Max Chernyak

Aforeposted meme aside, I think the new reality of building a viable business is: you need to offer something more tangible than software. A pass through a real barrier, like proprietary access to something, red tape, unique knowledge. Tech glue alone is a losing battle.

Opus just wrote me a script without using any of my coding, docs, and review procedures. I asked why, it said "I treated them as optional, even though it says _always_". This is in Claude Code. I don't think it ever did that in Pi via API. 🤔

Just loaded up my "dot" with a bunch of chores that I never got to tackle before. Let's see how it goes. The cool thing about a managed product is that I don't have to think about software updates and memory management.

Technical barriers are lowering now, so app-only businesses are losing value prop. However, there are still physical world, red tape, legal, regulatory barriers. If your business reaches into that, it is more likely to remain standing.

Everyone: “ai is good enough now” Also everyone: “let’s see who’s the dev building this slop to make sure it’s well made”

I wonder if the same people who never read code also never read their agent skills and configs, and end up with bigger and bigger mess that probably works worse than if they just removed all of it.

5% AI failure rate can feel worse than 20% human one. I think it's because the failures don't overlap. For humans it feels extra super bad to see an AI fail where you wouldn't have. Extra super duper bad when talking about self-driving cars.

That’s what I call solution hoarding.

One of the biggest insights I've picked up later in life: everything you own owns you back. It takes maintenance, and mental space. The more complex and obscure it is, the more it owns you. This is just as true of the software we build, especially now.

I don't want to build a million things with AI, because I'm not a solutions hoarder. I've grown to resist the urge to solve all perceived problems, because only a few of them are real, and almost never what they seem at first. Clutter taxes my mental well being.

"Stop bottlenecking development by reading code" say people who work on fun/convenience utilities and games. "We must review code" say people who work on finance, medical, low-level infra. Why so godddamn polarized? Read SOME code, use the right amount of due diligence per task.

There is still value for high level PLs and frameworks in AI world, but it’s mostly in critical industries (finance, medical, physical), where proving the outcome is not enough, you must also prove the steps taken. Clear code is that proof.

Removed a fancy work monitor from my wishlist because the way I interact with code is changing too fast.

The more I say to AI, the worse it gets at what I don't say. This skews it away from balanced decision making. Feels like picking stocks to try to win against the market.

I find it cute how ChatGPT suggests written confirmations for everything. Wonder what'd happen if I actually requested these. "Can I have a written confirmation that your kids pool will be heated?" "Can I have a written confirmation that you serve milk at your restaurant?"

Annoying that AI consistently messes up the changeset reference point for changelogs, commits, etc. I often have to remove notes about fixes to issues that were never published, but AI keeps documenting them. Wonder if this could be handled by a skill. 🤔

PLs are a workaround for human cognition. If we stop reading code, we could achieve maximum efficiency by having agents program at the lowest level. This would obsolete almost all PLs and unroll hundreds of abstraction layers, which could free up a ton of context. 🤷‍♂️

My motivation to blog about code maintainability has been dead since AI took off. Does it even matter anymore? Same goes for designing PLs. But then, how will AI know what is/isn't maintainable for humans going forward? Aren't humans still the backup plan?

Decided to finally blog about Ringbinder. It's been such a useful little tool, and I feel like README alone doesn't do it justice. max.engineer/ringbinder

Ringbinder 0.4.0 is out with some awesome new updates. It now supports Gemini 3.8 Flash for all your docs OCR, and can optionally fallback to Mistral. It also supports Gemini Batch API which makes OCR half price! github.com/maxim/ringbi...

Warning: a bitter kind of commentary on the performative social media "life lesson"-flavored marketing. max.engineer/human-slop

Current workflow: Pi via Paseo. Sol for planning -> Kimi K3 (Opencode Go) for plan review -> Sol for implementation -> Sol + Kimi K3 for final review -> Sol for fixes. Luna for explore subagent. Sol/terra/luna auto-choice for general subagents. No Claude.

For the first time my iPhone feels like a production rather than a consumption device, thanks to Paseo and tailscale. First time I didn’t take the laptop, because I prefer it plugged in at home, controlling agents remotely, without worrying about closing the lid.

My 8yo inspired an experiment. He wanted to listen to news like me, but more for his age. So we made an AI-generated news podcast for kids. So far this has been quite fun to follow. myfirst.news

The less programming skill you got, the lower is the AI intelligence threshold for you to abandon code review on serious projects, and the easier it will be psychologically.

If "code was never the hard part", why not everything else AI is swallowing? "Knowing a language was never the hard part". "Understanding math was never the hard part". "Knowledge of medicine was never the hard part". Only application is hard, right?

Why even upgrade, my old iPhone 13 Pro gets better every day. Guide Pi’s via Paseo while watching casts/vids in pip, researching in ChatGPT, checking news, tracking steps, snapping pics/vids, or taking calls, or playing a game. Kinda surreal how much you can do on the go.

Been looking for a way to control agents via my iPhone. Must haves: tailscale to my existing Mac Pi setup, but native iOS UI. And no forced Mac GUI. Herdr: 👎️ (TUI on mobile sucks) Paseo: 👍️ (daemon only on Mac) One wish for Paseo: let me connect to a live Pi session.

I like using pi with mcporter, but with stateful MCPs you either get one daemon for all pi sessions, or new mcp on each call. This pi extension gives each pi its own stateful "lane" with mcporter. github.com/maxim/mcport... - so e.g. multiple safari-mcp don't stumble on each other.

Is everyone just sprinkilng typos into their writnig now to show that they aren't AI? And then to hide that they are actualy AI, they ask AI to edit in some typos? Not sure I'm a fan of wehre this is gong.

"Narrator" (a pi extension I just made) uses a cheap model (DeepSeek 4 Flash in my case) to asynchronously narrate what's happening in pi, CRPG combat-log style. So much more fun to follow the agent this way.

Really liking the mini weekend projects I get to complete now. Like a staging/video pipeline for syncing kids movies and shows onto an SD card, formatted for the minivan tv. Or the iterm2 status bar pi session summarizer (to remind what this session is about).

Current setup: Codex plan, mostly GPT5.5-xhigh. Using pi (+custom ext/skills) on iTerm2, lazygits in splits. Obsidian on the side. Codex-5.3-spark for misc work (e.g. summarize pi session in iTerm status bar). /grill-me into obsidian, then build. Opencode Go models sometimes.

Removed from pi: ask-user-question, btw, supervisor, subagents, tasks. Paid for 5x codex plan. Installed herdr. Done with experimenting for now: clean context, one good model, simple guided process, iPhone access, let's go.

Created a well-grilled plan with GPT 5.5 xhigh, wanted to save by asking GLM 5.2 to implement it, got a broken solution, spent twice as much asking GPT 5.5 to fix the result. 😬

After using ponytail in pi (with GPT-5.5 xhigh) for a couple of days, I'm seeing much higher context usage, and consistently worse outcomes in both plan+implement. Feels like I just got an upgrade by uninstalling.

Got GPT-5.5 to build me a bunch of custom rubocops that are full of subjectivity. Probably the biggest step towards getting AI to code like me. Recommend.

As I'm sitting here, on question 31 of a /grill-me prompt, having asked GPT 5.5 to "make db path configurable", I can't help but marvel at the endless void of pedantic decisions we devs have been expected to make daily. Hard to imagine a non-dev surviving a /grill-me session.

I scan and shred lots of docs (kids' schoolwork, mail, etc), so wanted a simple way to let AI at them. Vibed this up and been using it happily for a couple months now. Maybe useful to someone else. Contrib welcome! github.com/maxim/ringbi...

Tried to have AI read my posts and codify the style, then made several different LLMs try to write as me. The result is unfortunately beyond cringe. Weirdly, simpler models (e.g. Qwen 3.6-plus) came the closest, but it's still painful to read. Hard nope.

A new way to recognize AI-built web frontends: it's broken in Safari. As a Safari user, this is getting old fast. I thought AI should help cover more ground for you?

Using plannotator with pi lately (liking it so far). Wrote a seeds file for my Rails app. Code review from Claude: "here're a couple of nits". Code review from Codex: "you're about to break production, because you're doing db:prepare which also runs your seeds".

Something's missing in these gstack-style approaches to AI. Why reupload millions of tokens of its own slop back into it? It's in there already. All it needs is what it doesn't have: your data, and some steering to apply its knowledge according to your preferences. Use pi too.

Added pi to the list. Unironically considering it as the base for AI setup. bsky.app/profile/max....

PSA for those caught in the AI hamster wheel, something I wrote in my college notebook 22 years ago, still holds up: Happiness is a cause, not an effect. Get happy now, in your mind, then do things. You don't work towards it, you start with it, and it enables everything else.

Here's where the puck is going: AI will replace software. You ask it to solve problems, you instantly get custom apps. All computing will become this. We will be sharing T=0 prompts that produce useful results. All software engineering will become about improving AI itself.

My attempt at deciding on an AI orchestrator. So many out there. Suggestions/corrections welcome. gist.github.com/maxim/c6ef7f...

If you're gonna embrace non-determinism at all stages of software dev, including verification, why even make software? Just have AI be the software. Why have an "if" condition when AI can decide. Unironically, that's where we're going. It's moving ground all the way down.

Of all the AI artifact-managing solutions I've seen so far, beads is the one that nails it.