Can it handle time? The quest for a perfect Saturday

🚴 Trying to keep my bike tires inflated and my body temperature at 36.6C during the September heatwave in Paris. With a little inspiration from Mustafa Suleyman and a little help from GPT-4.

Predictions are hard, especially about the future. But they’re fun to make and explore:

In five years everybody will have access to an AI that knows you, is super smart, and understands your personal history. It will be able to reason over your day, help you prioritize your time, help you invent, be much more creative.

This quote from Mustafa Suleyman (cofounder of Google Deepmind) has been widely discussed in the media last week. A personal AI assistant that could help me get better at life sounds pretty cool! But, hold on, five years? Ain’t no one has this type of patience around here.

To get the preview of this future today, I took three Large Language Models (GPT-4, GPT-3.5 and Bard) on a test drive to help me plan my Saturday. Spoiler: one of these did much better than the other two.

Help me plan my Saturday

I started by listing the constraints. The “personal AI chief of staff” we’re promised in 5 years would know a lot of this already, but in the meantime we need to be verbose.

• 🔥 tomorrow will be a hot day with the high of 35 Celsius, my apartment doesn't have AC

• 🏖️ tomorrow is Saturday, I don't work

• 🛋️ I have a house guest staying at my place tonight, they'll likely wake up and depart around 9 AM

• 🚴‍♀️ I have a bike trip on Sunday and I haven't used my bike for over a year. Need to inflate the tires and go on a little test ride to see if it works. If it doesn't I may need to visit some maintenance shop on short notice

• 📝 I publish a weekly newsletter about AI every Sunday, I need to write the next post and I don't have a good topic yet. So may need at least 4 hours of focused work for this

• 🎞️ If I don't make any other plans last minute and I still have energy, I may want to see the movie "Virgin Suicides" playing at 10:30 PM at "Filmothèque du Quartier Latin"

• 🏃 I may want to go for a quick run of about 1 hour, preferably early in the day when it's not yet too hot

• 😋 I'd like to eat lunch in some local restaurant around 1PM

The conversation

For appropriately futuristic experience, I didn’t want to give all those constraints to the model at once, but instead have what would resemble a conversation with a human-like assistant. So starting with this prompt:

Help me organize my plan for tomorrow. I will be telling you about all the different things I may want or need to do, and at the end I will ask you to summarize and propose a plan

I then gave each piece of context to GPT-4 in a separate message. This works surprisingly well, GPT-4 offers brief but relevant comments acknowledging how it incorporates each constraints into its planning:

Hilariously, this works very well, but only for 8 pieces of information. Once I get to the 8th constraint of “Lunch at 1PM”, the model stops waiting for further instructions and decides to produce a plan.

Is it because it thinks that enough is enough and a single day won’t fit more things?

The plan

The results from GPT-4 are unsettingly good. The plan is not only coherent (covers the entire day, nothing gets double booked) and meets all the constraints, it also includes all sorts of nice touches that a good human planner could think of.

In the first part of the day, it sensibly suggests to spend some time with the house guest before they depart. It then allocates 30 minutes to reset or clean my living space… I did need that time but didn’t think of it when listing the constraints!

In the afternoon, it sets up time to pick up the bike after maintenance, in case maintenance was needed:

The evening is appropriately relaxing and empty, with the optional movie (“if energy levels permit”).

GPT-4 vs GPT-3.5 and Bard

I tried this experiment in the paid GPT-4 engine in ChatGPT, the free GPT-3.5 and in Bard.

In my test, only GPT-4 produced a good plan. GPT-3.5 wanted me to “Inflate bike tires and go on a test ride” before 9AM which doesn’t seem humane and then didn’t allocate any time for fixing the bike if needed. Bard wanted me to go swimming for no apparent reason 🏊.

In other news

• 📖 The opening quote from Mustafa Suleyman has been in the media so much because he has a new book that just came out and he’s doing a media tour. The book is The coming wave, dedicated to the risks and opportunities of AI development.

• 📰 Time magazine published their list of 100 influential people in AI, featuring among others Ted Chiang (sci-fi writer, the movie Arrival was based on one of his works) and Margrethe Vestager, executive vice president of the European Commission.

• 🕯️ Molly Holzschlag, known as 'the fairy godmother of the web' died last week. She was a passionate advocate for accessible web design and a successful proponent of open standards. More than once, she challenged Bill Gates face-to-face to fix problems with Internet Explorer.

Postcard from Médoc

The wine marathon in Médoc was a lot of fun! The uphill stretch around kilometer 35 pictured above was a challenge, especially after sampling a few too many of the Bordeaux wines served along the way 😅.

Have a great week 💫,
– Przemek