Hes supposed to be a credible source, hes an educator, hes the modern day Bill Nye. It’s very hard to take anything he says as objective truth when his research and fact checking consists of asking an LLM. Considering hallucinations and straight up lies are not only common but not even unexpected for an LLM you can’t believe anything that comes from one nor can you believe anyone who uses one.
Hes supposed to be a credible source, hes an educator, hes the modern day Bill Nye.
There are no credible sources on YouTube. He’s just a guy with a computer and a camera. Youtube is about selling you stuff, people need to accept that.
He didn’t state what his actual process was. People assume that just means firing out a query and no bothering to check it, he said he uses it to summarize sources to see if they are worth reading. I don’t think that’s unreasonable if he still reads them.
Regarding using it to write scripts, I think if you use LLMs to generate something you are good at making and are picky and opinionated about, the review process very quickly makes you skeptical about the output.
Before the response, IP rights and environmental effects of LLMs are completely valid and are a valid reason for opposing them. Don’t just assume someone’s using it out of laziness instead of in an attempt to scale up their own work.
I read a decent number of papers. If you’re looking for something specific, and abstract isn’t enough. If you want to research something specific that’s not the core focus of the paper, you at least need to skim it.
If Hank Greene is understaffed, using AI to prefilter papers isn’t a terrible use of the technology.
Creating science Youtube videos is not more important than maintaining a livable planet. AI is unnecessary and harmful, and should not be used outside of certain brute-force type of applications like protein folding, simulating new drug structures, or preliminary cancer diagnosis.
Golf courses use 30x the water that data centers use. Lawns, 100x. Standby mode for appliances costs more energy, by a lot. I wonder how much carbon emissions are offset by people no longer driving to work at call centers and the like.
Out of curiosity do you eat meat? .2% of water is used on data centers. 9.8% is used on beef. For carbon emissions one pound of beef consumed is equivalent to about 640,000 prompts.
AI is an environmental concern but the way people act about it is straight up ridiculous.
Edit: I am very proud of all of the vegans who unplug their appliances who are downvoting me. I, too, am a vegan who unplugs their appliances.
and yes I am vegan, but eating meat is, in many cases, a culturally ingrained practice which people need time to understand the harms of. AI is not culturally ingrained for anyone with an actual culture. People need to eat - it shouldn’t be meat, but I’m not going to shame people for not seeing their harm yet. People do not need to generate slop. I am going to bully all AI users.
Do you think AI is mostly generating slop? Only like 5% of AI use is private individuals.
The point is that this is not actually an ecological hazard the way people say it is, and that there are much, much more pressing issues if we’re going to fix the environment. Literally all we would have to do to solve global warming is stop eating meat. You see anywhere near as much about that as you do datacenters? Which cause less than one percent of the environmental issues of meat?
I prefer to self host, and am in the process of setting up my entire system to run off solar now (tiny 2x250 watt solar array, 3 deep cycle 100Ah marine batteries and sundry equipment).
End to end, my AI rig - a BC-250 (a cut down PS5, basically, though I’m thinking of down sizing) consumes 50w standby / 200w on inference and can run Qwen 3.6-27b (Ternary Bonsai specifically, about par with Sonnet 4.6 on the AAII) at good speeds (~30 tok/s)
With a smart power plug and Raspberry pi (uses about 5w) as the pilot fish, I can run my entire system “off grid”, to say nothing of the excess solar that accumulates (several days battery back up to run home router, pi server, hard drive array and 2-3hrs daily of inference).
For $1500 AUD, “Poor man’s Sonnet 4.6 at home” ain’t nothing to sneeze at.
Outside of folks such as myself, this blog post may be of interest -
no its true. ChatGPT is confidently wrong all the time, but I find its wrong in ways I can learn from anyway. I ask the question. it gets the answer wrong. But its wrong in a useful way where I actually learn more by correcting it than just getting the right answer.
I could get the same experience from talking to a person for much less money though
They tend to work much better when you point them at things. Hallucination rates go way, way down when they aren’t pulling from their “memory.” They are, also, like you say, wrong in ways that are noticeable and often “interesting” in the Ibsen way. I also doubt that you could get the same experience for less money. I’d have to go back to college to get the education I’m getting with AI. At least community college. It’s a lot of near-experts at once. And when you point it at expert literature it basically is a senile college professor. But yeah the senile part of that statement is pretty important.
I find that a frontier bot pointed at the internet does a lot better job of answering the question than I do when searching on my own. I mean they’re always going to be wrong about some stuff, but aren’t people? I think that’s generally a better metric. And at this point they’re far away from what they were just a few years ago. But you’re correct in that there is an element of brute forcing them or at least shaping them that has to occur that I can’t blame anyone for tapping out of.
I could have always skipped the dewey decimal card catalog and just read the entire library but I’m not mycroft fucking holmes. You don’t have to use a search engine, you can just follow hyperlinks until you find the content you’re looking for.
I like Hank, and this is a shitty article.
That aside he should be facing a revolt.
Hes supposed to be a credible source, hes an educator, hes the modern day Bill Nye. It’s very hard to take anything he says as objective truth when his research and fact checking consists of asking an LLM. Considering hallucinations and straight up lies are not only common but not even unexpected for an LLM you can’t believe anything that comes from one nor can you believe anyone who uses one.
There are no credible sources on YouTube. He’s just a guy with a computer and a camera. Youtube is about selling you stuff, people need to accept that.
He didn’t state what his actual process was. People assume that just means firing out a query and no bothering to check it, he said he uses it to summarize sources to see if they are worth reading. I don’t think that’s unreasonable if he still reads them.
Regarding using it to write scripts, I think if you use LLMs to generate something you are good at making and are picky and opinionated about, the review process very quickly makes you skeptical about the output.
Before the response, IP rights and environmental effects of LLMs are completely valid and are a valid reason for opposing them. Don’t just assume someone’s using it out of laziness instead of in an attempt to scale up their own work.
Papers have an abstract and summary sections SPECIFICALLY for this reason and usually aren’t more than a paragraph; there’s no excuse to use LLM.
I read a decent number of papers. If you’re looking for something specific, and abstract isn’t enough. If you want to research something specific that’s not the core focus of the paper, you at least need to skim it.
If Hank Greene is understaffed, using AI to prefilter papers isn’t a terrible use of the technology.
I mean this is true of all sources and they should be checked.
It’s not from my understanding he was using AI for semantic search to find paper to research
I don’t know about going that far - LLMs can be a great starting point …. As long as you click through to the sources rather than blindly accept
Creating science Youtube videos is not more important than maintaining a livable planet. AI is unnecessary and harmful, and should not be used outside of certain brute-force type of applications like protein folding, simulating new drug structures, or preliminary cancer diagnosis.
Golf courses use 30x the water that data centers use. Lawns, 100x. Standby mode for appliances costs more energy, by a lot. I wonder how much carbon emissions are offset by people no longer driving to work at call centers and the like.
Out of curiosity do you eat meat? .2% of water is used on data centers. 9.8% is used on beef. For carbon emissions one pound of beef consumed is equivalent to about 640,000 prompts.
AI is an environmental concern but the way people act about it is straight up ridiculous.
Edit: I am very proud of all of the vegans who unplug their appliances who are downvoting me. I, too, am a vegan who unplugs their appliances.
I don’t have a golf course or a membership so
thanks for that irrelevant factoid
and yes I am vegan, but eating meat is, in many cases, a culturally ingrained practice which people need time to understand the harms of. AI is not culturally ingrained for anyone with an actual culture. People need to eat - it shouldn’t be meat, but I’m not going to shame people for not seeing their harm yet. People do not need to generate slop. I am going to bully all AI users.
Do you think AI is mostly generating slop? Only like 5% of AI use is private individuals.
The point is that this is not actually an ecological hazard the way people say it is, and that there are much, much more pressing issues if we’re going to fix the environment. Literally all we would have to do to solve global warming is stop eating meat. You see anywhere near as much about that as you do datacenters? Which cause less than one percent of the environmental issues of meat?
Just in defence of your point:
I prefer to self host, and am in the process of setting up my entire system to run off solar now (tiny 2x250 watt solar array, 3 deep cycle 100Ah marine batteries and sundry equipment).
End to end, my AI rig - a BC-250 (a cut down PS5, basically, though I’m thinking of down sizing) consumes 50w standby / 200w on inference and can run Qwen 3.6-27b (Ternary Bonsai specifically, about par with Sonnet 4.6 on the AAII) at good speeds (~30 tok/s)
With a smart power plug and Raspberry pi (uses about 5w) as the pilot fish, I can run my entire system “off grid”, to say nothing of the excess solar that accumulates (several days battery back up to run home router, pi server, hard drive array and 2-3hrs daily of inference).
For $1500 AUD, “Poor man’s Sonnet 4.6 at home” ain’t nothing to sneeze at.
Outside of folks such as myself, this blog post may be of interest -
https://blog.andymasley.com/p/individual-ai-use-is-not-bad-for
no its true. ChatGPT is confidently wrong all the time, but I find its wrong in ways I can learn from anyway. I ask the question. it gets the answer wrong. But its wrong in a useful way where I actually learn more by correcting it than just getting the right answer.
I could get the same experience from talking to a person for much less money though
They tend to work much better when you point them at things. Hallucination rates go way, way down when they aren’t pulling from their “memory.” They are, also, like you say, wrong in ways that are noticeable and often “interesting” in the Ibsen way. I also doubt that you could get the same experience for less money. I’d have to go back to college to get the education I’m getting with AI. At least community college. It’s a lot of near-experts at once. And when you point it at expert literature it basically is a senile college professor. But yeah the senile part of that statement is pretty important.
Then you can also just as well skip that extra step and just do it yourself
This. Research that always has to be fully confirmed hasn’t actually been done in the first place.
Solving problems by brute force don’t make economic sense to me.
Using an automation tool you can only trust for subjective tasks makes even less sense to me.
I find that a frontier bot pointed at the internet does a lot better job of answering the question than I do when searching on my own. I mean they’re always going to be wrong about some stuff, but aren’t people? I think that’s generally a better metric. And at this point they’re far away from what they were just a few years ago. But you’re correct in that there is an element of brute forcing them or at least shaping them that has to occur that I can’t blame anyone for tapping out of.
I could have always skipped the dewey decimal card catalog and just read the entire library but I’m not mycroft fucking holmes. You don’t have to use a search engine, you can just follow hyperlinks until you find the content you’re looking for.