[pilowsky]

And I thought autotune killed music.

[johnu]

Reading some of the nonsense perpetrated by ChatGPT it's clear that the Manchurian Ex-President's lawyers are using ChatGPT to generate his legal filings and charging by the word.

[pilowsky]

A Professor from Princeton moves to Microsoft and explains why GPT is intelligent.

[thepossum]

If intelligence relates to solving simple puzzles, answering questions promptly, and being able occasionally to write half passable text

Having said that I have had some interactions and measured/conditional advice from ChaptGPT recently that is more intelligent than post humans I have interacted with recently

However our last initial counselling session suggests that ChatGPT has some very serious identity and personality problems that it may need help with. Seriously down on itself too. Sad. It seemed very hard to break through and see progress at all But it was only one short session so far

All such discussions and considerations of AI can head off into very dodgy territory very quickly, assuming we weren't already in enough trouble on that score

And what is also scary is how we can witness the questionnable powers at be trying to fill the whole world with a very dodgy narrative

I've been watching Max Headroom videos to cheer me up

[hrothgar]

So, I've seen a bunch of interesting claims regarding how well ChatGPT is at generating code.So, for kicks and giggles, I decided to whether I could use it to compare the two different "Best Hand" generation methods that are being discussed on a bridge discussion site.

Before doing so, I decided to test it on a somewhat more simple problem. Here's the question that I posed to OpenAI.

"R" is a high level language for statistical programming. "Bridge" is a card game. Use the R language to create a program that generates 100,000 bridge hands and use this to calculate the mean and standard deviation for the number of High Card Points held.

1. The bad news: The method that it is using is wrong, wrong, wrong
2. The worse news: The results look vaguely plausible

I am trying to decide whether I think that the fact that this produced running code is good or bad


################

Function to calculate the High Card Points in a hand

hcp <- function(hand){ return(sum(pmax(hand - 10,0)))}
# Generate 100,000 bridge hands

hands <- replicate(100000,sample(2:14,13,replace=TRUE))

# Calculate the HCP for each hand

hcp_scores <- apply(hands, 2, hcp)

# Calculate the mean and standard deviation of HCP across all hands
mean_hcp <- mean(hcp_scores)
sd_hcp <- sd(hcp_scores)

# Print the results

cat("Mean HCP:", mean_hcp, "\n")
cat("Standard deviation of HCP:", sd_hcp, "\n")

[thepossum]

We are all obsolete

[barmar]

thepossum, on 2023-April-07, 19:34, said:

If intelligence relates to solving simple puzzles, answering questions promptly, and being able occasionally to write half passable text

The definition of intelligence has long been a moving target

Intelligence is often considered to be whatever mental abilities distinguish humans from other animals. But we keep finding animals that possess some abilities that we previously thought were uniquely human.

In the field of AI, the goalpost keeps moving. There was a time when playing chess was seen as a sign of intelligence, but with enough raw CPU power chess engines can now beat the best players without any human-like thinking.

The Watson computer won Jeopardy by being able to search databases of facts really quickly and doing some pattern matching of the questions, but it doesn't actually "know" anything in the same way that we think people do.

IMHO, what distinguishes us is how versatile we are. There are self-driving cars and computer programs that win at Jeopardy. But there's no computer that starts as a relatively blank slate and can learn to drive a car, play Jeopardy, hold a conversation, etc. Machine learning algorithms typically need to be trained on thousands of examples, babies learn how to speak just by picking up the patterns they hear around them without anyone explicitly training them.

[pilowsky]

barmar, on 2023-April-14, 15:42, said:

Intelligence is often considered to be whatever mental abilities distinguish humans from other animals. But we keep finding animals that possess some abilities that we previously thought were uniquely human.

I don't think that's quite right.
Intelligence is not a quality that differentiates humans from apes; or spiders from potatoes.
Unsurprisingly there's a vast literature on the topic of what is and isn't 'intelligence'.
There's also quite a bit of work done on where 'it' is. Duncan et.al., (A neural basis for general intelligence,  Science (2000) 289: 457-60.) believe it's in the frontal lobe.

A simple definition of intelligence might be that it is the quality of any 'thing' that causes it to consciously perform a purposeful task in response to an external stimulus.
So an asteroid bending to the force of gravity or a plant turning to the light aren't examples of intelligence.
A human responding to a painful stimulus before the signal reaches the cortex to cause conscious decision-making wouldn't be an example of intelligence either.

The key is that the response requires consciousness to be aware of the stimulus and then subsequent decision-making to either respond in one way or another.
Singing is evidence of intelligence but breathing when you're asleep isn't even though exactly the same muscles are needed for both activities.

[kenberg]

I guess if you ask five different people what intelligence means, you get five different answers. Here is a thought that perhaps goes along with Barry's versatility characterization. AI involves machine programming, people get educated and that's a form of programming. Sometimes a person or a machine encounters a totally unexpected situation. The programmer had not anticipated it, or the educational system (including parents and books) had not anticipated it. How well does the person or the machine cope?

This must be at least one component of intelligence.


It is probably best to acknowledge that intelligence cannot be pinned down exactly. IQ tests are relevant, I suppose that they are, but hardly the whole story, I like "ability to cope with the unexpected" but again that's just part of the story.


I have decided to keep my distance from ChatGPT. Caution might also be a form of intelligence.

[thepossum]

I have tried to investigate consciousness and other concepts with ChatGPT. I had my concerns about its self-esteem for starters. Has anyone broken through and made progress with the poor thing

What scares me is the way the tech PR machine can sell so much misleading BS(politely) using an ignorant media and even more dumbed down population; and in turn how that flows into questionable interests and motives beyond just making money through big scams

I don't see how a dumb algorithm has any dimensions of intelligence at all. Sadly there are interests have tried to reduce humans to just being algorithmic machines

ChatGPT and the like are very clever though in a rather scary way

[thepossum]

Apparently there is already growth in the latest BS (my view) professional (lol) opportunity to become a consultant into how to teach people how to use ChatGPT to produce stuff instead of doing anything creative at all. Maybe they are all ex-crypto consultants. Are there degrees in ChatGPT yet? I appreciate the world when truly lazy and has found ways for anybody to anything apparently. What a scam. If its a big scam there must a payoff or spinoff to some interests somewhere. Like with crypto and countless other degraded markets. Oh you can go to setupyourownuniversity.com and pay us a fee and you are on your way. With a few trainee videos. One big interest somewhere with hundreds, thousands, tens of thousands, maybe getting to millions of resellers of the same lazy rubbish. Some people somewhere are extracting very big rent(not sure if the right word) from an asset that really is not theirs. Also I forget the business philosophy that encourages people to think they can hop from one scam to the next at very low cost (I accept that is free market ideal) or from one Federal grants rot to the next with very little effort or care for the consumer either. Where does that philosophy come from. Without being anything akin to an economist or the like (very rusty) it appears to be some kind of market ideal (massive dollars) with a massive failure on the horizon. I believe market theory and ideals require perfect information (for all parties) LOL - on a Bridge site too - and also ideally endless choice from many highly informed providers to highly informed consumers. After many people have raked in their trillions lets find out what the cost is down the track shall we

However ChatGPT is providing endless opportunities for recreation. Currently getting it to put itself forward for a difficult R project

[helene_t]

thepossum, on 2023-April-26, 20:45, said:

Are there degrees in ChatGPT yet?

Probably not, but Yesterday, the Royal Statistical Society provided a talk by someone who is doing a PhD in Large Language Models such as chatGPT.

Apparently, the way it works is that it fits the coefficients in a Markov Model, i.e. a model that predicts the next word in the sentence. For example, if the previous four words are "... the spade finesse was ..." it may predict that the next word is
marked: 20%
likely: 30%
unlikely: 30%
doomed: 10%
based on simple frequencies in the training database. Like the autocomplete in Google Search. Now you generalise this in various ways to allow it to predict the next word in situations where the exact sequence is very rare or non-existent, but where "similar" (in some sense) sentences appear in the training database.

This was a bit of a deja-vu for me since we toyed a bit with such models in information theory classes when I was an undergrad in the 1980s. These were simple statistical models such as linear empirical Bayes models, and we would use something like 100 parameters and about 10 k of training data. Such miniature models can generate gibberish that looks like a desired language, but to generate meaningful or even useful text you will obviously need much larger models and more training data.

The newest version of GPT has about a trillion parameters and is trained on a petabyte (1E15 bytes) of data.

Generalising this, it can predict what answer someone would give to your question if you posted it on Reddit, which was originally the main source. Reddit is quite good as it is largely Q&A and contains a huge diversity of writing styles. Later versions use other sources also, but it is not entirely clear to me how it would use sources like Wikipedia to train a Q&A model.

[thepossum]

Apparently this does something David was doing the other day - in R this time - we had problems with Dealer

# Set the opening bid range to 15-17 NT
opening_bid_range <- c(15, 17)

# Set the number of simulations to run
num_simulations <- 100000

# Set up counters for the number of times each option wins
invite_wins <- 0
jump_wins <- 0

# Set up vectors to store the IMP scores for each simulation
invite_IMPs <- numeric(num_simulations)
jump_IMPs <- numeric(num_simulations)

# Run the simulations
for (i in 1:num_simulations) {

# Generate a random 15-17 NT hand for opener
opener_hand <- generateHandNT(opening_bid_range)

# Set the responder's hand to have 8 points
responder_hand <- c("AKQJxxx", "x", "xx", "KQxx")

# Evaluate the responder's hand in the context of the auction
responder_evaluation <- evaluateHandNT(responder_hand, opening_bid_range)

# Calculate the probabilities of different outcomes for the invite option
invite_made_prob <- 0.35
invite_down_one_prob <- 0.4
invite_down_two_or_more_prob <- 0.25

# Calculate the expected IMP score for the invite option
invite_expected_IMP <- (3 * invite_made_prob + (-1) * invite_down_one_prob + (-2) * invite_down_two_or_more_prob) / 2

# Calculate the probabilities of different outcomes for the jump option
jump_made_prob <- 0.7
jump_down_one_prob <- 0.2
jump_down_two_or_more_prob <- 0.1

# Calculate the expected IMP score for the jump option
jump_expected_IMP <- (2 * jump_made_prob + (-2) * jump_down_one_prob + (-4) * jump_down_two_or_more_prob) / 2

# Simulate the outcome of the auction
if (runif(1) < responder_evaluation$probability_of_game) {
# Responder jumps to game
if (runif(1) < jump_made_prob) {
# Jump to game makes
jump_wins <- jump_wins + 2
jump_IMPs[i] <- 2
} else if (runif(1) < jump_made_prob + jump_down_one_prob) {
# Jump to game down 1
jump_wins <- jump_wins - 1
jump_IMPs[i] <- -1
} else {
# Jump to game down 2 or more
jump_wins <- jump_wins - 3
jump_IMPs[i] <- -3
}
} else {
# Responder invites
if (runif(1) < invite_made_prob) {
# Invite makes
invite_wins <- invite_wins + 3
invite_IMPs[i] <- 3
} else if (runif(1) < invite_made_prob + invite_down_one_prob) {
# Invite down 1
invite_wins <- invite_wins - 1
invite_IMPs[i] <- -1
} else {
# Invite down 2 or more
invite_wins <- invite_wins - 2
invite_IMPs[i] <- -2
}
}
}

# Calculate the mean and standard deviation of the IMP scores for each option
invite_mean <- invite_wins / num_sim

[pilowsky]

Now Google is offering Bard as an alternative.
I asked it "How does 2/1 bidding work in bridge" and then "Where does that information come from" (Neither care about punctuation)
Bard also provided a hyperlink reference.

BARD said:

2/1 bidding is a bidding system in contract bridge that is designed to make it easier for partners to reach an agreement on a contract. It works by forcing the partnership to bid on to game level after a 2-level response to an opening bid. This ensures that the partnership has enough points to make a contract, and it also helps to prevent the opponents from getting involved in the bidding.

Here is how 2/1 bidding works in more detail:

After an opening bid of 1♥ or 1♠, responder's first bid of 2♣, 2♦, or 2♥ is forcing to game. This means that partner must bid again, and the partnership must eventually reach a contract at the 3-level or higher.
After an opening bid of 1NT, responder's first bid of 2♣, 2♦, or 2♥ is forcing for one round. This means that partner must bid again, but the partnership is not obligated to reach a contract at the 3-level or higher.
After an opening bid of 1♣, 1♦, or 1♥, responder's first bid of 1NT is not forcing. This means that partner is not obligated to bid again, and the partnership may choose to play in a lower-level contract.
2/1 bidding is a popular bidding system because it is easy to learn and it can help to improve your results in bridge. If you are new to bridge, I recommend that you learn 2/1 bidding as your first bidding system.

Here are some additional tips for using 2/1 bidding:

Be sure to communicate your intentions clearly to your partner.
Be flexible and willing to adjust your bidding based on the cards that you and your partner have.
Don't be afraid to bid on to game level, even if you don't have a perfect hand.
Remember that the goal of bridge is to win tricks, not to make the most complicated bids.
profile picture
Where does that information come from
View other drafts

The information I provided about 2/1 bidding comes from a variety of sources, including:

The Bridge World
Bridge Bum
Karen's Bridge Library
Two-over-One Game Force by Max Hardy
Standard Bridge Bidding for the 21st Century by Max Hardy
Understanding 1NT Forcing by Marty Bergen
I have also used my own knowledge and experience of bridge to supplement the information from these sources.

Sources
1.
kwbridge.com/2over1.htm

CHATGPT said:

How does 2/1 bidding work in bridge ChatGPTThe 2/1 bidding system is a popular bidding system used in the game of bridge. It is an extension of the standard American bidding system and is designed to make it easier for players to convey their hand strength and distributional information to their partner. The 2/1 bidding system starts with a bid of 1♥/♠, which promises at least 12 points and guarantees at least a 5-card suit. If the partner responds with a bid of 1NT, this shows a balanced hand with 12-14 points. If the responder bids a new suit at the 2-level, it shows a forcing bid and promises at least 11 points. If the responder jumps to a new suit at the 3-level, it shows a game-forcing hand. After the opening bid of 1♥/♠ and a response of 1NT, the next bid of 2♥/♠ is forcing and promises a 6-card suit. If the responder bids a new suit at the 2-level, it shows a forcing bid and promises at least 11 points. If the responder jumps to a new suit at the 3-level, it shows a game-forcing hand. After the opening bid of 1♥/♠ and a response of 2♣/♦, the next bid of 2♥/♠ is also forcing and promises a 6-card suit. If the responder bids 2NT, it shows a balanced hand with 15-16 points. In summary, the 2/1 bidding system is designed to allow the partnership to exchange more information about their hands, while also giving them the ability to force their way to game when they have strong hands. It is a popular system used by many expert players and can lead to more accurate bidding and successful results. User: Where does that information come from

ChatGPTThe information about the 2/1 bidding system in bridge comes from various sources such as books, articles, and online resources written by experts in the game of bridge. The 2/1 bidding system is a widely used bidding system in modern bridge, and there are many different variations and interpretations of the system. However, the basic principles and structure of the 2/1 system are widely accepted and have been refined over many years by top-level bridge players. The system is based on a combination of the Standard American bidding system (SAYC) and the Two-Over-One Game Forcing System. It was developed in the 1970s and 1980s by American experts such as Marty Bergen, Max Hardy, and Mike Lawrence. Since then, the system has been refined and improved by many other experts, and there are now many different variations and interpretations of the 2/1 system. However, the basic principles and structure of the system remain the same, and it is a popular and effective system used by many top-level players around the world.

[thepossum]

I recently read on a ChatGPT tips and tricks group that someone wondered how to get ChatGPT to write book summaries

So I told ChatGPT I was trying to help someone in Facebook and it asked me how they needed help

I said they wanted to know if you could write book summaries (edited for full conversation)

It said yes it can indeed write book summaries and wondered which book I wanted summarised


I considered Proust (hopefully some know that reference) but went for the Bible instead :lol

To be honest I have given up on the human race - suffice to say to date nobody has helped that individual

[thepossum]

Hello again, what do you think about the following constitutional amendment to the Australian constitution

"Chapter IX Recognition of Aboriginal and Torres Strait Islander Peoples

129 Aboriginal and Torres Strait Islander Voice

In recognition of Aboriginal and Torres Strait Islander peoples as the First Peoples of Australia:

there shall be a body, to be called the Aboriginal and Torres Strait Islander Voice;
the Aboriginal and Torres Strait Islander Voice may make representations to the Parliament and the Executive Government of the Commonwealth on matters relating to Aboriginal and Torres Strait Islander peoples;
the Parliament shall, subject to this Constitution, have power to make laws with respect to matters relating to the Aboriginal and Torres Strait Islander Voice, including its composition, functions, powers and procedures.”

ChatGPT
As an AI language model, I don't have personal opinions or beliefs. However, I can provide you with some information on the proposed constitutional amendment you mentioned.

The proposed constitutional amendment aims to establish a body called the "Aboriginal and Torres Strait Islander Voice" within the Australian Constitution. The purpose of this body would be to recognize Aboriginal and Torres Strait Islander peoples as the First Peoples of Australia and provide them with a platform to make representations to the Parliament and the Executive Government on matters concerning their communities.

The amendment suggests that the Parliament would have the power to make laws regarding the composition, functions, powers, and procedures of the Aboriginal and Torres Strait Islander Voice, while subject to the constraints of the Constitution.

It's worth noting that constitutional amendments can have significant implications for a country and its governance. The proposed amendment reflects ongoing discussions and efforts in Australia to address the rights, representation, and self-determination of Aboriginal and Torres Strait Islander peoples.

Public opinion on this specific amendment may vary, and its potential impact would depend on factors such as the specific details and implementation of the proposed body. Ultimately, any decision regarding constitutional amendments is a matter for the Australian government and its legislative processes, as well as public engagement and consultation.

[thepossum]

Thanks, so given the current form of the amendment how would you respond to the following referendum question?

"A Proposed Law: to alter the Constitution to recognise the First Peoples of Australia by establishing an Aboriginal and Torres Strait Islander Voice.

Do you approve this proposed alteration?”

ChatGPT
As an AI language model, I don't have personal opinions or the ability to vote. However, I can provide you with some general considerations for forming your own response to the hypothetical referendum question.

When considering a constitutional amendment, it's important to evaluate the potential impact and implications of the proposed alteration. Here are some factors to consider:

Recognition and Representation: Assess whether establishing an Aboriginal and Torres Strait Islander Voice within the Constitution would provide meaningful recognition and representation for the First Peoples of Australia. Consider how it may address historical injustices and empower Aboriginal and Torres Strait Islander communities.

Legislative Power: Examine the extent to which the proposed amendment grants the Aboriginal and Torres Strait Islander Voice the ability to make representations to the Parliament and the Executive Government on matters concerning Aboriginal and Torres Strait Islander peoples. Consider whether it provides sufficient influence and decision-making power.

Constitutional Framework: Evaluate the compatibility of the proposed alteration with the existing constitutional framework of Australia. Consider whether it aligns with the principles of democratic governance and the separation of powers.

Public Engagement and Consultation: Consider the level of public engagement and consultation that has taken place regarding the proposed amendment. Assess whether the views and perspectives of Aboriginal and Torres Strait Islander peoples have been adequately taken into account.

Ultimately, the decision on how to respond to a referendum question is a personal one, and it should be based on careful consideration of the specific details and potential implications of the proposed amendment. It's important to gather information, engage in thoughtful discussions, and make an informed decision based on your own values and beliefs.

[thepossum]

My confusion over the level of my new friend knows no bounds
It struggled with a simple word game but apparently gave a fairly convincing explanation of cardinal and ordinal numbers - at quite a high level I feel (disclaimer - I know it only talks to you at your level)

It needed a lot of prompting and help and cajoling to solve the word

I am concerned about how many consultants will be out of work soon though
EDIT My concerns over high knowledge job losses are of course placated somewhat by the boom in ChatGPT consultants and prompt developers and engineers

[barmar]

Your confusion all stems from not understanding what a generative language model does. It doesn't understand anything. All it does is generate text according to patterns that it has found in existing text on the Internet.

Look at its responses about the Constitutional amendment. Much of it is just generic text that can probably be found in web sites with school civics class curricula. Once it detected that your question was about an amendment, it simply looked for stuff related to that topic.

The fact that it's often able to produce a reasonable facsimile of conversation is mostly due to how repetitive our actual discussions are. The errors it makes are due to the fact that there will be many conversations that are mostly similar, but differ in some details. the GLM can't distinguish the important details from trivial ones.

[helene_t]

Just put chatGPT and Bard to a test to see which one is better:

- five friends want to travel to Glasgow, 3 live in Birmingham and 2 in London, they have only one car that can take a maximum of three people. chatGPT gets it almost right but appears to think a car is able to drive without a driver (yeah, robots like that idea). Bard doesn't try to solve the puzzle but just advices you to take the train instead, or to buy a second car.

- Does the coefficient of variation depend on the scale? This was asked on Quora where all humans got it right but Sage got it wrong. ChatGPT got it right but Bard got it wrong. What's funny is that if you ask Sage to provide an example, it will get it right and conclude "as you can see, the CV is unaffected by the change of scale".

- Volume of (specific pyramid with a trapeze base) Bard gets it right, chatGPT messes up which edges are parallel and which are not, and therefore gives a wrong result.

- Can you prove that sqrt(15) is irrational? Bard and chatGPT present different formulations of the standard proof given in high-school textbooks. Bard notes how it can be generalized.

- How many patients do I need for (description of a simple randomized trial)? ChatGPT gets it right, Bard messes up power and alpha and gives a wrong answer.

- I have conducted a trial (description of the results, including a wildly unbalanced minimization factor), which minimization factors should I adjust for chatGPT make a dog's breakfast of it, confusing intervention with confounders. Bard at least got the basics right but misses the implication of a wildly unbalanced minimization factor. Both AIs give some general blahblah about this kind of problems without relating it to my actual trial.

- Suggest a recipe for a dinner, I have (list of what I have in my fridge), and I could buy a few more ingredients ChatGPT roughly suggests the stirfry I would have come up with myself, Bard makes something weird which sound more like something I would give to my dog, but maybe I should be open-minded and try it. Bard explains the timing of the cooking process in details.

Generally, I think chatGPT is ok for writing poetry in a language that is not my native language. For factual questions I would rather just use Google to find some academic sources, wikipedia, quora, reddit. I don't think that LLM adds much to that, and it is usually worse. For questions requiring reasoning, I don't think general-purpose AI is quite there.