mjburgess

88 ms·

1321.
by mjburgess 3y ago
Lot's of people in this thread seem desperate to not think clearly about this. Whether these tests, verbatim produce the same response on any given version isn't the point. GPT4 doesn't engage in reasoning if it gets any answ
1322.
by mjburgess 3y ago
We here means "our species".
1323.
by mjburgess 3y ago
Reasoning by analogy is one of the ways we solve the framing problem. So, when explaining the universe we imagine it's an act of will by a conscious entity (ie., like how we invent). When explaining the mind we suppose it's like o
1324.
by mjburgess 3y ago
Yes, there is a dynamical equilibrium between engineering and science; with "analogical engineering" often occurring before "direct explanation". So, accidental discoveries -> lenses -> optics -> telescopes -&gt
1325.
by mjburgess 3y ago
I don't think the claim would be about quarterly profits. It has had profitable quaters before. Consider, eg, >Uber on Tuesday posted a profit of $394 million during the second quarter, compared with a loss of $2.60 billion a year e
1326.
by mjburgess 3y ago
It's descriptive, not normative. Engineers build "useful models", scientists as a matter of fact , build explanations. Whether those explanations prove useful is always an open question; newton's took centuries to have
1327.
by mjburgess 3y ago
You've misunderstood the point. The models we used do not predict data we have observed about almost everything They do predict data counter-factually: if a rocket were to do X, then Y would happen. They do this by a model of how the w
1328.
by mjburgess 3y ago
The last illusion sceptics are subject to is scepticism itself. The final illusion is that we are separated from the world. Without this premise, hume has nothing to say
1329.
by mjburgess 3y ago
I know of no such scientific model. Indeed, almost all of them do not predict observed data. Newton's universal law of gravitation didn't predict huge amounts of stuff in the solar system. Rather than say it's wrong, we suppo
1330.
by mjburgess 3y ago
Some scientists use statistics, sure -- by "doing science" I mean directly explaining how the world works. The way that (actual,) scientists use statistics is to "choose between possible vases". Suppose we're in pla
1331.
by mjburgess 3y ago
The problem of induction since Hume has been a massive ideological misdirection against the possibility of knowledge. We do not use induction. We use abduction. We first posit a universal, and then each subsequent data point is taken to con
1332.
by mjburgess 3y ago
Ah, then let me rephrase, it's actually this model: > P(A|B,C,D,E,F....) And with clever choices of B,C,D.... you can make A abitarily probable. Eg., Suppose, 'lolita' were rare, well then choose: B=Library, C=Author, D=19
1333.
by mjburgess 3y ago
An LLM is just a model of P(A|B), ie., a frequency distribution of co-occurrences. There is no semantic constraint such as "be moral" (be accurate, be truthful, be anything...). Immoral phrases, of course, have a non-zero probabil
1334.
by mjburgess 3y ago
`tidy` is similar to python context managers, or an automatic `defer`. So any resources within the function have their `.close`s run. `heap` is a semantics of `new`: allocate to the global heap `temp` likewise means allocate to a smaller te
1335.
by mjburgess 3y ago
I've recently arrived at a mixed declaration, which "distributes the information" across the line, I think in a way which allows it to grow without being overhwelming, (which i take to be a holy grail...) fn calc :: (x: i
1336.
by mjburgess 3y ago
> Infinite dimensions aren't needed for any real interaction because the number of particles are bounded, The number of particles in an interaction is irrelevant to the property of linearity requiring infinite dimensions. It is stra
1337.
by mjburgess 3y ago
1) ordinary classical physics is non-deterministic without infinite precision; this isnt observed 2) these "60 or 70" digit -type reasoning omit iterative problems, such as the number of digits required to specify the behaviour
1338.
by mjburgess 3y ago
Well that's a bad principle. Geometry isnt computable, and hence neither physics.
1339.
by mjburgess 3y ago
Well (1), the reason this distinction is relevant is so we can separate out whether the system has developed a capacity or an apparent capacity. Is the child a genius or are they just reading out of a textbook? Can the toddler really compos
1340.
by mjburgess 3y ago
Not shown: advertising revenue; debt load; rise of competing services; probability of bankrupsy over time; number of staff; market cap... It might , perhaps, almost be that a google trends graph isn't a reliable measure of business su
1341.
by mjburgess 3y ago
It isnt just the 9th. It's throughout the constitution, and indeed why it's a philosophical document, not a legal one. The people who wrote the constitution were not trying to enumerate a finite number of laws which were contempor
1342.
by mjburgess 3y ago
> competent at achieving those goals despite not being explicitly trained for them. Well (1) it doesn't achieve goals, since a "goal" is observer-relative. We have goals, the LLM has a formal optimisation objective which
1343.
by mjburgess 3y ago
> I don't believe the nature of LLM is to hallucinate. It's instead a side effect of how we train them. We train them to guess, to be close, but not to be correct necessarily. Thoughout this comment you speak about LLMs as-if
1344.
by mjburgess 3y ago
Here, "supposed to guess" means "having the goal of..." So no LLM knows what it's supposed to do. If you prefer, you could say it only ever has one goal: to generate a sequence of tokens which are jointly the most p
1345.
by mjburgess 3y ago
My point is that it wouldn't even matter. The historical conditions no longer hold, ie., Would "illegal abortion" be a tyranny for women in 1800? Well, it's going to be very poorly enforced; you may die anyway; etc. It&#
1346.
by mjburgess 3y ago
Hence the SC should throw-out constitutional interpretation which places unenumerated rights above enumerated ones --- that would be a genuinely libertarian act. Of course, the SC is only libertarian when it wishes. It's rather mad tha
1347.
by mjburgess 3y ago
Whenever i design a lang, i just give option semantics to types of the form `x|none`
1348.
by mjburgess 3y ago
Yes, this is essentially what i mean by "distributions are sampled from distributions", ie., there's a subjective uncertainty in the choice of model but also an uncertainty in the very determination of that uncertainty. You
1349.
by mjburgess 3y ago
I'm talking to a very rarified audience, for sure. Giving a bayesian gloss on moorean epistemology is not really a project for a hacker news comment.
1350.
by mjburgess 3y ago
It really doesnt matter what the "meta-distribution" is, bar trivial ones. My whole point is that we can augment bayesianism by a-priori choosing these meta-distrbutins. Is my hand in front of me? Is what's in front of me rea

More ›