They use LLMs for what they can actually do, which is bullet point core concepts to a huge volume of information, parse a large volume of information for specific queries that may have needed a tech doing a bunch of variations of a bunch of keywords, before, etc. Provided you have humans overseeing the summaries, have the queries surface the actual full relevant documents, and fallback to a human for failed searches, it can potentially add a useful layer of value.
They’re probably also using it for propaganda shit because that’s a lot of what intelligence is. And various fake documents and web presences as part of cover identities could (again, with human oversight), probably allow you to produce a lot more volume to build them out.
right, at which point you’re just better doing it the right way from the beginning, not to mention such tiny detail as not shoving classified information into sam altman’s black box
I’m not really arguing the merit, just answering how I’m reading the article.
The systems are airgapped and never exfiltrate information so that shouldn’t really be a concern.
Humans are also a potential liability to a classified operation. If you can get the same results with 2 human analysts overseeing/supplementing the work of AI as you would with 2 human analysts overseeing/supplementing 5 junior people, it’s worth evaluating. You absolutely should never be blindly trusting an LLM for anything. They’re not intelligent. But they can be used as a tool by capable people to increase their effectiveness.
the other thing about text like this is that many of the claims of what they’re doing will be completely false because someone will have misunderstood then will try to reconstruct a sensible version
it’s not airgapped, it’s still cloud, it can’t be. it’s some kind of “secure” cloud that passed some kind of audit. openai already had a breach or a few, so i’m not entirely sure it will pan out
My interpretation of what they’re saying is that it’s on their own servers in their own location that can only be accessed from specific access points.
Talking about networks as airgapped isn’t abnormal.
they probably do. I worked for a content-as-a-service company that had a contract to deliver our product, airgapped, to a three-letter agency on a regular schedule, and we were a tiny company. Microsoft’s biggest customer is probably the U.S. government; I’d be shocked if they don’t provide an in-house airgapped set of full Azure services for the entire intelligence agency system.
tbh I personally wouldn’t expect/suspect this to be using any of the flavours of govcloud for mass-market flavours (because that has implications on staff hiring etc)
the easy way to handle this is to have a backend/frontend separation with baseline access controlled simply by construction of routing and zone primitives. it’s relatively simple (albeit moderately involved) to do this on most cloud providers
They use LLMs for what they can actually do, which is bullet point core concepts to a huge volume of information, parse a large volume of information for specific queries that may have needed a tech doing a bunch of variations of a bunch of keywords, before, etc. Provided you have humans overseeing the summaries, have the queries surface the actual full relevant documents, and fallback to a human for failed searches, it can potentially add a useful layer of value.
They’re probably also using it for propaganda shit because that’s a lot of what intelligence is. And various fake documents and web presences as part of cover identities could (again, with human oversight), probably allow you to produce a lot more volume to build them out.
right, at which point you’re just better doing it the right way from the beginning, not to mention such tiny detail as not shoving classified information into sam altman’s black box
I’m not really arguing the merit, just answering how I’m reading the article.
The systems are airgapped and never exfiltrate information so that shouldn’t really be a concern.
Humans are also a potential liability to a classified operation. If you can get the same results with 2 human analysts overseeing/supplementing the work of AI as you would with 2 human analysts overseeing/supplementing 5 junior people, it’s worth evaluating. You absolutely should never be blindly trusting an LLM for anything. They’re not intelligent. But they can be used as a tool by capable people to increase their effectiveness.
the other thing about text like this is that many of the claims of what they’re doing will be completely false because someone will have misunderstood then will try to reconstruct a sensible version
it’s not airgapped, it’s still cloud, it can’t be. it’s some kind of “secure” cloud that passed some kind of audit. openai already had a breach or a few, so i’m not entirely sure it will pan out
My interpretation of what they’re saying is that it’s on their own servers in their own location that can only be accessed from specific access points.
Talking about networks as airgapped isn’t abnormal.
Iirc OpenAI uses Microsoft’s cloud?
If so, MSFT has a special airgapped cloud specifically for USGov.
they probably do. I worked for a content-as-a-service company that had a contract to deliver our product, airgapped, to a three-letter agency on a regular schedule, and we were a tiny company. Microsoft’s biggest customer is probably the U.S. government; I’d be shocked if they don’t provide an in-house airgapped set of full Azure services for the entire intelligence agency system.
They do. Source: I worked in at MSFT in Azure Identity. It’s completely separate, has its own rollout schedule for all products, etc.
There’s also a physically separate cloud for China 🙃
tbh I personally wouldn’t expect/suspect this to be using any of the flavours of govcloud for mass-market flavours (because that has implications on staff hiring etc)
the easy way to handle this is to have a backend/frontend separation with baseline access controlled simply by construction of routing and zone primitives. it’s relatively simple (albeit moderately involved) to do this on most cloud providers
I see you’ve never taken part in a FedRAMP audit. They’re brutal.