Yevhen Diachenko

🇺🇦🇳🇱 ML/AI engineer at Amazon

[Yevhen Diachenko @yevhendi]

@something_ua Do you meant when Claude Code is on fire?

[Yevhen Diachenko @yevhendi]

I just got The Welch Labs Illustrated Guide to AI. Some thoughts on the book:

If you already work in the field, you will be familiar with concepts in the book. But you can still have new "aha moments" on things you've used for years.

For people who are new to AI and ML, that's one of the best places to start. It explains ML concepts one by one: building from Rosenblatt’s neurons to LLMs and diffusion models in later chapters. If how AI works is a mystery to you, after this book it won't be.

Also, this is the most well-printed and beautiful tech book I've ever had. Great kudos to the @welchlabs team.

[Yevhen Diachenko @yevhendi]

@scaling01 That is something that can be solved with continuous memory. When a user asks the same question, the model needs memory to notice that this was asked before and stay more consistent

[Yevhen Diachenko @yevhendi]

@daniel_mac8 I feel it when I waste tokens on code style nitpicks like changing variable names

[Yevhen Diachenko @yevhendi]

@tickerplus @sama @OpenAIDevs The $20 plan with big limits is the reason. Claude Code is great, but the only way to use it for real tasks is either through the API or with a $200 plan

[Yevhen Diachenko @yevhendi]

Spotting failed AI coding tasks got much harder.

A year ago, when AI couldn't do something, it meant failed tests, obvious bugs, or complete garbage.

Now, if you reach the limits of the frontier, everything will be green, but with subtle bugs.

[Yevhen Diachenko @yevhendi]

@scaling01 Benchmarks clearly miss the full picture. People keep saying that Chinese models are now close, but honestly, will anybody prefer Kimi K2.6 over Opus 4.5 or GPT 5.2 for real work tasks?

[Yevhen Diachenko @yevhendi]

@kimmonismus In every interview, Dario says, "We've grown 10x every year for the last 3 years, but I don't believe it will happen again". Well...

[Yevhen Diachenko @yevhendi]

@kimmonismus That makes sense. There’s good hardware, but still no ChatGPT moment for a general VLA model. I believe it will happen soon enough with enough investment from top labs.

[Yevhen Diachenko @yevhendi]

@AdamHoltererer It is really good for writing feedback. It is less generic and can highlight non-obvious problems

[Yevhen Diachenko @yevhendi]

@SierraHotel28 @Curline1222 @chrisgpt I tend to agree. I wonder if ARC AGI v4 would be brute-force resistant without the need for a human baselining

[Yevhen Diachenko @yevhendi]

/btw command in Claude Code was so good, so Codex has just copied it and called it /side.

The idea is simple but powerful: you can ask something and discard the answer without putting it into the context. Less context means more time until the next compaction.

As a bonus, if the agent runs the terminal command it won't be interrupted.

The typical btw questions include explaining how a part of the code works, why the agent did something, or what the plan is.

The best command to kill curiosity without derailing the agent.

[Yevhen Diachenko @yevhendi]

@Curline1222 @chrisgpt Yeah, and probably they assume that removing the number of step penalty will open up brute-force possibilities

[Yevhen Diachenko @yevhendi]

@VraserX If it's bidirectional, it means it can interrupt a user when appropriate. Which will really feel like a real conversation

[Yevhen Diachenko @yevhendi]

@NickADobos Good to hear about /side
/btw is extremely valuable for such a simple idea

[Yevhen Diachenko @yevhendi]

@vsuda Driving in Amsterdam is the pure edge case experience. Glad to hear this problem is solvable at all😆

[Yevhen Diachenko @yevhendi]

@tszzl That's why I like coding agents in Tmux

[Yevhen Diachenko @yevhendi]

Today I wrote detailed requirements for Claude Code, switched it to Act mode.

I returned with my coffee after 15 minutes and found that it did not progress because it had stopped almost immediately to ask me for clarification.

The fix is simple: just tell your agent that you will be absent for some time, ask it to work until the task is completed and list all decisions made along the way.

[Yevhen Diachenko @yevhendi]

Another valid answer, but now it's about dragons:

[Yevhen Diachenko @yevhendi]

If you missed it, GPT-5.5 is obsessed with goblins and that's hilarious. Today it gave me a really good and accurate answer, but then finished with: