chrisra

I think so. I guess the weights could be identical in function, but offset or rotated. I assume a retraining of Astra would yield numerically totally different weights even if functionality was identical.

I don't know if inspection would be practical. How do you inspect the weights for an intentionally hidden trigger that writes malware?

Might as well stick around until the option is no longer there, as long as they enjoy it.

I realized recently that I no longer get stuck. I don't sit and figure something out for a few hours, trying thing after thing. "Banging my head against the wall" isn't really part of my vocab anymore.

I've used something like this on a phone to tell another car something was wrong with their vehicle.

Man, I would love a table.

What type of applications are you working on?

I think a "reversed" south-at-the-top is a good idea to at least be aware of. I don't think it makes sense to reverse it since it just flips the problem, but it is (very slightly?) unlucky for the southern hemisphere.

Maybe this is silly: what about zero-width spaces or other hidden characters? Add them in post-processing, and the response quality doesn't degrade (though rendering etc. might be wonky).

I have the same feeling but don't really know why. Why do I dislike reading LLM text so much? Weird

Because they have extra money to do whatever they want with. And data to use it on? Not that that is sufficient, but it helps.

Loaded it on a pixel phone, saw a list of countries with no context about what picking one meant, saw a map but didn't understand how to interact with it, and that was about it.

And what would they do if you refuse?

It might be worth looking into probabilistic programming languages. I'm out of date, but I remember webppl, stan, anglican, pymc (a python library).

Seems worth an investigation and maybe mention on the article.

Decisions by small groups should be the default. Others only need to be involved if the risk/consequences of failure are high.

I started ignoring all PRs from our large team because we had a similar policy. My teammates can handle, they don't need me to check on each PR.

I recommend finding a real project that is likely to have some complex needs. It's easier to learn when there are real problems to solve.

In general, the harder it is to undo and the broader the consequence, the more you need to check.

Agreed. I enjoy looking at and using a lot of these components.

The first picture is of overcooked pancakes. Hard to trust the rest :)

If Imagen generates a digital painting after my prompt, I didn't make the painting.

I don't want them to say "I wrote it." They're still responsible, but it's helpful to know they weren't deeply involved in its creation.

I have no problem with my credentials being revoked everywhere before I know about a layoff. I don't really care how I learn about it, just please don't make me come in to the office.

What models are you using, on what type of codebases, with what tools?

I feel the same way. In general, I prefer working on a couch with my laptop. My eyes aren't great and I end up ruining my posture at a desk, invariably.

Their decision to... use AI for coding?

"...the system described here does it all automatically: It proposes optimizations using LLMs, formally verifies safety properties, shadow-evaluates against real production traffic, and hot-swaps improved code into running services without human intervention or service restarts."

Next up: proportional fonts and font weights?

> To increase the contrast of our sampling vector, we might raise each component of the vector to the power of some exponent.

How do you arrive at that? It's presented like it's a natural conclusion, but if I was trying to adjust contrast... I don't see the connection.

Maybe try finding a software engineering job at a place that also uses your scientific expertise. You may be able to find or create opportunities there. The position might not exist, but you can maximize your chances of something coming your way that needs both skill sets.

Which LLM was used to generate that post?

> the real challenge is standardization and integration

In what sense? I'm a newbie, but curious because I'm working on stuff related to https://mesastandards.org/mesa-der-std/.