Most of the reference documentation is actually internal project/other google sources and internal docs/guides/design docs.
Todays LLMs are no suitable replacement for documentation, in my experience, because their knowledge is so sparse. You will not notice it immediately because they fill the gaps with plausible nonsense. Also training a model with domain specific knowledge is not a realistic option for most of us as of today.
For having all the reference documentation locally (possibly indexed in a vector database and accessible to the LLM) I'm doubtful as well, since it is so hard to determine scope beforehand. A couple of years ago I tried to program off-grid and prepared a MacBook with Dash (an OSX offline doc reader) and all the reference docs I thought I need. It was a nightmare, and that is from a dude who learned programming before the Internet, based on offline docs solely.
Not relying on network connectivity in my entire pipeline is still today a big discriminator in the stack I choose.
The biggest downside was, and still is, documentation. I had a script that scavenged the installed packages and downloaded all relevant documentation packages automatically, along with a local indexing service (I'm still using and recommend "recoll").
There is a humongous gap between libc/posix (anything instantly accessible with "man" with stunning quality) and pretty much everything else. Even python, which has a decent "online" (built-in) reference, is not as convenient and easily searchable when looking at the html manual itself.
A local search index doesn't have the inference power of something like google to point you to the right example you're looking for something in cppreference.
And once you pass the top 10 best documented packages you have on system, you realize how downright bad everything else is documented and how much you rely on the search engine to fill the gaps from unofficial sources.
For me at least, documentation was always THE problem.
I have not, will not, and am not allowed to copy and paste anything into chatGPT from my company that I wouldn't be comfortable putting on twitter. I agree with this policy.
The last thing someone expects is that when they're replying to a support ticket is Google or OpenAI being fed that info and it's being put into a training dataset.
We all value our privacy but it seems accepted to breach other people's privacy for the sake of using an LLM to make your job easier.
If GDPR were actually enforced consistently, or similar laws covering more specific information that exist in the United States, the economy would crawl to a screeching halt and there would be warlords fighting it out over what's left of civilization within a week.
I'm not really sure what the solution to this is, only that we are lying to ourselves that we are anywhere close to having figured it out, and anybody involved that tells you otherwise (people selling ztna, etc) is willfully ignorant or a conman.
There is also interesting discussion to be had about the meaning of all of this. It's not clear at all that we have ever had a society that guaranteed the level of privacy that some seem to expect. We always had other ways of monitoring people and violating privacy that did not require electronics. Community involvement to a level sufficient to ensure nobody is totally screwing everybody else over has always been a requirement if you wanted to not be living in a cardboard box or a prison. I happen to think that is better than a total state or the modern corporation, but recognize that others can have different views on this.
We have things like GDPR which are all about protecting your data. People are breaking the law and violating the rights of customers for the sake of making their job a little bit easier.
I’d like to know if chatgpt agrees.
godbolt would be a weird choice to whitelist if you go through the trouble of airgapping in the first place. All it takes is one accidental copy&paste of a sensitive code snippet and it's there for the world to see with no undo.