Goooooood morning! This is The Neuron, where just like Matthew McConaughey, we chase our future selves. Alright, alright, alright.
Today in AI:
- Is It Time for Google to Panic?
- GPT-3 Takes the Bar
- Around the Horn
- Leo Sends His Regards
Is It Time for Google to Panic?
ChatGPT blows people's minds → people start whispering about Google.
Here's the conversation so far:
- ChatGPT gives a more streamlined experience. Ask a question, get an answer vs. type 5 keywords and scroll through a bunch of links to find the right site.
- But, ChatGPT gets stuff wrong and would be too expensive. Plus, Google's PaLM is better than GPT 3. Today's ChatGPT doesn't come close to a threat.
- Still, that doesn't mean it can't happen.
Yup, that's enough to send Twitter spinning.
Some people think it's just a matter of time before ChatGPT takes over. Others think the reports of Google's death are greatly exaggerated.
Our take: In the short-term, Google isn't going anywhere. But it can't deny that this new technology is powerful. It'll have to figure out a few things:
- What searches is the ChatGPT way actually better for? Google's way better for "nba scores" but maybe it's better at "how to set up a vertical farm at home".
- Can you make it fresh, fast and cheap? Google discovers hundreds of trillions of webpages a year. Search results in an instant. Is it even possible for ChatGPT to get to the same level without breaking the bank?
- Can you continue to print that cold hard cash? It's risky to touch the golden goose. They'd only take a page from ChatGPT if the numbers work. #4 most profitable public company doesn't come easy, you know.
- (most importantly) Is Google the right company to face this challenge? Or is the 300,000 employee staff too big to stave off a startup competitor?
CEO Sundar Pichai just got a new $210 million stock package. Time to see if that's too much or not enough.
GPT-3 Takes the Bar
GPT-3 lost a bet and had to take...the bar exam.
Every year, 60,000 bright-eyed, debt-carrying law school grads set their sights on the final boss: the bar.
It's a grueling task - essay questions, a mock case and a 200-question multiple-choice test over 2 days.
To prepare, you study 400 hours just for the bar exam alone. About 60% pass.
Well, two researchers dragged a factory-setting GPT-3 model out of bed and shoved it in a room with a proctor and two pencils to take a part of it.
How'd it do? Better than we could, but at 50% accuracy, it didn't meet the usual 60-65% minimum needed to pass.
Good at Civil Procedure, Evidence and Torts, especially bad at Criminal Law and Procedure. Let's maybe keep GPT-3 away from criminal cases for a bit.
Fun side note: They tried to make GPT-3 study before taking the test and it actually did worse.
Even though, GPT-3 couldn't pass muster, we think its bigger bro GPT-4 is up for the challenge. We'll report back when they try again 🫡
Around the Horn
- Need a new listen for the commute? Check out Shawn Wang's best new AI podcasts of 2022. Heads up: these go pretty technical and deep.
- DALL-E 2 co-creator predicts car prices using the images alone. Just a side project, by the way.
- A charity uses Midjourney to create ads depicting poverty. Controversy ensues.
- Turns out research paper titles become beautiful pictures with Midjourney's help.
Leo Sends His Regards
