A few days ago I watched a recording of Michael Wooldridge’s talk titled “This is not the AI we were promised”. The talk was informative and really fun to watch. However, what caught my attention the most was not the technical content of the talk, but the communicational/pedagogical approach prof. Wooldridge took. Consequently, it made me reflect on my own style of communication and on ways to improve it.

From my perspective, the main points of the talk are as follows. The talk opens (at around 6:00 of the video) with a slide containing the sentence “This talk is not an attack on contemporary AI”. Later during the introduction, prof. Wooldridge says “I am not trying to tell you that the today’s AI is a nonsense”, and shows a slide with “Contemporary AI is remarkable”. What follows is a section showing successes of contemporary AI. In the middle of the talk, around 27:00 in the video, he discusses the concepts of “soundness” (an AI is sound iff whenever it gives an answer to a question, the answer is correct) and “completeness” (an AI is complete iff for every answerable questions gives an answer), and how LLMs are neither sound nor complete. Both soundness and completeness are desirable properties of an AI system, and historically, people were interested in these properties - hence the reference to “the AI we were promised” in the title. Around 24:00, the talk is about how LLMs are inconsistent; are incapable of distinguishing knowledge, fact, belief; hallucinate; and are overconfident. In the conclusion, around 47:30, prof. Wooldridge says that “the architecture of contemporary large models is a marvel of contemporary software enginnering”, and stresses that “the central scientific AI problem of our time” is understanding what LLMs are and what they are not, and simultaneously, that “Understanding how they can be safely & productively used is the key AI engineering challenge of our time”.

The content of the talk could easily raise questions like “How this is not an attack on contemporary AI?”. That is perhaps why the introduction explicitly addresses and negates such concern, and why the section on successes of LLMs is such an important part of the talk. It is often said that “before criticize X on one point, one should say at least three points on what they like about X”; also, I have been advised many times by people I trust that when criticizing a piece of research, it is best to frame the weak points as “limitations” of the research, rather than “flaws”. In that framing, I view the talk as a flawless execution of a well-meant criticism of contemporary AI.

Which, of course, raises a different question: is saying that “contemporary AI is remarkable”, while simultaneously reminding us that contemporary AI is also unsound, dishonest? In my opinion, it is not. On the contrary, it is the right thing to do, as it is a way of efectively communicating an important message to us - to people who might be tempted to ignore the information because of our biases and insecurities. Specifically, “ai fans” might be tempted to filter out the informational context of a talk like that, if the talk were too “raw”, because the talk would feel like an attack; similarly, “ai skeptics” might be tempted to focus on “look, an important researcher is saying what I think”, and not actually think about the content.

So maybe I would want to look at more examples of how Michael Wooldridge communicates, and take some notes.