Artificial-intelligence aide handles email, meetings and other things, but its price and limited use have some skeptical
Microsoft’s new artificial-intelligence assistant for its bestselling software has been in the hands of testers for more than six months and their reviews are in: useful, but often doesn’t live up to its price.
The company is hoping for one of its biggest hits in decades with Copilot for Microsoft 365, an AI upgrade that plugs into Word, Outlook and Teams. It uses the same technology as OpenAI’s ChatGPT and can summarize emails, generate text and create documents based on natural language prompts.
Companies involved in testing say their employees have been clamoring to test the tool—at least initially. So far, the shortcomings with software including Excel and PowerPoint and its tendency to make mistakes have given some testers pause about whether, at $30 a head per month, it is worth the price
Like many tools, there’s a gulf between a skilled user and an unskilled user.
What ML researchers are doing with these models is straight up insane. The kinds of things years ago I didn’t think I’d see in my lifetime, or maybe only in an old age home (a ways off).
If you gave someone who had never used a NLE application to edit a multi track video access to Avid for putting together some family videos, they might not be that impressed with the software and instead frustrated with perceived shortcomings.
Similarly, the average person interacting with the models often hits their shortcomings (confabulations, safety fine tuning, etc) and doesn’t know how to get past them and assumes the software tool is shitty.
As an example, you can go ahead and try the following query to Copilot using GPT-4:
It will get it wrong (despite two prompt engineering techniques already in the query), defaulting to the standard form solution where the goat is taken first. When GPT-4 first released, a number of people thought that this was because it couldn’t solve a variation of the puzzle, lacking the reasoning capabilities.
Turns out, it’s that the token similarity to the standard form trips it up and if you replace the wolf, goat, and cabbage in the prompt above with the emojis for each, it answers perfectly, having the vegetarian wolf go across first, etc. This means the model was fully able to process the context of the implicit relationship between a carnivorous goat eating the wolf and a vegetarian wolf eating the cabbage and adapt the classic form of the answer accordingly. It just couldn’t do it when the tokens were too similar to the original.
So if you assume it’s stupid, see a stupid answer and instead of looking deeper think it confirms your assumption, then you walk away thinking the models suck and are dumb, when really it’s just that like most tools there’s a learning curve to get the most out of them.
My problem with this is that your example replies on you already knowing the correct answer, so that you know it’s given you the wrong answer and you can go back and try to trick it into giving a different answer. If you’re asking it a question to which you don’t already know the answer, how would you know if this has happened?
Don’t use LLMs in production for accuracy critical implementations without human oversight.
Don’t use LLMs in production for accuracy critical implementations without human oversight.
I almost want to repeat that a third time even.
They weirdly ended up being good at information recall in many cases, and as a result have been being used like that in cases where it really doesn’t matter much if they are wrong some of the time. But the infrastructure fundamentally cannot self-verify.
This is part of why I roll my eyes when I see employment of LLMs vs humans presented as an exclusionary binary. These are tools to extend and support human labor. Not replace humans in most cases.
So LLMs can be amazing at a wide array of tasks. Like I literally just saved myself a half hour of copying and pasting minor changes in a codebase by having Copilot automate generating methods using a parallel object as a template and the new object’s fields. But I also have unit tests to verify behavior and my own review of what was generated with over a decade of experience under my belt.
Someone who has never programmed using Copilot to spit out code for an idea is going to have a bad time. But they’d have a similar bad time if they outsourced a spec sheet to a code farm without having anyone to supervise deliverables.
Oh, and technically, my example doesn’t actually require you to know the correct answer before asking. It only requires you to recognize the correct answer when you see it. And the difference between those two usecases is massive.
Edit: In fact, the suggestion to replace the nouns with emojis came from GPT-4. Even though it doesn’t have any self-introspection capabilities, I described what I thought was happening and why, and it came up with three suggestions for ways to improve the result. Two I immediately saw were dumb as shit, but the idea to use emojis as representative placeholders while breaking the token pattern was simply brilliant and I’m not sure if I would have thought of that on my own, but as soon as I saw it I knew it would work.
But that’s what the marketers are selling, “this will replace a lot of workers!” and it just cannot