Last week's big AI event was OpenAI's Developer Day, a splashy San Fransisco summit where the company announced two major features: a new way for users to create custom versions of ChatGPT, and a significant update to its industry-leading GPT-4 AI model.

In a typical week, either of those would merit its own column — so let's take the time to explore both in depth.

Big News #1: You can now build your own customized ChatGPT 

Up until now, everyone who logged onto ChatGPT was greeted by the same generic chatbot.  

To use an imperfect metaphor, it was as if, whenever anyone dialed "1-800-CHAT-GPT," they always talked to the same guy named Jim. That wasn't a bad  thing — Jim was reasonably smart, and he was unfailingly friendly — but he didn't know anything about you, your company, or your use cases. 

No longer. Now you can create your own version of ChatGPT called, confusingly, a "GPT": a customized chatbot that can access your organization's files, run custom functions, and respond consistently over time. Essentially, instead of dialing up Jim, you can call your own specialist. 

Creating a GPT is as simple as writing plain-English instructions for your chatbot and uploading any files you want it to reference. (More sophisticated developers can, of course, create far more complex chatbots.) 

What might you use a GPT to do? Among OpenAI's demos are "The Negotiator," which acts as a negotiation teacher; "Creative Writing Coach," which gives feedback on your stories and poems; and "Mocktail Mixologist," which designs recipes for alcohol-free drinks. 

Already developers are experimenting with healthcare-specific GPTs. A quick Google search reveals GPTs promising (implausibly) to diagnose diseases, provide AI-assisted therapy, and even serve as hospital strategists.

I played around with a few of these healthcare chatbots, and candidly, their outputs didn't seem any better than those of the usual ChatGPT … but it's early days, and I'm sure some will improve. 

Big News #2: OpenAI's most advanced AI model just got faster, cheaper, and — in some ways — better

Ever since GPT-4 launched in March, it has been the world's most powerful large language model (LLM), topping most leaderboards for accuracy, understanding, and logical reasoning. Unfortunately, it was also relatively slow and expensive. 

Imagine conducting a simple back-and-forth: You ask GPT-4 to summarize Advisory Board's "16 Things CEOs Need to Know in 2023," and it responds with a 500-word summary. At previous rates, that exchange would cost about $1 — which adds up if you're performing thousands or millions of requests. 

But on Dev Day, OpenAI launched a new model called GPT-4 Turbo, which costs only 17 cents for the same request. What's more, GPT-4 Turbo is  better  on important dimensions: 

• GPT-4 Turbo's knowledge of the world is more up-to-date. It's familiar with events through April 2023, compared to September 2021 for the original GPT-4.
• It can understand and respond to much longer requests. The most capable older version of GPT-4 could manage 32,000 "tokens" (basically, words or sub-word units). The new version quadruples that to 128,000 tokens, enough to fit hundreds of pages of text.
• It's faster. Developers report that GPT-4 Turbo is quite zippy in generating responses — perhaps 2x or 4x faster than GPT-4.
• It can provide consistently structured responses. For an LLM's outputs to fit elegantly into a workflow, developers often need its responses to follow a predictable format. They might, for instance, want the details of a prescription to look like "medication name, dosage, frequency." Most LLMs struggle to follow well-defined formats consistently, but GPT-4 Turbo can guarantee every response arrives in the right way.

The downside is that, according to early testers, GPT-4 Turbo is a smidge less intelligent than GPT-4: It hallucinates more, struggles to manage the most complex requests, and so on. Even so, it's better than nearly any other AI model on the market. 

Other top AI news from the past week 

A deeper dive into OpenAI's Dev Day. The most accessible write-up I found on Dev Day came from Timothy B. Lee of Understanding AI, who contrasts OpenAI's approach with that of Apple: :When Apple launched the iPhone in 2008, it offered developers a single clear vision for how apps should work. In contrast, OpenAI is offering a menu of possible approaches to developing AI-powered software and letting developers choose."

What can custom GPTs actually do? In this blog post, Ethan Mollick of the Wharton School dives deeper into custom GPTs, including a step-by-step walkthrough of creating two simple chatbots. The real value of GPTs, he argues, will come in connecting them to outside services: "GPTs show a near future where AIs can really start to act as agents, since [they can] connect to other products and services, from your email to a shopping website."

How a game of '20 Questions' reveals the deep weirdness of LLMs. This Nature article offers a provocative example of LLMs' inherent strangeness via a game of "20 Questions." 

Think of it this way. When you play "20 Questions" with a human, your partner first decides on the secret item they're thinking of, then answers your questions accordingly. But LLMs work very differently: Because they generate each word of their response in realtime, when they begin to answer your 20 questions, they  haven't actually "thought”"of an item.  Instead, they offer plausible-seeming answers to your questions — and finally, at the end of the game, generate an answer consistent with everything they've already said. 

I just tested this myself with ChatGPT: I posed several questions that established it was "thinking" of a domestic animal smaller than a refrigerator, then asked it to reveal its secret answer. The first time I asked, it said "a cat" … but when I regenerated the final message in our conversation, it claimed to have been thinking of a dog all along.  

It's just one more way in which LLMs are so much weirder than they appear.