Genuinely curious.
Why do you like LLMs? What hopes do you have for AI & AGI in our near and distant future?
You must log in or register to comment.
Because they are extremely useful.
They speed up bottlenecks in many areas. It’s like having an assistant with a broad, although superficial, knowledge. Which is fine. I can manage myself and oversee senior employees, judging chatgpt output is pretty trivial.
Future are agents. That’s what is really missing at this point. Something that can take over more complex tasks with less supervision.
The challenge for AGI is social, that’s the biggest challenge. AI will take over a lot of work, but we need better wealth distribution
I could not agree more. I really enjoy Andrej Karpathy’s model where in the future AGI does 99% of the technical work and the human in the loop does the creative and critical 1%.
Right now an LLM is basically a two year old that knows every language in the world and has the entire knowledge of humanity squeezed in its little head.
They’re also fun to work with. Error messages are boring when you can instead try to figure out where an LLM got the idea to say what it did.
What I find particularly exciting is that we’re seeing this evolution in real-time.
Can you imagine what these models might look like in 2 years? 5? 10?
There is a remarkable future on the horizon. I hope everyone gets an equal chance to be a part of it.
They will get better and might actually become a threat for software engineers, but I don’t think LLMs in their current form will get us much closer to AGI.
We need to do reinforcement learning in the real world to get there. And that will be hard, because right now we have the internet as an essence of human knowledge, mostly in text form so it’s super easy go work with. It’s basically easy mode in the context of AGIs (not to discredit the people working of SOTA LMMs, I just think that the way ahead of us will be even harder).
- I fucking hate googling. The internet has completely degenerated into a useless mess of SEO leeches. LLMs save so much time. Even if they are often wrong they are still 10x better than Google.
- Imagine how insane they will make future video games.
- I have a lot of unordered notes and the only feasable way to organize them at this point are LLMs.
- I’ve always been bad at grammar, spelling, punctuation, phrasing etc because of dyslexia/AuDHD, so they make professional writing much easier.
- Imagine how much time it would save to not have to read through manuals, Q&A sections etc and you can literally just ask the documentation.
- They are extremely good at recommending things like movies to watch or listing different options. It’s like having literally every expert in every field available at your finger tips at all times.
- They will be amazing at preserving our privacy/anonymity online by rewriting our texts to make them unidentifyable.
- They might allow us to communicate with animals more effectively. E.g. that one project to decypher whale language.
- Dying languages could be preserved for eternity.
- Lots of previously unused/unordered knowledge can be integrated into zeitgeist. Like indigenous knowledge.
- I would love to see LLMs trained ona giant conspiracy theory or esoteric corpus. Would be cool to talk to those lol.
I never considered 8 - 11. Those are really interesting use cases. I’m with you on every other point. I’m particularly interested in solving the messy unstructured notes scenario. I really feel you on that one. I’ll see what I can do!
I came up with a few more in case you need more inspiration:
- It would also be amazing for anamnesis (collecting a patient’s medical history). I feel like doctors aren’t very thourogh with this. Imagine if you could tell your medical history to an AI, and it will ask relevant questions expanding on that. E.g. “I have a stomach ache” “when does it happen? where is it? what does it feel like?” “After I eat, on the right side under my ribs, I can’t say” “is it sharp an piecing or more like a burn?”… etc. It could even give suggestions about what specialists to see and eventually you could integrate it all with statistical models that actually have diagnostic power. Ideally you would want one giant ensemble model you feed all your medical data into and it can diagnose you.
My experience is that you need to go to 10 doctors until you get the right diagnosis. If you have a model that gets it right only 50% of the time, that’s still better than the doctors. Especially if you train them to give you all possible diagnosis that fit your case.
- Same thing for psychotherapy. An AI is probably better than nothing. And I’m pretty confident that it could be a very good solution eventually.
I know these topics are sensitive, but the potential upside is enormous. Even if the model isn’t perfect, it just has to be better than the alternative. Which more often than not is no or shitty health care.
-
Also with plant and mushroom id LLMs could be helpful. They could ask you pertinent questions about the morphology of the plant/fungus, and give you a risk profile, what plants/fungi can easily be mixed up etc…
-
You could probably tell an LLM what food you have at home and it could suggest recipies.
-
You could make one talk with itself nonstop as two different characters when you’re out on vacation, to hopefully shoo away burglars.
-
Literally every kid would have the smartest, most patient teacher available to them 24/7
-
You could have a image to text model interactively explain the world around them to the visually impaired.
-
In general I hope that LLMs embrace reinforcement learning and become much more user customizable. Not just LLMs, other model too, like text to image or any to any. I could see models being the future where you basically get 9 outputs, you label the ones you like and the ones you don’t like, the you get 9 more, rinse and repeat, and the model output slowly moves to EXACTLY what you had in mind. Not just trough zero-shot learning, but actual learning too. Imagine you had a model that could generate movies. You just tell it what you like and what you don’t like and it will generate a first movie. Anything you don’t like during that movie, you just tell it on the fly. After the 10th movie hopefully it would only generate movies that are perfect for you.
-
Huge input length models like LongNet or retrieval based models would be great to generate entire worldbuilding scenarios. With detailed and coherent histories for many different countries/characters/arcs etc… You could have a video game that comes up with entire coherent worlds tailored to your likings, adapting to your decisions on the fly.
-
Similarly, a good night story telling model.
-
A parental advice model. So many parents have zero idea how to raise healthy children. Having an AI sidekick could help with that. An impartial force that will tell you when your parenting style is counterproductive.
-
You could let a bunch of LLM powered solar robots loose that prentend like they are spirits, causing mischief. (they will demand you solve a riddle, else they don’t let you pass the bridge)
-
You could train LLMs on the letters of famous people of history and talk to them as if they were still alive.
-
You could have your car guilt trip you if you are being a bad driver. Similarly it could join you in passionately complaing about other drivers.
-
You could have an LLM that is trained to plan your leasure time. Optimized to always make you do something different, go to a different local restaurant etc…
They’re super useful.
I use them at work a lot as a programmer.
They are incredibly good at boring repetitive refactoring, creating classes based on a SQL table etc.
If you post in an article they can summarize it or extract specific information really well.
People often complain they don’t get facts right but that’s not what they’re for.
I cannot understate how nice it is having a coding assistant 24/7.
I’m curious to see how projects like ChatDev evolve over time. I think agentic tooling is going to take us to some very sci-fi looking territory.
Semantic computation is the future.
LLMs are the patient, understanding teachers that adapt to your level that everyone wanted to have. They never get frustrated to why you didn’t understand something or that it’s taking to much time.
LLMs can summerise a long article, a youtube video without losing information.
LLMs* albeit still not good enough for usage with everything are able to describe an image pretty much accurately and it’s really great for people who can’t see. ( the state of the art is Llava if you want to test it out )
*The LLM is used along a visual decoder to describe the world but it’s been shown that the better the llm the better the vision assistant is able to describe the world.
I have this idea of using an llm to make legalese ( legal writing ) easier to understand and finally allow people to know their rights. Imagine asking an llm : I am a handicaped worker, what’s my rights ? And it give you a list.
I have high hopes for AGI. Either it wipes us ( Great ! ) or it make our lives better ( great too !)
I have learned everything I have about AI through AI mentors.
Having the ability to ask endless amounts of seemingly stupid questions does a lot for me.
Not to mention some of the analogies and abstractions you can utilize to build your own learning process.
I’d love to see schools start embracing the power of personalized mentors for each and every student. I think some of the first universities to embrace this methodology will produce some incredible minds.
You should try fine-tuning that legalese model! I know I’d use it. Could be a great business idea or generally helpful for anyone you release it to.
I have thought about the legalese Q&A and translation tool for a bit already.
The best way to have such tools be reliable using an llm without much hallucinations is via the use of embeddings of large quantities of legal documents and only ask the model to look for the ones that may answer the user question, a bit like with privateGPT. Also, always refer to the source and show it to the user.
Maybe even ask the user additional questions like : “is your question about finance, family, rights…etc ?” to decrease the error rate even further.
Then I’d have the llm put all the law articles and court cases it found in a list and use langchain to make it ask itself in each one? Is it really related to the user question ? Is it the right category ?.." to try and remove the most false positive possible .
Now that we’d have a cleaned list, ask the llm to combine what he got and transforme the legalese into understandable language. In this step the fine tuning ( honestly, i don’t know how yet) using legal documents could greatly help the model to understand the legalese better.
Could be a great business idea or generally helpful for anyone you release it to.
If you think it could be a good business idea then feel free to make it a reality. My main goal however is to allow people to know their rights. The government is quick to remind us of our duties but unless we seek to know our rights ourselves, they’ll be trampled by anyone and everyone. I’d like to Imagine a few kiosks scattered around town to help the citizens.
I researched creative AI and how AI can help people be creative, people thought it was a ridiculous and pointless topic. I’m biased.
Firstly, I think it’s important to see the non-chat applications. Goblin Tools is a great example of code we just couldn’t have written before. Purely from an NLP perspective, these tools are outstanding, if imperfect.
I’m excited to see new paradigms of applications come up when talented new developers are able to locally run LLMs and integrate them into their everyday programming, and too see what they can cook up in a world where that’s normal.
I’m interested in LLMs not to generate data on the fly, but to pre-generate and validate massive amounts of content or data than we’d otherwise be able to for things like games.
From a chat perspective, I like that it can support fleshing out ideas, parsing lots of data in a usable way.
And finally I’m excited for how lightweight LLMs could affect user interface design. I could imagine a future where OSs have swappable LLMs like they have shells that can allow for natural language interfacing with programs.
I don’t know, it’s just really accessible NLP, and that’s great.
What I find interesting is how useful these tools are (even with the imperfections that you mention). Imagine a world where this level of intelligence has a consistent low error rate.
Semantic computation and agentic function calling with this level of accuracy will revolutionize the world. It’s only a matter of time, adoption, and availability.
So I can finally create my own waifu irl.
I respect your honesty.
Thank you! But maan… working with AI stuff is expensive.
If you mean cloud computing or GPUs are expensive, you can do it locally on your CPU if your models aren’t too big. I like KoboldCPP. It’s not as fast but I only have to pay electricity for my AI waifu.
Thanks for the info. ik about llama.cpp and stuff but the problem is that I’m looking to run both speech to text, llm and text to speech all at the same time. I only have 8gb so yeah even CPU won’t cut it. I’m planning to upgrade once I get a job or smthing.
8GB of regular RAM? That’s not much. No, that won’t cut it if you also want all the bells and whistles. Maybe try something like the Mistral-7b-OpenOrca with llama.cpp quantized to 4bit and without the STT and TTS. It’s small and quite decent. Otherwise you might want to rent a Cloud-GPU by the hour on something like runpod.io or use free services like Google Colab or you really need to upgrade.
Search engines have become almost completely useless and AI can give me the answers I need immediately without having to wade through 17 piece of shit websites looking for a relevant answer. AI is a companion I never want to be without now. I hope that open source AI beats out all of the corporations and we end up with self-hosted private super AI that benefits all of us. AI already teaches me how to program what I want easily and effectively. Uncensored models are fun too for creative writing and naughty output. I use Spicyboros model with CPU only on my personal PC for queries that I want to keep private and ChatGPT everything else. The spicyboros-70b-2.2.Q5_K_M.gguf is very slow on CPU only and uses 50 GB of RAM but it is fun all the same. I made a script to generate text to speech once my query has processed using Coqui TTS
Google has absolutely tanked for me these last few years. It revolutionized the world by revolutionizing search. But ChatGPT has done the same, now better - and in a much more interesting way.
I’ll take a 10 second prompt process over 20 minutes of hunting down (advertised) paged results any day of the week.
I do roleplay with them. Program my own chatbot/waifu, just chat or download one of the character cards for SillyTavern. Sometimes use them to bounce ideas for creative things or let them assist me in writing a story. So I mainly tinker around.
I don’t use LLMs for anything “useful”. I can code myself. Okay, sometimes I give them small tasks. Re-word things, give them a draft and let them put it into more vivid and proper words, draft an email for me or translate something. Just a little thing here and there.
I hope they’re getting more useful. Once we solved alignment to a degree, and got hallucinations under control, I believe they can take a lot of tedious tasks off our shoulders. I for example hate paperwork. And organizing things. I’d like some AI to do it for me.
I believe they’re first getting useful in combination with a human and assist. You give it some input and let it do the writing, or fill out an excel sheet. Or have it as your pair programming partner. Or answer questions. Or draft something. And we have some kind of a labor shortage anyways. I believe this can be used to make things more efficient.
I don’t see it as a threat to human labor, but as a way to get more things done in a work-day and make things easier (if done right). As long as you bring something to the table, human brains are still (more) useful for quite some time. So I believe well qualified people don’t need to worry too much.
Once we get close to AGI I hope this will revolutionize economy in a good way. Make everything easier and more efficient so we can concentrate more on the interesting stuff the world has to offer. I still have my doubts, since there are a few technical problems in the way, we need better legislation, also with the way society and capitalism works, I doubt we’re getting there without major hiccups. And I hope we’re getting to be a part of it. AI is currently dominated by large corporations like >$100 million to the size of Meta. We need this to benefit the people and be democratic. Not go into the direction that benefits the large corporations the most. But I’m still very optimistic. We have governments to control corporations and provide at least somewhat of an level playing field. They need to step up their game, but I think this is doable.
