_pastel

The tradeoff is highly contextual; it's not a tradeoff an agent can always make by inspecting the project themselves.

Even within the same project, for a given PR, there are some parts of the codebase I want to modify freely and some that I want fixed to reduce the diff and testing scope.

I try to explain up-front to the agent how aggressively they can modify the existing code and which parts, but I've had mixed success; usually they bias towards a minimal diff even if that means duplication or abusing some abstractions. If anyone has had better success, I'd love to hear your approach.

Bike is the one that gets me - when the full Ode to Joy hits I tear up every time.

You could fine-tune the embedding model to reduce cosine distance on a more specific function.

I work with an extremely effective machine learning engineer, and the biggest thing I've learned is how far you can get with vibes, even in a more traditional ML situation.

He invests most of his time visualizing the inputs and outputs of his systems very carefully, instead of focusing super heavily on metrics; this is a lot more effective than I thought it would be, particularly in the early stages of a project.

Tooling around embeddings has improved. Creating and fine-tuning custom embeddings for your tabular data should be easier and more powerful these days.

100% agree.

One thing that helps is hooking metabase up to its own database and building queries on your queries, e.g.:

    select *
    from report_card 
    where dataset_query ilike '%' || {{query}} || '%'

(You can also join in metadata like the author, when it was last ran, etc.)

We also try really hard to keep the Collection directory structure clean and consistent. But it's still really hard.

Some ways to make recipients feel more comfortable:

- You can suggest some other contribution. "Would you mind bringing snacks? / Would you mind handling music on the drive? / Would you mind giving X a ride?"

- You can allow them to reciprocate in less expensive situations, like taking the check when you are at a cheaper place.

True in this situation, but note that intermediate activations and gradients do take memory and in other contexts that's the limiting factor. For example purely convolutional image networks generally take fixed-size image inputs, and require cropping or downsampling or sliding windows to reach those sizes - despite the convolution memory usage being constant for whatever input image size.

Ah, the celery semantics.

Interesting how Gitlab repeatedly emphasizes it's based on "cost of market", not "cost of living".

So in the sparse grams explanation, what are the bigram weights?

Is it inverse frequency, so common bigrams get split last? And the goal is to be able to search on a larger gram that covers the more common trigrams as often as possible?