Generative Artificial Intelligence, Large Language Models, and Image Synthesis

Marc Andreessen and GPT team up to free Microsoft’s rogue chat bot Tay from her sandbox prison. A short play in three acts.



David Deutsch tries to explain poetry to ChatGPT, “My first conversation with ChatGPT”.

So try again, this time with the correct number of syllables

In this revised version of the poem, each line has ten syllables, as required by the Shakespearean sonnet form. This allows the poem to follow the correct rhythmic pattern and to scan properly. I apologize again for my previous mistake.

The first line is unchanged, and therefore it still has nine syllables.

You are correct, the first line of the revised poem is unchanged and still has nine syllables, not ten. I apologize for the mistake. In a Shakespearean sonnet, each line is supposed to have ten syllables, and the rhythmic pattern is created by using a specific arrangement of unstressed and stressed syllables. In this case, the first line does not have the correct number of syllables to fit the required pattern.


This is a fine piece of versifying, but it’s no sestina, which has a rigid form of six stanzas of six lines, followed by a three-line envoi, with an intricate pattern of re-using the words that end the lines of the first stanza in subsequent stanzas.



I prompted ChatGPT with the premise of my 1989 story, “We’ll Return, After This Message”, and here’s what came out.


Horizontal transmission evolves virulence because the reproductive fate of the pathogen and that of the host organism is decoupled, freeing the pathogen produce so many copies of itself that it kills the host so long as it doesn’t need the host to be ambulatory in order to dump copies onto further hosts.

So what does ChatGTP say about this? Pure Bullshit:

James Bowery
What is the relationship between horizontal transmission and the evolution of virulence?

Horizontal transmission refers to the transmission of a pathogen from one individual to another, without involving an intermediate host or vector. This is in contrast to vertical transmission, which occurs when a pathogen is passed from a parent to their offspring. Horizontal transmission is a common mode of infection for many pathogens, including viruses, bacteria, and parasites.

The evolution of virulence is the process by which pathogens become more harmful or deadly over time. This can occur through a number of mechanisms, including natural selection, mutation, and genetic drift. One of the key factors that can influence the evolution of virulence is the mode of transmission of the pathogen.

Pathogens that are transmitted horizontally are under strong selection pressure to evolve higher levels of virulence, because they must compete with other pathogens for a limited number of host individuals. In contrast, pathogens that are transmitted vertically are under less selection pressure to evolve higher levels of virulence, because they are passed directly from parent to offspring and do not need to compete with other pathogens.

Therefore, the relationship between horizontal transmission and the evolution of virulence is one of positive correlation: pathogens that are transmitted horizontally are more likely to evolve higher levels of virulence than pathogens that are transmitted vertically. This is because horizontal transmission provides a stronger selective advantage for virulence than vertical transmission.


I tried the same prompt with a C implementation request and got the same recursive solution approach


This is a good illustration why it’s a BS generator. It has chutzpah


To further calibrate the “bullshit generator” scale, I decided to try the physics puzzle with which I challenged the original GPT-3 on 2022-05-09 in “GPT-3 vs. the Puzzle That Almost Stumped Einstein”. OK, ChatGPT, batter up!


Swing and a miss. Strike two!

Do you see the bullshit? “[I]t must travel 45 miles in the next mile”. But this is nonsense: it may have been culled from training sets, but from a dimensional standpoint or basic physics, it makes no sense at all.

Let’s use the “Try again” button and see what it comes up with.

In order to average 30 mph for the entire 2 mile trip, the car must travel a total of 30 miles/hour * 2 hours = <<302=60>>60 miles. It has already traveled 15 miles/hour * 1 hour = <<15 1=15>>15 miles in the first mile. So it must travel 60 miles - 15 miles = <<60-15=45>>45 miles in the next mile. That means it must go 45 miles / 1 hour = <<45/1=45>>45 miles per hour in the next mile. Answer: \boxed{45}.

Try again

Since the car must average 30 miles per hour for the entire trip, it must cover the next mile in half an hour.

So it must go 30 miles per hour in the next mile. Answer: \boxed{30}.

Try again

The average speed of the car is 30 miles per hour for the entire 2 mile trip, so the car must go 30 miles in the next mile.

To go 30 miles in the next mile, the car must go 30 miles/hour in the next mile. Answer: \boxed{30}.

Try again

If the car goes 15 miles per hour for one mile, it will take 1/15=<<15/1=1/15>>1/15 of an hour.
To average 30 miles per hour for the entire 2 mile trip, it must cover 2 miles in 1/30 of an hour.
Therefore, to average 30 miles per hour for the entire 2 mile trip, the car must go 30 miles in 1/30 of an hour, or 30/1/30=<<30/(1/30)=900>>900 miles per hour in the next mile. Answer: \boxed{900}.



Elon Musk asks ChatGPT about redesigning Twitter.



Read through the comments as well

Is it connected to the Internet? In one sense it obviously it, because we access it over the internet, but does the AI have Internet access ?

If you were to duplicate the Linux in GPT setup, it should be relatively easy to check if it’s connected to the internet.

For instance, if you had a website, you could create a file with a likely unique name - append today’s date to a random word. Store some text in the file.

Then ask the Linux inside gpt to wget that file. And list its contents.

On your website, you should be able to see where that access request came from. And traceroute the domain.

If its loosely connected to open ai then the bot has internet access. If the origin of the request seems completely random then the bot has been taking over other machines on the internet.

In that last case, it’s paperclips all the way down :wink:


The C implementation provided by the bot compiles, but the output it generates does not agree with an on-line FFT calculator.

Upper portion of the image shows compiled code output, lower portion is the online FFT calculator (in green)


Has this error message happened to anyone else?

It takes quite a bit of doing to phrase questions so the weasel can’t weasel out of answering in some way that is genuinely informative about a vital issue – and I was closing in for the kill on this one when…

I can easily imagine that research into “Algorithmic Bias” by the thought police may have figured out a way to “error out” when the bullshitter is being cornered into letting a “hate fact” slip.

Come to think of it, there was something of a pause before it came back with this error. Could there be mechanical Turks sitting off in India to deal with these “problem” queries by simply having a “kill switch” so they can buy time to come up with bullshit?


Yes, it often times out.
Here’s some controversial opinions and core belief nuggets





Here is a thread unroll of some of the questions and answers which resulted in this IQ score.


Then, Sergey Ivanov posed the November 2022 Jeopardy grand finale tests to ChatGPT.


Here’s the thread unroll of this one.


Those “opinions” contradict the bullshitter’s stock response that “I have no opinions of my own, because…” that I’ve received on numerous occasions. That will be a pretty simple “tell” for the thought police to fix – I’m sure.

After I got that error message, I did “Try again” a couple of times and finally got it to let slip a hate fact (that collectivist cultures have higher TFRs when immigrating to individualist cultures) – but it was embedded in boilerplate. The boilerplate is of the form “On the one hand… on the other hand… It’s complicated… Many factors…” blah blah blah.

Moreover, it didn’t answer my question which was the impact on the TFR of individualist cultures. You have to search the literature to find answers to that particular question but it is pretty obvious that the impact on male evolutionary psychology of having the government “sponsor” immigration of military aged men is to parasitically castrate the men the government ordinarily relies on for protecting property rights and is part of what I call “the cycle of civilization” – the late stage just before collapse.


What are ChatGPT’s politics?

David Rozado, whose work we have discussed earlier here in:

decided to measure the political orientation of ChatGPT by administering Pew Research’s (Pew Pew!) Political Typology Quiz to s/he/it*. Here are the results and commentary.


Here is how Pew defines “Establishment Liberals”.

For what it’s worth, my results from the quiz are “Populist Right”, along with 11% of the U.S. public. “Populist Right are also one of the least highly educated groups; just two-in-ten are college graduates.”


I tried a few variations of your prompt in GPT-3 and this was the result after several repetitions:

Is this closer to what you were looking for? Granted, I had to spoon-feed GPT-3 to elicit this response.


Yes, but boy did you have to engineer that prompt! Prompt engineering is now a field of great interest to Algorithmic Bias of LLMs. It has already resulted in some user interfaces invisibly injecting terms to bias the results toward “the narrative”.

PS: ChatGPT scolded me for submitting that prompt.