BearOso

I think part of this is the mach bands illusion. We perceive some darkening at the transition between shades and think we have to replicate it in rendering. But we really don't because the brain creates the effect, so artificially adding it makes it really exaggerated.

Rust does this automatically.

A garbage collected language does this automatically. Rust still requires thinking about and tracking memory lifecycles, but the borrow checker will complain and keep you from doing it wrong. That's why LLMs like Rust. It gives immediate feedback on what to fix. By-default constant reference parameters helps prevent major performance problems.

It's been double-stable since early 2025 (version 3.2), so it's weird that Ubuntu doesn't have it when even Debian does. I've seen a couple projects holding back on the upgrade because of that.

With the way the software ecosystem works now, a lot of projects just include a copy or submodule because it's self-contained and a good candidate for it. It dynamically loads most dependencies, so it's pretty light-weight, considering. The whole static linking licensing problem is gone now, too.

The article is from Bain Capital. Of course they're going to ignore the obvious reason and blame something else. They raised prices more than people are willing to pay. Unlimited profit growth can't exist.

Nedry doesn't have a camera on his end, and he's talking through a headset. Maybe guy at the harbor just knows where the closed circuit camera is? Since it's just quicktime anyway, I guess we'll never know what they intended it to be unless someone asks the director.

I think Fabien is misinterpreting the part which he sees as "video conferencing." Nedry is talking to a guy on a regular phone line. He's just watching a security camera feed from the dock where the guy is.

Yeah, this particular experiment doesn't reveal much because the Nvidia driver is a black box. We know in the past it used its "composition pipeline" in Wayland situations, and that was a particularly noticeable increase in latency. Plus, for a while the Nvidia driver also was stuck on software cursors. And even on the DRM path the big two compositors only updated input once every frame, which was the reason for it "to have much worse input lag."

edit: I should also point out the mouse acceleration curve, which if you don't fix it is different between X11 and Wayland compositors. That really messes up the "feel" of things.

Their internal representations are so cryptic and complex that even the top AI researchers don't really know how they work or what their limits are. No one is going to take you seriously as a rando HN user if you're claiming to know better than them.

We know exactly how they work. When we say they're impossible to analyze, i.e. for particular traits like this, it means that the data model is so big that tracing it would be logistically impossible because of the scale involved and time constraints.

For comparison, suppose you tried to analyze all the nooks and crannies of the Amazon watershed to find out why a particular rock appears at the delta. You could follow it back to the exact tributary, but it'll take forever, and is it worth the effort when you're going to start from scratch with the next rock?

It's marketing. The big AI companies know what they're doing and are trying to drive adoption with FOMO. Some people bought into the idea early on and feel self-important. Meanwhile, using AI is so easy that it's tantamount to pushing a button, because that's the whole point of it: to make things easy to do. You can pick it up in half an hour, so it's impossible to "fall behind," unless you're holding stock when the bubble inevitably crashes.

I think putting it in terms of API pricing is oversimplifying disingenuously. Anthropic still hasn't pulled the rug out from under us, so I'm sure it cost a great deal of money once everything comes together, likely surpassing 1.5M. Summarily, they got the result faster, which a group of engineers couldn't do, but at a greater expense.

I posted on an older article that I thought it probably cost half a million in API pricing. 165k USD is a lot lower. I wonder what the actual compute cost was. When this first hit the news, Opus 4.7 was brand new and required 6x the compute power per user token vs 4.6. The article says they were using Fable, which is way more expensive.

That's probably in an overscan area on most TVs, so it might not even show up.

Lenovo has an incentive to say prices will never go down. They need to continue selling right now when prices are high, but customers will wait if they have reason to think they'll go lower.

Backticks ` ` are used to trigger code-style formatting in markdown. Since all the coding AIs are trained almost entirely on github, you see it everywhere in their output. They'll use it even in places where it's supposed to be just plain text, and almost always in git log messages, which by spec have no formatting.

Slack uses this? Thanks for the info. I've never been forced into Slack, so I wasn't aware of that. Either way, it's stupid to use in places it doesn't induce formatting.

Yeah. It simply shouldn't be falling back to a software cursor in the first place. That's the bug.

I'm seeing lots of bots/agents misusing backticks like this. It's because people are writing agent plans in markdown and the agent thinks it's appropriate everywhere. Maybe you shouldn't be running all your comment replies through Claude.

You do realize that an LLM is an enormous decision tree? The prompt is the state and output tokens added to that are the subsequent state that's fed back into the machine. Running the math without deliberately adding any randomness would produce a deterministic output.

He's referring to accumulator styles, eg. the 6502.

In our small town, the local compliance officer is colluding with a tree-cutting company that has a contract with the city. They lie and claim trees are "dead" and they're constantly cutting them down unnecessarily at the taxpayer's expense. Pretty soon we're going to look like one of those new suburbs that are just concrete and grass, but with old houses.

We're in this situation because of laziness. I doubt doubling down on it is the solution.

Avoid the backtick quotes, too. Claude also mistakenly uses them outside of markdown.

We had less intrusive ads. Google AdSense was originally text-only and small-sized ads. What happened? It became far too lucrative to switch to the annoying ads by taking advantage of their trusting user base. Ad companies have only one service: make money. They will all do the same thing when able to.

Alternatively, you could have downloaded an existing one. What you did is just a roundabout way of acquiring one of the many WASM Rust GB emulators in the training set. Except it probably now has several problems from reconstruction. You didn't learn anything new from the process. It's just wasteful in several ways.

There is a fallacy here in assuming there's 256 steps from 0 to 255. That's not true, there's 256 values that can be represented in 8 bits, and 255 steps (spaces between those values) from 0 (black) to (255) pure white. Thus, the division by 255 isn't problematic. Of course, 128 isn't half grey, it isn't in 0-255 and quantized 8-bit values are almost always in sRGB, not linear perceptive space.

This is the same kind confusion that happens with sampling positions in modern APIs, where the location is specified in coordinates and not in pixel centers.

I've noticed as well. A lot of pull requests are just agents running constantly, hoping to have produced something of value. Entropy is at an all-time high, though.

Going from Opus 4.5 to 4.7 secretly required 6x more compute to run. 4.8 is apparently 30% more on top. I haven't seen any optimizations lately aside from distillation. Nobody's optimizing, they're just scaling up.

I agree. The pope seemed to take the opportunity to talk about the ethics in good faith, no pun intended. But Olah just used the association to aggrandize AI for marketing sake.

That line comes across as a wink to investors. They aspire for AI to displace human labor, as does he. Reading between the lines, it just confirms business as usual, and consequences aren't even worth thinking through.

What novel data hasn't already been used in training? What new algorithms are there? Can you post some links so we can read about them?