Anticodeguy

AI Shopping Assistant <written by a human being> Looks like I've finally found one more practical use for AI assistants: a shopping consultant in stores. I used to spend long hours looking for products that fit on every parameter. Had to dig up reviews, check if they're real, match specs, compare brands and a bunch of other stuff that really ate a lot of time. Now you can just voice-dump your picky requests into a chat with an assistant and it, right while you're standing in the store, picks what you need and even checks availability where you are right now. Especially helps me here in Thailand, where store staff often don't speak English, and talking through translators is still pretty janky. Some things, of course, I'll still research going all in, because I find it interesting. Computer hardware, components, photo and video gear, audio and other stuff where I get a kick out of the studying itself. But routine things, like a new set of bedding or a frying pan, I'm happy to hand off to an AI assistant. By the way, just yesterday I noticed a banner in ChatGPT offering to buy ads from them. So it's started. AI is now a full-on ad platform. Which means the choice won't be objective anymore. Curious to see how this will work and whether it'll be possible to talk an AI agent out of it, to convince it that this particular product is not what I need and that it's pushing it on me for its own profit.

The Long Game Of Money <written by a human being> My epic with building a home bookkeeping system, which has been dragging on since April this year, is smoothly coming to an end. Years of data history are taking their final shape: fewer and fewer gaps and mismatches, and of course it's all thanks to persistent LLMs that tirelessly keep cleaning it up. When designing the new version of the system, I set a goal to have answers in front of my eyes to the questions about finances I care about. For example, how much money I've earned over my whole life, how much I've spent on clothes, how purchasing power changed, how much my life cost in different periods, and others. And so, looking at these beautiful charts, it finally hits me how real concepts like compound effect are, and that investing, or more precisely growing capital through investments, is a game of time. When you can't see money visually over a long stretch of time, it exists just as an abstract construct that on one hand hugely affects our lives and everything around, but on the other hand stays in the shadows, hidden from view, just like dark matter. Anyway, after I saw these charts, I suddenly realized I'm not gonna delay opening a brokerage account any longer, so I went straight to do it. I don't know yet where exactly I'll invest my tiny, for now, capital, but that doesn't change the point. By the way, for a long time I thought that as a Russian citizen I only had access to the Russian market, which, to put it mildly, isn't particularly attractive for that. But turns out that's not true. Interactive Brokers is one of the few companies that officially and fully supports opening brokerage accounts for citizens of my homeland, which I'll gladly use. The first rule of investing: don't lose money. Remember it.

Fake It Till You Make It? <written by a human being> Just listened to the latest My First Million podcast, where the guys were sharing their takes on the documentary You Can See Everything, about Elizabeth Holmes, who's currently serving time in prison. For those out of the loop, this girl founded Theranos, a company that blew up worldwide thanks to pompous marketing and huge investments and promised to revolutionize medicine. They made a device that could run tests on literally a drop of blood and detect all kinds of diseases, helping to stop them before they develop. Turns out the whole thing was fake and the girl got sentenced to 10 years for misleading investors, breaking the law by using the Pfizer logo without approval, and other unpleasant stuff. So the documentary lets you look at this story from another angle, since the director spent the last days before her imprisonment living in Elizabeth's house, talking to her, her family, her close ones and the people involved in this whole drama. But I wanted to talk not so much about the film, which I haven't watched yet, though after all the rave reviews I really want to, but about an interesting observation from the world of venture business. There's this stubborn template, "fake it till you make it", which literally describes what Theranos was doing. Only in their case there was a clear tilt toward "fake", or maybe just not enough time to get to the "make" side. Either way, this line is thin as a knife blade, and holding the right balance here is where the real mastery is. For me, as someone far from the US, where venture rules the tech world, this approach is still more of a pretty fairy tale and a marketing story. Yet how many great businesses were built this way, when people sold something that didn't exist in reality yet, only in a presentation. And somehow we only hear about the failures when someone gets put behind bars...

Meeting Notes On Autopilot <written by a human being> Haven't done a practical AI-for-business thing in a while. Here you go, sign for it. Automatic transcription of work meetings became one of the first areas LLMs entered by kicking the door in, and nobody even tried to stop them. The convenience is so obvious to everyone, and the relief from boring routine work so big, that even stenographer-secretaries (do they still exist?) didn't mind handing their bread over to soulless machines. Zoom and its siblings immediately got plugins and add-ons that transcribe, summarize and even log tasks for call participants automatically. Which, of course, saves a ton of time, but the quality of the free solutions leaves a lot to be desired. My team and I noticed right away how bad Zoom's native transcriber is, so as soon as we had the resources for it, we decided to swap it for something better. After a short market research it became clear that the leader in transcription right now is ElevenLabs models, which we plugged in for this task. On top of that, we set up saving all transcripts, plus the recordings themselves, into an internal knowledge base. You can search it and find something to pin on the boss's careless talk, the one nobody pulled by the tongue three months ago. Simplified, the pipeline looks like this. When a call starts, recording starts automatically. A worker periodically checks the Zoom API for new recordings, and when it finds one, it triggers the following actions. The audio track is extracted from the video, saved to the archive and sent to ElevenLabs for transcription. The video itself is saved to YouTube. The resulting transcript together with the video forms a new page in the knowledge base, dated with the meeting time. After all that, the video and audio get deleted from the Zoom cloud, so they don't take up paid space there.

The Attention Game <written by a human being> Just listened to a Twitch streamer who wants to find his niche on YouTube. He's not a beginner anymore, around 1.5-2K people watch him, so things are going relatively well for him there. But YouTube is a different story, because content works differently there. It's built on grabbing and holding attention. If you don't do that, nobody will watch your content, no matter how interesting, high-quality and useful it is, three times over. And it's not even about the content maybe being uninteresting to the mass viewer, it's more that in this setup YouTube itself has no reason to show it to a wide audience when there's another video that's already popular. That's the one to push in the feed, since it means more retention and watch time. So the key question here is how to stay authentic and true to yourself while still making content that hooks and holds. When you look at some creators, it feels like they were literally born for this. They've got hooking charisma, a good sense of humor, they grab attention at first glance. And you can see it from the very first video, shot many years ago. But if you don't have that gift from nature, it becomes a skill, one that you can and should develop if you want to make it in this field. Either way it's a game for attention, which matters in pretty much any area, from business to politics, from relationships to a content creator career. And this development of the necessary skills, along with getting to know yourself to find and realize your inner potential, is in my opinion the real game of life that we're all playing.

Token burn depends on your repo <written by a human being> That's what I noticed while switching between projects. In one, a whole day of work with 4-5 sessions running in parallel on Opus 5.5 eats about 20% of the weekly limits. In the other, even with less intense work in two sessions over a shorter stretch of time, it's 35-50%. And it's clearly not the codebase. The project that burns less has a much bigger codebase, way more system components in use. Second suspect - project disciplines. And here too everything looks fine, both projects are set up the same way and are equally frugal with tokens. The devil was in the work agents do under the hood, invisible to the eye. The second project, the one mercilessly burning limits, is my home bookkeeping, where I'm reworking the logic of the bot that takes in transactions. The bot lives in Telegram, and that's where I send it receipts and messages about expenses. Then the bot uses an LLM to classify the transaction, figure out the right account, category, recognize items on receipts and so on. So yesterday Opus decided to run a regression against my whole corpus of messages already in the bot - there are thousands of them. And on every fix it ran the full pass over the entire message history to see whether the fixed function works. The funniest part is the model there gets called with the --no-session-persistence flag, so session transcripts don't get saved, and catching this thing right away wasn't that easy. Anyway, watch your usage - an abnormal burn rate clearly means something is working wrong.

Build Once Ship The Same Container <written by a human being> The web application you’re developing is, in its final form, packaged into what’s called a container. This container (or several of them) holds everything the system needs in order to run. It is then launched and runs on the resources of the production server. When a system needs to be tested before a release goes to production, that is, to the very server that real users are served from, it is deployed to a separate machine, isolated from production. That’s where all kinds of tests are run. But if the application is built in the test environment independently of the build in production, the test and production applications may end up being different. These are different builds, different containers, running on different machines. What can you do about it? This is where containers and their shared storage come to the rescue. The idea is this: the application is built once into a final container image and pushed to a dedicated registry, tagged with the corresponding version. Then that same container is pulled onto both machines, the test one and the production one, with no new builds. In other words, the container deployed to production is physically the very same one that was previously deployed to the test environment and thoroughly tested. Many hosting providers let you create such container registries with literally a single click, so you don’t have to bother with maintaining their functionality and reliability, which is very convenient. Once your project matures to the point of needing remote builds, you’ll now know one of the tricks of developer kung fu.

Local Dev Hit The Limit <written by a human being> As you dive into fullstack development with AI agents, you also run into all the joys that classic programmers deal with: setting up and maintaining a local development environment, running application builds, and running different kinds of tests. As long as the system isn’t too complex, almost any machine an average person works on today can handle all of this. But once you cross a certain threshold, the hardware starts to fall noticeably short, and at first you find yourself thinking about upgrading your components. Then it occurs to you that surely not every programmer upgrades their hardware each time their tasks get more complex, and there must be another way. I first ran into this a few months ago, when I noticed how heavy the builds of my working application had become, especially when I needed to spin up several environments at once to check new functionality. Back then I wrote a post about a staging environment: a dedicated server for isolated application builds. And this week an incident exposed one more gap in my development setup. Several agents simultaneously launched several Playwright browsers to test their builds. Together they ate all 32 GB of my machine’s RAM, which in turn killed the production build command that was running in parallel. So my new challenge is isolating testing and moving it off the local machine. More broadly, I want to make this a default principle of the development environment for any product, and remove the dependence on the local machine.

City Of Contrasts <written by a human being> Today I got back from a short trip to Bangkok, where I flew for visa stuff. So today there’ll be a travel blog note. Bangkok is one of those cities that won’t leave you indifferent. Here the slums, where poverty rules, where the homeless sleep on the counters of closed shops and the ever-present concrete rots from the constant humidity, sit right next to modern glass skyscrapers that lure shoppers from all over the world into huge, pompous malls and make billions of dollars for their owners. Here a Maybach with tinted windows drives past a creaking tuk-tuk, in which an old man, earning his food, has tanned his skin under the scorching sun so dark that he forgot what colour it really is. Just the other day I rewatched Bangkok Hilton, which I remembered as a very scary horror from my childhood, and now I took it in from the point of view of someone who understands Thai culture a little, having lived here for almost five years. And the 40 years that separate the film from the present moment radically changed my attitude to Bangkok, from fear to pleasant admiration. In a lot of ways it reminds me of Singapore: the glass and concrete, the greenery and humidity, the enormous number of buildings and people. But in Singapore you won’t run into monitor lizards clumsily rolling around right on the pavement by Lumpini Park. Or the huge number of colourful carts with street food, where you can find all kinds of exotica like fried bananas or pickled scorpions. There’s so much here that’s unusual and at the same time down-to-earth that you just can’t decide whether it’s the cyberpunk Chiba from Gibson’s Neuromancer, or some forgotten-by-everyone business centre that doesn’t try to compete with Dubai, New York or Hong Kong, and so lives a completely independent, original life, developing its own unique character. That’s why I sincerely believe everyone will find something in this city to love it or hate it for, to admire it or look at it with disgust, to keep it in their memory for life or just record a reel that will vanish from the feed in a day. A truly amazing and one-of-a-kind city of contrasts.

My Go-To For Websites Nowadays <written by a human being> One of my longtime and good clients, Irina Willems, a well-known in her circles executive coach from Monaco, has come around to redoing her website, because the business has changed since it was built and her media presence needs updating. Before, my go-to for websites in such cases was always WordPress: a well-proven platform, tested by time, easy to maintain and accessible enough for a non-technical user. But with the arrival of coding agents, my approach to development has changed too. I finally got access to all the modern methods and approaches that used to be closed off to me by the high price tag of programmers. I learned about static site generators a long time ago, back when Gatsby and Headless CMS were just starting to be talked about in narrow circles. They let you build sites that were incredibly fast for the time, loaded in fractions of a second and got scores close to a perfect ten in every possible speed test. But now my costly programmers are Claude Code and Codex, who handle such tasks without extra invoices beyond the subscription. So my stack for building websites has changed accordingly. Now my go-to is Astro, which leads among static site generators in speed and minimalism of code. As for the content management system, I decide depending on the situation. Sometimes it’s a full-fledged admin panel on something like Payload CMS, and sometimes, as in Irina’s case, it’s just a repository connected to an agent, which in turn will carry out her instructions for updating the content.

Dots In The Real World <written by a human being> Last night was the moment of testing Dots in “field conditions.” To remind you, my assistant arranged my trip to Bangkok: it picked and booked the hotel and tickets both ways. Online check-in opens 24 hours before departure. You get your boarding passes and can go straight to boarding, skipping the check-in desks and, therefore, all the queues. I’ve been doing this for many years, ever since airlines first offered it. That part is really convenient: I asked the agent in advance to check me in for the flight as soon as it opened, so I didn’t have to time it or break away from what I was doing. I just got a notification that check-in had started, and all that was left for me was to wait for the result. But then the stupid questions started, at every step of the form that the meticulous AI assistant was filling out. First it asked whether to use the same phone number for check-in as I’d used for the booking. Then a question about the passengers’ citizenship. Then several questions asking me to confirm checkboxes saying we aren’t carrying anything prohibited and, finally, that we agree to the airline’s policies. I generally don’t understand the point of these questions and checkboxes. As if I could choose differently and still fly. Like, “No, I do not confirm that I’m not carrying anything prohibited.” And they go, “Oh well, take your seat then...” The most annoying thing was that I couldn’t get two adjacent seats for me and my girlfriend. When we first booked, the airline issued two booking codes, so we had to check in one at a time, and we were given seats in different rows. But that’s not the AI assistant’s fault. Either way, all this fuss with a pile of questions turned out to be no faster than filling in the form myself, where I’d have ticked those boxes without any waiting. So I still don’t share the value or the hype.

Opus Beats Astra <written by a human being> I’m continuing my development diary for a home bookkeeping system, so far my longest project with AI. Over the last two weeks I managed to compare, back to back, how Opus 5.5, Astra and Sol 6.1 handle it. (As a reminder, this is a weekends-only project, so I don’t touch it on weekdays.) In these two weeks Opus 5.5 noticeably pushed it toward the finish line. It’s great at working through historical financial data in big batches, and pretty fast too. In 10-20 minutes it brings me a report on the next chunk of data, with a clear breakdown of how to classify everything and which errors to fix. I check every single transaction carefully and meticulously, but with Opus, in 99% of cases all I have to do is agree with the proposal. Everything is absolutely correct, and the pile of “dirty” data is shrinking fast and visibly. It’s the same story with improving the system itself. 1.5-2 hours of work and the needed features are written, the regression is fixed, everything is tested on live data and works as expected. But Astra and the latest 6.1-Sol don’t come even close to these results. With data processing it’s sad: 30-40 minutes of work, hundreds of thousands of tokens, some huge amount of code written for no clear reason, and the agent brings me one (no exaggeration) single transaction to approve. My regular question became “so in this session you’ve only sorted out one record in the database?”, and the agent sheepishly agrees. With feature work it’s even worse: the session ran for more than 9 hours, almost two million tokens were burned, and the task still wasn’t finished. I had to just kill it, because there was no end in sight. Again, these are identical tasks in the same project.

Dots Instead of OpenClaw? <written by a human being> The payment incident with z ai, which blocked me from using GLM in the Hermes harness, got me thinking about possible alternatives. And right around the same time, a new wave of AI assistants launched, including OpenAI's Dots. These AI assistants come with their own remote machines where they can execute tasks. And you don't even need to keep your own computer running. Wait a second, isn't that exactly what made OpenClaw so popular in the first place? Autonomy! Another key feature that made everyone rush to buy Mac minis and install OpenClaw on them, and later Hermes, was the ability to communicate with an agent through a familiar messenger, whether it's Telegram, WhatsApp, or corporate Slack. But guess what? Dots can do that too! The only thing these newly released AI assistants seem to lack is control over memory, skills, and instructions. Or rather, all of that is there, but hidden somewhere under the hood, and it's unclear how it works or how to control it. But maybe we don't actually need that control, and it's a good thing that the provider has taken care of it for us. In my case, I'll still have to do a bit of tinkering, since Dots doesn't yet have a native integration with our corporate messenger, Mattermost. But I've already solved that problem by writing my own MCP server and installing it as a plugin that the Dots agent can easily use. I'll say right away that I haven't finished setting everything up yet, and I don't have a full replacement for Hermes just yet. But it looks like all the pieces are there to make it happen. I'll share the results, of course.

Z ai Incident <written by a human being> An unpleasant thing has happened over the past couple of weeks with z ai, where I bought a Coding Plan subscription for access to GLM models. As a reminder, these guys are the only major model provider that lets you use their models with third-party harnesses like OpenClaw or Hermes (in my case) under subscription limits. Meanwhile, the flagship labs ban accounts for exactly that, nudging you toward their own harnesses (Claude Code and Codex), and if you want to run your own, you pay for API access, which is orders of magnitude more expensive than a subscription. So, after z ai updated its pricing, I moved from the $64 plan to the $80 one and paid for another month, but... the subscription never activated. The first thing I did was check for a scam: maybe I’d clicked the wrong link somewhere and paid on a fraudulent site? Nope, it was their site, and I checked several times. The SMS from my bank points to exactly the same merchant I’d paid before, and the payment itself went through as usual, but there was no subscription. At first I didn’t think much of it and assumed the automation hadn’t fired or something had broken and I just needed to wait for activation. But the next day it still hadn’t activated. That’s when I started looking for z ai’s contacts. First I found an email address and wrote a letter. Then their X account. I filled out the sales team’s form and, finally, made it to the Discord community, where I at last got a response. Z ai claims it never received my money, which is why the subscription wasn’t activated. And, of course, they sent me off to sort it out with the bank. I opened a dispute with the bank, and they in turn confirmed that, on their side, the money was sent to the merchant. So somewhere in the vacuum between the bank and z ai, my money got stuck. With two parties now looking into that transaction, I decided to pay for the subscription a second time, since I need to work and my Hermes was sitting idle. And again the same thing: the money left, but never arrived where it should. After a few days, the bank refunded me for the first transaction (the second one not yet). So be careful when paying z ai: something’s off there...

Where Skills <written by a human being> Somehow, all of a sudden, skills have fallen into the background. Or rather, turned out not to be needed at all. This week I caught myself realizing that my agents now use only local skills tied to the specifics of a given project. Global skills, on the other hand, I haven’t used in my work for a long time. The blame, in my opinion, lies with the intelligence level of the latest models. They handle tasks perfectly well without instructions from skills. At least in my projects, I haven’t run into a situation where I’d see a need to pull one in. Even the much-praised Superpowers, which is a permanent resident in all my projects, hasn’t been called for several weeks in a row. And honestly, I didn’t even feel the need. It used to be different - as soon as something went wrong, I’d call a skill that straightened out the agents’ behavior. Now there’s simply no need to do that. Yes, project-specific algoritms stay, of course. The ones that describe deployment rules or how to handle tasks in the internal tracker. Everything you need to know precisely as working rules. But even those are down to literally 2-3 per project. The latest models handle the rest on their own. Either way, it’s great news - time to clean out the unused plugins and skills, which, by the way, eat up your context window at the start of every session.

How To Claim Your $200 API Credits <written by a human being> This morning started with cash gifts from Anthropic. Every month they promise to hand out $100 or $200 in API credits if you're on a top-tier plan. In the app, open Settings, then Billing, and hit Create Organization, even if you have a personal account. Apparently it's technically easier for them to run the giveaway through organizations. Next, you need to log in to the Claude Console, which opens in your browser via a link. Then you'll be asked which organization you want to link your account to. I picked the individual one. And the next step is a screen with a big white button, Link and claim $200 (for my $200 plan). That's it. After that, the money lands in your account and can be used for API calls. Worth noting: $200 in API credits is nowhere near the same thing as a $200 subscription. As a reminder, usage limits on a subscription are equivalent to thousands or even tens of thousands of dollars if you run the same tasks through the API. I could easily burn through all $200 in a single coding session on one project, and most likely wouldn't even finish the task. So let's not get too excited: I still need to think about where these generous handouts can actually be put to use. But overall, it's clear the AI giants keep competing in generosity and doing everything to butter up their users. Well, we won't resist...

Cute Tamagotchis - Same Old Agents <written by a human being> The new personal assistants in the form of AI agents came out last week, and I still haven't written about them. But that's because I couldn't find a task for them that would let me test them properly and understand where all the hype around them comes from. Yeah, cute tamagotchis. But what's actually new besides that? They can do tasks for you? Yeah, like any normal AI agent we've been working with for months. Oh, it has its own computer in the cloud and can write in a messenger? Hi OpenClaw, Hermes, which turn out to be not that useful compared to an agent that works locally. And how is all this different from remote access to a running agent session, which you've been able to launch from mobile apps for a long time? Honestly, I don't get the hype and all the attention. I did give it one task though - planning a trip to Bangkok. I need to be there next week and so far this is the only case where these autonomous agents can be useful. I told Dots (the one inside ChatGPT) to find suitable hotels closer to where I need to be, with a convenient flight schedule. Here it saved a bit of time - yeah. I didn't have to mess around with searching and picking a suitable combination. It just sent me hotel options and I was picky about choosing. I gave it access to my booking system account, where passenger details from previous trips are already saved, and at payment all I had to do was enter the card number manually. So basically the AI assistant clicked around on the computer for me. Honestly, I don't buy the hype of this whole story at all. What impresses me way more are coding agents that in a week do the equivalent of several months of work of a full product IT team on my project. But somehow everyone's already forgotten about that...

Systems Analysis Is Back! <written by a human being> One unclosed gestalt has been hanging over me since the vibe-coding era started, and it's systems analysis. More precisely, the poor quality of it when you delegate those tasks to AI agents. Time to come back to the question and test the latest models. And right on my current project I needed to model a production process and calculate labor costs in detail to figure out the cost price of the product we're building. Wait, this is a textbook case for IDEF0! About half a year ago I wrote about my first attempt at this task - back then it went really badly, the models lacked knowledge of the notations, understanding of how to build coherent processes, the context window to grasp the whole system. This time I again started from the basics - compiling lists of objects and functions, interviewing process stakeholders - and fed all of it to the latest Opus. Then we brainstormed how to display process diagrams simply, no hassle, but still nicely and neatly. We landed on plain YAML files as the base, a Node script that checks the structural integrity, and a React Flow component for visualizing the diagrams. The result is a pretty solid, clean diagram, and most importantly - a properly consistent process that needed minimal edits from me. Of course there's a lot more that can be tuned, but overall the mechanism works, and AI agents pick it up easily thanks to the clear structure. So now systems analysts are definitely on their way out. Or rather, the ones who don't get along with AI in their work.

My Favorite Habit <written by a human being> Back in the day, before I started writing about AI, I used to talk about different aspects of my life philosophy and the practices I use. Today an idea popped into my head to reflect on that a bit. One of the habits I picked up back then is long daily walks. Honestly, I didn't think I'd get this used to it, since at first I didn't really like it. But I combined it with something useful, to satisfy the restless brain that constantly wants to be doing something, and started listening to podcasts and audiobooks while walking. But morning walks are sacred - that's meditation and just an empty walk with no content whatsoever. Pretty quickly I started enjoying the walks and didn't notice how they became part of my life. Not gonna dive into the list of benefits you get from this habit. But one of the most important ones today, in my opinion, is the chance to be alone with yourself and the world around you, ideally far from the internet and the endless information bombarding your attention. At least for a while. Now I can't imagine my day without a morning walk with meditation and without an evening one that wraps up the workday. And it's one of the few things I can confidently recommend to anyone who wants to change their life for the better.

The Real AI Stress Test <written by a human being> The real intelligence stress test for models, for me, is my home bookkeeping system project. The very one I started back in April and still haven't finished. The difficulty is not so much the mechanism of the system itself - that one's actually not that complicated - as the tangled data banks that need to be matched, stitched together, gaps found, and everything linked into meaningful financial chains. It's almost 20 years of my financial history, a ton of bank accounts, several accounting systems with different data formats, debts, loans, lost records and a lot more. All the Claude and ChatGPT models have been taking turns trying to deal with all this junk as they came out. And seems like they're slowly digging through years of history, but no light at the end of the tunnel so far. Not Sol, not the latest Fable, not even the much-hyped Astra handled the task in a way where the progress was actually noticeable. But Opus 5.5 really got things moving. It found an approach to analyzing the data where it now just cleans everything up in huge chunks, and does it with surgical precision, which its predecessors couldn't pull off. And I'm not talking about temporary patches, but absolutely legit changes, quality findings in the data and correct interpretation of loads of statements, receipts, old databases and other accumulated goods. Finally I see a straight path to the end of this great vibe coding epic, which I will report on with great pleasure, of course. Right now Opus 5.5 is state of the art for me.