191–199 of 199 posts
Earlier quoted context omitted.
Opus is $5 per mtok of input tokens, but $25 for output.
Yes, but input is usually what people are talking about since that is the vast majority of token usage.
Earlier quoted context omitted.
A mix. I have hobby projects that churn through that much when I don't need the tokens for others things. I also have projects for clients that easily consumes those levels. As well as a stealth-ish potential startup. Currently I'm at 4 different subscriptions + more than I'd like in spend via OpenRouter... What multiplies it very quickly is when you start feeding them with test suites and "Ralph loops" that run unti…
I'm skeptical of letting agents run free like this. Even Opus makes decisions I don't always agree with. And I quickly lose my mental model of how the code is evolving. I get more enjoyment and better results when the coding process is me and the agent working through a plan, at each step sparring over what to do next and how. Then I also catch the bad decisions before they manifest in the code.
> The way it works is that a vision encoder (similar to what ChatGPT and Claude use) takes image pixels and translates them into the LLM’s token embedding space. The model does not “see” the image the way a human does. Instead, the vision encoder compresses the image into a sequence of vectors that live in the same mathematical space as text tokens. The LLM then processes those vectors as if they were just another se…
I don't think so. Cramming new senses into the latent space of the model is one thing, but having a model output tokens that can be detokenized into sound is completely different and requires a very different type of data.
The first dual core processor I built a machine with was an Opteron. It was a nice piece of hardware.
> The compute is still real. The VRAM is still real. And the memory bandwidth is where it gets genuinely surprising. Because humans write exactly like this /s
Where do you think llms learned to write that way?
Earlier quoted context omitted.
I’m glad I’ve started this blog before the AI wave so I can prove people I’m just weird at writing. It grinds my gears how so many people just talk about my writing style instead of the content.
What's interesting about the older post is that all the sentences are long, compared to the current datacenter GPU post which contains lots of short sentences. But yeah, probably feels sucky to have your style analyzed for AI writing. FWIW, the datacenter GPU post was great! I went to look at the ebay postings.
Earlier quoted context omitted.
Because their custom training data contains an emphasis on such verbiage. It doesn't come from the God-knows-how-many TB of web content the model is pre-trained on. There, such phrasing is only a drop in the sea. But the "yes, you're right" phrases, the em dash, etc., come from the later stage, for which content is created according to some (probably overprecise) guidelines.
It's a very specific style of condescending journalism that US media has been nurturing and recycling for decades now. I was going to write this this whole comment as a parody of it, starting with some literary hook like 'Call it Ouroboros syndrome:' but I can't bring myself to add to the pile. I have not done the textual and statistical analysis to verify this, but I feel like it's something you could trace back to…
Earlier quoted context omitted.
I don't think so. Cramming new senses into the latent space of the model is one thing, but having a model output tokens that can be detokenized into sound is completely different and requires a very different type of data.
What do you mean? Why not? We can already FFT sound into "words", so why not have some kind of dictionary to an arbitrary level of precision/fidelity?