April 2, 2024 Voice twins in 15s

Hey folks! It’s models models models today. Three new models are giving a glimpse into the future. We’re covering…

Let’s get into it!

PRODUCT HIGHLIGHT

An open-source alternative to GPT-4 that performs

Everyone in the AI space has been talking about DBRX, a new open-source LLM from data intelligence platform and AI company, Databricks. It works similarly to OpenAI’s GPT-4, and it’s apparently very good.

A new milestone for open-source AI: Jonathan Frankle (Databricks) "shared data showing that across about a dozen or so benchmarks measuring the AI model’s ability to answer general knowledge questions, perform reading comprehension, solve vexing logical puzzles, and generate high-quality code, DBRX was better than every other open source model available," writes Will Knight for Wired.

Not only did DBRX outperform Llama 2 (Meta), Mixtral (Mistral), and Grok (X), it got close to GPT-4’s performance “on several scores.”

A win for Databricks: Last July, Databricks acquired MosaicML, an open-source AI startup, for $1.4 billion. The team has since spent about $10 million training DBRX, which has been at the top of Hugging Face trending board since it launched.

Databricks is betting this investment will pay off in its ability to help industries that want to use AI but have had their hands tied because of data security and privacy concerns (e.g. medical and financial companies).

What’s next: DBRX is available on GitHub and Hugging Face for research and commercial use. Databricks will be releasing a blog about how they created the model, which, if done well, will also be a refreshing change from the secretive world of training LLMs.

NEWS

OpenAI’s new voice model

Image created by DALLE

Last week, OpenAI revealed a new model called Voice Engine that can clone a person’s voice with just a 15-second audio recording. OpenAI didn’t share the tool itself — only learnings the company had during a small-scale preview of the tool.

Why the slow rollout? It’s not because the tool isn’t impressive, Just listen to the examples.

It’s because the tech can be dangerous in the wrong hands, especially during an election year, explains OpenAI in its blog. So for now, Voice Engine is only being tested by a “small group of trusted partners” who have agreed not to do anything shady with the tool like recreate voices without consent.

It’s all great reasoning, but between waiting on Sora and now Voice Engine, I’m chomping at the bit.

WOAH

An AI with emotion

Image: Hume AI

A startup called Hume AI launched a demo of its “Empathic Voice Interface (EVI)” — i.e. a conversational voice bot that is giving people the heebie jeebies.

Tell the bot about your problems (or wins) and he’ll respond appropriately, with enough emotional intelligence to carry on a decent conversation.

Hume AI is co-founded and is led by CEO Alan Cowen, a former researcher at Google DeepMind. It recently bagged $50 million in a Series B fundraise.

TOOLS

Circle to Search lets you circle content to get immediate results about it.

Living Images lets you A/B test your images and auto-replace them with AI.

Daytona is an open-source dev environments manager that automates your setup.

OpenDevin aims to replicate Cognition’s Devin, but make it open-source. 

Dodoboo is an iPad app that turns kids’ doodles into AI-generated art. 

Skeleton Fingers is a super simple and free AI audio transcription tool. 

MORE READING

“Many-shot jailbreaking” – Anthropic shares its discovery on safeguarding a new way to find a model’s vulnerabilities.

A little guide to building Large Language Models in 2024 by Tom Wolf, Hugging Face co-founder

Thanks for going deeper with us!

Written by Sarah Wright.

Here via forward? Subscribe here.

Reply

or to participate.