The AI generator museum of uncanny amusement
-
A good way to think about 'bad' datasets is maybe a weather satellite that's a bit borked. Maybe something has failed slightly in the software or the detection instruments. Could be a bit of dust or a crack in the lens or whatever. If it's trying to predict the future weather based on the dataset it can still do it because it'll compensate for the defect once it compares the prediction to the reality enough times. There's still a pattern there that is useful. The crack could be warping the light but that's fine once it retrains itself, which is automatic. It'll gradually forget the data when the lens was 'good'.
Training AI on AI manufactured datasets is just inefficient in many cases because the AI has to cut through the other AI middleman. It can be useful though when it comes to things like virtual roads for self-driving. It depends on the case really."Plus he wore shorts like a total cunt" - Bob -
Out of interest are you on the free version Gav? It’s much “dumber” than the paid for version.
And that is definitely one of the things people keep saying. It will never replace X because look at my anecdote. But yearly compute is growing at 10x and ChatGPT 5 is going to be even more intelligent than 4. Then those things that it “failed” at will be in the past. -
I'm quite keen to delve into how I might build my own LLM. How tricky is that going to be when I have a beginner-to-intermediate grasp of Python and understand the core concepts? -
Might start with that book
-
Elmlea wrote:Might start with that book
You should, it's perfect for beginners, but don't be thinking about LLMs just yet. You'll do the hello world of deep learning which is the mnist dataset. Start with that and work your way up!"Plus he wore shorts like a total cunt" - Bob -
CEO of Hasbro
Maybe AI will kill the nostalgic content recycling of the past couple of decades. Only non-AI IP will be copyright protected and only new stuff will be any good. -
Yeah, all depends on the size of the datasets you have. There's going to be some really cool stuff in the not too distant future."Plus he wore shorts like a total cunt" - Bob
-
AI could be useful for creating ideas for long running things based on established lore. Of course you can do that now by having a human that knows about the thing.
Do we reckon Paramount would pick Chat GPT or Jonathan Frakes? -
Diluted Dante wrote:AI could be useful for creating ideas for long running things based on established lore. Of course you can do that now by having a human that knows about the thing. Do we reckon Paramount would pick Chat GPT or Jonathan Frakes?
It might find cool patterns to base a new game on, or improve an existing one. Essentially, what makes a particular game good? You could leave the lore and character design to humans and let the AI concentrate on finding cool gameplay loops or satisfying complexity."Plus he wore shorts like a total cunt" - Bob -
What’ll happen first is AI writing all the porn. Cheapest option for production of content content content, innit. -
SpaceGazelle wrote:
You should, it's perfect for beginners, but don't be thinking about LLMs just yet. You'll do the hello world of deep learning which is the mnist dataset. Start with that and work your way up!Elmlea wrote:Might start with that book
Excellent, I'll read that. I bet I brush up on Python and get stuck in the cycle of being unsure what dev environment to use again. I lived out of cracked PyCharm before, and tried managing with just SublimeText for a while, but I'd like something full featured. Any recommendations for an env for Jupyter/Anaconda? -
True fact. I spent a big chunk of my November and December building a scraper to harvest Larry David's posts from this here forum. The intention was to train and build 'Larry dAIvid', my gift to the forum, who could be summoned to post in the sports thread with a tag.
Most of my time was spent on the logic of trying to programmatically deduce prompts from responses in the absence of consistent quotes. I think i just about got there in the end, but it amounted to nothing i) because there was nowhere near enough content (this would have been obvious, if i knew more about the topic to begin with), and ii) Larry David came back. Oh, and iii) my workload for stuff i actually get paid for increased a lot. Just before i gave up trying to do it properly, i tried the shortcut version using my GPT API key, and honestly it was utterly rubbish. Perhaps predictably so.
Naturally my intention was to keep this secret until the moment of unveiling, but given my workload, the seeming impossibility of the task, and accumulation of other priorities, i think that's unlikely to ever happen. -
Just point it at Harry Redknap quotes for the extra content. -
I have an idea for a game which would make good use of AI. Think it could be workable and pretty cool. Keep meaning to post about it in case anyone would want to collaborate.
-
@Elm
Just download Anaconda and use the default Jupyter Notebook at first. There are not many lines of code when you're building AI, at least at first. The frameworks like Keras do all the heavy lifting for you so you don't need all the stuff that comes with modern IDEs. You'll see as you go through the book. The author actually just uses the python shell direct and does it line by line.
Use Jupyter though. You'll want to play around with the data and this is best done by partitioning code. You can run segments of code individually or all at once. The fully featured bit is Keras itself, not the IDE. You can get a model built with just a few lines of code so no need for anything fancy, it'll only hinder.
Python Crash Course book is ideal and fast to brush up. Revise the basic data types again and have a look at classes and OOP. Classes are useful to understand libraries you'll be importing and what is and isn't part of the core language."Plus he wore shorts like a total cunt" - Bob -
b0r1s wrote:Out of interest are you on the free version Gav? It’s much “dumber” than the paid for version.
Avatar shenanigans -
equinox_code wrote:I have an idea for a game which would make good use of AI. Think it could be workable and pretty cool. Keep meaning to post about it in case anyone would want to collaborate.
PM me the idea. I'm not saying I'll help work on it but I can maybe advise on how to go about it."Plus he wore shorts like a total cunt" - Bob -
Diluted Dante wrote:Do we reckon Paramount would pick Chat GPT or Jonathan Frakes?
ChatGPT won't have his flair.
-
SpaceGazelle wrote:
PM me the idea. I'm not saying I'll help work on it but I can maybe advise on how to go about it.equinox_code wrote:I have an idea for a game which would make good use of AI. Think it could be workable and pretty cool. Keep meaning to post about it in case anyone would want to collaborate.
Will do... eventually. -
SpaceGazelle wrote:@Elm Just download Anaconda and use the default Jupyter Notebook at first. There are not many lines of code when you're building AI, at least at first. The frameworks like Keras do all the heavy lifting for you so you don't need all the stuff that comes with modern IDEs. You'll see as you go through the book. The author actually just uses the python shell direct and does it line by line. Use Jupyter though. You'll want to play around with the data and this is best done by partitioning code. You can run segments of code individually or all at once. The fully featured bit is Keras itself, not the IDE. You can get a model built with just a few lines of code so no need for anything fancy, it'll only hinder. Python Crash Course book is ideal and fast to brush up. Revise the basic data types again and have a look at classes and OOP. Classes are useful to understand libraries you'll be importing and what is and isn't part of the core language.
Brilliant, I did the Python Crash Course book (and Automate the Boring Stuff) back when I was learning, so they'll do as a refresher. Might just pull out SublimeText for those and then start a fresh env with Anaconda. -
You could always just download Jupyter Notebook if you're happy pip installing libraries like Keras and Numpy. Anaconda will give you more than you need but it's a handier but bulkier way to get everything in one place. Give me a PM if you get into difficulty."Plus he wore shorts like a total cunt" - Bob
-
Will do, I'm happy with pip but I like the full featured things, so will give Anaconda a wobble.
-
Elmlea wrote:Will do, I'm happy with pip but I like the full featured things, so will give Anaconda a wobble.
Cool. Might be easier to PM me your email address if there are lots of questions and don't be shy, I'm happy to answer them."Plus he wore shorts like a total cunt" - Bob -
Have started on the book I got from Amazon. It has more than just the AI stuff - the first half has many useful algorithms for other purposes. So I've dipped into the latter half. It's started with unsupervised learning techniques - covered some techniques for identifying clusters of datapoints.
Not sure why I'm surprised but I am at the fact that there's plenty of useful libraries for importing for doing the donkey work - calculation of euclidan differences, cosine differences, k-means clustering etc. Very nice.Holding the wrong end of the stick since 2009. -
The libraries are why python has become the language for ML. Because the python interpreter is written in c, the libraries can just add a load of pure code c without any conversion. They're fast af because you're basically running c in a very light Python wrapper.
Core Python is translated to a thing called Bytecode first so it can be quite slow, especially for loops. Also ints are immutable so it needs to create and assign a new int everytime you add one to some variable."Plus he wore shorts like a total cunt" - Bob -
Here you go Davy, it'll help get you up to speed after coming from other languages.
Youtube embedding not playing nice so search "Batchelder - Facts and Myths about Python names and values".
Python doesn't quite work like many languages - it doesn't really pass by value or reference. These things are handy to know. I fucking love Python."Plus he wore shorts like a total cunt" - Bob -
Never been a purist re languages. As long as it gets the job done. Performance depends on the application as to whether it matters.
Have worked with interpreted, partial like byte code and compiled. Played with assembler too. You do have to be careful with languages that have less guardrails but as long as you know the hazards its OK with me.Holding the wrong end of the stick since 2009. -
Python is pretty well guarded but it can easily trip you up when it comes to variable assignment. Check out that vid and the your experience will do the rest. You'll need to know for it ML. Also for ML check out python generators. You'll be dealing with big data (and I mean BIG) and these things will save you a ton of time for data manipulation."Plus he wore shorts like a total cunt" - Bob
-
What book are you using btw?"Plus he wore shorts like a total cunt" - Bob
-
Well we're using typing with python these days. With a good ide plugin for mypy jazz you'll be laughing.
Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Categories
- All Discussions2,718
- Games1,881
- Off topic837