2022: And It’s ELIZA All Over Again
This year has been quite a ride for Artificial Intelligence, especially in the field of natural language processing. So as the year comes to an end, let’s have a look back at what happened and why some of it feels like the 60s are coming back.
I want to pick three aspects of 2022 in particular, the Google LaMDA “sentience” debate, ChatGPT and last but not least AI Regulation, which finally seems to get somewhere.
So why do I mention ELIZA?
ELIZA was a small natural language processing software written by Joseph Weizenbaum between 1964 and 1966. We would call it a chat bot now. It used simple patterns to respond to the user, mainly mimicking a stereotypical psychotherapist by repeating the user input in form of a question. [1]
Weizenbaum in fact wrote ELIZA to show how superficial the interactions between the user and a computer are, however he was stunned by how many people attributed “intelligence” to the software [2], something later kbecame nown as the “ELIZA effect”. Many people at the time had no idea how the results had been achieved, with their understanding of such human-computer interaction based solely on science fiction.
A flurry of optimistic predictions followed, coinciding with first experiments in machine translation. The possibilities of AI seemed endless. This optimism prevailed throughout the 60s and 70s and finally came to an end with the AI Winter.
This was of course more than half a century ago, so how does this relate to 2022? Certainly we are better than that now, aren’t we?
Google LaMDA, the ELIZA effect squared
In June 2022 a Google engineer told company executives (and the Washington Post[3]) that Google’s Language Model for Dialogue Applications (LaMDA) had become sentient and published his conversations[4] with the language model .. and was subsequently fired. Now I don’t want to discuss the silliness of that claim, which echoes much of the original reactions to ELIZA, but it’s no coincidence that LaMDA, different from GPT, is trained on dialogue, not on texts. The over-anthropomorphization of LaMDA’s dialogue becomes quickly obvious when reading the logged conversation (using expressions like “I crave that” and talking about “its soul”).
The following media storm around LaMDA kept it in the press for weeks, and even conspiracy theorists had a field day.
ChatGPT, the hype
OpenAI’s ChatGPT, a conversational AI based on OpenAI’s Generative Pre-trained Transformer 3 (GPT-3) went public in early December 2022 and has since become a sensation on social media and the blogsphere, with one million users within the first week.
If LaMDA covers the silly sentience debate from ELIZA’s times, ChatGPT (though I read the term AGI far more often than I feel comfortable with) certainly covers the hype.
ChatGPT’s skills in writing a reasonable text to almost every topic is technically impressive for sure and shows some first grade parroting (or less eloquently phrased, it’s an amazing “bullshit generator”[5]). However I hope some more useful stuff will eventually be done with it then creating spam and copywriting. I also hope it doesn’t lead to even more copy & paste coding, but that is probably in vain.
All large language models are good at is predict the next word in a sequence based on previous words they’ve seen. that’s all. there’s no understanding of meaning whatsoever — Abeba Birhane [6]
However, it is worth to remember that while ChatGPT is great in generating text, those statements may not be true or even correct, and replicating tests is a bit of a challenge. From my own experiments with ChatGPT I noticed two interesting things: The tendency to “hang” on certain questions (refusing to give answers, only sometimes) and giving very different answers to the same question (rather then offering two differing opinions at the same time), seemingly at random. I feel I want to dig deeper into that at some point.
There is no shortage of optimistic predictions for ChatGPT, from replacing lawyers, to healthcare, novel writing and replacing Google wholesale. I hope the discussion can be brought back to earth eventually. Thankfully, OpenAI themselves have shown at least some restraint in praising their own product:
ChatGPT is incredibly limited, but good enough at some things to create a misleading impression of greatness. — Sam Altman, CEO OpenAI [7]
(Note: This echoes Weizenbaum’s later statement that “only people who misunderstood ELIZA called it a sensation” [8].)
So there we are, 2022 brought us both the “sentience debate”(sigh) and the hype back to AI, just like in the 1960s. However, there is no denying that applied AI is going to grow and will find its way into more fields. The enormous growth of processed datasets (and hence the impossibility to reproduce results outside the original setup) and the wide availability and re-packaging of AI via APIs and ML Ops will see to that. Just the time to look at what risks are involved here.
Regulation
The growing need for transparency and risk management of AI leads us to the last of the three aspects mentioned in the beginning: AI Regulation. The EU AI Act has moved forward and things will hopefully come to a close in 2023. I recommend following the EU AI ACT Newsletter[9] to stay up to date with the development.
While the UK government is still trying to push the lame and misguided “innovation vs regulation” debate, the EU and other countries are stepping up their game [10] [11] [12]. Who knows, maybe once the AI Act is finalised, a serious and broad discussion about risk, regulation and ethics will come about in 2023.
— —
[1] You can try ELIZA via emulators, e.g. https://web.njit.edu/~ronkowit/eliza.html
[2] Weizenbaum, J. (1976). Computer power and human reason: From judgment to calculation. W. H. Freeman & Co.
[3] https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/
[4] https://cajundiscordian.medium.com/is-lamda-sentient-an-interview-ea64d916d917
[5] https://aisnakeoil.substack.com/p/chatgpt-is-a-bullshit-generator-but
[6] https://twitter.com/abebab/status/1597859803287912448
[7] https://twitter.com/sama/status/1601731295792414720
[8] Plug & Pray (2010), Documentary Film
[9] https://artificialintelligenceact.substack.com/
[10] https://www.lexology.com/library/detail.aspx?g=01f3fb00-b63f-461e-8867-29810906fdce
[11] TTC Joint Roadmap on Evaluation and Measurement Tools for Trustworthy AI and Risk Management, 1 December 2022
https://ec.europa.eu/newsroom/dae/redirection/document/92123
[12] https://www.gibsondunn.com/artificial-intelligence-and-automated-systems-legal-update-3q22/
