PLUS: Anthropic's 100k context limit
Good morning human brains, welcome back to your daily munch of AI news.
Here’s what’s on the menu today:
Claude’s 100k Limit 🚦
Anthropic’s Claude can now process 75,000+ words.
StabilityAI just released Stable Animation for developers.
AI learns to bowl 🎳
A YouTuber trained an AI model to bowl strikes.
Anthropic’s 100k Context Limit 💯
Anthropic just announced that its chatbot Claude can process up to 100,000 tokens, or 75,000 words. That's right, GPT’s most direct competitor can now read and analyze a full novel faster than you can get to page 2.
GPT4’s context window is only 32k tokens long
During testing, Anthropic loaded The Great Gatsby (208 pages) onto Claude and made a single change to the text - turning Mr. Carraway into a software engineer at Anthropic. Claude was able to spot the change in 22 seconds.
You can probably hold onto your bookmarks a little longer though - Anthropic is only focused on the enterprise world for now. The expanded context window is set to help businesses quickly digest lengthy docs, assess legal risks, and comb through dense developer documentation.
Claude’s new power is only available with an API that’s behind a waitlist (request access here). Testers with API access say the new Claude is impressive, but performance drops the more you make it process.
One more thing: Anthropic also announced that it’s started training Claude based on a constitution it designed using the UN Declaration of human rights.
How this works: Anthropic uses a different AI model to review and discard Claude’s most disturbing outputs. Then, Claude's Constitution of 10 principles guides it toward avoiding toxic outputs and ultimately creating an updated system that’s more "helpful, honest, and harmless” than before.
This training method reportedly produces significant improvement in the AI's performance compared to models trained only on human feedback.
Our Prediction 🔮: OpenAI’s about to reveal its own guiding constitution for training models in the next two months.
BUZZWORD OF THE DAY
The amount of data an AI system can “think” about at one time.
The bigger the limit, the more data you can provide as input to the model.
When you hit the context limit in tools like Claude or ChatGPT, you’re exceeding the amount of information the tool can process.
If you’re interested in learning more about how large-language models like Claude and ChatGPT “think”, this video by Computerphile is a good place to start.
FROM OUR PARTNERS
Chat With Your Files
With WriteSparkle, you can streamline your content creation process by seamlessly integrating AI with all your favorite tools and platforms.
1/ Chat with your docs - extract, create, and summarize content.
2/ Create content with AI - create blog posts, reports, and more.
3/ Automate your workflows - integrate AI with all your apps.
4/ Semanticly Search - search through all your files in seconds.
StabilityAI makes pictures move
Bored of static images? StabilityAI just joined the text-to-video game with Stable Animation. Here’s everything you need to know:
1/ Stable Animation is a new way to create moving pics, with three different input options: text prompt, text prompt + image, and text prompt + video.
2/ Stable Animation has no cap on video length. It’s also compatible with all versions of Stable Diffusion and adds new overlays for anime, comic books, pixel art, and more styles.
3/ Every video will cost you between 3 - 18 US cents per 100 frames (considerably lower than both SDXL and Midjourney’s cost per image).
4/ Resolution goes up to 1,024 x 1,024 pixels (only with an upscaling tool, but a new high nevertheless) and while the default frame rate is 12 fps, StabiltyAI’s documentation claims it can go up to infinity.
5/ It’s still experimental, and StabilityAI mentions no plans for a chat interface. Right now, you can only play with it using StabilityAI’s SDK & paid API.
But why work on animation when you’re trailing months behind Midjourney?
Animation tools based on Stable Diffusion have already entered and validated the market - StabilityAI’s play with Stable Animation is killing off-the-shelf startups with a more comprehensive solution that they can control.
If you’re interested in installing the SDK and running Python scripts to generate psychedelic cats, here’s the how-to.
MEMES FOR DESSERT
YOUR DAILY MUNCH
Why you’re probably underestimating AI bots.
AI invents new bowling techniques (11m).
Microsoft backs software startup Builder.ai.
Together raises $20M to build open-source generative AI models.
Moving towards the third generation of AI: establishing a new explainable and robust AI theory.
The benefits, limits, and risks of GPT-4 as an AI assistant for medicine.
Slatebox: create and edit visualizations with prompts.
MealPractice: discover, follow & create menus and recipes.
Rosebud: journal your personal achievements and goals.
DrafterAI: build no-code tools and workflows.
What'd you think of today's newsletter?
TWEET OF THE DAY
An org working on digitizing humans for the past 14 years just demoed new tech that creates can perfectly relight faces in deep faked videos.
The Laboratory - @Julian_cona_
Until next time 🤖😋🧠