Igor Kolomiets

@HeyTolstoy I won’t be surprised if rewriting Bun from Rust back to Zig by agent (would be nice to see Codex in action) will bring even more benefits, because 1) SOTA LLMs in general are much better at programming than humans 2) Zig could be better for high-performance optimizations

A few thoughts on what Bun migration from Zig to Rust means for programming languages. Language choice is now an autoresearch hypothesis Historically language choice was mostly driven by subjective stuff. Personal taste – you like the language syntax and concepts or you don’t. Talent availability – pick e.g. OCaml and you’ve just capped your hiring forever. Meanwhile, objective factors like performance were hard to factor in. Synthetic benchmarks aren't really representative. Until you’ve actually built the system, you don’t know where your bottlenecks will be, and whether you’ll hit the performance ceiling of, say, JVM. That whole picture is flipping. When agents write the code, the human UX of a language stops mattering. Team experience matters significantly less – you don't need everyone to be experts. But objective evaluation actually becomes possible! Write the system, hit performance issues, profile them, rewrite it in another language, measure again. Language choice gets reduced to just another hypothesis in the autoresearch loop – a bit expensive in tokens, but that’s not a huge issue. On the evolution of language ecosystems Choice of programming language has always been one of the strongest possible vendor lock-ins. You pick a language, you stay with it forever and die on that hill if needed. And that was a driver for fixing the weaknesses of specific ecosystems. Shopify picked Ruby, which had no decent IDE support in VSCode? Then Shopify spins up a team to build that plugin and then open sources it. This kind of corporate open source pushed language ecosystems forward – such projects ended up pretty mature, with much lower maintainer burnout risk and company resources behind them. This actually was the core of our strategy for Kotlin Multiplatform. We optimised the technology for large teams with huge codebases, expecting them to invest in open source libraries and tooling in return. It worked greatly, and helped us to catch up with more mature ecosystems, despite having less users overall. If the economics of migration changes, that’s definitely going to affect this picture. Why pour engineering into fixing your ecosystem’s gaps when you can just leave? Where this ends up The migration cost is dropping. The motivation to invest in existing ecosystems is falling with it. We’re going to see a lot more rewrites like this. The languages that win long-term will be the ones that win on the metrics agents actually care about – runtime performance and agent-oriented UX, meaning strong compiler checks, simple build tooling, clear error messages, hard-to-misuse APIs. The endgame is stagnation of the majority of languages, with just a few of them staying relevant.

@braelyn_ai @relizarov @relizarov agree about TS. What is your take on CLI in Kotlin Native?

there is absolutely no reason to build your CLI or TUI in typescript. please stop

@relizarov @JamesWard So the latest change in Zig’s IO in its stdlib that requires passing an explicit IO parameter (context) is “co-effects”? Good to know.

@JamesWard To give a simple example: effect is declaring “this function can do IO”, then handling or propagating this effect on call sites. Co-effect is declaring “you can do IO in this context”, then requiring context parameters for functions that need to do IO.

It’s no surprise that many of the new AI Native programming languages are Effect Oriented (Algebraic Effects). If you are creating a new AI Native programming language, why would you go any other route??

@kotlin I'm interested in learning why GraalVM and not Kotlin Native? I evaluated Kotlin for AWS Lambda a few months ago. I wish there existed a stable Kotlin AWS SDK with Kotlin Native target.

How much can a two-person team accomplish with a 100% Kotlin stack? 📱 93% shared mobile code with Compose Multiplatform ⚡️ Fast, GraalVM-friendly Ktor backend 🤖 AI pipeline built with Koog Check out this awesome community deep-dive: @ofekteken/how-we-built-ishape-using-100-kotlin-43c3941aec83" target="_blank" rel="nofollow noopener">medium.com/@ofekteken/how…

@headinthebox So many in my feed are praising Codex for one-shotting tasks where Claude 4.7 just keeps looping. I would be interested to learn how Codex stands up against Claude on your tasks.

"... The disciplined way to find it is the ... I never finished ... It's a half-day of work but it ends with bit-close parity rather than guesses ..." If Claude Code keeps just fucking around trying random shit instead of doing what it promised, and continues to be as slow as a snail ("Cooked for 1h 22m 54s"), that estimate of a half-day might actually be accurate. On a more serious note, I feel I was more productive in January 2023 copying and pasting code between IntelliJ and ChatGPT 3.5 than with the current Claude Code. Sad, very sad.

@antirez This week Kimi’s team shared how they used their latest K2.6 to automatically implement inference though their custom harness for Qwen 3.5-0.8B from scratch in Zig. After 12 hours and 14 iteration they managed to outperform vLLM by 20%. Any idea why Zig?

Is it too cruel to let GPT 5.5 implement inference for DeepSeek v4 in llama.cpp? Since that's what I'm dong right now.

@xlab_os @rtroar @rikarends I’m new to Zig, but I’m stunned with what I’m able to do with it in Codex. Also, LLMs are great at explaining and effectively teaching idiomatic Zig.

@rtroar @rikarends I tried in Dec '25, probably too early. Can certainly give it another go, but it still will be a lazer focused use case and not general "vibe coding gui apps" experience I want..

Zig is not a memory safe language and does nothing to protect the runtime from mistakes with memory. I don't get the trend to make another "sexy" language out of it just to rewrite all existing software. Yes it is fast, but it also crashes fast. Bun crashes, Ghostty crashes, I don't want to know the name of next big thing that is written in Zig and also crashes. In 2026 many segfaults or memory corruption errors can be weaponized to exploit systems. If you take slow but safe software and rewrite it into Zig you are actually working backwards. LLMs are also terrible at Zig because there isn't much code base and documented mistakes to learn from. One of my few exploration attempts was about making a CLI app with a buffer and mouse handling. It was one-shotted okay, but upon launch it just started to print raw bytes from my RAM onto the buffer. Basically exposing adjacent apps state. Zig might sound cool for people coming from the web world (where everything runs in the most safe sandbox imaginable) or even non-CS-backgrounded people who just pick a language that is popular on X, but this is a really bad bad option for anything with >0 users. Before LLMs - maybe, but not in today world. If you want raw power - write in C, ASM, use ASan. Also, Rust is not perfect but at least it offers something.

@antirez @OpenAIDevs I’m on the cheapest “Go” plan. You can do surprisingly a lot with it. Recently, I did my taxes in Codex - it found all legit expenses and calculated all the numbers. It even wrote python scripts and tests for this 🤯

@ikolomiets @OpenAIDevs Usually for Redis / other open source work I never reach my Pro plan limit, not even near. But this reverse engineering requires Codex to read a lot of disassembled code, this requires a lot of tokens, very long context as well, so...

During the last week I executed very long autonomous sessions of Claude Code Opus 4.6 and Codex GPT 5.4 (both at max thinking budget), in cloned directories (refreshed every time one was behind). I burned a lot of (flat rate, my OSS free account + my PRO account) of tokens...

@antirez @OpenAIDevs I’m waiting until Wednesday for my token limit reset in codex too 😭

Now, unfortunately, my @OpenAIDevs tokens are gone for the next 48h :D So the reverse engineering must end here, I need the left tokens for Redis work. Will resume in two days.

@PAHoyeck Can this proof be formally verified using Lean?

A valid and sound proof of God's existence: P1. If God does not exist, then it’s not the case that if I pray my prayers will be answered. P2. I do not pray. C. Therefore, God exists.

@nntaleb I didn’t know this. x.com/i/grok/share/e…

Predictably, a victim of AI is Wikipedia. I picked mathematical subjects that may have an October seasonality, but are not affected by fashion/current events. Same for Math StackExchange. This, considering that a lot of the math learned by AI came from Wikipedia and MSE.

@nntaleb @k_sonin You won’t find intellectuals among professional “Kremlinologists”.

I wonder if this "distinguished" professor of pol science @k_sonin, when AI takes his bullshit job away, will have the mental abilities to qualify as substitute food taster at Costco.

@nntaleb Same is true for any diaspora from “authoritarian” states.

A perceptual bias: because the Iranian diaspora is naturally more heavily anti-regime than residents of Iran (people vote with their feet), some decision-makers may overestimate potential domestic support for the strikes.

@axboe @jorandirkgreef @TigerBeetleDB

In 2 days...

@skdh This doesn’t seem to bother IBM that brags to investors about imminent availability of their latest and greatest quantum processors. I wonder if they try to pull same PR trick they did with “Watson AI” more than a decade ago.

Today I talk about how more quantum computing use cases are disappearing, and an unexpected problem with quantum computing in general. youtube.com/watch?v=N-9muK…

@thegeeknarrator @jorandirkgreef @TigerBeetleDB When and what asserts to disable in PROD release?

Recording a new pod on “assertions in production systems” with matklad (from @TigerBeetleDB) in a couple of hours. Any questions? Drop here 👇 I will try my best to cover them.

@thegeeknarrator @jorandirkgreef @TigerBeetleDB When will episode on TigerStyle with matklad be available from Apple Podcasts?

@jorandirkgreef @jedisct1 Oh, so this is how TigerBeetle plans to support cryptography while renaming true to its principle of having zero-dependences!

This is something I’ve always deeply appreciated: @jedisct1 and others’ work in making Zig’s stdlib cryptography the vanguard.

@antirez This would be French classic 😂

I agree with Macron words but not with the sunglasses style (I understand it's for an eyes condition, but, go classic, dude). It is exactly the kind of situation where you want to look European and not the fucking Top Gun.

@omooretweets @grok what is WAU?

More than 5% of ChatGPT messages sent globally are about healthcare - and 25% of WAUs ask health Qs. (per OpenAI's new "AI as a Healthcare Ally" report) Usage is higher at times where doctors' offices are closed, and in "hospital deserts" where access is limited 👇