DALL-E 3 Preview | Is AI Getting Dumber? | RAG vs. Fine-Tuning | Google’s NotebookLM | Midjourney Goes Print
My Generative AI Digest Sep-25 2023
OpenAI Gives Preview of DALL·E 3 and … wow!
Midjourney has gained quite an edge lately — but with DALL·E 3 OpenAI has finally upped its game.
DALL·E 3 can accurately represent a scene with specific objects and the relationships between them.
– OpenAI
This is even more impressive, as almost all AI art models have previously struggled to accurately convert prompts to images, often simply ignoring some of the information.
Preview: DALL·E 3
RAG vs. Fine-Tuning
This is a question which has bothered me for eons (OK, since 2021): When should I fine-tune my LLMs and when is Retrieval Augmented Generation (RAG) to be preferred? While fine-tuning is the classical approach, RAG retrieves the most query-relevant documents, which the model can then access via the prompt context.
Heiko Hotz: RAG vs. Fine-Tuning — Which Is the Best Tool to Boost Your LLM Application?
Magic Google Notebooks
Any sufficiently advanced technology is indistinguishable from magic.
– Arthur C. Clarke
Google is beta-ing on a Harry Potterish student notebook which directly comments or answers my notes. Magnificent! Can someone please start working on Newt’s suitcase? You know, the one that can store an entire forest of magical creatures or at least the extremely extensive belongings of certain family members on vacation.
Waitlist? Yup. And set your VPN to the US.
Join: NotebookLM
AI Art with Text in the Image
Finally: AI Art with text that doesn’t look like someone smuggled a scroll of Babylonian hieroglyphs under their armpits out of Iraq. Ideogram is a Midjourney alternative capable of incorporating (legible) text. And even better yet: It’s free!
Try it out: Ideogram
Is AI Finally Getting Dumber?
AI apocalyptists can breathe a sigh of relief: A recent paper sees a strong drift in the capabilities of the same GPT models — in many cases the performance got worse over the course of time.
Secure LLM Responses
The age-old human dream of secure LLM responses. With Guardrails AI there is a framework that tries to get exactly there.
Essentially Guardrails enables a Guard that wraps around the LLM to validate and correct the outputs based on predefined Rails. So Flat Earthers, the next time your AI claims the earth is round, you at least now know where to fix that.
Documentation: Guardrails AI
Jesus Rodriguez: Inside Guardrails AI
Midjourney Goes Analog
We all knew the Internet would be a fast-fading hype. The AI art platform Midjourney has now pulled in the reins and switched back to print. This glossy best-of magazine will be hurled to your doorstep by the postman.
This sleek new magazine features a selection of the best Midjourney artworks.
Secure a copy here: Midjourney Magazine or if you prefer real communication, just write them:
Midjourney, Inc.
611 Gateway Blvd. Ste 120
South San Francisco, CA, 94080–7066
Learning Machine Learning. This week: Understanding and Applying Text Embeddings.
Featuring Nikita Namjoshi and Andrew Ng, this course deals with text-to-vector conversion, frequently employed in semantic search (like RAG) or classification. Structured in a compact 1-hour format, you’ll learn to measure semantic similarity between texts and how to incorporate it into your LLM. Requirements: Basic Python Knowledge.
Check out the course: Understanding and Applying Text Embeddings
Unsupervised Laughing
Based on a document or image, some AI Art models can extrapolate previous versions of it: Here, Stable Diffusion XL tries to unveil Early Drafts of the Declaration of Independence.
Other versions of this document by unintellisense on Reddit
Thank you, have a great week, may the world align with your wishes!
Maximilian
