Nobody Paced: Opus 5.5 & GPT-6 Launched 101 Minutes Apart (Both Cheaper!)

0:55 Hello everyone. Welcome to Thursday I today September 24th. This is Alex Walov talking to you live from another Thursday I an incredible week. Entropic says OPUS 5.5 does fable level work for 40% less and we've been testing this. OpenAI cut GPT6 prices in half and also released two new models. And Meta AI passed GPD chat GPT in the app store as the number one downloaded app in the fastest time ever all in one week. Plus, we had Meta Connect. We had Grock release. We had Gemini come back. It's it's a big big big big show. And I would love to welcome my host Peter Gustv, co-host Wolf from Raven Wolf from France. Peter, how are you doing, sir?

1:42 How is your week? >> My week's been good. Yeah, lots of testing, lots of new capabilities to explore. Yeah, I know we're going to talk about all of this, but yeah, Oppus 5.5 has been fun. It's always cool, you know, when when you feel like you know what the models can do and then they just throw something in. It's like, holy crap. Like, I didn't I had no idea you could even do this. And now, yeah, all of your timelines are filled with this stuff.

2:09 >> The timelines are the social media timelines are condensing. The timelines that we get releases though are just crazy condensing. Now, folks, there's so much in the show that is just like major releases that that I don't know if we're going to get to like open source, which is also very interesting. So, Peter, what is the one thing that's highlight of this last week for you? I think I know, but I would love to hear from you.

2:38 >> I mean, it really has to be OPUS because I think it >> Oppus 5.5. Yeah. >> Yeah. Oppus 5.5 because it is like they've clearly added something. the fact that they can create and if you don't know what I'm talking about is it can sort of create something like with in just plain HTML like we we had this Astra creating stuff in Blender but yeah now in HTML and I don't even see that it's using like some assets or libraries or something like that just handcraft stuff I don't know in JavaScript or something and just creates these amazing animations and I and it's like you know we've been looking at diffusion video and how much better it's getting.

3:20 And there's like >> I have something to show about this this week. Like my mind was absolutely blown. I don't know if you saw it, but yeah, diffusion videos going crazy. >> Yeah. And it's like, yeah, diffusion videos amazing, right? But then you you start seeing like, oh, the LLM can like create obviously not the super realistic ones, but it can create like versions of it. I I haven't shared it yet, but I was put in a recipe to cla to Oppus 5.5 and say, can you like animate it? And it's insane like it just looks real, right?

3:51 So now suddenly there's a whole category of capabilities that's opened up that could could be for children for fun. I created the the you know the rick roll kind of in kind of arty style. Amazing. It just took I mean took some time to make it but I didn't have to put in much work. So it I don't know who knows maybe it's all hyper goes away but it's amazing. It's so cool. While Peter talks, I don't know if you guys noticed that we're talking about new capabilities. We have a new producer today on the show. Opus 5.5 is live producing the show. It listens to us as we speak and put ups these like nice chronons as well. And also in our chat and also on X. So if our stream disconnects, you know who to blame. But Peter, I I'm absolutely with you. I think Oppus 5.5 for me was the absolute highlight of the week. Folks, Oppus 4.6 six is back with fable level abilities and and it talks like a normal person. I it is quite crazy. Wolfrram, I know you're traveling a lot. Hopefully you're back now with us.

4:58 how you doing? Tell us what's your highlight of the week. >> Yeah, sure. I'm here right at the AI engineer Paris hosted by Mistral and yeah >> thank you Wolfen for reporting live and covering that event and going there. Shout out to the AI engineer folks from France. Mistral I think is co-hosting that. So big shout out to them. All right folks three themes for the show of this week. Number one pacing the frontier does not mean stopping. Open AI Ananthropic releasing backto-back on the same day incredible models this week.

5:34 Oppus released 5.5. JBT released GPT6 Soul and GPT6 Luna. Terra is no longer there at half the price. So pausing the Frontier does not mean stopping at any any point. Number two, every assistant is an agent. Not every agent is an assistant. We're going to have to start talking about AI assistants in a different language. Assistants have a heartbeat. Assistants are proactants know context about you and assistants take care of their own context. They have tools. They have a browser. They work while you sleep.

6:10 Agents is a very generic term and I know it's confusing to many people, but we're going to lead the charge. We talked with David Palon from Assistant Benchmark last week and this week we're going to cover Metam Muse and the stuff that they did at Meta Connect. Those are AI assistants. There's a very thorough and distinct line between them and assistance is what people need. Agents is a scary word for many people. So, we're going to talk about AI assistance.

6:35 And I think at some point we're going to have like an assistant corner. And number three is voice. Gemini with a Gemini 3.8 Flash. and Gemini 3.8 flash TTS number one at voice design and sounds like just absolutely human. So, we're going to definitely test this out on stage. By the way, shout out to the beautiful folks from Gemini. I received like a mystery box and I was like, "Who is this from?" the folks at Gemini said, "Hey, we're launching Voice." And they sent me this beautiful microphone and a t-shirt and a cap. And shout out to the folks in there who who just thought about me, which is very nice. But I think first of all we'll start with should we start with TLDDR and then start going.

7:28 All right folks this is everything that we have for the TLDDR on the show this week. We'll start with the main three. Antropic launches cloud OPOS 5.5 claiming fable 5.1 level work at 40% lower cost than OPUS 5. I can concur. I've been running this model on ultra high yesterday in like four things and I barely hit 15% of my weekly quota. It's quite uncanny after all of what happened last week which I prefer to forget.

7:59 just a few hours later, OpenI rolls out GPT6 Soul and GBT6 Luna at half the price of the GBD 5.6 promotional API price. We got price cuts from both companies and it's very welcome and I think the number one news at least for me from the assistant perspective is that Meta unveils that Muse is central to their whole road map. Muse will now have a voice and a video chat that you will be able to talk with your muse like me and Peter are talking right now.

8:32 You'll see it like lips animate, etc. Muse will also come to the Meta VR glasses and a charm device that kind of looks like a Tamagotchi that you can buy for Christmas very soon. And Meta also announced VR glasses with Muse built into operating system. So AI is kind of permeating through all of what Meta released and Meta Connect coverage will do so very soon. In other big news from big companies and LLMs, Space Xi released Grock 4.7 and it's reported 46% on Cursorbench and yet and yet many people are very disappointed with that release. I think we expected more.

9:08 However, they also have distribution. So Tesla now connects with Grogbots and connectors. So literally you can now talk to your assistant while driving hands-free. And I think that that is more important news than Grock 4.7. I think that.1 release of models is now augmented by where we can use those models or assistants and Grogbot in a Tesla which I didn't get yet maybe after the show we will is just an absolutely mind-blowing experience and also open published a four area plan for independent safety assessment but no assessors of start date but this is a continued thread of what we talked about pacing the frontier and letting evaluators external evaluators in to your lab In this week's buzz, folks, we're one week away from fully connected in Moscone House next week. Pitbull, Sarah Gua, Fe Lee, Alex Walov is not nearly close to the level of those folks, but also we're going to be there with Wolfrram. It's going to be sick, folks. I really want you to come to fully connected. We have a free ticket for you. If you don't know where you're going to be after dev day on Open EI, we're going to welcome you to see Pete Boo with us. So, the Thursday I show next week is going to be a little different. We're going to start a little bit later and we're going to host a show folks on the floor. It's going to be exciting. we haven't done a live one since AI engineer in Musconi also. So this is you know backtoback live shows from Musconi. Not too shabby. Not too shabby Thursday. and also Corwe earns the semi analysis platinum cluster max benchmark for the third report in a row.

10:43 Shout out to the Corwe folks for this incredible incredible news. Plaster Max measures performance of Nvidia kernels on how we set them up and core. We for the third year in a row is platinum tier. Folks, last week we told you about Jeff. This is the deal. Diar. We're not going to go super deep, but Jev changed the world since last week. We told you about here with Ali Labs from type safe until this week. Jev is everywhere. But as we told you last week, the open-source community is going to show up. So apparently some folks already trained decision models and so there's like multiple classifier open- source stuff and Swix our friend of the pod the host of latent space and the creator of AI engineer which wolf from is now at hosted do yoga the CEO and co-founder of types safe and they talked about Dogo is not like he's really happy with this they're not like oh no competition is going to come for open source he's like yes this is a new category we all need to innovate it's amazing the open source is going to catch and I don't care about benchmarks. So, a lot to talk about how Jeffy this last week was and how it took over the feeds. but also we need to talk about Jeff Bench from my friend Florian. Maybe he'll show up that measures other classifier models and system one models. so we're definitely going to talk about Leia and and and Jeff. There's Jeff and there's Jeff. let's see how my transcription deals with me saying Jeff and Jeff backto back. Okay, folks. Voice and audio is also just blowing the up. Google debuts Gemini 3.8 flash and flashlight TTS claiming number one on the AI voice design benchmark. Folks, this is a cloning your own voice system from Google.

12:31 We told you about this three years ago that nothing is going to break in the world where audio models will be sounding like you. Nothing is going to happen. And now Google, one of the biggest companies in the world, is allowing you to clone your voice and play it back within like seconds. And it sounds really, really, really good. we're going to play around with this. We're going to play you some samples of including me speaking. It's going to be dope. Quen also debuts Quen 3.8 Omni Flash, a gentic audio video model with 1 million token context. And Quen also has a live translate model, cutting its reporting interpolation lag to 2.3 seconds live translating.

13:10 Quen has been all over the open source news as well with vision, image and video category. Alibaba Quen open sources Quen image 2.1 7B model for image generation and editing with native transparency. Black Forest Labs open sources flux 3 action a 7B a 7B world action model that it says tops rob at 42%. So you know Black Forest Lab is back and this is quite crazy folks. Karen Gwen unveils Act 486, a pre-rendered research demo that lets viewers talk back to a Joe Rogan episode. It's quite the eye openening demo. It's quite crazy. In open source LLM, we have a big one as well. Shiomi open sourced Mimo V2.6 Pro and artificial analysis rocks this the top open weights model at 46% artificial analysis intelligence index and there's a bunch more stuff that we haven't covered in the TLDDR but wolf from maybe one or two more items that you want to make sure that we need to at least mentioned.

14:18 >> Yeah. So Nvidia Neotron 3 it's a diorization model that came out open of course 100 million parameter model for speaker tracking and even overlapping speakers. So this is super exciting and yeah in other news open router has a batch API which can save you a lot of money if you are using open router and you don't need the response now if you can live with it 24 hours later then you can save up to 50%. Well, that is something cost is a very important aspect. Yeah, I think those are the two main things.

14:55 Step one also has a new agentic model with a million contexts, but it's it's supposed to come out as open weights in October, but it's ATI usable now. I think at least I saw it. >> All right, folks. I think this is the TLDDR and I think it's time for us to kick off our show officially with with Frontier Labs. Let's go.

15:34 All righty folks. I think let's let's start folks. Claude Oppus 5.5. Andropic just shipped a model that beats their own fable 5.1 that just recently was released on real work and it is 40% cheaper than Oppus 5 and honestly it just feels limitless. My week started a week ago with burning quotas all over Fable burning my quarters ranges and I was like oh no I cannot do this work because for Thursday I need a cloud for production. Then OPUS 5.5 feels limitless because I was barely able to get to the end of the quarter and I was running workflows and and agents whatever OPUS is back. It feels like OPUS 4.6 and you can talk to this model but it's as intelligent as fable as I was able to to see. So I think the timeline agrees. I would love to hear from folks here. for coding and agentic stuff, most of your bill is cash reads and so the cash read is a significant price drop and you can feel it. Also, it's fast and Oppus is also producing the show Oppus 55. So hopefully we'll see like a little banner that is going to put up for us. but I would love to hear from you guys. how was how's your few days with Opus 5.5?

16:53 Let's start with Peter. Yeah. Well, first thing to say is that we released a score for code arena which is measuring front end and all of that fun stuff. And guess what? It's on top again. So Astra was above everyone else. And now Oppus 5.5 is above Astra. And yeah, it's super impressive on on the kind of Kodium front end. We've got a bunch of new capabilities in terms of using the all of the HTML generations that seems to be just completely insane and all the 3D stuff. To your point about efficiency and cost obviously they did reduce cost by 20%. That is good. They said it's using fewer tokens. It's kind of interesting. I kind of had mixed experience with that. When I was running the harder prompts on Max, it was running for many hours and it was then I calculated the API cost and it was like $80 per generation. So I would guess maybe if you take a median maybe it is fewer tokens but then if you take the top 10% of proms is probably way more tokens is my sense. It's just kind of vibe based sense. I don't have data for it yet. So that feels kind of different.

18:06 I was comparing to GPG6 Soul at the same time and I had like Soul $1 and OPUS $60 side by side. Yeah, I had a few experience like that. I'm not saying that's universal and I do agree with you in just day-to-day use it seem to last longer than your sub, but yeah, I think it's not just completely across the board. >> What happened to me last week was for the first time the things flipped. I remember being very cautious about my kind of like quotas on cloud max account and I remember limitless in the GPT you know soul in the 5.5 days I remember limitless codeex just never thought about reaching my quota like never just never and then Aster came out and then within the span of half a day I finished my weekly quota and hit the reset finished my weekly quota not to mention I got bad results we let's not talk about this fine We're over and it flipped now. Now a randotropic cloud code for the longest time with just no end in sight and I this is quite incredible to me but also the performance of this model it works very well. Wol from would love to hear from you. You know you've been traveling in Non as well. I don't know if you already had a chance to try feel free to chime in. How has your experience been?

19:25 As much as I don't like the politics of anthropic, I do find I was a Typescript dev for like six or more years before this. It writes the best code. So, I've been using it the whole time. Opus 5.5. there are some drawbacks like it is still bad at DevOps. I would still revert to Fable for anything like configuring networking, DevOps, containers and stuff on Linux. But when it came to anything to do with the web, web GPU, web APIs, anything to do with Chrome, it's just been trained on much newer data and it understands even the int8 optimizations of web GPU and it was pretty crazy at writing kernels both like on SGLANG on hardware and making websites and also using webgu.

20:22 And yeah, the model is just an absolute banger. It's it it is that good. it it does everything. It does it fast. It knows how to troubleshoot better. There are a few things which you'd still revert to fable. But yeah, overall that is my default now on most things. Yeah. >> Folks, do you understand what it takes for Non to tell you that his default is not some open source obscure model that he runs himself on 17 GPUs? Do you understand what it takes for Entropic to win over Nista? I don't think you guys quite get what just happened here. This is my default as well. This is a fast model that is as smart. I used to love Oppus back in OP 3 days. Oppus was like Oppus and then they screwed it up.

21:08 Somebody from Tropic said, "Hey folks, we're sorry for Oppus 5. We hope this model would make up for it because Oppus 5 and we told you about this just like 5.0 was an absolute crap model. Nobody wanted to talk to this. It was slow. It was incoherent. It used claudisms. Claudisms are gone. Jargon douche is gone. All of that is gone. This model talks like a normal person. And here's a tip from me to you. If you haven't got there yet, if you don't use Pstack from Lauren Tan, I believe from the Gro Cursor team. Pstack has this beautiful skill called bro. All it does is when you get something that's very very confusing, you do slash bro and it asks the model to repeat it in human words. I I love using that skill. I was like every time even Opus 5.5 or even before it gives you like some cla you're like bro and it explains to you like a human. so you know we'll link that in the show notes for the folks and traffic please don't nerf opus. This is a strong request from us. There is this whole thing about models coming out and they're amazing and then after a few weeks they're not amazing anymore because the hype died down and maybe you tweaked inference. We ask you keep OPUS 5.5 the same it is it now. Please do not nerf OPUS for us. I think with this we can move on to the next news.

22:33 OpenI just cut GPT6 prices in half. And the funny part, the cheap models might be better to use than the flagship model of Astra. GPT6 Soul and GPD6 Luna, the faster, cheaper, and probably more performance siblings of Astramm dropped at half the price of GPD 5.6. So, this is a new next level model at half the price of the previous model that was promo. Soul is $2 in and $10 out. And folks, Luna GPT6 Luna is 10 cents in and 50 cents out, which is like it's nothing. This is the best cheap Frontier model that we can now have. It beats the previous model from GPT. Also, GPT6 family now is Astra, Soul, and Luna.

23:20 Terra is gone. No Terra anymore. And honestly, what needs to be renamed is Aster should be Soul. Soul should be Terra. and Luna should be asteroid or whatever. but okay on Deep Sweet Soul GB6 Soul gets 68.8 and Luna gets 66.6. And OpenAI says that Luna is roughly OPU 5 and Fable 5 at medium effort for 96% less per task. I don't know about that. I don't know. But I've been very very impressed with Luna specifically. Here's the sneaky big news for the builders. caching 90% of cached input. All these labs understand that we are now learning long ass agentic stuff for over 500 like thousand tokens or so and all of it needs caching and changing reasoning effort or tools doesn't break your cash anymore.

24:14 This is like the big highlight from both I think entropic and GPT and I think that's great. if you run agents or assistants this is huge. So here's my honest read on this and I said this in my notes. Astra has been kind of dumb for me in the daily use. Like honestly, it's been burning tokens and it's been giving me crappy results for the stuff that I care about. Soul is great and Luna is even greater. Luna you can integrate into your own products and it's very very very cheap and also fast.

24:41 These ones I'd actually now use. Red isn't fully sold on this. our codex crowd says that the performance gain on this is marginal and the real news is the price and couple of threads are showing soul scoring below GPD 5.6 on deep suite. all right folks is half the price with same score a win or did openai just lose the race a little bit? Let's go to let's go to Wolf. I know you posted some stuff about Astra Wolf. Let's go >> just quickly because I I switched to GBT 6 Z because I was using 5.6 more because Astra is just eating tokens too much and I noticed in my personal assistance that the humor has been John and the soul the writing it was different and I didn't like it and I've seen people online say the same thing and 5.6 soul remains my favorite model from open eye. That happens with GBT5 when it came out. It was so robotic in a way and I feel the same with this one. So actually I would rather stay with 56 or 5.5. Those two those were the big ones for me. It feels like the four oh keep four movement now and say keep 56 please. yeah unfortunately or maybe I do have to go to >> you should you should Peter you kicked off this category. You've been playing with obviously Astra, you're you're super fame. I think dude, I think your level of fame jumped because of your Astro demos. and now we got other GPT6. So now the GPT family of models, GPT6 family of models is complete. And may I just say just before we get to you, the Open Dev Day is next week, folks. And the fact that they released their agent platform which is directly for developers and then they released GPT6 soul and GPT6 Luna and Luna is priced for developers just before dev day means what the hell is coming up on dev day. What what are they going to I I I think I have an idea of like one thing but just to highlight that we'll be covering dev day. Peter's going to be there. I'm going to be there and we're going to talk to like a bunch of folks and we'll bring that all to you next week. But Peter about GPT6 soul and Luna please what what was your experience? Yeah, I it's they kind of went opposite ways I think with with Anthropic and the definitely the important part there is efficiency with obviously with price reductions. I think that's the kind of the obvious point and 50% is a lot, right? And the second part there is that it's genuinely using fewer tokens, right? We I looked at our data a little bit yesterday, so nothing inclusive yet. We need a bit more data, but it genuinely looks like it's using fewer tokens and it's going to rank higher than 56 on our leaderboards as well. So that is the important part there. So I say it's kind of iterative improvement in terms of quality. I think it is better than 56. I think it is at least for me personally it just feels like smoother model and for me it is the kind of the daily driver model if you just want to do normal stuff that makes the most sense to use it's not going to burn your sub and it's I think the problem was and Alex back to your point about just burning your subs is that we were used to pretty generous subscriptions at medium-sized models that were big at the time so like 56 and Opus and remember Sonnet like we used to use that that thing as well. So if you're using these models your subs so fine forgot to say this Sonet and Haiku 5.5 are coming. Yeah, please go ahead.

28:19 >> Okay. Okay. Okay. Yeah. So remember we were using these models and then your sub was fine and then they give us Fable and Astra. You want to use these models instead and then your sub just dies in in like 3 hours. And I think it's not and I don't think we should be blaming the labs like these are more expensive bigger models that they need to run and they genuinely I don't think they're like lying to us. I think they just genuinely don't have compute.

28:46 >> So they gave us more allowance. We just kill it quick and they just wouldn't have compute. So open already remember they had to stop the signups. Right. >> So there's no $200 a month pro tier anymore on OpenI. You only can get to like $100 on that. Also, if we're comparing the two releases, kind of like Opus is still the better deal. You can get $200 worth of like 20x usage and OpenI had to stop because so many people wanted Astra specifically Astra is burning. Question is, my question for you, Peter, is six soul right now replacing for most people what they thought they needed Astro for?

29:21 >> Yeah, I think it it's a little bit tricky. My general rule of thumb would be if you have a open eye sub, you should just use six soul for everything and then just switch the toggle when you need Astra when it doesn't work. It's not that difficult and I do that. That's kind of what I do myself and that feels good to me. I don't feel like it's terrible or anything and it and I my sub goes pretty long way now. but yeah, I think it was just never sustainable. I think it's just unrealistic for us to say, "Oh, I want to use Fable and Astra at my $200 in exactly the same way how a model used to be," which is like quote over a price. It's just not.

29:59 >> Can I say something? I don't know if it's controversial, bro, but here's my take on this after playing going very hard this week, folks. If you guys remember the the token billionaire thing that AI engineer invented and the card that there's no fingerprone token billionaires. I think I'm at like like 8 billion tokens this week. I'm working on something super cool. However, I've tried all the models and all the harnesses as well. So I've tried Corser Devon directly Codex directly cloud code. I have my thoughts on that.

30:33 And here's my take. I personally for my work I don't want an extra powered model that can solve Navier Stokes equations. I want a model that can understand me and build me a simple settings page without going over engineering writing a thousand tests. That's what I need right now. And it feels like what we just got with Opus and Soul. So like the the the one rank lower models that also are sparing your quota is quite that opposably better in understanding me and doing some stuff exactly like I like and soul still needs like very direct instruction and is very good at like computer use for my work. I need a reliable fast assistant model that can understand me better and run fast. I don't need navair stokes. So it's a very interesting thing that that we started the show with. We pacing the frontier after last week Dario Amod and Sam Alman agreeing Ela Musk whatever and we're still getting a lot of very incredible advancements in practical everyday AI brains and to me that's great.

31:45 N we haven't heard from you on the GPDS. Thank you Wolf from hopping on and taking wolf from space we have back on the show Yam Pel. Welcome Yam. We are now talking about GP6 soul and Luna and Niston is up. >> I use Astra, but mainly I just let Cloud Fable manage it, what it does. It's just for the work that I do and I'm pretty opinionated. I really don't like the taste it has or how many different files it creates.

32:17 >> You use on topics taste and then open execution. I mainly just use OpenAI for restyling because it does produce much better websites in terms of design now and I use it for reviews, but overall I don't like having it make a whole bunch of files in my ripple. I just don't like the style it develops at. And I'm but I've been very used to >> you remember we're talking about GP6 soul and Luna, right? like Astra is is a model that have you tried the new ones? Have you tried the the smaller brothers?

32:56 >> No, I just haven't have not used those >> and I would love to know why because I think first of all a lot of folks love your text here on the show cuz they're not always like going with the hype which is great. it's cheaper, it's faster. Why why didn't you just cuz you already have your workflows or do you think that your prompts will not align with the new models or just like too many news in the span of a short week? I mean, it's just for what I do. I know the majority of people do prefer it over Opus. It's just for what I'm doing. One, even Fable and Opus 5.5 are not good enough at DevOps.

33:32 And I find Astro worse. The only thing I find it better is that designing stuff. And if I'm doing kernel work, it just makes way too many files and it just feels it feels dumber to me. So, I just used it as a second opinion. But then again, I've never been a fan of how chat GPT models respond. And even when I've been working with people, I could tell right away when it was like a chip written report and I absolutely hate it. I hate the way it writes.

34:10 Sorry, it's just me. >> not the best writer. I agree. Yam Pelleg, welcome to the show. >> So, >> GPT 6, GPT6 Luna, no Terra anymore. What's your take? Have you used those models? How are you feeling about them? >> Which one of you is in what camp? Who is on on Opus? Who is in GP? So, >> I don't want to take sides. I want to run both. But I will say Oppus is doing just incredible stuff. And I definitely definitely have used the trick that I will now teach you. weights and biases hive mind is a tool that stores all your conversations from all your agents across everything in the cloud.

34:50 It has a very nice feature fork. So you can like you can download another's agent conversation regardless of which machine. I have definitely used fork my codex conversation into cloud and say hey review what he did and tell me what you you would do differently. And that's been my like this is the direction of the stuff that I use. I haven't used the other way around. So I haven't ever used open eyes stuff to go and review cloud but I definitely use cloud to go and review codex. That that's where I fall like true >> that thing that thing sounds useful as hell.

35:25 >> Yeah, it is very useful. Yeah. Where where do you fall? >> Look let me just say directly I wish I could use soul enough to give you an opinion. Mhm. >> It's just it burns the entire account really quick. >> So it does. >> Yeah. Un at least for me. At least for me. >> So here's here's the thing that I I complained about this on my timeline a lot. I don't know if you guys saw Astro just like decimated my limits. And then helpful folks from OpenAI. Shout out shout out to incredible deil folks by the way on OpenAI. Dominic Kundal, a friend of the pod was on the show. Eric Povventure.

36:05 >> He went to OpenAI to work on on the stuff. is very like adamant about no folks you don't understand this is a a skill issue so and Astra are amazing models Eric Povventure he was also replying to people and then I noticed that I had some settings in codeex that up with my whole >> setup >> here's the settings that I had I a long time ago played around with the context window and I set mine to like a million context window which openi models support this absolutely kills their cache system and So this like rips through your tokens. I also had experimental settings for having it keep it notes for itself. And I also played around with how many sub aents it opens.

36:45 After removing all of those and asking the model to actually open the sub aents in a lower tier model, I actually started seeing normal use back in codeex again. So if this is was your issue and you felt complete burning of tokens, see if your codeextoml or settingstol file is carrying some older stuff. all right. All right. I must admit I need to check, but I'm playing with the codeex tunnel file all the time. By the way, it's all the time tweaking things.

37:14 There are so many hidden hidden stuff that are like you should you should tinker with. just want to say look a soul that is more efficient like 5.6 soul that is more efficient even if it wasn't even better. I'm taking this seriously. It's a really good model. It's just that it eats the account really quick. And just on the contrary, I just want to say Opus, ever since 4.6, I had something to say about each of the models, but the new one 5.5 is just good. Hands down. There there's no denying. Yeah, >> I'm really confused about your burning the subcoin because in our in my personal use I don't think it's burning my sub anymore and also on in our data we can see it costing half so it's like more token efficient and cheaper so I don't see why I should be burning it more so I would also say this is such a common pattern for me whether it's clo code or codex or whatever else I find something super annoying something is not working and nine times out of 10 it's some stupid setting somewhere or some plugin is broken or something else.

38:29 So, it's good idea just as a general tip for everyone like if something is annoying you about something just ask your a to see what's going on and diagnose it and fix it. Just I come across something like that every week. They should everyone should have a heal button or something just like >> let them run through it and like reset your settings or something. Just way too common. I will say though, both Astra and the new models, I've noticed this as well, suffer from the pause syndrome.

38:58 Can we talk about the pause syndrome? The pause syndrome is where you talk to the model. You ask it for like a long thing to do and then you steer and then you interject with your thoughts in the middle because OpenI came up with this amazing steering concept where instead of stopping the model and inject your message, you basically steer. I think Entropic got there eventually as well. >> Oh yeah. >> Apparently for the Astra and Soul models, steering too much after compaction causes it to lose the original direction and the model just stops working. It will say something.

39:34 You would steer and ask it a question like a by the way question. The model will answer and stop. This pattern has happened to me and many folks on my timeline. And what Eric from open suggests is to ask the model to write down and jot down some messages for you when you steer and continue working because the steering is the problem. Openi needs to fix this in their own harnesses etc. But steering is absolutely the problem. So if you steer a lot and you get astro pause syndrome, they're not pausing the frontier. They just the model is confused after compaction. So this is a tip for you from the folks at Open EI.

40:10 these definitely are some issues that have plagued my timeline as well. Folks, I think it's time for us to anything else that's missing that we haven't said. The only thing I will repeat, OpenAI released the agents platform, managed agents platform that runs inside your stuff last week, which is incredible and sounds like a this should be the tool that they would release on dev day. They now released two new models and dev day is upcoming next week.

40:40 speaking of the Hivemind tool, producer put it up for us. Thank you, producer. Every agent conversation and forkable and it's free to use. Hivemind oneb.tools. This is the brainchild of Chris Montpelt, one of the co-founders of Weights and Biases because he used to work across many computers, across many harnesses. It's really damn useful for you if you run multiple harnesses over multiple computers. Right folks, I think it's time for Meta Connect.

41:12 Meta AI news app just passed CHBT in the number of people getting it. It's number one in the app store and at Meta Connect they gave it an email address, a control over your Mac and a place on your face. Meta Connect happened yesterday and I did my bingo card on Twitter and I got some of the stuff that I wanted for the bingo card. The the headline is Muse. Zuckerberg was up on stage and said that the central piece to everything they're building at Meta is Muse. Muse is Meta's personal AI assistant and it's number one downloaded app in the app store right now which a year ago. If you told me that Meta would be outranked JBT, I would probably I would believe you as we told you don't bet against the Zach but many people would would have not. As a reminder, last year Zach started hiring with some soup all of these incredible developers and started bringing them from different labs and the soup paid off. So now Muse as an assistant that sits on your Mac, you can control your Mac.

42:17 They're giving it its own email address so you can CC conversations and realtime voice and video conversations. So I assume by this time next month, Muse, my Muse specifically, could join the show as a participant on the show. It works in background while you talk into it. This is the what separates assistants from agents. And there is an avatar you can talk and voice and face design yourself. You can customize your own muse and voice design it yourself.

42:42 And Muse is coming to Meta Glasses. You will be able to walk around in the world and chat with your conversation. And the custom wake word is the highlight feature for me. You'll be able to say, "Hey Wolfred," instead of, "Hey Meta." I think that's super cool. I think if they're training custom wakeboard for me in my Ray-B bands I'm going to work around in the world. This is like us having our own you know Tony Stark assistant. Then there's the hardware and there's a lot of hardware. Rayan Meta Gen 3 with new six microphones and aviator styles. The first audio only glasses that people cannot see the cameras anymore. And not only that, the audio glasses, the video glasses too, they work with the FDA to make them like certified hearing aid, which is I think is a genuinely great move because of the ADA act, the disabilities act in the United States. If you're disabled and you wear meta glasses as hearing aids, you can sue people who tell you to turn them off because of the ADA act. I think this is like a genius play by Zuckerberg that's not talked about. so audio only glasses with no video. and the lightest frame they've made. They kind of Peter, they kind of look like your glasses. Honestly, they just look like there's no no indication that this is like listening to you. and they cost a fraction of normal hearing aid. And honestly, this is could be the most important thing that they announced. And VR glasses. Meta announced a VR headset.

44:06 That's not a headset. This is the first time. They look like glasses. In fact, the the big unveil for which many people compare Zuckerberg now to Steve Jobs is that the VR glasses that they've announced have been on the table with other Meta glasses and people didn't even notice. It's quite incredible. And then I think we're getting to a point where Meta is beating Apple and Open the Eyes whatever rumored hardware. They announced this little Tamagotchi like device that your Muse is sitting in a keychain like device with a screen at the size of the Apple Watch pretty much with a thumb fingerprint that is shipping in December by the time for the holidays and no price yet. I think that this is going to be sold out the second it drops as people think about which gifts to give other people.

44:56 Peter, have you tried Muse already? Is it available in the UK? No, I haven't tried it yet. No, >> it's coming. It's coming. >> I think what's been missing and like if you remember the trajectory, right, is that open eye went consumer, then sort of competing with Google, then open eye realized, oh, damn it. Anthropic is eating our lunch, so we're going to go more enterprise. It's not like they've forgotten about consumer, they're going to release devices as well. But we certainly and and Google's been like okay at consumer stuff. So we've been missing a a lab that is very very good, very focused and has a genuine consumer interest. And yeah, Meta is it. So it looks like they're really just going going all in on making consumer great and like not Freriedman, right? The he understands this a lot. He's one of the key people behind it, right? It's making things the CEO. Sorry, sorry to interrupt. Nat Freeman just was the CEO of GitHub and then Nat Freeman Daniel Grass became one of the more prolific AI investors in the valley a long time before the AI wave started. And then the biggest news about the meta super intelligence labs is that Zach was able to bring both of them on specifically getting Daniel Grass away from Eliaver SSI into the MSL. I think this is like a big big big big move and Natt Friedman is behind Muse and Alex Wang is showing up with Crocs and a crocodile thrifted t-shirt on the stage of Meta Connect and talks about how Natt Friedman worked his ass off on this muse thing. Nat Friedman doesn't talk on Twitter too much. He said that once openclaw dropped, they all stopped everything they're doing and he brought 400 Mac minis into the MSL lab to all run openclaws all the time to figure out what's going on. Just so you understand how early he was. This was in I think he said February, January. That's when we all bought our Mac Minis. Meta has a top level exec that now drives the ideation thing. As early as when we told you about stuff before it blows up on the internet, that's how they got to Muse.

47:03 Now, the security of Muse is they're going for unparallel. They understand that I'm going to put more and more trust in this device and unlike the folks who ran open claw on their Mac minis etc. and that didn't really care about those are folks who are putting their like bank account information integration with one password etc. This all needs to jump over the fear that folks had the meta has all my data for ads. So they're talking about a few things here. One is the business model is new is no more ads, no more your data. The business model is transactions fee. Muse and other assistants are going to do so much for people shopping on the internet, which by the way, they announced integrations with Best Buy and Walmart and other folks. At the same week that Amazon has banned Muse from shopping. So if you use your Muse for shopping in Amazon, you use computer browser, Amazon will ban your account, which is a ridiculously stupid move. But I think it's just muscles like they're still negotiating probably the cut with Zuck.

48:06 and this is these assistants will do a lot of shopping for people. And Zach is like, "Okay, this is the business model. No longer do we sell your data or train on your data. We take a small cut from the sellers fee to you if you use Muse to buy stuff." I think it's a genius model. It's a win-win for everyone. >> I think yeah, it's the problem is it's hard to make money that way. Charg backed away from this model just because it's the we can go deep on this but basically the way you make a lot of money is advertising because it's you know you can kind of raise prices on your suppliers without consumers paying more for it. But look >> I think that Facebook is going to stop advertising elsewhere. I just think that for me is to give it out for free. The business model that they're advertising for it is that hey we're going to have a secure VM which was co- architected with Moxy Marlin Spike the creator of Signal that secure VM as at the level of a cryptographically verifiable solution that even Zuckerberg himself would not be able to read your data because it's end to end encrypted.

49:14 That's what they're touting live on stage. So they are going after hey we know about privacy we know what people think about Facebook etc. here is industry level architects that say that I cannot read your data anymore, but I can like look out at your transactions and I can take a fee from your connectors. That's kind of the business model. >> Yeah, I think anyway, we'll see. I'm I'm a little skeptical and I think they're probably also doing that unnecessarily.

49:38 You know, our parents don't give a crap about that and they don't watch these keynotes and they're not going to know what that any of that means. So, I think that's maybe might maybe overeaging it. Anyway, but I think what I would like to add is that what's really impressing me about how they're approaching it is that I think >> it's the only like really big consumer app that where they don't treat people like idiots. They give the powerful tools.

50:08 It's all the agentic stuff. It's all of that and it's not like some crappy flash model. I think they're taking it seriously. Now obviously like open air to be fair to them I think the HR GPG is like the best it probably could be but if you look at Apple as in the free version and if you but if you look at Apple it's just like oh yeah it's going to like summarize your notifications like they're just not not particularly interested in in really giving the best toys to consumers. Yeah, they can't or they don't care or I don't know what.

50:40 But I think they also care about the experience or they just want to be like, "Oh, we want to be the last ones but do it the best." May I remind you of a famous actress, I think she was in Game of Thrones, that took ads for Apple and their Siri AI two years ago where Siri was supposed to be this like assistant thing that reads your messages from your mom and puts on calendar and does all these things and then Apple pulled those ads because they couldn't deliver on this promise and then Apple launched Siri AI and it fizzed completely in the air and iOS 27 is now on many many devices and many people will use Siri and we had such great hope for Siri as being that personal assistant with access to my stuff and that absolutely fizzle because Siri is not agentic in any sense of the word and I think for us what we need in assistant is productivity in addition to context in addition to connectors to everything I don't need it to be super private on my device whatever because my device cannot read I don't know go on my Twitter via connector so that's what missing for me in the apple stuff and I think that Whether or not all the folks care about privacy, I think it's incredibly important that the company that gives them that product does care about their privacy. And so I absolutely applaud the folks from Meta for following Apple and Apple famously has the secure cloud where your LM execution of Siri and then go if they need bigger LM to go into the secure cloud. Apple cannot read this.

52:06 Even FBI FBI comes knock on their door. Meta is working on that for folks that use Muse. Now that they're shoving muse in the face of the billion people across their devices, they're shoving muse in every Instagram ad and every Instagram story has a little muse thingy. obviously the Facebook crowd will going to get there very very soon and it's been exploding. Everybody who I know was not in our circle in AI knows about Muse. It's quite crazy how strongly meta has an effect on the populace. And I've been talking about this here. Many folks use Google and then Chad GBT came about and many folks use Chad GBT as the new Google. They just ask a question, get an answer, and they continue to the next one. We are like, you know, advanced there. We're now using a agent and assistant for coding. The next iteration is having an agent that's an assistant. There is a difference there. There is 24/7, has its own computer, has its own browser, and does for stuff for you proactively, reminds you about stuff proactively. U renews your driver's license online, which was what I had it do last night.

53:10 Gro went and renewed my expired car registration for example and did it on its own and assistant can do shopping for you and pay for you via the stripe link integration and the one password integration. So I think many many people will get there faster. I think they'll skip the agent part. They'll just go to assistance. Yam thoughts on Muse? Are you getting FOMO from not getting it? >> First let's let's just say that I already got it.

53:36 >> You already got it. Amazing. >> Absolutely. Yeah, I tried it. and brilliant because of you hyping it every every week. So I had to try and it's absolutely brilliant. >> Yes. >> But can we stop and talk about it overtaking chatbt number of users? Seriously >> not number of users the speed with which >> it acquires users and the speed with which it got to number one on the app store. >> Mhm. Okay. That's the number of but they will surpass number of users 100%. Meta has the biggest >> surface area products in the world. That Zach literally stepped on stage and said we have the most people using our products in the world and everybody else and they're going to shove music to all of them very soon. So I expected to like surpass JGBT in whatever. Also, we expect Openai to come back both with an assistant model, which GBD for work kind of already is, but people don't know because it's hidden in there.

54:38 >> And do you guys remember Johnny IV and love from and open the eye and the devices thing that they promised and Sam Alton putting his head on Johnny IV shoulder and collaborating together? That's also coming at some point. Open the eyes hardware is coming and where is it? yeah. Yeah. So, back to Muse. Look, I don't know about the the end device. I must say that why won't I use my phone for this? I I'm asking because you sound really excited about this and I'm like >> I don't know. My phone works great. I just want the software. I want the assistant. Like convince me that I want the device in general.

55:16 >> I will let Johnny convince you when their device comes out. I will just tell you that I cannot wait for Muse on my glasses. Why wouldn't you do this on the phone? Yeah, you could. You could. However, putting out a phone and clicking buttons, shoving this into people's face when you talk to it's like a whole another thing that you need to do. Whereas just saying a keyword and just going about your day, looking the person in the eye is a completely different experience, I believe. Non muse in Canada from last week. That's also big news. So, you got it.

55:49 >> yep. >> Tell us. I got news. I really liked it because I just put it to computer use at first place and it ran into a capture on Teeu. I was just asking look up products on Teemo and then it had a remote desktop session and I noticed it was a Chromium session running inside it and then it would have a inside the app on my Android it would have the remote desktop window and I can go and I I could click the capture with my finger and then it would just continue on. That was the best user experience so far. And I love how open they've left it.

56:29 I'm not going to reveal too much, but it is like when you have root on the whole machine, which is pretty nuts. They're just letting everyone have machines running as root and you can actually do whatever. You mean the the VM that comes with mute the virtual machine that they Yeah, >> it's an 8 gig VM with 100 gigs of storage space. >> The thing is that you told me last week they had a tail scale integration. So I was able to tail scale in. So now I have a complete shell. It can you can even like run a model inside the 8 gigs of RAM. It has if you wanted to. They've done a really good job understand what's going on.

57:06 >> It's it's insane. Like they're on another level >> for free. everyone a 100 million tokens a week plus a whole computer in the cloud that is a good computer in the cloud for free for everyone. >> Do you guys understand the scale of >> one more thing as for privacy that people you can just open a new Gmail account and you can just log in with that Gmail account and have nothing connected to it >> because this week they announced the Muse is going to get their own inbox.

57:36 They'll be able to just CC on your conversation. Okay, they're making it like one step less one step less. >> Folks, assistants are here to stay. AI assistants are here to stay. we're going to cover use cases on the show. I want to share two real use cases for AI assistants and I think that we will have a corner here that will just share our experiences because folks when they get those things and they don't know fully how to utilize them although meta did an incredible job in the ideas tab. So, number one, if you get Muse, go into the ideas tab. Number two, when you get Muse or Grockbot or whatever assistant, folks, go on a walk, hit the record button, and just talk about your day.

58:19 Talk about your business, talk about your workday, talk about your family, talk about talk about talk about what you do. Because if you try to imagine use cases that it can do for you, that's great. But the thing with assistant is that like this is a whole new paradigm for you of stuff that it can do that you don't even know it can do. And a lot of the stuff that you can do or need to do comes from your emails etc. So number one connect your inbox so it knows and number two just talk to it and then say hey here's the stuff about me. The beautiful thing about Muse and other assistants they have this memory file that they put about you and about your contacts and in that memory file this is how assistant knows how to help you. the assistance will only be as helpful to you as the information that you give them. And an additional tip is that there's an import memory feature. When you go to Muse and you click a button there, it will give you a prompt to send to your chat GPT, to your cloud, to other assistants to basically brain transplant whatever information they have into Muse. You can review that and you can send and then your assistant significantly better. And then the idea that they're surfacing there is going to also be significantly more contextual and stronger for you. So those are my tips for now. A real use case that I had is my weekend bot at in Grock, but Muse also does this.

59:35 I arrived very early to put my kids in a karat class on Saturday only to find out that I should have known that they are closed that week. I could have slept for another hour immediately. Then I talked to all of my assistants to first of all check all my plans for the weekend and their websites and emails because I don't look at their emails for cancellations and two plan out weekends ahead of time. So now my assistants go search local news ahead of time. So by the time that I arrive on a weekend, I have a list of plans that I could be doing with kids or with my wife. And that's a weekend planner prompt recommendation for you that you should you should add to your bots like hey every whatever Thursday go and look at the local stuff that are happening around my city. sort them out by free, paid, whatever, premium tier, and do this for this weekend and the weekend two weeks after because things do run out. Sometimes you do need to buy places. That has changed my weekends. I no longer have to like think about, oh, what are we doing? Peter, what's your assistant tip? What is one thing that you're like, oh, this is novel. People should know about this. Put you on the spot there.

60:40 >> For me, I know it's very boring, but just aligning my admin. So like any kind of emails, calendars, all of that flying around, it just I feel like I was always anxious about what's is my stuff aligned? Am I forgetting something? And just whatever bothers you like that, just try and give that over to an assistant that can just manage that stuff for you. I don't really otherwise do a lot of stuff like that, but just the simple things to just get off my plate. That's what I really appreciate. So, I just have more head space not to think about stuff along those lines.

61:21 >> Yep. all right, Non, maybe you've run this for the whole week. What is one use case is coming up for you with Muse that that you think people should know about? so it is structured very much like a very light and very nice Hermes type of build but actually a very lightweight Hermes. So it has the soul file and and memory and stuff. it does adopt personas quite okay. I just give it the persona of philosopher epic tetis and then the prompt is just first learn the meaning of what you have to say and then speak and then I tell it a whole bunch of stuff to say it in that persona and it actually is super nice by default like it's not it's so nicely done it is so nicely prompted it's not overbearing and it's not like chat GBT where it just makes assumptions is like right on the spot as an assistant.

62:19 >> It really is a good product, folks. It really is the product for the mass. We started the show with OpenAI, etc. being for the AI geeks and Openi quickly got to whatever billion people that use it as a chat. This is a good product for many people. Just the notification thing that shows you what it's doing and it's summarizing really what it's doing versus just like typing or showing you CRL command. Just that is worth like so many points. And shout out to Alex Conway I believe is the designer for that. that pattern has existed obviously in Telegram but Telegram only said typing. Here they actually tell you what the assistant does and I think that's incredible. right folks I think it's time for us to move on to Grock and Gro 4.7 was released. We waited for this. Grockbot is Xi's assistant right is running Gro 4.6 six fast and now after promises there is Glock 4.7 and it's kind of a disappointment for me. I you know I wish I had like a w do I have it? I think I had Yeah, I am sorry cursor folks. I believed that your coding mojo will be able to to beat whatever whatever stuff and Grock 4.7 is not that great. I expected more. I wanted my Grockbot to use Grock 4.7.

63:42 it'd be significantly better at writing at writing code at making decisions and that's not that did not arrive. However, what did arrive is Grogbot is now connected to your Tesla which I find is an amazing pattern because your car is basically kit from the 70s show and you can talk to it and it self-drives and it can do stuff. It's in incredible. So this is the jagged frontier of AI we're talking about. You're getting a new model and it's not the best. However, you're getting a connector and it can do stuff.

64:18 So for example, you can talk to your assistant without hands. You can hold it like this and say, "Hey Grock, what are some of my emails that I haven't replied to?" And it will say or you say, "Hey Grock, check in on that thing." And we'll go and do the connectors. And Grockbot in Tesla I think is significantly more important than any advancement in in AI that could have could have gotten. But back to Grock 4.7 Corser Bench is getting 46% back from 40 for Grock 4.6. It's the same price $2 per million tokens with half a million context window. The fast variant is 2x the speed and price but terminal bench is 37 and it's still behind GBD 5.6 soul on deep I trained this model for the longest time and still behind openi stuff.

65:15 Peter any thoughts on on Grock 4.7? Were you excited disappointed? >> Yeah I think yeah it's kind of a weird one. I'm with you that when the cursive team stepped in to do the training, I kind of thought, "Oh, yeah, that's it. They're just going to fly off into the distance and I think they didn't quite get there very quickly." I don't think we should write them off yet. I think, you know, these things are hard and I'm sure, you know, you never quite know they just not quite have enough GPUs at that point and they're just ramping up the next one and it will be all amazing. So I don't think we should make any big conclusions from that.

65:55 But I think well so clearly it's not an amazing model. I think it's going to be fine. Maybe it's like slightly better than the previous one. But actually 4 46 was not that great either. I the the question in my mind is you know they had this deal with anthropic right for 6 months of compute. What happens to that deal? It's I think we're coming up to five months now into this six months deal. So in my mind it goes two ways.

66:23 One is they take the compute back and say look our model needs more compute we need to train more. I think I can see that happening. >> Yeah. >> Or they you know or they say to anthropic you know what pay us double you know you need you need the compute you're building good models pay us more. >> I maybe this is rumors maybe I'm completely wrong but I don't think it's a scale problem. I think they have enough to train Grock as well. I think it's a talent and data problem. Elon Musk got incredible talent when the cursor acquisition came over. Obviously, Grockbot is a cursor product that they took over, slapped the name Grock on it, and now is on every surface. They're doing the same to their surface area that Meta is doing for Muse. They're putting Grogbot on top of every cursor cloud instance. You can have a Grogbot link now in every X chat. They're opening this up to like all of the all of the X users. They lowered the price.

67:24 It's not quite free yet, but I'm pretty sure they're going to look at Zach and say, "Hey, this is now free." But for the model training, I don't think that the cursor data did a dent yet. I don't I don't see it. Like I don't think that now I'll go to Grog 4.7 and start using this for my coding assistance. If it's better and I use it inside Grogbot, that's great. And I think here we're seeing like a battle lines are drawn where the interface/h harness/ass assistant where through which you will talk to these models matters more than the actual model. Like literally I am more excited this week about some of the features in Grogbot that launched which is one password integration. Folks, if you don't use a password manager, please, please, please sign up for one.

68:11 It's important even more in the era of AI assistance. Do not use the same password for all your services. However, an additional reason to use password manager, specifically one password, which I love with a passion over a decade, nothing happened, nobody hacked me, anything is that those assistants now integrate with your own password. You'd be able to share your passwords in a secure way that the assistant can see. So, I'm more excited, Peter, about this one integration with Grogbot than the whatever next step of intelligence that they didn't quite hit because it for for most users, this is like the most important stuff. Nissan, thoughts on Grock 4.7?

68:49 >> Yeah, look, there are 9.4 million Teslas or more out still in circulation as total and we're just we're >> 9.4 million. >> Yeah. Yeah. in total production over well it's been over like a decade >> I think there's way more than 9 for >> we we will we'll ask producer fact check Niston live on air go and find out the best estimates of how many Teslas are currently driving on the road but yeah N go make your point we'll see if >> it is around that that ballpark I mean as far as I only checked with the bots but you have to think that even the initial like Nvidia SOC's that were in those they might not be too powerful They might barely be holding the interface, but you don't need anything to run the agent on it >> because it's all just API calls, >> just the internet connection. Yeah.

69:39 >> Yeah. You've just offloaded all of that server side. So, it is pretty Yeah, it's pretty nuts that they reach those. on the other hand, as for having all the cursor data, yeah, you have a lot of enterprise data and code and stuff, but that might not actually be good code. You might need like very smart filtering pipelines and stuff like that. This is why I think Codex and Anthropic are are ahead on that. Like yes, they have the data, but they pro they might not just be using it that well or be able to generate that good of a synthetic set.

70:14 >> I think that I think we're underestimating what I just said before. I I I need to Yeah, I need to bring your energy about like Navier Stokes from last week. My car can drive itself and I can talk to it and it can read my email. My car can drive itself and while it drives itself, I can voice interface-wise ask it to do my emails for me. Do you guys realize what the amazing world we're living in right now?

70:43 >> Wait, do you have Grock bot in your car? >> That's what I'm saying. Grogot integration into the Tesla has dropped this week and it's insane. >> Oh, shoot. Okay. Right >> guys, this is a beautiful world we're living in. we get like amazing like discoveries and presence. By the way, our live AI assistant has fact checklist. I think >> I was not wrong. I was pretty there. I said 9.5. I was right in the middle.

71:14 >> Yeah. So Nissa, you want to read out the fact check? Tesla reported 9.2 million cumulative deliveries as Q1 of 2026 and 48,000 in Q2. So 9.7 through June with Q3 progress it's around 10 million delivered by now with a few% scrapped. So Nist is right a bit more than that. Yeah. So call out the 10 million Teslas. Nissan fact fact checked live by our AI producer. >> Yeah that's pretty cool. We should have it for our politicians now please.

71:42 >> Can I just say like Yeah. >> Listen. It just listens to us and and that's it. Like it's not it brother. It's not it. If you guys want to pause for a second and talk about how AI how there's this new concept called AX assistant experience or agent experience and I believe given where mus is going given grabbot etc that like many many many surfaces and services will need access to assistance because APIs is not enough MCP is not enough I am planning to make Thursday the most assistant build show in the world we have an AI assistant that is a live producer if you guys go to Thursday live we have an assist our AI assistant is also like marking off and checking off the the stuff that we're talking about and and and shows pace. So, pacing is very important, right, in in the world of live podcasting. So, if you go and type Thursday.live, you'll see all our notes and me talking.

72:35 Obviously, there's going to be a little bit of a window in the window in the window situation. but you can see the rundown like we're running through all of the stuff we need to start talking about fully connected. We're like a little bit behind. All of our links are here as well. we have a producer that's listening to our show in live real time and and putting up things. not only that, the the another producer will edit it and so you will get the show much faster after it. So we're working to being the most agent build show in the world.

73:04 >> Most agent pilled ever. Everything is live listening to us just doing what we're saying. I totally agree with you that transparently being in the background and acting on whatever is going on regardless of the AI. Oh yeah, absolutely. Absolutely. The nextG interface AI that just listens to you and does stuff. But you just talk about the actual just using Grock as it is not within inside any harness because >> I don't know if if it matters. Does it matter?

73:36 >> Okay. I like where where would you use Gro 4.7 and Gro code the thing that they launched that like I don't know many people who use cursor has fable and Astra >> it's a cursor model it's default default in cursor definitely I think it's I mean it is targeted at least somewhat to developers >> it's disappointing and this is also the reaction for everybody else however sitting in my car while it drives talking to my cars when it does guys my car has an connection Do you understand what I'm talking about? My car when I talk to it has an SCP ability to do anything in the world that has MCPS.

74:15 >> Okay, >> bro. You just you just want it to drive like you don't need >> put a robot arm. You have to put a robot arm in it and then when you get snacks or coffee, you can just to you. >> But but seriously, what can you do? What can you do with it in your car? Like what what the difference is? >> Open up your garage door. Yum. garage door connector smart home. You can say, "Hey, Tesla, open up my garage."

74:39 >> Can you tweet from it? >> 100% you can tweet. >> Oh yeah. Okay. Can you do that? >> What do you guys mean? MCP is an is a connector universal connector to everything. I saw one dude showing an example sitting in his car and saying, "Hey, get me my favorite thing from Starbucks and drive there." And the Grogbot sent a invite to the Starbucks order that he usually gets. He didn't even say what it was. And then sent a ping to his Tesla on the map. And all he we needed to do is hit drive.

75:05 >> By the time he arrived to Starbucks, his coffee was waiting for him already paid. >> That's pretty crazy because while it might not be that good at coding, it's more than smart enough to just handle other agents to do. >> It is really good for handling a lot of agents, specifically Groudbot when they talk to each other. And this whole show is also produced with Grogbots. All right, let's go to this week's buzz, please. tell you all about weights and biases and biases.

75:45 Everyone, welcome to this week's buzz, a corner of Thursday, where we talk about the company that makes it all possible, Cororeweave. Next week, Thursday I is live coming to you live from San Francisco Mosone South at the fully connected 2026 show with folks like Dr. Fale Lee from World Labs with Battlebots fighting out live on the show floor and yes this is 305 Pitbull is headlining fully connected is next week September 29th to October 1st Muski South in San Francisco the day after open dev day by the way. So if you are flying out for dev day, come to Thursday I you can get our tickets for free if you follow the show and our newsletter and we're doing the live show on the floor. I'm very excited to hear the keynote from Dr.

76:33 Faith Ali from World Labs. Sarah Gua is hosting a panel folks like Jerry Louu Tom Rochester Lucas Bewald and Corwave CEO Mike in Trader is going to talk about how we received the platinum tier on the benchmark from semi analysis. We also have Ian Buck from Nvidia and live battlebots and Pete Bull is going to close the show. The Claster Max 3.0 from Semi analysis is our little flex this week yesterday and we got platinum. It's a three reports in a row and we're only the only provider that's been platinum all three editions. Dylan Patel called Coreweave the operational benchmark for the industry in GPUs. And hey, we'll take it. I really do hope that you join us in in person. All right, this has been this week's buzz.

77:14 There's a lot for us to talk about. Let's move on to let's move on to the world of open source and Jev. I really want to talk about that correlation. Open source AI. Let's get it started.

77:47 All right. Bring it back everybody. bring it back to the show. All right. So, one week after Jeff dropped from Typesafe AI, the open source clones are here and one of them is already beating Jeff in its own benchmarks. Last week, I told you that Jeff is a ChibiT moment. Jev from Typescape AI, a new system one type model. We had Ali Labs the Devil for Jeff talk to us about the different the primitives that that you can use. I told you about the open source will copy it fast. Well, it took less than a week.

78:30 Over the weekend, more than tons of GF compatible project showed up and they all speak the same format, decision-making format. And so, you point it to the official SDK and it just works. one is classifier dev and and Jeffbench is something that I wanted to show you. So I will pull up Jeff Jev Bench. Meanwhile Nist would love to hear from you your thoughts about Jev because last week you were somewhat skeptical slash disappointed but now I think you've played with it. What what do you think about this new paradigm?

79:07 yeah, it actually brings the fun back in computing because you're thinking more in terms of like TypeScript or object-oriented thinking that yes, I have all of these tasks and I just need to categorize them really fast and they could be images on the screen, stuff you have to click, emails you have to write and it's now you just use an LLM in a much faster way for this and it's just excellent at feeding it big large JSONs and then coming up with other large JSONs and it does the accuracy better than an open source LLM often.

79:48 >> Yeah, >> even the ones that you can run at home. So this is this is actually super useful and I'm really surprised as to how close the open source alternatives are. It's just that jab itself does not really make mistakes and the open source onesh tend to have about like 1% or like a little bit less than 1% mistake rate which is still a lot less than using a natural LM even if like a somewhat good one. So yeah this is yeah it's a new paradigm in how you make the apps it'll just be in every app basically >> server side client side. Yeah, >> this is like a new paradigm in in software engineering where you can because of the speed and performance and cost of system one models, Jev specifically, but also open source ones, you can put them inside your execution chain in software and rely on the non-probabilistic how should I say execution of this that it will work the next time your software runs. It is the microprocessor of the next era of software engineering.

80:59 It is that big. Our whole timelines collectively were Jev build. It broke through the bubble. My friends who are not into I know about Jev. Every every engineer that I know is thinking and thinking, "Oh, how can I use Jeff more?" I've been completely Jeff build since before we got Alli on the show. And then Swix, a friend of the podsw hosted Diego Almeida on his podcast on latent space. And that is a conversation that I recommend everyone listening to. I would love to hear from you about how big is this deal and also the open source catching up.

81:38 >> Absolutely big. >> Absolutely. look the thing about LLMs and and and all all of that is yeah they can do general things and you can just talk to them and agents also figure out but the thing is that it everything becomes probabilistic and there are ways to mitigate this and so on. But like there are building blocks that if you had a way to like compartmentalize and and you know branch out and choose how to use e easily you can do many many of the things that you do with LLMs just without using the LLM that are much more you know out there in in terms of all over the place in terms of how probabilistic they really are and Jeff just I I didn't expect this. I kn I know about all encoders and and so on and how they work, but the fact I I didn't take into consideration that if you train a really good one, probably a really big one with all the modern inference tricks and so on >> and data a lot data >> if you do it well enough, you get to a point it works really well, near perfect and nearly free. I mean that's I think this is the big deal that it's basically fast and basically free to use. So you can just you know just spam spam decisions in like in insane rate like like you did in your demos. I mean it was the the rapid fire. It was the the speed. Look, I'm all for training open source ones. Absolutely go for it. No problem. I'm just saying that it's not easy to get to these results and I think we should wait a little bit before we claim an open-source replacement for gem. they are they are impressive. I must say that they look quite impressive, but I would wait for more benchmarks to to to measure them all out in all sorts of different ways.

83:53 >> Yeah. just because look the the people behind there told you we worked on this full time >> peoplehindo is the guy behind GPT and our LHF and his takes you should listen to the show because Swig takes him deep on the process of how did you get there. Basically the summary is he was thinking about who will use computers and and he's like okay AIS will use computers way more than human but chat GPT and LMS were trained for humans to use so who's going to talk to other AIS AIs but they're now built to write text and respond as a chat that says oh I'm so sorry about that my friend but computers talk language machine code so this is why he's like okay we need to build AI for computers to to talk to and how do computers talk formatted structured structure with JSON files. He has a story there where Sam Alman told him go work on this and he's like no and then he worked on it for two years after the pre-training was done. But also the important going back the important note about open source yum is that it may not get quite to that level but it gets very close. I want to continue the show a little bit. So while this runs I want to bring up Florian. Let's say hi to Florian. Everyone, Florian, say hello to the folks. Welcome.

85:18 >> Hi everyone. >> How are you? >> Florian, I want you to tell us what you built last week. I think this is a super exciting part. We just showed it off, but I would love to hear from what you're doing. >> Okay. Yeah. I mean, last week has been kind of insane. I've kind of barely slept like I've been obsessive. but I loved it. like it was like >> what were you obsessed about? Why didn't you sleep?

85:47 >> Yeah, Jeff. I still feel like Jeff is completely underrated still. Like I mean you can't call it underrated because it got like 40 million views and everybody talked about it for a few days but still I believe it's kind of underrated because >> what just so many people haven't really fully understood is it unlocks a lot of use cases like because of this unique combination of three things. It's incredible fast. So in fact they can use it to play Doom and win a level like you have to be really fast multiple times per second and then it's incredible cheapest. So you wouldn't do it with any normal large language model because it would just blow your bill and you couldn't pay for it. And then it's also incredible intelligence. So that's like the combination of these three. You can't just look at one of these dimensions and say, "Oh, I've got a CHF competitor. You got to do all three."

86:39 And so like like right after it was released, I started to build a benchmark because I was totally aware that the open source community will start to work on this right away. It's obviously a small model. If it's fast like that, got to be a small model and small models are kind of affordable to train. So basically at least if you if you believe you can just like hack an existing open source model and make it a chef. And yeah, it came as I expected. Like the next day it was like 10 or 20 competitors. Now it's more than a 70 on this benchmark.

87:13 I've I've like been benchmarking these things. I've been fending off approaches to gain the benchmark. People have tried to to hack me, try to exploit my machine. People have tried to steal the the benchmark test set. All that stuff. But it's super exciting and it's super fun. I've been like I' I've had like more than half a million views and I'm really a small small unimportant Twitter account, you know, >> very important Twitter account. Let's not let's not sell ourselves short. I think that benchmarking is incredibly important part. If Wolf was here, he would like very very much agree with what you're doing. Folks, I really wanted to shout out Laurian for benchmarkheaven.com.

87:55 Please go there. And Jeff Bench specifically jumping on this new thing. despite the creator of type safe and jev hates benchmarks as a concept I think it's a very incredible thing for folks realize hey not all open source are the same not everybody trains on the same data I fully agree with you that open source is important from that perspective because people may not want to send data into types safe despite types of being like the next maybe frontier lab people may not want to send or may not be able to send we're talking about medical stuff with nist we're talking about different things on prem like people run open source specifically for we cannot legally send stuff especially Floren I don't know where you're from I think Germany right >> so Europe has very string laws and like mistral the whole reason for the existence that they're the only one who can sell to European governments people want stuff on prem or approved versus this and when you when Jeff K5 on the benchmark gets 62% of the stuff that Floren is is testing it looks the most where's Leia on this by the way.

88:59 Floren, where is Leia? Cuz that's the one. >> It's not performing very well on this benchmark to be honest. it it used to be on place too for a moment, >> but like I I improved the the benchmark. You got to really keep up and now new new competitors have shown up. So it's pretty pretty down. I mean what what what Leia is is good in is it's incredible fast and and can run on on a CPU even. So, it's a very valuable thing to have. But if you look at the I' I've got these radio charts on the page where you can like look at multiple dimensions very very far down >> of of like tests and there you can see like layer is just not as smart you know it's it's >> yeah try to find it in this huge list layer 36 at the moment. So if you look at the different types of questions, maybe scroll a little bit further down even there is another radio left. The left one is is very telling on the hard questions like the red part is where it kind of yeah what it achieved and Jeff is just so much smarter in all these different types of questions.

90:10 Yeah. So like layer is great if you know that your tasks are not so complicated, but if you, as I've said, you need to fulfill all of the three accesses. You got to be int intelligent as well to be actually a chef class competitor. >> Yeah. So >> yeah, on on >> you have to Yeah, you have to benchmark it. So as long as I see that it gets over like 99.5, it's pretty good for that use like comparing that JSON or doing some computer use. But it's definitely not a general list. Like this is not anywhere there like to play games and stuff.

90:45 >> So folks, open source stuff will catch up. Already is catching up. It's incredible. if you want to follow along with this, go to benchmarkheaven.com. Floren, thank you so much for joining us. So Floren, thank you so much for coming. >> >> thank you. Have a good day. >> Yeah. Thanks, man. all right, folks. Let's talk about voice. Let's talk about voice. Google says it now has the best texttospech model in the world and you can clone your voice. Yes, Google, the big company, will let you clone your voice from 30 seconds of audio, which we should do now live on stage. Two new speech models, Gemini 3.8 Flash TTS and Flash Light TTS. Google says they're number one and number two on Humeaiisai's quality index with Flash at the top of the Hume voice design benchmark as well. So you can design voices or clone them. Flash is for voice design, character acting, and long narration. Flash light is for high volume dubbing and voice agents. And logul kinpatrick, our friend of the pod, says both are cheaper than the old 3.1 flashts. And you can do two speakers in one request. And you can drop tags like laughs and hm and listener and sounds.

91:53 The voice replication is the part people will pick up because it is a huge company that allows you to clone your voice. 30 seconds of audio only adults. they can they don't allow to to to change kids voices. and the voice owner has to record the consent and every clip gets a synth ID watermark which I think Google is very incredibly on leading and clone voices get a C2PA credential. So Google clearly thinking about the deep fake problem. Voice remixing is coming soon and somebody on our bard in Reddit already built a live AI podcast where listeners talk to the hosts and the top reply was notebookm had this a year ago. Will we have an AI assistant here on the show with us with using Gemini TTS or will it be Muse? I don't know. But we definitely need to show you and hear this. Okay, so let's listen to the meditation guide.

92:47 This is the design your voice thing. Okay, the naming is meditation guide. A soft airy female voice in her 40s. Lower middle pitch, very slow, unhurried pace with gentle pauses for breath. you can hit improve my prompt. It will generate it. But let's generate this voice. It's three voice options. Let's hear for the first one. >> And so in this moment, just breathe and find that space within. It's always here and you don't have to try.

93:23 >> This is great because of the breathiness of this. I could hear the especially in headphones you can hear this. but yeah, this is not our energy right now. So I want to I want to like bring up the late night DJ >> ready to make something amazing >> and the mad scientist >> got a project in mind. >> So folks, I I had my voice in there but I can't find it. So we will definitely talk with the AI studio folks. I don't want to clone this live on air because obviously the consent thing I need to record the consent in there and I don't want my consent being recorded although I think that that's going to be faked. but the the voice replication is a thing that we told you about what two years ago that nothing is going to happen in that like deep fakes the world etc and then the big companies will follow suit. I remember there was a fear in the industry about hey oh no voice cloning is coming. Yeah, man. I don't know if you remember that episode or not, Nan. And we told you nothing is going to happen. If anything, it's for the best because people will know the disability is out there. So, they don't trust phone calls from grandmas anymore.

94:33 And nothing has happened. And I think there's a parallel there. And I think we'll finish on this. There's a parallel there between the fears that we are getting told by oh AI is going to kill us all. And the reality that we see open sourcing GPT2 did nothing. Open sourcing stable diffusion didn't break the world. Open sourcing and allowing everybody to clone their voice and use that didn't break the world. We're still here better than ever. I think there's a very great parallel between this and all the AI is going to kill us all. Nobody knows. Nobody. As I say again and I keep saying this, this is my line. Nobody could have predicted a Austrian dude that has you know done his thing and doesn't need to work anymore creating the world's next pattern that's called assistant that everybody's cloning metamuse is a very like inspired thing from openclaw Hermes as well all of the assistants now do a thing there's like a great great visionary Peter Steinberger did in his like apartment nobody in the doomer world could have predicted that this is where it's going to go because nobody can predict where we're going and so this is I we're here to tell you about all the cool stuff in a positive way. And I think voice cloning is a very positive way to finish on this. A few housekeeping notes that of the stuff that we skipped this week that we didn't get quite to talk about.

95:55 Lightning round. Quen 38 Omni Flash released Omni model with a 1 million context window. Quen 3.8 live translate has released a real-time interpreter model that we can pull up here for every language. Maybe we'll build this into Thursday. those are the three main ones. Envision and video. Quen released a 7 billion parameter model Quinn image 2.1 which is going around and doing some incredible stuff and AI art in diffusion. Flux 3 action model. Black forest lab released the flax model open source an action model 7 billion parameter world action model model which is incredible. It takes camera frames and robot state and text instructions predicts the next 32 actions and how the scene will change.

96:42 And Peter, I promised you this in the beginning of the show and I promise people there's this thing called act 486. This is I think this is the thing that we'll finish on because I think it's incredible. So like this is a pre-recorded clip. >> It's actually just a roofing torch with an air rifle cover. >> I don't get that. >> What don't you understand, Leah? >> How a flamethrower works. So, I'll draw the propane canister feeding a pressure line.

97:12 >> I want to pause here and say two things. Beside the fact that they got an only fans model to to to market this for nerds, besides that point, what's happening on our screen right now is h somebody's watching the Lex Freedman Elon Musk podcast. They they pre-recorded this. They they pre-baked this. Presses on Elon Musk and asks a question. And so then they clone his voice. I don't think he gave him permission by the way to his voice and I don't think they used the Google TDS for that. They ask the character in that video questions. The character turns to them, starts answering the questions and then also shows >> works like the model throws like a realtime visualization about flamethrower stuff.

97:59 >> Gas exits the nozzle >> and he knows about the person that talks to them. It's like what do you don't you understand Lee? got Starship and describe it for me. >> Here's Starship version 3. The stainless steel shell is built for heat, strength, and reuse with four flaps steering it through atmospheric. >> Open it up and describe its components for me, please. So, this is like a beautiful interaction moment because she pressed on Elon Musk's face and then you can see him nodding while the thing generates the thing and he's fading in and out. I don't know if you guys noticed this.

98:37 He's fading in and out cuz he's in loading state. I find that this interaction is incredible. I want to play this again. >> Steering it through atmospheric. >> Open it up and describe its components for me, please. >> Like look at this. He's like acknowledging and he's fading in and out like a person in loading state. This is crazy. >> He just looks so sad. It looks like he's been defeated, you know. >> But look at this. He is now the the video model generates the starship model that he's talking about and he is like rotating. Folks, this is the dream. This is the Game of Thrones and you're editing your own ending dream that we want. This is the demo of that. This is a demo of talking to any type of like video and making it your own. this is I haven't gotten access to this act 486.

99:24 Let's see what else they have in their demo. oh, she she goes she goes crazy in here. Look. Okay, let's finish on this because this is crazy. Okay. >> Makes those two futures feel tangible. And one future is we're out there among the stars and things we read about and see in science fiction movies. So for folks who are listening, she now asked the video to put both Joe Rogan and Ela Musk in SpaceX suits on Mars.

99:49 >> Are true. >> What would life on Mars look like? >> They're both not. >> All right, let's see how daily life works. Life would happen mostly inside these habitats. >> How does that work? Do people get upset at you if you do certain things? >> don't smoke that. You do not want to smoke that. yeah, going for the viral moment. There's a video there's a famous thing where I must smoke some weed on Joe Rogan and she's like, "Oh, no, don't smoke this." And the result is really hilarious.

100:26 >> Okay, I won't. >> You're welcome. >> I I think with no reactions, folks, this is the world we're in. You can now talk to videos and have them just do whatever you want. If you enjoyed the show, first of all, thank you so much for our live producer assistant that did a lot of work, a lot of lifts, so all the notifications and stuff. second of all, huge thanks for guest improto guest Florian S for jumping on with Benchmark Heaven. thank you Niston, thank you Peter, thank you Yam, thank you Wolfam who joined us and from and Mazer Panahi from AI engineer at Paris.

101:04 And folks, if you missed any part of the show, you can re-watch the whole show as it was live on Thursday. Or you can subscribe to a newsletter or our YouTube, which if you aren't subscribed to our YouTube, we're doing a lot of work. It will really help us if you hit that sub button. It's completely free for you and cost you nothing. And you can see notifications when we post new clips. And next week show is going to be live from Osone's house, a little different time, so we're going to go 11 a.m. Pacific. So, a little like three-ish hours later than usual and go for around two hours. First hour is going to be all about news. The second hour is going to be all about core stuff. It was a pleasure of mine. What a crazy week to have a show that talks about AI, folks. Just an incredible crazy week. and we'll see you here next