From AI highway to cul de sac and back
Finding the smarts and the Dunning-Kruger effect in the battles breaking out as artificial intelligence gets a voice
Forget the speed of light, the past few months has given me a new measure for lightning quick innovation and iteration: the speed of AI.
It’s been four months — just four — since the public was able to get hands-on with OpenAI’s Chat-GPT (built on GPT-3, then 3.5), and earlier this month a fourth edition landed, promising better answers, multimedia processing and — crucially — fewer errors.
This is because, as magnificent at text generation as Chat-GPT is, it’s got a helluva imagination on it and has been shown to “hallucinate” ,a spookily humanised way to describe its proclivity for stating nonsense and misinformation with all the authority of straight fact.
This new version, we are told, is more than 80% less likely to be coaxed into creating “disallowed content” than its predecessor, and 40% more likely to create responses that are factual. Those are big leaps and bounds, and definitely in the right direction.
Still, last week OpenAI funder Microsoft was 100% more likely to publish codswallop. On March 22 researchers at Microsoft published a paper — on the arXiv open-access repository —
that claimed GPT-4 was showing “sparks of artificial general intelligence [AGI]”.
This is a big claim, and was never going to fly by unexamined. In simplistic terms, saying a machine or system is capable of AGI means it is capable of cognition on a level akin to humans, solving problems it has never seen before, extrapolating and working backwards to solutions — rather than the incrediblebut-inherently-procedural processing AI currently exhibits.
GPT-4, the researchers wrote, “can solve novel and difficult tasks …” with performance that is “strikingly close to human-level performance” that “could reasonably be viewed as an early (yet still incomplete) version of an ACI system”.
Exciting, ja? Well, it would be if the rest of the research paper didn’t go on to negate these claims — its own claims, to be clear. It has since been criticised for taking this rather nonacademic tactic to publishing, the backlash even prompting Microsoft to release a statement clarifying that with GPTintegration it is “not focused on trying to achieve AGI” but rather “platforms and tools that … help humans”.
Vice magazine has a great rundown of the contents and the failings of the paper if you’re looking for the full story, detailing the disconnect and calling the paper’s title out for the clickbait it is: “The Microsoft researchers write that the model has trouble with confidence calibration, longterm memory, personalisation, planning and conceptual leaps, transparency, interpretability and consistency, cognitive fallacies and irrationality, and challenges with sensitivity to inputs.”
That’s quite a list. In short, Vice continues, “the model has trouble knowing when it is confident or when it is just guessing”. Now why does that sound so familiar? It is not quite the definition of the famed Dunning-Kruger effect, but close enough.
The pair of psychologists proposed in the late ’90s that — and this is a broad-strokes description — those who know the least overestimate their knowledge the most. It is a little more involved than that. Plus, the research on which this
observation is based has been challenged recently. Still, the study and its pithy findings landed at just the time to define our response to internet comment sections ever since.
So, even if we can no longer apply it blithely to people, we will always have the early hallucinations of Chat-GPT to hang it on. Microsoft’s “AGI sparks” research paper may be already ignoble, but it is a technicolour demonstration of the push and pull of both commercial and academic interests in developing technologies — something we are going to have to get cleareyed about.
Last week Twitter and Tesla chief Elon Musk took aim at his former collaborators at OpenAI
— an organisation he helped to found — on these grounds, saying it had gone from “open source nonprofit to a closed source forprofit”. He also tweeted critically of the relationship with Microsoft, sharing a meme about “realising AI, the most powerful tool that mankind has ever created, is now in the hands of a ruthless corporate monopoly”.
It is not lost on this writer — and other commentators — that this is coming from the same man who is monetising all the features of the tool he once called “a common digital town square”, the very thing he now aspires to turn into “the most respected advertising platform in the world”. Nor does it escape my attention that Musk tried and failed to regain control of OpenAI a few years ago.
So, is this sour grapes or is it valid critique? Yes. There are a few points of fact — those tricky things that cannot be dreamed up, but rather uncovered — that Musk neglects: that Microsoft is a huge investor but doesn’t have a seat on the OpenAI board, and that OpenAI is not claiming to be a nonprofit. But I am loath to write off the valid element of his point, even if he does give the impression of sulking over his lost seat at the cool table, or his turn to play with the shiny new toy in the tech playground.
For what it’s worth, OpenAI’s Sam Altman considers it a “capprofit”, explaining that though it realised that scaling up necessitated corporate cash, the return on investment those investors can expect is not infinite even if AI’s potential is. Altman isn’t taking the jibes to heart, telling tech analyst and podcaster Kara Swisher that Musk was just being a jerk — his words, not mine. “Most of that is not true, and I think Elon knows that,” Altman said on Swisher’s podcast late last week.
Whatever tech leader you choose as your champion in this week’s fight, we must concede that there are valid reasons to be concerned about how funders can guide and shape projects, especially when these projects are — without hyperbole — transforming the world.
THE NEW VERSION OF OPENAI’S CHAT-GPT IS OVER 80% LESS LIKELY TO BE COAXED INTO CREATING ‘DISALLOWED CONTENT’ THAN ITS PREDECESSOR