One could refer to old-style flash powder lamps as "firecrackers" - which did give usable light for photography, quite analogous how a current LLM inference process generates brief "flashes" of intelligence, each time devouring the context window and spitting out a chunk of tokens. Just like with combustion engines or digital processors: at some point, a fast-cycled process is practically indiscernible from a continuous process.
For current transformer architectures, the weights don't change but the context window is where "it" "lives", so the "only" thing missing for continuous learning is either infinite context or another persistence mechanism.. something akin to "dreaming" has been proposed f.e. .. Pretty likely somethign will make it into roll-out within a year, probably crossing more thresholds into AGI territory.. Of course that will trigger more goal-post moving, but I don't see how even sceptics would come to the conclusion in 2029 AGI hasn't been reached.
This is the relevant historical context to understand how this ever got into big G's mission statement, it was to distance themselves from the then well-known anti-competitive mal-practices of the Redmond giant.. which seemingly they've learned a lot from. Don't no-one worry too much, they only have less than half of the world population's messaging metadata, schedules, realtime location and private pictures.
It was absolutely the end of the world for many folks not as privileged by sheer chance of birth place. And getting through this "just fine" will mean hundreds of thousands of starvations in addition to the "default rate" of ~10k just children dying of malnutration each day. So yeah, congratulations to your luck that you can just shrug it off as insignificant for you personally..
reply