AI Horizon

The OpenAI team, led by Sam Altman, has finally unveiled GPT-5, with around 600,000 people watching the launch livestream either live or during the first three hours afterward (and you can do the same too if you missed it). While that lags certain space shots and World Cup events, the audience demonstrates tremendous interest in the newest version of ChatGPT. YouTube isn’t totally forthcoming with viewership data, but only two or three Apple events have had significantly more people watching their livestream.

I previewed GPT-5 before the launch and pretty much everything that we expected was announced. Below, I will talk about what surprised/impressed me and what made me scratch my head.

Wow #1. They have come a long way in 2 1/2 years

Many writers including myself, have complained about hallucinations, errors and other flubs. All true. But we also need to take deep breath and acknowledge the tremendous progress that OpenAI has made since the first public version of ChatGPT. That version was more of a toy than tool. The currently available versions—4.0, 4.1 and 4.5— are very powerful and more useful. If GPT-5 lives up to even half of the promises that they made today, it will represent a major jump forward. Wow 1.5 came when Sam Altman said that they now have 700 million users. I applaud their ambition and am reminded of a quote from Leo Burnett (who is on the Mount Rushmore of advertising): “When you reach for the stars you may not quite get one, but you won’t come up with a handful of mud either.”   OpenAI is clearly reaching for the stars

Wow #2: They are shutting down old models and moving everyone to GPT-5

If you looked at at the menu at the top of the ChatGPT page before launch, you’d see multiple options for different models. There was a word or two about what each model is supposed to be good at. It was always confusing, at least to me. Now they are moving everyone to GPT-5 and are shutting down the older models. This chart summarizes the differences between GPT-4 and GPT-5. Altman said that if GPT-3 was a high school student and GPT-4 was a college student, GPT-5 will operate at a PhD level. This version will be available to all commercial users soon (enterprise and education later). Free users will have their usage capped and will be moved to a less powerful model when they exceed the cap.

Wow #3: Fewer hallucinations

OpenAI stressed that GPT-5 will spend less time in Fantasyland than prior versions. This is welcome news, assuming that it’s true. They made reference to percentages better but they didn’t show their work. My gut says it will be directionally (and probably significantly) better but not yet perfect.

Wow #4: Much stronger voice integration

Some people have used voice to access ChatGPT in prior versions. There has clearly been a major effort to elevate voice to the same level as text input. I am not yet clear how much of their usage this will represent but it is definitely a great option. My skepticism is based in part on being overly optimistic about how quickly Alexa would revolutionize commerce for Amazon. I envisioned a world where people in kitchens would be ordering groceries while they cooked but I haven’t seen that happen. Typing may still be the preferred input method for many people.

Wow #5: More powerful code development  

This will go into both categories. They tried to demo how easy it is to develop web applications using GPT-5. One fellow “developed” an app to teach his girlfriend how to speak French with his family. It kinda, sorta worked but looked like a high school class project. They claim that coders prefer working in GPT-5 but didn’t show their work or sources. They also showed how GPT-5 could help with debugging. Put this in the bucket of “great if it actually happens”.

Wow #6:  Big strides against bad actors

First, do no harm. It sounds as if OpenAI is taking its responsibilities to society seriously. There was discussion of how GPT-5 would respond to problematic searches. It will provide more context as to why something is wrong, unethical or dangerous. They are terming this “safe completion”. Will be worth continuing to track this over time.

Hmmm #1: This felt like a class presentation not a product launch from a major company

Maybe Apple and others have spoiled us, but the actual presentation was amateurish. They ran nervous product people on and off of the stage in three minute segments. A lot of the “jokes” felt unscripted and awkward. I know it’s about the product, not the packaging, but it feels like it is time for them to up their game.

Hmmm #2:  They need help translating their excitement and thoughts into English

One very interesting example showed how a parent might help their high schooler with a science report on the Bernoulli effect. So far, so good. The presenter then asked “Wouldn’t it be great to have an animation that showed this?” Yes! And, if she had continued to explain it in layman’s terms, it would have been very great. Instead she said “create an SVG using Canva” and showed a bunch of Python code. There is still an opportunity to make the power of ChatGPT 5.0 more accessible to the average user but it needs to be pitched at that level.

Hmmm #3: They are claiming to have jumped the evolution of AI ahead by two or three years

One of the leading AI scientists is Daniel Kokotajlo. He was a top engineer at OpenAI before leaving. He is also the main author of AI 2027, a very well documented and reasoned study of the development path for AI. According to that paper, it will still take two or three years for AI to really do much of a developer’s job. OpenAI pretty much claimed they are heading there with GPT-5. Not sure that I buy this. Again, directionally perhaps. All the way? Not until I see more.