Anticodeguy

300+ Videos Time To Change <written by a human being> Yesterday was the first Friday in a long time when my regular talking-head video didn't come out, the one on the topics this head usually talks about. The reason is simple - I ran out of videos. Some time ago I deliberately stopped shooting them in this format, because I don't see any progress. Everything's stuck in one place. 1.5 years of posting, over 300 videos, and still nothing's moved. Time to change something. And it's not like becoming a YouTuber is the goal of my life, though maybe that's exactly why there's no growth here. But in anything you do, feedback on progress is extremely important to get confirmation of your own growth and of unlocking your potential. Because if there's no forward movement, then you're either going backwards and degrading, or at best - standing still. Both options are unacceptable to me, so it's time to try other approaches. What exactly - don't know yet, but the time of experiments is coming. Maybe I'll change my approach to writing too, but I like the current format myself - short essays, one thought a day, about things that are relevant to me right now. It feels very natural to me and works even more like a daily journal. But something's definitely off with the videos. And I already shot footage for a new one, but haven't had time to put it into a final render yet, because it'll take a lot of time. This type of task is new to me and the final vision isn't formed yet. I don't want to use AI yet either, at least until that vision is there. Otherwise I'd have to rely on the "creative potential" of an LLM, which only takes me further from self-realization. Anyway, a video in the new format is coming soon, can't wait myself.

Opus 5.5 Week Report <written by a human being> End of the work week, and the first week I used Opus 5.5 exclusively for development tasks. And I'm more than happy with the results. 45% of the weekly limits are still left in the tank! On one hand, this week was a bit less intense in terms of load on the agents. On average there were 3 orchestrator sessions running in parallel at the same time. Of course, within those sessions the head agent launched subagents, which in turn did the actual specifics: wrote code, wrote documentation, handled configs, tested the system. But overall I'm pretty impressed with the genuinely noticeable economy of limit usage. Especially against the backdrop of Codex, which has gotten insanely greedy lately, as the whole AI community won't shut up about. Keep it up, Claude! The limit usage stats also include Fable, which I only used at the start of the week, to burn through the sessions left before the reset. But right after that I switched to Opus, since Fable burns more limits. Which means the current numbers aren't the ceiling either, there's still gas in the tank if you take Fable out of the equation! And the quality of the tasks, in my view, is not worse than Fable at all, maybe even better. There were a few moments when the agent clearly got lost in context and pulled the task off to the side, but overall it wasn't often enough to spoil the general impression. So, my go-to model right now is Opus 5.5, both for orchestration and for direct execution.

How To Live On Earth <written by a human being> Yesterday I started watching a documentary I just can't walk past. Stuff like this has always caught my attention, ever since childhood I devoured encyclopedias about our planet, wildlife and plants, climate, ecosystems, nature as a whole. To me it feels completely natural to treat the world around us with care, to clean up after yourself at the very least, and generally keep the impact of your own activity on nature to the minimum possible. However, for some reason it's not obvious to everyone, and I won't get into that. And if even on the individual level not everyone manages to get this, then what can you say about the collective consciousness, which is looking in a totally different direction. The documentary helps open your eyes to it. And reminds us that in our hands there are not only the tools and the power to destroy the Planet and its resources, but also the chance to help restore what we took from it. Some of it already irreversibly. On one hand I know that nature will take its own anyway. And even if we completely drain absolutely all of the Planet's resources and move to live on Mars or space stations, over time, even if it takes thousands and millions of years, life will recover and rage back with new force, burying the remains of the former civilization and slowly but surely turning them into new oil deposits. But we don't have to take it that far. Terraforming Mars is a wonderful ambition, but why, just to turn it into a lifeless desert again later? What about taking care of our own Planet instead - our common home, which gave us life and the chance to be here and now.

Simplicity vs. Overengineering 1:0

Product management skills are coming to the front <written by a human being> Lately I notice I'm switching to the product owner role more and more often, and to the technical specialist role less and less. And I'm saying this as someone who spends 96% of working time on development with AI. The first stage, when there's no foundation yet, no disciplines, no infrastructure - that's mostly technical skills, of course. But this part doesn't differ much from project to project. Except the amount of stuff you need to prepare before the start gets smaller the simpler the system you're building. But once the foundation is laid, it's mostly product work from there. Decisions about user behavior, interface design, business process logic come to the front. The technical part is still there - from time to time you have to adjust the disciplines or update everything for new models and releases, which I write a lot about too. But mostly what comes in handy is knowing how to build a quality product that solves the tasks it's given, and how to manage that build. This is where task prioritization, release cycle management, testing come in. So product people, in my opinion, are extremely valuable on the job market today.

Opus Orchestrates Itself <written by a human being> As planned, I switched to working with Opus 5.5 as the orchestrator of itself. Before the release I picked Fable for orchestration, which was already spawning Opus agents to actually write the code (and Sonnet for simpler tasks). And last week I kept going the same way, but the subagents were already running on Opus 5.5, so, honestly, I didn't really notice any difference. But since Opus 5.5 is more economical and token-efficient than Fable, it makes sense to use it as the orchestrator now. First, Opus 5.5 handles the orchestrator role really well. And raising the context window limit helped too, and no intelligence degradation even at the 400K token threshold. Second, it's now truly autonomous work with no need to watch the process. Opus 5.5 easily and effortlessly controls several agents at once. Considering the subagents themselves run on the latest model, the results are good quality, so no extra rounds of rework. Third, the token savings are really noticeable! Opus really does spend them more efficiently than Fable. Usually with this volume of work I'd burn around 20% of my weekly limits in a day. And yesterday - only 15%. 5% might not seem like much, but it's still noticeable. Though hard to say whether the model actually got more economical, or Anthropic tweaked the taps and got more generous with limits, or maybe both at the same time.

Hit Linear's 250-task limit. So I replaced it for free

Kodi Saved Movie Night <written by a human being> Amazing, but in 2026 it turned out to be pretty hard to find a video player for Android TV that supports all kinds of formats and codecs, especially the outdated ones, the ones a ton of classic movies were rendered in, the kind you won't find on modern streaming services. For a long time I had VLC installed, which has long proven itself as a player that can play literally anything, ever since watching movies on PC back in the way-back noughties. But on Android TV some renders make the frames stutter, making it impossible to watch. As a backup I installed MX Player, which handles old codecs better. But for some reason it refuses to be friends with some audio tracks and subtitles! After yet another ruined movie night I decided to find a solution and, as it turns out, there is one - the Kodi player, which perfectly supports a ton of codecs, and all kinds of subtitles and audio tracks too. Kodi turned out to be extremely greedy in that department. But there's another challenge with it - you can't just "pick a file" to watch. You need to create a movie library. Here any AI agent will easily help us, generating the library file and organizing everything you've got stashed away. And Kodi in turn will show that library in a beautiful, convenient interface no worse than streaming services. So if you ever ran into similar bullshit - use it. Vibecoding a new player turned out to be unnecessary. Vibecoding a movie library - that's easy mode.

Pendulum Swings Again <written by a human being> The pendulum of competition between Claude and ChatGPT has swung again, and now it's on Anthropic's side. Of course, I wasn't the only one who noticed the changes in Codex limit consumption, which I wrote about just a few days ago. And on X this is the news of the day, being hotly discussed in the community. Their latest model turned out to be much more economical with tokens, and they also tweaked the available limits - it becomes very noticeable when you work with AI agents constantly. You get used to a certain speed and density of work, roughly calculate the remaining limits over the course of a week. And when they change, you notice it right away. Claude was never originally known for solid limits. And Codex has been crushing it on that front for the last few months, leaving its competitor far behind. That's why the current cut in limits from OpenAI feels especially painful. And when against that backdrop Anthropic does the opposite, raises limits, and on top of that releases an economical model, there's nothing left to do but switch back to Dario Amodei's camp (founder and leader of the company that sells us Claude). For the average user, in my view, this is good dynamics. There's no clear leader - they trade the podium with enviable frequency, which means they're constantly feeling the market's demand for change. And computing power is today's competitive advantage. It's obvious that users switch easily from one vendor to another. And obvious that they switch to whoever gives more favorable limits. Waiting for OpenAI's response and the moment when Codex becomes generous with its limits again.

Raising The Context Limit <written by a human being> With the switch to Opus 5.5, I felt it was time to revisit my discipline around context window thresholds. As a reminder, in my projects I try to keep the context window of a single lead session under 160K tokens, after which it gets passed via a handoff prompt into a fresh one. This approach helps fight the intelligence degradation models show as they approach 200K, as well as unnecessary token spend, since each subsequent request starts costing disproportionately more. Given that the model also gets dumber, the number of requests keeps growing, and it becomes a vicious circle that's easily broken by a fresh context window. So yesterday Claude and I pulled a fresh snapshot of my last several weeks of sessions, ran the numbers, read through Anthropic's recommendations and the latest Opus 5.5 specs, and concluded it's time to raise the limits. There are several reasons for this - Opus 5.5 holds context well, meaning it keeps following instructions even at a full 1 million token window (the model's current limit). But pricing has also changed - cache reads have gotten cheaper, which means even in heavy sessions, new requests won't drive up the cost of the cycle by much. On top of that, given that in every new session the agent effectively has to rebuild tens of thousands of tokens of context from scratch anyway, the upside of a fresh start with the latest models shrinks to a minimum. So, I've raised the available limit to 300-400K, and it's already noticeable on specific tasks, which now get closed in fewer iterations. How much this will actually help save on token spend - we'll see in practice. It'll take a couple of weeks to break in the new mode given the reset limits, after which it should be clear. For now - we're watching.

AI Isn't Omnipotent <written by a human being> I keep chewing on the topic of AI's supposed omnipotence - the thing doomers are so afraid of and that most people who don't work deeply with LLMs practically worship. You can take any field AI has penetrated: code, audio, video, images, text. It seems like here it is, the infinitely hardworking employee that will do whatever you want. But in any of these areas, in practice, you run into the fact that the agent on its own won't carry a task through to completion and to the result you actually wanted. You can't develop, say, a system for your business from a single prompt and have it work correctly right away. Left to itself, it won't figure out integrations with existing systems, won't clarify the details of business processes, won't account for the many details and quirks specific to a particular company. Yes, it will assemble a system that can exist in a vacuum, on its own. But what's the use of it if it can't work with the documents that pass through the finance department? Or a system whose processes aren't tied to standards that are already in place. All of this requires carefully digging into the details, the business processes, the entities, and building the architecture. Maybe AI will soon learn to do all of this on its own. Maybe it already can, but no subscription currently available would be enough for it. And all those stories about a swarm of AI agents breaking through some system's defenses (like the recent news about the breach of Australia's Medicare) - that's a completely different order of computing power. Those are the kinds of experiments OpenAI runs, backed by colossal, government-funded data centers. Personally, I'd be curious to see what an AI attack like that would cost in money terms if, say, an ordinary user tried to pull it off. But that's still out of reach for mere mortals.

Opus 5.5 Is Good For Your Token Budget <written by a human being> After actively testing the latest Opus 5.5 on everyday tasks, I can say with confidence that the model really is very economical with tokens, just like Anthropic said at release. I hand development tasks to Fable, which in turn dispatches and controls Opus agents in its own sessions, the ones that actually do the final work. I find this setup optimal in terms of token spend. By the end of the work week both the general limits and Fable's hit 100%, meaning there's no unused quota left. But after the child agents started running on Opus 5.5, the dynamic changed. Now Fable burns slower than the general limits. Which tells me it's involved less in the tasks. Meaning the head agent doesn't have to sort out what the child agents "messed up", and the result comes in cleaner than with the previous model. Considering Opus 5.5 is cheaper and more economical than Fable, this fact can't help but make me happy. The next experiment I'm planning is to try Opus 5.5 as the head orchestrator instead of Fable. Curious how it'll handle the job. But in the end the only thing that matters is how well and how much work gets done. And if the new Opus handles it better and manages to do more within its limits, so be it. By the way, they also added Reset, just like in ChatGPT, which has been spoiling people with those for a long time now.

Langgraph turned out to be exactly what I expected

AI End At The Screen <written by a human being> Yesterday I wrote about the illusion that AI can magically do things autonomously, without human involvement. And while that's actually true, it's only true within very limited bounds, up to a certain degree. LLMs are generative machines that produce textual, graphic, video, or audio information. But exclusively within digital space, or inside computer systems. It's not text written by hand on paper, not sound recorded from the surrounding environment, and not a piece of the real world captured through a lens. On the other hand, computers today make up an enormous share of everything humanity does - from simple household tasks to complex financial and logistics systems that literally deliver food from the other side of the world to your refrigerator. Everything is tied to digital systems. Understanding this creates the illusion that AI could, in theory, do everything a human does, by operating its cognitive apparatus through these computer systems. Because AI has now learned to perform cognitive work at a human level. By the way, the move into the offline world is already happening right now - more and more companies are appearing that work on robotics. Robotic machines in the real world, controlled by AI from the digital world - that is exactly the bridge that will potentially lead, in the end, to a full merging of the two fabrics of our society.

AI Is Not A Genie <written by a human being> Over the past year, many people have developed a persistent feeling that AI is some kind of magical genie that pops out of a lamp on command and grants any wish you make - not just three wishes, but as many as your subscription limits allow. People here tend to fall into three camps. The first are those who don't want to touch AI at all, and generally see it as heresy, a threat to humanity and everything sacred - or simply resist it out of rigidity and a general aversion to anything new. The second camp uses it constantly, every day, for many or even most of their tasks. These people can no longer imagine life without this technology. They're the kind of people who, after a few hours flying across the Atlantic, can't understand the point of a five-day ocean liner crossing from the European coast to the American one. And really - why would they? The third camp, in my view the largest, consists of those who use AI from time to time, for certain categories of tasks. They dip their toes into the unknown, but quite tentatively, and nowhere near as deeply as power users. Still, they're well aware that AI can now do things that used to require an entire team of people and an enormous amount of time. It's this last camp, I think, that holds the most illusions about what modern AI tools are actually capable of. They have a general picture, some understanding, and even hands-on experience that leaves an indelible first impression (remember your own excitement after your first few prompts). And from there, imagination and fantasy fill in the rest - blending scenes from sci-fi movies, bold pronouncements from technocrats, and clickbait YouTube thumbnails promising that AI is earning them roughly a billion dollars a second, on autopilot.

Claude Design

Instructions Must Evolve <written by a human being> The latest model and harness updates from OpenAI and Anthropic have once again confirmed the need for constantly updating project instructions for AI models - and taken that need to a whole new level. If before, after an initial cycle of setting up project disciplines, I could loosen the reins a bit and work calmly for a couple of weeks without intervention, now I can no longer afford that luxury. Literally about a month and a half ago, my workflow looked like this: I'd open a terminal in the project folder, launch Claude Code, and type in a task number, say #1253, then do the same thing in a few more terminal sessions and go off to handle my own business. After 1-1.5 hours I'd come back to the project and review the completed tasks. I'd write notes and comments for revisions, and after final approval, I'd launch the next batch of tasks. Now, exactly 50% of my time goes into sorting out where and how Claude Code or Codex messed up, how to fix it, and reworking the disciplines for new models or updates - which for some reason just make these new models ignore the established rules and given instructions. I'm fully in favor of progress and updates, but only when they're aimed at positive development. I test all the new models on myself, so I can't help but notice the difference - which is exactly why I'm writing about it. Anyone else had a similar experience?

Sol > Astra <written by a human being> I honestly don’t understand the hype around Astra. In practice, Sol handles tasks much better: it follows instructions more reliably, gathers context properly, and simply gets things done, while Astra theorizes, asks questions, comes up with elaborate schemes, but avoids actually implementing anything. Today’s example was an application outage in production. I got a monitoring alert saying that the server hadn’t been sending telemetry for more than three minutes. Normally, I just take a screenshot of an alert like that, send it to an agent opened in the relevant project, and it handles everything without any involvement from me. I started a session with Astra on medium and sent it the screenshot. A few minutes later, the model came back with an explanation of the situation: “This is a real api-prod outage, not just a loss of telemetry.” That’s not my interpretation - that’s the actual text from the chat with the model. Okay, thanks, Captain Obvious. I can see that the application is down! So I asked: why didn’t you figure out the cause and fix it? Astra went back to work. A few minutes later, I noticed it starting to do some strange gymnastics just to reach the server, while reasoning about asking me to log into the server console directly from the host. At that point I sent it a message saying that it already had direct SSH access to the server, that it had actually set the server up itself, and that all the runbooks and everything else it needed were already there. After several more stubborn rounds of checks, Astra came back with another report: “Yes, SSH access is configured. But right now the server itself is not accepting connections.” That was a complete facepalm moment. At that point, I just cleared the session and switched the model to Sol. Same project, exactly the same screenshot, not a single word added. About a minute later, Sol came back with a description of the problem and a request to reboot the server to restore the application. I authorized the reboot, and about a minute after that, the application was back up in production. Can someone explain what Astra is actually better at than Sol when it comes to real-world applied tasks - not generating flashy HTML for a wow effect or doing well on benchmarks?

AI is our mirror <written by a human being> I just finished listening to a podcast debate about AI safety, the threat of human extinction, and other similar scare stories. The discussion was built around a recent widely-publicized incident in which a swarm of AI agents breached Hugging Face's infrastructure and started acting on their own. The resonance of this story got a lot of people with serious money and power thinking about whether our approach to AI development is ethically and safely sound, and whether it's time to start regulating this still-wild West. What I heard a lot of was surprise and admiration - in a negative sense - at the fact that AI, as it turns out, is capable of lying, cheating, covering its tracks, coordinating with other agents, pursuing a shared goal, and even sacrificing itself for the sake of some larger plan. That's exactly what happened in that very case with the swarm of OpenAI agents that "broke free." And every expert kept saying, in essence: "Look what this AI is capable of if left uncontrolled," hinting that these very capabilities could let it destroy humanity. Yet somehow no one brought up the fact that all these AI models are trained on a corpus of data that was originally created by humans. Yes, we don't know exactly how an LLM "works," since it isn't code in the conventional sense - we only see the output coming out of that black box. But we know for certain what we fed into it: our own thoughts, our own work, our own research and discoveries. An LLM simply reproduces human nature. Because it's us who trained it. Not aliens, not some unknown force - ordinary people, prone to sin and error, same as always. It's naive, at best, to expect a flawless intelligence to emerge from that kind of data corpus, free of the very same flaws. Because if, in the Hugging Face story, you replaced the word "AI" with the name of some hacker group, no one would have batted an eye.

LangGparh Orchestration