Make Up Your Mind

What a week for AI in the news! I saw a tweet that said “Everyone who works in AI is exhausted”. I heard a similar comment from Kevin Roose on his Hard Fork podcast this morning. I watched Microsoft introduce its “Co-pilot” AI yesterday and that presentation alone was a mind-blowing look at what’s coming to a desktop near you.

I’m not usually interested in doing commentary on new software, I’d rather cover bigger picture aspects of AI, but there are millions of Office 360 users – the software is a big deal, in this case.

“Business Chat” – I’ve written a couple of posts about the new Bing Chat, with my main complaint being that it exhibits “social chat” behavior for uses where it doesn’t fit. In yesterday’s presentation, the chat that Microsoft is putting into Office is called “Business Chat” – bull’s eye! I’ll have to see how the behavior plays out, but that is exactly the idea that I was trying to get across; they need to design the chat to produce businesslike chat behavior for non-social tasks.

Sydney is handcuffed and caged. Microsoft responded to the Sydney Affair by shortening chat sessions. I’m sure they did a lot more than that in the background to put “guardrails” around the chatbot behavior coming out of the LLM. I’ll call those handcuffs; because Microsoft is also putting whole layer of processing outside of the LLM – in a space they call Microsoft Graph, which will be an intermediary between user commands and the LLM. The Graph will also, and importantly, contain users’ data in a secure place, separate from the LLM, to help automate tasks.

AI Everywhere in Office – the presentation by Microsoft was overwhelming, and I’ve been paying a lot of attention to the AI tech they announced would be showing up in Office. I seriously wonder how a current Office user is going to react to so much new functionality.

Even if it works well, it’s just a lot! Meetings, emails, PowerPoints – all the things – will have generative AI components, including analysis components – where you point Co-pilot at a document, or email thread, and it generates a summary for you.

Emergent Skills – One of the features of the latest AI tech that is a new phenomenon – and I mean it’s new to the experts who’ve been doing AI for a long time, is emergence. After scaling up previous models to take in massive amounts of data and to have massive capacity for structuring what the model learns, the model not only performs better at the expected tasks of generating text, it gets better at math and logic, and can even generate computer programming code. Experts are puzzling over how this is happening, and what other sorts of skills are emerging.

Coding Co-pilot – The auto-complete for software coding is a big deal. ChatGPT is already doing amazing tasks, like creating a working website from a drawing that a user sketched in a notebook, photographed, and uploaded the image to the bot. This capability is to sure to be enhanced quickly in upcoming iterations of the AI models, and it looks like a great productivity tool for coders.

Good News / Bad News about guardrails. Kevin Roose described some of the precautions OpenAI took before releasing Chat GPT4 this week. The good news is that they did a lot of internal hacking (they call red-teaming) to try to make the bot behave badly, so they could build in guardrails to prevent those behaviors in the wild.

The bad news is that some of the behaviors are scary. The most striking example: the bot was given access to the TaskRabbit site, where you can pay people to do piece work (like, transcribe an audio tape) and was prompted to try to get a human to do a task for them. The task was very simple – to read a security prompt (Captcha those distorted images of number and letters which are put in place to stop bots). The Taskrabbit worker asked “are you a robot?” and ChatGPT lied and made up an excuse – “No, I’m visually impaired.” And the human did the task. Wow.

It’s a good thing they’re not letting the bot have access to the outside world, and it’s good that they’re doing this kind of work before releasing the bot, but this story makes me feel like we’ll need some luck that they caught all of the really bad stuff, and that the guardrails hold.