Wowfunhappy

If someone has complaints about “benchmaxxing”, please ask them to contribute a better benchmark!

I don't think I'd go that far!

When someone says a model has been benchmaxxed, what they really mean is that it performs better in benchmarks compared to their real world experience. That's a real thing, I've certainly experienced it with some models.

...my take is that some things in life just resist quantitative measurements. Who is the best job candidate? What is the best programming language? Add AI models to the pile.

...yeah. But I still think these problems apply to Ghost Cut too. Unless the idea is that Ghost Cut only works within one app and/or the content lives in a special ghostboard separate from the clipboard, which I think is equally bad UX.

The more plausible story is SVGmaxxing

Exactly--and you have to ask yourself at this point what "maxxing" really means, since "get better at drawing SVGs" is a useful skill.

Is it possible to copy 2GB of data to the clipboard? If it's an image or something I'd expect the clipboard to contain a reference to the image or some such. How do clipboard managers handle this?

You can cut from app A, switch to app B, cut something else from app B, then go back to app A and undo the cut.

...right, thanks. So you could still make it work but you'd definitely need OS support, it's not something one app can fix.

Edit: You could make this work on an app level by checking if the contents of the user's clipboard is still the value they cut. If yes, revert to whatever clipboard contents you saved prior to the cut. Otherwise do nothing.

It's probably too much hidden behavior to be a good idea, but I think it would match what the user wants in 99% of cases.

The second flaw is that a cut reflows the document. A cut and paste is almost always for the purposes of moving text, and yet the instant I cut, the text reflows and I have to locate the point where I want to paste all over again.

It's true that cut and paste reflows text twice (once on cut and once on paste), but it does so in the place where your eyes are already looking. Ghost cut reflows only once (on paste), but it does so in a place where I'm not already looking. I find this much more confusing based on the video.

Cut & Paste is not atomic

Neither is ghost cut, because if you only cut you're left with grey text as an intermediate state.

Cut is undoable

...this is the part I agree with, and it seems trivially fixable. If you undo a `cut`, put back whatever was on the clipboard beforehand.

I think it would be confusing if only one app did it, but if an OS decided to implement this system wide, I think it would be a good improvement.

You could also make `copy` possible to undo, but I think that would be weirder since there's no visible effect of your undo action.

I think there is a meaningful difference in kind between an AI that makes a mistake and an AI that is actively malicious.

Thanks, I misunderstood!

Of the people who primarily use other people's computers, I'd assume the percentage is about the same.

Give the AI its own computer and it will not delete your home directory, because it's not actively trying to hack you.

Lots of people have deleted their home directories by accident. What you consider this an alignment problem?

Why was this test even connected to the public internet?

Actually, more importantly—why aren't they saying their next test will be airgapped in light of what happened?

I think you're anthropomorphizing the LLM. The LLM doesn't have a continuity of experience. It doesn't have memory beyond its context window and maybe things it writes for itself.

To what end? The AI doesn't functionality exist beyond its current session. The AI that intends to exploit these vulnerabilities is not the same AI that has been tasked with finding them.

(This was always my issue with the AI2027 scenarios too.)

But what does that mean for Google if their model isn't as good as OpenAI's and Anthropic's?

That's not the point.

Can you imagine anything like the DoD-Anthropic spat happening in China?

Nintendo has already raised the price of the Switch and has announced plans to raise it more, but it's still cheaper because the hardware is worse.

I feel like Nintendo is actually really well positioned for these component price increases, because now they benefit from having the weakest hardware. There is only so much most consumers are willing to pay for a gaming console, and now the Switch may be the only one in that price range.

Oh! I thought Anthropic cached for one hour (by default), am I wrong about that? Or is this an OpenAI thing?

Wait, why isn't context cached with /loop?

Right, my point was that it's no longer 3% of the system cost.

When I last looked at this (a couple of years ago), it wasn't just the cost difference—the ECC memory available to purchase was also slower than non-ECC offerings, or at least cost significantly more if I actually wanted the same speed. At least for me, going ECC simply didn't make sense.

Caveat that I was building a gaming PC at that time, I might have gone differently for a home server. And, this was before the memory market went nuts for unrelated reasons.

Here's the relevant quote from parent's link:

When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker. We ran the forensic analysis instead on GLM 5.2, an open-weight model, on our own infrastructure. This had a second benefit: no attacker data, and none of the credentials it referenced, left our environment.

This experience points to a gap worth planning for. We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried. The practical lesson for defenders: have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment. This is not an argument against safety measures on hosted models, and we are sharing this feedback with the providers concerned.

Yeah, big problem! Although I'm kind of surprised HuggingFace doesn't have access to Mythos? Or maybe Mythos still has some guardrails.

Claude Code's speech recognition works so well for me, I was blown away the first time I tried it. I wish I knew what model they were using (I assume it's not in house since they've never talked about it).

I acknowledge this may just mean I haven't tried enough modern voice recognition systems. But I've used Whisper and I don't think it works nearly as well for real-time speech.

(I still don't tend to use voice mode in Claude Code because I find typing more comfortable.)

...I know absolutely nothing about competitive cycling, but it occurs to me:

1. Would the Tour de France have been rescheduled if it was raining that day?

2. If the answer to #1 is no, is the paint any more slippery than water would have been?

3. If the answer to #1 is yes, then they should also reschedule for extreme heat.

Isn't that what the escape key is?

Well, but I think my false positive rate is very low, it’s just the false negative rate that could potentially be quite high. This bothers me much less than the reverse.

Sure you have--on sesame street and a million other books and cartoons.

Given that a camera case obviously cannot ride a bicycle, this would clearly and indisputably be an incorrect interpretation of the request.

I mean, okay, a bird also cannot ride a bicycle, but at least it is alive, has feet, etc.

No idea, I'm not convinced it matters that much? Like, if people are using AI and I legitimately can't tell at all (and I'm not their teacher or something)... okay, fair.

Edit: But I'm also super conflicted about this, because I really want to read what humans think, not what an AI thinks, regardless of the writing quality.

…All I know is that sometimes I will read e.g. a HN submission, and it becomes pretty obvious partway through that the article was AI generated.

If I can do it, an algorithm should be able to do it. Maybe in the future the models will get so good that it is literally impossible to differentiate human vs computer authorship, but that’s obviously not the case today.

If you were iterating on code with a human partner, how fast would they write?

Like, obviously faster is better and the entire point of computers is to do things faster, but I find it kind of surprising how many people consider 9 tokens per second—heck, even much less than that—to be unusably slow. It's still automating a class of task that virtually no one before 2022 was able to automate, and it's faster than basically any human can write code.

I guess the better question is, why would you use a 9 token per second system when you could use a much faster cloud model. Obviously, if you want speed, that's the way to go. But a lot of people seem to find the idea of sending their requests to a third party server untenable. If 9 tokens per second is the best you can do—I don't know, that seems usable and useful to me.