@Jamey Sharp
An advocate of nuance and historical context
"not the rockstar you want, but the janitor you need"
Relays (1)
- wss://relay.ditto.pub – read & write
Recent Notes
Big software companies have always wanted to use browser history surveillance for their own purposes and I think that makes us so disgusted with the concept we don't think about how powerful it could ...
@nprofile1q... There's so much more we could do in a purely-local fashion with browser history today, even without LLMs. Several years ago I discovered that (at least in Firefox) the browser history records referrer information that I think isn't displayed anywhere in the UI. I made a quick proof of concept WebExtension that offers a sidebar which shows you, for the current page, which pages you've visited it from, and which pages you've followed links to. I think people more clever than me could find all sorts of user-directed applications for that information. https://github.com/jameysharp/contextual-history
someone apparently made a metal 4 to vulkan translator mature enough to render a MacOS 13 VM at 60fps on a Linux host "without writing a single line of code"
and I'm skeptical because I don't think t...
@nprofile1q... I suppose the inverse thing (https://github.com/KhronosGroup/MoltenVK) is in the training data, so maybe that helps? I dunno though
I'm thinking perhaps the fact that it uses a dedicated WASM also explains some of its gains... because it likely gives the JIT an easier time proving more memory accesses to be in bounds.
regardless ...
@nprofile1q... I'm confused about why the wasmfft approach might make it easier for the JIT to prove bounds-checks redundant. I could see that if the wasm was specialized for different FFT sizes or something (which seems like it might be an interesting thing to try), but as far as I can tell it's the same wasm every time it's instantiated. I don't know enough about browser JIT, so I'm curious, what leads you to think separate instantiations might be helping performance in this case?
btw I love your benchmark page for this!
btw I love your benchmark page for this!
yeah very cool but what about WebUSB :neofox_cry: serial is nice but not all USB devices are serial
@nprofile1q... did you already see https://github.com/ArcaneNibble/awawausb, a WebExtension implementation of WebUSB for Firefox? obviously native in-browser support would be better, but this is something, at least. I haven't tried it out yet but I may need to soon
If there was a Fedi-Dating platform...
1. what features would you find most valuable?
2. what features do you like from existing dating sites?
3. and what features do you dislike from existing dat...
@nprofile1q... What I think is particularly interesting about a fediverse approach to dating is that I've never seen a dating site use the existing social connections between people to help. For example, I want to write public recommendations for other people ("hey, I think this person is cool and here's why you should date them"). Horrifyingly, I think the easiest way to describe what I mean is "LinkedIn for dating". (Maybe Fedi-Dating and Fedi-Hiring should be the same protocol, but with different profiles?) I think it's often easier for people who know you to talk you up than it is to promote yourself.
That said, safety is already difficult in both social networks and dating sites, so identifying the right safety features for a combination of both is particularly intimidating.
That said, safety is already difficult in both social networks and dating sites, so identifying the right safety features for a combination of both is particularly intimidating.
@nprofile1q... I knew about QPSK but hadn't heard of this "offset" variation of it. How do you do carrier recovery for that? Does a Costas loop still work?
I have a hypothesis about device drivers that I haven't quite worked out how to test yet.
First, I suspect many drivers would benefit from model checking as a particular kind of formal verification; I think that might be particularly good for eliminating hangs and race conditions, especially on cancelation and error paths. (I'd appreciate pointers to people who've done this; somebody must have tried, right?)
Second, I think the models to be checked can be split into an OS-behavior model that's generic across a whole class of devices, plus a device-behavior model that's OS-independent, as demonstrated by the Termite research project (https://trustworthy.systems/publications/nictaabstracts/Ryzhyk_WKLRSV_14.abstract). Termite 1 leaned fully into program synthesis to completely automatically generate device drivers; Termite 2 came from the observation that human judgment was still necessary for quality but that synthesis could be used as a specialized auto-complete; and I'm suggesting that even the auto-complete isn't necessary. Formal verification alone would be a huge improvement over the current state, and I think existing tooling is closer to supporting that goal on real drivers. (Maybe spin and modex can do much of this today? I also considered CBMC but I think some support for temporal logic might be important.)
Third, I bet OS-independent device models would be useful for capturing and validating beliefs formed while reverse engineering closed-source drivers. I envision tools for checking that a register trace conforms to the model, or model-checking a reverse-engineered binary.
Fourth, I hope those OS-independent models could be shared between well-established operating systems like Linux and the BSDs, while also enabling new experimental operating systems to get up to speed more quickly. I feel like the biggest barrier to building new operating systems is the lack of device drivers, but sharing C source code between different kernels is super messy because OS-specific design decisions get mixed in, the result is hard to review or maintain, and of course I'd prefer not to be limited to C. If sharing is at the level of "this is how this device works, do whatever you want with that", but it's machine-checkable that each driver matches the model, I think that might be the best possible option for sharing this knowledge and effort.
I'd love pointers to related work or just discussion of pros and cons of these thoughts!
First, I suspect many drivers would benefit from model checking as a particular kind of formal verification; I think that might be particularly good for eliminating hangs and race conditions, especially on cancelation and error paths. (I'd appreciate pointers to people who've done this; somebody must have tried, right?)
Second, I think the models to be checked can be split into an OS-behavior model that's generic across a whole class of devices, plus a device-behavior model that's OS-independent, as demonstrated by the Termite research project (https://trustworthy.systems/publications/nictaabstracts/Ryzhyk_WKLRSV_14.abstract). Termite 1 leaned fully into program synthesis to completely automatically generate device drivers; Termite 2 came from the observation that human judgment was still necessary for quality but that synthesis could be used as a specialized auto-complete; and I'm suggesting that even the auto-complete isn't necessary. Formal verification alone would be a huge improvement over the current state, and I think existing tooling is closer to supporting that goal on real drivers. (Maybe spin and modex can do much of this today? I also considered CBMC but I think some support for temporal logic might be important.)
Third, I bet OS-independent device models would be useful for capturing and validating beliefs formed while reverse engineering closed-source drivers. I envision tools for checking that a register trace conforms to the model, or model-checking a reverse-engineered binary.
Fourth, I hope those OS-independent models could be shared between well-established operating systems like Linux and the BSDs, while also enabling new experimental operating systems to get up to speed more quickly. I feel like the biggest barrier to building new operating systems is the lack of device drivers, but sharing C source code between different kernels is super messy because OS-specific design decisions get mixed in, the result is hard to review or maintain, and of course I'd prefer not to be limited to C. If sharing is at the level of "this is how this device works, do whatever you want with that", but it's machine-checkable that each driver matches the model, I think that might be the best possible option for sharing this knowledge and effort.
I'd love pointers to related work or just discussion of pros and cons of these thoughts!
the more I'm forced to interact with gcc, the more I wish we all finally moved to LLVM
@nprofile1q... if we're only picking one, I want to vote for Cranelift. (I fully expect to lose that vote.) anyway, yeah, I can understand that feeling. I'm curious if there are specific things that are causing you grief recently? I almost certainly won't have solutions, just interested in hearing about it
I just threw together a website visualizing this ICE contract data! You can browse through the companies and their contracts, and filter them by state https://micahflee.github.io/ice-contracts/
@nprofile1q... Oh, I see this is from all of DHS, not just ICE, and also that it includes contracts which ended years ago. That's good to be aware of. I actually worked on part of this one contract, over a decade ago, and it was purely about improving cybersecurity for people's smartphones; it had nothing to do with immigration. https://micahflee.github.io/ice-contracts/?state=OR&modalType=contract&modalAwardId=937
@nprofile1q... Did you get the answers you were looking for? I'm not a writer but I've been studying web-serialized fiction (webcomics and prose) and talking with authors for a couple of decades, so depending on your questions maybe I can help some
@nprofile1q... I forget who shared it but I recently saw a post recommending a book called "The Authoritarians", available from the author at https://theauthoritarians.org/. I read it and I think it actually does a pretty good job of answering your question about "How do they ever wash the taste of boot out of their mouth?"
@clacke: exhausted pixie dream boy 🇸🇪🇭🇰💙💛 Oh, if you control your data sources and can dictate JSON, I guess that works too. I think my confusion is because I'm missing the context on what you're using jq or jinja for here, but that's okay.