paretolaw

If claude only did what we tell him...

Why almighty openAi doesn't solve PvNP problem :(

I guess solution had not yet appeared in training set.

I got one better puzzle. Predict Jane algo moves when they try to manipulate market and frontrun their orders.

Let them taste their own medicine. :)

Can you please share your exhaustive research supporting the predictions?

What else should they follow? Would recommend reading synthesis of form.

MS lawyers might visit you soon

So, long samsung and short nvidia

Would you rather be doing GB with 70s hardware or TB today?

I would not understimate what folks in 70s had done.

Tried Cline with a local Qwen2.5-Coder 14B model through llama.cpp.

Same story. The model works perfectly when called directly, but the agent integration doesn't actually execute the tools. It generates things like:

<tools> {"name": "read_files", ...} </tools>

as plain text and then just sits there until I tell it to "do it" again. It repeats the exact same fake tool call.

So instead of:

Qwen → tool call → grep/find/read → result → Qwen

I get:

Qwen → pretend tool call → nothing happens → repeat.

Aider-AI/aider review: how many times do I need to click D here to stop being prompted this each time I open session?

> Open documentation url for more info? (Y)es/(N)o/(D)on't ask again [Yes]: D

Update: after actually trying Aider, I don't understand the point of it.

It feels like browser AI chat moved into a terminal, except now it's on your own dime, with no freemium. Ask it `find test_main location` and it won't actually run `find` and inspect the repo. It tells you it can't access files you haven't explicitly added.

That's basically browser chat wrapped in a terminal with a few pre-existing skills.

I hoped anomalyco/opencode would be a decent Claude Code alternative, and that it would work with a relatively small model like `qwen2.5-coder-14b-instruct-q5_k_m` for simple coding tasks.

Unfortunately, my experience was extremely disappointing. This happened on Fedora.

Instead of responding to simple prompts, it got stuck in what appeared to be an infinite loop producing nonsense like:

> Objection > work state > active > blocked > next move...

Even the response-copying functionality didn't work, so I couldn't easily copy the output. (Using Terminator Terminal Emulator)

And no, the model itself isn't the problem. I manually called the exact same llama.cpp OpenAI-compatible endpoint with `reply with Hi`, and it correctly replied `Hi`.

I also checked the OpenCode configuration and corrected the model ID to exactly match what the OpenAI-compatible `/v1/models` endpoint reports. Same result: endless nonsense.

So something in the OpenCode integration/agent prompting is clearly going wrong in this setup.

And please don't tell me to "read the docs." I've never read a single line of Claude Code documentation and have been using Claude Code without issues for several months. If Claude can make the experience that seamless, I'd expect a project with 200k stars to provide at least a usable experience.

Maybe I'm missing something obvious, but I wasted ~20 minutes trying to get a basic setup working.

Posting this in the hope it saves someone else 20 minutes.

"...and then train..." and those dummies in openai and google wasting billions, while all they need is some 17 year old and few pizzas

Hi agent, please maintain this fork :)