@garybernhardt I think the issue is with the “implicit” word.
Implicit asks for every possible thing that is a state machine on some level, including esc key closing a dialog.
Maybe “all *impactful* implicit state machines” would have focused the dev work on the highest ROI parts
@emojibakemono@garybernhardt Yes I would say every app is a state machine if it runs on deterministic software.
But whether you want to model it in your programming language explicitly or implicitly is up to you.
@denbvk Google: C++ is too complicated to onboard our junior engineers so we created Go
The same motive applies to LLMs, it seems.
I have been wondering about the cost of generating Rust code with lifetimes. Surely the LLM requires extra computation vs letting the runtime GC handle it.
I find it strange that some developers are training their AI to produce code using the styles and layout that they would use as a human. Those are designed for human efficiency in reading and writing. LLMs do not have those same human time and energy costs.
Some of those abstractions are thin and light weight syntactic sugar. If you will never write the code directly again, it makes no sense to use those abstractions. I think LLMs will work better by having more freedom at the lower level abstractions.
Eg an ORM vs SQL directly. The LLM can now statically map SQL tables to state, and get the full benefits of the SQL language and no runtime cost or limitations of the ORM.
Just an example, but it applies to many kinds of abstractions they the LLM works with.
My new project
https://t.co/FK2tW5cy2K
Pick your favourite server and upload it's private key to the web UI.
Join a virtual "table" with 4 others doing the same.
One SSH server is chosen at random, the system SSHs in and does `sudo rm -rf`
You spend the rest of the day thanking your lucky stars your server is fine
Or doing apt-get install
Care to join me for a game?
@jarredsumner It’s still one dependency whether you write it or not?
But your new obscure one will have significantly less tokens spent on finding its bugs.
I agree: Both SQLite and the “direct filesystem write” are user space programs.
It’s comparing:
1. SQLite
vs
2. the worst possible file write program
They both write to the filesystem but SQLite is allowed to apply optimisations like caching and batching
There is a 3rd app: something that is not SQLite that applies all possible file write optimisations
I think the benchmark just says “SQLite is faster than the default way to write files”, not “SQLite is faster than the actual filesystem”.
@lauriewired Well technically filesystems are specialised databases.
And the file system is still underneath SQLite so there could be a way to game the benchmarks with optimised direct file writes that maxes out the disk write speed that would beat SQLite.
An attacker would need your authenticated hardware device
So they need two factors:
1. Device password
2. The device
If passkey syncing is insecurely implemented this breaks down.
But still much stronger than a string of chars used for passwords which can be used from anywhere on the network