Daily Digest: The end of GPUs

PLUS: New RAG specific model and consistent characters.

Want to get in front of 100k AI enthusiasts? Work with us here

Hello folks, here’s what we have today;

PICKS

  1. Beff Jezos, the founder of the e/acc movement launched his AI startup focusing on making new-age chips for AI. The company is called Extropic and the founders claim that this new chip will unlock exponential efficiency for generative AI. They did a video with Garry Tan, but we tried to 🍿break it down further in simpler language.

  2. Another day, another model. Cohere's new release, Command-R, is built for serious business, focusing on RAG and "Tool Use" features for automating crazy complex workflows. It beats GPT 3.5 Turbo, Mixtral and is available via API.🍿Our Summary (also below)

  3. OpenAI released a Transformer Debugger - OpenAI’s Superalignment team has been using this internally to analyze transformers for interpretability.

  4. Midjourney just announced the support for consistent characters, so you can now create multiple images with the same person across them. Here’s why that matters and how to use it.

from our sponsor

Need real-time data for your AI project?

The Brave Search API can fuel your AI applications with high quality Web data. 

  • Empower your AI with the ability to search the Web

  • Improve LLM responses with context from extra search snippets

  • Get real-time info like weather and news

  • Autosuggest, Spellcheck, and more…

Get started for FREE (up to 2k calls/month) at brave.com/api. Paid plans as little as $0.50 CPM.

TOP TOOLS

  • VoiceFlow - Automate customer support like Klarna.

  • Agentboard - The easiest way to use an AI agent in the browser.

  • Repo to Text (for LLMs) - Convert your repo to an optimized text file for sending it to LLMs.

  • AICI - Constrain and direct the output of an LLM in real-time.

  • Dola - Sync your calendar and life in seconds on a messaging app.

  • Semiform - Replace your forms with AI agents.

  • Travelicious - Your personalized guide to your next destination.

NEWS

Unclassifieds - short, sponsored links

QUICK BITES

Beff Jezos, the founder of e/acc movement launched his AI startup focusing on making new-age chips for AI. The company is called Extropic and the founders claim that this new chip will unlock exponential efficiency for generative AI.

What is going on here?

Extropic AI is building analog-thermal chips for AI.

What does this mean?

Guillaume Verdon and Trevor McCourt, the founders of Extropic chatted with YC’s president Garry Tan, about what this new chip does. But it’s still filled with words from physics. So, I tried to break it down based on that video, Extropic’s Litepaper and a sprinkle of my own tech understanding.

Training machine learning models include creating random sequences all the time. But, creating and storing those random sequences of numbers is inefficient with digital circuits.

Extropic's approach taps into the 'thermal noise' of electrons for a direct source of randomness. Since electronics already speak analog, this makes things way more efficient for storage as well.

Extropic is building hardware—analog stochastic (random) chips and software to make this work for machine learning.

Take the explanation with a pinch of salt, or maybe electrons.

Why should I care?

You know how everyone's been saying AI progress is going to slam into a wall because chips can't keep getting smaller? Well, Extropic might just be the workaround we need. Extropic claims that their analog approach could squeeze a whole lot more performance and efficiency out of the hardware we already have. If this works... it means AI progress could keep accelerating way longer than we expected.

QUICK BITES

Another day, another model. Cohere's new release, Command-R, is built for serious business, focusing on RAG (think supercharging your internal knowledge for customer service) and a neat "Tool Use" feature for automating crazy complex workflows.

What is going on here?

Cohere releases a new model that beats Open AI’s GPT 3.5

What does this mean?

Command-R is built for serious tasks like retrieval augmented generation (RAG) – think accessing your private data to power smarter applications. It also plays nice with Cohere's other models and external tools, meaning you can move from cool experiments to actual products faster.

Cohere claims it beats the competition, with enterprise users finding Command-R's responses almost 2x better than Mixtral. It also edges out OpenAI's GPT-3.5 Turbo. That's impressive, and the pricing makes it even sweeter: just $0.5/$1.5 for a million input/output tokens.

You can start using it right now on Cohere's API, and it'll be hitting major cloud platforms soon. Plus, Cohere released the model weights for research and non-commercial use. They're not chasing every single commercial use case, but the CEO's open to discussing licensing if it's a fit.

Why should I care?

For most business use cases, LLMs fight in two weight classes. There's the big ring where the heavyweight champs like GPT-4 flex their muscles. Then there's the smaller, scrappier ring where models like GPT-3.5 duke it out—they're not the strongest, but with the right tools, they pack a punch. And importantly, they're a lot more affordable than the big guys.

Command-R is stepping into that smaller ring—same cost, similar size—but it's bringing its own gear. Think of it as the boxer who knows how to use the speed bag, the punching bag, AND can strategize with their coach (that's the RAG and tool use stuff).

Bottom line: if you want to build real AI applications without breaking the bank (basically every enterprise), Command-R just became a serious contender.

Ben’s Bites Insights

We have 2 databases that are updated daily which you can access by sharing Ben’s Bites using the link below;

  • All 10k+ links we’ve covered, easily filterable (1 referral)

  • 6k+ AI company funding rounds from Jan 2022, including investors, amounts, stage etc (3 referrals)

Join the conversation

or to participate.