Fable 5.1 Features and Astra Release  🔗 **Source:** [Fable 5.1 Features and Astra Release](https://share.snipd.com/episode/70a52686-66dc-4fc5-925f-44281a5969cf) ## Highlights - **Anthropic Shifts AI Toward Scientific Discovery** - Anthropic is positioning Fable 5.1 and Mythos 5.1 as tools for scientific discovery, not merely general-purpose chat. - Demonstrations include analyzing a Venusian shield volcano and modeling hearts and arteries. Transcript: Jaeden Schafer The first thing I wanna talk about today is the new Fable and Mythos 5 story, what they've kind of put out in regards to this. There's a blog post they dropped and I'm liking their blog posts. They're very aesthetic these days. You can like go in and switch the color palette on the banner image on their blog, which is completely unnecessary, but it's very cool. Their blog is titled Claude Fable 5.1 and Mythos 5.1. It's kind of interesting. They're syncing up the naming conventions here with Fable 5.1 and mythos 5.1 i feel like that's something that open ai really just let just went wild with in the last couple years all Of their different numbering schemes and it feels like they're getting a little bit more stable now but the naming conventions have been something that's been pretty wild one thing In particular that i'm excited about with this if you go into some of the scientific research they have a whole bunch of very cool demos of ways that they're helping with scientific research. They're showing off different things that they're doing. You know, they have like a small shield volcano in on Venus that they've been helping researchers look at. They have a lot of work around hearts and arteries. And it's just like the capabilities of this obviously are getting a lot better. They're going to go and focus on how it can help science instead of how it can help hacking. ([Time 0:02:16](https://share.snipd.com/snip/eda87be1-4f80-49b8-9ec5-66599277368d)) - **Mythos 5.1 Remains Restricted Despite Breakthroughs** - Fable 5.1 is broadly available through Anthropic's API and cloud platforms, while Mythos 5.1 remains restricted to cybersecurity and life-sciences partners. - Mythos reportedly set records on Terminal Bench 4.0 and Humanity's Last Exam while producing a high-resolution Venus map. Transcript: Jaeden Schafer It is fable 5.1 is available today on the any cloud platforms and on the anthropic api i really appreciate when it comes out on the api at the same time because then you know you're going To get it on all of different platforms mythos 5.1 is still restricted to cyber security and life sciences partners so certain companies are going to get access to that. But because it's, you know, so capable for hacking and all that other stuff, it's still not available to the general public. Enterprise Frontier Safeguards, which is Anthropics zero day retention tier for Fable, is going to roll out to users in is rolling out earlier or rolled out. And Mythos 5.1 is setting a bunch of different records on Terminal Bench 4.0 and Humanity's Last Exam. And it also produced a custom GPU optimization and high resolution Venus map pre-release that I was kind of talking about earlier. Right now, the system card rates Mythos 5.1, it's very low risk for automated AI R&D acceleration. This is kind of in line with some current trends. Overall, I'm really excited on their announcement. They said Anthropic has never trained on enterprise data without explicit permission. And we never will. This is something that they've been kind of really putting out there there. You know, a lot of people are concerned that this is going to be getting their company's data and whatnot. But overall, really excited. This is a big model. And there's a lot of big updates, you can go check out their release, their release and their blog post to see kind of all of the different things. ([Time 0:03:24](https://share.snipd.com/snip/203f20fe-2624-4e5e-94ff-89fe694eb7f8)) - **Astra Finds Zero-Day Vulnerabilities Autonomously** - OpenAI's Astra reportedly achieved a near-perfect ExploitBench score and found two previously unknown vulnerabilities without human guidance. - Its zero-day performance pushed Astra into OpenAI's highest internal cybersecurity risk tier. Transcript: Jaeden Schafer Release their astra model and in in anticipation of this i mean this is basically their response to what uh anthropic has come out with their new with their mythos 5.1 or you know five Which was kind of the big thing that got in the news but this is really something that you know they're saying look this is just as capable at security research and i mean technically hacking But they don't like to put it that way. It had basically a perfect score on exploit bench, which is benchmark measuring how well these AI models are hacking known system vulnerabilities. And what I'll say is more important than this is like, you could say, okay, we know there's this vulnerability, we're going to give it we're going to see if it can hack it out. Okay, got a perfect score. But there's a lot of security vulnerabilities that are day zero, meaning, you know, no one knows about them. And this model has actually gone into that as well. In a modified version of this test, which was built by OpenAI engineers, Astra actually discovered and exploited two zero day vulnerabilities without any human guidance, meaning They said, you know, here's the code base or here's the tool, go and hack it. And it found stuff that no human had been able to find before. So it's not like, oh, we these vulnerabilities and it gets it. It's going to finding brand new stuff. OpenAI says that Astra is the first LLM to cross their quote, critical cybersecurity threshold. So this is basically their highest internal risk tier. ([Time 0:05:02](https://share.snipd.com/snip/b63e4764-d414-4160-b596-99e0c24528f5)) - **Gate Access To Frontier Cybersecurity Models** - Treat frontier cyber-capable models as high-risk systems and restrict access according to likely misuse. - OpenAI plans chain-of-thought monitoring and account-level risk scoring, while Astra's advanced capabilities remain limited at launch. Transcript: Jaeden Schafer And they're like, look, this model is good enough that it is at the absolute front front of the line for like this is very high risk and so because of that they're putting a bunch of guardrails On it they're you know when they're trained if it's able to hack something they go back in and like try to go fix it and add guardrails so access to astra's most advanced cyber security Capabilities are going to be restricted at launch this is basically what anthropic is doing with their mythos and kind of what they did earlier this year opening eye is deploying astra With additional chain of thought monitoring and account level risk scoring to catch misuse that's interesting to me right account level risk scoring it's like how high risk is your Company or is your account and depending on that is like that kind of like determines what we're going to do with the model or what we'll give you access to. And so, yeah, I mean, this is pretty wild. If you go over and look at their blog post on all of this, you can see just how much better they did on the exploit bench, which, by the way, if you're on Apple, you can now watch. This is also video is now published on Apple and Spotify. So you can see the video of this if you're interested in the screen shares of all the stuff I'm talking about. But, yeah, if you go and look at it, if you're if they give it, you know, 80,000 output tokens, the Astra model got like a 39 or 40% completion rate on exploit benchmark. Whereas when you're using something like the GPT 5.6 soul, which was their previous best model, it was doing something not quite as not quite as good, which was closer to they had a closer To 11.5% of completion. So we're really bumping up. I mean, that's basically 4X on this internal port on exploit benchmark. ([Time 0:06:18](https://share.snipd.com/snip/82897a55-ba9a-4a84-a182-2314a3f37740)) - **Hollywood Characters Become AI Licensing Assets** - Google is reportedly offering major studios up to $40 million per popular character to license copyrighted properties for Gemini. - The potential deals treat characters like valuable datasets and could include studios sharing YouTube AI advertising revenue. Transcript: Jaeden Schafer Google is pitching Hollywood Studios $40 million per character when it comes to certain AI licensing deals. And this isn't going to be like every single one. It's not like every character is going to be paid $40 million. But I think for the most popular ones, this is something that they're going to do. Disney, Warner Brothers, Discovery, and Universal all have hundreds of millions of dollars that they're being offered to train Gemini on their libraries. And they the per character price tag. It's so interesting to me because it feels a lot like how you would, I mean, go and buy data, any other data set, right? Like you, you would go buy data sets based off of the value based off of how rare it is based off of the quantity inside of the data set. And so now that's like, we're putting these intrinsic values on the characters of, you know, Disney or something like that. And this is going to be for, of course, copyrighted characters. The total deal value across multiple characters could run into the billions with studios potentially getting a cut of YouTube AI ad revenue, right? So if they are like, hey, look, we're going to, I'm assuming they're going to roll this out as YouTube AI or YouTube's able to generate AI videos. And if it's making ad revenue from a lot of people generating these AI, you know, generated videos, and it's got Mickey Mouse on there, Donald Duck, or it's, you know, they've got Peter Pan or whatever their trademarked characters are, the studios are going to get a percentage of the revenue. Google's DeepMind already made a $75 million investment deal with A24, which was earlier this summer and disney warner brothers discovery and universal none of them are by the way Saying anything nobody has signed anything as of the beginning of this month so this is all in talks these offers are in talks i'd be curious to see if they come back wanting more or if you Know what kind of happens there lions gates 2024 licensing deal with runway has yet to produce any sort of released product. This was, you know, a couple years ago, they had this licensing deal, and we haven't seen much come out of it. But we know that, you know, there's there's a lot of money in this. Over on the LA Times, there was an article kind of talking about the story and Tom Noonan, who's a former studio and network executive talking about this story said, Google's enlisting Some of the smartest people in Hollywood to try to unpack what's inevitable, which is Hollywood and AI are going to become reliant on each other and how best we can understand that without Just terrifying people. So he views this as something that is, you know, it's going to happen. ([Time 0:07:58](https://share.snipd.com/snip/05997570-9f1b-43f9-b21f-64a3b5b772e6)) - **Grok Enters Government Through Starshield** - The Pentagon's genai.mil portal has onboarded 1.7 million users from a workforce of roughly three million. - Grok for Government runs under SpaceX's Starshield AI, pairing the model with secure satellite infrastructure for military use. Transcript: Jaeden Schafer Google is launching ChatGPT Mill, or short for military, and Grok for government on genai.mil, which if you don't know, this is kind of a special site only for the military. If you go there, you know, if you're curious and you go to genai.mil, it's going to hit you with a, you have reached this page because you're not authorized to visit genai.mil from outside Of the DOW networks. If you believe this is an error please contact your company's it's you know your commands it service desk and basically yeah you can you're not gonna be able to get into this if you're A civilian this is only for the military and this is something that the department of defense is rolling out they launched chai gpt mill and grok for government which is acts which is Accessible to about 3 million military and civilian personnel. Of course, even civilians, right? These are people that have to be inside of the department. Genai.mil has already onboarded 1.7 million users, which is more than half of the Department of Defense's 3 million person workforce. Anthropics Claude is not in that portal. We know that Trump and the whole administration kind of labeled them as supply chain risk and they designated Anthropic as such. And then Anthropic kind of had the whole court battle. And I think it got overturned at some point. And I'm not sure kind of where we are back and forth on this case, but it's, you know, missing right now from that list. Grok for government is shipping under SpaceX's Starshield AI, which is kind of a secure satellite network, which is built on Starlink infrastructure. This is kind of interesting. This is Elon's way of getting into supplying the government. And honestly, I mean, I think it's a very smart move, kind of pairing not just the AI model of Grok, but also the Starlink infrastructure, right? Because if you're in some remote area that doesn't have cell reception, or that's really hard to get, or it doesn't have internet service, accessing chat chippity might be difficult, But you could get Grok, you know, theoretically through Starlink. And I'm sure the military's got all of this, their own satellite, you know, connectivity stuff. So it's probably not as big of a deal as it would be for your average civilian. ([Time 0:10:25](https://share.snipd.com/snip/a55510a9-4693-4449-b561-2fade528c34b)) - **Satellite AI Could Improve Crop Forecasts** - The USDA is testing satellite imagery and machine learning to improve crop-yield forecasts that influence corn, soybean, and wheat prices. - Better estimates could reduce market distortions affecting farmers, although hedge funds already use similar satellite-plus-AI models. Transcript: Jaeden Schafer The USDA is going to test satellites and AI to help with crop estimates. There was a bunch of backlash from farmers. So now the agency is piloting this kind of new remote sensing and machine learning. And their idea is to get better at their yield forecastings. Farmers say that the markets have moved against them a lot because they say that the yield forecasting is not accurate. And so it makes it, you know, the price of commodities or the price of food changes and it's bad for the farmers, they're saying. So the government's like, okay, well, maybe we can come in and make this more accurate. The agency's monthly, it's called the WASD, and it's a report, and it basically shows that most market-moving data is kind of going through this, and they release this for global agriculture, But it basically drives corn, soybean, and wheat pricing. So when they say this is what the yield is going to be, there's going to be this huge oversupply, all of a sudden the price might drop, or if they're like, hey, there's a huge shortage, then The price for corn might go up. So anyways, if there's any sort of government agency that's affecting the price of something like corn or soybeans, obviously, the farmers that are growing that if they feel like it's Negatively impacting them are going to complain about it. So at the end of the day, the best thing I can do is make it as accurate as possible. Using satellites and using AI is probably the best way to do this. This is also mirroring what the private sector does. They use satellite plus AI yield modeling. So it's kind of funny that USDA has their own forecast. A lot of hedge funds have their own, you know, in-house satellite or drone fleets that are going and measuring all this stuff. They're bringing it back to AI to model it correctly. And they're for arbitrage opportunities where maybe the government overestimated the supply or underestimated the supply. And they'll take the over under on those different bets. They'll make money off of the arbitrage opportunities of basically what the government got wrong on that and what the markets were forecasting ahead of time. So anyways, hedge funds have been doing this for a long time. ([Time 0:12:38](https://share.snipd.com/snip/ff8fb388-d7f4-4a26-970f-08ae4ba428b6))