The Generative Edge Week 18
DeepFloyd shows you beautiful text, VectorDBs are rolling in cash and can ChatGPT replace your datascientists?
Welcome to week 18 of The Generative Edge. Here is the gist in 4 bullet points:
New image model DeepFloyd IF generates high-quality images and legible text; try it at https://huggingface.co/spaces/DeepFloyd/IF.
Vector databases gain traction as essential tools for AI, with Pinecone and Chroma raising significant funding.
ChatGPT's code execution plugin allows data analysis and visualization through interactive dialogues, similar to data science notebooks.
"Godfather of AI" departs Google over concerns about the existential risks associated with AI technology.
For everything else, let’s hop right in!
A new image model in town: DeepFloyd IF
Dall-E, Midjourney, Stable Diffusion - the list of generative image models is ever growing, not to mention the models that haven’t even been released to the public yet (Imagen, Parti). A novel model called IF has been developed by the research team DeepFloyd, which works under the umbrella of Stability.AI:
DeepFloyd IF is based on a different architecture compared to the well known diffusion models
It is much better at generating legible text and it’s very good at following complex prompts
You can try it easily here: https://huggingface.co/spaces/DeepFloyd/IF
Enter a prompt (e.g. an astronaut riding a horse, with the text “to the moon” in the background), click generate, pick an image, then click upscale
The vector db gold rush
Vector databases are shaping up to be an essential tool in the LLM/generative AI ecosystem. These specialized dbs allow for search tasks like “based on this query, find me documents that are semantically similar:”. Often, they play the role of memory in large language models, as well as providing various functions (such as semantic search) to help the language model access relevant and current information.
Chroma raises $18 million in seed funding for their OpenSource vector db
Semantic search was already gaining popularity before the generative AI craze, but it has only accelerated since
VectorDBs are “shovels” in this AI gold rush, possibly essential tooling in this space, which is reflected in the investment volume.
Datascience and ChatGPT
AI is coming for a lot of cognitive/mind jobs, but is it also coming for datascience? We’ve talked about ChatGPT’s plugin system before. In addition to plugins, ChatGPT also has other tooling available, like code execution, which can help with a lot of datascience tasks:
The code execution plugin (still in closed alpha and invite only) lets you upload data (e.g. CSV data) into ChatGPT
It will suggest sensible data analytics based on the uploaded data.
Using the ChatGPT typical dialog based back-and-forth, you interact with the assistant, data and visualizations at once.
The resulting output is somewhat similar to a datascience notebook
Check out these Twitter threads for more examples:
https://twitter.com/emollick/status/1653069121704058883?s=20
As with all generative AI tools today, you will get the best results if you actually know what you’re doing. Will a layman perform these analysis in the future?
… and what else?
“Godather of AI” leaves Google, citing “existential” risks involved with the technology.
And that’s it for this week!
Find all of our updates on our Substack at thegenerativeedge.substack.com, get in touch via our office hours if you want to talk professionally about Generative AI and visit our website at contiamo.com.
Have a wonderful week everyone!
Daniel
Generative AI engineer at Contiamo