binarymax

There are no more than 218 ways to interpret this title

The real question is why do all UK politicians, no matter their party, hate people so much?

But is OpenAI recouping this? I remember seeing reports a year ago that it was in the realm of $700M/mo in inference costs for them - are they earning that now?

Of course the strategy of taking a loss and reinvesting - but I don't see how OpenAI is making enough money to pay for all this, now or in the future.

Well sure. But maybe the token logprobs can be used to help give a confidence assessment.

Saying “I don’t know” to 30% of queries if it actually doesn’t know, is a feature I want. Otherwise there is zero trust. How do I know that I’m in a 30% wrong or 70% correct situation right now?

A loonnngggg time ago when I was green, and wasn't taught about constraint solving in my State University compsci program, I encountered the problem when trying to help a friend with his idea.

He wanted to make an app to help sports club owners schedule players for the day based on a couple simple rules. I thought this was going to be easy, and failed after not realizing what I was up against. At the time I didn't even know what I didn't know.

I often look back on that as a lesson of my own hubris. And it's helped me a lot when discussing estimates and timelines and expectations.

There's only one thing that would throw me off this email and that is DMARC. But I didn't get the email, so who is to say if I actually would have been caught.

It’s the most popular in the UK because that’s what NHS GPs tell you to take for absolutely everything.

If you live in the US, you're welcome to move to one of the states that doesn't have property tax. But the states that do have it is a choice of the states. Federal doesn't have a say.

GPT-4 was rumored to be trained on 13 trillion tokens. https://www.kdnuggets.com/2023/07/gpt4-details-leaked.html

Not sure where you’re getting the 45Gb number.

Also, Google doesn’t use GPT-4 for summaries. They use a custom version of their Gemini model family.

I approach this from a technical perspective, and have research that shows how Google is unfit for summaries based on their short snippet length in their results [1].

Google also has to support AI summaries for 200k to 500k queries per second. To use a model that is good enough to prevent hallucinations would be too expensive - so they use a bad model since it’s fast and cheap.

Google also loses click through ad revenue when presenting a summary.

All of these factors considered while Google opting for summaries is an absolutely disastrous product decision.

[1] https://maxirwin.com/articles/interleaving-rag/

I didn’t misunderstand. I tried to vibe code, and now I don’t. Not sure how you misinterpreted that.

Even then I’ve mostly given up. I’ve seen LLMs change from snake case to camel case for a single method and leave the rest untouched. I’ve seen them completely fabricate APIs to non existent libraries. I’ve seen them get mathematical formulae completely wrong. I’ve seen it make entire methods for things that are builtins of a library I’m already using.

It’s just not worth it anymore for anything that is part of an actual product.

Occasionally I will still churn out little scripts or methods from scratch that are low risk - but anything that gets to prod is pretty much hand coded again.

I guess I’m an older developer.

But I’ve come full circle and have gone back to hand coding after a couple years of fighting LLMs. I’m tired of coaxing their style and fixing their bugs - some of which are just really dumb and some are devious.

Artisanal hand craft for me!

I think about failure modes. What happens if cloudflare decides you are a bot and you’re not. What recourse do you have? What are the formal mechanisms to ensure a person is not blocked from the majority of the web because cloudflare is a middleman and you are a false positive?

I agree in principle, but I disagree that it should be designed and mandated by a private gatekeeper

This is chaos.

I for the life of me cannot understand why everyone just rolls over. There is no fight. Media, CEOs, Lawmakers, there’s no protest and no pushback. It’s insanity.

The thing about spaCys API is that it perfectly aligns with how NLP worked at the time with actual programming paradigms and allows you to be very pythonic. For example, you can use list comprehension to get all the nouns from a document in a one liner.

These days NLP is quite different, because we look for outcomes rather than iterating over tokens.

What does your NLP library need to do? The way I design APIs is I write the calling code that I want to exist, and then I write the API to make it work. Here’s an example I’ve worked on for LLM integration. I just wanted to be able to get simple answers from an LLM and cast the answer to a type: https://www.npmjs.com/package/llm-primitives

Some low hanging fruit: SpaCy makes an amazing chunking tool for preprocessing text for LLMs.

I’ve had success with fine tuning their transformer model. The issue was that there was only one of them per language, compared to huggingface where you have a choice of many of quality variants that best align with your domain and data.

The SpaCy API is just so nice. I love the ease of iterating over sentences, spans, and tokens and having the enrichment right there. Pipelines are super easy, and patterns are fantastic. It’s just a different use case than BERT.

I’ve been a user of SpaCy since 2016. I haven’t touched it in years and I just picked it up again to develop a new metric for RAG using part of speech coverage.

The API is one of the best ever, and really set the bar high for language tooling.

I’m glad it’s still around and getting updates. I had a bit of trouble integrating it with uv, but nothing too bad.

Thanks to the explosion team for making such an amazing project and keeping it going all these years.

To the new “AI” people in the room: checkout SpaCy, and see how well it works and how fast it chews through text. You might find yourself in a situation where you don’t need to send your data to OpenAI for some small things.

Edit: I almost forgot to add this little nugget of history: one of Huggingfaces first projects was a SpaCy extension for conference resolution. Built before their breakthrough with transformers https://github.com/huggingface/neuralcoref

Democrats should lean on this hard for 2026 midterm elections.

You might be getting downvoted because you editorialized your own title. If it’s obviously not the only thing then don’t add that to the title :)

Classic HN dismissive comment.

The talent displayed here is immense. I challenge you to do better.

Really enjoyed the article and the illustrations. Lots of care went into publishing this.

My primary use case for LLMs are running jobs at scale over an API, and not chat. Yes it's very slow, and it is annoying. Getting a response from GPT-5-mini for <Classify these 50 tokens as true or false> takes 5 seconds, compared to GPT-4o which takes about a second.

Just plus

This doesn't seem to be the case for me. I have access to GPT-5 via chatgpt, and I can also use GPT-4o. All my chat history opens with the originally used model as well.

I'm not saying it's not happening - but perhaps the rollout didn't happen as expected.

I've seen that, but I've also seen (from both sides) a total lack of understanding on what to do that consultants came in to help. I am/have been a search relevance consultant for years - I've seen lots of teams just without a clue how to make things better, and we did. I've also been on the other side where for the life of us we could not decide how to price a product, and had pricing consultants come in and get us out of our paralysis.