markerz

Speech-to-text models are a lot less open than text-generation models. There's a lot less competition in the space. I can also see this space being enterprise heavy, and could command more dollar premiums. I noticed this when shopping for a STT model for a vibe-coded transcription app. You'd need to pay to get the good ones that support diarisation. There's okay ones that support simple transcription that are open. The good ones charge a premium.

This game is quite interestingly fun despite how basic it is.

The models are likely to check certain places. So either their training overrepresents it or the models were RLHF'ed to go there. I don't think it's that far away from checking StackOverflow for some bug, or reading Wikipedia for some trivia.

Property rights are not the only way to create profit. That's a very capitalistic mindset. Plenty of societies have functioned adequately without our current form of property rights, either physical or intellectual. Before AI, it was the internet that really challenged intellectual property rights. When information flows freely, what happens? Example: many musicians make their pay by performing live, not from royalties on the properties they sold to producing companies.

Intellectual property rights rewards the first-comers, not the innovators. For example, you give a drug researcher a patent on their discovery, that discourages innovation on making that new drug cheap and readily available. The same can be said about AI models: the heavy investments by OpenAI and Anthropic to make these frontier models forces them to charge pretty expensive prices, but if we only had them, we would not have much innovation in making fast tiny models that are more cost and energy efficient. That drive comes from China because they are resource constrained and price-conscious.

Exactly my and top commenter's point. "Temporary price reduction" and "production workloads" are two different worlds.

I'm against the idea that "there's really no point in locking in model choice for anything more than a month or two these days". At a minimum, enterprises are going to lock in a provider for a year due to enterprise contracts, which restricts their model choices. You sign for Anthropic, but now OpenAI models are "better". Or, you signed for AWS Bedrock: Oh no, you don't have access to deepseek-v4 because they're behind.

What about companies purchasing enterprise contracts? Most contracts are minimum 12 months. At a minimum, to secure enteprise requirements like zero-data retention, you'll need to lock into a single provider.

These price reductions are mostly targeted towards self-serve customers on individual or small team plans, where individual choice matters and the friction of changing models/providers is low.

Yeah. I've never done that. I can see the value in it, though.

Personally, I do a Kagi search, or I make an LLM do it for me.

I highly prefer word-of-mouth recommendations like Reddit, HackerNews, or Lobsters. It's quite easy to browse threads where people mention hating some package and switching to something else. It's also a lot easier to see gamified posts (astroturfing), whereas it's really hard to judge projects on GitHub. So many forks are distractions, and stars mean so little. As such, I highly value a personal blog post.

Similarly, I look at what the projects I find link to. They'll often be "inspired by" another project, or a rewrite in a different language. It's a great way to see how different creators consider trade-offs between different projects.

I only browse someone's GitHub when I see an insightful blog post, and I want to learn more about the author and what they work on. Usually I close that tab within 10 seconds cause I can't be bothered to figure out which of their 100 repos is interesting.

My best analogy is that GitHub is the TripAdvisor/Yelp/Google Reviews of finding a good restaurant. Those platforms have been gamified aggressively. The signal is low, and you can't flatten judgement of a restaurant into a 1-5 star scale. Your best bet is to find opinions tied to real people that aren't tied to points.

I had a Japanese teacher a few years ago who explained that Japan's ramen is the US's "burger and fries". It's calorie-rich processed food that's available everywhere. They are prepared in advance, served very quickly (usually within a few minutes), and people finish their meals within 15 minutes. It's far closer to "fast food" than most Westerners realize. The average ramen outside Japan is far more elevated than within Japan.

Are we using GitHub differently? I don’t understand how the social network part of GitHub has to do with having your code discoverable and to discover.

I’ve never stumbled upon a project by browsing GitHub casually. It’s always through finding the product first and trying to find the source code. If the source code is on some other website, I’ll probably figure it out, but it’s be a new muscle. Same with making an issue or a PR. The barrier is higher but not so high that I wouldn’t complain about a bug.

If anything, I find the social tools on GitHub like PR reviews, issue management, and notifications to be terrible for their job. Perfectly fine for a 1 person project and absolutely awful for a team.

In my world, most multilingual speakers are children of immigrant families, and that transcends socio-economic boundaries. Plenty of immigrants were already wealthy before coming to the SF Bay Area, where we import highly educated, specialized workers. On the other hand, we also import physical laborers who are also generally multilingual but not in the same social class.

Some of these immigrants are very well supported with a strong social network, while others struggle with isolation.