The original gpt4 is just an LLM though, not multimodal, and the training cost for that is still estimated to be over 10x R1’s if you believe the numbers. I think where R 1 is compared to 4o is in so-called reasoning, where you can see the chain of though or internal prompt paths that the model uses to (expensively) produce an output.

The thing is that R1 is being compared to gpt4 or in some cases gpt4o. That model cost OpenAI something like $80M to train, so saying it has roughly equivalent performance for an order of magnitude less cost is not for nothing. DeepSeek also says the model is much cheaper to run for inferencing as well, though I can’t find any figures on that.

Algorithm is just a fancy word for rules to sort by. “New” is an algorithm that says “sort by the timestamp of the submissions”. That one is pretty innocuous, I think. Likewise “Active” which just says “sort by the last time someone commented” (or whatever). “Hot” and “Scaled”, though, involve business logic – rules that don’t have one technically correct solution, but involve decisions and preferences made by people to accomplish a certain aim. Again in Lemmy’s case I don’t think either the “Hot” or “Scaled” algorithms should be too controversial – and if they are, you can review the source code, make comments or a PR for changes, or stand up your own Lemmy instance that does it the way you want to. For walled-garden SM sites like TikTok, Facebook and Twitter/X, though, we don’t know what the logic behind the algorithm says. We can speculate that it’s optimized to keep people using the service for longer, or encouraging them to come back more frequently, but for all intents and purposes those algorithms are black boxes and we have to assume that they’re working only for the benefits of the companies, and not the users.

Algorithms can be useful - and at a certain scale they’re necessary. Just look at Lemmy - even as small as it is there’s already some utility in algorithms like “Active”, “Hot” and “Scaled”, and as the number of communities and instances grows they’ll be even more useful. The trouble starts when there are perverse incentives to drive users toward one type of content or another, which I think is one of the fediverse’s key strengths.

Drat, a missed opportunity!

Protip: Adding “: a study” to virtually any title instantly makes it sound more informed and scientific.

When did we get away from saying “X - formerly known as Twitter” ? I liked seeing that gentle nudge in every headline.

Maybe true, but even at $3500 the Vision Pro would be about the cheapest thing in the operating theater anyway.

It’s actually 1.58bits weirdly. The addition of 0 here was the significant change/improvement in this experiment. The paper isn’t too dense and has some decent tables that explain things fairly accessibly.

I read about 25% of the book this is based on before giving up (too manifesto-y for me, and needed a different/better editor), but the thought of coupling a book with a game like this is pretty interesting — get the point of your argument across to people who might otherwise never engage with it (if not for the title, anyway)