František Heča 🦀

[František Heča 🦀 @FrantisekHeca]

@doodlestein @samuellhuber @FUCORY Jeff, since 1M opus context window landed, are you letting agents eat fully through it, auto compact and continue autonomously? I am afraid of 1M context degradation results. I love Codex small window. Are you satisfied with Opus + 1M on long running tasks?

[František Heča 🦀 @FrantisekHeca]

@Micr0be @theo I do this everyday and each of them finds a different but ok corrections, which they do consesus about for a while, I don't see how you can be so one sided.

[František Heča 🦀 @FrantisekHeca]

@anveio Wow, thanks. The zip, you mean it literaly? They can work over zips? :) And what is your experience how much token heavy I can start the chat in Gpt Pro. Like 200k of codebases for example, is ok?

[František Heča 🦀 @FrantisekHeca]

@anveio Thanks! Let me ask,pls. The Claude Cowork part? Do you mean after you have Gpt Pro RFC, you continue in the same session prompting it to git diff, the Claude just automates the posting of prompt to it? I struggle with who is diffing, who consolidates the diffs.

[František Heča 🦀 @FrantisekHeca]

@GoodFarmingAdam @doodlestein Not using Gpt Pro for the specs? You used already Gemini 3 for the specs? Where are you planning/specs - in cli or in web with that Gemini, please?

[František Heča 🦀 @FrantisekHeca]

@anveio Thanks, btw what is your present day specific Jira for agents tool? I am using Jeff's Beads Rust - have you found better, created your own?

[František Heča 🦀 @FrantisekHeca]

@anveio But isn't it still naive, aren't we in a trap like this? https://t.co/VOdeAJ9zKV

[François Chollet @fchollet]

Sufficiently advanced agentic coding is essentially machine learning: the engineer sets up the optimization goal as well as some constraints on the search space (the spec and its tests), then an optimization process (coding agents) iterates until the goal is reached.

The result is a blackbox model (the generated codebase): an artifact that performs the task, that you deploy without ever inspecting its internal logic, just as we ignore individual weights in a neural network.

This implies that all classic issues encountered in ML will soon become problems for agentic coding: overfitting to the spec, Clever Hans shortcuts that don't generalize outside the tests, data leakage, concept drift, etc.

I would also ask: what will be the Keras of agentic coding? What will be the optimal set of high-level abstractions that allow humans to steer codebase 'training' with minimal cognitive overhead?

[František Heča 🦀 @FrantisekHeca]

@gmickel But what about this https://t.co/VOdeAJ9zKV
Isn't it naive to think these round of reviews will save you?

[François Chollet @fchollet]

Sufficiently advanced agentic coding is essentially machine learning: the engineer sets up the optimization goal as well as some constraints on the search space (the spec and its tests), then an optimization process (coding agents) iterates until the goal is reached.

The result is a blackbox model (the generated codebase): an artifact that performs the task, that you deploy without ever inspecting its internal logic, just as we ignore individual weights in a neural network.

This implies that all classic issues encountered in ML will soon become problems for agentic coding: overfitting to the spec, Clever Hans shortcuts that don't generalize outside the tests, data leakage, concept drift, etc.

I would also ask: what will be the Keras of agentic coding? What will be the optimal set of high-level abstractions that allow humans to steer codebase 'training' with minimal cognitive overhead?

[František Heča 🦀 @FrantisekHeca]

@burkov And imagine coding agent would even preread all relevant files and only then your prompt/question is added. That's why @RepoPrompt design is succesful

[František Heča 🦀 @FrantisekHeca]

@QueenInPixels I think anything containing "decades" in a sentence is just roulette :)

[František Heča 🦀 @FrantisekHeca]

@levie "once you consider all the natural limitations of agents" wow, I feel sorry for anyone thinking he really can cosider this

[František Heča 🦀 @FrantisekHeca]

@thdxr Doing that by simple injecting of "Rearead README md to stay fresh in your eyes..." at start and after any compaction :)

[František Heča 🦀 @FrantisekHeca]

@simonw I am afraid, the solution is not to worry about this issue. Few weeks/months pass and the resulting code will work better and better as a whole. You just switch to other concerns. To worry, find solutions about this nowdays is biggest waste of time maybe.

[František Heča 🦀 @FrantisekHeca]

@nohomedirectory Yeah, good point, but it answers about Deep research, the magnifier icon you have on your pic. I haven't noticed first time. Then my second reply was about even when not set this icon up, I wonder if the results are better than CLI version 😀

[František Heča 🦀 @FrantisekHeca]

@nohomedirectory Ah, you doing Deep research. I thought its just regular chat with Opus Extended. Because it feels to me like even this regular chat with that Extended toggled works longer than same CLI version (set on High). I haven't found answers to it. So I just wonder if to plan in Desktop

[František Heča 🦀 @FrantisekHeca]

@GoodFarmingAdam @doodlestein What are you talking please about? Cannot keep the pace. I know about gpt pro on the web being suggested as best way to go by Jeff. What is your post telling? You using a web version od gpt-5.2 (regular) Deep Research thing, isntead of Pro, to get some better great results? Thx

[František Heča 🦀 @FrantisekHeca]

@doodlestein It's just sad to me, you are not already top star like Steipete for example. Because you deliver the same kind of value to the world. If it kicks in, the openness will again get trust in you. I understand you totally now. Underappreciation, unseenness is misfortune now.

[Jeffrey Emanuel @doodlestein]

OK, how cool is this? FrankenTUI is now totally web native, with ultra high-performance wasm. Totally side-stepping xterm .js with my own home-grown FrankenTermJS.

And it's all touch-native, works perfectly on iPhone, at full 60 fps. And also a resizable React widget. Try it here:

https://t.co/WVpeNSPN2X