#52 Webflow Bloodbath, Opus 4.8, Codex Sites, Microsoft Models, RTX Spark

Are we live? Can you hear me? Okay. Are we live? [gasps] Let me just check the stream. I'm using a new microphone today which is a bit sensitive. If I can just whisper a little bit. but hopefully a bit more suitable for a podcast. And I'm just waffling on to check the levels of it all. So bear with me a sec whilst I see how it sounds. typing isn't very good, is it? I think that sounds pretty good. I think that sounds pretty good. I will try to not type so much. We'll have to figure that one out, won't we? I'm just listening through. Yeah. Right. We are in. I'm happy with that. Happy with the audio. and we're good to go. Uh Kabaza is not with us today just because he is in Italy again. So, well, not again, but he was he's away again. So, you're stuck with me solo, but we've got a really good I'm I'm I'm pleased with the lineup that we have today. Pleased with the lineup. Um going to go through something that Anthropic just dropped yesterday. I haven't prepared it so much, but we're just going to go through it a bit more, you know, relaxed and all the rest of it, but it's pretty cool. But what news do I bear? Well, obviously I have this microphone here. I'm kind of not happy with the the volume of it. I'm trying to figure it out. And test test. One, two. No, that's not it, is it? Whatever. We'll figure that out. So, I've got my new microphone. Very excited about that. I have a surprise. I have bought a new computer, and it's it's a beast, but it's not arrived. It was supposed to arrive. In fact, actually, the delay on it was so long. It was an M M5 Max. Um, I have you can you can check my channel, my channel 0x50M5. Um, sounds interesting for the video input. What do you mean by that? Um, yeah, M5 Max, uh, 120 gig of RAM. And originally it's it's backdated until like the 17th of June, end of the month. So, I just bought it because, you know, whatever. But um I also checked the Mac Studio and they only sell the 96 gig of RAM and to be honest like the Mac Studio is just not um not for me. Like I need a laptop like I have to combine like I can't just I'm not at that level where I can just buy a bunch of computers just for YouTube. I need to actually do work as well. So anyway, I just thought it was really interesting that you uh you couldn't buy any other RAM configuration option than than 96, which is which is bizarre. Anyway, I was surprised to see that my laptop shot from the 17th of June to the 4th of June, which is oh, sorry, 5th of June, which is today, but then today arrived and it actually is going to get pushed to Monday. Uh so looking forward to that. Going to do some local model stuff and whatever. And yeah, looking forward to it. You can actually put video in the chat and they can read it. Oh, because it's multimodal. Yes. Yeah. Yeah. I didn't know you could put video in it, but yeah, you can you can definitely upload uh images to it and uh yeah, but there's a video on my channel because they actually sponsor me. So, there's a there's a sponsored video where I just kind of like I add it to open claw and stuff like that. I was coding with it yesterday as well. Um which which went okay. They are built like a little app using music because they they've got a a music generation endpoint. Let me just show you actually uh Mini Minax. Let's go to uh Mini Max docs. We go me share this [snorts] share this tab and go here. So yeah, you can um generate video. That's not what you're you're on about here, Void. Um, but you can generate video, you can generate speech, you can generate images, and you can generate music. And I actually wanted to generate some music for the podcast and for my own channel and stuff like that. But there was a really weird timeout error that I was uh getting that I just couldn't generate any music. and Crawford inspected it and I had um Miniax inspect it and they both concluded that basically it's it's an endpoint thing like the endpoint is timing out after 60 seconds because we're not getting a response and I've contacted Miniax. Let's see how we get on there. Um but all that to be said is that I used Pi um and and Miniax have their own Miniax code as well actually. They have their own coding thing. Uh, M3 for coding tools. Um, yeah. So, I'm guessing you just run Miniax or something. It's it's it's somewhere in here, which is really cool. I didn't know, but I was using it in Pi. Um, because it's just such a barebones harness. And I was just I built this little TUI little application that generates music from a prompt in in uh Miaax. And yeah, unfortunately I can't release it or do anything with it because there's a timeout issue on the endpoint web flow blood bath. Yes, indeed. I'm surprised you haven't heard this to be honest. It's it's crazy uh what they've what they've done. But we'll get into that. It's it's first up, but we'll go through an anthropic thing as well. uh for for no X if you want to debug a bug and I want to record a video and then show to it what's happening I can actually do it with other model yeah yeah yeah yeah multimodal so you can give it a video like Gemini can do that and stuff like that um yeah I understand what you're saying you can you can give it a video and I don't know tell it to make an well they gave examples in the past Gemini gave examples in the past where they asked to generate an app from a video or something so Yeah, really cool. Um I I'd hate to think how much tokens it consumes, but Miniax is cheap. It's pretty good. Uh but yeah, that sort of thing multi um Gemini has been doing as well. So yeah. Um what was I going to show? Oh, that was it. I was just going to show you that. But other than that, what are we going to go over today? Let us know if you hear anything, by the way, because so much happens on Friday. and then I I go, you know, we finished the pod and then we've missed a bunch of stuff. There's been a bunch of stuff added, so let us know. Um, but get my notes up here. We have got a slightly short episode, but I think a pretty good episode today. So, first up, we'll go through the Web Flow blood bath. That's just insane. Um, [clears throat] I can't quite believe they approached the their recent layoffs in such a way. It's crazy. So, we'll get that over sooner um rather than later. Anthropic shipping Opus 4.8. Again, something that got released while we're on the pod last week, so we're covering it this week. Uh I'm sure you've all been using it. Great new model. However, I've got some questions about the benchmarks a little bit. So, we'll get into that. And the new Deep Suite benchmark just to, you know, bring that full circle. Uh codeex sites, which Web Flow are playing a part in. I'm not quite sure how, but there will be a web Web Flow mention in inside of that. uh looks like a really cool tool and I do wonder where it's going to go. So yeah, codeex sites Microsoft ships mi code one flash their own coding model basically really interesting that they've brought I think they released about 18 models or something crazy like this but I've just been looking into the code one and uh yeah we'll do a little bit more digging when it comes around to doing that. Nvidia drops the RTX Spar a one petlop Windows PC built for personal agents. So, that's going to be a pretty cool one. Very tempting. Very, very tempting, but I don't think it's going to be cheap. Um, and then we will jump straight into this article that Anthropic shared, when AI builds itself. So, I've quickly gone through this. It is a mega article to be fair, but it's kind of what it says on the tin. Let's kind of go through it and we can uh we can dig a bit more deep. So loving all these little graphics and stuff like that that they do. I uh yeah, this isn't procedurally generated. It's the same animation over and over again, but uh still still pretty cool. So basically this article is is as it says on the tin. It's about recursive self-proven Oh, sorry. I'm reading my note there, but I'm not giving you the lowown. It's about recursive self-improving agents and the inevitability of it and what we can do about it or what we need to think about. And they propose some really drastic uh solutions to it in here which we'll get into but ultimately it's breaking down the states of Claude and how they got to this point and their visions for the future. So they kick off most for most of AI history, humans drove every step in its development cycle. But at anthropic, we are delegating a growing share of AI development to AI systems themselves, which is speeding up their work. Duh. Taken far enough and even enough compute, the trend points to an AI system capable of fully fully autonomously designing and developing its own successor. This is called recursive self-improvement. We are not yet there. We are not there yet. and recursive self-improvement is not inevitable, but it could come sooner than most institutions are prepared for. So, that's kind of like the lay of the land. By the way, again, I'm using a new microphone right now, so if you could give me the heads up on how it sounds, if it peing or whatever, then I can sort that out. Um, AI that can build itself would be a major development in the history of technology. One that could bring enormous good for the world in science, healthcare, and beyond, of course. But for recursive self-improvement also might increase the risks of human losing control of AI systems. That's that was a Terminator 2 reference. So let's look at this little breakdown which again it's a beautiful little graphic here. I wonder how how do they this how do they do this? Is it Lotty? Is it Rive do you reckon? Oh it's genuine SVG animation. Look that's that's cool man. They've they've got literally they've built this whole thing as an SVG which in of itself isn't too um impressive but yeah they're building it out. Um so first of all 2023 to 2025 where most people started their AI journey including myself when chat GBT came to the came to the forefront and we are we discovered that people can interact with their computer through an Asian. The the ideal way to communicate funny enough is this chatbot. People used early chat bots to help with parts of the process like generating short code snippets, copying the output into text editors. I remember those days. They were uh they were tedious, but also I think a really nice learning workflow if you're learning about AI and stuff like that. Uh it's a great way to really have a bit of um learning basically. Hang on, let me just get water. Uh two seconds. And then we get into coding agents which again person talking to their computer. Um in fact actually I think we are this is misaligned at the moment. I think this is this is co this is chat bots. This is coding agent. So yeah through a chatbot we are communicating with our computer. [snorts] Then we get agents coding agents. As the agents become more capable they were able to write and edit code on their own. Sometime sometimes entire files and this is talking about uh 2025 2026. And that's the stuff we all know and love at the at the moment. now and with an update that we'll go through uh that they announced last week, autonomous agents. Agents can now run code themselves and dedicate hours of work to other agents. So um person talking to their computer via a chatbot, spinning up an agent who then subsequently can spin up other workers and agents and things like that and actually handle that delegation completely within here. So you're starting to get a sniff of everything now. Closing the loop. In future, agents could become more capable of enough uh more cap become capable enough to build and train models themselves. And this is again there's no human in this loop. It's literally just this endless cycle. So you can kind of see how we got there with all of this. So it makes sense. Like it doesn't you know, as I said at the beginning, it's not necessarily a guarantee. However, you can see how they kind of got to this point. [snorts] Um, and then they kind of go into again the history of of uh anthropic about how um the sort of the stuff that they were doing like uh where is it here? Um uh about four minutes to complete. So, Opus 3 Opus 3 could complete software tasks that take humans about four minutes to complete. Think about where we are now. Think about where we are now. Because yeah, a year later, Claude Sonet 3.7 managed tasks that took about an hour and a half. A year after that, Claude Opus 4.6 managed 12hour tasks. If this trend holds, this task could uh this trend holds tasks that take a skilled person days could come into range this year. And to be honest, I think they're already we're already there because remember uh hands up any [clears throat] how many devs are in the audience, you know, give us a shout because when you used to plan out tasks, it would give you a um it would estimate how long it would take and it would take days and weeks, week one and week two and do this and do that. They've since removed that from the um pro from the feedback but it used to like yeah it would be something you would get done in about an hour but it was estimating weeks and weeks and weeks of work in the response. Um I for sure am getting work done in maybe not minutes but certainly a couple of hours like I'm getting days worth of work done potentially even weeks worth of work done. So, I'm not quite sure how they're measuring this, but either way, like it's insane. It's insane what you can what we can do nowadays. Um, and uh, building a frontier model takes two broad categories of work. There is engineering, which is writing the code, standing up the architecture and overseeing the model training, and then there is research, deciding what experiments to run, interrupting what what comes back and figure out ideas to try. So there's like still that kind of human in the loop thing. Um this they go on to talk about how much code is actually written and you can see here like about you know average kind of stayed the same for the first few years and then you've got this use kicker um the second quarter of 2025. What was that? Was that Opus 4.5 maybe? [snorts] Um and then Opus 4.7 came out probably around here I think. and then it's just kind of gone crazy. These are um mythos which we're starting to see sniffs of but point is more and more code is written. They do they go this is an interesting point they lines of code is an imperfect measure as it measures quantity over quality. So eight times a eight times lines of code engineer a day in the second quarter 2026 is almost certainly an overstatement of the true productivity gain. Nonetheless, it indicates acceleration. Anthropic don't reward people for how many lines of code they write. Rather, team members are producing more code simply because they're using AI assistance to write more code. So again, they're sort of saying don't don't take this with a pinch of salt, but it's it's what other factors are at play that that can um that can justify or explain this other than people are being more productive. you know, so you know, and then we get into sort of uh how capable they are. Trivial problems have always been very very easy. That's kind of stayed, but as you start to get more broader, the the the success rate it massively drops, but they are uh converging at the top here being more consistently dealing with uh bigger challenges and more open-ended challenges. So anyway, the thing Oh, we've got some more stuff here. Um the evidence suggests that human role is narrowing at each step in the AI development process. Once human and AI authored code quality reach parity, humans will stop writing code entirely and shift to only reviewing it. But if they can't review code as quickly as core can generate it, human review will become the bottleneck to AI development. Dun dun dun. Um, an [snorts] area of human comparative advantage for now is research, taste, and judgment, including choosing which problems matter, which results to trust and when an approach is a dead end. So, we're basically just we're guiding the AI, you know, we're just shifting it and it's deciding uh how to accomplish things based on our broad goal, you know. Uh, so yeah, we we're getting we definitely get to this point. [snorts] This is where it gets interesting in my opinion. What should we do? Actually, they do go on to say they've got three potential outcomes which I'll highlight. Like so there's three possible futures here. The trend stalls, but today's AI capabilities are widely diffused. This article features many exponential trajectories, but these trajectories may actually turn out to be Scurves. We may be approaching the bend in the S-curve where it returns to scale diminish and the line straightens and then flattens. Then you've got the AI labs continue to see compounding efficiency gains, so getting better and better and better. [snorts] Um, AI systems themselves become more capable of full recursive self-improvement and begin building their own successes. Again, self-explanatory. So, what should we do? If it were possible, this is where it gets really interesting. If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that that would be a likely a good thing. Slowing down, which I was thinking about this today actually to be honest, slowing down. But if a slowdown simply lets the least cautious actors catch up technologically, it could leave everyone else less safe. So you've just got this, what do you call it when you when you're at the stalemate? You've got this stalemate situation where it's like, you know, well, we've got to keep getting better because they'll overtake us, but then we're also expediting our our own downfall of humanity. Without a global coordination mechanism, companies and governments will have to make difficult decisions about safety while under the competitive and geopolitical geopolitical treasure uh pressures. Um the Anthropic Institute will conduct research in collaboration with many others and take actions to help build the systems that that a credible slowdown or pause would require. So they're actively looking into this potentially pausing of things, but they just need to make sure that everyone is is capable of doing it or everyone is going to follow it. I you know, you've got some treaties out there now. I think they've got they mention it actually um uh the intermediate range nuclear forces treaty uh we you know we have these treaties and stuff like that that people are adhering to however you know it's so much is on the line when it comes to AI like literally people they're doing everything like look at anthropics history where they stole a bunch of books and they're just like look we'll deal with the consequences later on do you know what I mean like everyone it's it's crazy how much they how desperate they are. [snorts] In the coming months, we'll organize conversations where policy makers, researchers, civil society, and other AI companies can help answer some of the questions this piece raises, especially around the full recursive self-improvement. How and how to create better options for coordination and deliberation. We'll publish what comes out of it. The window to investigate the questions together is here, and people outside AI companies should be involved in this deliberation. [snorts] So, kind of crazy. Which reminds me, Mini Max uh M3 reminds me uh the M3 does have this uh I don't know why it reminded me of this because it's kind of different, but they've introduced this new sparse attention thing which is and they do mention kind of different paradigms and stuff like that. Sorry, I'm not showing the right tab. this thing um uh anthropic do mention different paradigms like changing how we develop um or how we do AI and and get more efficient and stuff like that. This is a really interesting aspect to the new M3 model which again I don't I don't see anyone else doing or or whatever. So yeah, that was the article. Um really interesting to see and it's also interesting here I thought Claude would have developed it or written it but there's a big old team of people doing it like unless they are you know um uh when they say co-authored do you know what I mean they're refraining from saying opus wrote it but yeah crazy crazy crazy and I like that they do publish this stuff I really do it seem like I don't see open AAI doing this sort of thing so hats off to Anthropic for continuing to you know think about this stuff um alongside of their deliverables and stuff like that. So Sachin Webflow bloodbath please elaborate. Um, [snorts] so I don't know if you saw this, but Linda Tom, the CEO of uh, Web Flow, released this article. Earlier today, I shared this news with Web Flow employees. I'm sharing a version of that message here because this is an important moment for Web Flow and our customers and our community. Let me make that bigger for you. Today we made the difficult decision to restructure Web Flow's team and operating model. As a result, many of our Web Flow teammates are leaving the company today. She goes on to say here about how obviously AI is changing the way we build things and we, you know, we've got to really consider the product and what it enables people to do blah blah blah blah blah basically. And then we get to the specifics of what's um what's going down. Um well, first of all, we cannot achieve this vision through prioritization alone. We have to change where we invest and change the size and shape of the team. That's why I'm making this decision today to lay off a bunch of people. They get 16 weeks of severance with an additional week of base pay um for every completed year of service. Six months of COBRA coverage. I don't know what that is. I'm assuming it's a US only thing. Um, continue benefits and career resources where legally permissible and the ability to keep your laptops. Sounding okay for now, right? Sounding okay for now. Bear with me. I guess this is more specific thing. We're moving to smaller, more focused teams that can move faster, drive impact of our customers. We're move we're moving to a simpler structure where leaders who can stay closer to the work teams will operate [clears throat] more with more speed and direct ownership while reducing the collaboration tax that comes with too many handoff and appro approvals. Doesn't mean you're going to avoid that just as a side note because people will come and go. Um and we're going to lead by example. Our own marketing team will become the first proof points for what a modern marketing organization can look like when it is paired with AI agents in its day-to-day work. Basically, AI is just replacing pretty much everyone. I don't know the numbers of people that are being laid off. Um, and then she finishes by saying, "We flow is financially strong and we'll continue investing in the teams, products, capabilities, and internal international hubs that support where Web Flow is going. The decision about building the company around strategy, customers, and ways of working will define Web Flow's next chapter. So, as I say, seems okay. You're like, well, obviously people are using their job is not, you know, great." and people losing their jobs and and this is a clear sign of AI replacing people in their work is not good but the approach seems uh fair. In a now deleted post on LinkedIn, this was sent out by Mark W on LinkedIn tagging her on on LinkedIn. I'm knocked out of my Web Flow since 7 a.m. this morning. We flow laptop since 7:00 a.m. this morning. Rumor has it that we've been laid off. Rumor has it, but I don't have any email or any message to confirm anything. And I actually said I don't think Linda would lay people off again without letting them know first, especially if they're on a closed work permit in a foreign country and would mean relocating their whole family. I'm certain she would have the dignity to let folks know in a better way given the de debacle debacle last time. Tell me it's not true. Obviously there's a lot of tongue-in-cheek humor here. My manager Christopher Harup, do you have any news? Do I have a job? So what happened was is Linda um wrote that wrote that article seemingly without letting anyone know about it. And she does say in the in the thing right at the top that this is a edited version. I'm showing a version of that message here. I shared this news with the Web Flow employees. What Web Flow employees did she s um share this with? Because clearly Mark did not get the memo. It's now deleted which I I don't know about you gives me a bit of a Was he slapped on the wrist? Did he check his junk and it's in his junk folder or something like this and it he's been told to take it down cuz he was off the mark there. That could be inevitability here. But it was the it was also the um the the fact that uh the foreign work permits. So, how this works is, and you guys probably know more about it than me, but from what I understand, if you are hired, if you have to relocate, whether it's across state or even um in another country, I'm I'm pretty sure this in from another country, if you come for that company, you have to you're you're tied to that company. And then once your once your contract ends or whatever with that company, you have to go home. Obviously, I think it's a bit more intricate than all of that. I think that there is an element of um they could help you hand you off or there's a certain amount of years you can do and then that wipes it off or whatever. But Mark seems to think that whether he's speaking for himself or or whatever that um uh that that he's in he's in he's up shitz creek basically with it. So this is terrible form from Web Flow. Absolutely terrible form. um really sad to obviously see that people are losing their jobs to AI and stuff, but we kind of knew this was coming, but I don't know. I I I think that um I think this is this they've not had a good track record with sort of layoffs in the past, which again what Mark alluded to. So yeah, that's the that's the bloodbath. Um it shows how indispensable you are as an employee. You know, you they make you feel like a family. They make you feel like this and it's like you know there is this middle point where we don't have to feel like a family but there is a transactional relationship between us and we you make us feel good about working here subsequently we do good work for you. Um, yeah, there's some definitely some um lines overstepped here. And I think and and here's the thing as well. Linda has not responded. Last time I checked, unless anyone has any, let us know in the comments. Like, unless anyone knows anything different, Linda just has not responded. And again, this is really really bad form, you know, and it does kind of I I've witnessed this like one of my first jobs, my boss just kind of like I was like made redundant and I just couldn't get through to my boss. It's like he I don't think I even spoke to him again after that. He was really just like turned his back on me, gave me the cold shoulder and it's like, you know, wow, that's it's a surprising surprising behavior from from Linda from from my boss in that respect. So yeah, hope I didn't let you down there, Sachin, but uh the blood bath is is real. Let's talk about Opus 4.8. If you've got any questions about that, let us know. Or if you have any updates, let us know about that. So, this is a bit of an old one because we're a week behind, but Opus 4.8 8 just dropped, which is a significant improvement over 4.7. It introduces a brand new feature called workflows, I think, which people don't like because it's you you type workflow and it's going to trigger off a workflow, which can cost a [ __ ] ton of money. [snorts] But 4.8 is the next big release from Claude and I am absolutely 100% here for it. I'm a big Claude user. We're upgrading Claude Opus to a new version, Claude Opus 4.8. It builds on Opus 4.7 with improvements across benchmarks and is more effective collaborator. It's available today for the same price which we love. Cloud code and cloud code has a new dynamic workflows feature allow allows it to tackle very large scale problems and fast and fast mode for Opus 4.8 eight where the model can deliver work at two and a half speed um is now three times cheaper than it was for previous models. I think that's fast mode is three times cheaper than previous models. I think it's it's still double the price but yeah pretty pretty cool. So looking at the benchmarks which of course we have to do. Um let us know what your I don't know about significant let us know what your experience has been. I uh I definitely have felt a bit um more success rate uh with 4.8, but we will get into some doubts that I have and I'll be more than happy to talk about those in in a bit. So, here are the benchmarks. 69.2, which is uh better than uh Opus here, massively better than the GPT 5.5, humongously better than 3.1 Pro. Um, terminal coding here. It's never been great at terminal. GPT 5.5 is the leader there. Uh, humanity's last exam, 49 versus 46. Uh, 57 versus 54. Uh, beating out uh, GPT5.5. Uh, a genetic computer use, which I would like to know. I guess I suppose a gentle computer use is is pretty um different but yeah uh a point improvement there but always it was still uh Opus 4.7 was better than GP 5.5 knowledge work blah blah blah all that kind of stuff. I'm I've got my keen eye on this one. I like to do my finances through Opus. So yeah, pretty cool from the benchmark perspective. Early testers have found that for Opus 4.8 to be more reliable and sharper in its judgment when performing agentic tasks. One of the most prominent improvements in Opus 4.8 is in it's in it is in oh my god I can't speak is its honesty Jesus is in its honesty I could have said that anyway uh we train all our models to be honest for instance to avoid making claims that they can't support now what I found and going to your point here about um being um more significant I was finding and I tweeted about this that it would be like yeah you know I challenge it like 4.7 wouldn't do something or it make a mistake or something like that and maybe maybe it's made and not fix the mistake after I've tried correcting it and then it would respond and be like okay well let me do some research and rather than trying to guess stuff to that effect and it's like well hang on thinking mode is on high here which by the way is the new default they've made high the new default which we'll talk about pricing two things we're going to talk about benchmarks and pricing don't let me forget um they made high the default and It's like, well, why are you even guessing in the first place when thinking is on? Why are you why are you um guessing in the first place? So, anyway, 4.8 is more likely to flag uncertainties about his work and less likely to make unsupported claims. So, that's I've not seen that since since 4.7. So, that's good. The assessment also showed that 4.8 8 to have rates of misalignment behavior such as deception or cooperation with misuse that are substantially lower in than 4.7 and similar to our best aligned model claude mythos preview. So this is the misalignment. Obviously higher is worse. So we've got 4.8 basically little bit higher than mythos. So that's pretty good. So yeah, I'm I'm pretty stoked about it. Um, dynamic workflows. As I said, the new feature in research preview allowed to claude to take on bigger tasks in claw code. Claude can plan the work and then run hundreds of parallel sub aent in a single session. This completely eats your cost. I've I've heard people like destroying their their limits within minutes on this one. And you know, it it's crazy, but for some people that's a that's a a good thing to do. I don't see myself using it mostly because of the cost and because I'm a traditional like developer. I like to be probably slightly more hands-on than most people when it comes to writing code. Effort control in uh co-work and claw which is quite nice. Why not? And the messages AI now accepts system entries inside the messages array. So developers can update clause instructions mid task without breaking the prompt cache or routting the update. So I think that's going to change the um uh make it cheaper basically. So um and there we go. Defaults to high high effort. Now I would uh finish off here. Croous is available everywhere today. Pricing for regular use is unchanged from 4.7 5 in 25 out. So yeah, two aspects here that I'm uh that I'm interested on. The I don't know whether you about you but like it was a few I mean I know it's hard to differentiate between the compute restrictions that they had um and uh yeah just not having enough compute and the and the limitations the the rate limitations and stuff like that. They were like the the default was Sonic. I think it was was it default was like Opus and then it went down to default being Sonic because Sonic overtook Opus for a bit. Sonic was on like 4.5 and then Son um sorry, Sonic was on 4.5 whereas Opus was on four or something to that effect. Correct me if I'm wrong, but the point is Sonnet then became the default. And then they've obviously got a lot more compute now, which muddies the water in this kind of little um opinion that I have here. But the the the default now is Opus. It's it's like, you know, you select the model inside of of Claude Code or whatever and it's like the recommended default model for, you know, your your everyday tasks, everyday tasks, because that's that I thought that's what Sonic was good for, everyday tasks. But they've increased that and they've bumped up the default thinking mode to high. Um, in fact, oh, we'll we'll get into the the benchmarks in a second. Um, and and you've got the new XI model. So, is this a marketing effort to push you to spend more, jump up from that pro to the max and then the max x5 to the max x x10 or 10x or whatever it's called? Is this a marketing ploy or is this a byproduct of compute because at the same time they've lowered they've raised lifted the limits and all that sort of lot. So, it's just I don't know. I don't know whether whether um we could probably get away with Sonnet is is kind of what I'm saying and we can reserve a lot of um we can go a lot further with our usage because it's in their best interest to make you use to spend more with them to use them more to to be more ingrained into their ecosystem. I'm an absolute sucker for it, don't get me wrong. Even though I know it's like that, I'm still like, you know, oh man, should I be using Opus on high? And I've set my default as Opus on high and I am using Opus way more. I still drop down to like Haiku when I'm doing basic like get commits and like asking it questions and things like this. Like I'll still drop down to Sonic. HiQ. Um but I'm like Sonic just feels a little bit redundant. And they've actually there's a preview of um what do they call it? Uh [sighs] there's basically another model that they've they've uh that's leaked uh that's better than Mythos. So I don't know. It's it's all getting a bit confusing. saying that all of these amazing benchmarks are you know that Claudia claiming um let's just remind you of those the you know smashing them and a significant improvement of a Gentic benchmark of coding over GPT 5.5 and [snorts] even though I'm a Opus uh Claude fanboy this doesn't smell right to me like GPT 5.5 is a brilliant coding Like I used it for a while when they gave they gave out crazy ton usage and this and that um like for half the I had a free you know account for a bit but it still doesn't smell right. It's still that significantly better. I don't know. And people really do swear by GD5 and you can't you can't blame them for liking that. Enter Deep Suite, which really makes sense. Like fight the corner for opensource benchmark tools because that's where you get into benchmarking territory where people are training their models or tweaking their models to actually score really well on these benchmarks that they know what they're going to be tested on. They know the questions or however it works. I'm not quite sure how it works, but they're able to benchmark against them and score really well. Deep Suite is a closed source benchmark for obviously software engineering. You don't know how it's going to do it. You don't know what the questions are. You don't anything about it and they will they will run the tests and then you it's easier to read I think here. But ultimately GPT 5.5 X high, you know, this is scoring like 70 and 62 and it's contradictory to this idea that um max and X high are pretty much scoring the same. So I wouldn't necessarily I suppose the average cost goes down on X high which is quite oh no sorry max is max max. So why you would spend you know based on their scoring like getting the same result as X high is arguable. However, for me, like even though I don't want to believe this, cuz again, I'm a I'm a Opus Claude fanboy, this starts to make sense as to why this would be more accurate than open-source benchmarking tools. Um, and from actual use because as well, you know, I'm reminded here about these open-source tools, um, like Miniax 3, like Kimmy K 2.6, like Quen. I think there was one of them. Um, what one was it? I I Yeah, I don't know. Basically, Quen is like scoring the same as Opus 4.8 and all this sort of stuff. And it's like, really, dude? Really? Like, I find that very hard to believe. So, that's where you start to call [ __ ] on these benchmarks, these open source benchmarks. Deep Sweet by data curve seems to be the step in the right direction. Um, as much as I don't want it to be true that, you know, Opus is is falling short of these results against GBT 5.5. [snorts] However, it's it's there and it is tempting me to try and use GBT 5.5. Brucey bonus discussion point as well cuz and again, I forget where I see it. So it's another trust me bro story from myself but the score that people are getting out of clawed code versus using the API basically called code is nerfing opus's potential. Um, it might even be, let me even find it on deep suite. Actually, it might even be on deep suite, but uh let's try basically uh they managed um GBT 5.5 was getting the same scores in codecs versus through API. Like there's no difference there. whereas Opus or or Sonet or whatever anthropics models were scoring worse inside of Claude code than they were um is this it here we go this is it so Claude Opus 4.7 in the mini suite agent which is a barebones harness really really bare bones hardness it's scoring 50 it's got a 50% pass rate and it's using more tokens but 50% pass rate whereas inside of clawed code it's a 40% % pass rate using less tokens mind. Um but you know it's it's notable [snorts] whereas you've got 40% for GT 5.5 you got 40% 40% it's using less tokens in the minu agent but that's kind of to be expected in some ways just because it's a bare bones. This is more this is actually quite surprising because claw code is notorious for handing so much context even before you've typed your pro prompt. So yeah, 16 to 25. Not too much of a not too much of a difference there. And then minuite. Uh look at the Gemini CLI. Jesus Christ. And this is an open source CLI as well. A lot of the um opensource models like Miniax like Quent, they're all using the Gemini CL. I think it's Gemini CLI is what they're using to to build their two UIs and stuff like that. But yeah, Craw code is nerfing Opus. So, I get like and it's so crazy cuz now um I'm building more things like I'm building more of my own tools and stuff like that. And literally days before they announced this cut off point, we've got a video on it. I was like, hm, I should be using Claude-P and I should be using Claude agents SDK to build some of these tools so I could leverage my Max subscription, my pro subscription, whatever it is I had. And that was the day that they announced that they're going to be introducing this programmatic use credits, which by the way, I haven't received. I don't know how to get that. Let me know if you've managed to get that. But um this programmatic usage thing and it I wanted to start using uh Claude with um outside of claw code you know so it's not um bloating the context of all of the skills all the MCPS and this and that. I wanted to use something a bit more lightweight. So it's a bit of a shame. It's a bit of a shame because we're not getting the most out of Anthropic's own models using their own tools. Yet, their Max subscriptions bind you to using their own tools called code. So, bit of a shame there, but um all that to be said, it's definitely is tempting me to to look uh elsewhere or to try some new models. I mean, I have used uh GT 5.5. It is a great tool, but I I just I like claw code, man. I like claw code. So, yeah, there's that. There's that. Speaking of codecs, open AAI have released codeex sites, which is a pretty interesting progression that have web flow entangled up in the in the joy of it all. So more than five million people now use codeex every week. Codeex started as a tool for software development but is increasingly useful for more kinds of work. Non-developers including analysts, marketers, operators, designers, researchers, investors, and bankers make up about 20% overall of codeex users and are growing more than three times as fast as developers. The normies are getting involved, man. The normies are getting involved. Today [snorts] we're introducing new ways to do more work with Codex plugins which is we won't go over that but it's plugins ecosystem and a preview of the ability to create interactive sites and apps you can share with your workspace using a URL. So they're really gearing this towards the workforce, you know, the workers, bluecollar workers, uh, by it's not for you and me, which is what I thought it was, but it ain't. It's for internal URLs. Inside of OpenAI, non techchnical teams use coders to build internal apps, prepare exe executive materials, create dashboards, and turn creative briefs into work that reflects the brand and design constraints. So this is the stuff they're sort of build. Oh, here's the sort of hype video. They've got it. They've got going on. I don't want to put the music on because we'll get thing. So, I think we all do this. I think we all uh we certainly should be and especially because we covered the uh unreasonable effectiveness of HTML. We all are building apps and um dashboards to help us get our own work done in the workplace. And you might have seen that it looked like just by tagging sites, it's able to then create that research that you've done or or this, you know, whatever it is and then, you know, build a ticker, build a revenue dashboard. Um, being able to post that as a private URL to share with your team is really expediting the communication within a workforce, which is pretty cool. [snorts] Um, so yeah, that's kind of it. I think it's it's pretty self-explanatory there. Um, and that's all of that stuff. Starting with improve for businesses and enterprise customers. Codex can now create and share interactive hosted websites and apps. Sites can be shared with anyone in your workspace via URL. And there's [snorts] some examples there. We're also working and this is where it got interesting because I saw uh this from web flow. We're excited to partner with open eye sites as we help bring the next generation of web create creation to this web creation to life. This was probably maybe the same day as the uh the layoffs to be honest. So it was kind of weird seeing that and I was like what is going on? And then [snorts] yeah reading this. So I can imagine for now it's sort of like a testing ground. I don't think it's going to remain as an internal tool. I think what this probably mean is is that you'll build something and then you will um you will just just be able to publish it. It's like lovable. I think it's lovable. Even though lovable I mentioned here, they're working closely with these platforms. Is this a what's the word? you know, Codeex is becoming the what do they call it like a master app like WeChat in China? Um, are they digging their own grave by partnering with Open AI, you know, because it be it people be like, well, I prefer to just work in Open AI why should I pay for lovable if I can just do exactly the same thing in open air. So, it's going to be really interesting to see how these two things come together. Um, and what that looks like. All we've got for now is, like I say, that mention and then this tweet from Web Flow. TBC on that one, guys. But it's still it's still pretty cool. Sites are rolling out in preview for business and enterprise teams through Codeex app. Enterprise admins can enable sites in the admin settings. If you've got access to this, let us know. Let us know what's going on. if you have any integrations with web flow or loveable or whatever or it's strictly just sites at this point. Um I'm not part of an organization so I can't tell you. If you are on YouTube, shout out to you guys. Give us a like on YouTube. I really appreciate it. We are We have a new website actually. We should probably say this. We have a new website commandow.com where you can read a little bit more about the actual episodes themselves. um just a bit more of a a holding ground for everything that we talk about and obviously SEO and all that kind of nonsense. But yeah, we have a website. Let us know what you think. It was uh it was uh put together by Kabaz and myself and um yeah, we look forward to your feedback. Also, go to where is it now? If we go we edit this, if you go to insiders.comandai.com uh command show.com. We are we want to launch some um uh community. We want to we want to do something a bit more behind the scenes than what these shows do. Oh, didn't know you could have a timer on it. We want to keep that up forever and ever and ever. Um, yeah, we want to build a community, but we're not quite sure what you guys want from it or what you would want out of a community. So, we're just, if you go to that URL, it's just a form. We're not tracking you or anything like that. You can put in a fake email if you really want. Um, we just want to know what do what should we talk about in this community? What should we how can we help you? How what can we do to help you more? Is it teaching you code? Is it, you know, app discovery or something like that? I think there's a dropdown in fact let's go to it now in size command show this is when the website breaks um let us know your name would be really nice to meet some of you guys but as I say it's not really important uh what do you want to learn about AI tools and oh you can't see it AI tools and productivity building with AI AI and design creative work understanding how AI works AI for business and career AI ethics in the future and then if you don't see your topic which I've just selected there you'll be able to type in something give us a little bit more detail. So, we really appreciate that. Even if you don't want to join the community, help us help us out. Do you know what I mean? Help us out. [snorts] Um, so yeah. Yeah. Give us a shout out. Give us a like on YouTube. And let's get on with the next article. How we doing? We're an hour. [snorts] The next one is Microsoft shipping um an a coding model, which is quite cool. Also appreciate if you could let me know how the microphone is doing because I'm banging my table there which you can probably hear but there's a f a fuzz happening. I don't know whether that's just a connection. Can you hear that like fuzz that that fizzing or is it just a rumble? I appreciate cuz I keep putting my hand down. I can hear like that fizz. If if you can't hear it, then I don't mind. It's all good. Okay, let's clean up some of this stuff. Cut group. Um, so yeah, Microsoft announces a ton of new AI models. Um, a ton of them. I'm most interested in the coding model which they call M AI code one flash. Not not the nicest uh name for a model but heyi ho is here and uh I'm pretty um it's a new model. Why not get excited about it? They've got a lot of co-pilot research from there. [snorts] So and it's the build developer conference in San Francisco. Microsoft announced MAI code1 flash and it its inaugurable model for the AI coding space. So um in addition to MI code one flash, god damn it, Microsoft is introducing MI thinking one a reasoning model and is playing up the efficiency for both offering uh offerings. interesting thing here is why I feel like thinking is just you're you're talking about levels of thinking. You're not like thinking should just be on at this point. I don't think I used to have thinking off a lot actually. Um it'd be interesting to kind of go back to that workflow, but right now yeah thinking is just on by default. So why are they releasing a model that uh you know the reasoning model is medium size built for high efficiency and performance but importantly a low token cost like I don't think the this reasoning model is is available yet. Um it should have been released with reasoning in my opinion but I wanted to give this a go. I wanted to try this out. You'll have to let me know if you can because I think you have to subscribe to Copilot. you get it through co-pilot. Um, I tried using the auto mode and I said, "What model are you?" And it kept using Raptor Mini. I don't know whether Raptor Mini is actually the uh the the model itself. However, yeah, I don't know. I want to be able to see it. I want to see it in action. I I'm guessing it's a very fast model. I bet I'm guessing it's a competitor compos to Composer 2. So, um, yeah, if I ever get my hands on it, I want to give it a run for its money, see where it it pans out. I don't think it exists in any benchmarks. Let's go on, uh, artificial analysis, actually. Um, let's go intelligence and let's go flash. Uh, ring flash. No, no. No, they don't they don't have it. Um, so not a lot of information about it. Um, and as I say, I think they've released about eight models, but they're all very specific to their use case. I think, correct me if I'm wrong. And this is quite and this is quite um sorry, I wasn't sharing my screen then. Um, I was looking at artificial analysis here. Um, [snorts] the thing about like uh I want to get into this and let me know if you're stoked about it as well, but uh I want to get into tuning my own models and it was kind of inspired by composer 2.5 because that is tuned off of Kimmy K2.5. So they've refined an existing model to make it really good for coding. You think about it, right? A coding model doesn't need to know about the history of Battle of Hastings in 1066. It doesn't need to know the chemical properties of hydrogen or whatever. You got to think that these models must have all of that information in there. Um, and it's all wasted compute. It's all wasted um, uh, noise for when you're you're trying to get work done and whatever. So, it's a really Yeah, I'm I want to try and get into fine-tuning models and and seeing how that goes about. I want to do it when I get my new computer, so stay tuned for that on my channel, but obviously I'll share stuff I have here. Um, yeah, build your own models. you know, I don't think you'll ever obviously compete with Anthropic, of course, but if you can have a really fast, cheap um or even free model uh running on your local machine that does 90% of the grunt work, maybe you maybe you create a plan in Opus and then set off your local coding agent to just run for a day just doing all the work. Like, I don't know. I sort of see a future in that if at least for a fun little experiment, let's be honest. So, yeah. Um, nothing much more to say on the co-pilot um, Microsoft models. They ain't backing down from the AI race. Uh, but they could be a little bit too late because I feel like we've all got our favorites at the moment. Um, [clears throat] maybe similarly, I'm not too sure, but Nvidia just launched the RTX Spark, a one petaflopk win one pock a one petaflop Windows PC built for personal agents. So, they've released this computer effectively, or is it the chip? I don't know. But the [snorts] RTX Spark, a new be a new beginning for PCs, announced at GTC in Taipei at Computex along with Nvidia Open Shell bringing secure agents to Windows with two times inference performance on llama CP [snorts] whether it's the chip or whatever because you've got this I think that's the DGX Spark this little computer thing which is a five grand. I think it's got 128 gig of of VRAMm in it, but it's just a dedicated little PC for for um AI agents. They brought out a laptop version. I don't think this is it, but just to paint a bit of context here. RTX Sparks one paflop of AI compute and 128 gig of unified memory can meet the computing demand of ondevice agents. I will be absolutely putting my Mac to the test on what we can actually feasibly expect from ondevice agents with 128 gig of of VRAM because I I have my doubts. I'll be honest, but I'm determined to try and make something work and that could Sorry guys, I'm not sharing my screen again. That is the ball a lake of doing this all on my own. Um, yeah. So, I'm I've got my doubts and I've got my my tests I want to run to see if I can get a decent model running on on a machine or what is the bare minimum that we can actually do. [snorts] The RTX Spark brings Nvidia's 30 years of technology innovation to slim Windows laptops with all day battery life and ultraefficient desktop PCs. This to me sounds bonkers. I read this and I was like, h like Apple took the the um thin laptop route and the the thermal throttling was just insane. Like if you bought one of those super thin laptops and then put stuck an i9 in it like and I know these are i9's and I know we've we've um you know we've grown a lot um since since the Intel days when it comes to Macs but still man like there is physics at play here. So, I don't know. We'll see. Saying that I'm get my laptop on Monday. Um the what day? Tuesday. Is it five, six, seven, eight? I think it's like Tuesday or maybe even Monday the uh Apple conference is. So, we'll see. Maybe I'll be kicking myself because they tease an uh a MacBook Ultra. I don't know. Which everyone thinks is going to be AI focused, which it probably should be because they are absolutely killing it right now when it comes to AI stuff. Nvidia collaborated with Llama CPP community to enable features and optimization such as multi-token prediction, a speculative decoding technique where a smaller draft model proposes multiple comp multiple tokens at the time that the target model verifies in a single pass. So that's interesting. I thought we'd already kind of done that, but you've got like a little model, a tinier model that's kind of working slightly ahead and doing some sort of pre-processing basically. uh to help speed things up. This coupled with other optimizations such as programmatic dependent launch delivers two times performance on Quen 3.6 and 3.527B and a one sec 1.6 performance boost on Quen 3.6 and 3.5 30 oh sorry I read that kind of wrong. Quen 3.6 and 3.5 27B and Quen 3.6 and 3.5 35B. So, this is obviously a bigger model than the 27B. These updates are available via Llama CPP. So, you can you can get access to these updates right now. Uh, but of course, these laptops aren't out yet. So, yeah, they've they've shown some improvements there, which is quite cool. [snorts] So, yeah, as you can imagine, all of the benchmarks, blah, blah, blah. This is an amazing laptop. And this is the landing page. They are not, I mean, you can see all the specs here. Um, one petlop, 128 gig of unified memory, all day battery, game ready. This sounds too good to be true to be honest, but the only glimpse I could find are these these really slim metal laptops. Um, no price on them. This is the DGX Spark, I think. Then I again I'm confused. RTX Spark desktop PC. Is this is this a this is different from the DGX Spark or whatever it's called? I don't know. Um and then I went on to Twitter and then I found this N1 systems which I guess is the lower-end model starting around 1.7. Uh N1X systems reportedly starting around 2.8. So I don't know is this cheap? uh Windows on ARM CPU in Blackwell 20 cores two cores 128 gig unified memory 600 gig bandwidth um and then you've got people um but nearly 3,000 for NX1 28 2.8 you are competing with high-end gaming laptops that already have a proven x6 compatibility and discrete uh RTX GPU. So yeah, this is all we have right now. Um, pretty cool um, exciting times for technology and and AIdriven laptops. People are definitely interested in this stuff. However, yeah, is this out of reach for the normal person? How what can you feasibly get done? I know they're trying to bring down the price and you know 128 gig seems to be the cap that you can really put in a laptop but is that enough to get real work done you know or substantial work done using agents and stuff like that. So yeah RTX Spark [snorts] um that is pretty much all we have time for. However, one last thing before you go. Um I thought this So, I've been talking a lot about um AI tools for designers because I don't think we're there yet. [snorts] Um designers want, you know, that fine touch when it comes to creating the designs they want. They don't necessarily need AI to finish a design, but AI is there to do so much more for them. You know, it's there to do workflows. It's there to create variations. It's there to create drafts. It's it's there to inspire and, you know, be a conversation partner. And mostly in the stages up to the final deliverable, AI can be very, very powerful when it comes to design. um it's not going to replace designers anytime soon. I don't know whether it ever will because a designer will have something in their head that they want to create. It's it's so specific. Whereas developers are like, "Yeah, just make sure that feature works. I don't know. I don't care how you do it." Saying that, is something I've been talking a lot about. And then Julian from I think he was from um uh Memberstack released ship studio and this really really interested me. It's an IDE that you've got clawed code running. So it's just it's just your subscription of clawed code inside of the clawed code harness. If you earlier on in the show I mentioned that actually claw code harnesses slowing claw down but regardless at least it's using your max plan. You can spin up agents and build your website. You can see a preview of it here. I have actually got it running. It's just a bit of a nightmare to share my screen on the on the stream. So uh you'll have to bear with me. But what I really liked which I think they've nailed it's all in beta. Maybe I'll have to show you. Maybe I will show you uh ship studio. [clears throat] [clears throat] So, this is it. Really, really nice and clean. I have to say they've done a really good job at a first uh installation or a first release of this. You can join the Slack and obviously chat with them and recommend features, but it works with your default your your coding agent of choice. Right now you've got open code, codeex and claw code. So it's doing that. And if I open up my personal website here, you can create branches. really really smart and wellthoughtout tool. If I go to It's starting the server now. Look, you haven't even seen the good side of this yet. You haven't even I haven't even shown you that. Let's try and make some more space for this. There we go. Uh, close that. What we've got is an editor here. [clears throat] Can't really see it there. Let's try um these are dynamic. Um let's try I don't know. Let's make these bigger. You've got that fine grain control that doesn't use any AI. This isn't using any AI whatsoever. This is actually changing the code underneath. So you can use AI to build or get your website kind of up and running and then finesse it with with the with the tweak panel here and you don't have to use up any credits or anything like that. So really really nice. You can write some custom CSS here. Look as well. A few caveats and I did get a message from Julian because I've got a website that's just a static HTML website with classic CSS. This is limited to Nex.js applications that are running Tailwind. I did see something. Here it is. Works with any Nex.js or Astro project that uses Tailwind. So, it's obviously needing it's obviously got a knowledge base of Tailwind. So, it's just basically swapping out classes. However, this is a first um look at this tool. I can't I I can't get over how clean and wellfeatured this tool is. It's it it's really well thought out. There's a lot of stuff I would like to show you as well that um what was it? There was something Oh, I added a website here. If I go to my home, I added this website here. In fact, I'm going to do it right now. Um Oh, look. Plain. Actually, no, no, sorry. I imported it. [snorts] Uh local folder. If I go to this one here, open this website, just this project appears to be inside template subfolder. And that's true. Like there's no HTML CSS files. I don't know how it does it. And I know this is this is stupid. Like the point is these little touches that this app has in v1, I think it's v1, um 0.67, 67 um carries through the rest of the rest of the experience. So obviously when we go here we got clawed code we can add skills MCP servers and things like that. Like it's just a really nice app. So shout out to Julian for building such a cool app. I'm I'm very happy to to to show it on to have it on the show. Um I'm not working on anything right now. NextJS or or Tailwind Wendy at the moment. But for sure, this is this is a really nice thing. If you are a designer, let us know if this solves your problems or or whatever it is that you want to use AI because I know designers typically have been a bit, you know, they're quite anti-AI, you know, because they're probably threatened by it. I don't think they need to be threatened, but they are threatened by it. I think they're threatened by it because um uh clients are choosing to use AI over hiring a designer, which just that's just the way it is at the moment, isn't it? But yeah, the only thing I would say with this tool is um it's the research aspect. It's the canvas. It's the experimentation areas that is is not missing from this tool for sure. And if they can figure out a way to do that, fair enough. But the Figma release seemed to lean more into that experimentation and laborous tasks and stuff like that. This might just be the transition from Figma into the website and then you know how when you build a website and you actually see it in a browser, it doesn't quite look like how you want it to look. This might be like a gateway into that or or an IDE that that satisfies a developer. Oh, sorry. As satisfied as a designer to feel more comfortable in making changes without having to prompt their agent, you know, they literally just look for the element and change it, whatever. So, let us know if you're going to use this. Definitely. Um, I wonder if I've got a referral link or something because I'd like to let him know that I'm sending people. But, really cool tool. Really, really cool tool. So, that's it. That'll Let me just double check all my notes here. I think we've gone through everything. As I say, Kabaza is not here this week cuz he's he's had to make a trip to Italy. A client work stuff happens. You know how it is. So, he's gone to Italy, but I'm sure we'll be back next week. Um yeah, I think that's that's everything. Give us a like on YouTube if you're on YouTube before you go. Um, check us out next week um where we'll be back here same time week of AI news and uh AI news and design and dev. Check out insiders. Let us know if you want to start a community. We're up for that. I think that's it. Cheers for tuning in. Cheers for your comments and I guess until next time if I can find my mouse. Here we go. Keep on vibing.