image description (contains clarifications on background elements)
Lots of different seemingly random images in the background, including some fries, mr. crabs, a girl in overalls hugging a stuffed tiger, a mark zuckerberg “big brother is watching” poser, two images of fluttershy (a pony from my little pony) one of them reading “u only kno my swag, not my lore”, a picture of parkzer parkzer from the streamer “dougdoug” and a slider gameplay element from the rhythm game “osu”. The background is made light so that the text can be easily read. The text reads:
i wanna know if we are on the same page about ai.
if u diagree with any of this or want to add something,
please leave a comment!
smol info:
- LM = Language Model (ChatGPT, Llama, Gemini, Mistral, ...)
- VLM = Vision Language Model (Qwen VL, GPT4o mini, Claude 3.5, ...)
- larger model = more expensivev to train and run
smol info end
- training processes on current AI systems is often
clearly unethical and very bad for the environment :(
- companies are really bad at selling AI to us and
giving them a good purpose for average-joe-usage
- medical ai (e.g. protein folding) is almost only positive
- ai for disabled people is also almost only postive
- the idea of some AI machine taking our jobs is scary
- "AI agents" are scary. large companies are training
them specifically to replace human workers
- LMs > image generation and music generation
- using small LMs for repetitive, boring tasks like
classification feels okay
- using the largest, most environmentally taxing models
for everything is bad. Using a mixture of smaller models
can often be enough
- people with bad intentions using AI systems results
in bad outcome
- ai companies train their models however they see fit.
if an LM "disagrees" with you, that's the trainings fault
- running LMs locally feels more okay, since they need
less energy and you can control their behaviour
I personally think more positively about LMs, but almost
only negatively about image and audio models.
Are we on the same page? Or am I an evil AI tech sis?
IMAGE DESCRIPTION END
i hope this doesn’t cause too much hate. i just wanna know what u people and creatures think <3
I honestly am skeptical about the medical stuff. Machine learning can’t even do the stuff it should be good at reliably, specifically identifying mushrooms/mycology in general.
that is interesting. i know that there are plenty of plant recognition onces, and recently there have been some classifiers specifically trained on human skin to see if it’s a tumor or not. that one is better than a good human doctor in his field, so i wonder what happened to that mushroom classifier. Maybe it is too small to generalize or has been train in a specific environment.
Do not trust AI to tell you if you can eat a mushroom. Ever. The same kinds of complexity goes into medicine. Sure, the machine learning process can flag something as cancerous (for example), but will always and forever need human review unless we somehow completely change the way machine learning works and speed it up by an order of magnitude.
yeah, we still very much need to have real humans go “yes, this is indeed cancer”, but this ai cancer detection feels like a reasonable “first pass” to quickly get a somewhat good estimation, rather than no estimation with lacking doctors.
I haven’t looked closely enough to know, but I recall medical image analytics being “better than human” well before the current AI/LLM rage. Like, those systems use machine learning, but in a more deterministic, more conventional algorithm sense. I think they are also less worried about false positives, because the algorithm is always assumed to be checked by a human physician, so my impression is that the real sense in which medical image analysis is ‘better’ is that it identifies smaller or more obscure defects that a human quickly scanning the image might overlook.
If you’re using a public mushroom identification AI as the only source for life-and-death choice, then false positives are a much bigger problem.
yes, that is what i have heard too. there was a news thing some days ago that this “cancer scanner” thing will be available in two years to all doctors. so that’s great! but yes, we very much still need a human to watch over it, so its out-of-distribution-generations stay in check.
Having worked with ML in manufacturing, if your task is precise enough and your input normalized enough, it can detect very impressive things. Identifying mushrooms as a whole is already too grand a task, especially as it as to deal with different camera angles, lighting … But ask it to differentiate between a few species, and always offer pictures using similar angles, lighting and background, and the results will most likely be stellar.
Like I said, I’m just skeptical. I know it can do impressive things, but unless we get a giant leap forward, it will always need extensive human review when it comes to medicine (like my mycology example). In my opinion, it is a tool for quick and dirty analysis in the medical field which may speed things up for human review.
From what little I know if it, it’s sorta twofold what it does:
-
It looks through documentation across a patient record to look for patterns a doctor might miss. For example, a patient comes in complaining of persistent headaches/fatigue. A doctor might look at that in isolation and just try to treat the symptoms, but an AI might see some potentially relevant lab results in their histories and recommend more testing to rule out a cancer diagnosis that the doctor might have thought unlikely without awareness of that earlier data.
-
Doctors have to do a lot of busywork in their record keeping that AIs can help streamline. A lot of routine documentation, attestations, statements, etc. Since so much of it is very template-heavy already, an AI might be able to streamline the process as well as tailor it better to the patient. E.g. the record indicates “assigned male at birth” and an ER doctor defaults to he/him pronouns looking only at the medical birth sex marker, but the patient is also being seen by a gender clinic at which she is receiving gender affirming treatment as a trans woman and brings up that earlier data to correct the documentation and make it more accurate and personalized for the patient.
In reality, I am sure that practices and hospital systems are just going to use this as an excuse to say “You don’t need to spend as much time on documentation and chart review now so you can see more patients, right?” It’s the cotton gin issue.
LMs give the appearance of understanding, but as soon as you try to use them for anything that you actually are knowledgable in, the facade crumbles.
Even for repetitive tasks, you have to do a lot of manual checking to ensure they did not start hallucinating half way through.
I haven’t really used AIs myself, however one of my brothers loves AI for boilerplate code which he of course looks over afterwards. If it saves time and you only have to do some minor editing then that seems like a win to me. Probably shouldn’t be used like this in any non-hobby project by people who aren’t adept at coding however
I’m a programmer as well. When ChatGPT & Co initially came out, I was pretty excited tbh and attempted to integrate it into my workflow, which kinda worked-ish? But was also a lot of me being amazed by the novelty, and forgiving of the shortcomings.
Did not take me long to phase them out again though. (And no, it’s not the models I used; I have tried again now and then with the new, supposedly perfect-for-programming models, same results). The only edgecase where they are generally useful (to me at least) are simple tasks that I have some general knowledge of (to double theck the LM’s work) but not have any interest in learning anything further than I already know. Which does occur here and there, but rarely.
For everything else programming-related, it’s flat out shit.I do not beleive they are a time saver for even moderately difficult programs. Bu the time you’ve run around in enough circles, explaining “now, this does not do what you say it does”, “that’s the same wring answer you gave me two responses ago”, “you have hallucinated that function”, and found out the framework in use dropped that general structure in version 5, you may as well do it yourself, and actually learn how to do it at the same time.
For work, I eventually found that it took me longer to describe the business logic (and do the above dance) than to just… do the work. I also have more confidence in the code, and understand it completely.
In terms of programming aids, a linter, formatter and LSP are, IMHO, a million times more useful than any LM.
for even moderately difficult programs.
My brother uses it to generate templates and basic structs and functions, not to generate novel code. That’s probably the difference here. I believe it’s integrated into his text editor as well? It’s the one github offers
Edit: Probably wouldn’t be useful if it wasn’t integrated into the editor and therefore the generation being just a click away or some sort of autofill. Actually writing a prompt does sound tedious
this matches my experience too. good IDEs or editors with LSP support allll the way.
also wanna add that it’s weird to me that we turn to LLMs to generate mountains of boilerplate instead of… y’know, fixing our damn tools in the first place (or using them correctly, or to their fullest) so that said boilerplate is unnecessary. abstractions have always been a thing. it seems so inefficient.
you’re right, it doesn’t do classification perfectly every time. but it drills down on the amount of human labour required to classify a large set of data.
about the knowledge: it really comes down to which model you are talking to. “generalist” models like GPT4o or claude 3.5 sonnet have been trained to know many things somewhat, but no single thing perfectly.
currently companies seem to train largely on IT-related things. these models are great at helping me program, but they are terrible at specifically writing GDScript (a niche game-programming language) since they forget all the methods and components the language has.
Even with LMs supposedly specialising in the areas that I am knowledgable (but by no means an expert) in, it’s the same. Drill down even slightly beyond surface-level, and it’s either plain wrong, or halucinated when not immediately disprovable.
And why wouldn’t it be? These things do not possess knowledge, they possess the ability to generate texts about things we’d like them to be knowledgable in, and that is a crucial difference.
I’ve heard this argument so many fucking times and i hate genai but there’s no practical difference between understanding and having the appearance of such, that is just a human construct that we use to try to feel artificially superior ffs
No. I am not saying that to put man and machine in two boxes. I am saying that because it is a huge difference, and yes, a practical one.
An LLM can talk about a topic for however long you wish, but it does not know what it is talking about, it has no understanding or concept of the topic. And that shines through the instance you hit a spot where training data was lacking and it starts hallucinating. LLMs have “read” an unimaginable amount of texts on computer science, and yet as soon as I ask something that is niche, it spouts bullshit. Not it’s fault, it’s not lying; it’s just doing what it always does, putting statistically likely token after statistically liken token, only in this case, the training data was insufficient.
But it does not understand or know that either; it just keeps talking. I go “that is absolutely not right, remember that <…> is <…,>” and whether or not what I said was true, it will go "Yes, you are right! I see now, <continues to hallucinate> ".
There’s no ghost in the machine. Just fancy text prediction.
What I think is missing from your viewpoint (and from most people’s, this is [IMO] a problem at scale) is the distinction between “simple” and broad machine learning, and the very specific things that are Large Language Models.
For example, there are no small Large Language Models, and I think that the oxymoron speaks for itself. Machine learning is a very good thing, and automated classification is definitely its best use case, but they are not a small version of ChatGPT, the same way that the average Joe is not a smaller version of a billionaire.
For more details, these small models are trained on a small set of data, how small depending on how specific the task is; as an example, I worked with models that detect manufacturing defects on production lines, and theses need a few hundreds images in order to produce good results, this make it very easy to produce the data ourselves, and it is relatively cheap to train energy-wise.
Compared to that, Large Language Models, and their audiovisual counterparts, operate on billions of data, and work on a task so general that they provide incredibly bad results. As a little statistical reminder, anything below 95% confidence is a bust, LLMs are way below that.
It’s very important to distinguish the two, because all of the positives you list for AI are not about LLMs, but about simple machine learning. And this confusion is by design, techbros are trying to profit of the successes of other form of artificial intelligence by pretending that AI is this one single thing, instead of an entire class of things.
Otherwise, I generally agree with the rest of your points.
i completely agree. training an actually small model on your specific task almost always results in WAY better output.
current LLMs might be great at PhD questions, but are still bad at way simpler things, which shows that they have been trained on these questions, rather than generalizing to that level.
training a “cancer recognizer” will be way more efficient and accurate than a general, much larger VLM trying to do the same thing.
wait no, there are small language models! like the one in the phone keyboard, suggesting the next word. sometimes there are rule-based but in many cases, they are real neuronal networks, predicting what you will type. in my case it even trains on what i type (an open source keyboard i got, running locally obv)
I’m pretty sure that phone keyboard use heuristics and not Machine Learning. Basically, it does not create a neural network through trial and error, but whenever you type, it saves the context of each word, and when it sees the same context again, it “knows” what the next word is.
For example, if you type this big brown fox, it might saves something like "{ fox", ["big", "brown"], 1 } (assuming two words of context, and the 1 being the number of times it was encountered).
Then when you type my big brown, fox will be suggested.
Using the technology of LLMs for keyboard suggestions is impractical, as your typing habits would be drowned in the initial training data, and would yield worse performance as well as results compared to the simpler approach.
Sounds like a pretty nuanced opinion to me. AI’s biggest problem is that it’s being driven by profit seeking companies that, as usual, have no interest in the damaging “externalities” that are involved. But anyone that says it has no application and all it does is hallucinate and waste energy is not aware of the progress being made.
My biggest problem with AI is how it was pushed and marketed to us in ways that don’t make sense / are unethical. Even the environmental concerns would be ameliorated if AI weren’t being pushed into everything. (Using “AI” here to refer to things like LM, image, and art generators,etc.)
yes, i completely agree.
having some LM generate “comment suggestions” for content creators on youtube is such a genuine waste of compute and the environment. (yes this is a real thing)
it was marketed as this “smart machine” which ends up being too dum for most people wanting to use it.
