_kulang

Not my claim, what I am saying is that as long as there is a physical connection between these domains we can’t be surprised about universal representations. In a spatial Fourier transform, spiky things have high frequency information. In a time Fourier transform, spiky sounds have high frequency information. In space, spiky things have sharp edges (attack). In time, sharp sounds have faster attack. The presence of a physical analogue here means that all these experiments are testing is whether brains identify this existing connection.

It would be more surprising if the connection was indicative of genuine universality of brains and their representations, rather than ability to identify a cross-domain mathematical/physics-cal connection.

I think programming was a proxy for understanding, that is, only by understanding a system well could you produce a good program. The actual act of programming is arguably like speaking. I wouldn’t call speaking a skill, but oration is, as is speech writing.

Assessing programming in a uni environment was effective because it was a proxy for understanding and implicitly tested things like design and debugging capability. Perhaps teaching programming required so much repetition because we are not so equipped to teach the actual skills. AI accelerates the pedagogical concern because it can generate programs without the user understanding, so assessing programming no longer by proxy assesses actual skills like design and debugging.

Sure it is. Fourier transform - Bouba and Kiki are far from the same there. And as someone else mentioned, time domain content - attack and decay. There are physical differences in the sounds which are pretty fundamental and can be definitionally linked to sharpness and softness across domains, especially as sharp objects are defined in nearly exactly the same way (except spatially).

It would be more interesting if we had two sounds which didn’t have such apparent objective differences giving rise to specific and universal subjective differences, or if something that was not so cross domain elicited a different response, e.g. linking volume with sharpness as opposed to something like size.

I guess in embedding space it would be white noise? Perfect compression would result in something indistinguishable from white noise. Or at least this is my basic understanding, that degardless of data perfect compression looks like white noise

As someone who is writing a university course that is allowing the use of AI, I think that the value of an in-person education is underestimated. I reckon that a solid chunk of what you are taught in a good university environment is soft in some sense, and that technical knowledge is less valuable than it was as compared to technical ability. The knowledge problem is what MOOCs (and any open resources) solve. I am not sure that technical ability is fostered by MOOCs as much as it is merely enabled by them.

They likely also want to have the users generate more data. They are paying people to use the subscriptions essentially, as compared to API billing.

Thank you, it reads like a psycho hazard

What’s the difference between sloppiness and fraud when you stand behind the paper all the same? I would argue that sloppiness in academia is a form of fraud. It’s one thing to be bad at writing, and another to have poor methodology

At some point during this progression, the poster lost most of the “AI-ness” and I didn’t get such a visceral reaction to it. Cool!

What’s wrong with milliwatt hours?

You’d have to live there, in the presence of a data centre for AI which actually wastes water. The numbers people throw around on social media are absurd. I’m pretty sure they do not take into account the circular reuse of water for cooling, and just look at flow rates passing over chips.

Why argue with it? Why not just edit the memory yourself?

I have quite a bit of success with Claude on Jira tickets as long as the tasks are scoped well. Put effort into the ticket, and fire Claude at the ticket with MCP. I essentially review the PR.

But this is only possible because I have created a completely new python development workflow with fixed formatting, style, and unification of tooling around uv. We’ve also split our projects into smaller repos, some which function as libraries and some which are applications depending on those libraries.

I think the key to success here is to limit what context is required to do development. I’d say it’s slightly more on the extreme side of “modular” than a codebase I would write otherwise

It’s the theory of colour charge essentially, counterpart to electrodynamics. Quarks have colour, rather than electrical charge. Mediating particle is the gluon, analogous to (but entirely different from) the photon

Sure – but I’m entitled to make a judgement on what I consider to be “good” and “bad” output from an LLM, where “good” just means “helpful to my process”.