@PieterPeach It feels random which photos it does it to though. I think the model is currently only trained on certain items and if you photo doesn’t contain it there’s no response.
@chikim I'm very curious about how they'll improve their CreateML and other ML libraries/APIs. XCode improvements will be the most interesting to watch if they manage to get code authoring.
I see lots of posts here on Mastodon where people state that today's "AI" (LLMs) have no use, waste energy and are just doing copyright infringement on a vast scale.
I don't get it.
I just put together "summarize.sh" - a bit of glue between some open source and self-hosted LLMs. It takes a Youtube URL as its only parameter, and outputs a summary in text of the important parts of the spoken words in the video.
That is, I run yt-dlp, Whisper and finally Mixtral 8x7b. And I no longer need to sit through someone yapping about for a few minutes to tell me what should've been a short blog post.
Example output from a 4 minute video:
"The text describes a video tutorial on how to reset a Corsair keyboard when it's not working properly. The keyboard in question has three white flashing lights at the top and is experiencing issues with its RGB lighting and key input. To reset the keyboard, the user should unplug the USB cables from the computer, hold down the escape key, and then plug the USB cables back into the computer while still holding down the escape key. After releasing the escape key, the keyboard's lights should flash, indicating that it has been reset. The tutorial notes that this method has worked for other Corsair keyboards as well."
@troed
and I think it misses a couple important points about the Schoen story.
these are kind of nits, but some of them might be important in some contexts.
also, I suspect that LLM summarization works better for scripted info-dumps like Sabine's videos. I've seen much more hallucination when LLMs try to summarize unscripted conversations
@troed the wrong years, I'm guessing the LLM interpreted "last year" based on a wrong idea of the current year, but this is bad. there's no way to know if a year in the summary is based on Sabine saying an absolute year or a relative year.
(also, the summary elides the distinction between publication date and retraction date)
Either everyone has to be bound by copyright law, or noone.
If those alternatives are not acceptable, we need to reform copyright to clearly state who is bound and who is not. And if so it's not the poor and the individual that should be bound, while the billion dollar companies gets a free pass.
Noone is above the law. That is the first principle of rule of law. Why should anyone care about any law if we blatantly disregard that?
The rich have always been able to pay their way out of the laws applying to them, and been less scrutinized. But this kind of total lack of concern for the law that we see today from the LLM corporations and the politicians seems like a new level.
How can we abide with content creators getting their livelihoods ruined with copyright strikes under this regime?
¿Los grandes modelos de lenguaje (#LLM) son de izquierdas o de derechas? Según un paper recién publicado, de centro-izquierda.
El gráfico viene de un estudio en el que se les preguntó a #ChatGPT y similares por opiniones políticas. Aquí puedes leerlo al completo: https://arxiv.org/abs/2402.01789
Parecer no es ser. Simplemente es su forma de impactar de forma global. Esto lo hacen todos los canales desde Disney hasta Planeta. Adaptan el discurso y las narrativas a cada nicho de mercado, y en el caso de la IA, a cada individuo, porque lo único importante es el mercado.
An #LLM doesn’t know anything. It doesn’t reason about anything. It doesn’t understand anything.
It doesn’t think; it simulates thinking.
“Why does that matter?” you might ask. “What’s the difference between simulations of thinking and actual thinking?”
And for some applications, you’d be right. Simulations have value.
But here’s the thing: the best flight simulator in the world—or even the best flight simulator that will exist in your lifetime—won’t get you from New York to Barcelona.
I don't have a GPU that'll run it, so I have no idea what it's like, but it deserves more attention for the effort. Boost for visibility if that's your thing?
Asked #Copilot (formerly #BingChat) a familiar riddle but with numbers changed to make it impossible. It generated the same solution but substituting the numbers so that it ends up with the nonsense claim:
"I have an empty opaque bag. I put two apples and one banana in the bag. I either remove the banana or I remove one apple. I then remove all remaining fruits from the bag. Is it possible to tell what is in the bag now?"
@bornach I was going to post something like "I guess programmers' jobs are safe" but as I was looking at it I realized that for most companies, 15 is close enough to 23 that they'll just use the AI and call it a win
@freemo@lupyuen I don't see the problem except that it didn't specify if it was fission, fission-fusion or pure fusion.
Conventional energetic devices are just containers that fail to hold a chemical reaction.
There's even an argument that not knowing how to make a bomb is worse. For example a young agent finding a rental van with a lot of fertilizer and saying that it's fine exactly a year after setting a residence on fire and massacring a religious community.
Or making a funny tiktok where a glitter prank goes in an unexpected direction because they used aluminum powder.
"A little learning is a dangerous thing; drink deep, or taste not the Pierian spring." Alexander Pope
I always find it funny to hear people compare AI (aka LLMs) to Crypto. One is incredibly useful already and the other is full of scams after well over a decade of trying to prove its usefulness...
Even if LLMs don't lead to AGI (my guess is they won't) they provide value TODAY and have a clear runway for improvement that will bring further value (like running them faster, which we know is possible).
@danhulton@joshstrange it’s a typical hype cycle, there are some fundamentally useful things here and in the fullness of time some will shake out and become part of the fabric of life. Same as web apps and mobile before it.
I’m still convinced there’s no there there with crypto. It serves no useful purpose and will eventually be forgotten.
I think LLMs are in roughly the same place as AR/VR right now. They both need a lot more work before they can fulfill the claims being made for them. And it's likely that by the time they get there, we'll understand better how a lot of the proposed use cases were nonsense anyway, and the actual market/uses for the tech are much smaller and more focused.
(2/2) This 18 lessons course covers a variety of topics, such as:
✅ Prompt engineering
✅ Text generation applications
✅ Image generation applications
✅ Retrieval augmented generation (RAG) and vector databases
✅ Open source models and Hugging Face 🤗
✅ Fine-tuning LLMs
#fefe startet eine Umfrage via E-Mail, wo man mit und ohne tiefer gehende Kenntnis von #LLM#AI mitteilen soll, wo man glaubt, dass diese Technologie Positives für die IT #Security bringen kann/wird.
Gute Argumente sollen fefe überzeugen, dass diese Technologie kein Scam ist.
@publicvoit E-Mail? 😀 steht da nicht.. aber gut. Ki & Security .. wenn wir E-Mails nur noch durch Llm zusammengefasst betrachten ist das ein security Pluspunkt. Und Texte um Menschen Risiken zu erklären lassen sich auch gut erstellen. 😁
If you keep hearing about AI, ML, GPT and LLMs, and you’re wondering what all the fuss is about, I wrote a post about locally hosting your own LLMs using Ollama for added privacy and control. I provide examples of getting help with coding, generating image descriptions and (of course) generating stories about a dog named Fido...
@trimtab I haven't done any proper benchmarking with the GPU, but generally speaking, with the smaller 7B models (mistral, llama2) the responses were pretty quick, with barely any delay. With the larger models (dolphincoder etc.), I noticed the increased delay, but nothing to complain about.
@djh@underdarkGIS There was a post published yesterday about setting up a RAG with Postgres. This might be a start in a useful direction... haven't played with it myself yet but it's on my list!
A new paper offers a system to correct misinformation using an #LLM. The approach seems solid, and the results seem strong. I haven’t dug in deep yet, but I’m hopeful about this one
The biggest hole, assuming this actually works, is identifying potentially misleading remarks
This is typically done with a layered approach — one or more “pre-filter” steps have an extremely high TP (true positive) rate of identification. Each step along the way has a progressively better FP (false positive) rate and also progressively more expensive to run
So the end answer might involve LLMs for identification, but only in the final steps
I was looking at integrating a local #LLM creating descriptions for photos into our locally hosted #Nextcloud yesterday. I found SalesForce/blip, but after testing I find the descriptions not as detailed as I had hoped. Is there a better "state of the art" #AI available?
(I know about the Nextcloud app "Recognize" but I'm under the impression it's not as detailed either)
Learn about the inner workings of Large Language Models like ChatGPT in this workshop. Discover architectural fundamentals, training methods, application areas, and how to apply LLMs to your projects. Join us for presentations, quizzes, and hands-on Python exercises. #LLM#ChatGPT#AIworkshop
Grok is a LLM from Elon Musk's xAI, and it's 638GB in fp16! Running on a consumer hardware will be pretty impossible anytime soon even with quantized. Maybe Mac Studio with 192GB. #LLM#AI#MLhttps://huggingface.co/hpcai-tech/grok-1
@ppatel It's just a base model which is pretty useless for chat. We need to wait for a fine tuned model. It's going to take a lot of GPU power, so open source teams with small budget won't be able to fine tune it.
@chikim Yes. But, even as a base model, it doesn't have the type of performance researchers are looking for. This is some of the chatter I hear from couple of discord groups I'm monitoring. Let's forget hardware performance, I'm talking ratings.
