:14:27]Uh new models are a little tamer, so the AI psychosis stories have lessened. And these were really you know, sorry to say, but there was a lot of comedy in in these stories. The funniest thing about my previous comedy-only videos were that it's essentially people stepping on a rake and hitting themselves in the face.

[00:14:43]It's funny, it's like slapstick. Then it evolved to companies stepping on a rake.

[00:14:48]And that was funny, it's like companies overspending on AI, the dude that spent $1,200 to build a demo. That was really funny. That's going to ramp up, I believe. More and more and companies are sobering up, spending less on AI, um and stating really strict budgets, but it's it's just not that funny after the first one. So, well, I bookmarked this.

[00:15:05]Someone picked up a good quote from my video, which if you idiots will which if you lovely people watched to the end, you would have found it was a very balanced video. I just called the model beautiful because it was beautiful relative to other models, but um the reason so I said in that video from yesterday, the reason everyone talks about all the crazy demos they've been able to cook up, but no one's shipping anything is because of issues like this. LLMs can get you 80% there, something we talked about a lot, but the other 20% man, if you're not an expert, you're going to have a hard time.

[00:15:39]I mean, dude, is that not what I've been saying? So, now if you want a much more sobering engineering perspective, Mitchell Hashimoto, uh, working on Ghosty, he's encountering some good technical problems. He says, "The problems with the it works, who cares what the code looks like mindset for agentic work is that it assumes the agent has a perfect understanding of works.

[00:15:59]Realistically, things are under specified, agents make bad assumptions, etc. To be fair, agents are pretty good at unit test coverage, they're pretty bad at designing human experiences like API, CLI, flags, etc., especially cohesive ones for future roadmap plans they may not have visibility into unless your backlog is perfect and vision fully laid out, which I doubt. True.

[00:16:20]They're bad at knowing where performance matters and what type, CPU versus memory tradeoffs. They're bad at where compatibility matters and where it doesn't. Unless you have all this specified, you can't possibly claim it works without taking a look and thinking about it. Now, this is a really good take.

[00:16:36]That, you know, the thing about evaluating AI for engineering is it takes you can't do it with one shot. And I gave so many disclaimers in the video yesterday that I was making demos, and demos are highly AI is known to build highly impressive demos. I I was I made that abundantly clear in the video. No one watched the video, they just left comments. Um, once you get to real engineering, which I am not doing right now. I'm not doing real engineering cuz I don't have a product to work on, unfortunately. I'm looking for one. I'm looking for something to build so that I can have actual deep experience and see these kind of things, but luckily we'll have these people who are who still have something to actually work on that, you know, there's deep stuff that you can only really figure out once you go super deep with these agents and it's an LLM thing, right?

[00:17:20]It's just an LLM quirk. It's it's a different type of intelligence. It's it's it's pseudo intel not pseudo.

[00:17:27]God.

[00:17:28]Here here I am trying to get into the nature of intelligence. Nope, not doing that. Look, it's it's its own thing, right? Let's not play word games here, but there's a fundamental nature of LLMs that's that hasn't been cracked, that hasn't really advanced and no one's really no one's figured out how to really replace humans fully with AI. Like I understand it's a war and but you know, the the war has subsided a little bit.

[00:17:49]Like I don't know if you guys know this, if you're still very tense about AI, everyone everyone sort of is now. It's like you don't have to so actively be outspoken against AI because everyone has sobered up. The hype people, the pro-AI boosters, they've calmed down a little bit. Everyone is I mean there are exceptions, but you know, even the companies are like more realistic now.

[00:18:09]The hype is slowly becoming more nuanced. You know, if you mourn the loss of my roasting videos because you felt that there was a counter there was a need for a counteracting force. I think more reasoned videos, more nuanced videos can serve that purpose much more.

[00:18:24]So I think we'll end it here. Thank you for spending your morning with me. I'll try to get this out before your coffee.

[00:18:29]If there are any questions that you would like for me to talk about in tomorrow's video, if I'm alive to make it happen, share in the comments and maybe I can sort of discuss them in the next video.

[00:18:41]Thank you for watching.