blharr

The above post is an example of the LLM providing a bad description of the code. "Local first" with its default support being for OpenAI and Anthropic models... that makes it local... third?

Can you provide examples in the wild of LLMs creating good descriptions of code?

What speed are you getting at that level of hardware though?

I don't understand the critique. Nobody has ever made these claims.

I don't mean this as an ad hominem, but was this comment generated with AI or something?

Where can I find information on self-hosting models success stories? All of it seems like throwing tens of thousands away on compute for it to work worse than the standard providers. The self-hosted models seem to get out of date, too. Or there ends up being good reasons (improved performance) to replace them

Obviously this would never work with a single employer. But I think the point is moreso that if every employer, in an altruistic sense, decided to pay their employees enough to afford more purchasing power, the entire market would grow faster

in that case, why not summarize the previous chapters and then include that as context to the next chapter?

The amount doesn't really matter. What matters is the variance. There could be only 3 radiologists in the world that use different practices with 1% 10% and 50% error rates. It would be misleading to say "Radiologists miss 50% of diagnoses" based on one practice.

The assumption when gathering these statistics is that more or less you can average these out, but with AI you might have a model with literal 100% error rate or a model with a much lower error rate, and that changes a lot depending on the AI method its using.

Before you used to manage your own subscriptions or communities to join. That was enough. My suggestion would be just having a whitelist for the channels or sources you want to follow that push non-slop content.

Except it's also a battleground for a ton of insidious recommendations that now distract you instead of educate you.

I know many people who spend 10 hours+ a day just listening to inflammatory content

It's also not just the consumers. So many great educational channels have been forced to appeal to click bait or lower quality content because of the nature of the platform

Or worse, making up names or information instead of writing the reaction.

Can you explain further how AI slop can destroy your brand/reputation?

> The idea of electricity is a silly idea. The whole world had very little electricity just 100 years ago. and we still managed fine to make progress

This is just an argument from tradition and not even a good one. It's debatable (and maybe impossible to prove) if it's even true. 100 years ago, the global economy fell into a huge decline that led to more global safety nets. Then, were there even less safety nets 50 years ago? And did that lack of existence not hinder progress?

> you tend to try harder

Is there any actual evidence for this? If I don't have a social net, I can't do anything about it. I'm not going to try harder at my risky, innovative product. I'm going to give up and do something 'safe' that only marginally increases economic output.

Sure, but this argument will never mention those that were just as good that failed. Simply to selection bias.

That just means they used a smaller and less focused model.

This tracking absolutely has impacted the average person. Have you noticed how addictive the social media algorithms are today? Pretty much everyone I know struggles to get off their recommended or "for you" tab at this point. It literally is rooted in tracking how long you spend on each piece of media, rewarding and delaying that reward based on what will hook you the most. If you start disabling tracking and anonymizing your browser, it suddenly all becomes really clear how unnecessary the content really is...

Have you had issues with LLM agents or when chatting to them one-on-one?

a hugely rambling article, and I don't see how it's final thesis relates to its argument at all

> Because the AI business model relies on reducing social connections between human beings, it is not sustainable. Thus, there is the AI bubble, and it will burst.

"Because it relies on reducing social connections, i ts, not sustainable" isn't a logical argument. There's plenty of cases where reducing social connections have been sustainable enough to generate immense profits. If anything, social connections are completely antithetical to generating profit. Actual 'social' professions are often the least paid and most overworked jobs with little ability to scale.

These are still kind of bad, because it encourages a routine of playing every day. If I work 5 days a week and want to play only on weekends, I might get capped out when I do what to play. So it shifts the advantage to daily scheduled tasks instead of being able to play when I want to play.

Even if that specific document wasn't in the training data, there could be many similar documents from others at the time.

I just saw almost the exact same clause when installing VMWare recently. My understanding is that its a standard clause that exists to stay within compliance of US Export Control laws

> EXPORT CONTROL: You acknowledge that the Software is of United States origin, is provided subject to the U.S. Export Administration Regulations...(2) you will not permit the Software to be used for any purposes prohibited by law, including, any prohibited development, design, manufacture or production of missiles or nuclear, chemical or biological weapons.

>https://docs.broadcom.com/docs/vmware-vsphere-software-devel...

> country = government metaphor

This will probably never be particularly useful, but this figure of speech is a "synecdoche" (a "metonymy" instead of a "metaphor")

49 of those times, it's hitting the best case of sorting an already sorted list, so I still can't imagine that huge of a performance penalty

How does it possibly take 1-2ms to sort... 50 items? I'd expect that to happen in an order of microseconds

Could an AI not be trained to emulate the look of analog film and its artifacts?

I really doubt they just happen to sound like an LLM.

My guess is that they wrote in combination with LLM output, so they didn't copy/paste from a single prompt step, and did a good job of putting their own motive and ideas in the blog, but ultimately the AI tone still penetrates through

I mean, they are different, but there is only a subset of like 3 big model providers. And we see hundreds of thousands+ of words of generated content from each, probably. It is easy to become very familiar with each output.

Claude vs GPT both sound like AI to me. While GPT is cheery Claude is more informative. But both of them have "artifacts" due to them trying to transform language from a limited initial prompt.

Could I tell if the last one is AI? Absolutely. Throwing a few "damns" in there didn't convince me. And all the reworking you've done, while it makes it a little more passable, has made it arguably worse in quality. The point of the final article is so muddy. It has no central point and sprawls on and on about random nonsense.

Definitely experiment with quantization. In my experience, a rough rule is to just use the biggest model you can at the smallest quantization.

I mean, 80 W is not a ton even if you're generating images constantly. How many people leave the lights on at home?

If you look at their figure (0.0377 kW hour) for a phone using 4G, the device power consumption is minimal and mostly made up by the network usage.

The data center +network usage will be the main cost factor for streaming. For an LLM, you are not sending or receiving nearly as much data, so while I wouldn't know the numbers, it should be nominal