Microsoft Is Using GPT-4 Wrong
I’m sure it’s been said but I haven’t seen it said: GPT-3 seems way more useful for converting user input into some computer-readable form than for publishing its output directly to users/customers/etc. I would not trust it for that.
(Michael Knepprath, Mastodon)
I want to expand on my thoughts from a month ago based on recent events and my own experience working with GPT.
It’s easy to lose sight of what GPT is in all the chatter about how GPT “feels” or what it “wants”. GPT is closer to a human language emulator, it has no intelligence to call its own. When asked, “What is the meaning of life in one word?”, it will predict an answer letter-by-letter based on similar patterns of letters in its corpus. In this case it might return, “Purpose.” What it’s returning here is not all that different from the autocomplete suggestions that appear in your messaging app as you type.
So, what’s the problem with how GPT is being used right now?
“Liu, who took a leave from studying at Stanford University to found an AI search company called Chord, said such easy workarounds suggest ‘lots of AI safeguards feel a little tacked-on to a system that fundamentally retains its hazardous capabilities.’” (Washington Post, Meet ChatGPT’s Evil Twin, DAN)
GPT has no concept of accuracy or correctness, so it’s wild to see Notion, Microsoft, Snapchat and other well-known companies placing this volatile service in front of their customers with relatively few precautionary measures. When Microsoft announced their new GPT-powered Bing Search, I assumed Microsoft had cracked it. Surely Microsoft wouldn’t release a search engine that told their customers it loved them. Or compared them to Hitler. Or straight up lied and then argued with their customers about the lies it told.
Maybe it’s too early to judge. This is a beta, after all.
But I’d like to posit this: These companies are using GPT technology poorly. In most cases, they should not be outputting GPT results directly to customers. At best it’s a novelty. At worst you have Bing, which uses GPT-4 to serve Microsoft’s customers inaccurate information with an authoritative voice.
Where do I think GPT would be most useful?
Instead of adding separate GPT-enabled experiences to interfaces it should be used to augment existing interfaces. A few examples off the top of my head: voice-activated home assistants (Alexa, Siri, Google Home), chatbots, search inputs and text games.
Customers don’t need to be aware of the fact that GPT or “AI” is involved at all. When the hype wears off, customers simply won’t care (unless it’s making their experience worse).
The theme I will keep coming back to is this: GPT should be used to infer user intent and convert it into a valid format the service understands.
Home assistants have been around for nearly a decade now and still suffer from a core issue they’ve had since day one: they lack an interface. This makes using them completely opaque. You have to know which words to say and in which order to say them to get any kind of consistent result.
GPT should be used to infer user intent based on what they say and convert it into a valid format the home assistant understands.
Instead of saying the strange incantation, “Alexa, stop in the kitchen,” you could say, “Alexa, please stop playing music in the kitchen.” GPT will convert the latter phrase into the former without the user’s knowledge. It will just work.
Facebook led the charge when chatbots had their 15 minutes of fame 7 years ago. They quickly faded away when they hit the same wall as home assistants: they’re completely opaque. Users had no idea which messages would yield useful replies and the chatbots weren’t flexible enough to accept arbitrary user input.
GPT should be used to infer user intent based on what they message and convert it into a valid format the chatbot understands.
This frees users by letting them send messages in any format they feel comfortable with.
Search inputs on websites are typically quite naive. They often don’t do much more than text-matching across the contents of a site. If you don’t know the exact text used on the site, you’ll struggle.
GPT should be used to infer user intent based on what they search and convert it into a valid format the site’s search engine understands.
Not only that, but the search input could be supercharged with its newfound natural language recognition abilities. If GPT is provided the site map, FAQs or even the customer-facing API interface, a user could type, “How do I change the location in my profile to Cleveland, OH?” and be provided with a list that includes a link to a step-by-step guide, a link to the profile editing page, or even the option to make the change directly from the search input itself.
Some text adventure games get around their lack of interface by listing options for the player to select from. This is a big trade-off, as it limits the options a player has at any given moment and prevents the player from exploring. Historically, the trade-off has been worth it because letting the player enter arbitrary commands and fail repeatedly is a frustrating experience. With GPT, this trade-off is moot.
GPT should be used to infer player intent based on what they enter and convert it into a valid format the game understands.
This is a particularly fun example for me because I have already implemented it in my text adventure game, lilt. That’s right, I have a live example of the concept I’m proposing here and I’ll be writing a separate post about how I implemented it. Spoiler alert: It was not hard to do and makes the game infinitely more pleasant to interact with.
I’m not surprised that these companies are currently implementing GPT in quick, hacky ways. There is a ton of hype around the technology right now and hopping on it before the competition garners a lot of attention. I do think this is how we should view a lot of what’s out there right now, though: marketing gimmickry.
I do believe GPT and others like it will be a major part of our technological future. It will become an invisible layer that closes the human/computer communication gap.
Thanks for reading! 👋