riskable

We do already have poop emoji gas.

...and burns people's homes down due to lack of safety features.

...and children choke to death from easily removable small parts.

...and people get electrocuted because of a lack of warning label telling them not to use it in the bath.

Mongo is appalled!

Free shipping to send him away? I'll pay that subscription 👍

RVA23 is a big deal because it allows the big players (e.g. Google, Amazon, Meta, OpenAI, Anthropic, and more) to avoid vendor lock-in for their super duper ultra wicked mega tuned-to-fuck-and-back specialty software (not just AI stuff). Basically, they can tune their software to a generic platform to the nth degree and then switch chips later if they want without having to re-work that level of tuning.

The other big reason why RISC-V is a big deal right now is energy efficiency. 40% of a data center's operating cost is cooling. By using right-sized RISC-V chips in their servers they can save a ton of money on cooling. Compare that to say, Intel Xeon where the chips will be wasting energy on zillions of unused extensions and sub-architecture stuff (thank Transmeta for that). Every little unused part of a huge, power hungry chip like a Xeon eats power and generates heat.

Don't forget that vector extensions are also mandatory in RVA23. That's just as big a deal as the virtualization stuff because AI (which heavily relies on vector math) is now the status quo for data center computing.

My prediction is that AI workload enhancements will become a necessary feature in desktops and laptops soon too. But not because of anything Microsoft integrates into their OS and Office suites (e.g. Copilot). It'll be because of Internet search and gaming.

Using an AI to search the Internet is such a vastly superior experience, there's no way anyone is going to want to go back once they've tried it out. Also, in order for it to work well it needs to run queries on the user's behalf locally. Not in Google or Microsoft's cloud.

There's no way end users are going to pay for an inferior product that only serves search results from a single company (e.g. Microsoft's solution—if they ever make one—will for sure use Bing and it would never bother to search multiple engines simultaneously).

You want political toilet paper?

So that's why sales were up.

Some day, a lucky archeologist will unearth the one true archive from an innocent-looking tarball.

Vegan Linux users can compile their own protein from source.

Their purity level is so high that they can kill -9 anyone wearing a leather belt with just a glance.

Not at this point, no. Not unless you know how to setup/manage docker images and have a GPU with at least 16GB of VRAM.

Also, if you're not using Linux forget it. All the AI stuff anyone would want to run is a HUGE pain in the ass to run on Windows. The folks developing these models and the tools to use them are all running Linux. Both on their servers and on their desktops and it's obvious once you start reading the README.md for most of these projects.

Some will have instructions for Windows but they'll either be absolutely enormous or they'll hand wave away the actual complexity, "These instructions assume you know the basics of advanced rocket science and quantum mechanics."

It depends on the size of the content on the page. As long as it's small enough to be contained within the context window, it should do a good job.

But that's all irrelevant since the point of the summary is just to give you a general idea of what's on the page. You'll still get the actual title and whatnot.

Using an LLM to search on your behalf is like using grep to filter out unwanted nonsense. You don't use it like, "I'm feeling lucky" and pray for answers. You still need to go and open the pages in the results to get at what you want.

AI models aren't trained on anything "stolen". When you steal something, the original owner doesn't have it anymore. That's not being pedantic, it's the truth.

Also, if you actually understand how AI training works, you wouldn't even use this sort of analogy in the first place. It's so wrong it's like describing a Flintstones car and saying that's how automobiles work.

Let's say you wrote a book and I used it as part of my AI model (LLM) training set. As my code processes your novel, token-by-token (not word-by-word!), it'll increase or decrease a floating point value by something like 0.001. That's it. That's all that's happening.

To a layman, that makes no sense whatever but it's the truth. How can a huge list of floating point values be used to generate semi-intelligent text? That's the actually really fucking complicated part.

Before you can even use a model you need to tokenize the prompt and then perform an inference step which then gets processed a zillion ways before that .safetensors file (which is the AI model) gets used at all.

When an AI model is outputting text, it's using a random number generator in conjunction with a word prediction algorithm that's based on the floating point values inside the model. It doesn't even "copy" anything. It's literally built upon the back of an RNG!

If an LLM successfully copies something via it's model that is just random chance. The more copies of something that went into its training, the higher the chance of it happening (and that's considered a bug, not a feature).

There's also a problem that can occur on the opposite end: When a single set of tokens gets associated with just one tiny bit of the training set. That's how you can get it to output the same thing relatively consistently when given the same prompt (associated with that set of tokens). This is also considered a bug and AI researchers are always trying to find ways to prevent this sort of thing from happening.