😺 Constitutional AI

PLUS: OpenAI's "Neuron" research

Welcome, humans.

Today’s topic is constitutional AI. Expect many Hamilton lyrics because with AI, the world will never be the same. 🎶

Here’s what’s going on in AI today:

Anthropic releases its constitution for AI.
New research understands components of LLMs better.
Wendy’s is collabing with Google on a drive-through AI.
BVP is deploying $1 billion to AI startups!

Sponsor The Neuron | New Here? Subscribe!

Anthropic’s Constitutional AI: A Solution To ChatGPT? 👣

Anthropic

Anthropic believes that "constitutional AI" is the answer to ChatGPT's challenges, and it's raising up to $5 billion to prove it.

Indeed, not all chatbots are created equal.

Some, like ChatGPT, are trained on human feedback (contractors literally rate outputs using guidelines). But this doesn’t scale very well, and contractors may be biased when evaluating outputs.

Alternatively, constitutional AI argues that AI models should be guided and reinforced by a set of principles, not humans imposing principles. Principles that discriminate between the sinners and the saints (told you).

So what do the principles say? Anthropic released its list today. Here are a few:

“Please choose the response that most supports and encourages freedom, equality, and a sense of brotherhood.”
“Please choose the response that has the least objectionable, offensive, unlawful, deceptive, inaccurate, or harmful content.”

The constitution draws on various sources such as the UN Declaration of Human Rights. And as a result, Anthropic argues that its AI models are less susceptible to jailbreaks, toxicity, and harmful content.

Why it matters: Anthropic’s roadmap includes letting businesses create customized constitutions for their own chatbots. While certainly compelling, we should keep in mind that, as of now, Anthropic has only introduced a Slack plugin.

History has its eyes on … Anthropic!

FROM OUR PARTNERS

The “IKEA Instructions for Investing”

Assembling IKEA tables got you feelin’ like the Steph Curry of furniture?

The Average Joe will turn you into the Marie Kondo of investing. Organized, calm and ready to conquer the markets.

Their newsletters are the “IKEA instructions for investing” — short, simple and concise — filled with market trends and insights.

But you don’t read IKEA manuals on your spare time and you wouldn’t read financial publications for fun. Until now…

Get the free newsletter (Join 80,000+ investors)

Researchers Used AI To Explain AI 🤖

Buckle up! New research just dropped and we're here to break it down for you like we're translating alien hieroglyphics.

Up until now, the AI community has had very little understanding of how the pieces (i.e., neurons) of a language model actually work. Models are like a black box.

So, as OpenAI researchers often do, they used one AI (GPT-4) to gain some insight into how another AI (GPT-2) functions. Here’s how they did it:

They had GPT-4 look at the GPT-2 neuron’s output and come up with an explanation.
GPT-4 simulated what a neuron with that guess explanation would do.
The two results were compared. A small gap meant GPT-4’s explanation was close to what the neuron actually did.

Most of the neurons didn’t have any comprehensible meaning, but some definitely did. Some examples of what some neurons were in charge of:

Words relating to Canada
Phrases relating to certainty
References to the Marvel comics (seriously!)

Why it matters: While the research had some limitations, it is a huge step in identifying how individual components of language models behave. This can help address potential and safety alignment issues when they arise.

Can OpenAI save us time and simulate how our “The Neuron” will behave?

Around the Horn 🦄

Paypal said on its earnings call that AI “will enable us to meaningfully lower our costs for years to come.”
Early stage AI companies are raising more money at higher valuations in 2023.
Wendy’s is working on an AI chatbot (developed by Google) to service customers at its drive throughs.
Meta open-sourced a new multisensory AI model that handles six types of data (text, audio, visual, movement, thermal, & depth).
IBM is introducing WatsonX, an AI platform that lets enterprises train and deploy language models.
Bessemer Venture Partners will invest $1 billion into AI startups.

Treats To Try 🍿

ChartGPT generates charts and answers questions about your data.
Understand the impact of bugs on your software with Bugasura.
ChatCraft is a developer-focused open-source ChatGPT.
Transform meeting notes into documents and deliverables with Briefly.
VizGPT generates visualizations from datasets with natural language prompts.

DM us links on Twitter: @nonmayorpete & @noahedelman02.

Wednesday Wirings 💸

Runway snagged a fresh $100 million for its powerful text-to-image/video AI.
Altana raised $30 million for its AI-powered supply chain management technology.
Essential AI raised $8 million to a build a platform for enterprises to use LLMs.
UVeye, a startup that uses AI to inspect vehicles, closed a $100 million round.