Are you including cyber-testing by UK AI Safety Institute in this list?
Sure, the models are capable (for some test tasks, though they are not omnipotent yet) but does it mean the actual OAI sandbox is adequate? Could have a competent engineer done better and made the escape less likely?
It may be even simpler: plain markdown files that explain architecture including C4 diagrams generated from text using plantuml (automated via invoke/fabric/whatever). Plantuml includes C4 in stdlib https://github.com/plantuml-stdlib/C4-PlantUML
Yeah, haha funnily enough I'm a huge fan of both C4 and plantuml. We actually do have an experimental 'software map' feature, turned off in settings by default, which draws a hierarchical / C4-based architecture diagram of your codebase (even highlighting the nodes which changed in the PR!) Warning, it is a little rough around the edges at the moment.
As far as plantuml/graphviz/mermaid/similar UML tools, it's really cool stuff, I agree. My issue is that those tools are mainly built to render static diagrams (output to svg, png, etc.) Some eg mermaid are perhaps extensible for interactivity, but it's a little wonky. I think a lot of the value of whiteboard comes from the interactivity -- being able to click on a node to see the underlying code, being able to go through a guided walkthrough of a given flow, etc.
But using their UML authoring format is certainly worth looking into, since the LLMs are already so well-versed in it. i'll put some thought into this!
Middle may be wrong (it is a marketing trick to add 3rd outragesly expensive option, to make 2nd option look reasonable). The correct answer is “it depends” (even how long you should spend on choosing may depend on context too).
For example, write in whatever language you know best, then translate to a more appropriate language using LLMs once the desired behavior can be checked automatically. It is a tactic that works in some cases.
It can be used as a marketing trick, possibly because of the logic in GP post. Marketing people use this to trick people, which is different from it being a marketing trick in and of itself.
At the moment, the hurdle is [missing] context [aka shared model/understanding of the business, common sense].
Though it is not hard to imagine that any and all communications being recorded for AI consumption in the future.
The Next one is the relative lack of prompt feedback (expect the blowup in finite time like Navier-Stokes ;) [there is not much feedback even for humans at middle management positions].
The cost [tokens] might become prohibitive unless LLMs improve further [not a guarantee].
Is the implication that those who impose these sanctions can MITM anybody in the world except for countries under sanctions (which have to use their own root CAs)?
I've only heard of OpenAI and Anthropic agents hacking during training. Are open weights model not smart enough to hack Hugging Face to copy solutions? (doubtful)
MAD is a belief system. It doesn't matter whether AI can effectively stop the retaliation, it is enough that people making the decisions believed that it can (are we there already?).
Sure, the models are capable (for some test tasks, though they are not omnipotent yet) but does it mean the actual OAI sandbox is adequate? Could have a competent engineer done better and made the escape less likely?
reply