Chatter

I have several emails about ChatGPT, mostly from the feed from the MIT Technology Review, waiting for my attention, but they have been pre-empted by a correspondent telling me how easy it is to use. So I logged on, somewhere in the vicinity of reference 1.

And it is indeed easy to use. You type in some text and it types something back. Sometimes it is spot on, it knows what you are talking about and tells you some stuff, in quite respectable English, that you did not know. Sometimes it does not know and tells you so. And sometimes it clearly just makes stuff up - but in my ten minutes or so with it, the story seemed to be that you needed to pitch your text right to get it to do this. Put in something too outlandish and it stalls, maybe with soothing words about local myths and legends which it does not yet know about. I thought it did quite well with panning for gold in the Ouse, snapped above.

Texts which did produce plausible nonsense included 'Toller play writer', 'battle of Chatteris Cut', 'Huntington 5th hussars' and 'Bulbeck sluice'.

A product which appears to be pitched somewhere in the space presently occupied by the likes of the Bing and Google search engines and the Wikipedia encyclopedia. With the difference that while Bing and Google may turn up rubbish - much less likely with Wikipedia in my experience - there are usually some clues to alert you. Whereas here it is all dished up just the same.

Annoyingly, it failed to find anything about the famous Toller giant and told me so. For whom see reference 3. Maybe in a few years time Openai will be offering a reading service, whereby you pay them to include your stuff in their databases. Rather as you can pay Google and Bing now to appear in their hit lists.

So no doubt clever, but I hope people in high places are worrying about how to manage and control such beasts. The track record with the likes of Facebook, Twitter and (most recently) TikTok being poor.

PS: in Anil Seth's book at reference 4, say in Chapter 5 about wizards, there is much talk of perception being, in some large part, a matter of top-down processing (aka feed back) rather than the bottom-up processing (aka feed forward) favoured (say) twenty years ago. So perhaps here, when ChatGPT is in fantasy mode, it latches onto a few facts - true or false - at the top of the heap and then drills down to generate the details which plausibly join up the dots. Not so unlike what television detectives do in the middle reaches of their stories. Or large florid gentlemen when holding forth to their cronies in the saloon bars of old.

References

Reference 1: https://openai.com/blog/chatgpt/.

Reference 2: https://openai.com/.

Reference 3: https://psmv2.blogspot.com/2014/08/a-disappointment.html.

Reference 4: Being You: A New Science of Consciousness – Anil Seth – 2021.