TomatoCo

What do you think about putting the shopping cart back in its corral?

In some respects, having your entire reputation on the line matters just as much. And sure, someone might have a server cage in their residence, or maybe they run their own small business and it's there. But the vagueness is troubling, I agree.

A picture of the "living conditions" for the server would go a long way.

I don't think he claimed that "some guy setting up a server rack" is similar to a rapist. I think he's blaming the corporations. I don't think that individuals can have that big of an effect on the environment (outliers like Thomas Midgley Jr. excepted, of course).

I think "you people" is meant to mean the corporations in general, or if any one person is culpable, the CEOs. Who are definitely not just "some guy setting up a server rack."

So LLMs predict the next token. Basically, you train them by taking your training data that's N words long and, for X = 1 to N, and optimizing it to predict token X using tokens 1 to X-1.

There's no reason you couldn't generate training data for a model by getting output from another model. You could even get the probability distribution of output tokens from the source model and train the target model to repeat that probability distribution, instead of a single word. That'd be faster, because instead of it learning to say "Hello!" and "Hi!" from two different examples, one where it says hello and one where it says hi, you'd learn to say both from one example that has a probability distribution of 50% for each output.

Sometimes DeepSeek said it's name is ChatGPT. This could be because they used Q&A pairs from ChatGPT for training or because they scraped conversations other people posted where they were talking to ChatGPT. Or for unknown reasons where the model just decided to respond that way, like mixing up some semantics of wanting to say "I'm an AI" and all the scraped data referring to AI as ChatGPT.

Short of admission or leaks of DeepSeek training data it's hard to tell. Conversely, DeepSeek really went hard into an architecture that is cheap to train, using a lot of weird techniques to optimize their training process for their hardware.

Personally, I think they did. Research shows that a model can be greatly improved with a relatively-small set of high quality Q&A pairs. But I'm not sure the cost evaluation should be influenced that much, because the ChatGPT training price was only paid once, it doesn't have to be repaid for every new model that cribs its answers.

I've heard The Looker described as "If you liked The Witness, you'll like The Looker. If you hated The Witness, you'll love The Looker."

So learnable, in this context, rhymes with reverse-engineerable?

I wonder how it relates to the health warnings on tobacco products?

To continue the carpenter analogy, the issue with LLMs is that the shelf looks great but is structurally unsound. That it looks good on surface inspection makes it harder to tell that the person making it had no idea what they're doing.

If mathematicians aren't occasionally saying something that's obviously wrong to a 300-level student, are they really pushing the envelope? I'm just a programmer, but I find that my, and my coworkers, biggest insights always come right after we've said something seriously dumb.

Half Life 2 recently got a dev commentary track where Valve reflected on their decisions from 20 years ago. One of the things that stuck out to me was that, apparently, Valve called up Microsoft and said "Hey, what percentage of desktops have DirectX 8 compatible graphics cards?" and Microsoft had no idea.

And thus the Steam Hardware Survey was born. The specs automatically sounded a bit anemic to me, too, but seeing them placed on the hardware survey I don't think they're making an outright mistake, per se.

LLM performance depends on doing a lot of math on a lot of different numbers. For example, if your model has 8 billion parameters, and each parameter is one byte, then for 256gb/s you can't do better than 32 tokens per second. So if you try to load a model that's 80 gigs, you only get 3.2 tokens per second, which is kinda bad for something that costs 3-4k.

There's newer models called "Mixture of Experts" that are, say, 120b parameters, but only use 5b parameters per token (the specific parameters are chosen via a much smaller routing model). That is the kind of model that excels on this machine. Unfortunately again, those models work really well when doing hybrid inference, because the GPU can handle the small-but-computationally-complex fully connected layers while the CPU can handle the large-but-computationally-easy expert layers.

This product doesn't really have a niche for inference. For training and prototyping is another story, but I'm a noob on those topics.

Looks like the first digit of each part of the date was incremented by 1. I think it's meant to say 1025-10-19.

I think you could make your fair-use defense a lot stronger by stripping the ROM. Shouldn't even be that hard! Only let it run for a few seconds, record all the memory accesses, zero everything that wasn't hit.

You'll be thrilled to hear what they're doing with higher wattage USB-C cables, then! https://www.startech.com/en-eu/cables/s2cepr2m-usbsl-cable

I'm imagining a text-to-speech model that only speaks in the Transatlantic Accent

> I don't think we'll ever have a zero-crime society, neither should we aim to be one.

This reminded me of https://www.bitsaboutmoney.com/archive/optimal-amount-of-fra...

If you can't make your own protons store bought is okay.

There's a lovely youtube historian, Drachinifel who covered this https://www.youtube.com/watch?app=desktop&v=eQ5Ru7Zu_1I

A cop telling you it's okay to do something, and then getting arrested for it, might be one of the only times you can correctly claim entrapment.

Ah, I think I wasn't clear enough and you misinterpreted me! I'm absolutely not suggesting an LLM for solving the "describe to aliens what our system looks like" problem or doing the assessment of our description.

I was trying to describe how, in the future when we have surveys of thousands of other star systems, it would be fun if there was a website that played the role of the aliens. We would describe to it in plain language what our system looked like and it would tell us if we did a good enough job to get home or not. To me, this means finding how many rows match in a database. I'm not sure how to turn plain language into a database query but, if pressed, today I'd reach for an LLM.

Once more extrasolar surveys are done it would be cool to see how unique we are. If something (possibly LLM based) could rate your description and see how many systems you'd have to visit to find Earth again.

"Eight major planets, the outer four are gas giants. Planets 2 and 3 are nearly the same size. All of the other planets, edge-to-edge, fit just inside the orbit of my planet and its moon."

> She realized how important it was to track the state of variables as they changed, introducing a notation to illustrate those changes.

The thing that really stuck out to me was how similar it was to static single assignment. https://en.wikipedia.org/wiki/Static_single-assignment_form#...

I think this is a state-of-the-art technique today and she had it, what, 180 years ago?

This thought reminds me a lot of the Tesla unit. https://en.wikipedia.org/wiki/Tesla_(unit)

It goes from 3.2 x 10^-5 to 1200 fairly gradually and then suddenly to 10^9. My partner at the time looked at this and remarked "There are some unexplored orders of magnitude" which feels like a good, dramatic phrase.

Maybe it's worthless because it's all already been scraped?

No, I don't really write C. I was remarking on something I saw once and I wanted to share it because the original article mentions trigraphs.

I think my favorite C trigraph was something like

  do_action() ??!??! handle_error()

It almost looks like special error handling syntax but still remains satisfying once you realize it's an || logical-or statement and it's using short circuiting rules to execute handle error if the action returns a non-zero value.

It's amazing what you can do with just gears, cams, and levers. Project-Rho provides a quick survey of mechanisms. https://www.projectrho.com/public_html/rocket/astrodeck.php#...

Well, would the anecdote work if you bumped it back a few centuries and swapped trains for ships?

Obviously the remote code could send the data to whatever three letter agency the operator wants, so the remote operator need to publish the server's code. But how do you prove that the remote operator is running the code they claim they are?

That's what SGX does, it lets remote systems provide a cryptographic proof that they are running certain code. Including the ability to have a private key protected by the SGX, so you can public key encrypt your data, send it to the remote server, and know that only the code they've already published is processing your data.

But sometimes you want to be a massive beacon! https://en.wikipedia.org/wiki/ADM-160_MALD