Cuda also compiles to PTX, which makes it much easier to distribute and therefore also easier for users to actually use. Doesn't matter that much when you're writing code for specific hardware like MI300X, but it's part of the developer story.
Frankly, I find the attitude towards AI coding here on HN to be both disappointing and a bit disgusting. Not long ago places like this where software developers gathered were full of various texts about how important it was to be able to reason about your code, how tech debt crept into your projects, and how skillful you had to be to write good software, various smart algorithmic tricks to squeeze more performance out of your hardware, etc.
Now? Seems like code quality is outdated and uninteresting all of a sudden. Everything is about agentic coding, harnesses, paying hundreds of dollars to Anthropic to let their LLM do the coding for you or perhaps using a 128 GB Mac to run a local model. Do you know your code base? Doesn't matter, if there are any bugs in the future Claude will fix them! Tokenmaxxing is the new paradigm, who cares about the end result as long as it's runs for now and passes all (AI written) tests!
But don't suggest these people shouldn't get $100k+ salaries, after all, they still "software engineers" in their minds, they're running the agent orchestration harness in the terminal after all, not everyone anywhere in the world could do that! They're special and deserve to be well compensated for their hard vibe coding work!
I don't understand why the height and weight errors aren't 0 when they are known inputs? If I say how tall I am, why is the model estimating something else?
That's a common phenomenon in model fitting, depending on the type of model. In both old school regression and neural networks, the fitted model does not distinguish between specific training examples and other inputs. So specific input-output pairs from the training data don't get special privilege. In fact it's often a good thing that models don't just memorize inputt-output pairs from training, because that allows them to smooth over uncaptured sources of variation such as people all being slightly different as well as measurement error.
In this case they had to customize the model fitting to try to get the error closer to zero specifically on those attributes.
Yes, but why are they estimating the features when they are already available? They can estimate the other measurements from height etc, and just use the known inputs as is. I don't get the point of passing them through a model at all.
The previous response was exactly right. The estimated features are impacting height, so the height can't be set then do the rest. It also cannot be tuned afterwards because it would change the mass. So vicious circle.
A version of this comment is posted in all submissions about Low Rank Adapters. I don't see how "Learning to reason in 13 parameters" would apply to low power radio communication, so it's even less relevant this time.
> Please don't complain about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage. They're too common to be interesting.
The cropdetect example made me wonder if they're thinking about including support for yolo or similar models. They're including Whisper for text to speech already, I think yolo would enable things like automatic face censoring and general frame content aware editing. Or maybe Segment anything, and have more fine grained masks available.
On the other hand, when I compared the binaries (ffmpeg, ffprobe, ffplay) I downloaded the other day with the ones I had installed since around September, they where almost 100 MB larger. I don't remember the exact size of the old ones but the new ones are 640 MB, the old ones well under 600 MB. The only difference in included libraries was Cairo and the JPEG-XS lib. So while I think a bunch of new ML models would be really cool, maybe they don't want to go down that route. But some kind of pluggable system with accelerated ML models would be helpful I think.
reply