different system prompts, codex will have system prompt telling the model to gather a lot more context before starting work

for that you would need to compare the same task implemented in two different languages - C# and Python for example, no?

The advantage of their dataset is it's large enough that you have a statistical number of PR's, allowing you to treat the average cost of Python PR's vs the average cost of C# costs (or etc.) as statistically meaningful.

they explain this is a benchmark, all models/harnesses receive the same prompt

don't forget "where are all these beautiful apps that supposedly everybody vibe codes now?"

Who says this? "Beautiful" vibecoded apps are a dime a dozen. Getting support or continued feature development for those beautiful apps after the developer's AIDHD moves on to their next half-baked idea is usually the differentiator between a good vibecoded app and a bad one.

easier to plan/estimate compute for 5 days than for a month. worst case you only have 5 "unprofitable" days

there are many sources saying that Anthropic has 80% margin (profit) on API tokens

and "accidentally" they forgot to disable it when releasing

Wouldn't be the first nor the last time someone is asked to ship something and it gets rushed through for reasons XYZ...

This being said makes the situation for an attacker awfully convenient...

Believe it or not, shit happens in the software business.

I know this from personal experience.

good approach, but your security should not depend on your router anyway, you should be immune to attacks from it

Not exposing your management interface to internet and running a guest network which doesn't have access to said management interfaces can block 95%+ of the attacks, I believe.

yes, defense in depth

an LLM can't access its high dimensional vectors any more than we can access whatever the brain is doing at a low level

all kind of math structures were found in mammals brains - fourier transforms (well, not exactly), ballistic equations, Gabor filters

who knows how exactly we approximate the magnitude of a math operations, maybe we also use helices

my point is that we dont know if what we discover the neural networks doing (helical manifolds) is actually the same thing a brain converges on, or not

and there is an implicit bias here - evolution created language, and we forced neural networks to also evolve to be good at it. so it wouldn't be surprising to find some convergence, this particular kind of language turned out to work well (words, linear sentences, grammar)

Interestingly the paper (I finished scan reading it now) does say that the models can move data from their 'automatic' circuits into the J-space if they need to reason on it reflectively. So in some sense LLMs actually can access their own vectors, at least some of the time.