In 700 words
Is the future of LLMs (large language models) and of AGI (artificial general intelligence) now? or just looming in a thicket of uncertainties?
If the promises of a multitude of commentaries hold, no one knows for sure but many appear to have an inkling about incredible opportunities (and dangers)
To separate the grain from the chaff, naïve questions will shed some light on these novel issues
- Do simple facts regarding AI, undisputed and broadly shared, form a foundational understanding?
- Is AGI – with superhuman reasoning capacity – credible as the ‘next frontier’ to be conquered at any cost (literally) and quite soon?
Simple facts
Language models (LLMs) are probabilistic machines, producing words by splitting language into long sequences of vectors, in a space defined by hundreds of dimensions, making it possible to detect the most probable connections between those words
Assuredly ‘artificial’, the underlying models become ‘predictive’ of those connections with training on massive data sets, but that does not make them ‘intelligent’: they will never be
The mathematical concepts are straightforward, but the sea of data accumulated to train the models and ensure predictability, and the massive investment in datacenters to turn ‘black board’ assumptions into reality, remains impossible to fathom
Looking past the legitimate excitement generated by such a monumental breakthrough, the business plans supporting investments of billions of dollars (turning into trillions?) are wanting
In confronting Business software with AI agents, the market seems to intuit, that indeed AI apps, precisely formatted for tightly defined contexts, will upend business software niches, generating new revenue streams for AI – likely to be built-out over time, starting now
However, by taking on the large consumer market, Open AI and its ilk compete to make themselves globally indispensable, and dominant – raising hard questions
- The step from “nice to have” AI gimmick to indispensable feature of everyday life has yet to be taken by the targeted consumers
- Costs of individual AI requests do not scale the way they do in Google searches for instance, and have to be computed and paid for one at a time, putting the onus on AI providers for now...
- Amortization of the hundreds of billions sunk in ever more - and ever more advanced datacenters – require shortened timeframes (5 years, maybe less…) as Nvidia
launches GPUs an CPUs upgrades at a rapid pace - putting urgency on user adoption rates, which may or may not, come true...
...assuming LLMs prove themselves to be reliable at all times, which is not true (yet)...
Superhuman intelligence, neither intelligent nor ‘human’
To save the day, Sam Altman, CEO of OpenAI, has defined the 'next frontier' of artificial general intelligence (AGI) as “highly autonomous systems that outperform humans at most economically valuable work.”
Mr. Altman’s statement suggests that AGI would expand the range of LLMs’ vector analysis by exploring the trove of accessible research, to provide new insights on causality chains and to offer guidance for novel combinations between established scientific fields
In his telling, the use of brute force relying on immensely powerful computation power does not require the input of human ‘intelligence’ to make potential breakthrough discoveries in medical research, climate change and many of the most vexing issues of our time
Probability driven AGI computation does not align with ‘reasoning’ in its ability to 'see' patterns and connections beyond the scope of the human mind
Reasoning, at pole opposite, is not ‘probabilistic’
Reasoning is circumscribed by the search of an 'equilibrium' to reach the most sensible compromise solution, accounting for multiple factors under consideration…but not in the infinite vector dimensions of AI
AGI, in the stead of LLMs, is predicting the probability of the next token, a process of ‘feed forward’ neural networks, in a space with exponentially many options, of which only one is correct
Such a feat appears to be close to miraculous - finding a proverbial 'needle' in a multi-dimensional haystack...
Taking on the immensity of available research, collected indiscriminately in many formats, text, video and voice, AGI may not be disputed conceptually but practically, today, has been, and remains, severely questioned
Even so, promises of AI system ‘outperformance’ remain real in a 'pedestrian' way, deployed to address specific problems within boundaries and on selected data sets
AI agents competing with business software services might be showing the way...
Such frameworks allow for clarity in the search of an optimized solution of a well defined problem (the target)
Frameworks, boundaries and selected data sets will carve out specialized AI applications
These trends forshadow an evolutionary world, not the revolution promised (and still seriously in doubt) of all-encompassing LLMs and AGIs
The stakes for the near future of AGI - and for U.S. research - have never been so large, as immense sums continue to be invested in the slipstream of U.S. driven 'Stargate'
Sharp differences between the two leading drivers of AI implementation, the U.S. and China, might be setting the stage for just such a showdown between AGI and constrained frameworks, as I hope to discuss soon