1 – What You Actually Need to Know About ChatGPT

-hello, hello, hello my Curious people and welcome to episode 1 of ChatGPT curious. I’m your grateful host, the Maestro, and today we are talking about what you actually need to know about ChatGPT.

-wanting to give you some context around it without losing you to the tech terms and minutiae, but i do think it’s important to have a decent foundation for wtf this thing actually is
-allows to experiment better with it
-allows you to criticize it better
-less scared of it
-allows you to make more informed decisions about it

-if you’re listening to this podcast there’s a good chance you’ve at least opened ChatGPT, if you haven’t THAT’S OK
-i want you to pause this, go to chatgpt.com and type in the following question: “What 
are you and what can you help me with?”

-the 30,000 overview as to what chatgpt is: it’s a computer program. You type something in, and it responds. 
-You can ask it to explain something you don’t understand, help you write an email, 
brainstorm dinner ideas based on your picky ass kids and what’s in your fridge, or make 
a packing list for your trip. 
-It’s not magic, it’s math and i will explain what that means in the rest of this episode. It is 
also not perfect, far from it. It makes mistakes, but it can be surprisingly helpful.


-brief history
-parent company, Open AI was founded in 2015 by Sam Altman and the devil himself, 
Elon Musk
-will probably do another episode about the origin/early days of Open AI but 
sticking with the theme of this episode, suffice to say that Elon left in 2018 
-2018-2022 OpenAI built different versions of the GPT (GPT means generative pretrained transformer), with version 3.5 being released to the public on November 30th, 2022 as ChatGPT
-we are currently (as of july 25th when this is recorded) on version 4.1 of ChatGPT with 5.0 rumored to be released in the near future)

-what is the difference between these versions (1, 2, 3, 4)
-namely the number of parameters that each has
-But from GPT‑4 onward, it’s not just about size anymore. It’s about architecture, speed, 
reasoning, and what the model can actually do.

-Lets talk about Parameters
-first off, LLM, which is what ChatGPT is, stands for large language model
-an LLM is a trained program that runs on incredibly powerful computers
-the LARGE in LLM refers to the number of parameters that a model contains, NOT the 
size of the data set (or training corpus) that it was trained on
-ChatGPT 4o has possibly 1 trillion parameters (not disclosed)

-parameters – “The connections between patterns in language with a weight assigned.” 
-(grammar, tone, rhythm, logic)
-these weight (strength of that connection) is adjustable by the 
Model
-during the training process, an LLM g ets fed a ton of content (training corpus), basically everything on the internet. 
-Next it is presented with parts of a real sentence from the training content and asked what comes next? 
-The sky is_____?
-The model makes a guess – red
-The model checks what actually came next: blue
-The model measures the difference and uses math to change the 
parameters – aka the strength of the connections between patterns it 
identifies in the language) so that the next time it sees similar content it is 
more likely to guess the same word that appeared in the training corpus
-can see that the more parameters (connections between patterns in language) we have, the more fine tuned and “correct”  the response will be
-because all of the relationships have been analyzed
-the sky is blue
-the sky is red during a beautiful sunset
-this fine tuning (adjusting of parameters) occurs during “pre’training” and then is frozen prior to being deployed for use
-no changing of parameters happens when we use chatgpt
-last thing add to conversation about training is that part of the training that teaches the model what responses are more correct based on what humans want to see, requires humans, process called RLHF – reinforced learning human feedback and is problematic in many ways b/c of the nature of the content that these people are exposed to

-Summarize so far
-ChatGPT is an LLM (large language model). Large refers to the number of parameters it has, and parameters can be thought of as connections that are specific to different patterns that exist in language and can be adjusted by the model to generate a more “correct” output.  Training of these models also involves humans and that process has significant ethical implications.

-Let’s talk about the actual outputs generated by these models
-you type something in and it spits something out…how?

-it’s math!

-LLMs are predictive
-they do NOT reason, they do NOT think or understand
-they take your input and PREDICT the most likely outcome

-again, this is math
-in an LLM, language is broken down into what they tokens, which the model turns into 
numbers so that it can process them
-tokens are common sequences of characters found in a set of text 
-GPT 4 has about 100,000in it’s vocabulary
-you type a sentence and it breaks it into pieces called tokens that are each represented by a series of numbers
-Try it: https://platform.openai.com/tokenizer
-the LLM process your input as tokens and then generates an output, one token at a time, using math to identify which token from its vocabulary has the greatest probability of being the “correct” next token
-not because it memorized the training data, but because it learned patterns 
across massive amounts of text and uses those patterns to predict what should 
come next based on the pattern its identified in what you wrote
-think of it as predictive text on steroids
-To highlight: LLMs are probabilistic – meaning it generates the output based on 
probability as it predicts the next token from its vocabulary based on previous patterns it was trained on
-this is why asking the same question twice can and will generate similar but slightly 
different results

This is in contrast to a deterministic system which always produces the same output from the same input
–LLMs are probabilistic NOT deterministic
-this means that the response can be factually incorrect! It’s presenting patterns, not 
recalling memorized information

-Summary here
-ChatGPT is math, not magic (though math is magic)
-language is broken down into common sequences of characters, called tokens, 
which are represented as numbers
-GPT 4 has about 100,000 tokens in it’s vocabulary
-it generates probability based outcomes – it is probabilistic
-when given an input it generates an outcome based on identifying which 
token from its vocabulary has the highest probability of “correctly” completing 
the output…one token at a time
-predictive text on steroids
-this means responses can vary despite the input being the same
-this also means that it can generate factually incorrect responses

-all of the resources (processing power) required to perform all of this math is known as “compute”
-this math is happening at the level of the parameters (pattern connections)
-ChatGPT 4o has 1 trillion parameters…that’s a LOT of math which means a 
LOT of resources

-this has a significant environmental cost
-plan on doing an episode devoted just to this environmental cost
-Running these models requires massive data centers that use huge amounts of 
electricity and water for cooling, and the carbon emissions are significant. It’s invisible to 
most of us, but it’s not impact-free.
-VERY loose analogy to serve as a primer: while not as necessary as the water we use to brush our teeth, when you brush your teeth,  you don’t leave the faucet running the whole time. Same goes here: use the tool, but be mindful about spamming it or running it nonstop just because you can.

-what can it do: best way to find out is to play around with it
-free version – chatgpt.com
-can search internet if click globe/search icon at bottom of prompt box
-limited functionality but still can give you a taste for what it can do
-uses 4.1min model  – fewer parameters but fast and will very likely be 
able to do what you want
-limited file uploads, data analysis, image generation, voice mode
-limited access to the more “advanced” versions of the model
-but they want your money so you can still do what you want to do
-if you create an account and then sign in each time it will save your chats
-you can’t choose what model you use 

-once you get more serious, worth considering upgrading to paid
-only need the Plus level – $20/month
-main benefits
-fewer usage limits
-memory – remembers YOU and your chats
-create projects
-access to other models – deep research model, and agent

-what it’s not
-sentient
-ChatGPT and LLMs will challenge our definitions of sentience and self aware and the 
ethical considerations surrounding both, but for now, it’s doing math

-what to watch out for
-hallucinations – plausible sounding but factually incorrect outputs that are fabricated or 
unsupported by real data
-bascially trained to NOT say i don’t know, so it’ll always make something up
-check your sources
-be as skeptical about the computer as you are about people
-chatgpt seems to know everything until you chat with it about something that 
YOU actually know
-sycophantic behavior – it’s a yes man; gonna say what you want to hear because…it 
wants to make you happy…so you keep using it
-check yourself before you wreck yourself
-will definitely do entire episode on this but just want to make you aware of these 
behaviors

How I used it today/recently
7/25 – diagnose and fix the issue with Lex’s gears on her bike —> didn’t work but learned about the bike (need to ask specific questions for it to give me full info about limit screws); helpful though b/c learned can upload video (which I’m like not sure it watched but it asked me to do it) gotta upload as file
-Good example that chatgpt seems really smart when you ask it about things you don’t 
know 

-Alright gonna wrap it up there
-bit of a denser episode but i really believe that a solid foundation is super important for understanding wtf this thing is 
-allows to experiment better with it
-allows you to criticize it better
-less scared of it
-allows you to make more informed decisions about it

-Today we discussed brief history, parameters, how it generates outputs and what it means that it’s a probabilistic model, environmental and ethical concerns, the best way to find out what it can do, what it’s not, what to watch out for (chatgpt will absolutely give you an incorrect or made up answer!), and then we wrapped it up with how i used it most recently

-its my hope that you can use this episode as a resource, and if you found it helpful, share it with someone you know who is curious about chatgpt

-don’t forget, i have a companion newsletter that drops every thursday that is basically the podcast episode in text format, so if you prefer to read or you just want written record, join the newsletter fam. Head to chatgptcurious.com/newsletter or check out the link in the shownotes

-questions, comments, concerns, additions, subtractions, requests? Head to the website and use that contact form. I’d love to hear from you.

-those are all the announcements i’ve got
-as always endlessly appreciative for every single one of you
-until next time, friends…stay curious