We have essentially two approaches in computing, deterministic and probabilistic, whose mathematical counterparts are respectively algebra and statistics.
Some problems, by their nature, need deterministic solutions. Some problems, by their nature, only have probabilistic solutions, and therefore need such solutions. Some problems happen to fall at the boundary between the two.
(1/4)
=> More informations about this toot | More toots from jbqueru@fosstodon.org
It can be tempting to believe that probabilistic solutions are new and that most of the history of computing was deterministic, but that's not the case. In fact, Colossus, arguably the first digital electronic computer, was meant to solve arguably probabilistic problems ("find the likely settings of the first two wheels of the encryption key").
It's important to recognize that applying the wrong approach to a given problem can lead to dead ends.
(2/4)
=> More informations about this toot | More toots from jbqueru@fosstodon.org
As an example, don't try to apply deterministic computing to speech recognition. Leave that to probabilistic computing, which does an excellent job.
In the other direction, don't try to apply probabilistic computing to compilation: that's an area that needs isolation, hermeticity, reproducibility.
Passords: deterministic. Auto-completion: probabilistic. Finances: deterministic. Automative automatic lane tracking: probabilistic.
(3/4)
=> More informations about this toot | More toots from jbqueru@fosstodon.org
I simply do not trust anyone who claims that deterministic approaches are the only viable ones, or anyone who claims that probabilstic approaches are the only viable ones. Both have their unique applications, with some overlap where we can have interesting discussions, but neither will ever cover all categories of problems.
And, please, please, please, pretty please, stop claiming that anything probabilistic is inherently intelligent. There's no intelligence in throwing dice.
(4/4).
=> More informations about this toot | More toots from jbqueru@fosstodon.org
@jbqueru generally agree with the sentiment (god knows we could do with a lot less moronic attempts at getting LLMs to solve calculus 101 problems)…
A few quibbles on the details though: plenty of deterministic statistics model. Plenty of cases where the best solution is deterministic and statistical: stochastic finance models are an obvious one, but even genomics, and by extension systems biology, rely on statistical models…
=> More informations about this toot | More toots from deivudesu@mastodon.social
@jbqueru as for adding “intelligence” to the mix: setting aside AGI and other snake oils, and if we’re merely content to look for a capacity to generalise, or do transductive “reasoning” (/inference) then the Bayesian and other probabilistic approaches, are quite likely to be involved… (but doesn’t mean the reverse is true, obviously)
=> More informations about this toot | More toots from deivudesu@mastodon.social
@deivudesu Yup. I can't claim that I know to recognize intelligence, but, in my domain (software engineering), I can recognize a lack of intelligence (e.g. in job interviews).
-Someone who can't recognize and admit that they don't know something is not intelligent.
-Someone who can't work from first principles (e.g. documentation) and needs to see examples instead is not intelligent.
And I can apply the same criteria to rule out intelligence in systems that claim to replace software engineers.
=> More informations about this toot | More toots from jbqueru@fosstodon.org
@jbqueru @deivudesu
"-Someone who can't work from first principles (e.g. documentation) and needs to see examples instead is not intelligent."
Depends on the documentation
=> More informations about this toot | More toots from bjb@fosstodon.org
@bjb @jbqueru @deivudesu sad but true
=> More informations about this toot | More toots from numeredevs@fosstodon.org
@numeredevs @bjb @deivudesu well, intelligence will make someone notice that the documentation is not good enough, and ask for help.
=> More informations about this toot | More toots from jbqueru@fosstodon.org
@deivudesu Very fair, thanks for reading and commenting.
I think that some of the problems you're mentioning are exactly in that area that falls in between: those are problems that have no known exact solution. At some point those problems amount to optimizing for a metric that we hope reasonably correlates with the quality of the solution, but the hope itself is not exact.
=> More informations about this toot | More toots from jbqueru@fosstodon.org
@jbqueru but what if I run the dice results through some matrix multiplies?
=> More informations about this toot | More toots from StompyRobot@mastodon.gamedev.place This content has been proxied by September (ba2dc).Proxy Information
text/gemini