• The Neuron
  • Posts
  • 🐈 How a safer ChatGPT actually happens

🐈 How a safer ChatGPT actually happens

PLUS: Google responds, Boston Dynamics is back again

Good morning! This is The Neuron. We're the "fast casual" of AI, guac included.

Today in AI:

  • Behind a Safe ChatGPT Are Kenyan Workers

  • Google Responds to Microsoft

  • Watch: A New Video From Boston Dynamics

  • Around the Horn

  • Leo Sends His Regards

Behind a Safe ChatGPT Are Kenyan Workers

This one's uncomfortable.

Yesterday, TIME reported that OpenAI paid Kenyan workers to do the dirty work of making ChatGPT safe.

The idea: For ChatGPT to be "safe", you need to teach it what "safe" is. It's a computer - it doesn't really understand what anyone is saying.

To do this, you find a bunch of the bad stuff, label it with "this is bad stuff, and it's this specific category of bad", then feed that into ChatGPT.

How OpenAI did it:

  • They scraped content about topics they didn't want ChatGPT talking about (e.g., murder, self-harm, sexual abuse, incest)

  • They sent it to Sama AI along with a $12/hr contract to get that content labeled

  • Sama AI paid Kenyan workers $2/hr to read and label 150-250 passages each nine-hour shift.

Pay is a common first debate: Was this a well-paying job (Kenya avg. being ~$1.25/hr) or an exploitative one? Where you stand will depend on your values and frame of reference.

For that, the Hacker News discussion thread is one of the few places on the Internet where you can hear from Kenyans and non-Kenyans alike. It's worth a browse.

The bigger discussion: This is the reality of today's models. They all need this approach to be safe. Plus, as ChatGPT-like tools get more popular, the more people will try to make it say bad things, the more you have to clamp down.

Someone has to do that psychologically brutal work. Is this how things are going to work? And is the ultimate payoff worth it? It's worth thinking about, especially if the answer is "yes".

Power and responsibility. AI is getting better. That's exciting. We also need to acknowledge how we keep it safe - both in concept and how it actually happens.

Google Responds to Microsoft

Here we go.

In the rap battle that is Big Tech AI, Microsoft spit a crazy first 8 bars by declaring they're putting OpenAI everywhere.

It's Google's turn to take the mic, and they've showed up with *check notes* a 7,000-word blog post titled, "Google Research, 2022 & Beyond: Language, Vision and Generative Models".

It's the first of a series, which will eventually cover robotics, health and quantum along with today's hot topics in language and generative models.

This is Google saying, "Let's not forget that we're absolutely stacked in the R&D department. We've got models."

  • Multiple language models that are bigger and better than yours

  • Multiple models that can generate images from text

  • A model that is getting real good at doing math and science (Oh, your little ChatGPT can't add two numbers? Cute.)

  • A model that can handle text, images *and* over 100 languages

  • A model that can generate video from text

It's not a flashy announcement, but it's a strong pitch to prospective research talent and a solid reminder of just how much firepower Google is wielding.

Microsoft fanboys: you think this battle is over? It's only getting started.

Watch: A New Video From Boston Dynamics

The robot overlords are back.

Every 6 months, Boston Dynamics humbles us with a robot that's way too good for comfort.

Anyways, they're right on time (click to play):

Around the Horn

  • Great demo from Vowel.com of their AI-generated meeting summaries.

  • Writer: AI copy generator as a Figma plugin.

  • Millie: Your AI dating assistant.

  • MagicForm: Replace Intercom with ChatGPT. Product aside, he voiceover is so hyped up!

  • Sam Altman's full interview with StrictlyVC: "The GPT-4 rumor mill is a ridiculous thing"

  • Wharton innovation professor leans in hard with ChatGPT in his classroom.

  • New study: People generally found ChatGPT answers more helpful than human answers.

  • Reid Hoffman and Elad Gil talk AI. Video, notes.

DM me links on Twitter: @nonmayorpete

Are you new to all this AI stuff? Here's The 3-Minute Guide to Slaying Your Dinner Convo About AI to get you up to speed. Or at least smart enough to impress your family.

Leo Sends His Regards

That's all we have for today. See you cool cats on Twitter if you're there: @nonmayorpete

What'd you think of today's email?

Login or Subscribe to participate in polls.