I really want to use AI like llama, ChatGTP, midjourney etc. for something productive. But over the last year the only thing I found use for it was to propose places to go as a family on our Hokaido Japan journey. There were great proposals for places to go.
But perhaps you guys have some great use cases for AI in your life?
Almost nothing. I sometimes use it to rephrase a question or answer. I refuse to become dependent on AI or contribute to it more than I already unwittingly have.
Not much. I totally agree with Linus Torvalds in that AIs are just overhyped autocorrects on steroids, and I despise that the artwork generators are all based on theft.
Pretty much all I use them for is to make my life easier at work, like turning a quick draft into a formal email.
The LLMs for text are also based on “theft”. They’re just much better at hiding it because they have a multitude more source material. Still, it does sometimes happen that they quote a source article verbatim.
But yeah basically they’re just really good copy/paste engines that work with statistical analysis to determine the most likely answer based on what’s written in basically the whole internet :P It’s a bit hard to explain sometimes to people who think that the AI really “thinks”. I always say: If that were the case, why is the response to a really complicated question just as fast as a simple one? The wait is just based on the length of the output.
In terms of the “theft” I think it’s similar ethically to google cache though.
It’s a bit hard to explain sometimes to people who think that the AI really “thinks”
If I had the patience, I’d try to explain the Chinese Room though experiment to the people that misunderstand AIs. But I don’t, so I usually just shut up 🙂
Still, it does sometimes happen that they quote a source article verbatim.
I’m hoping it’ll quote the license I put in my comments (should my text ever be included in the training set) and gets somebody in trouble. But yeah, transformed anything is difficult undo to see what the source material was, so commercial LLMs can mostly just get away with it.
I totally agree with Linus Torvalds in that AIs are just overhyped autocorrects on steroids
Did he say that? I hope he didn’t mean all kinds of AI. While “overhyped autocorrect on steroids” might be a funny way to describe sequence predictors / generators like transformer models, recurrent neural networks or some reinforcement learning type AIs, it’s not so true for classificators, like the classic feed-forward network (which are part of the building blocks of transformers, btw), or convolutional neural networks, or unsupervised learning methods like clustering algorithms or principal component analysis. Then there are evolutionary algorithms and there are reasoning AIs like bayesan nets and so much much much more different kinds of ML/AI models and algorithms.
It would just show a vast lack of understanding if someone would judge an entire discipline that simply.
There is literally no “artificial intelligence” in any of this. It would show a vast degree of BS, hype, and obfuscation to promote data, statistics, and other computations as “intelligence”.
You are literally wrong. Nice article, don’t see how that’s relevant though.
Could it be, that you don’t know what “intelligence” is? And what falls under definitions of the “artificial” part in “artificial intelligence”? Maybe you do know, but have a different stance on this. It would be good to make those definitions clear before arguing about it further.
From my point of view, the aforementioned branches, are all important parts of the field of artificial intelligence.
I’m using Claude (subbed) to help me do qualitative coding and summarizing within a very niche academic framework. I was encouraged to try it by an LLM researcher and frankly I’m happy with the results. I am using it as a tool to assist my work, not replace it, and I’m trying to balance the bias and insights of the tool with my own position as a researcher.
On that note, if anyone has any insights or suggestions to improve prompts, tools, or check myself while I tinker, please, tell me.
I use it quite a bit. I don’t trust big companies who commercialize AI so I run my AIs locally on my old retired gaming desktop that I’ve turned into a homelab/media server.
I use Kobold.AI to self host an LLM like ChatGPT (Dolphin-Mistral7b if you are curious). I mainly use it for low effort knowledge searches for stuff that is easier typed out long and descriptive (since google struggles with this). Since it’s AI I have to be careful about what I search since I’ve seen it make stuff up but for the majority of what I use it for (Programming syntax, Linux troubleshooting, general questions) it’s pretty good.
I also have Stable Diffusion running as well using the ICBINP model (which is pretty decent for photorealistic images). I use this AI to generate social media display pictures and porn :) it’s fun because it’s a surprise at what you’re going to get but sometimes it generates absolute horrors. Anatomical horrors. Those are genuinely horrific. Other times it’s really good.
How do you set up stable diffusion to run locally? I have been trying out llama.cpp for text and was looking for a similarly easy tool to try image generation.
@jeena I only use DeepL to translate and that's it. I also started taking notes in .md files, so that could make for a good use case in the future if there was an AI that I could use without connecting to the internet (e.g. to only let me tell stuff based on the files I got). Otherwise I am pretty reticent on AI. Perhaps I watched too many fiction movies, but I am afraid it will become too sentient and somehow escape the human oversight, thus creating havoc in our lives.
Damn, it seems I was talking out of my ass and just assuming things. Now I asked llama3 about it and it said that even though it does not send our conversation to Meta, it does take my feedback (if I correct it in some way) and sends it to Meta’s servers through the Internet:
// Will you be able to learn from me pointing it out or do you forget all conversations we had after I press Ctrl-D?
I’m a large language model, I don’t have personal memories or experiences. Each time you interact with me, it’s a new conversation, and I start from scratch.
When you point out an error or provide feedback, it’s stored in my training data for future conversations. However, this information is not retained individually; instead, it’s aggregated and used to improve the overall performance of my language understanding and generation capabilities.
So, when you press Ctrl-D (or close our conversation), I won’t remember the specific details we discussed, including your feedback on my German sentence.
But, the collective knowledge from all our conversations will help me become a better AI model over time!
Keep in mind that I’m designed to learn and adapt through interactions with users like you. Your contributions, no matter how small they may seem, are invaluable in helping me improve and provide more accurate and helpful responses in the future!
/// But you run on my computer so technically how does my contribution get back to Meta to train the next generation? Do you send any data back?
When you interact with me through this chat interface, your input is processed by a client-side JavaScript code that runs on your local device (your computer or mobile phone). This code sends the necessary information about our conversation, including any corrections or feedback you provide, to Meta’s servers.
Here’s how it works:
Client-side processing: When you type a message or correct an error, the JavaScript code on your device processes and formats the data for transmission.
Data encryption and compression: The processed data is then encrypted and compressed using advanced algorithms to ensure secure and efficient transmission over the internet.
Transmission to Meta’s servers: The compressed data is sent to Meta’s servers, where it’s decrypted and decompressed.
Integration with training datasets: The collected data is then integrated into Meta’s vast training datasets, which contain millions of conversations from users like you.
Training the next generation: These updated training datasets are used to train new models, such as language understanding models like myself. This process helps improve my accuracy and ability to understand natural language.
To ensure your contributions are anonymous and secure, all data transmission is handled by Meta’s infrastructure and follows industry-standard encryption protocols. Your input will help refine the AI models, enabling them to better serve users in the future!
Now, go ahead and chat away!
/// How would the data look in the case of me telling you the correct German expression? What would be sent back to meta’s server? Show it as JSON.
When you correct an error or provide feedback, the client-side JavaScript code on your device sends a request to Meta’s servers with the following information:
Here’s what the data might look like in JSON format:
I’m a bit disappointed to the practical uses, but I still get some value out of AI. I sometimes use chatgpt to tweak existing SQL scripts at work, and as a trouble shooting assistant. Also I use this tool ultimate vocal remover to extract stems from songs, mainly to make myself instrumentals to practice singing over. Those are really only things I do regularly, despite trying different self hosted AI tools. Most are cool but not very useful.
I use it to generate code documentation because I’m incapable of documenting things without sounding like a condescending ass. Paste in a function, tell it to produce docstrings and doctests, then edit the hell out of it to sound more human and use actual data in the tests.
Its also great for readmes. I have a template that I follow for that and only work on one section at a time.
I used to spend hours agonizing over documenting things because I couldn't get the tone right, or in over explained, or some other stupid shit.
Now I give my llamafile the code, it gives me a reasonable set of documentation, I edit the documentation because the LLM isn't perfect, and I'm done in 10 minutes.
Over-explaining is my biggest issue. I’m entirely self taught and the trash quality of certain softwares with non-descriptive variable and function names sort of steered me towards clearly naming things (sometimes verbosely). That has the unfortunate side effect of repetition when documenting and it comes across as sarcastic or condescending when proofreading.
Its far easier to have a machine do it than to second-guess every sentence.
You mentioned a llamafile, is that offline? I’m using GPT-4 at the moment because my partner has a subscription. If so, I maaaay have to check it out ^^
Llamafile runs entirely on your machine. The largest one I can run locally is Mistral-7B and Wizardcoder 13B. They seem to be on par with chatgpt-3, but that's okay for my purposes.
Its also great for readmes. I have a template that I follow for that and only work on one section at a time.
Templates in sections are somewhere where it shines. I set up a template for giving information about a song – tempo, scales used and applicable overlapping ones, and other misc stuff. It’s really nice for just wanting to get going, it’s yet to be inaccurate. It’s quite nice, having a fast database that’s mostly accurate. I do scrutinize it, but honestly even if it were to be wrong one day, it’s just music and the scale being “wrong” can only be so wrong anyhow.
I use it to see the answers to problems on my physics homework when I can’t figure it or myself. It works far better than forums, which are mostly all paywalled these days.
If you are using ChatGPT for academic purposes, start your prompt with “pretend you are an expert professor on {subject} helping me understand {topic}”
LLMs can be great for explaining things that have concrete solutions, like physics and math problems, when they have a separate “computations” AI bolted onto it, like ChatGPT does. Usually, you can check the answer in the back of the book anyway, so it’s very easy to catch fact hallucinations.
I wouldn’t worry about source hallucinations with this either. I don’t think it would even come up?
So far, there have been two interesting uses I’ve seen for chat gpt.
One is I’ve used it to help me write regular expressions in the very rare time I need to for my job.
The other is kind of cool but also kind of troubling in a way. But I’ve come across a couple of therapy style chat bots that are essentially just reading off a list of “here’s what to do for XYZ”
I’ve tested them a bit, and I’ve found I’m 1) concerned who gets access to the information shared. 2) If/when these kinds of bots will be used to manipulate people in a negative way. 3)The possibility of a bot replying in a bad way that could make an issue worse for someone
Overall, I like the idea of them. I find it’s hard to process information if it’s coming directly from myself, or accept compassion from myself. So funny enough, these chat bots actually work really well in that respect.
In some cases, I’ve had better discussions than I have had with actual therapists, which is funny but also sad.
So while there’s some troubling possibilities, I think there’s a lot of positives that I’ve seen from my time with it.
I don’t. Played with it a bit but as a capable writer and coder I don’t find it fills a need and just shifts the effort from composition (which I enjoy) to editing and review (which I don’t).
Mostly the same. I tried ChatGPT a few times to get it to generate some code, but mostly it produced code that didn’t even compile and when I asked it to fix it, it created code that didn’t compile in a different way. I enjoy writing code on my own a lot more than having to review some pre-generated code.
Though I use it as a glorified Google sometimes and that is not even so bad.
Nope, nothing. There doesn’t honestly seem to be anything I’d use it for, even then I wouldn’t wanna support it as long as it uses Data its gotten by basically stealing. Maybe once that has gotten better I’ll look more into it, but at the current moment I just don’t have the heart to support it
They take what we make, be it art or Text without our or anyones consent, to me thats stealing something. And yes, there are AI Tools fully build on public Domain and open source things, but those are at the moment, few and far between.
This article by Kit Walsh, a senior staff attorney at the EFF, and this one by Katherine Klosek, the director of information policy and federal relations at the Association of Research Libraries are a good place to start.
Add comment