Dave Rahardja
@skyfire747 @nomenloony The problem with AI is you don’t know when it’s wrong.
Feb 5 at 8:42pmWeb
Show threadElias ProbstFeb 5

@drahardja even AI itself doesn't know, since it doesn't understand, conceptualize, think...

It just replicates and generates what, statistically most likely, matches the recipients expectation.

@skyfire747 @nomenloony

Show threadChinga la migra! πŸ‰πŸŒˆFeb 5
@eliasp @drahardja @skyfire747 @nomenloony It then adds its own slop in to the training pool
Show threadFrank HeijkampFeb 5
@fiend_unpleasant
What could possibly go wrong?
@eliasp @drahardja @skyfire747 @nomenloony
Show threadBigDFeb 5
@eliasp @drahardja @skyfire747 @nomenloony
Exactly.
Show threadπŸ‡ͺπŸ‡Ί Paul Schoonhoven πŸ‹πŸ‰Feb 5

@eliasp @drahardja @skyfire747 @nomenloony
And as ai doesn't understand that people warn others about AI errors, AI thinks it is right.. πŸ™ˆπŸ™ˆ

When you check dossiers over time, you will find that errors get much more attention -as they have to be corrected-. Just skimming the subjects can lead to wrong (opposite) conclusions.

Show threadTomFeb 6
@vosje62 AI doesn't think.
Show threadπ”Έπ•Ÿπ•”π•šπ•–π•Ÿπ•₯ π•Šπ• π•¦π•Ÿπ••π•€ πŸ”‰Feb 6
@eliasp @drahardja @skyfire747 @nomenloony
But in this case it doesn't even achieve the statistically most probable answer (which is surely on gazillions of β€œtraining images”), so EPIC FAIL!
Show threadKennyFeb 6
@ancientsounds @eliasp @drahardja @skyfire747 @nomenloony Thanks to the β€žtemperatureβ€œ that results in β€žmore variedβ€œ responses. ❀️
Show threadSuranFeb 6

@ancientsounds @eliasp @drahardja @skyfire747 @nomenloony

Why?
When mistakes are more often shown then the correct one (because the correct one is obvious to any human and doesn't need explanation), then the wrong one is statistically more likely ;)

Show threadThualFeb 7

@eliasp @drahardja @skyfire747 @nomenloony Just to open the debate...

Define precisely "understand", "conceptualize", "think", and show why AI does not do it.

A theory about life in general is: we are just a set of billions of molecules working together mechanically.

Perhaps we are more efficient to "understand", "conceptualize", "think" than a probabilistic model, but... do not we, all the time, replicate and generate what, statistically and most likely, matches a situation?

Show threadThualFeb 7

@eliasp @drahardja @skyfire747 @nomenloony

Other point of view, less wide. Several philosophers thought that the "thinking" comes with the language. Like, it's not that language is used to share what our brain thinks, but what allows our brain to think in the first place.

And AI talks, clearly. Leaving from there, why wouldn't AI "think"?

Show threadThualFeb 7

@eliasp @drahardja @skyfire747 @nomenloony

I'm not telling any of this to say that AI "thinks". But in my eyes, the debate is way more complex than just "AI dos not really understand, it just computes statistics".

Show threadA Republic, while we kept itFeb 17
@eliasp @drahardja @skyfire747 @nomenloony I like to say that AI is good at generating plausible answers, but don’t use it when you need a *correct* answer.
Show threadTaran RampersadFeb 5
@drahardja @skyfire747 @nomenloony the problem with ai is that if you're smart enough to ask the question, you should be smart enough to find the bullshit answers. 🀣
Show threadJuan PerΒ’ent,πŸ‡²πŸ‡½ πŸ‰,πŸ‡»πŸ‡ͺFeb 5
@knowprose @drahardja @skyfire747 @nomenloony Feeling lazy? Use AI.
Show threadTaran RampersadFeb 5
@dacig @drahardja @skyfire747 @nomenloony feeling lazier? Get another ai to prompt the ai. 🀣
Show threadTaran RampersadFeb 6

@dacig @drahardja @skyfire747 @nomenloony And before anyone beats me to it, I'll up it to multiple AIs prompting each other and shout, #Moltbook 🀣

Where people are even too lazy to interact with each other. So their AIs talk to each other.

Show threadyour auntifa liza πŸ‡΅πŸ‡· πŸ¦› 🦦Feb 5
@drahardja @skyfire747 @nomenloony πŸ›Ž πŸ›Ž πŸ›Ž
Show threadRaineer  πŸ¦†Feb 5

@drahardja @skyfire747 @nomenloony @blogdiva

My favorite quote about LLMs is that they are extremely effective for people who already are domain experts on the topic

The UX fail of the century is for every one of these companies to not append β€œteach me how I would verify this is true?” to every single prompt.

The skill it takes to use the tech is higher than the avg user is going to invest. Every single maker of chatbots has ignored this from day one.

Show threadPete Alex HarrisπŸ¦‘πŸ•ΈοΈπŸŒ²/∞πŸͺβˆ«Feb 6

@raineer @drahardja @skyfire747 @nomenloony @blogdiva

Domain expert here: they aren't extremely effective for me in my field (programming) because to tell an LLM precisely enough what to do, I'd have to use a verbose English description, and I've already constructed the solution in my head in a more structural form that isn't words. Turning it back into words to tell an LLM what code to generate is a step backwards.

Show threadMatt Hardy 3.11 for WorkgroupsFeb 6
@petealexharris @raineer @drahardja @skyfire747 @nomenloony @blogdiva Don't you do that anyway when you add comments to your code?
Show threadPete Alex HarrisπŸ¦‘πŸ•ΈοΈπŸŒ²/∞πŸͺβˆ«Feb 6

@technicaladept @raineer @drahardja @skyfire747 @nomenloony @blogdiva

No, a duplicate explanation of what the code is doing is not what comments are for.

Comments are more for "why" one method or choice than another, sometimes "why not", than "what".

Except doc comments, which are closer to spec i.e. "what". The code itself is "how", and it doesn't necessarily come via a verbal description in natural language.

Show threadyour auntifa liza πŸ‡΅πŸ‡· πŸ¦› 🦦Feb 6

comments have to be about the why but also the WHERE, when you're patching. you want to remember not only why you are patching, but where you got the patch. this has saved my ass soooo many times when doing Drupal upgrades. unfortunately, due to the modularity, not all patches for modules can be replicated with their default installation, so keeping a comment on the patch saves me debugging time.

@petealexharris @technicaladept @raineer @drahardja @skyfire747 @nomenloony

Show threadbufalo1973Feb 6
@petealexharris @technicaladept @raineer @drahardja @skyfire747 @nomenloony @blogdiva just like you don't explain to the client why you use a nail and not a screw when building a chair. You explain how to use the chair and if there's something that works different to a normal chair.
Show threadJonFeb 21

@petealexharris to point out, this is exactly the plot of _Danny Dunn And The Homework Machine_ (1958), in which Danny gets access to a new and powerful minicomputer and he and his gang of nerds try to use it to do their homework for them.

https://en.wikipedia.org/wiki/Danny_Dunn_and_the_Homework_Machine

Danny Dunn and the Homework Machine - Wikipedia

Show threadAdvanced Persistent TeapotFeb 6
@drahardja @skyfire747 @nomenloony @blogdiva @raineer they don't care whether users are *effective*. They care whether users are paying.
Show threadJeff McNeillFeb 5
@drahardja @skyfire747 @nomenloony that's the problem with content in general, unless using actual experts, which Google pushes down the page because that makes it more lucrative for them, showing more ads.
Show threadJustinFeb 6
@drahardja
Which is quite often it seems.
@skyfire747 @nomenloony
Show threadKevin RussellFeb 6

@drahardja @skyfire747 @nomenloony

Always wrong, in many ways levels and categories including wasting lives.

Show threadGerrit van AakenFeb 6
@drahardja @skyfire747 @nomenloony only let AI do things you really could do by yourself, for saving some time. Control the result!
Show threadJames TweedieFeb 6
@skyfire747 @nomenloony @drahardja So you have to assume it is wrong. And then look up an Electricians web site which shows you how to wire a plug.
Show threadPaul CantrellFeb 6

@drahardja @skyfire747 @nomenloony

Yeah. I have a set of dice that get it right some percentage of the time