stephantul

Love this guy and how committed he is

Thank you! I’ll check that out

Where did you get this from? Searching for it, in a weird irony I guess, just leads me back to this post.

I don’t really believe this is a paradigm shift with regards to train/test splits.

Before LLMs you would do a lot of these things, it’s just become a lot easier to get started and not train. What the author describes is very similar to the standard ml product loop in companies, including it being very difficult to “beat” the incumbent model because it has been overfit on the test set that is used compare the incumbent to your own model.

“Normal search” is generally called bm25 in retrieval papers. Many, if not all, retrieval papers about modeling will use or list bm25 as a baseline. Hope this helps!

I fully agree, except that I think this will still be a very “power user” thing. Perhaps this is also what you mean because you reference Linux. But traditional search will be very important for a very long while, imo

It does not run on Google’s cloud. You can download the model and host it yourself, locally or using a provider you trust.

I recently did some work on making tokenizers greedy.

Hello!

I recently did some experiments in decasing models. I noticed our users not using casing at all, while the base models we use tend to be cased. So I investigated, and found a nice way to decase transformers tokenizers and make them lowercase internally, which I call decasing. I wrote a blog about it!

Hello everyone, we recently benchmarked fastText, an older library by Meta, versus our Model2Vec library and were surprised that Model2Vec is both faster and better.

So if you currently use fastText somewhere, consider switching them out.

Haha yeah, it comes off as pretty insecure, from the looks of it the person writing has no reason to be insecure either.

I think this article misses the main point by focusing on removing pydantic. The main point is that you should convert external types as soon as possible to decouple them from the rest of your code. Whether this involves pydantic or something else is not really important I guess

Pydantic (the company) owns logfire, a logging service. There’s a lot of money in logging/observability. The pydantic library itself is not monetizable, as you indicate.

What a hit piece. The only thing the author seemed to have on his mind while writing it is revenge. Oh, and he’s also selling a course btw

Literal attack vectors

Think of crawlers: a crawler typically makes hundreds or thousands of requests per second. The owners of the crawler then sell this data for X$, or gain X$ profit.

Proof of work adds a very small cost to each individual request, increasing the cost of crawling to a number higher than X. Because actual humans make very few requests, we don’t notice the increase in cost.

“Colleagues”, the first prompt on the main page, is misspelled as “collegues”

Nah not really, the information content is what counts of course. It’s just a bit cringe to see it happen.

This post is 100% rewritten or fully generated by gpt-4o. It has the gpt smell all over it.

It doesn’t break text into chunks at all. These models can handle sequences of arbitrary length.

This is cool! I think that, in general, generating test cases “offline” using an LLM and then running them using regular unit testing also solves this particular issue.

It also might be more transparent and cheaper.

Very cool! Very much looks like one of those projects where you first come up with a fun name/connection and only then write the thing. Was this the case?

The dspy tutorial is amazing. I think dspy is super difficult to understand conceptually, but the tutorial explained it really well

Very curious about this

Very nice! One confusing prompt is “not a valid word” when the word is actually valid, it’s just not the word you are looking for.

During refactoring some stuff in sentence transformers, I dove into overloads and decided to write a blog post about it. Enjoy!

What is the meaning of the word “boxed” in this article? Bit confused about it.

Thanks for responding. I guess I was getting at the fact that 99% of people won’t consistently hit high reader numbers. So to pick this as your goal, or starting point for an article, is dangerous because it just leads you to stop writing at some point.

But fully agree on the advice if the starting point is getting a lot of views/front page.

Counterpoint: writing blog posts so that they are read by someone else completely defeats the point of writing for 99% of people. I do not mean to say that this the advice in the post is bad advice, just that if you focus on being read (i.e., checking rankings on HN, only writing articles that don't exist yet) you probably just will just stop writing at one point, because most of the stuff on the web just isn't read, and writing just to be read is probably not very motivating.

Writing, even if no one reads what you write, is super valuable, and fun! Writing something down is to structure your own thoughts so that you learn more about the topic and about yourself. In my experience, publishing a piece of your writing just ensures that you double check your thinking, but most of the benefit is learning more about what you intend to write about.

So here's my advice: just write posts on what you think is interesting on your personal blog. Don't install analytics, just write it down, publish it, and put it on your LinkedIn. Someone will see it someday and will like it.

Thanks, I will check it out