Genuinely Artificial
Posts
Anthropic gets billions in investment and finds alternative to RLHF

Anthropic gets billions in investment and finds alternative to RLHF

LLM companies continue to battle for market dominance.

Chris Dedow
October 30, 2023

Read time: 4.1 minutes

Welcome to Genuinely Artificial!

This newsletter will augment your reality with AI insights, automations, and antics.

I’ll be operating as your own personal researcher, data scientist, data analyst, and tech journalist all rolled up into one.

With your help, we should be able to reach 100% worldwide subscribership within 34 newsletters if we double each time. Seems doable.

It’s the 2nd newsletter release and I’m happy to say we’ve officially doubled our subscriber base reaching one step closer to worldwide subscribership. 🤝

1 × 2 = 2 subscribers achieved (currently pacing ahead of this 😁)

2 subscribers × 2³³ to go. 🤷‍♂️ 2 ³⁴ ~ 8.6 billion population 🌎️

Anthropic gets a massive investment from Google and also announces an alternative to reinforcement learning with human feedback. Spotify open-sources one of its libraries.

Let’s get into it with…

TODAY’S DOWNLOAD

Anthropic gets investment from Google
Constitutional AI set to improve human feedback training
Waymo partners with Uber
Spotify open-sources library of how they know your taste in music so well

AI NEWS & INSIGHTS

INDUSTRY NEWS

Google follows Microsoft’s playbook and invests $2 billion in Anthropic

Google agrees to invest up to $2 billion in Anthropic as the companies leading the charge with large language models continue to stay red hot.

the sitch:

Google follows Microsoft’s playbook by also signing a $3 billion multi-year deal for Anthropic to use its cloud platform
Both giants are managing to essentially pay themselves while getting exposure to two of the hottest investments currently available which directly impact their larger business models at scale.

INSIGHTS

Human feedback for model training swapped out for AI-conditioned feedback

Reinforcement Learning with Human Feedback may not be enough to curb problematic behaviors in conversational AI. Anthropic is taking an alternative approach through the creation of an AI model that would provide feedback based on only written principles like “do what’s best for humanity.”

Our thoughts:

The paper genuinely suggests “both general and specific feedback principles have value for steering AI safely.”
Initial principles seem generally agreeable but since these artificial intelligence models are controlled by a select few does that have them playing Moses?

AUTOMATIONS & ANTICS

AUTOMATION ALLEY

Launch your morning go-to websites with Python

Do you ever wake up having several news sources, social sites, and machine learning research papers to visit before you even start work?

Your time is precious and Python is here to protect it.

Get a list of your favorite websites
insert them into the code below,
hit-run
watch your favorite browser come to life with your list of favorite websites.

I check 20 websites every morning.

How many are you going to check?

# python package to open web pages on 
# default web browser
import webbrowser

urls = {'website 1: 'www.{insert_site_name}.com',
    'website 2: 'www.{insert_site_name}.com',
    'website 3: 'www.{insert_site_name}.com'
}

for company, url in urls.items():
    # open links in browser
    webbrowser.open(url)

ANTICS AVENUE

Waymo partners with Uber but who’s going to buy Lyft?

Google has loved not picking sides by giving money to both Uber and Lyft over the years in investment wanting to be “right” with whoever reached market dominance.

Well, as of today Uber has a market capitalization of ~$84B while Lyft is valued at <$4B.

I think the war has been won.

The unanswered question: Who will buy Lyft to gain access to its 20 million+ users to leap forward the accessibility of autonomous vehicle taxi rides?

TOOLS, TUTORIALS, & TESTIMONIALS

TOOLS & TUTORIALS SPOTLIGHT

Spotify improves how well it already knows you

image source: Spotify R&D Engineering

Searching in a galaxy far far away Spotify has come back to us with a new nearest-neighbor library that helps us find items based on their similarities. Playlists have never sounded more uniquely you! 🎵

Spotify engineers have been using it internally for almost a year with some of their products to focus on improving four major areas:

flexibility
statelessness
language support
cost

The Voyager library has now been open-sourced to be accessible to all with popular Python and Java options.

To try out Voyager, check it out at spotify.github.io/voyager or

pip install voyager in Python to get started.

FROM THE READERS

No feedback yet but I sense it’s coming soon. The creativity is starting to brew. 👨‍💻

EVENTS & EPILOGUE

UPCOMING EVENTS & OPPORTUNITIES

10/30 Apple Event 8 pm EST (link)
10/31 Google: Open LLMs on GKE - Llama 2 and Beyond (link)
11/2 Google: Gen AI 100: Getting Started with Gen AI (link)

QUICK LINKS & RESOURCES

ML Papers of the Week (link)
Langchain: chat with your data, 1 hour course (link)

FEEDBACK WELCOME

How can I make the Genuinely Artificial newsletter more entertaining, educational, and useful for YOU?

I’ll attempt to read your minds but most of my prediction capabilities are still centered around churn models. 🫣