- GenAI360 - Weekly AI News
- Posts
- Generative Search Engine Faceoff, DeepMind + Meta’s New Biotech Models, You Only Cache Once
Generative Search Engine Faceoff, DeepMind + Meta’s New Biotech Models, You Only Cache Once
Plus, an Exclusive Meetup at LlamaIndex's New HQ
Before we start, share this week's news with a friend or a colleague:
Key Takeaways
Last week’s key developments include:
DeepMind’s AlphaFold 3 enhances protein folding models, while Meta’s RadOnc-GPT improves precision in radiation oncology treatments.
DeepSeek introduced its V2 model, a highly efficient 236 billion parameter language model with advancements in handling complex computations.
Prometheus-2 is a new model used for LLM evaluation and offers an alternative to proprietary models.
The XGen-MM model series handles image-text tasks and SQLCoder specializes in converting Postgres text to SQL.
Cornell University researchers tested Siri and Alexa's ability to show empathy across 65 human identities and found some major shortcomings.
Got forwarded this newsletter? Subscribe below 👇
San Francisco, an exclusive event with Activeloop + LlamaIndex + Tryolabs on Tue 5/28
We all know vanilla RAG is like the plain yogurt of the AI world. But why settle for plain when you can have a sundae sprinkled with advanced retrieval techniques, fine-tuning, and agents? Come to our in-person meetup to learn the secret sauce of building production-grade RAG engines from speakers from Activeloop, LlamaIndex, and Tryolabs.
Rumor has it this is one of the first big meetups at LlamaIndex’s swanky new HQ. So, apply now to be one of the first people to check it out since the spots are limited!
Coming up Next Week: Google I/O vs GPT-4-O
But first, some hot news fresh off the press that we will dig into deeper next week. Seems like everyone and their grandma is jumping on the multi-modality bandwagon.
Both OpenAI and Google asked their respective models to guess what's the announcement going to be about (spoiler alert: it was about the respective models being multi-modal and great). GPT-4-o is 2x faster, 50% cheaper, has 5x higher rate limits GPT-4-o and a more efficient foreign language tokenizer. The rest will come next week, alongside with I/O updates (we will also update you how many followers Google's Sundar Pichai has gotten after I/O on his newly-opened LinkedIn profile and if he accepted my invite).
The Latest AI News
Biotech advancements continued last week with DeepMind and Meta releasing new models.
We also saw OpenAI make a whole range of moves, including the release of The Model Spec, negotiating deals with publishers like Axel Springer, and confirming the development of a search engine for ChatGPT.
Multiple Moves From OpenAI
OpenAI released “The Model Spec”, which sets guidelines for its AI models to follow - especially when using reinforcement learning from human feedback (RLHF). It aims to shape behaviors in complex scenarios by outlining objectives and rules so that users and developers have more control over the models.