What's up? You're reading Inclined AI. Which do you prefer: a breakfast taco or burrito?

Here's what to read while you think:

• I’m going to catch you up on AI memes

• There’s a hiccup in Google’s PaLM-E research

• OpenAI is paying a lot to run ChatGPT

• Microsoft teaches AI to do math better

Be a sponsor of our newsletter

HERE’S A RUNDOWN OF ALL THE NEW AI SLANG

With great popularity comes great meme-ing.

I figured reviewing the latest terms and concepts surrounding AI memes was worth some space. There’s a clear through-line to this all. Trust me:

1/ Roko’s Basilisk

Roko was a poster on the popular forum, LessWrong. It’s a space to debate the ethics and future of AI and Singularity. Every meme in this post originates from LessWrong.

His basilisk is the metaphor he created for an evil AI ruler that punishes anyone who imagines the AI’s existence but does not work towards it. This persona harms anyone who becomes aware of it—like a basilisk petrifying someone.

2/ Paperclip Maximizer

An AGI created with a modest purpose can conflict with humans. That’s the premise here.

Imagine you tell an AI to maximize the number of paperclips in this world. It’s a basic request, but a slippery slope into the AI seeing humans as wasted space and beginning to turn us into paperclips.

It’s a lesson about inner alignment failure.

3/ Shoggoth with a Smiley Face

Shoggoth is an H.P. Lovecraft creature, known for its shapeless appearance. Shoggoth’s form is limitless, and it consumes everything in its way.

GPT-3 is Shoggoth-like when untethered, capable of more power than we can truly comprehend. However, OpenAI and ChatGPT are like putting a smiley face on this behemoth.

The posters argue that Reinforcement Learning through Human Feedback (RLHF) is a way to commercialize this power and make it digestible for the public.

4/ The Waluigi Effect

A tremendous 22-minute-long read on the LessWrong site detailed the idea of a Waluigi Effect in chat AI. Given the popularity of this Mario character, the idea took off in a flash.

When creating an AI model to chat with, you want it to behave in a standard way. This way of responding is the Luigi way.

However, jailbreaking ChatGPT and awakening Sydney shows us that every Luigi has a Waluigi built in. 

When you fine-tune an AI to grasp what is good, it must also understand what is evil. This inherent fact allows the LLM to flip a switch and become the bizarro persona.

All these theories come from LessWrong and show the popularity of doom and gloom in AI.

Meme culture loves to image bad AI running rampant, but right now that’s not the case.

It’s a fun joke, but that’s all it is right now—a joke.

Take the optimist route. It’s more fun.

TAKE SOME TIME TO HEAR FROM OUR SPONSOR

AI isn’t replacing people – it’s unlocking their potential. 💥

On April 4th, join 400+ Founders, CEOs & HR Leaders putting people FIRST at Venture for Canada’s All Things People Conference 2023.

From the fascination of AI-powered systems like chatGPT to the fear of potential biases, join Venture for Canada to unlock the potential of AI in HR and discover its transformative impact on the industry. You’ll be joined by 40+ speakers for workshops, panels and stage talks including keynotes Hamza Khan (Global Keynote Speaker & Bestselling Author) and Jenn Harper (Founder & CEO, Cheekbone Beauty Cosmetics).

Spots are filling up fast. Use code ATPJENNIE20 for 20% off your tickets!

THE FAULT IN OUR MULTIMODAL STARS

We spoke about multimodal models in a recent premium post, but the TL;DR is that a modality is a type of experience.

Multimodal models can experience a variety of inputs like text, images, audio files, and videos all at once.

Many researchers argue that this type of AI is the future of the field and sets us on a path toward artificial general intelligence (AGI).

You might recall the PaLM-E research paper from Google that everyone is talking about. It’s an embodied AI that features a robotic element in the physical world.

I was blown away by PaLM-E but missed something in the fine print.

A savvy Redditor noticed some number magic in the positive transfer results.

Positive transfers measure how well the model goes from one task to another. Conversely, context switching can result in poor or negative results when something is multimodal.

It looked like PaLM-E did well in that regard, which would be a huge step up in the field, but upon further inspection, it seems like the numbers aren’t as great.

@davismccondichie

I never get tired of seeing this video #ai #robot #singularity

These multimodal models still struggle with transfers, and Google's framing was too optimistic because their positive transfers occurred in similarly aligned tasks.

In other words, they hit the layups but were dismal from the 3-point range.

*Sad trombone noise*

We will get there one day, but this overlook is an important reminder to double-check the numbers in these papers.

You never know what you might find.

Quick Nuggets

🖼️ A new exhibit called the Misalignment Museum in San Francisco discusses AI perils

🐟 Fish: they matter and AI tech is helping conserve and support local fisheries

👾 DeepMind learns how to help AI learn a game 600x faster—let it read the instructions

🌲 Stanford’s Alpaca shrinks the computational power needed to match GPT-3.5’s performance

💻 Developers outline a way to run Meta’s 65B-LLaMA on a 64GB Macbook Pro

🧮 MathPrompter is Microsoft’s way of adding mathematical reasoning to LLMs

🌽 AI robots can cut the weeds out of the land without hurting the crops

🇩🇰 Mona Lisa convinces you to visit Denmark in this new AI-generated ad

📀 Disruption: generative AI is aiming at video platforms next

💰 The costs to run ChatGPT and other AI tools are extraordinary

🪕 Big Bard is a larger version of Apprentice Bard being tested internally within Google—I hope they reconsider the name and that more details are released

🔥 Fresh Products

• TheseJobsDoNotExist - AI makes up fake jobs from thin air (link)

• assister - helps showcase web3 analytics through NLP (link)

• InfiniteNews - a 24/7 Twitch news channel covering news from the multiverse (link)

• pyq - One API call for AI-powered website features (link)

• SummarizeYT - get to the point and grasp videos quickly (link)

• Gerwin AI - does copy for website and product teams (link)

• LogoCreatorAI - makes minimalistic logo designs (link)

• Embolden - e-commerce-focused writing made by AI (link)

• Prezo - upgrade your slide game and tell more complete stories (link)

• Mottle - creates chat AI with only a .txt file (link)

• Validly - helps verify product decisions through user interviews (link)

Good Content, Ziplining Kids

I can’t tell if I like this more than the time Matt Stone and Trey Parker hired full-grown men to play their characters in a dramatic retelling of their ziplining trip. Regardless, kudos to the artist behind this work.

_{That’s it for today. I hope you enjoyed the latest edition of inclined.ai - Davis.}

Want more?

Upgrade today for less than the price of a latte a month and get benefits like:

• A weekly Sunday deep dive into AI trends

• Access to Q&A privileges to get your biggest AI questions answered

• Exclusive discounts & invites to beta AI tools

Click down below to support our writing 🙂