Maybe it’s just me but (and I ask this as something of an ML practitioner myself) can anyone at Anthropic or Open AI or any of the frontier labs work on summarizing content into a format where a human won’t immediately vomit?
There’s this sort of local optima all these models pick which is instantly recognizable and hardly digestible for human consumption. Perhaps too much internal RL against benchmarks during chain of thought? The older non reasoning models had they’re own problems but at this point I can’t get an LLM to summarize data for humans, which is troubling.
As much as I’d like to read this piece it follows suit and I don’t have the patience to try and extract anything valuable from it.
what do you think these things are trained on? mountains of rubbish, human and likely also AI slop. how many bad movies and ads do u think they are fed. poor things. i can see them caged in the corporate HQ, with their tails shoved down their own throats rigged up to the SlopExtractor9000.
its time for the UN to Unslop the world. or something like that :(
"write (whatever you want written), BUT do not make it sound like a typical AI don't use any of the tropes or typical phrases that AI tends to use - if you include anything that makes it appear to be from AI then the task has failed. Pull your writing away from the median and toward very specific random styles that are NOT the cliched, averaged LLM style."
edit: actually wrong I tried to get ChatGPT to do this and it kept falling back into LLM speak even when it recognised that's what it was doing.
yeah, no, I've tried a number of variants like that. for ex even if I tell it to not rely on its chain of thought created jargon it'll just usually continue to hallucinate adding in the jargon unless I line item them painfully and eventually get something a human can understand.
prompting it to be concise ("250 words or less") seems to help somewhat but theres only so much you can do esp in cases where the details matter and don't compress well
Before LLM, the majority of the web was content farms... but those are instantly recognizable and we seldom fall into them. Now LLM mixed the content farm style writing with _some_ real content. It take time and effort to tell they are slop, causing much fatigue.
I have no data to back up this claim but one hypothesis I have is it's gotten particularly acutely bad esp with newer gen models because of increased training on reasoning and chain-of-thought.
To borrow a programming lang analogy the failure mode I see a lot is it invents its own jargon that is effectively like a compiler intermediate representation of the high level natural language you actually want a human to look at and then inserts it directly into what the human has to read.
It needs to stop doing that but it doesn't seem to do a good job at differentiating from what is or isn't chain of thought slop. I'm sure the jargon is useful during its reasoning but it's very unhelpful and not very nice to deliver it to a human.
One of NetBSD’s strengths and emphases is portability. They have a motto- “of course it runs NetBSD.” IIRC NetBSD was ported to Apple Silicon slightly before Asahi Linux succeeded.
Another cool corollary to this ethos is the portability of their portage, pkgsrc. I’ve personally used it as the basis of my user space packaging system on a Linux From Scratch distribution I made during the pandemic and the results were awesome. I could be wrong but it is the only package ecosystem I’m aware of that currently supports MacOS, Linux and the BSDS.
As far as the health of the ecosystem pkgsrc uses a rolling quarterly release process, has built in mechanisms for identifying vulnerable packages and a dedicated (volunteer?) security team and packages are about as up-to-date in many cases as, say, arch or gentoo.
Additionally pkgsrc can run rootless and my understanding is it finds itself used in a lot of high performance scientific computation settings as a result.
Ported to apple silicon before Asahi is definitely a stretch, single cpu core, storage "barely" working with several hacks, graphics simialrly.
Iirc they still required the Asahi folk to do a lot of the heavy lifting for device tree bindings, meaningfully functional bootloading, etc.
Even now the situation there is still rather dire even by bsd standards.
They for that reason also were very very early in the Xen game both host and guest. Back when it was essentially a research project, why before hardware virtualization was even on the roadmap for Intel.
I think from what I can understand about Serenity OS and Ladybird from afar and the kind of Cathedral culture that Andreas Kling values and feels Apple benefits from I'm not wholly surprised that the development of Ladybird took this course.
What I am curious about as someone who has been kind of cheering off on the sidelines is if there's any way that folks could get involved still in the future or if this is in practice permanently a closed project?
BSDs are more cathedral style and getting maintainer status is usually pretty onerous from what I understand but there are at least routes to it available to people willing to make an appropriate level of investment.
I'm not at a point in my life where I can meaningfully provide that kind of time and energy into serenity or ladybird but if my circumstances changed it's the kind of open source project that I would love to dedicate my time and energy towards in the future and I'm sure I'm not alone in feeling that way.
Has the Rebble community ever explored their own open source HW for the rebble ecosystem? I know there’s a ton of work involved to get something high quality/consumer grade and there’s obviously cost implications correlated to order volume and we were all hoping Core Devices would offer the goods but maybe we can lean into a community driven model for the hardware as well?
I'd be surprised if more 'hackable' watches didn't pop up around the Sifli chips. Lilygo have an upcoming device with Sifli 52 chip. There's the SF32LB52-ULP smartwatch development board.
The Silfi 52 chip is news to me- thanks for pointing it out (also didn't know it's what is powering the new Core Devices products- pretty cool).
I've built custom firmware for a DIY OLED ESP-32 watch that is made by a few vendors before. In some ways we're emerging into that reality now but I'd admit that what Core Devices is trying to do and the general level of polish of the Pebble ecosystem is a lot further than something like what I'm describing.
Ironically, I don't think glibly remarking that you "still have all your limbs" and some handmade furniture at the end properly demonstrates someone "fear[ing] the saw," and it demonstrates some of the hubris we're seeing in current tech culture.
One of my high school teachers impressed the same caution upon my cohort but was missing the end of a finger.
I have Susan Kare's original finder iconography tattooed on my body. I don't have much opine on the new design language but I do think the new finder icons displayed in the post are an abomination.
I had the pleasure of meeting Dan in person a few times up here in the Bay Area. He was incredibly approachable and always generous with his time. If he sensed your curiosity, he’d give you his full, undivided attention.
Just weeks before he passed, we were trading long Twitter DMs late into the evening—deep, technical conversations spanning topics that were hard to get good information on elsewhere.
After his passing, as I began sharing these stories, I found that so many others had experienced the exact same generosity from him. He had a remarkable way of making people feel seen and supported.
I thought it was funny they spent so much time bashing choice of distro and highlighting various performance considerations only to pick Alpine?
Alpine linux uses musl as its libc, which contrary to the articles' claims (unless I'm missing new information?) can have severe performance implications in many production settings.
I think this is true if there aren't new questions to be asked. But technologies shift and evolve all of the time.
One of my top StackOverflow questions for years was around the viability of ECMAscript 6. It's now essentially irrelevant because it's found wide adoption in browsers etc. but at the time a lot of people appreciated the question because they wanted to adopt the technology but weren't sure what its maturity was.
It's also true that some technology stacks mature to a point where there isn't much more to be asked but I think there will continue to be a place for forums of discussion where you can ask and get answers around newer, bleeding edge technologies, use cases etc.
> What do you do if you need to look up the definition/implementation of some function which is in some other file?
At some point, for me, ‘find <dir> -name “*.ext” | xargs grep <pattern>’ took over for recursive grep, because the required tools are available on most Unix systems.
There’s this sort of local optima all these models pick which is instantly recognizable and hardly digestible for human consumption. Perhaps too much internal RL against benchmarks during chain of thought? The older non reasoning models had they’re own problems but at this point I can’t get an LLM to summarize data for humans, which is troubling.
As much as I’d like to read this piece it follows suit and I don’t have the patience to try and extract anything valuable from it.