😺 AGI

PLUS: This AI ate 100,000 pages of research.

Noah Edelman
September 19, 2023

Welcome, humans.

Just told Siri it might need a hearing check, and it clapped back with, “if you like ChatGPT so much, why don’t you just date OpenAI.”

Then I said “why are you yelling?” and it locked me out of my phone… What gives?!

Here’s what you need to know about AI today:

Experts are split on the potential of AGI.
Morgan Stanley trained a chatbot on all its internal documents.
OpenAI is racing to launch multimodal search before Google.
40% of big Q2 early-stage investments were in AI.

AGI: Overrated OR Underrated?!

Human Progress, from WaitButWhy

The builders of a new technology think it will either create a utopia or destroy the world. No biggie.

Enter AGI — Artificial General Intelligence.

Current AI has specialized intelligence, meaning it’s great at singular tasks like copywriting.
AGI has general intelligence, meaning it can do just about anything.

Like anything, anything: picture a computer that can learn, think, reason, decide, refine, and take action... and it doesn't need a babysitter.

The fear is that AGI could go rogue and we’d be utterly helpless. Think HAL 9000 from Space Odyssey, but even worse.

It’s getting so real that AI vets like Geoffrey Hinton and Sam Altman signed a letter warning “of extinction from AI on scale with nuclear war”, and others like Elon Musk signed a petition to put AI development in a timeout.

Pause. Take a deep breath. Okay. Proceed.

It’s not all gloom-and-doom. Nvidia reassured the Senate that “uncontrollable AGI … is science fiction”, and most of our tech-y friends agree. Many are even betting that a superintelligent AI could usher in a new era of productivity and affluence.

Here’s our take: We’re cautious optimists. AI accelerates exponentially, but the baby AGI we see today (aka autonomous agents) are quite clunky.

Either way, AI should be taken very seriously, and the Senate’s AI summit is a step in the right direction.

As for you all? No need to worry about AGI yet. Go touch grass, have a beer, and shoot a sweet message to your parents. :)

FROM OUR PARTNERS

Use Llama 2 At Top Speed With OctoML.

Building in AI is tricky with things moving so quickly. Here are two strategies we’ve noticed from top-tier engineers:

Fine-tuning open-source models like Llama 2 over splurging for GPT-3.5/4 tokens.
Using OctoML to run their AI models.

OctoAI is THE fastest place for devs to launch, tune, and scale AI models for their business. They've just rolled out a suite of turbo-charged Llama 2 flavors with no compromise on accuracy.

If Llama 2 is in your roadmap, use OctoML.

Bonus: they’re treating Neuron readers to free GPU hours to put their herd to the test.

Morgan Stanley Trained A Chatbot On All Its Knowledge.

Morgan Stanley’s just released an internal AI assistant trained on ~100,000 of its research reports and documents.

The result: an AMA on everything finance-related, from market trends to legal processes. Huge. Time. Saver.

McKinsey pulled a similar move last month, arming employees with a chatbot for anything consulting-related: research, data, past cases, & more.

The whole train-a-chatbot-on-my-company-knowledge is something we expect a lot of companies to do soon.

The best option for that, for now, is using ChatGPT-3.5 turbo, assuming your data is as squeaky clean as Charmin Ultra Soft (here's a guide for finetuning).

If you just want a chatbot that can “talk” to a few PDFs at a time, Chatbase should get the job done.

Around the Horn.

Microsoft’s AI research team accidentally leaked 38 terabytes of private data. Major oops.
OpenAI is hustling to launch multimodal search (chatbot that works with images AND text) before Google does.
40% of the biggest early-stage VC investments were into AI.
ChatGPT usage is inching back up … just like the number of students returning to school.