chapters transcript notes
click any line to jump to that moment in the video
0:14 Hello everyone and welcome back to Agentic Thinking again. We have some more topics to talk about in the agent world. Matthias, welcome. Hello. Welcome. Yes, happy to be back here and 0:26 what's what's new in in in your AI life? we've got a lot of new things to talk about. It's a surprising I talk about this with other people, but to have an a calendar invite that's 0:37 we talked about things last week and already this week we have a number of new announcements, things that are changing, things we need to investigate and go look at. , we're 0:47 going to unpack a lot of news items or or things that are happening in this week. It's a roundup of what we're seeing. Matthias, you're playing with different models. I'm doing some different things , 0:57 but I do want to point out from last week's Friday episode. , Matthias, we did a a a show and tell on Friday and this was around working with an MCP co-pilot inside Microsoft Fabric 1:09 using GitHub co-pilot and installing the NPM package there directly. , this was helping you build a system around GitHub co-pilot talking to your MCP 1:21 server for modeling and doing actions with it directly inside the GitHub co-pilot web interface. that's just something the repo's available and the first link in our 1:32 description below here. Yeah, absolutely. and we're going to be doing more demos around that particular setup and 1:42 that particular repo. we we only scratched the surface last week, but what we've got public are very 1:52 detailed descriptions how to get everything set up if you want to replicate the same setup on your end. cuz there are quite a few things to jump through there, but 2:03 with with our read me in the in the repo, you can replicate it quite quite easily. The link to the repo is also in the chat window. I'm going to put there 2:13 and if you have any other topics, ? we're just starting to scratch the surface on what this is going to look . This is your show audience here. If you have a particular area or 2:24 want for some knowledge or some training or education that's specific either to this or something else that you're working into, please let us know down in the comments or let us know in the chat here what you're interested in looking at and learning about cuz 2:34 we're all over this space anyways and we're interested to learn about what you're what you're thinking and we'll teach you what we know in these other questions. , this is open forum. You're welcome to ask questions and as 2:45 we go through the discussion here today. All , let's go wrap around here to some news items. just generally, let's talk about a lot of things that are coming out. Two big things I think came out last 2:56 week or maybe it was a maybe 2 weeks or . One was Opus 4.7 was released. , we have this new model Opus 4.7. It was, dropped. we have this new way of 3:09 interacting with a new model. It's starting to appear in other places. Matthias, have you been playing with 4.7? What's your initial first reaction to it? I've definitely been playing with it 3:20 a fair bit. It seems to be very good indeed, but it also seems to be really pushing 3:30 on on the pricing. as in you you really go through your budget quite quickly. it's 3:41 reported or it's it's it's officially documented by Anthropic that 4.7 comes with a new tokenizer that consumes a larger number of tokens 3:52 compared to previous versions of Opus. if if you were to use use it with the exact same inputs. and I think that really shows. I'm getting 4:04 to the point where I'm wondering whether we need the very latest iterations and 4:16 versions of those models. , perfect I personally, I'd be very happy to be fair to stay on 4.6 for instance. 4:26 . cuz I've had no complaints with that, ? you get a lot of mileage out of it. far, I 4:36 haven't really seen a a a pressing reason why I would want to migrate to 4.7 necessarily 4:46 and pay the penalty pay the penalty of having having a higher bill there. 4:57 In the link here, I'm going to put here I'm having the link. I found the actual item that said updated token counting for what's new in Opus 4.7. Talks about the new tokenizer. 5:08 The suggestion here that thing is kind of comical here as . They're , we suggest you update your max underscore tokens parameters to give additional head room including compaction triggers for 5:19 Opus 4.7. , they're being very blatant around the new tokenizer consumption method. they're increasing this they're saying here it will 5:30 this new tokenizer will use roughly and this is silly 1x to roughly 1.35x as many tokens. , a third more tokens is potentially going to be used 5:41 in this new 4.7 arena, which would make sense why it becomes more expensive across the board, why you're seeing the GitHub 7.5 multiplier as 5:51 opposed to just a the Opus 4.6, which is a 3x multiplier in the GitHub premium request space, which doesn't really count tokens. It's slightly different. 6:01 I've been thinking a lot in that space and just want to want to share, some some thoughts here. Let's see what you think about that. 6:11 I it almost feels at this point that 6:21 the the the the quality those those the latest coding models can turn out really high quality code is almost too good for us to 6:31 for for it to to to be meaningful in in the sense that these models are incredibly good at producing code much that the 6:44 real challenge we have is a completely different one. The real challenge is how do we tame this whole thing? How do we orchestrate this whole thing? How do we manage 6:56 having this army the swarm of highly productive coding agents to turn into something that's ultimately 7:07 a shippable product or to something that's ultimately , working software. , personally and again, this is something I'm I'm working on and 7:18 thinking about a lot, but I personally, I don't think writing code is an issue anymore. the the the real issue we have and 7:28 moving forward and even more is what is the whole product development life cycle around that, you 7:39 know? how do we get from ideation to to shipping something and and supporting it in production? And that's where we 7:50 don't really have the real answers yet and that's where that's where we need much more innovation moving forward and that's that's where I I I think 8:02 the the real bottlenecks are. you know, writing code is solved and as I said before, ? I don't even need it. I I don't even need it Opus to go from 8:14 4.6 to 4.7, ? 4.6 already did way more than at this point I can possibly make good use of. It's It's everything 8:24 else that's where we have massive gaps still. , there we go. I'll I'll leave that as an opinion. I I think I think 8:35 your opinion your opinion's justified here and I think we're not again, Matthias, maybe you observe this as , but sometimes when you build software, you build just what works to 8:45 get it out the door because we have to deliver some value to stakeholders and get things running and operational. And I I think I totally agree with that that mindset and I think that's how it works. I believe though also you're 8:57 irresponsible unless you come back and start doing the optimizing phase. what Let's come back , I could do it this brute force way and I could spend a lot of computer, turn on a lot of VMs, but often what we do after the first 9:09 pass of did it work? Is it adding business value? We come back and say, should we optimize? This is the same story I think I have with , Power Query and M. Very good first pass 9:19 on getting some data engineering done, but then I really moving that over to notebooks because it's more efficient to run. It's less costly on my CUs. I'm going to bring up another article here and I'll put it here both 9:29 in the chat window and in our private chat here as , Matthias. this is an article up from Yahoo Finance. take that for what you will cuz it's from Yahoo and it's from finance side, 9:39 but it's talking about Uber's Anthropic AI push hits a wall. Their CTO is coming out and saying there's a budgetary struggles because they're 9:49 spending they said here for $3.4 billion on AI things at this point. , I'm not sure if the number's quoted, but 9:59 they spent 3.4 billion on research and development inside the company, and it has already exhausted its planned budgeted AI in just months in 2026. , 10:10 the article goes on to describe some things that I think is really interesting as a leader in AI or agentic spaces inside your organization. There's going to be a balance of there's this 10:21 non-deterministic cost that's coming into our organizations around how much AI do we implement, where does it go in whose hands, 10:32 and just cuz you just encourage people to use it doesn't necessarily mean it's being used the most efficient way. To your point here, ? Wow, 4.7 comes 10:42 out. Everyone runs to 4.7, starts using a lot of the 4.7 modeling, and we start seeing costs 7x what we would have been using on a chat GPT 5.4 model. 10:54 Still pretty good, but a 1x cost. we've 7x our costs in a week of time. , that's a very fast pivot to see costs grow 11:04 using AI and agentic tooling. , I think my mental model here is these large very premier models, I think are going to be good, but we're going to really 11:14 start trying to shoehorn them into planning exercises and how do we leverage them to help distribute workload to these cheaper, more efficient models, but give, , 11:25 they're still good. They just need a lot of direction and a lot of instructions to these other models. I don't want to write the instructions. I I want some higher-end reasoning model 11:35 to think through this, write out very detailed instructions around all the APIs, the tests, the things that we need to build there. The agent can handle that, and especially when we're talking the pricing model for GitHub Copilot, 11:47 ? If we give the GitHub Copilot really good instructions to build very detailed instructions because you're going to handle hand each one of these issues to a sub-agent that's going to be a much cheaper, more efficient run. 11:58 I think there's this concept, and Matias, maybe I would curious your mind on this, maybe you're already thinking this way, there's this idea of the prompt router. . 12:08 How do I bring a single prompt in, have the reasoning model think about the the structure of the request, and then be able to dole out to cheaper, 12:20 more efficient models to get the execution of things done? And then, , this is part of the agentic loop, which is once those smaller models complete their tasks, ? We have ample testing, we have 12:31 some quality assurance that we can be given there, and we can then run another higher-end reasoning model against the output and say, does this conform to our standards? Is it ? Is it good? And 12:42 then there's still a human in the loop here somewhere, but it it it's more of the agents doing a lot of the lion's share of the work and the testing and the QAing back to a human to say, this is what's happening. 12:53 This is what we're building. Let me just pause here. There's still a lot of things. Is this resonating with you? Do you see this as a trend we're going to start trying to move towards? Absolutely. And and I think doing using 13:03 AI capabilities efficiently is is one of the hardest tasks nowadays, ? It's it's incredibly easy to, for instance, 13:14 throw everything at the latest Opus model, and you're you're you're almost guaranteed you're going to get good results out of it. Yeah. But then coming back to your CFO point, ? You're 13:25 . They may have a very different [laughter] perspective on that, and and possibly rightly , because I think we alluded to it a little bit on Friday as . 13:36 Things are no longer as cheap as they used to be in terms of, , getting AI workloads to do something for you. if if anything, they're going to get 13:46 more expensive rather than anything else. , it it takes a lot of good engineering, and it takes a lot of 13:57 good engineering mindset Sure. to to use AI capabilities in an efficient way. And and and I would and if you look at 14:07 this at scale, , across huge corporations or enterprises, that's really where if you have a massive employee 14:17 multiplier in the mix there, where you can see huge differences. And I think this is, moving forward, what will distinguish successful and unsuccessful 14:28 companies. , but there's a massive learning curve here and a massive education need. , and yeah, what you said is absolutely 14:38 part of the equation. Do not throw all tasks at at the latest, most expensive frontier model, ? . use cheaper models whenever 14:50 possible, whenever suitable, whenever applicable. make sure you 15:02 don't send tiny fractured thoughts thoughts Yeah. you know, as a prompt. 15:12 maybe use a a a cheaper model to to work out what your big prompt should be that you send in Yeah. , 15:23 to to an expensive model. , Yeah. and most agentic systems nowadays don't make it very apparent to you is what the 15:33 actual cost is in terms of dollars, ? , maybe that's one of , obviously, as far as the the the providers is concerned, maybe that's intentional, but that's 15:45 the mindset we really need moving forward. Agree on that one. , I think this is going to bring us into our next conversational, . a lot of things we're mentioning is is going to be a 15:56 very natural flow of the conversation today. , Opus 4.7 design comes out, and we have Claude design, which runs on Opus 4.7. , if you're on one of these higher-end plans, 16:08 Claude design is here. , Matias, have you played with Claude design? Have you opened it up? And initial reactions that I've I've done a little quick video with my friend Armando. We've done a 16:18 little a initial investigation. In in our video, we had it build the agentic thinking podcast website for us with Claude code Opus 4.7. , 16:29 very impressive. initial reactions for me was very positive. , really liked the in in traditional ways, you have 16:39 this question and answer, you give it some prompt and says, hey, I need I have some a couple additional clarifying questions. I liked how they did this ill-based 16:49 response piece of this. the other part that I thought was really interesting that I'm as I'm looking at Claude design, there's knobs. They they give you the option to make 16:59 knobs on whatever the UI is. Light mode, dark mode, font types. You pick the knob you care about, and you can adjust the feeling, ? The knobs to me 17:10 look or the equivalent of these are the feelings of the website. Changing a font on a website can greatly impact how you perceive it and what's going on there. , what was your initial 17:20 impression? Have you played with it at all? I have not played as much as I would have loved. I've definitely opened it, , I've I've looked at it. I've I've got pretty good established workflows 17:31 including using lovable and and using Gemini and nano banana for for images and and graphical stuff. 17:41 . , I don't I don't have a pressing need to to move anywhere else, but I'm very keen to explore it. One thing I noticed when I opened Claude design, and and I 17:52 went through and and I saw the getting started tutorial showing up there, they mentioned data dashboards as as one of the tutorial items. Yes. , coming from a Power BI 18:04 background, obviously, that that really intrigued me, but I have yet to go in and and play with it. I'm I'm assuming you've done a bit more with it than I 18:14 have far. , far it's been really around website and web web building at this point. I haven't gone in to try and do data dashboards. This is another area that I I've seen new articles around there. , I'm also 18:25 seeing people saying some aggressive things very much around, , the there's a huge there's a terminal. There's a real-time data terminal that's used for stock trading and stock 18:35 investing. , I don't remember the name of it. It's it's It's a really big company. It's a Shoot, I can't I can't remember the name of it, but the article What was that? Yeah, Bloomberg terminals. Yes, correct. 18:45 the Bloomberg terminal is is someone is , oh no, Claude's going after the Bloomberg terminal with live dashboards and and graphics on page and 18:55 build whatever you want, and the data can just live update and do the man, I'm I'm telling you, Anthropic is just throwing bombs 19:06 all over the industry to really establish places and saying and throwing AI at these things to really disrupt these additional thoughts here or how we've 19:16 been traditionally going to business. this is really interesting me to see where this is going to go and how this is going to continue to shape our industry. 19:26 But I think this brings us to our next really article or point we wanted to talk about here as was, in lieu of that, did you see the announcement for Databricks and Lovable? 19:36 I didn't know. Tell me about that. , this is one that I found just recently. I it came across my radar late last week, and I was , oh my goodness. one of the challenges I think we've 19:46 that Databricks has primarily had is their AI BI dashboards is . It seems fairly limited. We didn't We had some flexibility there. But we really needed this ability to to 19:56 build an experience and a data application, ? And this is what we've been building with Power BI for a number of years , bookmarks, pages, all these things. you can source the data from 20:06 Databricks and directly integrate it to a Lovable UI front-end piece. And , to your point earlier about the Claude dashboards and Claude report 20:16 development, this is a really interesting partnership. And where I'm interested most in this is I really want to understand what this implementation is going to mean for Microsoft. What's 20:27 the What's the response to what this looks because this is a really compelling -established Lovable's making a ton of money. There's not a lot of employees there. They're 20:37 building really cool stuff. And I honestly, I love It's a fun experience to build with Lovable. They've got it dialed in. And , 20:47 it feels Claude is trying to or Anthropic is trying to start step on this space a little bit with what they're coming out with. Databricks has realized they're not strong in the UI 20:57 space. They're building agentic dashboards or agentic application pieces. , why not partner with one of the best leaders in this marketplace, Lovable, at this point? , I really this integration. I have not 21:08 directly played with it. I do have Databricks. I do have a lighter weight subscription in Lovable as to play out with these with these things. But admittedly, I have not done a lot of deep dive on building 21:19 Lovable items with Databricks data. . . I feel this is the direction our industry's going. Yeah, certainly a big fan of Lovable as . one of the reasons why I I 21:31 got hooked to it was when I discovered I had Git integration or specifically GitHub integration, you know. obviously, it's a it's a SaaS 21:41 product , lots of these AI products which obviously we have way too many of nowadays lock you in, you know. You You may have a nice interface, 21:51 but then you can only do something with whatever you're creating within that interface. With Lovable, it's different, ? you do have a fantastic interactive responsive 22:03 interface, but whatever it produces gets committed straight to your GitHub repository of choice. And you can and it it it works both ways. You can You 22:14 can commit to the repo and Lovable will pick it up. and you can You can kick off a project in Lovable and then take a snapshot of the repo and and 22:26 continue, let's say, with Claude Coda or ? I really really love that. And that was the the deciding factor for me. But 22:36 when you talked about things Claude design and many innovations coming out of Anthropic , it it made me think because clearly 22:49 at the global scale, we we have more and more capacity issues , ? Yes. It's quite apparent in GitHub for 22:59 instance, they've introduced new rate limits recently. they they they kicked out lots of Git lots of Copilot capabilities from from their 23:09 student offerings. when you try to host models in in Microsoft Azure Foundry, there are more and more restrictions 23:19 there which require you to jump through significant approval and sign-up processes that they allow 23:29 you to even use those ? , even if you wanted to spend money with them, often times you may not even be allowed just because they need to manage their seemingly very limited capacities. 23:41 there are those two trends there, ? At on the one hand, Anthropic is is wanting us to spend and do more and more. On the other hand, 23:52 it seems that the world's infrastructure can't really catch up. and ultimately, , coming back to the pricing point we made earlier, this 24:03 is this can only result in things becoming more expensive cuz we've got a scarce resource that needs to be distributed 24:13 amongst more and more demand. I think this the story we are on the cusp of two fronts here that's fighting 24:24 against us. One is this idea of there's a spend from the organizational capacity, ? We can't just keep throwing bigger and bigger models at organizations. At some point they're going to yell uncle. This is too much money. I don't want to spend any more 24:34 money on these tokens. Here's our budget. We're going to hit a threshold. And I joked about this on the Explicit Measures podcast a long time ago, ? The beginning of the month is going to be great. Everyone's going 24:44 to be productive. The first 3 weeks are going to be good. And then all of a sudden, the last week of the month, everyone's going to be , , we're out of tokens. I'm going on break, ? , we're going to we're going to 24:54 inadvertently work ourselves into the 3-week work month as opposed to a 4-week work month because the tokens and things run out and, , productivity would just grind to a halt. 25:05 maybe maybe there's a new predictive cycle here that's happening where the first 3 weeks are build code crazy. And then the last week of the month is code review. Because then 25:16 then you then people have to stop building and step back and go look at what everything has been written. And it's going to grind to a much slower release cadence. But you do need someone to rationally 25:27 think your way through the code, think think through things that are going on there. , one area that I want to point out here that I think is a great challenge to this and 25:38 let's also talk about , , another one another announcement I want to bring up here is a new a news announcement from Salesforce. , Salesforce is a really big leader in the CRM space. 25:48 And they're offering what they're calling Salesforce Headless 360. when the big players, large software companies, start ditching their own UI 26:00 and saying we're building agentic first with APIs, MCP servers, and if that's if that's the move the largest players in the market are starting to move 26:10 towards, I think we need to take notice here because I think this goes back down to the money, the spend, all these opportunities here. We're going to need to become more efficient. Routing of 26:21 requests to the model is going to be very high of high importance. And I think to your point, the competition is is starting to shift here a little bit. ? I feel a couple months ago it was, 26:32 can you wield a single agent? And then it turned into, can you wield multiple agents using a CLI? Can you have a swarm of agents doing something? 26:42 And I feel again, we're still those skills are still essential, but I think we're moving into another era which is, can you efficiently wield a swarm of agents? Can you efficiently 26:53 route things and still accomplish the same amount of work with a more optimized token route? And this is some other observations I've seen here as . 27:03 In that vein, one news item I'm quite keen on Google released Gemma 4 recently. [laughter] 27:14 that's an open source model that works very on edge devices as in your own hardware, your your own GPU at home because it's 27:26 very small, but at the same time and I've been playing with local LLMs for a while . This really from my point of view, this one really stands out because the reasoning capability you get in in 27:39 in a model that size is is really fantastic. And , maybe, , that's a that's a move 27:49 we're going to see much more of moving forward, ? it's would definitely be a a if if if people and companies and and 27:59 and engineering teams invested more in smart routing, , between cloud models and and local ones, that would 28:10 solve both the cost and the capacity issue quite nicely, ? , maybe that's that's a space to to watch moving forward if you want to if you want to be 28:20 ahead of others not just in terms of output, but also in terms of efficiency. I think there's a token war coming, honestly, ? I I I think there's 28:31 going to be a substantial effort here for companies to acquire One, it's there's not enough capacity to go around at some level. 28:41 but , everything we see moving forward, whether it's image generation and I've seen a lot of these funny memes show up on social media , which is someone has building 28:51 this massive Higgins video, ? , it's a a funny video about people interacting and doing something funny on they're just political events or whatever they're doing, but they're making these funny 29:01 little videos that are animated. And I I'm hearing people comment go, , there's where all my tokens went. [laughter] You're making funny and stupid videos and I'm unable to write code and do my 29:11 daily job because you're eating up all the compute in these other spaces, which I think is quite comical because, , we're we're using this really expensive compute to build 29:22 memes and laughing about videos and our people that are building stuff. Meanwhile, I can't get enough tokens to run jobs on things and and get our companies to build things efficiently. a an interesting dichotomy 29:33 there. but Gemma 4, I think is really interesting here. And I think this is let's think of this from the big players, ? Anthropic, Microsoft, 29:44 and OpenAI, Open AI, ? Those teams want to compete on compute, ? Because they own the compute. You have 29:54 to use their models on their compute. They get money when this occurs. The open-source arena, and particularly I think maybe some of this motion around open claw, harnesses that aren't 30:05 controlled by these main central pieces, we're already starting to see a shift in Anthropic has directly cut off open claw usage on their models at this 30:15 point. I think they had a a public announcement. They're actively pushing people off of it because of this. And I think this is a I think there's a reason why this is occurring is because I think 30:25 the big players are seeing potential threats. Their biggest threat is an open-source model that can perform as as their 30:35 premier models and run at a fraction of the cost. For whatever reason, the way it's being built, the way they're quantizing the numbers. I've even heard some really interesting things around when 30:45 xAI is building chips , they're building them on integer-based solutions as opposed to floating-point solutions. And , their chips that 30:56 they're building are all integer-based, which makes them really cheap to build, really cheap to make, and can really drive down costs on what they train and how they train in their systems. , I I 31:07 I think there's potentially a war coming here of integer-based systems, floating-point based systems. Who can really drop the price down to the token level to make it extremely efficient to 31:18 run? and I think in the same way that we saw all the same I I again, I'm just throwing some random predictions out here at this point. But in the same way we saw , 31:28 mining Bitcoin turn from I just have a graphics processor card on my main computer to I go buy an ASIC miner. there's dedicated hardware to doing 31:38 the thing. I think we're going to start seeing more and more of these chips that are coming with very tuned, very optimized solutions specifically for that model. 31:48 It's not We're going to build a chip for the one model. I was even reading an article that was talking about companies are burning the model into the chip. 31:59 this is This is a 4.7 maybe not 4.7 at this point. But , hey, this is an Opus 4.6 chip. We're going to make a chip that's only 32:09 going to run 4.7. And when they do that, they get huge throughput on token output. And the cost to run that query 32:19 drops substantially. , I think we're getting close to a space where we're going to start seeing fab shops and things turn on where we're going to start seeing 32:29 once we get past more of this general acceleration cuz I don't think you can I don't think you want to do that cuz you're going to lock yourself into a chip. And then you're , , great. 32:39 Why would I go buy a chip and only use it for 6 months when it needs to last a little bit longer? At some point we're going to hit some gain max of gain returns here, and it's going to make more sense to start building dedicated chips per 32:50 burned model. I don't know, Matthias, this is pretty lofty thinking here. Do you see this going this route? Is this am I way off base here? 33:04 I don't think . and , the CFO would definitely love it, [laughter] ? If you Yes. My pocketbook would love it. If If you If you put in some big boxes Yeah. 33:14 with no a bidding mechanism for tokens, ? Yeah, correct. and yeah, that's we're probably going to see more of that 33:24 moving forward. Very interesting. All , any other final wraps we're just about time. We're going to try and keep these as quick conversational pieces here. the only 33:34 other part maybe you we want to note here, Matthias, you brought up a really good point around the Open AI pricing before we were doing the show. Do you want to call that out too cuz that's one of the only items we haven't really 33:43 talked about yet. yeah, it a while ago, maybe 3 weeks or , Open AI announced a a a new 33:57 two-tier system of their of their pro plan. which replicates Anthropic's pricing for Claude code. , they have a 34:08 $100 plan, which gives you 5x budget and a $200 one with 20x. 34:19 And that came out precisely when Anthropic reportedly had a lot of capacity issues, presumably cuz they were 34:29 getting things getting their systems ready for the 4.7 release. But interestingly enough, the Open AI announcement, I could only 34:40 find , it's on their official pricing page, but in terms of an announcement, I could only find it on one of their community forums. , there we go. just 34:54 definitely I mentioned last week or even last Friday, from my point of view, Open AI's GPT-4 35:04 is very, very close to where Opus is in terms of coding capabilities. Agreed. incidentally Peter Steinberger, 35:16 the Open Claw maker, reportedly is always been a huge fan of Codex models. It's apparently suited his particular 35:27 coding style much better than Anthropic models. definitely something to watch. I I use both of them in conjunction. I have 35:37 subscriptions for both, and I bouncing tasks back and forth almost for for peer review 35:47 purposes. that suits my own workflow quite nicely. single-vendor lock-in is never a good idea. No, I agree with that one 100%. 35:57 and in the same vein here as we wrap here as , one other area that I thought was interesting is xAI has also jumped on this bandwagon of opening up their API or, , Claude 36:08 Anthropic has locked out the the Open Claw arena, ? Here you have a harness that is Open Claw. , they they said, "Look, you can no longer use non-Claude harnesses with 36:20 their API or or signing in and using them." They put the kibosh on that. But xAI is stepping in the space and say, " they're open for business." , xAI is also 36:31 touting that you can go use xAI and their API platform with Open Claw. I think I think there's some really interesting 36:42 new opportunities here for other companies to really I I feel we're at a moment where everyone has come coming up with good ideas, and as quickly as we 36:52 can, we're seeing these ideas copied. my team just got wind of VS Code has a remote option, ? , Open Claw's this idea of you have a 37:03 a harness on your computer, you can do things, you can talk to it through Telegram. Anthropic within a month, not even, A week, I think. A week, ? It was 37:13 very fast. , Anthropic comes out with, "Hey, you can talk to you can chat with Claude code, , run it on your machine, turn it on, and you can chat with it with your mobile 37:23 device, ?" And I love that. I use it all the time. Really? . Yeah, yeah, yeah. Absolutely. Awesome. I I don't use that one very much from the from from Anthropic cuz I'm I'm more in the GitHub Copilot space. . And , we're 37:34 seeing remote open up. , it's GitHub CLI, GitHub Copilot CLI with remote add there as . , you can start using remote on your CLI that 37:44 runs on your machine, you can then talk to it through the GitHub mobile app. , it's not through Telegram, it's through a sanctioned chat channel where you can have that agent session available to you on the mobile device. 37:54 But I I got to be honest, a lot of these lightweight changes of things and apps and building and a big unlock for me, this is maybe a lesson that I've learned a little bit. A 38:04 major unlock is if you get your CICD established early in your project, ? If you're building software, if you're getting a way to publish those 38:14 changes and make it robust, it's much easier for me to step into an agent session CLI, my machine, and then talk through it through my phone and walk away from my computer and 38:25 have it making changes and committing things and making branches where I can go test them on my mobile device. , I'm I'm doing a lot of website design, building designs. I'm 38:36 going to be really interested to see how this translates directly into dashboards and BI design as . Same pattern would exist. Go build something, publish it, it kicks it through the 38:46 system. I have a something I can react to on my phone within a minute or two. I this, I don't this. I can make changes, and then I just go back to talk to it. "Hey, I want you to update this, this, and this." Screenshots from 38:57 my phone, I circle things, give it back to the agent. It knows what's going on. I really this handshake between the desktop needing to have all access to the computer and more on my phone. 39:08 . Even much that I'm talking I'm finding myself using the little microphone button a lot more on my devices. I'm talking to my devices a lot more and letting it transcribe my 39:19 language and talk it to that way. , I I'm feeling a trend that I'm starting to shift how I build. Are you seeing the same thing, Matthias? I have never done as much 39:30 deep software development over my mobile phone Oh, yeah. as I have over the past 12 months or . It's insane. , it's 39:40 really shifting how you work. Yes, it's I I spent probably I spend a lot more time interacting with agents through my phone and and driving 39:51 forward projects I'm working on then on my machine I I sit down at my machine when I need some real deep focus sessions 40:06 do some prototyping and and writing plans but then iterating and and executing those plans it's mostly the phone wow I need to we're going to unpack we're going to unpack your workflow on this 40:16 one cuz I'm starting to move more in this direction but this seems a really fun workflow to divulge into a bit more and get some more hands-on let's let's go this might be a good opportunity for 40:26 some tutorials in the future absolutely yes but maybe not today looking at the timer for today's session yeah there's much more to unpack moving 40:36 forward for everyone who's been joining us thank you much we hope you've enjoyed this session this is just us talking about the new things that are occurring in the market we're trying to react to them as we 40:46 touch them as we interact with them some key observations from this episode headless is here we're going to get more agentic interfaces from applications I think Gemma 4 is 40:58 promising I really what we're seeing here Matthias you've given some really good insight around it's powerful and I think maybe in the overall overarching part of this is 41:08 models and agents aren't always going to be free and cheap forever we're going to start seeing the cost ramp up and I think the new skill that we're going to need to start unpacking here is how to efficiently build things 41:19 model routing prompt routing I think this is a new conversation that we want to have more conversations around if you this conversation if there's something you want to learn more about please let us know in the comments or 41:29 the the description here below our description you'll find all the comments please let us know what you'd to talk about and if any of these topics are interesting to you and you want us to dive deeper let us know 41:40 Matthias it's always a pleasure talking with you about all this anything else next time we go live we're going to be a bit more hands-on again and do have some demos 41:51 I think that's a really good sequence of events for us I agree as I think thank you all much and we'll see you next time looking forward to it bye 42:04 agentic thinking [music]