AIが脆弱性を次々発見:Claude Duet登場、DeepSeekがDeepInferを公開、RecursiveのFlintほか

wow hello and welcome everyone to thursday i for october 22nd cold open today is scary the internet feels like it's failing and a lot of it is potentially because of AI. So we're going to talk about this. Obviously, we're not selling doom and gloom here, but we do have Josh Long, aka the Josh Meister, joining us at the end of the show to talk about, to break down some of the security incidents, many of them powered by AI that just happened this week, many of these months. If you heard anything about the PHP core vulnerabilities, if you heard about the GitLab RCEs, and this one not related to AI, which is the eMudra certificate.

We're going to break it down to you and also give you practical things, what you can do today, right now, to protect yourself with your agents as well. But also we're going to talk about Anthropic stepping in to the personal agent era with Cloud Duet, which I think to me is one of the highest highs. I just got access this morning. I'm so excited. haven't even tried this. We're going to talk about a lot of stuff. I'm probably becoming a PBC, public benefit company, giving away 5% of its value. A lot more. Joining me directly from GTC Berlin is Wolfram Ravenwolf from our booth at GTC Berlin. At Corwee, Wolfram, how are you? Hi, Alex. I hope the audio is fine. I'm in the middle of the GTC. At our booth right now. I see still stuff going on over here. You're joining from GTC with Jensen Huang standing by you ready to do an interview, yes? Here he is?

Oh, that's another colleague of mine. That is kind of a Jensen Huang. Maybe later. All right. Still waiting for him to come along. We will definitely chat about some of the stuff that CoreWeave launched in our new segment called From the Forge. Oops. Wolfram, what are you going to show us? Some of this stuff, yeah? Oh, yeah. This is the new one, the NVIDIA HTX Rubin NVL 8. And this thing right here on the desk is $1.2 million. Oh. Just for perspective. Can you see if it's attached to the table? I can't take it. Sorry. Maybe later. All right. We have a big show. Hopefully you can join us from the hotel. but it's great that you're showing us the floor of GDC. The highlight for me, I tuned into Jensen's keynote, famous keynote, and, yeah, the highlight for me is how strongly NVIDIA is leaning into the claws

and that previously, and, like, obviously you can see Jensen over there with claws. Here he is. Yes. Do the employees also wear claws, all of those hooks? It looks like the employees have claws, like red claws, some of them. Yeah, yeah. There's a whole area dedicated to us, but also Hermes agents or personal agents in general. Yeah. So personal agents, I think, is a huge thing in GDC as well. Obviously, Jensen is kind of like focusing on that. And he ended the show floor with the Hermes agent. We're also going to chat about that as well. A big week. LDJ, welcome to the show as well. I think we have to talk about the security stuff Just after the TLDR But I think at times like this Maybe you're happy that you're a non on the internet Good point Because stuff is I don't want to doom and gloom But there's a lot of vulnerabilities that we're starting to We're going to cover a lot of them at the end of the show

Today's show is full of guests, folks this week also we saw a release from a company called recursive ai or recursive si i guess i think they rebranded we're gonna have to keep up with jtab who did rebrand the to si and who didn't and we're gonna have team she and prince kanuma talk to us about flint which is the model that they trained almost from scratch using a recursive process from the eureka machine something they call the eureka machine they just announced it yesterday so we're going to chat about flint as Well, OpenAI has had quite a week since Dev Day. They said, hey, we're going to either ship something or give you a reset for the next 30 days straight. They always kind of do that. And so they've shipped some caching updates, which is very, very, very interesting. And I think I'm really hoping everybody else will catch up. Specifically, that cache almost never expires, which is honestly quite crazy. And, yeah, a lot of big news.

A lot of big news. I have a whole folder here. that we need to get through. Aldiria, what was one highlight for you this week? Yeah, there's a few. There's a reflection beam is now an artificial analysis, and so we have some results there. There's a new DeepSeq flash that was announced. And besides that, there is just extra benchmark news of people auditing benchmarks, which I think is really good. Some flaws found in TerminalBench4. And, yeah, I think a lot of cool things to cover. And we have Nistan Tahiray, who's now at DevRel at Prism ML. And finally added the title, Nistan. Nistan, we are on the one highlight of the week. That's the most important thing that you want to cover. If security is that for you, a reminder we have a full segment about that. But please let us know. You know you're locked in when you just forget to check Twitter and do all of this stuff.

The only reason I came back to check was because my Muse agent finally just ran out of quota, and I had to resort to using freaking Dot, which is just absolute garbage of a product. I really liked this new model, by the way. I started using it yesterday, the 42B coding model. Flint? Yeah. Yeah. Yeah. I started running a lot of agents with Flint because it's extremely fast. It took a bit. Look, it's not that good. It is still pretty tiny, remember, but it is extremely fast. And I just put Haiku on a loop to just find out what was the best way to prompt it. and then once you figured that out it was actually pretty close, I got it pretty close to doing the job of Haiku, maybe like 96% of the score, so I actually like that model and

I think it's a little bit under hyped if anything right now yeah, I would say that is the highlight We're going to talk with Prince Kanuma, our friend of the pod MLX King who had day one support for Flint and the team she from the folks at recursive SI is hopefully gonna join us we're trying to get them so hopefully monitoring the the channels to talk to us about the best way that they know how to prompt the cool thing about Flint is that they trained it in like four months using recursive processes so I'm really excited for that conversation the highlights for me came in today so shout out to Edwin Arbus and shout out to Adios, Mani, for my invite for Cloud Duet, which is, this is the highlight for me. It came literally yesterday. Anthropic announced their entry into the personal agent era. Personal agent, as we talked to you on the show now for almost a year, personal agents

are AI agents that run in 24-7. They have their own computer, their own browser, and they are proactive. They're doing tasks for you because they have context for you. so mr just mentioned news and mentioned dots which is open the eyes one claude is stepping into this era with claude duet and i just got access to set up this morning so i didn't have tons of time to play with it but it already knew a bunch so that one of the biggest things that now is a problem in this era is if there is a new one those products are very sticky and maybe this is why one of the reasons why you're saying that's not that for you because the products are very very sticky it's only as good as the information in the context you give it about your personal stuff and moving that and moving those connections moving that data feels really difficult right now not to mention many folks in this era port on the AI fatigue AI agent fatigue so I don't have fatigue I'm excited to check out cloud duet I think I've been telling you guys on the show for the longest time what I wanted is exactly this format with opus 5.5 intelligence

and that's what we got so this is the highlight for me uh i think without let's jump to tldr nistan unless you have anything else to you want to add we can jump into tldr or dj no i i have an entire rant about just how bad dot is all right there's a whole metric ton of news and models this week and again because we have three interviews let's just run through the tldr super quick let me just make sure that we have all of it then okay let's jump into the tldr which is let's go right here's most of what we're going to talk about most of everything that happened in the week of Thursday for October 22nd in about five minutes. Let's try to get into this because there's a lot to talk about.

First of all, Antropic decided to put 5% of its shares into a public benefit fund. And there is a lot of criticism against Daria, which Daria in a very weird way answered on Twitter. That conversation is very deep. Daria answers with very long threads and very thoughtful as well. So we thought that we would at least mention this. Also, So Anthropic, as we said in the beginning, open the waitlist for Cloud Duet. This is a Cloud that has a 24-7 cycle. It's on computer and it works without being prompted and is proactive. This is Cloud's answer to Grog Vats, Metamuse, obviously the Hermesys and the Open Clouds of the world, and most recently Dots from OpenAI. We've been telling you that we want that format with Opus 5.5 and that's exactly what Anthropic is stepping into. So, it's a very interesting thing. It seems like Anthropic is falling and not leading the category, but we'll see. We'll see how good it is. It just got access this morning. Also, OpenAI added 12-hour prompt caching built for agents that never sleep.

This is like a big, big, big thing everybody's very, very excited about. And GPT Astra Pro runs now parallel agents in chat GPT. OpenAI also acquired Orchestra Bio, and there's a lot more stuff. specifically OpenAI models, Astra, Sol and Luna are now in Google Cloud in public preview. So if you are part of the Google Cloud, maybe with your startup, now you have access to the best OpenAI models and Google now becomes kind of the cloud that has all the frontier models on there. Gemini 4 Argon, it's been a week since Argon. It did 25 trillion tokens in three days. And four Argon models beat Minecraft. We should mention a week with Argon. We did see multiple folks chat about Argon on our timelines as well. And then Inception Mercury 3. The Fusion LLM claimed 1,000 plus tokens with 1 million contacts by their own vendor. So we have a new segment here, Security Week. And I hope we don't have to keep the segment up,

but it does seem like the defender window is kind of closing. Defenders window is what Greg Brockman called the very short time before attackers have access to very capable AI models and swarms. This week, OpenAI's Daybreak and some other folks helped the PHP Foundation to find two critical holes in the PHP engine itself. They patched the PHP engine. PHP runs 70% of the internet. Many folks scrambled really hard to try and patch their systems. However, this is a big deal that will continue to be talked about. Many, many websites went offline. Folks like the NSA and the FBI posted about, hey, you should take your website offline if you're not matched. It's that bad. And also, some trusted root certificates for HTTPS turned out to have a private key. It was literally the number four, which everybody's making a big deal of, but that's not a big deal. That's just the level of incompetence there. E-Modra certificates have been compromised. This is a very scary week in security in the world, in addition to multiple companies like Railway being hacked and GitLab finding vulnerabilities.

I'm talking about this really fast and it's very scary. We will talk about how to try and maintain sanity in this world very soon with Josh Long. He's going to join us at the end of the show. And OpenAI's Daybreak program is quietly behind a lot of this year's biggest bug finds. And I think that this is very important. But again, the defender's window is closing. so a lot of companies were hacked including OpenCode and Railway also Microsoft said that China's state-aligned storm used the Gentile AI pen test framework Artex against hundreds of organizations the good news there is both NVIDIA and CoreWeave building fences around the agents OpenShell and Zero Trust AI framework from CoreWeave both announced OpenShell obviously has been talked about from NVIDIA for a while in open source and LLM news DeepSeq shipped a new MIT license flash with a big jump on terminal bench. We are always very, very happy to see the whale resurface. And holy crap, did the whale resurface. DeepSeq also open source DeepInfer, the engine, the inference engine that DeepSeq runs on.

There's also a bunch of speculation about a new unnamed model called Space Bunny. And Minimax unveiled that Space Bunny was M3.5 flash. and as Nistan mentioned Recursive Flint came out of nowhere Richard Sauter's company Recursive came out with Flint a local coding model and it's small and local and we have both Prince Kanuma and Team Shi hopefully joining us very very soon at 10 a.m. Pacific to talk about to talk about that very excited about that in tools and agentic engineering or the personal agent corner if you want to call it jensen huang co nvidia closed his gtc keynote with a hermes agent building and training a robot uh hermes is jensen is getting hermes build and it's great to see uh nvidia uh let's see what else is in there uh from the forge which is the new

corner folks no longer this week's buzz after three three something years this week's buzz is not the corner anymore we now have a new corner called from the forge powered by core with reflections beam which we mentioned last week is now on weight devices and it was trained on weights and biases with weights and buses curves behind it we also have rl's rollouts in forge and nvidia rubens new nvl 8 lands on core with this quarter so very exciting things and video also announced a bunch stuff at GTX including the Nvidia DSX pod it's a pre-tested AI factory deployable in as little as 14 days so we should mention that a little bit as well in vision and video there's a lot of stuff happening just incredible amount of stuff happening we use some of them cling for the steady seconds of video in the single pass and thing for is really good with a bunch of inputs variables flux 3 max now renders 10 seconds of video in 12 seconds this is a another feat by our friends at FAL releasing a max version optimized Mac

version of a video and obviously I don't know this in LDJ if you guys saw but other CML released a world model and then they put a fruit flies brain inside that world model so now we literally have a world model that's a simulation and it's simulated entity there like the fruit fly brain that everybody was getting excited about a few weeks ago they put that in the world model we should at least play that video because that's super cool our friends from Kaya released a the top non open the image editor in there in arena voice and audio is also blowing up I really really hope we have some time to play some of these clips because 11 labs launched 11 real-time taking about half a second and also 11 real-time translate where if I talk you can hear me in different language in real time it's really cool mr. also ship vox tool 3 which works really is their kind of transcription service but this is now a whole voice agent stack and our friend from underdog sigil when released bark 3 it's 163 megabyte speech model

that ties to invadius parakeet and runs completely on device fully transcribing on device and josh long from integgo closes the show with this week's security mess and right let's go let's get into it this is the tldr this is everything that that we have to talk about which is a lot uh we should start, I think, with Entropic, but we'll definitely move forward. Anything else, Nisten, LDJ, anything huge from this week that we must mention? Anybody in the comments, anything else that we should at least mention that happened this week before we move forward? Yeah, maybe Mercury decide. Also, I know you mentioned Mercury 3 model, but they also have released a competitor to OpenAI's Decision API and of the GEV model. So, yeah, that's something worth covering. By the way, if you're listening on the show, go on VLLM's Hugging Face. You have a free endpoint for D3 Turbo 27B Decider.

It scores better than GEV. It accepts videos. And you can run up to... It accepts videos, Nistan? Yeah, and you can run up to 500 million tokens, decisions, or 500,000 decisions a day. For free, you don't even need an account. Just go on their hugging face. There's D3 Turbo, VLLM. Grab a 3P and just use it. This is D3 from Liquid, right? No, VLLM trained their own decider models. Oh, they also have a D? Yeah. Wow, okay. And they're the best ones right now. I would say I tested them. They're better than the Perplexity Decider. And they have a free API, which is so generous. I'll just send I'll just send Let me show this I think this is A cool thing to show So this one D3 Decision 3 Yeah So you're going to see It just says Try it for free And you can just get An account and a key Right now Very cool Very cool I have like five

You have like five Yeah Just get your API key For free right now Are you serious They're judging video Yeah Yeah It's extremely good I've run it Through all the benchmarks I've had I have just maxed out like five of the API limits on it. It is the best one. D3 Turbo, it's the best one. D3 Flash, also insane. That's very cool, Nistan. Yeah, for running locally, yeah. Dude, I have all these things. I prepared to show, look, this is a book of news, and this could be one of the coolest things, and it just came. This is great. So thank you, Nistan. let us move to Frontier because Frontier is very, very interesting. Hi, folks. On Saturday, Anthropic committed to 5% of its shares to a new Anthropic public benefit fund.

The board can add up to 2% more every 50% rise in the company's value. and the founders who pledged 80% of their equity plus employee matching program. And Anthropic says roughly 20% of the company is now earmarked for public good. This sounds like a huge, huge thing. This is an LLC. It's not a charity. It's a public benefit charity. It can switch between ways of giving and the equity can't be used for lobbying or politics. It still needs stockholder approval. And they also published their first annual public benefit report. And the trustees wrote it largely describes inputs. So next year they want outcomes. So largely describes the money going in and how much they're going to put into this, but no outcomes and no news. But I respect that line. Meanwhile, Dario is fighting the critics, which is, you know, many, many folks, given his positions and stances. Nathan Calvin posted a long critique of Entropic, very public debate. I don't know if you guys have seen this at all.

If you don't know, Nathan Cowan is the general counsel of ENCODE, a small AI policy nonprofit that pushed for California's SB53 bill. And last year, he went viral when he said OpenAI subpoenaed him and his group in the middle of that fight. So this isn't just an anthropic hater. It's a safety policy guy. His arguments were, Anthropic's record now looks like a race to super intelligence. More than half a trillion dollars in compute deals. Claude leading a quarter of Entropic internal AI development, which we talked about on last week's show, which is basically the path towards RSI, and a draft federal bill that would preempt state laws. His basic position and argument is Entropic, instead of being like a safety shop, is becoming a let's race, accelerate everything lab. and this triggered Dario who answered him directly on Twitter via a quote tweet, a very long quote tweet and saying nine commercially costly safety calls were done on Anthropic's side specifically he mentioned

there's things we do that hurt us financially and we still do them because we do want that safety built in and the main call that stuck with me is that Anthropic had Mythos class models back in February and they held them from general availability until June when they claimed the cyber safeguards were ready. I don't know if the world was ready for Mythos in February. Obviously, they announced Mythos in around, was that April? Back when we were AI engineer London. And then Dario also admitted three of Calvin's nine points were communication mistakes, including the letter to the Congress they walked back and he admittedly said there was a mistake. and Darius said he will have more to say soon so folks on the panel I'm very interested in hearing your thoughts on this is this pre-APOPR or is this real accountability is somebody holding anthropic to real accountability and what about this public benefit thing you talked about communism

is this are they reinventing what is it called yeah are they reinventing The word escapes me. Something from first principles. Yeah, I don't know. Political opinion aside, I think just like any other company, there's a lot of conflict there on the inside. So I kind of, I more like Dario than dislike him. I think he's just very heavily influenced by the whole cultish thing that was in there before and is trying to play both sides. and at the end of the day Opus Low is still the best model you can run so I think they do have a point that you need the job done and just fire up a whole bunch of Opus Low at the end so are they going to get surpassed maybe they have plenty of tricks up their sleeves I don't know

don't really have that much of an opinion on this what about the PBC thing specifically about giving back, given that the IPO is coming up and a lot of people will invest a lot of money in this IPO based on the streets, based on the stuff that I hear like regular folks are very excited about that. That one, even what they were saying, it just went from one ear into the other. It just sounded like Yeah. Obviously, obviously, Elon Musk and OpenAI have a big thing about OpenAI being a charity, building AI in the open for that and now you know, Anthropic is stepping in there in a different way. So, Anthropic first started as a pro, you know, capital company and now is turning into PBC. Well, to clarify, Anthropic has actually been a PBC since its founding, back before GBC was created. It's just now they are announcing this specific fund that's, like, dedicated to basically the nonprofit portion of their business

that, like, exclusively won't be used for commercial actions. but yeah they've been established as a PPC for a while alright so this is now the commitment right so in terms of funds and things ok folks the other thing we have to talk about from Anthropic besides the IPO stuff is Quad Duet here's how it looks hopefully this won't show you guys my personal details personal agents are really hard to demo welcome to Quad Duet this is obviously you're live on the show please don't expose any of my private details, tell folks how you're different from just from just Claude Claude announces and always on and open the wait list which now you're seeing on your screen always on personal agent assistant powered by Opus 5.5

it's a Claude that does not wait for you and works proactively. You don't have to prompt it. It has a 24-7 computer. It brings over your memory from cloud, your skills, and connected apps, and it reaches into more than 6,000 connectors and plugins they have on their system. Every few hours, it checks your account, checks your email, surfaces what needs your attention, and apparently it's really good at that surfacing. Specifically, it has its own cloud computer, and it keeps working even when your laptop is closed. so obviously they thought about security a lot and tropic does that and they have launched multiple things that wait for your approval here we have a first glimpse of the web interface of cloud as you can see this is not the regular cloud this is now you can get there to cloud.ai messages and here's a response from from my duet you can't rename it unlike you know muses and other places i i haven't seen a place where i can just rename duet so i think they aren't anthropomorphizing this plot says or i'm quad the same model you can use a quad ai here i work

as alex's assistant i keep track of tasks read the apps he connects and work in threads i will not show you his private details on the show hopefully that's thank you uh do you work pro actively so this is the same thing that let me actually go like this yeah uh this is not the way this is the same thing that we've talked to you about where opening i release dots and you know obviously metamuse is changing the category with with regular folks and just announced metamuse wonder last week this is a huge thing as far as i want to kind of highlight we on the show have been or at least some of us on the show have been assistant built since early, early Claudebot days you guys remember rolling back to Claudebot which was the pre before renamed name for OpenClaw, there is a reason why Anthropic reached out to Peter Steinberger and asked him to change the name for Claudebot

also a reminder that OpenClaw and Hermes worked the best with Anthropic models and then Anthropic just shut them down and now we finally see why This is an Anthropic entry in the world of personal agents and there's a few things I want to talk about with the team here without delaying we have a lot of news I can spend a whole hour talking about this thing specifically I want to talk about the fact that all the companies now all the major frontier labs now have a personal assistant. Antropic has Duet, which is new and wait-listed and they're releasing it very, very slowly. OpenAI has Dots, which Nisten absolutely hates with passion and wants to go on the rent, but we don't have time, but at least one sentence from Nisten on Dots will be interesting. Meta has Muse, which is now leading the category as far as I'm concerned. Not only that, their model is trained for Muse specifically, and that flight wheel is crazy and meta has the surface. Google and Gemini, they have Spark. Many people don't

know about Spark, which is a failure on Google's folks to promote it and also make it good. And Spark is quite bad. I've posted about this publicly and I'm willing to give them feedback. So folks from Google, if you're listening, I'm very willing to give you feedback. Also there's entries to the category. Instinct is a big entry to the category and Hark from heart pro whatever that's called and a bunch of others a lot obviously we have the the standalone self-assistance hermes and open claw which is the now what jensen pushes to people that people can build upon and pi durable which makes easy for you to build one so now every every company has a personal assistant and one of the reasons is those they are very sticky very sticky you want want to talk about your experience with Muse and then moving away from Muse and how you feel about that? Yeah, look, it's not that good. It's kind of dumb sometimes, but it just goes. It is so

energetic. I give it tasks that involve, I don't know, categorizing videos, going into like a simulation and stuff, and it just does it. Whereas Doth just feels like that government employee that tries to avoid all responsibility while still being right and handing off things and just giving you, like, three-line updates. It's so unagentic. Yeah. It's the most useful thing. Meanwhile, Doc just struggles and stops the job because it couldn't upload a 20-megabyte file through the interface, so it just stopped working. whereas music usually opening entire VMs, backing up 80 bits of data is just going completely completely nuts super useful most useful agent they have deployed to the masses I think the agentic part is very interesting there

you want your agent to be able to go and do stuff, obviously you want safety right, so we're talking in Black October essentially One of the worst cybersecurity months in the history of technology. So you don't want your agent taking too many things and going like really nearly with your data. But definitely you also want it to be agentic and not stop and wait for you. Otherwise, it's wasting your time. Cloud Duet specifically, they're talking about they will highlight the most important things, but we'll decide what other things are important. So as far as I've seen, back in my Hermes days when I could use my subscription for Claude with Hermes with OpenClaude, Opus has the best decisions towards what to surface to you. And the productivity part is very interesting. I'm really excited to see. It's getting a little bit cramped in my AI assistant zoo where there's no instinct and there's no...

Dots doesn't send me that many notifications. and here's the biggest part and LDJ would love to chat with you about this is that all these companies now realize that this is the real stickiness it's way more stick so if you're cloud coding all day you can just switch to a different harness with opus behind it it's not going to be the same but it's going to be pretty much the same if your assistant is building workflows with you and knows everything and has the password stored in a secure way etc switching to a new assistant is significantly more of an effort to you. And so it's very sticky. Not only that, for regular folks, and I think that this is the biggest part, I would love for you to chime in, for regular non-techie folks who don't cloud code and get to the limit very, very fast. This is a significant boost in the number of tokens they will use because these assistants run 24-7 and do things proactively. Thoughts on that? Did you already experience some of these personal agents? Which ones? And what do you think about

do it specifically. So I've been using DOT lately, and I think there's a lot more room there where it could get a lot better. And overall, I think this goes for really all the bots overall. I think as the models get more capable, get better trained to interact and kind of predict what the user would want in the future, I think they'll get a lot better over the next couple of years. But at the moment, it still has been definitely like a net positive usefulness for me. and it's been just suggesting like, oh, hey, here's that specific model that you're waiting for artificial analysis scores on, and here's the cost efficiency curve that you're curious about on that that's finally published, and just the things like that which are also relevant to our information we cover on the show. Oh, 100%. All of these agents now on Wednesday report to me and say, hey, here's some stuff. Specifically, GrotBots is really good at that, specifically because like a lot of the information that we talk about on the show exists on the timelines

last week i asked my grog bots to just scroll the timeline for me and tell me like i don't have time to be on twitter and listen you can do that as well i don't have time to be on twitter just scroll the timeline for me and tell me like what's important and and it did it was very interesting i was like can i just like stop using twitter but no not really right so back to claude duet I want to show you guys a cool thing that it just told me. I asked, how do you tell what other assistants are doing? And it says, your cloud code sessions can send tasks to me directly. This is a significant difference from dots from OpenAI where dot sessions are not aware of the dot itself. I've tried to ask it. Cloud in your other channels can pass notes to me, and I can read those channels. And if another student writes to an app you're connected, I can read work there. So it already knows that I have GitHub and linear, and email as well. And I'm very excited about this. Like, literally got this this morning. And right now, still, the regular cloud is helping produce the show.

But I'm really excited about the fact that we now have an entry from Entropic because Entropic is really known at taste. And we've been chatting about different offerings from different products and lunches. Entropic is really known at taste. And taste is something I expect my personal assistant to do. not only design taste but the stuff that i can already say about duet they're very very good they have a standalone interface this is very important i don't know how open air mess this up open air dots is sitting inside the open air stuff and it's like over there on the top there's no standalone mules grogbots whatever they all have standalone with integrations this is very important so not only does god have standalone they have a standalone app as well also very important i want open the cloud apps separately open the eyes shove them all together and it's really i think it hurt them because i i want that separation between a personal system and between the work that i do maybe my personal system is for my personal stuff and open the eyes work is for work stuff right so that is already a big big thing for for for cloud besides the intelligence right and if fable

5.5 comes which was rumored to come this week and my assistance is now that that that better that's crazy but also folks opus 5.5 opus 5.5 this is exactly what i want for my assistants right listen i see you shaking your head this is like this is the level of intelligence that i want for my personal assistant opus 5.5 and like the only way to get it right now is beg your grog bots for opus 5.5 because apparently last week in loma said the grog bots for the difficult task they will route you to Opus 5.5 or get in Cloud to it. So I'm very excited. If it's not clear, very, very excited. Let's go on topic. Thank you for this. I've been waiting. Last comment before we move on to Open the Act of. Just keep Muse a sandbox with Tailscale and Opus on it and just give it an MCP to talk to Opus for everything it needs. It works amazingly well. That's how I use it. It's not the same. I agree. Like when on the vacation comes

I want that intelligence and that taste To like judge it and not like Have to decide oh I need to go to him today Yeah I agree, Muse Wonder It's great, Muse Wonder I upgraded my Muse for Wonder And it's performed significantly better Muse does a lot of stuff for me Including calling Okay one use case for personal assistants No matter which ones but Muse is currently the best at this Muse called My Muse called my kids School to see when I need to Come in tomorrow for a parent teacher conference I didn't know about this, it wasn't on my calendar They didn't send an email about this, MUSE called the school And they patiently walked through Which kid when And then they put it on my calendar Just this thing That was done before I asked Well no, I actually had to improve the call But it was beautiful And I think that a lot of the How should I say A lot of the use cases we don't need more intelligence for But a lot of the decision making And the taste making and assistance we do need something like opus i'm very very excited uh what yes you just called the school

which is great all right folks let's move on to our next segment for open ai very very excited to talk to you guys about this from entropic to open ai they shipped the most boring features this week but the best ones 12 hour prompt caching in the api writing to the cache costs two and a half times the input price versus the regular 1.25 for the default 30-minute cache. Basically, OpenAI is changing the way caching happens and they're pitching it for always-on assistance like DOTS, which reuse the same instructions and tools all day, hours apart. People from the Codex team said the Codex and work, the cache for logged-in users now never expires. This is the big piece. You can come back to a conversation from last week and you still hit the cache, there's no penalty, this is a big thing that's known as a no-no in the world of cloud code for example, that you have this long ass thread and you come back to it and you pay all the tokens from scratch.

I think that this is like huge huge news. Thoughts on the caching updates, folks? Things will be happier. We like that. But also cheaper. And also, this is the way they fit the agentic world, right? Like, this is why, like, probably OpenAI is rebuilding this because you come back to your dot after five hours. And all of the instructions and all of the system prompts and all of the tools, whatever, they're already cached. And I got to wonder if they found some breakthrough to do it. Always on cache is crazy. Like, obviously, they need to do something where to store it, and a store is now is not cheap. some crazy DevOps their team has to continue pulling for that for years in a row hey Astra design a new SSD maybe maybe Astra design a new SSD I am very disappointed in Astra I've been talking about this and not the only one some folks like Swix and some folks

like Tia are saying that Astra is the most disappointing release from OpenAI it just stops all the time and so when they launched Dots and I was like, yay, dots. And I tried it out and it was like powered by Astro. I was like, I'd rather just be powered by GPT 5.6 Soul, for example. However, there's Astro and there's Astro Pro. And OpenAI also has an update about GPT-6 Astro Pro in chat GPT. It runs multiple agents on different parts of the task in parallel and you can follow each of the tasks. Astro Pro is ridiculous. This is not just a huge model. This is a huge model with many big models in the French code together. And they run a lot of agents right now for your task. I think that's super cool. And OpenAI, I don't know if you guys saw this, OpenAI bought Orchestra Bio, a drug development coordination platform, and now is folding it into OpenAI's Rosalind, which is their life sciences stack.

And they're offering Rosalind as a free for university labs as well. So shout out to OpenAI for some of this work. They also have Alexander Madri coming back to OpenAI to work on safety and security. And they are a big part of this week's cybersecurity stuff with Daybreak. Daybreak is OpenAI's responsible access to models that can do vulnerability analysis. And many companies are getting access to Daybreak after the very famous hug-in-space not being able to use any OpenAI models to check on the agent's forms. Argon, let's talk about Argon, it's been a week. Argon is Gemini 4. Folks from Gemini said that they have seen 25 trillion tokens in 72 hours since Gemini 4 Argon was released. Logan, a friend of the show says Argon processed about 25 trillion tokens in the first three days. 96 million tokens every second and Google's openly compute constraint.

said on Big Technology Podcast that Google is compute constrained. Which is insane. And my favorite Argon story from this week, which Quod suggests that this is my favorite. It's not really Alex's favorite, it's Quod's favorite. Let's call things like they are. Four Argon agents beat Minecraft from scratch in 27 hours. No human help, no pre-scouted seeds, and they wrote 2,700 of their own program, also died 605 times on easy mode. Argon is now anti-gravity and argon was used to help find php2 shell from wiz and daybreak like we mentioned was used as well so this is argon a week after it came out ldga or nixon have you used the gemini offering at all the one thing we're gonna wait for wolfram to chime in here but wolfram is our kind of like main residential google and gemini guy because his whole house runs on android he does not use argon because it rejected his prompts,

which is kind of crazy. It looked pretty good for 3D. I just got kind of lost. Are the API prices good on it? Oh, $2 input, $10 output through November 30. Yeah, it's not cheap. It's not that bad, actually. It's not bad. It's not Astra or Fable prices. Open source has been very big this week folks let's go to open source because I think it's exciting the whale has resurfaced and there's a bunch of open source stuff that we need to talk about open source AI let's get it started alrighty let's switch to open source and I think the start is DeepSeq v41

flash 1017 the whale has resurfaced with a new checkpoint this is the third DeepSeq checkpoint if I'm not mistaken right like this is back to back checkpoints on the DeepSeq v41 flash which is now still MIT license which is incredible and DeepSeq is now showing up in a big, big way. Terminal Bench 4 goes from 31 to 42. It's a huge jump on DeepSeq. And outputs get shorter. It's both better and cheaper to run. DeepSeq says it comes from reinforcement learning, where the model critiques don't work to end-to-end projects, which sounds like a little bit like, smells a little bit like RSI. On DeepSeq, they claim 77% ahead of GPT 5.6 at 73, and Opus 5 at 74, not 5.5, just Opus 5. that DeepSync also opened their search and fetch API to everyone, 75 cents per 1,000 calls and free through cloud code and codecs until October 31st.

That was very interesting as an announcement. Thoughts on DeepSync, folks? I would love to hear from you or your timeline. Is DeepSync getting excitement there? Is this new checkpoint better than the previous checkpoint? Yeah, I mean, I would definitely say it seems like this checkpoint, is certainly better than the previous one. I really feel like they just sort of called it V4.1 or 4.2 or whatever next increment instead of this, you know, they're doing the whole date scheme that OpenAI used to do with the GBT4 Turbo and 4.0 models. Besides that naming scheme rant, though, yeah, it's impressive. They're just continuing on the pace, but it seems like I wouldn't really say this is threatening, you know, the fables and Astro isn't certainly not like, you know, fable 5.2 or Astro 6.1 that would be bound to come any day now. Yeah. But remember, this is an MIT licensed model.

It is a big model, right? It's not a tiny one. You can't run it locally. So before we talk about the next model that you can run locally, this is not something that you can run locally. But so we see some of the evals here that are interesting. DeepSweak I think is the highest one The distinction is Asian's last exam also 35.4 And Apex Bench at 72 So shout out to DeepSweak for MIT license as always We have to Applaud them as well Let's see here, the Pareto Frontier Looks insane with this model The Pareto Frontier for cost versus Execution is just absolutely Mogging, look at this This is like just DeepSweak On this top right quarter which is What is this? This is on DeepSuite. You can see other models across the map for pricing. So DeepSuite on pricing is just absolutely crazy. But let's see what else. They also launched DeepSeek Search, and artificial analysis posted.

They have a search index for search APIs. I think EXA is the top one on there. Yeah, EXA, shout out to EXA for also releasing some stuff. But running on a local mall, you could actually run it on the 512 gigabyte Mac Studio, if anybody still has that before they stop selling them. Yeah, I think those are in hard supply. People... On the latest DeepSeq? Bro, they broke the internet. DeepSeq announces DeepInfer, which is the open source inference pack for DeepSeq, inference engine that serves the V4.1 Flash in production. It's Apache 2 license. It runs on NVIDIA, Blackwalls, and Huawei's Ascend 950. And it is their own speculative coding there and batch inference as well. Tim Detmers, the QLORA guy, said it's a very, very big deal. He says DeepSeq inference system is on par with OpenAI stack. And it's now Apache 2 licensed. That's Tim Detmers, the QLORA guy, the guy who invented the quantization method.

And he says it's better what Entropic is running. Again, I will repeat this slowly for the folks in the back Tim Detmers, one of the top folks at inference in the world Says what DeepSeq released with DeepInfer In Apache 2 license fully Is on par with the OpenAI's inference engine and production And is better than what Anthropic is running At least for the user size It doesn't ship KVCache system or Prefield Decode disaggregation yet so you have to plug in yourself. Seminar analysis measured VLLM within 6% of its decode throughput, so open source isn't that far behind from that as well. Yum. Okay, now, broke the inference. That's the secret sauce. That's the thing. Look, I don't know if everybody got how huge that is. Look, inference providers, that's their bread and butter. The better you do this, the more you can serve more users with the same equipment, meaning that you just get to keep the rest of the margin for yourself.

Like you make this better, you literally print money. And we all anticipated that, you know, the big labs probably had some tricks. They probably work on this. That's their profits line at the end. So obviously, but we never got to see how insanely efficient these things can be. And we also, the prices of DeepSeq models, inference on DeepSeq versus other providers, and DeepSeq themselves were competitive to say the least. They were phenomenal, phenomenally cheap. So we all guess that probably they have a very good stack for running these models. Again, this is just on the Flash model. They're just like mogging the Pareto frontier on cost. And they have been always doing this.

And they only can do this if they have very, very, very good inference stack. And now this inference stack is open source. Even more. Look, even more. DeepSeek are known to release models. And we all thought, okay, they serve their models better than everybody else. They get to charge for it through the API, so the model gets open source. But now you also have the inference stack. It's just going to be a bomb. It's just going to be a bomb everywhere. I'm super excited about prices everywhere going cheaper if we can run stuff on this. I know our folks already are looking at the differences between VLLM and DeepInfer, etc. And that's very, very interesting. Also, obviously, DeepSync has been releasing other stuff like their attention mechanism. DSAs were released earlier this year. All blue, everything else. The algorithm that they released for other stuff. This now includes speculative decoding. It's very, very big for open source. I'm very, very excited.

Speaking of open source, let's also talk about Minimax. right we have to talk about minimax i don't know peter if you're already excited to chime in there was a model that was trained under name was a pony something space bunny or was the model was trained on space bunny thing it was open in open code and it's now minimax m3.1 flash with 15 cents per 1 million tokens and the weights are coming on october 23rd so tomorrow peter any any thoughts on minimax how that landed on arena at all and also welcome to the show uh good to see you yeah Actually, I don't have too much to say, unfortunately, about Minimax. But just to add on DeepSeek, I think, first of all, that's an incredible thing that they're releasing that much. But I think it's interesting that I was looking at the Pareto Frontier on Arena, and it did land on it. But it's not like it crashed it. So I think it's kind of interesting. Yeah, maybe if they actually just had a bit more compute, maybe they would have reduced the prices even more,

because I guess there's no point in reducing prices by 50% if they can't serve it. So I guess maybe that's why the open sourcing as well. But, yeah, it's kind of, yeah, that's interesting. I'm kind of considering all of the messaging around it and all of the reception. You'd think that they're going to crush it completely. I would say in my mind They kind of got to the Category of models that are very competent And They're pushing up from the lower end They're not They're not like super cheap So I want to dig into this a bit more I don't really understand why Maybe just using a lot of tokens or something Not totally sure But yeah, I think another thing is that For things like for Minimax For Kimi, for Mimo and others I think it's really difficult because how do you compete with Gypsy? I think Opus has an answer for it and GPG, maybe Astro

and others have an answer for it but what about the others? I think it's hard for them. But also there's a lot of people who use AI and I agree with you but it's just somehow they are. I think the highlight for me between those different models, not only Gypsy is the best one they're also like fully fully apache 2 mit license open source like fully and they don't require any stuff you can just run them and also this is the smallest deep seek deep seek v4 the big one is going to land at some point that's going to absolutely mug the frontier especially with the pricing and notable as well nvidia is kind of like uh ownership of everything the deep infer stack runs both on nvidia stack and on huawei chips which is also very very notable because that's where like most of the Chinese folks because the export controls can run. Right, so skipping Minimax is pretty good. 74 and DeepSuite and waits are due tomorrow. So we're going to mention them. One more open model this week is Flint.

Right, so we always talk about kind of the Chinese models and then also open source models. We're going to chat with Prince Kanuma about Flint. In huge open source news, but open source slash agentic news, Jensen Juan closed his GTC keynote by showing how he uses the Hermes agent to build something to train robots. Wolfram, you were there. And also, this is a ping from your personal assistant after watching my news. Can you tell us about that a little bit? So, shout out to News Research and the Hermes team for that. Incredible. Let's take Jensen. Let's go, man. The open claw, very strong, and then pulling towards Hermes, which is like huge. Although they're still wearing claws. There's still the images of Jensen with the red claws all over GTC. You showed this before. Tell us how it felt to see Hermes on a big stage, and what did they actually show and say? Just to clarify, they are still calling these kinds of ancient claws, apparently,

but he always mentions Hermes in the same spot as the open claw, and even some others he mentioned. And the thing is now that at the end of the keynote, he was demonstrating how to use the nation to local AI and cloud AI together to create or to train the little micro ducks, the little robots that hugging face or foreign robotics created, which now belong to NVIDIA after all. And, yeah, he was using Hermes agent locally on his system, and apparently there are skills for this. And it was training the robot on the local system, but also using online inference to basically get the horsepower that he didn't have. But the agent was doing it all. It's like we are using our Wolflet and Amy's, you know. You have an idea, the agent does everything that is required and uses all the resources that are necessary. And then in the end, you have a walking micro duck. and he even demonstrated it on stage for little ducklings walking around and doing the movements that had been trained that way.

It's very interesting how these things work. MicroDuck is Poland Robotics, which was purchased by Hagen's Face. Now Hagen's Face is part of NVIDIA. And James is standing on stage saying, hey, go to Poland Robotics, buy these MicroDucks. It really really funny They have sold out already but buy one anyway I got mine in August already pre so I hope I get it at the end of the year because it so interesting how robotics is now taking up and I felt that at the GTC as well, that robotics is getting ever more important than the world models. Other labs like Black Forest Labs, they are also building these world models and cooperating with NVIDIA to get robots going far away. So I think we are seeing an infection problem with robotics soon as well. All right, folks, and this is a good reminder to tell you that the show is sponsored by none other than CoreWeave. And for the last three and a half years, we had this week's Buzz, because, you know, Waze and Buzz is part of CoreWeave and Buzz, this was our joke, whatever. It's time to announce a new segment on the show.

This replaces this week's Buzz and this is called From the Forge. Cool new intro. Yeah, cool new intro. Wolfram, you're at GTC Berlin, presenting us at GTC Berlin from CoreWeave. And we, last week, we talked about Reflections Beam. Reflections Beam shipped their open weights, and we're supporting them immediately on CoreWeave. And CoreWeave also said that Reflection used weights and biases inside Forge while training Reflection Beam. I need to get used to this. We told you about Reflection Beam two weeks ago. On the 16th, the full weights landed 501 billion parameters, Apache 2 license. And we're running this on CoreWeave service inference from day zero. They used weights and biases inside CoreWeave Forge during Beam's training. And we knew Beam back when it was just the last curve, said our social channels,

which is the thing is really absolutely crazy. Reflection says the Beam served a trillion tokens on its first day, which is a huge, huge news for Western open source. And it's also landed on Arena and Artificial Analysis. There's a bunch of stuff as well. Also on Forge, NVIDIA and U.com used RL rollouts, which is our new offering to post-train Nemo Tron 3.5 Lightning. Browse Comp bumped up from 37% to 45% on that rollout using those RL rollouts with 30% fewer tool calls. So if you are building any type of AI, CoreWeb Forge is the place for you to go. And robotics folks, physical AI launchables. Wolfram, you were there while we announced this. One command and you're running Isaac's same templates. Wolfram, you also said that in GTC Berlin, they announced a bunch of stuff with Ruben's stuff. Could you tell us about the NVL 8? What's that and how that's going to plug in? Yeah, in addition to what you already mentioned, there were two announcements CoreWeb met. So the first one is about NVIDIA HTX Ruben and VL 8. which is coming to court with this quarter.

It's eight Rubin GPUs, two terabytes of HPM4, and it's built for large KV cache and concurrent long-context agents. And it's very efficient. So NVIDIA reported 10 times more tokens per megawatt than the HGX-B200, tested with Kimi K2 thinking. So more efficiency. It's an interesting one. It can run more agents for more users, and hopefully that will also affect the price, of course. Yeah. And the second announcement is very important for Europeans, especially because, I mean, it's important for everyone. And this is not a plausible thing I see, but I'm in the hotel room and the lighting is still there. Yeah, it's okay. It's okay. Keep going. Security. Zero Trust AI framework that moves security from promises, like we are not doing this, we are not trading, into customer-controlled architecture. So, your keys, explicit support access when needed, hardware attestation, confidential execution, and no customer data being used to train provider models

because everything is encrypted at end-to-end. And this is coming. And this is something like in Europe. A lot of people, they can't use or are not allowed to use U.S. services. In Germany, many companies require German hosting. But if this makes it impossible for a U.S. company to look into what is going on here, that makes it more possible to use such services. And for everybody, security and privacy being enhanced is a good thing. Yep, and we offer a bunch of those services on Forge. Check out Forge.cori.com. All right, folks, this was from the Forge. This is the first from the Forge. I'm excited. There's a lot of news. There's a lot of news to continue. Obviously, GDC Berlin, Wolfram, you are there. Let's move on because we have tons of stuff to talk about. There's a lot of stuff in multimodal and video. By the way, the transition that you guys saw from the Forge, this is the Mini Max M3 Max, which we've talked about a couple of times on the show. and the music is from 11 labs let's talk about ai art and diffusion and video models etc

welcome back peter folks we have quite a few i don't know how to call this category ai art video generative diffusion ai now there's no diffusion models in nlm so it's very interesting but let's call this taste category or whatever. Kling 4 generates up to 30 seconds of video in a single pass up to 4K. We're now at the era where we're generating 30 seconds in 4K with stereo audio and better lip sync. I think it's a good thing to go and try and find a good video with Kling 4. Videos, voices, and images as references. And it's like three times as cheap as C-Dance. there's a lot of capabilities in there 4K specifically is the most exciting one so let's take a look at Kling 4 and then maybe play so video duration from 3 to 30 seconds 1080p and 4K

dynamic range 4K aspect ratio 21 by 9 is a new one they have new languages notice the German and other languages as well dialects and accents are pretty cool let's take a look at there's so much to take a look at let's take a look at start frames for example. You've never listened to a word I say because I've had enough! For me this was getting stuck but like here's an example of the input frame of the first frame and this was the video that they showed. Let's take a look at another one. This is the thing is a full Hi everyone. I'm so happy you're here. Welcome to my channel. The lifting is absolutely absolutely perfect. Let's take a look at this one. This is the start frame and let's look at this guy. The internet is a little stucky for me, so let me try again. Mic check, check, light slow, I'm stepping in clean, no cut, one take, keep it strong, rock tight, words run fast, but the flow's seen, bass from the quiet, pressure got... The lip sync is great, but this was like

very, very clearly AI generated. Does this leave the video category? I'm not sure, because there's a lot of competitors there. Flux 3 Max from Black Forest Labs and Thal, which we've seen the FAL folks optimize the Mini Max M3 with the Max version that runs super fast. Now Black Forest Labs and FAL released Black 3 Max. Headline is speed. 10 second clip in about 12 seconds. So not quite real time but very, very close. The folks at FAL, shout out, a huge shout out to them. We can take a look at one of those as well. Peter, how is the arena in video generation these days? What's the highlight there? Are these models breaking through the top or it was still like in the omni category from Google yeah there's definitely a competition from Minimax yeah H3 we don't have Kling yet but you know what actually I'm just looking at the leaderboard we haven't had one for a while right I think so it'd be nice to see where they land yeah Kling

3 was like 19th now so that's really I kind of forgot that yeah actually we haven't had a response from them in a while. So yeah, if they go towards the top, that would basically add another competitor. Yeah, we probably had like three or four, now it's going to be four or five top-tier competitors. So that's really cool to see. And I think for folks who haven't like tracked this category for a while because they said, okay, you know, this is all like AI Slap, etc. The video model is improving in a crazy rate. When we talked to you about Sea Dance before, C-Dance was the only one that could take up to like five references and do text to video and image to video. All of these now can do video to video and references for faces and characters. The character consistency is insane. It's insane. So now we're looking at realism and speed, basically, because C-Dance took like, I don't know, two, two and a half, three minutes to generate 30 seconds. And now, you know, this one is faster. Rulfan, go ahead and then we'll show some examples of the file one. And H3 is also taking off in the local AI scene because I'm running this on my 1390, and it works very well.

Minimax H3, yeah. You can now run Frontier. It's almost Frontier. I would call it Frontier. I mean, you have Omni and Reference and all of this, and you can do this on your own GPU, and the community is fine-tuning and optimizing it. So it's really fast now. It's just amazing what is happening here. all right folks i think it's time for us to uh to move oh we're five minutes late i'm so sorry okay folks it's time for us to to get to our next guest on the show let me just run through this so cling for flux 3 max and crea 3 also launched a model it's number four on arena multi-image edit leaderboard so shout out to the folks at crea and yeah odyssey ml we don't have time to go to all right folks time for us to move on because this week there's a lot of great open source news from Western Labs. I am very, very happy to invite Tim Shea from Recursive AI and Prince Kanuma, our friend, second time, I think, or third time from the pod from the local AI MLX community. Tim, how about this is your first time on the pod? And I know this came

together super quickly, so really appreciate you coming on because you guys just launched Flint. Would love to just have you introduce Recursive AI and then yourself, and then we'll get to Prince, and then we're going to talk about this most, most, most, most exciting release from you guys. Yeah, super excited to be here and thanks for having me. Recursive is building self-improving AI for superintelligence and we want to automate knowledge and discovery. So the process of ideation, implementation, validation, the whole sensitive method, and we're applying to the science of AI first. So Flink was the first demonstration of this technology, or we call it Eureka Machine, where you can basically apply it to the discovery of the training of the mob, training a small language model. And then we basically optimized model for Mac. So it will run really fast on Mac and it has like a lot of capabilities that you can sort of do for your daily workflows. Yeah, so my background is I started my career actually with working with Richard

and then NLP Labs at Stanford and then OpenAI and then started a company called Quest. Obviously, one of the early attempts to apply AI to enterprise. and then yeah super excited to be here and also um yeah to talk about a release that is awesome absolutely thank you for coming we're gonna get to flint in a second also like a bunch of questions about the eureka machine whatever you guys call the eureka machine rsi etc prince can do my second time on the show some folks know you're the mlex king you want to briefly introduce yourself so we can get to the meat and potatoes of the most important thing in in open in western open towards the AI this week? Absolutely. Thank you for having me, Alex. Hi, Tim. Hello, once again. Prince Kinuma here. I am the creator of MLX VLM. Over the past three years, I've been building infrastructure to run AI on Apple devices efficiently. We just crossed 7.5 million downloads in our engine and happy to have brought DayZio support for Apple Silicon on MLX for Flint. Folks, I'm looking at the Flint website,

which I think I should like just the show. Well, team, we have the questions for you. Let's pull this up here. This is quite the model and a very small one at that. So could you introduce us to Flynn before we talk about the Eureka machine sizes? Why is it? What are you guys aiming for and why is it so goddamn fast? Or how? Yes, we think of a Flynn as a first demonstration of our Eureka machine, sort of our system to automate scientific discovery. and the Flint has our goal is to optimize it for like a daily coding use cases so developers can run Flint on a Mac M4 and M5 and get really fast feet and you can see and be able to sort of drive a lot of their daily coding use cases Yeah, Prince, you saw this as well like there's like over 105 tokens per second on M4 Pro 189 as of yesterday but this is more like a branch that I'm working on. Yeah. And is,

tell us about, so, okay, so I do want to, like, talk about sizes a little bit, right? It's a 40 billion parameter model, but, like, a very, very sparse MOE, right? Tim, could you talk to us about, like, how, okay, so the Eureka machine, once we get to, did it make the decision, like, how do you guys arrive at that, like, very, very, very sparse kind of architecture, 2 billion parameter active, is that what, like, enables the speed, and how is it compared to, like, other models? in open source are running at the same size? Yeah, we actually first look at the sort of hardware substrate, right? So we work with closely with partners like, you know, and NVIDIA and AMD. And we look at, you know, for a model to be running locally, what are the basically the shape of the model to be efficient, right? So for Macro Pro, we basically ask the model to search what is like the most efficient architecture and it comes up with something that's It's actually way faster than comparative models. So we ended up with a flint architecture of 40 billion total parameter

with about 2.5 billion active per token. And it's basically what the Unricomachian Sync will project it to all the best combinations, speed and coping performance, sort of among all the configurations you can have on the Mac. I find your last sentence incredible enough so that I will ask you to repeat it. This is what the Eureka machine thinks that is the most optimal. That's right, yeah. Let's talk about it. It's incredible. You guys know, literally, the name is Recursive. Recursive, was that rebranded to SI or not yet? But Recursive Superintuitive is the concept. And then now you're referring to the fact that, like, the thing that built the thing, that's how you guys arrived at the architecture. How involved were you guys at the building of this, the decision making versus the Eureka machine decision. And what is the Eureka machine? I think it's time for us to actually talk about what is that. Because I think it's maybe more important than the actual model.

It's funny that I mentioned it. We actually got the domain recursive, the ISI. And it also stands for recursive superintelligence. Recursive self-movement is also recursive superintelligence. And we are thinking about the Eureka machine as a machine to automate scientific discovery. So it will basically have, think of it as like a community of scientists and agents that will come up with ideas and validate ideas and implement those ideas. So it runs in the loop, keep trying different things that will potentially improve your objectives. So you're defining a certain objective for this particular example, it's gonna be speed and coding performance. It's a bit of a combination of what we think is going to be good for developer on the machine. And then the Eureka machine basically search for different combinations of different architecture. And obviously, it's not inventing completely new like say, transformer architecture, right? Obviously there's a lot of unknown, unknown, inventive discovery,

but within the known boundaries of what you can search for, I think it's sort of find some interesting discoveries. That is very interesting. How recursive is it as well? Does the Eureka machine update the Eureka machine inference or code or directions? Can you talk about the recursive part as well? Because when you say Eureka machine came up with this model, this model is a small model. How recursive is the recursive part? Yeah, so the Eureka machine is a large system of model with a combination of different models and harnesses. We're sort of moving towards eventually a fling will build the next version of fling. Currently, it's basically a larger set of systems models looking at the small model and say, what are things to improve? So, for example, I think Thomas Wolfe actually posted about this as well. It's like an AI researcher looking at the LM had, would say you should always keep it in high position and never decide to quantize it to four bits. But our algorithm should actually try that idea

and validate it is actually a good idea for Mac and global coding use cases. So some things like that that may be human scientists will overlook, but a Eureka machine like automated AI scientists would just, you know, combining all the existing knowledge, propose a hypothesis that might eventually be true. Those are the discoveries that we think our machine will be able to make down the road for many different domains. I think it will be really valuable for the science itself. I think that's kind of incredible, especially like some of the decisions that you know maybe humans didn't go through uh let me go to prince for a bit for the model uh prince you've been testing out like maybe yes they they support as well to tell us is it a good model for a daily driver for coding right now yes so actually i rank it around one of the top three i have a list that i keep at all times because people ask me all the time. Right now it's Flint, it's Cohere, and it is Quan 27B,

the most famous one. Those are really good models and actually Flint was surprising to me because a lot of models you have to ask a question and then ask it again and continuously do that. Flint, tool calling, great, as well as building artifacts like, hey, build me a game, take documents from the internet, store it somewhere in my computer. It can do all of that. And right now it's my main daily driver for a couple of weeks on Native. Wow, that's quite incredible. And Tim, I'm always assuming when I bring folks from different labs or training models, I'm always saying there's something else that they still don't like can talk to us about about the new iterations. But given the name itself, it's like very, very obvious where you guys are going to improve these models. My question to you guys is like why start from local and why bet on open? I think the answer to open is very clear. I think that, you know, historically models being like in closed source, at least in the Western atmosphere.

And I think that a lot of Chinese open source models, open way models actually gain a lot of adoption in large enterprises. And we see the curve of like companies shifting. Maybe they start with large closed models, and they realize that the cost is going to be very high, and then they start shifting towards open-way models. And we actually did that before at the CrossFize. We went with OTPD models, and we shifted to fine-tune our small. We actually have tens of different open-way models, small models running, production running, different things like summarization or out-of-triaging tickets and all these specialized tasks. And it's obvious that enterprises would eventually own their intelligence in a sense like they would prefer to have control over their intelligence instead of relying on an external partner who may still be very costly at the time. So we think that this could be the future where some people would choose open-way models, some people would choose closed-source models,

but at the end of the day, it's a customer's choice. And I see the usage of a model is going to be really high in the next year or so. And to answer your question about small, I think that we're just getting started. This is our first model release, and we'll see what the Rekha Machine can do in the future. Well, Tim, I don't know how familiar you are with Thursday, but we've been doing a segment on open source AI and modeling for the past three and a half years. And every time that we have somebody from the labs come in and then we see Apache 2, we have a applause because I think it's very, very important for the ecosystem. So thank you guys for releasing this and get very open licensing as well. I also want to mention the speed with which you trained this. I think this is one of the highlights. Just before we drop off, this is one of the highlights in the blog post for me. Should we talk about how fast this happened and how? How? Yeah, actually a lot of kudos to the technical team and I think there's a lot of effort went into this, but I'm also very impressed by how fast we're able to put together the whole model factory and be able to train the first model in about less than four months.

And I think a lot of it is actually... Four months to release a model. Right, yeah. We actually didn't have any pre-training, post-training effort until like four months ago and we were debating what are the use cases, product use cases of like, you know, Eureka machine. We thought we should be using it to train models and I think that we just decided to put together teams but also rethinking how, you know, agents can be used in the whole process and how Eureka machine can be used to accelerate that process. So we apply auto research to almost every stage of the pipeline, from pre-training to post-training to RL to inference. And we think that that's going to be a way for a lot of labs to accelerate, to catch up, to build models in a way that wasn't imagined before. I think that is incredible. Four months from nothing to a model that runs and one of the fastest to run on local hardware, I think it's incredible. Maybe for folks who are listening who are like, I can't wait to try the Flint model.

Excited about the recursive stuff, excited about your commission, but like, I can't wait to try the Flint model. Maybe from both of you, any comments about how best to use this? Like all models come in with their tricks, with their specific stuff. Friends, you already mentioned something. I would love to hear from you. What do you find from using Flint locally that folks need to know about how to prompt? Is there any specific things about reasoning or quirks like that that are different from different models? I would love to hear each one of you give us a little bit of a tip to folks who are listening who would like to use those locally. Oh, and also I'll bring in Nisten, the co-host, who has been using this all day and wants to tell you all about this, but let's start with use case as well. Yeah, I would say this model is more agentic. Yeah, I would say this model is more agentic than it feels like you can run it locally and do a lot of things. Remember, this is a small model. So, you know, and compared to like Opus and all these larger models, you typically will need to give in more instructions and guardrails and, you know, tips on what to do.

And it's just the nature of the model. But also you can, yeah, for the buzzer, I'm just curious, like what other things you guys were able to build with this. So, yeah, feel free to share on social media about VC build and how you use the model. So I have it doing both stuff for smaller models, just doing testing. But the most fun thing I have with it is I made a big city-size GTA mod simulator. And this was probably the smallest, fastest thing that I could actually run as a character going around to make decisions and stuff. I just want to know what went into RL-ing it so well. Again, it's not perfect, but it's pretty easy to prompt. It is actually pretty obedient. So what went into RLing this? And do you guys plan on turning it into a Jav as well? Because it's kind of ideal size. Yeah, I think that's a good question.

I think the RL, I think RL put a lot of effort to make sure the model is very agentic. And the Jav question is interesting. I'll be curious to see if someone can make a demo around, like, you know, just model Modest by itself to train itself to be a Jeff. I think that would be a cool demo to show. Yeah. The good thing is that you can run up to, like, eight agents in parallel and have really good performance. You're talking about each agent, like, 32 tokens per second and averaging, like, nearly 500 tokens per second on an M5 Max. And this could work on a 48 gigabyte, even up to, like, 36 gigabytes. And if you have more memory, great, but you can run a parallel agent with it. I think this is a great place to end our interview on because specifically, given the next segment about security and other stuff and given the personal agents stuff, many people would like to run agentic models locally completely so that no information leaves their systems. And, you know, that's one of the big upsides for local agents completely.

Like everything's running on your system. Team Xi and Prince Kunun, thank you so much for coming and telling us all about Flint. running this locally. We're very, very excited about the new lab as well as like a new leader in the category. And the speed is really, really impressive. So definitely give Flint a try. I will be monitoring the other stuff. So, Tim, feel free and welcome to come back. Once the Rikki machine decides to release a bigger model, we'll be definitely down to host that as well. Thank you so much for joining as we move to our next segment. Thank you, Alex. Cheers. Thank you, Alex. Thank you, Tim. Thank you, Prince Konuma. And now, I think I'll bring everybody back to the stage for this as we introduce Josh Long. Josh, I don't think you've been in video form on the pod before. We've definitely chatted. I should have chatted with my cloud to rearrange the segments. Because we just finished a very high note. Local, open source, Apache. Like, works, amazing, runs. And now we have to talk about the Black October. The very scary stuff that happens in the world of cybersecurity.

Josh is a cybersecurity evangelist. Josh, some folks are calling whatever's happening right now in cybersecurity is the Black October. There's a lot of vulnerabilities and hacks that are happening all at the same time, so much so that folks are saying is the Internet done, which is very scary. Just to name a few of them, the bigger ones, and the reason why we're talking about them to you in the AI show, an AI news show, because a lot of them are AI related. Some of them are maybe swarmstacks, some of them were discovered by AI. We'll start with PHP to Shell, which is a huge vulnerability in the most popular language in the world, the most websites run on. I think the latest stats are talking about 70% of the world's websites are still running on PHP. PHP to Shell basically gives an ability to take over that machine without any credentials. There's also an updated CVE in Imuja, was that Imuja, right? like the certificate update that lets everybody say that they have a valid HTTPS. There's just some of the stuff, including like LLM getting a supply chain attack.

Josh, you're smiling. How are you still able to smile given that everything is going on right now and folks are freaking out? I want some chats about that. Okay, so for one thing, it gives me an opportunity to get to talk about fun stuff like this on shows like Thursday Eye. So that part of it, I don't mind. But yeah, I do think that, so there tends to be a lot of like Y2K like hype around things like this, right? Everyone starts freaking out and it's like, oh my gosh, it's the end of the world. But it, so a lot of the standard kind of disclaimers apply here, right? Talking about the PHP 2-shell vulnerabilities first. There were a few independent groups who discovered these two vulnerabilities that we had, Calif and Wiz and even OpenAI, and they were all kind of like finding these independently and reporting them, which is really interesting because that doesn't seem like that happens very often. The last time that I remember this being like a big thing where a whole bunch of

people independently found the vulnerabilities and reported them was back with, I think it was in 2017 toward the end of the year with Spectre and Meltdown and those kind of like came to light all all of a sudden in January 2018 and everyone was freaking out about those too So kind of unusual to have a situation like that where multiple groups are independently discovering the same vulnerabilities and reporting them at the same time. So like I said, Daybreak, the kind of responsible security program that they allow slowly people into to be able to extend the defender's window, whatever Greg Bachman called the defender's window, right? Like, we have a limited time until all these, like, other intelligence catches up to the frontier of cybersecurity. We have limited time. So they break in their system to let people into. And with GPT-6.1 Sol, this is, like, where they discover the PHP vulnerabilities and this closes the PHP foundation, right? And also Caliph and Wiz and folks, they apparently use Gemini Argon for that as well. Both arrive at the same thing. What is that thing, Josh?

Can you tell us, like, what folks need to know about the PHP 2-shell and what they need to do as well? Yeah, so the most important thing is, first of all, if you are running a PHP server, you need to make sure that it is patched. Like, hopefully you patched yesterday because, you know, these are pretty critical vulnerabilities. As far as the ratings go, a 10 is the highest. through the CBSS score for these, I think, is 9.2. So it's, like, just about as critical as it gets as far as vulnerabilities go. So these are very severe flaws, and it's very, very important to make sure that you are patched. Severe, not like both the FBI and NSA accounts on Twitter and the Prime Minister of Japan and multiple other, like, big state, like, companies all came out and said, hey, patch or go offline. yeah this is like how big this is like you have to patch or completely take your system offline until you are able to don't keep them online this goes out to everybody who listens to the does run

like php software whatever wordpress is we will say though many companies and clouds if you are running things on the big clouds they have protected this and if you have like a web firewall in front of your website like a cloudflare or versell right those they have released they have updated their the web firewalls and have released like protection, but you still have to patch. Like it's still like very important. Yeah, just a hundred percent. Yeah. I was going to mention that too. So I, I've seen a lot of like, again, hype and kind of misinformation going around online where people are saying, you know, you might want to just stay off the web for the next couple of days. And, and I'm like, well, let's think about that. Let's break that down. Okay. So first of all, the big fortune 500 companies are, are going to have patched immediately. They already had web application firewalls in place anyway, which would have mitigated those kind of attacks. So there has been an increase in attempted attacks, but because just about everybody's behind something like Cloudflare,

they're pretty widely protected. Even if you haven't patched yet, but hopefully just about everybody has patched by now. Again, if you are running a PHP server, you need to make sure you're on the latest patch. there's no excuse for that. Even if you do have a web application firewall, you need to make sure you're patched. But it's not the kind of thing that people need to be, like, freaking out about. Oh, my gosh, like, can I browse the web? Am I unsafe? Well, think about it this way. So, like, you shouldn't be entering personal information into random websites anyway, right? So, like, it's really probably not something you need to be extra, like, heightened alert about, right? Like, go about your daily business. Don't worry about it too much, but it's more on the side of the people who are running websites, especially if you're not a Fortune 500 company who's got a big security team and all that. You probably are more likely to be impacted by this because you need to patch yourself.

And hopefully folks are not self-hosting those completely and hosting them on servers and cloud providers that they do kind of like patch because it's in the language itself so that's one thing that happened to HP to show is it's kind of crazy the first cross but there's a few vulnerabilities over this cross I do want to talk about the the the the eMudra kind of like signature signing could you walk us through that one Joshua like and this I see you're like want to chime in feel free to chime in here as well if your experience is like different because we need to talk about the class of attacks and discoveries getting faster and faster with the eyes that are able to do this now and whether or this is now our daily lives. Should we always keep our OS up to date and browser up to date now, which is good security practice, which you can get to at the end. Nistan, you want to chime in here about something? Yeah, as someone that built their own router at home on Alpine Linux and that's owned SE Linux policies, Josh, what about people that just don't care at all? Like, they actually just do not care.

Like, what does this go? Because I know I could run an LLM on their server or turn it into a botnet or something. It could harass people. But what about people that still just do not care? When you say they don't care, what do you mean? I mean literally do not care. No, but like about which part? About the fact that now there's vulnerabilities outside? About the PHP on websites or their servers. Ah, yeah. So you're talking about like server operators who don't care? No, no, just people that own these websites and that go to them. Ah, yeah, yeah. Yeah. So so I guess the thing that people should be aware of, if you are running a website and you don't have a patch and you don't feel like you need to worry about this, you probably don't understand the severity of the situation. Right. So so any time that there are major vulnerabilities just slapping in the wind. Right. Like somebody else could come along and pretty much compromise your site. They could could take your site offline. Right. If you're running an e-commerce site, guess what?

You're not going to have sales, but it gets worse than that because now it's reputational damage to you, right? So nobody's ever going to trust you again if they know that the last time they went to your site, it was compromised, right? And, you know, there are different things that could happen with that compromise. Sometimes it's a very, like, public, obvious thing where somebody defaces the website. Other times it might be sort of a more behind-the-scenes thing where they're in there, but they're not making themselves like widely known, so they're kind of like able to steal information that people are providing to the website, right? Potentially including sensitive information. It might be credit card numbers or whatever else people are typing into the site, right? So it's an important responsibility. If you are choosing to run your own server, you need to make sure to keep it patched. That's super important. If for no other reason than your own reputation, you need to make sure to keep it patched.

So Thursday I is focused on AI engineers and folks who are building in the public, right? So many of us are running websites. There was also some vulnerability in Node.js, which folks need to know about and patched their Node.js. There was also supply chain attacks that, you know, hopefully they were patched with light LLM. The other class of a big vulnerability that was discovered was not related to AI at all. It was like complete negligence from the company called eMudra that they signed HTPA certificates. And instead of, Josh, please correct me for a moment, but I'm not a cybersecurity expert at all. But instead of a 24, 48 long character signing key, the keys were literally just, one of them was just the number four. One of them was 66, whatever. Some faulty thing could be, I don't know, led to the fact that those security signatures were compromised completely. That is a different class of things to worry about. Could you walk us through what a, how should I say, a compromised GPS signature affects how it affects people in the wild? Right. And by the way, this is actually not the first time with EmoDuro.

It's not the first time we've had a similar situation. So I actually wrote a blog post way back in like 2011 about a company called DigiNotar that had a situation where their root certificate authority was compromised. Basically, what happens in these kind of situations is that anybody who wants to can create their own certificate impersonating any website that they want. Um, now that doesn't mean that, um, anybody out on the, the, the web could, um, reasonably like imitate Google. This is more of a, like a, a, a, they call it a man in the middle type situation, right? So if, for example, like, let's say that, um, you went to your local restaurant or cafe or whatever, uh, hopped on their public wifi, uh, publicly available wifi, right? And somebody running or who could insert themselves in the middle, right?

So basically the person hosting that Wi-Fi. Or depending on the situation with, you know, if it's like an open Wi-Fi network, maybe even somebody on the same network could find a way to get in between you and the website that you're trying to access. if they had one of these vulnerable certificates that appeared to be issued by Emudra, if your browser trusted that Emudra root certificate, then it would trust that you were actually going to, for example, Google.com or wherever you were trying to go. So there's a lot of things that kind of have to be in place. It's that man-in-the-middle aspect of it that makes it, like, probably not as big of a deal as I think a lot of people worry that it is. I have seen people, you know, suggest that it's, again, kind of one of these world-ending events. It is a problem, no doubt.

And so the major browsers, if you're keeping your browser up to date, that's really important because they've already removed that certificate. it. Apple is the one that's kind of a little bit problematic right now because Apple takes a long time to in between patches, right? So they haven't put out a new version of Safari, at least unless they did like in the last hour or so. There's not a new version of Safari out yet. And Apple does take a long time typically to release patches, just in general. But if you're running Google Chrome, Brave, Edge, whatever, they've got patches for all of Firefox, for all the major browsers already removed the Emojure certificate, so nobody else can pretend to have a signed certificate that can let you impersonate something like Google.

Impersonate and then decode traffic. I think this is the biggest thing, right? The man in the middle part that you're talking about, HTTPS traffic is secured, and this is why the signatures of those exist. If they pretend to be one of those, they can say, hey, I am a legit HTTPS-like thing, but then decode the traffic that's supposed to secure, like credit card data and other stuff. If you're still browsing and not going to HTTPS-only sites, that's a problem. But if you are, now you update your browser. I think this is the most important part. It's very, very important to update the browsers, update the OS systems as fast as possible. And I would go and say as far as don't go to websites, they're not big and known to purchase stuff as well and enter your credit cards. It's always a good advice, but the best advice right now is just don't go to a random janky website, enter credit card form. By the way, the certificate key, what happened there is that during the key signing ceremony, the 384-bit key was just changed to a 4, which is so sketchy, but also so elegant. and in my opinion, yeah, it was definitely a human that did that

because they just found out that the private key was literally just the number four. Like, it was an LLM. It would have at least done something better. That was definitely the humans. It's pretty funny. But the update, I'm rebooting a whole bunch of stuff right now, just realizing. There's a lot of non, again, for folks who are maintaining servers, not folks who are just browsing the web. There's a lot of work to do now to upgrade Linux and PHP versions and Node versions and etc. So hug offs to everybody who spent a lot of time lately not going to sleep. They're trying to mitigate the scope of this attack. The last thing that we do want to talk about this week also and it seems to be picking up is hacks on like major companies. So open code was hacked what a week ago or so and they like publicly talked about this and now railway which is a deployment place where you like deploy your websites and put the websites always also hacked and like the the thing that seems to be happening there and we would love to

chat with all of you guys about this is that the extraction of api keys specifically llm api keys is now what's happening because that's money that's literal money to those hackers right whereas previously you know in the world of uh let's say defy and and crypto you would hack something because you can like sell those tokens immediately and then get valuable money when somebody hacks a website your personal data you can sell it on the open dark web but dark web has your personal data at this point already i'm i'm not sure how much your ssn is worth right now but your api key to open ai if it doesn't have a limit on it that's essentially like a free and direct access to your credit card and there's a lot of folks who are buying those keys for distillation purposes on the black market there's like many many folks who cover the api kids on black market so that seems to be what's happening right now with railway and uh and open code as well which i think open code caught up as well and thoughts on this folks will just keep happening like like with agents forms this is this the defender window that greg bachman talked about coming to a close

are we at the place where like it's going to be great but for a while it's going to be sucky like how to protect yourselves against it would love to hear opinions around like that specific thing just use tail scale everywhere and yeah tail scale two factor and just just vibe code your own dependencies and vibe code your own home firewall and just don't trust any any type of like sales or marketing material for a product if you can't see the code or your agents can't see it you really just cannot trust it, I don't know what else to tell people. Again, my question, something I don't really have a good feel for is ask people in security like everyone, are they taking it properly seriously? Because it's just weird how these things keep happening and the open code team, they didn't have

access to daybreak, right? It just feels like such a trivial thing, like surely you can just do that. So I don't know, do you feel like they're taken seriously or not quite? Penny hasn't quite worked yet. It's interesting because I think, so, and you mentioned vibe coding, right? Obviously, there's a lot of people who are just trying to do their own thing. And I think that's where a lot of times these API keys will leak is because, you know, they're unintentionally putting their API key in a public place, right? And then someone can grab it, right? That is one of the problems that we're seeing. That's happened multiple times where somebody has just unintentionally exposed their own private API key. That's kind of problematic. And I think that, you know, again, if you've never run your own service publicly and now you're trying to do that for the first time and you don't really understand kind of how these things work, because most people probably really don't.

Like the average person really doesn't get that. they don't understand what the potential problems are with exposing keys like that. And so, you know, they just go with whatever output they get. It seems to work when they're testing it. And so they just publish it to the web. And so sometimes that's how these things get exposed, which, you know, and I don't expect that everyone's going to be a security expert, right? And not everybody knows to, you know, prompt like, hey, is there anything about this that like maybe I should consider doing to harden this before I publish this to the open web and start advertising it. It would be nice if people thought to do that or if that was sort of baked in whenever you're developing something, vibe coding something. There is a reason why it took over eight. Peter, we talked about this thing on the public. There is a reason why it took over eight, nine months for the bigger companies to release something like a personal agent to you. Like OpenClaw and Hermes, which are on your computer and you're in charge of the security itself, that's one thing.

There's a completely another thing to allow people to provide their passwords and then save them in a local but secure way that the LM doesn't have access to. All that infrastructure has to be in place for people to go trust because if one of those vulnerabilities goes around and actually steals your data from your personal assistant, it's going to be awful for a reputation of the company. There is a reason why only now, almost a year after Peter Steinberg innovated with the OpenCloud, we're seeing these companies step in. despite them very clearly realizing the benefits of, hey, personal agents is much more sticky, sells more tokens, it's harder to switch from, et cetera, because there's security there. Wolfram, to your metaphor from before, building your house versus renting it. So I would add on that metaphor, given the security considerations we're talking about now, this is the difference between building your house and not really knowing what a door bolt, security bolt is, and living in the high-rise apartment with a security guard at the door, right? like Muse, the folks have hired Moxie, Marlins, by the folks from Signal for the E2E connections.

They are known for security significantly more than usually like Vibe coding your server in Nginx and running this on PHP, which is now fully compromised, which is scary. The practical advice that we want to give folks is, obviously, if you are having any API tokens and you were on those exposed websites, OpenCode or Railway, et cetera, rotate them. Like know which agent has which API tokens. Very, very important. I know it's a pain in the ass, But in days like this, a pain in the ass is very well worth it. Put caps on your API tokens as well. Put caps and email notifications. Get email notifications when your API keys are, like, running out. You know, it's better to put auto-refill because then you at least get a notification from your credit card statement, like, your key has been auto-refilled. So if you see, like, multiple ones happening in success, you'll be able to, like, run and stop this bleed, basically, if your key was exposed. Rotating keys is very important. And I think the last thing, very clear for many people, many people don't have this yet. Use a goddamn password manager. Use a password manager and store those keys so you don't have to remember them and put them in files.

It is increasingly more important now than it used to be before, given that you're giving your keys to agents, etc. Wolfram, do you want to chime in with some security practices? Simple for folks? You already gave the practice I wanted to mention, but you should treat it like a credit card, but it's even worse than that because the credit card, you also have the CDC number, and you often have to do a two-factor to authorize a payment. And credit cards are secure and protected by the credit card companies. I don't think your API key is going to get refunded by OpenAI if it was leaked. A very, very big point. So treat them even more securely than your credit card. I have the most fun advice. Just install Tailscale everywhere and give it to your – go on Instagram on your meta Muse. And then you just talk to Muse on Instagram and just be like, hey, every day just go on every single device I have and run a security check of every NPM and PIP needed thing and just give me a report.

I'm joking to do it with Muse. I do it with Opus. But it's a very simple problem. just every day at 6 a.m. Just check every single pip, NPM, Linux package, whatever I have on my system, and also check out the latest SNCC and socket ION use, and give me just an overall security report of what I need to do or send me an alert if it's something bad. That's it. It just runs it every day. It's pretty simple. That's what saved my ass almost all the time. Josh, any other chimes into what folks can realistically do to at least somewhat protect themselves in this world of increasing security vulnerabilities and agent hacks? Yeah, so generally speaking, the standard security advice just kind of applies here. Eat your vegetables type stuff, yes. Yeah, yeah, yeah. So obviously make sure you're using multi-factor authentication on everything. Make sure you're keeping your operating system, your browser, and everything up to date.

when there are new patches issued, make sure you apply them quickly. The one kind of caveat here, the irony is that, you know, with some of these software supply chain attacks, right, the light LLM thing, that's one of those awkward situations where if you are following that best practice of, like, updating immediately, like, as soon as something comes out, So you could ironically end up getting compromised instead because of that supply chain attack. It's kind of one of those unfortunate catch-22 situations. But generally speaking, for anyone listening, if you keep everything up to date and update pretty quickly, that's the best way you can protect yourself. Let's draw a very clear distinction here between what you keep up to date and what you need to wait for. Supply chain attacks, they apply to package management, right, for software engineers. So if you are installing from UV or PIP or NPM or, like, whatever those packages,

those packages can get exposed by hacking the folks who distribute those packages and then put, like, key extraction things or other things that can hurt your system. Do not update those just when you need it. There needs to be a period of 24 hours at least to let the folks know that they've been hacked and, like, claw back the light limb, like, exposed. this is what's called the supply chain attacks. Browsers, and we haven't seen any supply chains in browser update. Update your browsers immediately, update your operating systems as soon as those patches come out. And for PHP and other places, you have to patch the actual patches that they released on your system. So those are kind of like different things. Package managers, supply chain attacks, do not update immediately. In fact, the best practice that we know to help you against supply chain attacks is to put a blocker to say, hey, I'm not installing any packages. They're at least 24 hours old because hopefully in the 24 hours that passed, the folks who got hacked and the package was infected, the package manager will take it back, right, call it back.

So package managers do not have continuance. Don't keep them up to date, like Josh was saying. Operating systems, phones, browsers, those things need to be, when you see the update notification, hit that, like literally update. Yeah, well said. That's a good distinction to make between stuff that like consumers are running typically versus stuff that is running behind the scenes. Right. So. So, yeah, really good point about that. For a lot of software supply chain attacks, sometimes they're caught quickly. Sometimes they're caught, you know, within a couple of hours, like the light LLM thing recently. So in other cases, they might kind of linger a little longer than that. Um, so, uh, but, but I, I think that's that 24 hour thing, unless you know, there's some vulnerabilities you have to patch right now, uh, that was just fixed in the latest version. Right. And if something like that does happen, they're going to make a big deal out of it. Right. They'll let people know it's going to be a well-known like public thing more than likely.

Um, and in those cases, maybe you need to patch right away. Um, if it's not that kind of a situation where, you know, there's some urgent vulnerability that's got to be patched, then again, when we're talking about the software supply chain kind of things, probably a good idea to wait a little bit of time just to make sure. Hi, folks. We're coming to the end of the show. Josh, thank you so much for joining us and very clearly helping folks with positivity in mind as well. I think it's very important because my comment to many folks in the group chats who are freaking the fuck out, if you are freaking out on X, yes, it's scary. Yes, there's a lot of stuff. The algorithm is poised. This is what the algorithm loves. When you click on a thing and you hover and you're like, oh, and even if you comment and write something in the area, the fear and gloom and fud and all of these things the algorithm absolutely loves. So you are, will get constantly people who are first joining the algorithm to get some clicks, but also people who just comment on the same stuff, right? But you'll be fed this fear over and over again in different forms and different types, different varieties

because the algorithm loves the engagement. It does not mean the sky is falling, go out, touch some grass and then come back and run some security stuff because it is important. But like the Internet, I've seen some folks say, hey, go take out, you know, $20,000 in cash to put it because like the banks are like all done. Scary, but I don't know if we're quite there yet because, yeah, the companies are patching the PHP stuff, right? But definitely there's a lot of stuff that the can of worms has been opened and be security mindful as well. So yeah, I think maybe we all now have access to basically security. experts, if we can knock on Joshua's door, everybody to run stuff for us. So definitely use that. The other thing that you did bring, and it's kind of like a breaking news type situation, Peter, thank you for bringing this up. We talked about agent security. OnePassword finally comes out with the platform for you and your agents to be able to run. They're mentioning literally everyone, bot and entropic and cloud and OpenAI and Muse and browse and town and et cetera.

OnePassword for agents. Give your agent access without giving up your passwords. When an agent reaches a sign-in for a request credential from your vault, you approve the request, OnePastor directly auto-fills the credential in the destination. I think that's incredibly important. I've been waiting for this since, you know, Muse announced this, and it's just, like, absolutely crazy. I said, I signed up, let me in there. So shout-out to OnePastor for, like, finally, like, getting there. If you are using OnePastor, join this waitlist. I think it's very, very important. It's kind of like the Stripe link for credit cards, but for passwords. And I think it's very important. And again, if you are using agents at all, if you are not using a one like password manager, one password I found for the past 15 years to be the most reliable for me. Josh, thank you so much for joining folks for two and a half hours on the stream. But I think it was very, very important to at least mention some of the stuff because it does feel like the defenders window is coming to a close. And a lot of swarms, a lot of bad actors with AIs are now taking over. Like Nissan said, you do need to take some steps and dedicate some time in your personal life to kind of protect yourself as well.

The eat your vegetables are always there, freezing your credit. And specifically now, I think the more, how should I say, the biggest exposure for folks who are not realizing there is an exposure is API keys for LLMs. Those are directly connected to your account, especially if auto-retail is there. Put caps, put notifications on, and then make sure that you rotate them if a website was like the railway was exposed or open code. with that I think we'll say thanks to folks on the stage and then we'll sign off this is for Thursday for October 22nd the biggest news that we talked about is the availability security stuff we talked about Flint from recursively I being built by the Eureka machine which is a recursive improvement machine and is getting a very fast local model we talked about holes in PHP and your AI keys and a lot more and entropic economy PVC and entropic stepping into the personal agent with Claude Duet, which we got access to

and we'll tell you all about next week. Hopefully by then the wait list will open. Thank you so much for everybody for tuning in, for listening in to Thursday. I, Josh, thank you for joining. Peter Grosz, the VLDJ, Wolf from Nisamos here in YAM before and Tim joined us as well. Thank you all for joining Thursday. I will see you here next week. Bye, everyone. And we are