The hidden chain-of-though inside the process, from the official statement about it, I infer / suspect that it uses an unhobbled mode of the model, puts it in this special mode where it can use the whole training, avoiding the intrisic bias towards the aligned outcomes.
I think that, to put it in simple terms, "the sum of the good and the bad" is the secret sauce here, pumping the "IQ" of the model (every output in the hidden chain), to levels apparently a lot better than they could probably reach with just aligned hidden internal outputs.
Another way of looking at the "sum of good and bad" stuff, is that the model would have a potentially way bigger set of choices (probability space?), to look into for every given prompt.
IT salaries began to go down right after AI popped up out of GPT2, showing up not the potential, but the evidence of much improved learning/productivy tool, well beyond the reach of internet search.
So beyond, that you can easily can transform a newbie into a junior IT, or JR into a something ala SSR, and getting the SR go wild with times - hours - to get a solution to some stuff that previously took days to be solved.
After the salaries went down, that happened about 2022 to the beginning of 2023, the layoffs began. That was mostly masked "AI based" corporate moves, but probably some layoff actually had something to do with extra capabilities in improved AI tools.
That is, because, fewer job offers have been published since maybe mid-2023, again, that could just be corporate moves, related to maybe inflation, US markets, you name it. But there's also a chance that some of those fewer job offer in IT were (and are), the outcome of better AI tools, and the corporations are betting actively in reducing headcounts and preserving the current productivity.
The whole thing is changing by the day as some tools prove themselves, other fail to reach the market expectations, etc.
ASI is the endgame where it is profitable to be in the OpenAI position, or even in the next first 20 market players capable of getting there a bit later.
But if ASI isn't achievable finally, the intelectual properties obtained in the way to it will probably be valuable, because the SOTA still works and can be re-reployed anytime in the future when the new hardware becames available and cheaper (think Cerebras stuff level). We would be in a new kind of AI winter, just waiting a couple of years till spring breaks up (cheaper, faster hardware).
Even in that winter, bigger players would still be available, think Gemini or Copilot products, getting bigger, better year after year during the winter, just as fast as the new hardware begins to be buyed/deployed. And minute by minute the market share for those bigger players will playing along with bigger reveneaus every quarter, preparing the way to full profitability in a couple of years.
Think automobile industry or oil extraction industry, going from manual work to fully machine assisted tasks, as the technology became available, from the 1900s to the 1970s. Quite a lot of years, but in the AI winter probably coming, you get even the chance to double check the countdown every now and then, just looking at what is cooking/assembling Nvidia/TSMC, Cerebras, etc.
Claude Sonnet 3.5 at least works awesomely, you can just talk to it asking stuff, it will infer your knowledge level from your questions and start answering according to it, proposing a follow up path for further insigths about the subject. If you follow it up, usually it will take you to a somehow similar path as you would follow if you just happen to having been reading the wikipedia page for the subject you're asking Claude to explain about.
but if the subject isn't something that obvious as something you can find in the wikipedia, you'll be good too, Claude will take to a sort of "shortest path algorithm" of knowledge about the subject you're looking about.
if you side it with web searching, you'll see it takes some few keywords, concepts about the subject, and explains them, and you can go deeper searching them and looking other sources (blog posts, answers in reddit, etc.).
In Claude I found almost no hallucinations in the deepest explanations it answered in some research I've done with it, maybe some non-human focus on trivial details while not looking at more relevant stuff about the subject (I infer that the training data could be cramped with the less deep data, more people answering about the subject on internet from a shallow level of knowledge than a few experts answering really good explanations > you'll find these answers first when you begin web searching).
The industry always has information before-hand, all those AI capable datacenters aren't being built just because a hunch.
It is possible that the next iteration of GPT4 level technology has already ocurred a year ago, august 2023. The Q* thing was built around that time, probably went online some time after that. Word "in the street" says it is an exponential jump from GPT4 level of "AI cognition".
The whole conversation - papers, press articles - about datawalls, billions sunk "for no reason", etc. could be just empty words since maybe october 2023.
Yes, this happens, there's happening some throttling, I've seen questions like this one regarding the same issue across several LLM providers ("works faster, better, solves better at night").
Lots and lots words flow about this. For me, it is very simple. LLMs do a complex process, quite analog to human thinking/understanding/speaking/writing, but they're doing all those - most probably - in an alternative way to what humans do in their brains.
Directly comparing LLM's outputs with human output is like comparing a F22 flying with an eagle flying. Both fly obviously, but using entirely different processes to do so (different requirements, capabilities, despite the simplest similarity of both systems - the eagle and the F22 - at "doing fly").
You don't automatically say "an eagle is more capable at flying than a F22, because it flies with very little energy requirements while deploying quite better, reliable take-off / landing capabilities".
You actually don't usually go comparing these systems just because both can fly.
but many out there are pulling their hair trying to compare side by side the obvious mathematical systems that LLMs are to - most probably again - the completely different in nature systems that humans are.
nor OpenAI or any of the prompt-based AI companies actually "need" the reveneu from the services they sell, the whole point of having a public (free or not), prompt facing the entire planet is just having live humans doing RLHF 24x7x365, that information, that dataset is more valuable than any symbolic amount of money anybody is willing to pay for any GPT or clone suscription.
Anybody noticed already that any current or near future reveneu won't make a dench in the actual costs of running giant models, anyway, the models keep chugging along just fine. And the ("free") money keeps flowing in.
Moroever, those millions of non-paying clients prompting the models are 24x7x265 working with 100% real world problems are inputing the models with valuable prompts, generating valuable content (originated in real situations, actually distilled by unvaluable billions of human sensory input).
That content can be and is used to train models, effectively cancelling the "data-wall", bit by bit, all the time.
There’s so just much missing information or changing contexts. Forget any specific model. You will not become particularly good at any domain specific task by reading prompts of people asking questions. That data sucks.
Not really, China is just a step behind, in a year they will be at current US AI state-of-art, without competition, from there they'll have all the GPUs in the world to keep improving their models.
Or most probably, US national interests will step in, just like it did with SpaceX, and will provide any amount of billions required to keep in the fight, at least till China steps out of the AI race.
Beside the (naive?), naysayers, until the technology actually shows it is failing to keep the promise of super powerful AIs, it is a global, geopolitical race to get there (AGI/ASI), or till the point it fails (new AI winter).
To make quite sure it really doesn't work, and not dropping out of the race just to see a chinese ASI (Artificial Super Intelligence), emerge a couple of years later.
China isn't even close to state of the art. Take a look at the chips they put out, the Chinese repeatedly “claim” they are building chips equivalent to a 4 or 5 year old Intel chip (Which might as well be a century ago in terms of technology) until a tech YouTuber in the US actually gets their hands on one and realizes the benchmarks are completely fabricated.
the prompt interfaces + smartphone apps were (from the beginning), and are ongoing training for the next iteration, they provide massive RLHF for further improvements in already quite RLHFed advanced models.
Whatever tokens they're extracting from all the interactions, the most valuable are those from metadata, like "correct answer in one shot", or "correct answer in three shots".
The inputs and potentially the outputs can be gibberish, but the metadata can be mostly accurate given some implicit/explicit (the tumbs up, the "thanks" answers from users, maybe), human feedback.
The RLHF refinement extracted from getting the models face the entire human population for to be continuously, 24x7x365, prompted in all languages, about all the topics interesting for the human society, must be incredible. If you just can extract a single percentage of definitely "correct answers" from the total prompts answered, it should be massive compared to just a few thousands of QA dedicated RLHF people working on the models in the initial iterations of training.
That was GPT2,3,4, initial iterations of the training. Having the models been evolved to more powerful (mathematical) entities, you can use them to train the next models. Like is almost certainly happening.
My bet is that one of two
- The scaling thing is working spectacularly, they've seen linear improvement in blue/green deployments across the world + realtime RLHF, and maybe it is going a bit slow, but the improvements justify just a bit more waiting to get trained a more powerful,refined model, incredible more better answers from even the previous datasets used (now more deeply inquired by the new models and the new massive RLHF data), if in a year they have a 20x GPT4, Claude, Gemini, whatever, they could be "jumping" to the next 40x GPT4, Claude, Gemini, a lot faster, if they have the most popular, prompted model in the market (in the world).
- The scaling stuff already sunk, they have seen the numbers and it doesn't add by now, or they've seen disminished returns coming. This is being firmly denied by anyone on the record or off the record.
I think that, to put it in simple terms, "the sum of the good and the bad" is the secret sauce here, pumping the "IQ" of the model (every output in the hidden chain), to levels apparently a lot better than they could probably reach with just aligned hidden internal outputs.
Another way of looking at the "sum of good and bad" stuff, is that the model would have a potentially way bigger set of choices (probability space?), to look into for every given prompt.
reply