Klathmon

I feel like we have the technology to prevent a car from starting if the driver is significantly impared, but "solving" addiction, cancer, and heart disease are much harder problems.

Obviously it would need to be implemented carefully, but I personally would be more than fine blowing into a tube to start my car if it meant saving 12k lives a year with a very low false positive rate.

That being said I know it would never be implemented in a sane way in the US, and you would probably have situations like your car insurance automatically increasing due to a faulty sensor, so I'm ultimately against it unless a lot of other stuff changes.

But it's still much more possible than "preventing drug addiction"

If they're using a static threshold of 200 views and there aren't many of these left it could pull them all out of that category.

It could obviously be fixed but I was more curious than anything

How many of these images are there? I cycled through a few and ended up hitting at least one duplicate [1] that I do actually enjoy (but I can't find the name of it now that I refreshed the page, I know it had the verrazzano narrows bridge in the title)

Is there a chance a site like this could ruin their metric by inflating all the views for these lowest viewed items? Or do these not count?

[1] https://www.artic.edu/iiif/2/67395c18-c83c-865f-b0db-4736574...

AES is so cheap with hardware acceleration that I could see an argument that it prevents casual sniffing of the traffic. Personally I think the false sense of security would outweigh those benefits, but I'm not in their shoes.

But DES is so broken that it's more of a giant flashing beacon saying "look here there are terrible decisions being made!"

Yeah it wouldn't have changed this specific case but still why use it? It's slower even without AES NI!

Is there any valid reason to still be using 3DES in 2026?

It was formally deprecated in 2018 and has been surpassed in just about every single way by AES long before that.

At this point I feel like it's use is such a huge red flag

I also have no idea what I'm talking about, but to me this seems closer to the "caveman mode" that some people use to compress info into fewer tokens. Going through the image tokenizer allows you to leave the source text untouched while still gaining (some of?) the benefits

That's like saying quantization isn't real because the frontier labs aren't using it in their production inference.

This is a lossy process, it produces worse results. It might be worth it for some situations, but applying it to everything would just be making your SOTA model worse

But it's not just him, it's the entire party aggressively supporting him and everything he does.

There are also cases where breaking a "rule" is the right thing to do.

I've had several instances where I told the model to do something that was accidentally impossible if taken at face value. The most memorable one is when I told it to re-run just a specific CI job, but it didn't have any way to do that, so it just ignored that part of the prompt and re-ran all CI jobs by pushing another commit.

Ultimately I preferred what it actually did, but technically it violated what I told it to. I have a feeling in a benchmark that would be points against it

I've been pretty damn happy with codex and vscode.

Between the codex app, cli, and vscode extension there are options for most ways of working

This is what I'm advocating for.

Give each dev's AI agent its own identity with its own access controls and tokens and everything.

It helps solve both the access control and attribution issues

I don't see 'grandma' building here own calendar app via Claude Code that reminds her of the family birthdays.

If you think of apps in the traditional sense I think I agree with you, but I have a feeling things are about to become a lot more messy.

Grandma might not even know she's building her own calendar app.

I don't think we are that far from being able to ask a general purpose AI to "help me not forget my family's birthdays" and it creating and maintaining code for that purpose. Not quite an app, but more than a one off script, I think AIs are going to unlock this weird situation where they're running a bunch of barely organized code almost as an extension of thinking.

In our internal metrics you can see a clear increase in PRs and CI runs in general that tracks with agentic coding adoption, and it's significant, so I absolutely buy that GitHub would be struggling to take the brunt of that without big changes

There have been a few instances, IIRC there was an Apollo mission that had a head cold spread among the whole crew.

But that's unlikely to be the case here because they've been up there isolated for over 6 months now

And if you like it, picking up any QMK or ZMK compatible keyboard would let you do it in firmware too!

Should you never use a calculator because you want to keep your math skills high?

There are a growing set of problems which feel like using a calculator for basic math to me.

But also school is a whole other thing which I'm much more worried about with LLMs. Because there's no doubt in my mind I would have abused AI every chance I got if it were around when I was a kid, and I wouldn't have learned a damn thing.

For code modifications in a large codebase the problem with multi-shot is that it doesn't take too many iterations before I've spent more time on it.

I've found voice input to completely change the balance there.

For stuff that isn't urgent, I can just fire off a hosted codex job by saying what I want done out loud. It's not super often that it completely nails it, but it almost always helps give me some info on where the relevant files might be and a first pass on the change.

Plus it has the nice side effect of being a todo list of quick stuff that I didn't want to get distracted by while working on something else, and often helps me gather my thoughts on a topic.

It's turned out to be a shockingly good workflow for me

If the LLM is able to handle it why do you need to maintain those specific skills?

Same, it's nice to see my username on the leaderboards.

Even though all I did was setup the docker container one day and forget about it

I've found most models don't do good with negatives like that. This is me personifying them, but it feels like they fixate on the thing you told them not to do, and they just end up doing it more.

I've had much better experiences with rephrasing things in the affirmative.

PHP's approach is simple though, and in my experience that simplicity pays off when you do start scaling the systems.

In other systems once you get beyond a single machine you need that external communication mechanism anyway, and now you have multiple classes of comms which introduces bugs and complexity and performance cliffs.

In PHP you just throw another server at it, it'll act the same as if you just added another process. Nice linear scaling and simple to understand architectures.

Man I miss working in PHP

Given it's passing retrograde (is that even the right way to say that?), would that make it easier to catch up and intercept?

Assuming you don't want to do anything but fly by or smash into it

Our group once badgered our DM at the time into allowing the parties pet goat to deal some minimal amount of damage in combat. Then we backtracked and bought a hundred of them from the local shepherd and had a small goat army for a bit.

Unfortunately there was a flood shortly after and our goat army was lost

We propose adopting diffusion language models for text embeddings, motivated by their inherent bidirectional architecture and recent success in matching or surpassing LLMs especially on reasoning task.

I didn't realize diffusion language models were at this point yet. But what's the catch? why aren't diffusion models (or some kind of hybrid) taking over?

LLM autocomplete is an entirely different beast.

Traditional auto complete can finish the statement you started typing, LLMs often suggest whole lines before I even type anything, and even sometimes whole functions.

And static types can assist the LLM too. It's not like it's an either or choice

So is the big improvement here simply skipping the unembedding/embedding step for internal thoughts? Or is it mainly in the training methods to teach the CoT and how to switch between "latent thought" and text output?

It's really interesting that a fixed number of "latent thoughts" performed as well as a binary classifier! I didn't expect that at all, the way OpenAI talks about CoT it seems the ability to let it "keep thinking" let's them continually score higher on benchmarks while throwing eye watering amounts of compute at the inference.

As someone who switches between macos and windows on a daily basis, custom keyboard firmware has been such a game changer!

QMK can auto detect the OS and can easily remap some keys based on it, and since it's baked in the firmware it'll work with any program on any machine I plug it into!

If you do ever want to dive into the keeb firmware hacking world, you absolutely can build very silent mechanical keyboards.

Between linear switches, quiet mounting options, and even o-rings if you have a tendency to press keys harder, you can get them to being only slightly louder than a MacBook keyboard (or maybe not louder, just different sounding)

I've been loving Copilot Edits (their take on the multi file edits stuff). I'm personally still doing the process of finding and adding files to its context/working set (letting it try to do that part just doesn't work great yet), but then it quickly applied edits and gives you a review/checkpoint UI.

The UI is still changing slightly every couple weeks as they improve things and polish it, but it's become a big enough part of my day to day that it's pretty much always open on the right pane of vscode for me.