HN user
silennai.com github.com/SilenNaihin
world-model-harness makes it easy to go from agent traces to faithful replication of your production environment where your agents run.
Basically, an LLM pretends to be a virtual machine executing instructions. Based on GEPA and Qwen AgentWorld.
Just - clone, - `uv sync` - `uv run wmh build` and you'll get a wizard that will help you create your own world model from your traces.
Thanks! I do have a section on this in the article "Why genetic algorithms aren't state of the art"
"Physics simulation involves discontinuities (contacts, friction regimes), long rollouts, and chaotic dynamics where small parameter changes lead to large outcome differences. Even with simulator internals, differentiating through thousands of unstable timesteps would yield noisy, high-variance gradients. Evolution is simpler and more robust for this regime." "The real tradeoff is sample-efficient but complex (RL) vs compute hungry but simple (GA). DQN extracts learning signal from every timestep and assigns credit to individual actions."
DQN likely would have handled this much better.
Simply, it's when your output embedding matrix = input.
You save vocab_dim*model_dim params (ex. 617m for GPT-3).
But the residual stream means that the weight matrices are roughly connected via a matmul, which means they struggle to encode bigrams (commutative property enforces symmetry).
Attention + MLP adds nonlinearity, but it still means less expressivity.
Which is why they aren't SOTA, but are useful in smaller models.
I disregard any comments like this one as baseless hate unless given an example or something constructive.
See what you don't understand is that you need to coordinate the deacon to take the witness out back and talk to the mayor. It's actually quite trivial.
Yes! I went to a meetup for it as well
This is probably true and not emphasized enough. Many are making their value judgements with token limits (which is fair, it is expensive).
Good point. I'll consciously be trying to write more "out of distribution" now.
I think you read my comment, but you didn't really READ it.
That section doesn't include a lick of AI writing. One tell (maybe?) is that I switch from past tense to present tense mid sentence.
Either a) I sound like an LLM when I'm writing articles (possible) or b) turing test AGI something something.
Lists point is fair, I did use Claude for formatting. Where did it off put you here?
Appreciate it! And agree. I added a longer comment above saying basically this.
Outside of the story at the start I intentionally tried to make it information dense. Would appreciate feedback!
Would appreciate feedback, see my comment above!
What sections where you most off put by?
I used Claude to expand on my ideas for a few of the purely informational things, and for formatting, but this article is largely written by hand.
For example "Interface tests are the ability to know what's wrong and explaining it." is in hindsight a confusing sentence. Many such cases.
Try Cursor Composer! It's the most natural transition. Exactly what you're currently doing, but it inserts the code snippets for you from within your IDE.
How can I improve it in the future? The intention was to start as a story, and become more of a structured guide as you read on.
The mayor doesn't communicate to the deacon or refiner well enough. Sometimes my polecat acts like a witness so I have to restart it...
Nothing competes with Opus 4.5 in Claude Code. Codex & Gemini are arguable.
Yes, the technical preview of Github Copilot. I rounded up.
Agree! Don't let it get too bad. See my other comment on debugging and refactoring being a core part of your workflow.
^this
Agree. And with the comments in the thread.
I'll caveat my statement, with AI ready repos. Meaning those with good documentation, good comments (ex. avoiding Chestertons fence), comprehensive interface tests, Sentry, CI/CD, etc.
Established repos are harder because a) the marginal cost of something going wrong is much higher b) there's more dependencies c) this makes it harder to 'comprehensively' ensure the AI didn't mess anything up
I say this in the article
There's no "right answer." The only way to create your best system is to create it yourself by being in the loop. Best is biased by taste and experience. Experiment, iterate, and discover what works for you.
Try pushing the boundary. It's like figuring out the minimum amount of sleep you need. You undersleep and oversleep a couple times, but you end up with a good idea.
To be clear, I'm not advocating for canonical 'vibe coding'. Just that what it means to be a good engineer has changed again. 1) Being able to quickly create a mental map of code at the speed of changes, 2) debugging and refactoring 3) prompting, 4) and ensuring everything works (verifiability) are now the most valuable skills.
We should also focus more on the derivative than our point in time.
Any suggestions on what to add to answer the question better? I tried to cover this in "Why I switched", "When to use Cursor", and "My current setup" sections.
This. Also I'm quoting the email they sent me.