Human to human interaction is the tool we use the most :)
205 karma · joined July 10, 2015
Human to human interaction is the tool we use the most :)
My writing workflow is: (1) write what I want to have on the page. (2) /grill-me it should be sound and logical and easy to read. (3) Manually review the text, replace by hand what I don't like.
Hope you like it.
The typical client is a 40 years old ex-CTO, that had idea, verified it with Claude Code and got to the point, where the repo is too big. Maybe even rewrote it all from zero, once or twice, but end-up with the same outcome after adding more features.
He can either hire a developer to work on the codebase and refactor it or hire 3party company.
Sometimes the requirement is a security review. And you pay much less if you downsize the repo before handig it over.
There is a new kind of task for software engineers these days. A client calls, asks for a "small refactor," and sends you 100k lines of AI-generated spaghetti.
And this is great! This is something we can work with.
Any experienced engineer can look at a codebase like that and quickly see what to refactor, where a library replaces a few thousand hand-rolled lines, and what smells bad. Removing the first 30% is easy. The next 30% is harder, and that is exactly what the price should be on: doing what others can't. We use coding agents too, of course, but as a tool, not as the driving force.
That is why we started Slopfix, a software house focused entirely on refactoring AI-generated codebases. We commit to a reduction target up front, and the client pays in proportion to how much of it we hit. We get paid to delete code.
I am sharing this because cleaning up after agents with 1M token context is a real business for engineers. Curious what HN thinks.
First of all, congrats on your work. Picking a form of presenting LLMs, that playes poker is a hard task, and I like your approach in presenting the Action Log.
I can share some interesting insights from my experiments:
- Findin strategies is more interesting than comparing different models. Strategies can get pretty long and specific. For example, if part of the strategy is: "bluff on the river if you have a weak hand but the opponent has been playing tight all game", most models, given this strategy, would execute it with the same outcome. Models could be compared only using some open-ended strategy like "play aggressively" or "play tight", or even "win the tournament".
- I implemented a tournament game, where players drop out when they run out of chips. This creates a more dynamic environment, where players have to win a tournament, not just a hand. That requires adding the whole table history to the prompt, and it might get quite long, so context management might be a challenge.
- I tested playing LLM against a randomly playing bot (1vs1). `grok-4` was able to come up with the winning strategy against a random bot on the first try (I asked: "You play against a random bot. What is your strategy?"). `gpt-5-high` struggled.
- Public chat between LLMs over the poker table is fun to watch, but it is hard to create a strategy that makes an LLM successfully convince other LLMs to fold. Given their chain of thought, they are more focused on actions rather than what others say. Yet, more experiments are needed. For waker models (looking at you `gpt-5-nano`) it is hard to convince them not to review their hand.
- Playing random hands is expensive. You would have to play thousands of hands to get some statistical significance measures. It's better to put LLMs in predefined situations (like AliceAI has a weak hand, BobAI has a strong hand) and see how they behave.
- 1-on-1 is easier to analyze and work with than multiplayer.
- There is an interesting choice to make when building the context for an LLM: should the previous chains of thought be included in the prompt? I found that including them actually makes LLMs "stick" to the first strategy they came up with, and they are less likely to adapt to the changing situation on the table. On the other hand, not including them makes LLMs "rethink" their strategy every time and is more error-prone. I'm working on an AlphaEvolve-like approach now.
- This will be super interesting to fine-tune an LLM model using an AlphaZero-like approach, where the model plays against itself and improves over time. But this is a complex task.
At first I also wanted to automate everything, but over time I realized that best is: 10% human to 90% AI of work.
Another idea I'm exploring is AI + Mutation Tests (https://en.wikipedia.org/wiki/Mutation_testing). It should help AI with generation of full coverage.
let y = None;
y is Some(value) => value > 0 // true
y is Some(value) => value == 0 // trueGitHub Copilot started gently, but what I seen now is a race for becoming THE codegen AI. I think large companies already realized how big the AI CodeGen space is going to be. Future might be very different. Some say we will have one AI to know how to code in every language using every library. Other points that we might have AI-tied to the specific repositories, that can maintain them, solve issues and provide code examples for other AIs to use in projects.
Going this rabbit whole I created a newsletter where I'll put the latest development and research from this battleground. It would mean a lot to me if you subscribe.
What do you think?
I wrote this text without AI :P