Founder Webshrinker, VocalVector, #Homebrewer 🍺│https://teapaw.org/t.co/rZrdvGUx0v │#MachineLearning #ArtificialIntelligence #BigData #Linux #AWS #SoftwareEngineer
Greater NYC Area
Joined August 2009
- Tweets306
- Following250
- Followers489
- Likes302
AGI has also been overhyped. It's interesting but also not needed for the vast majority of applications where current generative AI can be used.
A good example of this is how it's being used at vocalvector.ai
Many people have been recognizing that AGI won't be coming from mere scaling of current tech, and that generative AI has been severely overhyped. This is all true. But those people often conclude, "therefore AI is not going to be transformative -- this is a nothingburger". Absolutely not.
AI (both current and near-future tech) *will* transform nearly every industry, and we're still only in the very first steps of that process. Generative AI may be a bubble, but AI is going to be bigger in the long run than what almost all observers currently anticipate.
Adam Spotton ☕ retweeted
LLM model size competition is intensifying… backwards!
My bet is that we'll see models that "think" very well and reliably that are very very small. There is most likely a setting even of GPT-2 parameters for which most people will consider GPT-2 "smart". The reason current models are so large is because we're still being very wasteful during training - we're asking them to memorize the internet and, remarkably, they do and can e.g. recite SHA hashes of common numbers, or recall really esoteric facts. (Actually LLMs are really good at memorization, qualitatively a lot better than humans, sometimes needing just a single update to remember a lot of detail for a long time). But imagine if you were going to be tested, closed book, on reciting arbitrary passages of the internet given the first few words. This is the standard (pre)training objective for models today. The reason doing better is hard is because demonstrations of thinking are "entangled" with knowledge, in the training data.
Therefore, the models have to first get larger before they can get smaller, because we need their (automated) help to refactor and mold the training data into ideal, synthetic formats.
It's a staircase of improvement - of one model helping to generate the training data for next, until we're left with "perfect training set". When you train GPT-2 on it, it will be a really strong / smart model by today's standards. Maybe the MMLU will be a bit lower because it won't remember all of its chemistry perfectly. Maybe it needs to look something up once in a while to make sure.
GPT-4o Mini, announced today, is very impressive for how cheap it is being offered 👀
With a MMLU score of 82% (reported by TechCrunch), it surpasses the quality of other smaller models including Gemini 1.5 Flash (79%) and Claude 3 Haiku (75%). What is particularly exciting is that it is also to be offered at a cheaper price than these models. The reported price is $0.15/1M input tokens and $0.6/1M output tokens.
With such a cheap price for input tokens and its large 128k context window, it will be very compelling for long context use-cases (including large document RAG).
@OpenAI have clearly made a very high quality model relative to its size (pricing can indicate size due to the direct relationship to compute cost). The model seems a worthy successor to GPT3.5 Turbo as OpenAI's smallest model and the model used for ChatGPT's free version.
Had a chance to test the latest Fooocus image generation model. It's free, quick, not that resource intensive, and can run on Colab which is great.
It seems to still have issues with rendering faces, fingers, and artifacts around eyes though.
Worth keeping an eye on its progress!
#stablediffusion #midjourney
Adam Spotton ☕ retweeted
I just found out the @CERN uses @FastAPI (and @pydantic) 🚀
...to control particle accelerators 🧪
...including the Large Hadron Collider 🤯
What is that word for humbled and frightened? 😅
epc-ccs.web.cern.ch/content/…
English is the new programming language for LLMs with positions like Prompt Engineer.
What is social engineering of LLMs called? Adversarial attacks don't seem to fully describe it.
#Serverless Framework with Python feels like an unholy union of #npm and #python but there is no denying that it's 🔥
Django is Python based which is great and there's huge community support for it. In the past it seemed very bloated so I avoided using it for any main projects but seems to have improved a lot since then.
If you want to check it out it's called Rift and it's open source: github.com/morph-labs/rift
Adam Spotton ☕ retweeted
Stop using ChatGPT in beginner mode!
No one talks about these 5 secret settings that can instantly improve your AI writing game.
Don't waste your time! Steal these and start using them NOW.
I struggled with ChatGPT sounding robotic and with the persistent use of weasel words and repetitive phrases.
After countless hours of research, I found the solution to my struggles with ChatGPT!
Here's what I discovered about the solution - Using of parameters!
1) Max tokens
Max tokens determine the maximum length of the generated text. By setting a limit, you can control how much the AI says, ensuring it doesn't give too long of an answer.
👉Example: max_tokens = 50 - If you want a short response, like a quick answer or a tweet.
2) Temperature
The temperature parameter controls the randomness of the text generated by ChatGPT.
Generated text with a lower temperature will be more focused and conservative, while generated text with a higher temperature will be more creative and varied.
Temperature ranges from 0 to 1.
Here’s a quick breakdown of how it works:
▪️ Low temperature (0 to 0.3): More focused, coherent, and conservative outputs.
▪️ Medium temperature (0.3 to 0.7): Balanced creativity and coherence.
▪️ High temperature (0.7 to 1): Highly creative and diverse but potentially less coherent.
👉Example: temperature = 0.7 - If you want a fairly creative response, good for brainstorming or idea generation.
3) Top_p
This parameter, ranging between 0 and 1, controls the nucleus sampling, a method to introduce randomness.
▪️ Values closer to 1.0 make the output more diverse and random
▪️ Values closer to 0 make it more deterministic.
👉Example: top_p = 0.8 - If you want diverse options while generating multiple names for a new product.
4) Presence_penalty
This parameter, between -2.0 and 2.0, discourages the model from introducing new topics. Higher values make the conversation more focused but less interesting.
👉Example: presence_penalty = 0.6 - If you want the model to stick to the topic, like discussing details of a specific project.
5) Frequency_penalty
Between -2.0 and 2.0, this parameter discourages the model from using common phrases or responses. Higher values make the output more creative but potentially less coherent.
👉Example: frequency_penalty = 1.0 - If you generate a slogan or a tagline and want it to be unique and not a common phrase.
Let’s see it in practice! Use case time:
📌Twitter and other social media
Writings on social media need to be catchy and engaging. You should experiment with higher creativity to stand out from the crowd.
Use these settings:
▪️ Temperature: 0.7 to 1 for more creative and diverse outputs.
▪️ Frequency_penalty: 1 to 2 for maximum diversity without sacrificing too much coherence.
▪️ Max tokens: Set a limit corresponding to the platform’s character limit.
📌Creative storytelling
You should explore more diverse and imaginative outputs for creative writing or storytelling.
Experiment with these settings:
▪️ Temperature: 0.7 to 1 for maximum creativity and variety.
▪️ Frequency_penalty: 1 to 2 for various unique phrases and ideas.
▪️ Max tokens: Set according to your desired story length or scene.
Here is an example of how you can craft your prompts using the parameters:
1) Focused on topic, minimal creativity
Prompt: Temperature = 0.0 generate a tweet about chatgpt
Result: "ChatGPT is an advanced AI developed by OpenAI. It's capable of understanding and generating human-like text, opening up new possibilities for automation, content creation, and more. Explore the future of AI with ChatGPT today."
2) Balanced creativity and coherence.
Prompt: temperature = 0.7 generate a tweet about chatgpt
Result: "Discover the power of language and creativity with ChatGPT! Harnessing advanced AI, it weaves together complex data into insightful narratives. Dive into the realm of endless possibilities, one conversation at a time. 🚀 #ChatGPT #AI #LanguageRevolution"
3) Highly creative and diverse but potentially less coherent.
Prompt: temperature = 1.0 generate a tweet about chatgpt
Result: "Unleashing the magic of words, one byte at a time! ✨ With #ChatGPT, every conversation sparkles with intelligence and surprises. It's like having a digital companion who's always ready for a brainstorm, a debate, or a delightful chat. Can't wait to see where the conversation leads! #AI #ConversationsWithAI"
If you enjoyed this post, I'd really appreciate 🙏 if you would:
👉Follow me @AIFastTrack for more of these
👉Please RT the post to help your friends
One of my favorite "tricks" about Linux is that you can delete a file that's still open and being used by applications. As long as your application keeps the file handle open, even when the file gets deleted you can read and write to it just fine.
For Windows, there have been entire blog posts about how to track down which application has the file open so that you can terminate it in order to delete the file without an error
Serious Linux flex 💪💪💪
A great foundation for building the next ML/AI tool on top of! Hope to see even more embedding datasets made available like this in the future 🤓
1/ 🚀 What could you build if your app had the entire knowledge of Wikipedia? This post by @JayAlammar & @Nils_Reimers shows you how to use 100 million embedding vectors that cover Wikipedia in 10 languages.
Download them and start building today!
txt.cohere.com/embedding-arc…
You can now talk with #ChatGPT via a phone call. The low latency audio is pretty impressive and makes this useful, thanks @CallAnnieAI
US: 1-640-225-5726
IMO the paper on testing the capabilities of #GPT4 and its reasoning is a must read.
Interesting how it was able to pass the Amazon Software Engineer screening test in mere minutes instead of hours, amongst other tasks!
arxiv.org/pdf/2303.12712.pdf
I look forward to this report every year and it has never disappointed! Congrats to @amywebb and the team at @FTI for another excellent report
Our 2023 Tech Trend Report has officially launched! This is not only our biggest report to date, we also cover new verticals, and we have added an Executive Summary book, where we cover the key insights of each topic. buff.ly/2NwQyHm
