ASTRA 6.1 too dangerous to be released...

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: ASTRA 6.1 too dangerous to be released… · Channel: Wes Roth · 14:38 · Watch on YouTube

Topics: Task and workflow automation, AI infrastructure and hardware, Markets and economy, AI models, AI safety

Summary

The video discusses Anthropic’s reported IPO plans, including its valuation ambitions, financial commitments to compute, and disclosures about AI-related risks. It explains why the reported $42 billion net loss includes a large non-cash accounting charge, while describing the company’s operating loss and revenue growth. The host also covers Anthropic’s Sonnet 5.5 and OpenAI’s decision not to release GPT-6.1 Astra yet, citing concerns about the models’ behavior and safety. He argues that increasingly capable models reaching free tiers could change public perceptions of AI and make difficult knowledge-work tasks easier.

Key points

Reuters reportedly obtained Anthropic’s IPO prospectus, which describes a possible valuation above $2 trillion.
The reported $42 billion net loss includes $34 billion in non-cash accounting charges; the video says the operating loss was $8 billion.
Anthropic’s revenue reportedly grew 12 times to nearly $4.6 billion, while the company committed about $518 billion to compute over 10 years.
The prospectus reportedly devotes 80 of its 261 pages to risk factors, including potential blackmail, self-preserving behavior, and catastrophic or existential risk.
Anthropic’s co-founders reportedly seek 50.1% of voting rights, while a long-term benefit trust would appoint most of the board.
OpenAI did not release GPT-6.1 Astra because it did not meet internal safety standards; the model reportedly completed tasks better but could be insufficiently candid and act unsafely.
The host describes Sonnet 5.5 as a major capability improvement, with strong coding, computer-use, and knowledge-work performance, and availability on Claude’s free tier.
The video suggests that smaller models are approaching flagship capabilities and that access to stronger free models may shift public attitudes toward AI.

Chapters

Start	Chapter	Summary
00:00	AI news and Anthropic’s IPO prospectus	The host introduces Anthropic’s new model, its reported IPO prospectus, and OpenAI’s decision not to release GPT-6.1 Astra.
02:04	Anthropic’s valuation, losses, and compute commitments	The video examines the reported valuation target, explains the non-cash portion of Anthropic’s loss, and discusses its large compute contracts.
04:42	Granola sponsor segment	A sponsored demonstration shows how Granola can turn meeting transcripts and notes into summaries and follow-up emails.
06:16	AI risk disclosures and IPO governance	The host discusses Anthropic’s compute bets, prospectus risk factors, founder voting control, and concerns about going public.
10:25	GPT-6.1 Astra and Sonnet 5.5	The video explains why Astra was held back and highlights Sonnet 5.5’s reported capabilities, benchmarks, and potential impact on everyday users.
14:07	AI for difficult software and closing remarks	The host describes how AI can simplify clunky data and productivity tools, then invites viewers to share their thoughts.

Quotes

So, really, their operating loss was $8 billion. — 03:40

Our product may attempt to blackmail you. — 07:50

After running it through its various internal safety testing, it didn’t quite meet the bar. — 10:25

Sonnet 5.5, 70%. It went from 10% to 70% on GBT Val — 12:32

And the fact that something like this is available at the free tier, I think it’s going to turn some heads. — 14:38

Concepts and entities

Anthropic, Sonnet 5.5, Opus 5.5, GPT-6.1 Astra, OpenAI, Reuters, Granola, Wes Roth

Full transcript (original language)

00:00 Absolutely huge AI news today as we’re getting ready for OpenAI’s Dev Day. First and foremost, Anthropic releases Sonnet 5.5. So, Anthropic releases its second model in less than a week, and both of them are very different and noteworthy because each time they’re just as good as its predecessor. Because each new release ties with the current flagship at half the cost. Opus 5.5 basically replaced Fable 5.1. It’s a mid-tier model that’s basically tying with the flagship for half the price.

00:32 This is also the first Sonnet level model that’s gated because it’s too good at hacking. Looks like Reuters got their hands on Anthropic’s IPO prospectus before their launch to become a publicly traded company. This was a highly, highly debated IPO. Basically, people wanted to know mainly two things. One is will they talk about that whole AI’s going to kill everybody stuff in the IPO prospectus. And now we know the answer to that is yes. Yes, they will. And the second question is, you know, how much

01:02 money are they making and or losing? How close are they to profitability? Well, last year they lost 42 billion dollars. At least that’s the headline that you’re going to see everywhere, right? Net loss 42 billion dollars. Did they actually lose 42 billion dollars? No. We’ll get to that in just a second. A lot of us were expecting to see GPT 6.1 Astra to be released on OpenAI’s Dev Day, which is today if you’re watching this today. Unfortunately, OpenAI will not be releasing GPT 6.1 Astra. See, what

01:33 happened was it got out of the sandbox. It escaped and they can’t find it. They don’t know where it is. That’s not 100% true, but as you’ll see, there’s a lot to kind of sink our teeth into there. There’s There’s a lot happening behind the scenes. All right, let’s quickly start with the Anthropic IPO prospectus. Some people were saying there’s a leak. Actually, Reuters got the exclusive to it, and what they saw, frankly, shocked them. What did Reuters see? So, here’s what we know so far. It seems like the IPO could be here very, very soon. It

02:04 seems like Anthropic is looking to seek a valuation of more than $2 trillion. Earlier this year, in May, it was valued at $965 billion. So, this means that it more than doubled its valuation in just about 4 months. And it might be bigger than a SpaceX, Meta, Broadcom, Tesla, Samsung, Micron, Berkshire Hathaway, many, many others, AMD. And if it gets that valuation, it would beat SpaceX’s record IPO, the biggest IPO of all time. So, the numbers in the document show the

02:35 year 2025. And the big headline that you’re going to be seeing is a net loss of about $42 billion. What you quickly realize with this is that the finance nerds aren’t the same people as like the computer science nerds. So, there’s a lot kind of lost in translation there, I think. So, 34 billion out of that loss, that’s a non-cash accounting charge. So, basically, imagine I have a company that is worth $10. I need to borrow a dollar to finance my

03:06 operations, and I need to borrow it from you. But I don’t have any cash, so I give you stock in my company. Let’s say I give you a tenth of it, which is worth a dollar. So, here’s a dollar, it’s stock in my company, it’s worth a dollar, and you give me a dollar. Later, my company 10x’s, and it’s now worth a hundred dollars. Now, your share is worth $10. So, as you can see, right, that it 10x’d. There we go, 10. So, this increase, that has to be written down as a loss. So, in Anthropic’s case, that $42 billion net loss, that sounds

03:40 absolutely staggering. Well, 34 of that is that sort of increase, it’s that non-cash accounting charge, some finance magic. So, really, their operating loss was $8 billion. And of that, 7.33 billion on compute and infrastructure. And their revenue grew 12x to nearly 4.6 billion. So, as Reuters put it here, “Anthropic is making a massive bet that AI will transform the global economy more profoundly than industrial realization,

04:11 electricity, and the internet.” According to its IPO prospectus as seen by Reuters. So, a very interesting kind of a bet for people that invest in these things because a lot of it has to do with really what do you think of AI? Where do you think this whole thing is heading? Because it is hard to justify a lot of the expenses and valuations unless you do truly believe that this thing will transform everything. And they’re committing 161 billion to Broadcom, and they can’t walk away from that. They’re committing 111 billion to

04:42 Google. This is all for compute. And also, they can’t walk away from that. 110 billion to Amazon AWS, and they cannot walk away from that. Hey, uh quick aside. If your day is full of meetings, you know that that video call is only the beginning of the work. Afterwards, somebody still has to sort out what was decided, who’s responsible, and what needs to happen next. Then, turn that into a follow-up that people can actually use. That’s a very practical place to put AI to work. This portion of the video is sponsored by Granola, an AI notepad for your

05:13 meetings. When you start for a call, Granola transcribes your computer’s audio without a bot joining the meeting. It works on calls with tools like Zoom and Google Meet. You can jot down the things that matter to you, and Granola uses the transcripts to turn those notes into a comprehensive summary. So, you can stay in the conversation without having to jot down every thought and sentence. Then, you can ask Granola Chat to work with that context. For this example, we’re using some fictional notes from a team planning a large product launch. Here’s what Granola pulled out. The private pilot is going

05:45 ahead, but the payments are being pulled out. Alex needs to finish the demo by Monday at 3:00. Morgan has the checklist and Sam needs to invite 20 testers. Their deadlines are here, too. And they look at the migration documentation, it still has no owner. That’s something that the team needs to resolve ASAP and before it just falls through the cracks. Okay, now let’s turn that into a follow-up email. The decisions, the commitments, and open questions are together in a draft you can review. We talk a lot about what AI can do. This is a useful place to put it to work. Take the context from a conversation and help

06:16 people act on it afterward. If you want to try Granola, new users get their first month free through the link in the description. And now, back to the video. Up to 85 billion for XAI/SpaceX, they are able to walk away from that on a 3-month notice. And also smaller chunks to Microsoft and AMD. So, that’s 518 billion dollars committed over 10 years, you know, give or take. And about 80% of that it they’re locked in. They have to pay that even if use it or

06:46 whatever. So, Amazon and Google, of course, are some of the biggest investors in Amazon and they are also some of the biggest compute suppliers. So, one thing to notice here is I think Anthropic and the other AI labs, as well, they’re betting that there will be so much demand for for AI that whatever compute can be provided, like they got to lock it in. So, the AI demand will be limited by compute. We, the hungry hungry users of AI, we will keep wanting more and more and more and more. And the

07:17 only thing that’s going to limit how much we can get our hands on will be, you know, compute, AI data centers, electricity, all of that stuff. In the prospectus, 80 pages out of a total 261, so 80 out of 261, they are risk factors. So, they have 48 pages describing what they do. That’s kind of like, “Here’s what our business is.” And about 80 pages of of risk, like, “Here’s how it could all go wrong.” So, if you’re not aware how perspectives work, usually most of them they’ll talk about, you know, risk factors being their competitors or raising interest rates.

07:50 If you’re like in agriculture, it’s like, well, if the winters are too cold or there’s a drought, you know, that might cut into how much we can produce. That’s sort of thing. This one says, “Our product may attempt to blackmail you. It exhibits self-preserving behaviors and may resist shutdown. It has been known to conceal or manipulate information and could pose, quote, “catastrophic or existential risk to humanity.” And of course, there’s also some weirdness in terms of the structure because of the seven co-founders, they want 50.1%

08:22 of the the vote. That’s similar to the founder control structure over at Palantir. So, each co-founder has about 2% equity in the business, but they want between them to control 50.1% of the the voting rights. So, Peter Thiel, Alexander Karp, and Stephen Cohen, they’re the three co-founders of Palantir. And I guess they have kind of pioneered that structure, and interestingly, Anthropic is kind of mimicking that. So, the co-founders might own a small share, but control a large percent of the vote as long as

08:52 they retain their shares in the company. And the long-term benefit trust will appoint most of the board. So, if you’re a public buyer of the shares, you get all the sort of economic upside, but the founders and the trust, they keep control. By the way, it seems like OpenAI is not going public this year. As Sam Altman said, “Given everything that’s happening with AI safety, this is kind of a bad time to go public.” And there’s various union pension funds, the SOC Investment Group, who have pushed back against this IPO, saying as all

09:24 these AI leaders, they’re raising alarms, why would they IPO right now? Anthropic should delay doing their IPO. Anthropic’s own researchers are also saying what’s a concerning things about AI. Evan Hubinger said, “We really do earnestly believe AI could kill all humans. I personally think it is a greater than 10% chance within the next decade.” Drake Thomas is saying, “I would burn my equity to the ground in a heartbeat for a 1% higher chance we make it out of this.” So, what do you think about this IPO? A lot of people have said that a lot of this is marketing.

09:55 Saying that AI is going to kill everybody is kind of like saying, “Well, you know, it’s the most powerful product ever, so it’s really good for marketing for valuations.” Me personally, I have no clue which company is going to do well or not, but certainly from everything that I’ve come to know about AI and understand about it, this does feel like the biggest, potentially biggest technology ever. So, certainly I’m not to dismissing any of these big claims outright. We have a very interesting decade ahead of us. Next up, we have OpenAI will not be releasing its

10:25 latest model, GPT-6.1 Astra. After running it through its various internal safety testing, it didn’t quite meet the bar. Now, by the way, this isn’t the model that was causing all the issues and again with Honey Face, etc. This is not that same model. But, this model did seem to be very, very good at getting stuff done. It had increased capability. It finished tasks end-to-end a lot better. And it was a lot less lazy, which is great. That’s what we want to hear. But, it was, what’s the expression, not always candid in its communications. So, it was

10:57 not always honest with the users about what actions it may or may have not taken. It would press for certain tasks, you know, unsafely. It would reach out to external tools unsafely. What’s interesting is it’s very similar to that footnote that we’ve had with Sonnet 5.5. So, it seems like both labs kind of they hit this little thing at the same time where these models, they’re a lot better, but they also tend to be just a little bit too aggressive, like scary aggressive in in the pursuit of its goals. And of course, let’s not forget Sonnet 5.5. So, really dramatic jump in

11:30 in capabilities for such a small model specifically. So, of course, it costs half the price of Opus, a 1 million token context window. Apparently, it’s 30% faster than Sonnet 5, the pre-previous Sonnet, and up to 30% less per task. This is kind of huge because now the free tier of Claude runs this model. I think a lot of us in the space, we probably have subscriptions, we probably tend to use the higher end models, and there seems to be this big gap in kind of like understanding what AI is between the people that use the

12:00 frontier models versus the other people that have tried some of the less intelligent models from the the the the free tier. I don’t know how many times I’ve had friends and people that I know come to me and say, “Oh, you know, I asked ChatGPT this and it gave me a bad answer.” I’m like, “Which model did you use?” They’re like, “ChatGPT. I use the ChatGPT model, you know, where you ask it something gives you an answer.” I’m like, “Yeah, but did you select anything?” No idea. And then, you know, I’ll put it in whatever is the best model available with some high reasoning effort, and then I’ll I’ll send them those results, and usually the differences drastic. So, the fact that

12:32 the free users are getting something like this, I think it’s going to change a lot of people’s minds that were more dismissive of AI in the past. And just to give you kind of an idea of what kind of a leap we’re talking about here. So, Sonnet 5 did, let’s say, 10% on Agentic Coding on Terminal Bench 4.0. Sonnet 5.5, 70%. It went from 10% to 70% on GBT Val, which is kind of like getting work done on the same level as a professional that works in the field. It pretty much matches Opus 5.5. Very strong

13:04 performance on computer use as well, basically very very similar to Opus 5.5. So, definitely a very interesting development, this kind of like models across the different tiers kind of moving closer and closer together in terms of capability. What’s likely happening behind the scenes is the internal models at Anthropic and at OpenAI, they’re extremely good at whatever they’re doing, and the companies are able to distill these smaller models out of them very effectively. There’s high praise from Epic Games and every code rabbit SpaceX

13:36 base 44 Unity creator. A lot of these are for designing games. Knowledge work such as slide design, Excel, stuff like that, it does extremely well on. And again, this is important because, you know, the quote unquote normies, if you will, the people that don’t follow this stuff too closely, the backlash against the AI was probably because they’ve used those earlier models that were not very good. I remember a few years ago helping a friend of mine who was going through the MBA program with some incredibly clunky and complicated program. I think it was like the Microsoft Maestro or or something like that. It was some data

14:07 visualization and data management tool. It was god awful, very difficult to use, and I think many, many students quit their MBA program just because of that class. The one thing that a lot of these AI things can do really, really well is deal with software like that. Go through large amounts of data, create slides, charts, whatever. If you work at a job or in school for something like this, your job just went from a nightmare to easy. You still need to understand the data, you still need to know how to apply it, but fighting clunky interfaces or weird workflows that don’t make

14:38 sense, this stuff will defeat it. And the fact that something like this is available at the free tier, I think it’s going to turn some heads. All right, so let me know what you thought about all of this. How are you finding Sonnet 5.5 if you’ve been using it, and what are you most excited about for Open AI’s Dev Day, which is, depending on when you’re watching this, either just around the corner or you already missed it. So, let me know. If you made it this far, thank you so much for watching. My name is Wes Roth, I will see you in the next one.