I’m Max

🃏 Former professional poker player

🎓 Dual degree in Computer Science & Philosophy

🧠 Worked in machine learning & AI research

💻 Went to the Recurse Center, called "Hacker School" at the time

♠️ Built a poker platform

💸 Ran out of money, switched to doing consulting: Monadical

🔧 Build in a broad range of verticals: webdev, academic research, crypto, gaming, and AI

🚀 Just launched an AI investment corp in Q1: DevCap

🎙️ This week launched a new podcast: Paradigmatic

What is this talk?

• Hoping that you come out of this with an intuitive understanding of how AI might affect your business.

​

• I’ll start by looking at a historical case study of how technology can shift an industry.

​

• Then I’ll get give a mildly technical intuition for how modern AI works.

​

• And then finally we’ll look at some examples of new AI applications that are emerging today.

American Airlines Evolved

• SABRE required the creation of new specialized departments, such as data processing, systems analysis, and programming

​

• It led to the centralization of data (the mainframe) and the decentralization of control (localized 8-person crews to many parallel operators)

​

• It led to data-driven reporting & decision-making, and algorithmic pricing

​

• SABRE project teams would involve coordinating with reservations, pricing, marketing, and operations to develop new features, creating a “matrix-like” org structure

Planning a trip remains time-consuming. Why?

“Hard” AI Problems

Reasoning/deduction

Planning

Natural language processing

Perception

Knowledge representation

Learning

Reasoning as Search

Any reasoning problem can be framed as a search problem, given the right representation.

​

(Plan a trip? Win a game of chess? Evolution?)

​

Early AI programs generally used the same classes of algorithms:

​

- Proceeded step by step towards some goal

- If stuck, go back a step and try something different

​

​

​

Combinatorial Explosions

Combinatorial Explosions

• The game-tree complexity of chess is estimated to be at least 10^123, based on an average branching factor of 35 and an average game length of 80

​

• As a comparison, the number of atoms in the observable universe, to which it is often compared, is estimated to be between 10^80 and 10^81

​

• Go is even worse: with a 19x19 board, the game-tree complexity is something like 10^700

Even two very “similar” images can be quite different...

Structure

OR gate

0 1 (x)

(y)

​

1

​

​

​

0

threshold = 1

x

y

x-weight = 1

y-weight = 1

https://cowpig.github.io/netviz/viz.html

“Structure” in latent space

Embeddings

Attention (transformers)

More AI Applications

Copilots

Knowledge Management

Crop monitoring

CV for Quality Control

Traning/Documentation

Translation Services

Personalized Education

Triage/Assessment

Dictation & Notetaking

Accounting, Invoicing

Knowledge Extraction

Data Conversion

Drug Discovery

Automatic medical consultations

Talk therapy

Fraud detection

Automated billing

Schedule optimization

GPTs are Also Quite Stupid

• Hallucinations
• No self-correction
• No mutable long-term memory
• No model of their environment or how it can change
• Almost no logic/reasoning

The scale of the human brain

http://youtu.be/F37kuXObIBU?t=8m35s

​

The human brain has an estimated ~86 billion neurons.

​

In the human brain, each of the hundred billion neurons typically has 1-10 thousand synaptic connections to other neurons. The most recent estimate of the number of synapses in the brain is ~150 trillion.

​

The largest model right now has 1.8 trillion parameters, so we’re still about 100x away from the size of the human brain, but even that isn’t a really sensible comparison because biological neurons are much more complex.

​

Processing estimate (also pretty hand-wavy) sets the brain at ~1-10 thousand petaflops. Next-gen NVIDIA super-computer chips are going to be 20 petaflops!

New AI Applications

Structured information

AI is more like the PC than the internet

AI is at an inflection point

• Major tech companies realized they hit a growth ceiling, started doing mass layoffs
• What do out-of-work engineers do? They tinker, build
• AI has created a brand new greenfield of opportunity
• The aftershock of the pandemic & Ukraine war disrupting supply chains led to an economic slowdown
• Valuations are down, investment capital is