I’m lead 365 admin for a major corporation and have been working with MS to identify if Copilot would be beneficial and secure for my org. Some major takeaways from my recent meetings with them:
There’s two parts to Copilot. 1. Copilot 2. Copilot for 365.
The first is basically Chat GPT. It reaches out to the web to get info and essentially works as a search engine.
The 2nd part is internal only. It can do things like summarize meetings, compare documents, and search your emails. It abides by the same security, compliance, encryption, and DLP policies as the rest of your tenant.
You can open up access to one or both.
Government tenants are a unique case. There’s a specific 365 license for government entities, and their offerings are different from other organizations. This news article isn’t surprising - all new 365 offerings take a while before they’re available to government licenses. It will eventually be available.
Few questions about that, unless they’re literally taking their model and putting it into your own box using it’s own compute power, I don’t see how that’s possible. They can call it “your” copilot all they want but if they’re reading your data and prompts and computing that on their own box then they’re using your data, right?
Major organizations use encryption where they hold the keys so Microsoft is unable to read their data. They can have thousands of servers running on Microsoft’s Azure stack and yet Microsoft is unable to read the data that is being processed.
Thanks for the explanation
If all auditors are uncorrupted, highly competent and have full overview. Boeing was able to corrupt it’s government auditors to save some money on redundant sensors. With Microsoft pushing big on gathering and selling data I wouldn’t trust a byte that passes their server.
You clearly do not understand encryption or corporate auditing.
Microsoft has to compete with other cloud providers on security. Unlike Boeing who has no domestic competition. Any of Google, Amazon, or Oracle would love to find out that Microsoft is decrypting user data to sell to partners because they would be screaming to the high heavens that O365/Azure is insecure and enterprises must switch to their solutions. SaaS/IaaS subscriptions are much more profitable than selling user data, there is a near 0 chance that Microsoft is improperly handling enterprise data (on purpose)
Microsoft cannot decrypt your data when you hold the keys.
Until we either solve the problem of LLMs providing false information or the problem of people being too lazy to fact check their work, this is probably the correct course of action.
Imo the human laziness is the issue. Every thread where a lot of people chime in about ai, so many talking about how it’s useless because it’s wrong sometimes. It’s basically like people who use Wikipedia but can’t be bothered to cross reference… Except lazier. They literally expect a machine to be flawless because it seems confident or something?
I think you’re missing the point. I don’t like copilot/chat gpt for important stuff because if I have to double check their solutions I barely gained any time. Especially since it’s correct more often than not because it will make me complacent over enough time (the professors who were patient enough to actually explain why we shouldn’t be using Wikipedia as a primary source also used the same point which I thought made a lot of sense).
You’re going to need to fact check any code you get online anyways, why not have it hyper specific to your current use case? If you’re a good developer, review does not take nearly as long as manual implementation
I very rarely grab code online because I work in videogames and it’s very hard to find good code for the things I struggle with since all the publicly available stuff is for hobbyists and thus usually very basic/unoptimized as hell
Most of the time the stuff I can’t figure out myself isn’t even mentioned anywhere on hobbyist forums because it’s not needed for these applications (for a recent example: assets management. For hobby projects you can usually get away with hard references to all of your assets, so it’s not even a thing)
If what you want is difficult to find publicly, then that also means an LLM is going to be weak in that area as well
What you want is a “general AI” LLM, something capable of stringing together a solution based on past somewhat related solutions. We’re not here yet, so basically you’re asking it to do something beyond what it is capable of and it’s trying its best anyways
Alternatively, you could try fine tuning your own LLM, if you have access to some sort of large repository with non-public solutions or something
I can’t imagine using any LLM for anything factual. It’s useful for generating boilerplate and that’s basically it. Any time I try to get it to find errors in what I’ve written (either communication or code) it’s basically worthless.
My little brother was using gpt for homework and he asked it the probability of extra Sunday in a leap year(52 weeks 2 days) and it said 3/8. One of the possible outcomes it listed was fkng Sunday, Sunday. I asked how two sundays can come consecutively and it made up a whole bunch of bs. The answer is so simple 2/7. The sources it listed also had the correct answer.
All it does it create answers that sound like they might be correct. It has no working cognition. People that ask questions like that expect a conversation about probability and days in a year. All it does is combine the two, it can’t think about it.
Did anyone expect otherwise?
