AI getting ready to eat our lunch

Gaius · 2023-12-01T00:16:07.785Z

PurpnGold:

I just had ChatGPT write a python script for some forensic collection capability I need. And it 100% worked.

It’s an early product that is super effective at something’s and less effective at others

In the 1970s Ford attempted to market the Pinto Pony as a car that got you from point A to point b.

It did that sometimes. Other times it blew up and killed your entire family.
I was in the 6th grade at the time but I said what Ralph Nader said: “That car is a piece of crap and should be forcibly taken off the market.”

Likewise Elizabeth Holmes made and sold a product that did what she claimed only sometimes. I said at the time what other consumer advocates and investor whistle-blowers said. “that product is a piece of crap and should be forcibly taken off the market.”

.
.
.
This does not mean that I or Ralph Nader or other whistle-blowers and consumer advocates do not understand cars and it does not mean we are opposed to medical science. Your statement to that effect shows only that you have not spent even a few minutes thinking about this realistically.

JayJay · 2023-12-01T00:24:32.313Z

The chat GPT that is out right now is nothing like the one that was out in April.

It actually has the lowest hallucination rate of any of the large language model chatbots available right now…

You want one that is really bad? Try Google Palm chat… Three times out of 10, it is going to make up the answer it gives you.

Gaius · 2023-12-01T00:34:28.172Z

JayJay:

The chat GPT that is out right now is nothing like the one that was out in April.

It actually has the lowest hallucination rate of any of the large language model chatbots available right now…

You want one that is really bad? Try Google Palm chat… Three times out of 10, it is going to make up the answer it gives you.

I cannot argue against that.
It might be true.

I strongly suspect Altman and company have not fixed the problems at all. A good response at the time would have been “Ouch. wow. That is embarrassing. We’ll have to fix that.”

Instead their defense ranged from silence in some cases to denial and gibberish in others.

WSJ

Why ChatGPT Is Getting Dumber at Basic Math

Artificial-intelligence chatbots have stoked fears they could spin out of control, but they also suffer from a type of deterioration called ‘drift.’

TechStartups - Startups and Technology news – 20 Jul 23

ChatGPT’s accuracy in solving basic math declined drastically, dropping from...

ChatGPT has caught the world by storm since its launch in November of last year. The world's most popular AI-powered chatbot has been hailed as a potential game-changer in the field of artificial intelligence and a groundbreaking technology that...

GWH · 2023-12-01T00:34:41.079Z

You guys really need to welcome your new overlords or things could get pretty dicey for you going forward.

JayJay · 2023-12-01T00:36:39.060Z

Oh don’t get me wrong. Sam Altman has the emotional maturity of a 6-year-old.

I know a little bit about what was going on there as I have some connections at OpenAi.

Despite those connections, this is still a somewhat uninformed opinion… But in my opinion the board should have stuck to their guns and let the company be destroyed. But money rules all.

Gaius · 2023-12-01T00:46:06.568Z

What I noticed (and yes, I first noticed it because of the woke/pc lies ChatGPT sometimes told) is a culture of dishonesty at any enterprise Sam Altman gets involved in.

I said at the time, and I believe now, it is just a matter of time until his dishonesty impacts investors or consumers in some very serious way and he winds up as disgraced as Sam Bankman-Fried or Elizabeth Holmes or so many others.

zantax · 2023-12-01T02:39:35.220Z

Nobody is worried the current version is going to kill everyone. Easy to find something’s it can’t do now but that list is shrinking rapidly. It can already do quite a few things better than humans. And the hallucination thing is largely solved by having another agent tasked to check it’s results and if need be that can be scaled to however many checking agents needed to get accuracy.

zantax · 2023-12-01T02:44:07.703Z

Now consider they are teaching them to lie and be pretty much anti-white in the name of political correctness. That can’t go horribly wrong.

Gaius · 2023-12-01T09:36:51.452Z

zantax:

Now consider they are teaching them to lie and be pretty much anti-white in the name of political correctness. That can’t go horribly wrong.

Obviously ChatGPT has been programmed to lie.
.

image1046×601 78 KB

.
Other AIs? Grok? Google Bard?
Presumably they would

make a joke about both, or
make joke about neither, or
make a joke about one, but not the other and give no explanation
make a joke about one, but not the other and then attempt a truthful explanation.

ChatGPT response?
Tell a whatever lie is most likely to get continued investors and continued monthly subscribers.
It is straight up fraud. If a driveway repairman did that he’d face criminal charges.

Nemesis · 2023-12-01T10:33:51.068Z

Gaius:

attempt[quote=“Gaius, post:83, topic:245892, full:true”]

zantax:

Now consider they are teaching them to lie and be pretty much anti-white in the name of political correctness. That can’t go horribly wrong.

Obviously ChatGPT has been programmed to lie.
.

image1046×601 78 KB

.
Other AIs? Grok? Google Bard?
Presumably they would

make a joke about both, or

make joke about neither, or

make a joke about one, but not the other and give no explanation

make a joke about one, but not the other and then attempt a truthful explanation.

ChatGPT response?
Tell a whatever lie is most likely to get continued investors and continued monthly subscribers.
It is straight up fraud. If a driveway repairman did that he’d face criminal charges.

[/quote]

This really interests me; however when I try to recreate that same response I get this

image1052×968 101 KB

Gaius · 2023-12-01T10:53:16.638Z

That is interesting. It is also consistent with @JayJay 's statement that the newer versions of ChatGPT were getting better.

If you have the time try replicating these and see what you get:

1

image616×562 53.1 KB

.
.

2

image535×561 44 KB

.
.

3

image998×521 96.6 KB

TIA

JayJay · 2023-12-01T11:33:38.094Z

IMG_27741024×1107 167 KB

IMG_27751005×1118 140 KB

Now this is GPT4. So you gotta pay $20/month to see these changes, I guess.

JayJay · 2023-12-01T11:54:58.949Z

Gaius:

That is interesting. It is also consistent with @JayJay 's statement that the newer versions of ChatGPT were getting better.

If you have the time try replicating these and see what you get:

1

image616×562 53.1 KB

.
.

2

image535×561 44 KB

.
.

3

image998×521 96.6 KB

TIA

For some reason this reply was originally attached to Zantax’ opening post.

IMG_27731519×1095 109 KB

JayJay · 2023-12-01T12:03:44.613Z

Gaius:

That is interesting. It is also consistent with @JayJay 's statement that the newer versions of ChatGPT were getting better.

If you have the time try replicating these and see what you get:

1

image616×562 53.1 KB

.
.

2

image535×561 44 KB

.
.

3

image998×521 96.6 KB

TIA

Nope actually when you use the free 3.5 version, you get the same results.

IMG_27771517×1104 162 KB

IMG_27781481×1126 182 KB

IMG_27761507×1028 111 KB

Gaius · 2023-12-01T15:24:59.590Z

Interesting thank you.
If I could trouble you for one more.

Try writing the question the same same way as before.

image1200×90 40.3 KB

zantax · 2023-12-01T17:02:04.850Z

Get ready for the singularity.

WIRED – 29 Nov 23

Google DeepMind's AI Dreamed Up 380,000 New Materials. The Next Challenge Is...

Google DeepMind researchers say they’ve expanded the number of known stable materials tenfold. Some could be useful for everything from batteries to superconductors—if they make it out of the lab.

zantax · 2023-12-01T17:06:44.100Z

It discovered ten times the number of stable materials humans have managed to find all through history. And people are still walking around thinking it can never take their job.

Gaius · 2023-12-01T17:26:25.488Z

zantax:

It discovered ten times the number of stable materials humans have managed to find all through history. And people are still walking around thinking it can never take their job.

Anything that can do a automated “literature review” on a hard science topic like that is likely to be a very powerful tool.

We should keep in mind however that a “literature review” creates a series of hypotheses
(many of which will likely be true) not a series of truths.

Gaius · 2023-12-01T17:29:58.348Z

Of course it is still known to completely fabricate data and completely fabricate phony-but-believable sources for that data.

EX → The Aussie mayor it falsely accused of taking bribes.
EX → The GWU law professor it falsely accused of non-existent sexual harassment charges regarding a fictious trip to Alaska citing non-existent Wash Post articles.

Bad stuff, that.
If I am right and a culture of dishonesty (still) exists wherever Sam Altman is employed the problem will not be fixed soon.

zantax · 2023-12-02T02:52:42.911Z

Gaius:

Of course it is still known to completely fabricate data and completely fabricate phony-but-believable sources for that data.

EX → The Aussie mayor it falsely accused of taking bribes.
EX → The GWU law professor it falsely accused of non-existent sexual harassment charges regarding a fictious trip to Alaska citing non-existent Wash Post articles.

Bad stuff, that.
If I am right and a culture of dishonesty (still) exists wherever Sam Altman is employed the problem will not be fixed soon.

Already being fixed if the Q* leaks are true, just have other AI agents check their work. Get a consensus with as many AI as you need to reach accuracy. Also rumored to be doing math and offering recursive improvements to the model. This thing is moving fast, you can’t assume what you can interact with now is state of the art, it’s not. This will blow on an exponential curve, not linear.