We have a guy on our team of software developers that does literally nothing. I'm not even kidding. He probably has another job or two. Maybe he's an extreme case, but it occurs to me that at least he does no damage. It's worse when you have a below-average performer that is prolific. M'fer can bring a whole organization down. I'd prefer Chat-GPT to either of these all too common personas.
-
01
-
no signs of intelligence wrote:
It's a glorified database + search engine + autocomplete. Move along, nothing to see here.
that's all humans are too, but it takes us 20+ years to "train"
21 -
zxcvzcvx wrote:
If you were to provide a fair comparison in which the LSAT candidates also have full access to any information they want on the bar exam, would the AI program still beat most of them?
It didnt have access to information any more than seeing something and having it in your brain counts as "access" to information.
There are some seriously bad efforts to re-define intelligence here.
12 -
2600 bro wrote:
zxcvzcvx wrote:
If you were to provide a fair comparison in which the LSAT candidates also have full access to any information they want on the bar exam, would the AI program still beat most of them?
It didnt have access to information any more than seeing something and having it in your brain counts as "access" to information.
There are some seriously bad efforts to re-define intelligence here.
It had access to all the information it was trained on, an internet archive going up to 2021. This is not the same as having learned things. That is unless you have the ability to process all the information easily available on the internet and immediately memorize it. a good comparison would be letting someone take the bar with internet access.
30 -
7 ways ai fails. One of them is math lol! And common sense.
00 -
Steve Baller wrote:
This again? wrote:
Automation isn't only coming for low skilled workers. It's literally going to wipe out middle class professional jobs at much greater speed than anyone anticipated.
Careers that are going to disappear in 5 years:
-Copywriters (really any career that relies solely on the written word)
-Many attorneys (doc review can be done by AI for free with basically 100% accuracy, AI has demonstrated a very strong ability to logically reason within a legal context)
-Most software developers - CPT 4 produced bug-free code based on natural language inputs
-Anyone in the white collar "knowledge economy" is severely at risk now.
Real-world jobs that can't be done from home might be safe for a while.
I hope someone bumps this in 5 years. You’ve gone off the hype deep end.
40 -
Harambe wrote:
hkjhkh wrote:
Pretty sure the BAR isn't open book. But I get what you're arguing. That ai is like a human in the sense that it has to learn and retain info (which isn't the same but whatever). The only thing ai has going for it is it is faster and it "remembers" better. It really isn't doing that much more than ctrl+f. it already has a leg up on the human brain in the sense that it doesn't push info out to make room for more. Unless, of course, a virus is introduced.
A 'thing' that can recall information far better and faster than humans and can now breeze through relatively advanced tests of reasoning competency is great evidence of transformative impact.
Large parts of very skilled jobs are open to automation now or in the reasonable future (a few years).
You make a great point about the risks of these models out in the wild -- how vulnerable will they be to malicious actors (for one)?
We are decades away from any large scale automation.
10 -
it got a 2 on physics and calc bc
00 -
rickey jmike wrote:
attorney is basically a glorified paper pusher, it's a fluff job like an administrative assistant. Of course it'll be one of the first jobs that AI can make redundant. Thank god. hate lawyers.
If only!
I could then just sit back and let AI do the work, and get paid for signing documents. That would be a profitable retirement.
30 -
JoggyHobber wrote:
2600 bro wrote:
It didnt have access to information any more than seeing something and having it in your brain counts as "access" to information.
There are some seriously bad efforts to re-define intelligence here.
It had access to all the information it was trained on, an internet archive going up to 2021. This is not the same as having learned things. That is unless you have the ability to process all the information easily available on the internet and immediately memorize it. a good comparison would be letting someone take the bar with internet access.
No this is incorrect. There is no evidence that it is simply “memorizing” things. It’s not public how many parameters GPT-4 but it’s possibly on the order or 10-100 trillion. Each parameter is kinda like a “neuron” - but is really just a single number than can adjust during training.
There is no bit-for-bit encoding and storage of all the training material! That would be possible since the training data was far larger than could be represented in 100T numbers.
Just like the human brain, you can’t open up GPT4 and point to the cluster of neurons/parameters that stores the LSAT info!
11 -
2600 bro wrote:
JoggyHobber wrote:
It had access to all the information it was trained on, an internet archive going up to 2021. This is not the same as having learned things. That is unless you have the ability to process all the information easily available on the internet and immediately memorize it. a good comparison would be letting someone take the bar with internet access.
No this is incorrect. There is no evidence that it is simply “memorizing” things. It’s not public how many parameters GPT-4 but it’s possibly on the order or 10-100 trillion. Each parameter is kinda like a “neuron” - but is really just a single number than can adjust during training.
There is no bit-for-bit encoding and storage of all the training material! That would be possible since the training data was far larger than could be represented in 100T numbers.
Just like the human brain, you can’t open up GPT4 and point to the cluster of neurons/parameters that stores the LSAT info!
Stephen Wolfram describes machine learning as information compression in a sense, so "bit-for-bit" is a red herring. GPTs are doing nothing more than predicting words in the context of a prompt based on their training data, and there's lots of discussion of the bar exam/AP exams/etc. on the Internet. It's not any kind of generalized intelligence.
Good article here:
20 -
my 2c wrote:
2600 bro wrote:
No this is incorrect. There is no evidence that it is simply “memorizing” things. It’s not public how many parameters GPT-4 but it’s possibly on the order or 10-100 trillion. Each parameter is kinda like a “neuron” - but is really just a single number than can adjust during training.
There is no bit-for-bit encoding and storage of all the training material! That would be possible since the training data was far larger than could be represented in 100T numbers.
Just like the human brain, you can’t open up GPT4 and point to the cluster of neurons/parameters that stores the LSAT info!
Stephen Wolfram describes machine learning as information compression in a sense, so "bit-for-bit" is a red herring. GPTs are doing nothing more than predicting words in the context of a prompt based on their training data, and there's lots of discussion of the bar exam/AP exams/etc. on the Internet. It's not any kind of generalized intelligence.
Good article here:
So, ‘ctrl-f’ + choose the next highest-ranked word (based off millions of human written copy) + a bit of randomness thrown in to make it “interesting”?
00 -
in the ai biz wrote:
Harambe wrote:
A 'thing' that can recall information far better and faster than humans and can now breeze through relatively advanced tests of reasoning competency is great evidence of transformative impact.
Large parts of very skilled jobs are open to automation now or in the reasonable future (a few years).
You make a great point about the risks of these models out in the wild -- how vulnerable will they be to malicious actors (for one)?
We are decades away from any large scale automation.
Major goalpost moving. I guess it's what you mean as 'large-scale.'
00 -
Steve Baller wrote:
Basically the same results on the SAT and AP subject tests
Still think GPT is "just another tech bro fad"? Wake up.
Oh yeah, it can also code up a website in HTML 5 / CSS / Javascript from handwritten notes on a napkin:
This is not surprising. GPT searches all the information available to make a response to a question. In situations like this where the answers are widely available, it will be able to query all sources for the answer.
GPT could make life easier for a lot of lawyers. So much of what they do is pretty mundane. So using something like this could pull the necessary citations and the lawyer then double checks the work and saves hours (and perhaps the client money). We have seen a move toward this already with online attorney services where things like wills can be pretty easily done and cheap for probably 90% of people.
10 -
'Next word predictor' is, IMO, the most meaningless way to dismiss LLM 'intelligence.'
Point me to a model of intelligence that shows we are drastically different from an LLM? You can't because we don't know enough about the foundations of human (or machine) intelligence.
Anyone who says a 'next word predictor' is fundamentally unable to become 'intelligent' is just wrong because we don't know enough about human intelligence to say that!
I certainly think that LLMs will probably 'max-out' at some point but... who knows. Very few people successfully predicted how much better LLMs would get 5 or 6 years ago. Probably wise to operate with plenty of uncertainty going forward.
00 -
Vcvnb wrote:
7 ways ai fails. One of them is math lol! And common sense.
This is what we have to fear from our AI overlords:
Artificial stupidity, indeed.
00 -
Harambe wrote:
'Next word predictor' is, IMO, the most meaningless way to dismiss LLM 'intelligence.'
Point me to a model of intelligence that shows we are drastically different from an LLM? You can't because we don't know enough about the foundations of human (or machine) intelligence.
Anyone who says a 'next word predictor' is fundamentally unable to become 'intelligent' is just wrong because we don't know enough about human intelligence to say that!
I certainly think that LLMs will probably 'max-out' at some point but... who knows. Very few people successfully predicted how much better LLMs would get 5 or 6 years ago. Probably wise to operate with plenty of uncertainty going forward.
No one here has been able to tell me why human intelligence is different from a vast recall bank + next word predictor, to an outsider we could be classified the same way!
Lots of dismissing these models... so tell me what would count as machine intelligence?
01 -
Vcvnb wrote:
7 ways ai fails. One of them is math lol! And common sense.
Posting an article from 2021 about limits of LLMs and AI is... stupider than something an LLM or AI would do.
02 -
Vcvnb wrote:
2600 bro wrote:
Posting an article from 2021 about limits of LLMs and AI is... stupider than something an LLM or AI would do.
Yep, still bad at math.
or is 6 months not current enough?
Yes LLMs are not perfect at interpreting written math prompts. They certainly "knows" enough math to do well on the SAT :)
GPT4 is way better than than GPT3/ChatGPT3 at "figuring out" math. Look around at examples or play with them yourself.
It has improved a lot in ~2 years.
This is like watching the train coming at you saying "it hasn't hit me yet so no big deal."
We already have computers that are quite good at math, don't we?
00
What People Are Talking About On LetsRun
Hot Now
Of General Interest
-
RIP: D3 All-American Frank Csorba - who ran 13:56 in March - dead
-
RENATO can you talk about the preparation of Emile Cairess 2:06
-
Great interview with Steve Cram - says Jakob has no chance of WRs this year
-
Hats off to my dad. He just ran a 1:42 Half Marathon and turns 75 in 2 months!
-
2024 College Track & Field Open Coaching Positions Discussion
-
Running for Bowerman Track Club used to be cool now its embarrassing