Before we get started, two PSAs:
First, next Monday the new AI Academy courses will be out. This means that your opportunity to save up to 70€ with the pre-launch sale ends on Sunday. You may have noticed AI is exploding, and this is your chance to join the party. Don’t let it go, join now.
Also: Next Tuesday I’ll be hosting a workshop titled “Democratising access to AI education - why and how” for the Mozilla Festival.
I’ll guide you through an exercise to understand how you can kickstart your AI education and help others join the revolution too. Join me at MozFest! Get your ticket here.
This week’s news:
GPT-3 was released almost 3 years ago. Barely anyone cared.
GPT-4 was released two days ago. Everyone knows this will change everything.
Today I want to help you navigate what’s going on covering three aspects:
The fundamental difference with GPT-4 is that it’s multimodal, which is a fancy word to say that it’s not limited anymore to text, but it can “understand” images too.
It doesn’t sound that impactful until you see some applications. Here are a few:
Be My Eyes has built a product to help people who are blind or have low vision. They can now take a picture of their fridge and ask GPT-4 “what can I cook with this?”, GPT-4 will understand the items in the fridge and develop a recipe.
This demo blew everyone’s mind though: you draw a mockup of a website on a piece of paper, send it to GPT-4, and ask it to turn it into an actual working website. It’ll do it.
Greg Brockman (@gdb) of OpenAI just demoed GPT-4 creating a working website from an image of a sketch from his notebook.— Mckay Wrigley (@mckaywrigley) March 14, 2023
It’s the coolest thing I’ve *ever* seen in tech.
If you extrapolate from that demo, the possibilities are endless.
A glimpse into the future of computing. pic.twitter.com/1QB6wbQkld
The other differences are more incremental than transformational. In short, GPT-4 is smarter and safer than GPT-3.
Normally, I wouldn’t say that “incremental improvement” is something to be extremely excited about. The first iPhone was a transformational innovation, but the iPhone 14 is a small incremental improvement over the 13 that few people care about.
However, if technology improves by an increment large enough to surpass some threshold for adoption, things can get really exciting because some applications may go from “impossible” to “possible”.
So how much smarter is GPT-4? The most outstanding example is the US Uniform Bar exam: a series of questions to test the knowledge that every lawyer should be able to demonstrate prior to becoming licensed to practice law.
ChatGPT was in the bottom 10% of humans taking the test. GPT-4 is in the top 10%.
In the International Biology Olympiad (a biological olympiad for pre-university students under the age of 20), ChatGPT was better than 31% of participants. GPT-4 is in the top 1%.
So yeah, it’s much better. Is it better enough to enable it to have reached a tipping point?
It definitely feels like it, especially because it’s made great improvements also in safety (less likely to say BS) and “steerability” (how well it follows prompts).
My bet here is that we did reach a tipping point IF this “steerability” is given in the hands of subject matter experts. In simpler terms: GPT-4 alone may not be enough to replace a lawyer. But we are in a moment where a lawyer could build around it through prompt engineering (+ other resources) to confidently let it run part of his job.
And by the way, I used “lawyer” as an example, but you can replace it with many professions today.
The attention around the release of GPT-4 is 1 million times higher than what I saw when GPT-3 came out. Yet, GPT-4 is not 1 million times better. Why did we all wake up now?
I think ChatGPT was the key. While not smarter than GPT-3 (it’s actually dumber), ChatGPT was easier to understand and use. GPT-3 had a very odd pitch that I failed to understand too: give it text and it’ll continue writing text. So what?
ChatGPT was much more user-friendly: talk to it like you’d do with a person, and it’ll answer you. A 5-year-old kid could understand it and use it.
ChatGPT’s main role was to show humanity what this technology can do and inspire us on what we could build with it. From that point of view, we could say that ChatGPT’s biggest marvel and contribution was not the technology itself, but the user experience and marketing outcomes.
I think this underlines a very important point about generative AI that too many people are ignoring: people don’t care about the technology. People care about the problems technology solves them.
While GPT-4 is amazing, it’s still entirely useless unless companies include it within compelling, delightful products that solve real problems. It’s a tool.
You’ll see in the Pizza Bytes section of this newsletter how many companies released products that integrate the “GPT-X” experience (spoiler: a lot).
What impresses me is the speed at which everyone is integrating this experience within their products. To me there are a few questions that still need answer:
I could write a giant essay about it, so I’ll keep it short for now and you reply to this email if you want me to go deeper on one of those topics in the future. In short:
GPT-4 is freaking amazing, but it’s not the product that will change our world.
The products that will change our world are the ones that entrepreneurs will build on top of it.
On that note…I’m working on a couple, stay tuned 👀
Other quick news.
The sweet part of the newsletter: fun stuff from the crazy world of tech.
Get ready for this:
I’ll talk to you next week.
This post has been sponsored by the Mozilla Foundation.
I totally recommend it, great info and also great fun! As you can deduct from above, I declare myself a fan🙂 happy to read this condensed newsletter about tech!
Cultural Change Director - E.ON
I feel more empowered intellectually 🙂
Business development - FNA SPA
Your 10 min shortcut to understand how tech impacts society and business, available in audio, video or text format
Llama, a version of GPT-3 open-sourced by Meta. An algorithm like this has potentially destructive outcomes for society...
Meta built a better copy of OpenAI’s models and gave it out for free. It beats GPT-3 (OpenAI’s flagship model, the basis for ChatGPT)...