Sunday Rundown #60: Many Mini Models & Sketchy Superheroes
Sunday Bonus #20: Adjusting ChatGPT's creativity with a simple command.
Happy Sunday, friends!
Welcome back to the weekly look at generative AI that covers the following:
Sunday Rundown (free): this week’s AI news and an AI fail for your entertainment.
Sunday Bonus (paid): a goodie for my paid subscribers (a guide, tip, walkthrough of a tool, etc).
Let’s get to it.
🗞️ AI news
Here are this week’s AI developments.
Lots of players have launched smaller, more efficient models at the same time.
👩💻 AI releases
New stuff you can try right now:
Amazon’s AI shopping assistant Rufus—in beta since February—is now officially available to all US customers.
Anthropic now has a Claude Android app, which you can download from the Google Play store.
If you’re using Claude 3.5 Sonnet within Anthropic API, you now get twice as many output tokens (8192 instead of 4096).
Google is starting to gradually roll out Vids—its “AI-powered video creation app—within Google Workspace.
Hugging Face released SmolLM: a family of tiny language models that can run on local devices. Available in three sizes: 135M, 360M, and 1.7B parameters.
Microsoft expanded its Microsoft Designer with new features (Restyle Image, Replace Background, stickers, and more) and made it available as a free app for iOS and Android.
Mistral has launched not one, not two, but three smaller models in one week:
MathΣtral, a 7B model designed for “math reasoning and scientific discovery.”
Codestral Mamba, a 7B model specializing in code generation.
Mistral NeMo, a state-of-the-art 12B model with an impressive 128K token context window.
OpenAI also has a new state-of-the-art smaller model called GPT-4o mini that beats Gemini Flash and Claude Haiku on most LLM benchmarks.
Spotify now has a Spanish version of its AI-powered DJ.
YouTube Music is testing two new features: prompt-driven AI radio ala Spotify and a song-recognition tool ala Shazam. Available to select Premium users in the US.
🔬 AI research
Cool stuff you might get to try one day:
Meta AI is set to release the biggest model in its Llama 3 family— Llama 3 405B—on July 23.
OpenAI is reportedly working on new reasoning tech codenamed “Strawberry” that can perform deep research, potentially approaching Level 2 on the company’s internal AGI scale (see #2 under “AI resources”).
📖 AI resources
Helpful stuff that teaches you about AI:
GPT-4o Mini: How ‘Mini’ Is Its Intelligence? [VIDEO]: a look at the state of LLM reasoning and what’s coming by AI Explained:
OpenAI Scale Ranks Progress Toward ‘Human-Level’ Problem Solving: summary by Axios of a Bloomberg-reported OpenAI scale for tracking the five levels of AGI.
Prover-Verifier Games improve legibility of LLM outputs: research paper by OpenAI into making LLM output more understandable for humans.
The Future of Knowledge Assistants [VIDEO]: a talk by LlamaIndex CEO Jerry Liu about moving towards AI agents:
🔀 AI random
Other notable AI stories of the week:
Andrej Karpathy is starting an “AI-native” school called Eureka Labs that leverages an AI assistant to guide students through courses developed by a human expert.
🤦♂️ 10. AI fail of the week
Ah yes, my favorite Marvel show: “Super America Falcon vs. Special Spider-Bat”
Anything to share?
Sadly, Substack doesn’t allow free subscribers to comment on posts with paid sections, but I am always open to your feedback. You can message me here:
💰 Sunday Bonus #20: Weird trick that lets you tune ChatGPT’s creativity
Now, if you want an AI chatbot to give you more creative and imaginative answers, you can simply ask it to “give more creative and imaginative answers.”
It’s not rocket science, folks.
But I stumbled upon a curious trick that gives you a sort of virtual “knob” to adjust ChatGPT’s creativity in a more nuanced way.