Santa, (A)I’ve been good all year! 🎄

Author

Matthew J. Sánchez
December 22, 2022

“It’s Santa!! I know him!!!” - Buddy the Elf

I’m not Santa, but I come bearing gifts inside today’s issue 🎁 Christmas is on Sunday! On Monday we’ll be back with another issue for you and it will also be the first day of Kwanzaa. What are your Generative AI / Tech predictions for 2023? 🔮 We’re curious, let us know! Let’s jump in…

What do we have for you today?

A primer on Generative AI for founders and investors — the essentials as we gear up for 2023 and what promises to be a fascinating year for Gen AI. The Generative AI revolution of gaming — how the technology will innovate the world of gaming. Lastly, a reminder that robots aren’t human.

Welcome to Gen AI for VCs and Founders 🍎📝

2023 will be hot for investors and generative technology ($2.1B has been invested in 2022 alone). To enter the New Year ready for fundraising and building, here’s a short list of 10 things you should know as an investor interested in Gen AI or as a founder looking to create in Gen AI.

  1. All generative AI have the same foundations. The building blocks are:

    1. Model (e.g. Stable Diffusion, GPT-3, etc.)

    2. Performance Criteria (e.g. “predict the next word” which is done behind the scenes using a method like a Loss Function for instance)

    3. Dataset (e.g. they’re huge and can’t fit on one laptop! Data could be anything from “all documents on the internet” to “all publicly available Gov’t data”)

    4. Optimization Method (i.e. the way to steer and train the model, done using certain techniques such as Gradient Descent as one example)

  2. Expect industry disruption and innovation! There are no business best practices formed yet within the industry, don’t let this alarm you. New business applications, uses cases, models, systems, and dynamics are coming. Common monetization strategies today are: usage/credit based, subscriptions, and ad-supported (e.g. like integrating certain links to be included in ChatGPT when certain terms or queries are inserted by users).

  3. A sample of known Verticals and Use Cases

    1. Media & Entertainment: hyper personalization of content (e.g. images, video, music, video game assets); optimizing streaming bandwidth; custom browsing

    2. Health Tech: MRI/CT scanning and imaging; medical training (e.g. dermatology); toxicity analysis

    3. Finance: financial simulations for risk reduction, volatility, and overall robustness; fraud detection and prevention

    4. Manufacturing: material design; material strengthening and durability; sustainability

    5. Synthetic Data: creating synthetic databases for a variety of modeling needs which addresses the issue of how pricey training models can be

  4. There are few known “Moats” and competitive advantages right now. A few include (from most likely to least likely): attractive branding/marketing, improved UX, excellent data fly-wheels, custom prompts, custom datasets, “better” custom models. We have yet to see how strong each moat is for differentiation. Determine what will be your unique value proposition. Keep in mind with AI becoming mainstream: if you don’t own your query interface, you’re essentially assembling training data for those who do.

  5. Legacy companies can benefit from adopting generative AI to revitalize their existing product lines. Leverage Gen AI to create novel designs and revamp your presence; automate synthesis of customer feedback and sentiment; automate reporting and summaries; simulate user testing and behavior; Customer Service and Documentation/FAQ needs.

  6. Models are expensive to train. Capital will need to be raised in order to sustain computing costs and hire a team for proper execution. Some generative models can be made for as little as $100k, for example if you were interested in generating or synthesizing music or medical documentation. More complex use cases will require richer and larger data and undoubtedly will be > $100k.

  7. You will encounter engineering hurdles and training is messy. Thankfully, more than likely you will be taking existing models and fine-tuning the data or model to fit your needs (mitigating hurdles and adding some simplicity). But if you’re hoping to fully train or start from scratch, take this into consideration.

  8. Training takes time! Training can take as little as one month, one quarter, or as long as +1 years.

  9. Intellectual Property (IP) for Generative AI

    1. You can protect your custom models with patents (usually good for 20 years) if you want it to be novel and unique to your startup

    2. Depending on your Gen AI business, consider licensing your models. This will incentivize collaboration while also monetizing your startup

    3. Note: there’s class-action litigation right now surrounding the training data for Gen AI models (who owns the IP and data that trained the model)

    4. Disclaimer: I am not a lawyer and this should not be taken as legal advice. It’s recommended you consult with an attorney for legal counsel.

  10. Ask if your generative AI is equitable and mitigates negative impacts. For example, correct your models if they are perpetuating biases, exploiting or manipulating individuals, creating hateful language or images, and so forth.

Thank you to Andrew Carr for recently uploading a series on YouTube aptly named “Generative AI for Investors and Founders.”

A Reminder That Generative AI Aren’t Human

It sounds silly but this was a timely reminder for me. As I have fun in ChatGPT and GoCharlie I have a tendency to write my prompts in kind, well-mannered ways. “Please…” “Could you…” “Thank you so much.” Basically as if I’m talking to another person.

With certain prompts the models might respond in ways that feel human-like. Giving me the illusion of conviction or character. Using words like “I believe”, “I know”, “I think”, bordering anthropomorphism and having people, like me, see or treat them as more human than they are.

Shared in his academic paper Talking About Large Language Models (Dec 2022 v2), Professor Murray Shanahan of Imperial College London reminds us that Gen AI are merely LLMs (models) and systems of code. He clarifies in this tweet that we should practice caution to not fall into the understanding that there’s conscience in them. Unlike humans, when we see AI use words like “believe”, “think”, and “know”, it’s to be taken figuratively not literally.

This will be important especially for younger generations coming up who will directly be immersed in a Gen AI world. Dr. Shanahan adds that this is not to say what’s capable in the future à la I, Robot.

Gaming is Going Generative 🎮👾

AI is about to supercharge the creativity and productivity of game industry workers. From Indie developers, to studios, to game designers, generative AI is ushering in a gaming transformation.

What was first sparked by this Twitter post in November, a16z dug deeper and created this write up highlighting what is expected to come. Artists will be able to create high quality content and renderings that used to take weeks by hand in a matter of days! Not only does this Gen AI tech democratize game development, it also beats the old “Choose 2” paradigm of cost, quality, and speed. Finally, we can have our cake and eat it too 🎂

Because games are highly iterative and interactive, Gen AI is capable of disrupting the gaming status quo and opening the door for newer game developers to enter the ring. With speed, cost, and quality no longer being a major barrier to entry, it will be no surprise if we see more game releases per year and more indie (”micro”) game studios formed in 2023 and beyond.

The opportunities are abundant (like building an all-in-one suite for gaming similar to how Runway ML is to video) and the Gen AI Gaming Market is barely taking shape. Heck, only a couple days ago OpenAI released Point-E, a 3D model generator which can be used for gaming down the line! It’s a wonderful time to be into games.

Gen AI Deals that make your eyes (and mouth) water 💰

What caught our eye? 👀

The AI-powered photo editor addresses the pain point of post-production for photographers acknowledging that it’s repetitive and time-consuming. Imagen AI theoretically improves with each use as it learns your personal editing style. Imagen AI also offers “Talent AI Profiles” which are pre-trained profiles based on renowned industry photographers and their unique editing styles.

The icing on the cake? They’re already profitable. Imagen is raking in $10M ARR according to Imagen Co-Founder & CEO, Yotam Gil.

New developments to spam your #random Slack channel 💬

Things to learn when you need a raise

Want access to the Generative AI investors database? Share AI In the Middle (below) to receive it! ⬇️

Before you go

A classic Christmas movie gets the Gen AI treatment. Ever thought “what if Home Alone was made in the styling of the opening scenes of White Lotus S2?” Yeah, me neither. Until now. Look how mind-blowing Runway is!

That's it for today folks! Merry Christmas & Happy Holidays 🎄

— Matthew J. Sánchez (@matthewjsanchez)