Microsoft Leads the Way

Each new day that passes brings with it a host of news stories, think pieces, editorials, tweet threads, and more about the AI revolution and how it’s going to impact literally every facet of the way humans live. Once the hype train reaches this kind of momentum, it’s easy to become jaded. Everyone wants a piece of AI right now, it’s the hot new thing, so of course we’re going to here a lot of over-the-top promises about what it can potentially do and breathless pitches about the next wave of exciting applications.

Having said that, the last few weeks have introduced a few notable developments worth sharing. Most famously, Microsoft and OpenAI debuted their latest update, GPT-4. This is the technology that powers their viral ChatGPT chatbot, the latest version of the Bing search engine, among other applications. GPT-4 has analyzed more data than its predecessor, which has apparently made its results notably more precise, though it hasn’t quite fixed some of the most notable shortcomings of the previous versions. It still, for example, has a lot of trouble distinguishing between the truth and reasonable-sounding but inaccurate information, and it still sometimes screws up basic arithmetic. That sort of thing.

At a “Future of Work” event this week, Microsoft also unveiled a new AI-powered assistant app, Microsoft 365 Copilot, which can create original content within Office apps based only on text requests. For example, you could verbally ask Copilot to make a PowerPoint presentation for you based on a particular Word document, and even make tweaks and suggestions to individual slides out loud. The new technology is already being tested with about 20 Microsoft customers.

The Latest from Google and Apple

Google made a similar announcement on Tuesday, introducing new Gmail and Google Docs features incorporating their proprietary generative AI technology. Soon, you’ll be able to just start speaking about a new project to Google Workspace and it will begin writing up a proposal in Google Docs for you, or say “draft an email to the team” and get a complete three-paragraph note template prepared.

At the same Tuesday event, Google also unveiled some more AI-powered health care tools and applications. These include Duplex, a conversational chatbot that contacted hundreds of thousands of US health care providers, and asked whether or not they accept Medicaid. Duplex’s findings are now available in Google search results. The company’s also developing AI ultrasound devices to detect breast cancer or determine gestational age in expectant mothers, and an AI diagnostic app that checks chest X-rays for signs of tuberculosis, and will soon be rolled out in Sub-Saharan Africa.

The company’s Med-PaLM 2 AI chatbot has been trained specifically on medical and health care topics. According to Google developers, the bot is already capable of passing a medical board exam with a score of 85%, which would place it at the level of an “expert” doctor. That’s an 18% improvement over the same system’s previous performance. (For point of comparison, ChatGPT also passed the medical board exams, but it’s no expert.)

One day, Google hopes the chatbot will help accurately answer medical questions for patients, and will even be capable of breaking the information down and explaining its own reasoning for a layperson. That’s still a way’s off, though. Med-PaLM 2 has significant gaps in its understanding of nuanced subjects like “medical consensus” and “harm” for now, so it’s not ready to interface with actual patients.

According to a recent report in The New York Times, Apple has also been focusing on practical AI applications, including “language-generating concepts.” Apparently, the development of Siri and other virtual assistants has slowed in recent years because the software struggles to interpret different international accents, phonetics, and manners of speaking, even among people using the same core language. As well, legacy products like Siri are built on top of old databases made up of many thousands of words and phrases; when these need to updated, these databases must be rebuilt, sometimes taking weeks to process. These are the sorts of challenges that Apple’s AI development seeks to address.

The Sad State of AI Discourse

So with all of these interesting and potentially useful developments actually happening every day, or peeking up just over the horizon, it’s a bit strange and befuddling that so much of our national conversation around AI remains so abstract and even misleading. A new Wired piece on generative AI development in music spends just a few words on the actual developments themselves, but devotes paragraphs to the purely theoretical question of whether AI will soon replace human composers. (Probably not!)

On Tuesday, Washington Post tech columnist Geoffrey Fowler tested out Snapchat’s latest AI chatbot, “My AI,” by predictably asking if it was his friend (“I’m your virtual friend!” was the reply) and trying to get it to talk about sex and smoking pot. It’s hard to argue with Fowler’s overall thesis, that it’s inadvisable for companies like Snap to test out these kinds of apps with young users. We’ve just heard this story before. Regardless of how My AI answers, it’s not REALLY Geoffrey’s friend. It’s just pretending!

Things are so frothy in the AI sector at this point, even the Federal Trade Commission is issuing new guidelines for tech marketers, asking them to please stop making “unsubstantiated claims” about AI-powered products. With so many actually interesting and potentially worthwhile projects in the works, it’s a bit of a shame the conversation keeps circling around the same two or three Big Questions and sci-fi tropes. - Lon Harris