Time: 2025-07-11 11:15:24 Source: Codora.ai Author: rust Reading: 864 times

would like tounderstand what how good what theprobability confidence in any particularresponses which right now okay yeahright now it doesn't tell you right itjust gives you an answer know yeah Inever thought about that yeah so but sarah cindy white killerIguess like when you're usinguh uh models direct directly would theygive you probability or or would youhave to do some uh some uh postcalculation to get that in predictive AImost models give you some sort ofprobability of confidence some sort ofmeasure of how well it thinks itdid but that's that's but we weretalking about just general models likeany kind of ml model is that what you'resaying but with LMS would they have towould they also explicitly return themas well or no so uh most llms do notbecause they're ating sequences wordafter word they do not tell you whatthe what the modelsum how well the model thinks it did butin other probably if you're doing aclassification for example it would tellyouum so I guess that's a good thing it's agood thing that you said there becauseuh when we talk about um llms you haveto everyone talks about having uhevaluations right but is it because thenature of LMS makes it hard for it toproduce that number so you have to havean additional step whereas when you'reusing a classical machine learning modelit's producing a single value and thenit can also give you the probabilityback at the same time kind of so inpredictive machine learningtraditionally it was very task specificso you trained it on a very specifictask and so it learned that task and itgave you some sort of metric of itsperformance on that result nowgenerative AI tends to be andfoundational model llms tend to be moreum their general purpose so they can doa lot more things and they can dodifferent types of things beyond thethings that they were um trained to dothat's where all these evaluationmetrics come in is is if you see and wetalked about this at some point um isthey would tell you what how does itBenchmark on coding how does itBenchmark on reasoning how does itBenchmark on accuracy um because theythey they can do a lot more things thanwe train them on because they're veryvery large models and so there'semergent properties that come withthat um andso I'm just going to iterate here thatAI power assistance is I think where alot of people uh start learning it butuh what happens is that when you'reusing AI power assistance you find outthat it doesn't have a programmatic wayfor you to start implementing it foryour own Solutions and you know thepurpose of this course is to show folkshow to go beyond the AI power assistanceof course you can leverage this courseto do the best that you can do with apowered assistance but um you reallywant to move Beyond in this kind of parthere I almost I feel that I should havea little line here because this is wherethe real story starts right after oopsthat's not what I wanted but if I can uhget this open here it's not I don't knowhow to get my shapes here's my shapesbut I feel that right here uh yeah if Iif I was to draw a line right here thisis where things start to get seriousit's obviously an arrow not a line umand this is where you start get intoworkbenches and playgrounds surpristechically I showed where all theworkbenches were and playgrounds were inthese various Solutions but I didn'treally use them I just skipped rightover to code because there's not much totalk about they're they look a lot likea assistance they do have more uh knobsuh knobs to turn like uh what top K andtemperature and things like that onceyou learn one they all kind of feel verysimilar but uh some of them will haveinterfaces to their extended uh extendedfunctionality so like if you're using aCO here they would have a little box forthe response format for Json if you'rein um uh anthropic uh then if you wantto do tool use you could do that inplace and some of them will generate atthe code for you so you can startutilizing them and sometimes they generout good code I can't say they always doI think in some of them they're actuallyproducing older code I'm actually reallysurprised um with how many AIServices uh like they'll have a versionto uh API and and then their mainservice that Jour toota is still doingversion one uh and it almost kind offeels like uh these companies kind ofeven keep up with their own their ownchanges which is really surprising umbut I'm not sure what your experience isabout um documentation and how howaccurate it reflects the currentversions of things usually they'rethey're behind I think part of it is thethe speed of things but also I thinkpart of it is cost right the the oldermod model is likely smaller and it'slikely cheaper uh and so some some of itiscost um so you know once you startworking with things programmatically atleast in this way before you even startwriting code you might start to thinkabout what kind of models you want touse and there's all kinds of models outthere there's foundational models wehave models that are fine- tune whichtechnically Foundation models are finetune models models that are instructmodels we have embed models we have opensource models we have Edge optimizedmodels or we could call them slm smalllanguage models or we could call them ondevice models or we could call themsomething else there's a lot of namesfor this one in particular um and thenwe have uncensored or you could sayunrestricted I don't know the worduncensored doesn't sound good butunrestricted models uh and we have a lotof models why do we have so many darnmodels um so I think these are alldifferent terminologiesto um to describe different metadataabout the models so open you know opensource models are models where theweights are open um plus the data plusthe code open weight models are modelswhere the weights are open but not thedata and the code so each word descrieach term describes a veryspecific scenario so Edge

(Editor in charge: c)

Previous article: Learn Blockchain, Solidity, and Full Stack Web3 Development with JavaScript – 32-Hour Course - Ep154
Next article: GenAI Essentials – Full Course for Beginners - Ep186

• 
• 
• 
• 
• 
• 
• 
•