ghostpepper

> I don't really care if someone can chain together zero-day exploits to "pwn" my lightbulbs.

The point of pwning your lightbulb is not to decide whether it's bright or dark in your house. It's valuable as a residential proxy with a good IP reputation, to be sold to the highest bidder who wants to do something they don't want traced to their own IP.

Ironically very similar to LLM output.

Sorry, which backdoor(s) exactly are you referring to?

The theory is that open models cannot follow regulations and will therefore be banned or not eligible for many of the large contracts that the frontier labs will win.

I like the part where Frog and Toad are held accountable instead of blaming the machines they built.. oh wait

Any time someone shows me something an agent did, the first question I ask is whether there's plausibly something comparable already in its training data.

OpenAI has 5 hour limits on the $20 plan. I agree about cybersecurity refusals though.

Or the industry standard will change

It reminds me of when people will use filler words because they haven't actually decided what thesis they're going to commit to yet. Just-in-time thinking.

One thing they could try eating is glue on their pizza, helps the cheese stick to the toppings

They don't generate keyword search queries very well. They can overcome this by brute force but if you watch what they search you will cringe.

nhl toronto scores nhl hockey toronto scores "nhl hockey" toronto score today nhl "hockey score toronto" "hockey" who won toronto

etc.

Somehow being good at semantic search makes them bad at keyword search, for whatever reason.

it's not just you, there is research that people are starting to talk more like chatGPT. you have a gift for parody though.

Your argument in favour of TUIs is that they can be, in fact, not TUIs and actually CLIs

why is that staggering? I think we're at a point now where the use of AI is not really much of a quality signal in either direction

Did not read the paper so apologies if this is covered but isn't it possible that there is some recognizable semantic pattern in the training data where an "aha" is often followed by a subtle semantic shift that proves closer to the original premise in some critical way, and by emitting the "aha" token the model causes itself to produce such a subtle semantic shift that pushes the subsequent reasoning closer to the desired response?

They finally killed the `old.reddit.com` subdomain for logged-out users this week, which was the final push I needed to stop even idly browsing. I've been using the site since at least the digg exodus in 2010.

in my experience, gemini is easily the most grating, condescending, stereotypical LLM voice between opus/fable, codex-5.6, glm-5.2, etc

I believe this is also why 911 doesn't show up in iOS/Android call logs

I guess this is a command in the GUI app? I still don't see it in the latest official codex CLI app (0.144 I think)

maybe I'm misunderstanding but I don't want to sign into the app on a mac - I want to run the CLI on a headless linux server and control the sessions from my iPhone. Does Codex allow that now?

I am curious about the claim that the subscription models are different. Has anyone benchmarked this?

I have used claude and codex extensively but only from their CLI app (heavily sandboxed using rootless podman, network filtering, etc), so I don't really know what I'm missing with the GUI apps.

One killer feature that Claude has, and AFAIK Codex still lacks, is the ability to start a session in the terminal and then hand it off (actually just remotely control it), from the iOS app.

Last time I tried Codex on iOS it required a ton of set up to link a github project etc. The way claude lets me remote into a session I've already started on my actual machine is much better IMHO.

This is also the dichotomy I'm seeing. This meme sums it up https://fixupx.com/danhockenmaier/status/2021617680525172840

I like to imagine a similar dynamic happening when previous transformative technologies were invented: when the power tools like the chainsaw were invented, I wonder if there was a cohort of carpenters that dismissed them because of how much damage they could do in unskilled hands.

what I was taught (and what still works for me) is to look out the front window, never the sides, and pick a as far away (ideally on the horizon) to focus on.

the theory being, at constant velocity in a straight line, your body feels at rest, so you want to look somewhere that reinforces that. looking out the side window has scenery rushing past, which is the opposite.

turning sideways and reading sounds like a nightmare.

I am a big fan of computer science education for kids that uses a light touch, or even completely offline - things like Computer Engineering for Babies, Turing Tumble, or any game that introduces basic concepts with intuition as opposed to showing the final product of decades of abstraction.

I'd love to hear more about the STEM week one, if there's anything online you can share. What's the target age range?

which quants of 3.5 vs 3.6 did you compare? I guess you're saying that whatever quant you were using, going one lower was worse? ie. 3.5 Q6_K at 22.5GB versus 3.5 Q6_K at 22.9GB?

Supporting everything forever is how you end up with Windows

would it blow your mind if I told you there are already idiosyncrasies in the way you write? not you specifically; everyone has them