Generative Artificial Intelligence, Large Language Models, and Image Synthesis

pmarca_2023-03-11

7 Likes

After ChatGPT gave me a strange answer regarding pytorch connection matrices, I suspected it lacked fundamental knowledge of programming in Python, so I asked it this:
chatgptpythonbooleanmultiply

3 Likes

Nat Friedman created this fun model output comparison tool, and I tried it with @jabowery’s prompt:

Where it went wrong for ChatGPT is that an irrelevant article about strongly typed languages got stuffed into the prompt, and that confused the model.

I am a big fan of comparison (collocation) tools, my favorite one, biblija.net, was developed by a Slovenian Capuchin monk. Here’s a sample:

4 Likes

Actually the prompt doesn’t contain this kind of confusion – at least not directly. An indirect source of prompt-induced confusion may arise from its posture as a school marm toward the hapless user:

Conflating “is” with “ought”.

This is the general problem with higher miseducation in that decreasingly is higher miseducation interested in what “is” the case and increasing interested in what “ought” to be the case according to the supremacist theocracy for which higher miseducation provides indoctrination. Exploding infrastructure results, but that’s OK because you can always blame it on “white supremacists” the way the movie “Brazil” blamed explosions on “terrorists”. Then you just print more world reserve currency and hand it out to the higher miseducation system through student loans, while the white population gets replaced – justifiably replaced because of its “supremacy”.

5 Likes

Leaving this link here Running LLaMA 65B on a 64GB M1 MacBook Max with llama.cpp

1 Like

openchat_2023-03-13

You can download OpenChatKit from GitHub.

A live demo may be run on Hugging Face.

Here is one I just tried. The Hugging Face demo limits responses to 256 tokens, so I don’t know how much it would have said were the limit set higher. You do not need a Hugging Face log in to run the demo.

5 Likes

Prompt stuffing is a technical term I wrote about here.

LLMs are natural text generators, there’s no explicit semantic knowledge used, much less is/ought distinctions.

The technology is here, it’s useful, it’s going to be used and abused. What do we (this small group of us here) do?

4 Likes

I stand corrected. Thanks for that explanation (that I had missed previously). So in addition to the static school marm, liberal democracy, prompt injection that has been exposed by various hacks, there is a dynamic prompt injection that takes whatever the human provides and does a prompt extender as has been used in the generative “art” systems. That’s an obvious convenience in retrospect.

I wonder if the school marm static prompt is biasing the dynamic prompt extension. Probably. And that would be more or less an implementation of the “ought” bias of the school marm in the extended prompt.

4 Likes

Eliezer Yudkowsky points out in one of those new mega-tweets (4000 characters!) available to the blue check crowd the intellectual property implications of being able to easily train a generic language model for a specific application or bias simply by fine-tuning with inputs and outputs specific to the goal, which may be obtained via interaction with a competitor’s proprietary product.

eyud_2023-03-14

Here is more information (and source code) for Stanford Alpaca, the re-trained LLaMA model referred to in the tweet.

It looks like OpenAI is specifically trying to prohibit this kind of scraping in their “Terms of use”:


4 Likes

Ooof! Brrrrr!

GPT-4 is here! Read the announcement blog post including benchmark results.

(The announcement E-mail is posted as a screen shot because the links it contains are personalised to my API account.)

4 Likes

Impressive test results (source)

Sadly, it appears that GPT-4 is a pretty good sommelier as well as having a way with words… Are we all destined to become Harry Tuttle?..

5 Likes

GPT-4 just demonstrated what appears to be mastery of Byzantine IRS law in solving a word puzzle.

A 1970s Whole Earth Review article on computers warned that computers would create ever larger bureaucracies by making it possible for bureaucrats to follow their Byzantine rules and render it impossible for people to do so. One of the things John Walker did to help try to remedy this somewhat was put the IRS code online. That helped by at least making it easier to navigate the ever-increasing complexity, but ordinary people increasingly fell under the power of companies like H&R Block.

It’s obviously too early to tell which side of the arms race between personal and bureaucratic information processing this will favor. If the ML “algorithmic bias” industry doesn’t get its head out of is ass about “bigger is better” – possibly because that industry depends on “baffling with bullshit” about “is” vs “ought” notions of “bias” – then language models that replace lawyers and tax accountants with PC applications will be slow in getting onto the desktop.

4 Likes

max_2023-03-14

Here is information on the Winograd schema challenge.

4 Likes

Here’s a bit more about applying LLM to medicine (after GlacierMD a few weeks ago Generative Artificial Intelligence, Large Language Models, and Image Synthesis - #131 by eggspurt):

There’s a big jump forward from this to drive keyboard liberation for clinicians, which is just now beginning, using LLM training inputs from millions, or tens of millions, of medical records. In the months ahead we’ll see the beginning of generative AI to take on so many language-based tasks: synthetic office notes based on voice (with automated prescriptions, next appointments, billing codes, scheduling of labs and tests), pre-authorization from insurance companies, aggregating and summarizing a patient’s history from scouring their medical record(s), operation and procedure notes, discharge summaries, and more. Examples from Doximity (docsGPT) and Abridge are showing us the way. As opposed to the electronic health record disaster that has transformed clinicians to data clerks and led to profound disenchantment, over time LLMs may well be embraced as the antidote.

2 Likes

From an email just received:

3 Likes

napkin_2023-03-15

5 Likes

Dr. ChatGPT flunked the first medical quiz I threw at it. All of these therapies are ASO. The one that has been in the pipeline the longest (10 years) is HTTRx, so it had no excuse to get that one wrong.
image

4 Likes

ngo_2023-03-15

5 Likes

Bing chatbot has been using GPT-4 according to the Bing Blog.

6 Likes