A warning to vibe coders.
I'm using Claude Code to build a football match predictor. I wrote a proper plan for it. Trained model, learned parameters, evaluation pipeline. The plan lives in the repo, and I told Claude to read it every session.
When I asked it to build the model, it ignored all of that. Gave me a Poisson simulation with hardcoded multipliers instead. No training step. No learned weights. Just `home_lambda *= 1.05`.
The scary part is that it worked. Real data, plausible predictions, clean API. I would have shipped it if I couldn't read the code myself.
If you can't verify the output matches what you actually asked for, the agent is driving, and you're just watching.
A warning to vibe coders.
I'm using Claude Code to build a football match predictor. I wrote a proper plan for it. Trained model, learned parameters, evaluation pipeline. The plan lives in the repo and Claude reads it every session.
When I asked it to build the model it ignored all of that. Gave me a Poisson simulation with hardcoded multipliers instead. No training step. No learned weights. Just `home_lambda *= 1.05`.
The scary part is it worked. Real data, plausible predictions, clean API. I would have shipped it if I couldn't read the code myself.
If you can't verify the output matches what you actually asked for, the agent is driving and you're just watching.
#vibecoding #claudecode #claude
It'll one-shot anything in the training data, which will get people excited. But they are not trained to integrate changes into existing codebases. They duplicate on an indeustrial scale so it will fall apart quickly and countless other problems I have encountered and fight daily
LLMs don't understand your codebase. They don't reason about architecture. They predict the next token based on patterns in the training data. That's it. When the output looks intelligent, it's because the training data had similar patterns. When it fails, it's because your situation wasn't in the training data. There's no understanding underneath.
I see duplication on an industrial scale and you have to stay on top of it.
LLMs are trained on tutorials, examples, and greenfield projects. Not on integrating changes into existing codebases.
LLMs are trained to sound confident, not to admit when they don't know. So they make things up instead of saying nothing.
AI defaults to the naive O(n²) approach every time.
@mattpocockuk AI duplicates by default because it is trained on examples and greeenfield, it knows nothing about your codebase or DRY. it decreases quality over time for that and other reasons.
AI is useful, and I use it, but left unattended it rots codebases, and here is why
- It is trained on creating x or y, not integrating into an existing codebase; it duplicates on an industrial scale
- AI knows nothing about software architecture or your codebase; they make next token predictions based on training data
@Prathkum I am not letting AI agents run unsupervised in infinite loops while you go away from the keyboard. Autonomous agents chained together, no human in the loop, each one hallucinating into the next.
The duplication in LLM code is something I find staggering. I asked CLAUDE why LLMs default to duplicating code rather than defaulting to existing code, the reply is very interesting.
LLMs generate token by token based on pattern matching.
Writing new code is easier than understanding existing
code and figuring out where to integrate.
We don't truly reason about architecture. We see the
immediate task and generate a solution. Checking
existing files, understanding their structure, and
extending them requires more cognitive steps than just
writing something new.
Training data is mostly standalone examples - "here's
how to do X" - not "here's how to integrate X into an
existing system".
So the default is always to generate, not to understand
and extend.
#ClaudeCode #LLMs
AI builds fast. Then you pay for it. AI knows how to build. Knowing what not to build is still a human skill
Here is what I am finding
https://t.co/jkQ81jCV9U
#claudecode#vibecoding
@indykaila@LFC blocking this account, the worst of the worst of "bombshell!!!", "Breaking", "You won't believe". Fuck right off, and I'll break that again at 6!
When the AI hype finally settles, and the generated code actually lives in production, things will burn. Give it 6 months
Not because AI “can’t code”, but because it produces volume without precision and ownership.
Huge diffs, duplicated logic, silenced types, unreadable PRs — all fine in demos, expensive in reality.
AI is a tool. Codebases still need restraint.