reified intelligence on a chip, almost too cheap to meter,
“On a chip” is also stretching the truth. The kinds of models that you might point to in support of “reified intelligence” run on things the size of a desktop computer and cost more than a car, which is way off from the scale that “on a chip” suggests.
And “almost too cheap to meter” is aggressively false. Users of this service are known to talk incessantly about being metered, it is a daily fact of life for them. Individuals who have free usage for a project commonly report that, had they paid, it would cost five or six figures. Companies have seen enormous bills, some approaching the size of their payroll. And all of this is true for tokens that are dramatically subsidized, by one of the most intense and largest concentrations of capital in history. It is the polar opposite - “almost too expensive to even do, and absolutely must be carefully metered”.
In summary, let us indeed not forget what we got back for this: extraordinary distortions of reality evenly intermingled with bald-faced lies.
> It is certainly closer to intelligence than anything we’ve come up with before, it can certainly match or beat actual intelligence in some domains, but just as certainly it clearly lacks many features and capabilities of intelligence.
It's not complete or that well-rounded. But it's something that was the domain of speculative science fiction only 5 years ago, and it's rounded enough to be applicable to ~everything to some degree.
> The kinds of models that you might point to in support of “reified intelligence” run on things the size of a desktop computer and cost more than a car, which is way off from the scale that “on a chip” suggests.
By "on a chip" I meant more "in silica" than literally on a single chip" - though this actually is* true, but those chips aren't cheap.
> And “almost too cheap to meter” is aggressively false. Users of this service are known to talk incessantly about being metered, it is a daily fact of life for them.
You are looking at power users that use LLMs in agentic coding sessions. Most people just run off free tier of ChatGPT, which is free for them. There are equivalent open-weight models at this level, and while hardware to run one for yourself is expensive even for most westerners, the marginal inference cost is literally dirt cheap, which is why you can get that for near-free from smaller inference providers - or pony up some money, rent a bunch of compute with friends, and become an inference provider yourself.
(It's only a tough market because the major vendors are giving out better models than you can run for ~same or lower price than you can offer. Which either way is too cheap to meter in terms of solving useful problem for real people. Again, developers are a special case of power users, as usual.)