Lightning Development of AI or Will GPT Get Obsolete Soon?

Unlock the Secret Potential of ChatGPT Plugins! [Video]

Video uploaded by The AI Advantage on May 18, 2023

Lightning Development of Artificial Intelligence Advancement or Will GPT Get Obsolete Before Its Prime?

While we don’t agree with the current “AI panic” – with all due respect to the signatories of the recent petition to hit the brakes on its development – it’s starting to look like their worries are not completely without reason. The leaps and bounds with which this particular technology is progressing actually do make one stop and consider the possibility that our human intelligence could indeed fall behind the artificial one in that race, and one day we’ll develop it beyond the point where we’ll understand what to do with it.

See for yourself. Just days ago the founder of a new tech firm introduced a device which, he claims, will eventually drive the smartphones as we know them into obscurity – first and foremost thanks exactly to yet another AI-based function – namely, life translation, enhancing on “mere” ChatGPT conversation. (We covered this topic as well). And now the GPT software (and concept) itself appears to be threatened by the alleged next step on this road. At least, according to the newly-published joint work by the Canadian MILA and Stanford University researchers, including such AI celebrities as Yoshua Bengio and Christopher Re, who announced a totally new code, ominously named Hyena. They claim that even the latest version of the aforesaid GPT – the GPT-4 – will crash and burn on the amount of info that their tech beast can handle easily, without even a slight lapse of speed or any memory restriction.

The Hyena code. This new technology could blow away GPT-4 and everything like it.

And it’s not just a quantitative improvement (like a speedier CPU or more powerful search engine) which is behind it, but, in true intelligent spirit of the subject, a whole new approach, or, as the authors call it, a different “hierarchy” of the search and selection. (They liken this way of getting data to the pecking order of hyena packs, hence the title). Whereas the traditional GPT (sounds oxymoronic, doesn’t it?) uses, so to speak, their own “stationary” base of knowledge (words, sounds, images) which, simply by the laws of nature, is finite and gets cumbersome with the growing amount of stored info, the Hyena exercises another way. Namely, something that its creators – first of all Michael Poli, whose name is at the top of the paper in question – call “convolution”. Or, in (relatively) layman terms, combines the usage of all these massive data sets with applying an array of filters, depending on the fields / topics of searches and queries. And thus the so-called “quadratic complexity” (meaning, simply speaking, running out of space and resources) is broken: these filters, thanks to their ability to vary spontaneously, dramatically reduce the size of both the data pool and the amount of pre-set parameters.

Ironically, this particular convolution method, on which the topical breakthrough is based, is hardly a new idea: it goes back as far as the 80s, when this type of filters were invented and perfected. But in a way, it also looks in character with the idea of the advanced intelligence, artificial or not, being an example of inter-dimensional thinking: after all, time is a dimension too, is it not? Just like Poli, Re, Bengio and others merged the utilization of existing databases with flexible filters usage, they integrated the current cutting edge technologies with the invention of the past… and came up with something game-changingly new as a result. Talk about an effect surpassing the sum of its parts.

Just like Poli, Re, Bengio and others merged the utilization of existing databases with flexible filters usage, they integrated the current cutting edge technologies with the invention of the past… and came up with something game-changingly new as a result.

And mind you, this team’s claims are already backed by some practical tests. Poli and co. run quite a few comparative searches using several AI programs (i.e. Transformer, SuperGLUE etc.) including their new software – and the latter is reported to achieve equal results on a drastically smaller dataset with the same (if not greater) speed. Which may extrapolate to a really revolutionary AI system: the one that can deliver answers and information of whatever kind (textual, audible, visual) almost instantly regardless of location, and draw complex and far-reaching logical connections in the process. Sounds quite scary indeed, does it not?

But at the same time, it also sounds exciting. And since – despite all the efforts of the above-mentioned distinguished experts – there appears little to be done to put that rather smart genie back into its bottle, we personally choose to be excited rather than scared. With all due sympathy to poor, soon-to-be-outdated (not likely, in fact) GPT. Especially considering that existing – and advanced – artificial intelligence does not yet cancel out the human one. Of which the AI developers are certainly not the worst examples.