Probably worth mentioning Cline, a VSCode extension that let’s you pair with a LLM of your choice right from your IDE. It produces diffs when you chat with it about aspects of your project which you can then merge in or reject for a different version.
For folks lacking context, this is a reference to this video: https://youtu.be/0mCsluv5FXA where the creator emulated DOOM using just TS types. It took 12 days to render the first frame of the game.
The thing that bottlenecks Doom (like 60% of total time spent by the type checker) is serializing the multi-megabyte-type to a string, which I bet is going to be much faster than 10x – because it’s not something a typical typecheck has to do at all. But honestly, even if it’s just 10x (look at us… “just” 10x! what a dream today is!!!), that’s still gonna get it down to sub-1-day for the first frame, more than likely.
I don’t mind some of these things on AI, but why are they always so darn long? I think it’s like 10,000 words or something. I kind of scrolled to the end and oh well. Deep fakes are most likely to kill us. I think the summarization might have really helped this one.
I’ll be honest, I dumped this one straight into my Claude summary project even before I saw your comment.
I won’t post the summary, but the custom project instructions I use at the moment are:
You summarize the pasted in text
Start with a overall summary in a single paragraph
Then show a bullet pointed list of the most interesting illustrative quotes from the piece
Then a bullet point list of the most unusual ideas
Finally provide a longer summary that covers points not included already
I don’t think there’s nearly enough discussion about there about how to craft a good summary prompt. Mine here works well enough but I bet there are dozens of ways to improve it.
Does the model really “use” the highly subjective qualifiers “most interesting” and “most unusual”? What is your reasoning for including those? Have you tested results of the prompt with and without and compared amounts of “interestingness” and “unusualness” being highlighted?
Not Simon, but anecdotally, yes. Seeding the LLM via explicit instructions or implicitly through certain prompts can radically alter the response. For example, I’ve taken to prefacing questions with “if you aren’t confident in an answer, I’d prefer that you do not give one” and (again, anecdotally) it makes the LLM less likely to hallucinate false-positive responses. Using certain phrases; the royal ‘one’, ‘whom’, and grammatical constructs that are more unusual outside of technical discussions also tends to elicit more specific answers (you could write a paper on such biases, but it’s not something I care to dive into here).
The thing I’ve noticed most is that the LLM is not a human and does not inhabit a fixed persona. It sees the quality of the response as similar-in-kind to the form of the response. Questions that are structured in a more specific manner are more likely to elicit a similar response in kind. Ask it to provide more accurate information, and it will: because it does not understand accuracy as a semantic property, only an abstract syntactic one.
The “most unusual” thing definitely works. It’s my favorite part of the prompt, because it highlights the things that don’t fit with whatever the model’s idea of “widely assumed already” might be.
I haven’t done a formal evaluation of it, but I’ve run hundreds of ad-hoc documents through that prompt now and I often find the “unusual ideas” section is the thing that provides me the most value.
As for “most interesting illustrative quotes” I think it’s the “illustrative quotes” piece that’s doing the heavy lifting there - my goal is to jump straight to the one or two sentences that best illustrate the overall point that the author is trying to convey. Anecdotally it seems to work well.
I always like to ask for quotes because I can very quickly fact check them against hallucinations by searching for them in the source text.
Kinda tongue in cheek but here’s a summary from Gemini:
AI’s impact on work raises concerns about job displacement, particularly for knowledge workers, due to automation. However, AI also promises to augment human capabilities, changing the nature of work itself. This necessitates adaptation, new skills, and addressing ethical concerns like inequality and potential misuse. The rise of AI-operated firms and a potential deskilling crisis add further complexity. Ultimately, successful integration of AI requires proactive adaptation, ethical consideration, and a focus on equitable outcomes.
Aliases are a bad solution. Abbreviations in Fish have taken over 90% of my shell aliases, including even many of the git aliases I had. It’s just superior in every way.
It’s also better than the script in $PATH approach as they autocomplete and are more easily searchable in reverse history, plus you can template the command line and move the cursor where you want. Scripts or aliases can’t do that.
What I don’t use abbreviations for I use the script solution (rarely) or simple shell functions.
An abbreviation in fish expands after being typed, so you see the full command. There are some more advanced features such as templating, but I’m not experienced enough on that to speak too much about.
Thank you. You made me realize that that I need to use abbreviations way more often. I already do, but I still have some shell script in PATH that could just be abbreviations.
What I especially like about abbreviations is that when I copy&paste the outputs e.g. to document something or share instructions, my personal abbreviations are already expanded, so they don’t confuse other people.
Last I checked, Zen was mostly patches applied on top of Firefox — presumably these “features” could also be patched out. And since it is being built from source it should not be violation of their ToS.
Here’s to the crazy ones. The misfits. The rebels. The troublemakers. The round pegs in the square holes. The ones who see things differently. They’re not fond of rules. And they have no respect for the status quo.
Great one! Happy to support the project by paying a small fee. Like a friend once said “why are you complaining for a 2e app and buy 5 coffees at 0,40e each? One you’ll use often, the other is gone in 5 seconds!”
Looking forward to the login feature so I can comment and like from the app!
Unfortunately since it’s opt-out there will be some amount of time where it is on, and I have not seen a way to prevent it from being on ever. Appreciate trying to help.
Sure thing let me take a look. I have created a issue on GitHub tracking this – feel free to add any context that might be useful (things like what iOS version you’re presently on etc).
Spamhaus.org lists the .work domain as one of the worst TLDs (more than half of the websites with this domain are malware). I am not removing that exception in my DNS server, sorry. Do you have an alternate domain?
That’s a fun project - like Claw, could you please note in your app store title and description that it’s an unofficial app? I got some support requests and other confusion before Claw added that. Thanks. :)
A simple update of about 200Mb for the server version of the OS takes about 1-2 hours.
The default Debian version detects only 4Gb RAM from 8Gb installed, so I’ve tried Ubuntu and it works much better and faster. However, as I see, it is only for IoT.
Switching to NVMe SSD is not described well, so I used Google and more than one guide to do it.
Maybe I had to google it more time, but it’s not described in the official documentation or the getting started guide. I know this will sound strange, but I believe the documentation could be improved in the past two years, considering the fact that this question has been answered several times on the forum.
TL/DR: my experience is not so cool as some of other hardware nerds here, sorry about that =)
milk-v oasis is currently suspended, as sophgo, the company behind the SoC, was rejected by TSMC citing US trade restrictions.
It doesn’t mean there won’t be similar SoCs from other companies (there are likely many in the pipelines), but that specific SoC is probably not happening at all.
The best recent news is SG2044 exists and has been benchmarked (results have been posted in Geekbench). That’s a followup of SG2042, a 64-core monster, but using C930 or whatever the name of the newer core is.
C910 was RVV 0.7 + proprietary extensions, the new one is standard-compliant RVV 1.0 and thus a big deal.
I’ve had VisionFive 2 since early 2023. As of a few weeks ago, it is a Debian Trixie server with ZFS root. Runs Debian’s standard kernel, which is much better situation than most non-x86 boards out there.
But before that, I played with it plenty as a desktop. It’s more or less like this:
CPU is faster than rpi3, but slower than rpi4.
IO is much better than rpi4.
GPU is fast. 2-3x relative to rpi4’s. An open driver is being worked on with Imagination Technologies funding the effort, but it isn’t ready. Running the proprietary driver is a major pain.
Firefox and Chrome are both fine. With experimental video decode support, youtube is smooth.
Almost everything is already upstreamed from server pov, but desktop use needs patches for at least HDMI support, or using vendor’s old kernel for GPU or codec acceleration.
Long suspected this, but interesting to see it called out explicitly.
Companies House used to be another UK Gov property that had a website with opening hours, although they appear to have re-engineered it to be at least read-only during the night now. Means company history/officers can be looked up in the pub when ranting discussing now!
I think this comes into contention with the is-even/is-odd genre of packages on npm, what a gem!
The cherry atop the cake being:
Service degradation: Our API is occasionally returning 2024 instead of 2025. Our intern is investigating and working on a fix. We apologize for any inconvenience and would appreciate if you could stop emailing us about it.
required me to perform manual updates every time a new Immich release came out
Not sure if you’re aware but containers like watchtower automate the process of updating your containers – although it does not protect against breaking changes in the services that you’re pulling.
Using NixOS does have the benefit that you only have to take care of breaking changes twice a year. And inbetween, you’re fine with letting your system auto update (and auto reboot in case of Kernel updates). I’m aware this is much more often than e.g. on Debian or Ubuntu, but less often than rolling releases like Arch, so I consider it a good compromise between maintenance effort and frequency of feature updates.
As for auto updating docker containers, I wish docker tags had some kind of built-in semver resolution. Unfortunately Nix Flakes suffer the same issue, and now commercial providers try to piggyback on this lack of an essential feature by moving semver resolution into URL redirects.
Probably worth mentioning Cline, a VSCode extension that let’s you pair with a LLM of your choice right from your IDE. It produces diffs when you chat with it about aspects of your project which you can then merge in or reject for a different version.
Does this mean that the first frame of Doom can now be rendered in only ~1.2 days?
For folks lacking context, this is a reference to this video: https://youtu.be/0mCsluv5FXA where the creator emulated DOOM using just TS types. It took 12 days to render the first frame of the game.
The author of that project replied in the comments.
I don’t mind some of these things on AI, but why are they always so darn long? I think it’s like 10,000 words or something. I kind of scrolled to the end and oh well. Deep fakes are most likely to kill us. I think the summarization might have really helped this one.
I’ll be honest, I dumped this one straight into my Claude summary project even before I saw your comment.
I won’t post the summary, but the custom project instructions I use at the moment are:
I don’t think there’s nearly enough discussion about there about how to craft a good summary prompt. Mine here works well enough but I bet there are dozens of ways to improve it.
I use Perplexity for this and shared my prompts: https://kyefox.com/using-perplexity-ais-spaces-as-a-life-raft-in-an-age-of-ai-slop/
Does the model really “use” the highly subjective qualifiers “most interesting” and “most unusual”? What is your reasoning for including those? Have you tested results of the prompt with and without and compared amounts of “interestingness” and “unusualness” being highlighted?
Not Simon, but anecdotally, yes. Seeding the LLM via explicit instructions or implicitly through certain prompts can radically alter the response. For example, I’ve taken to prefacing questions with “if you aren’t confident in an answer, I’d prefer that you do not give one” and (again, anecdotally) it makes the LLM less likely to hallucinate false-positive responses. Using certain phrases; the royal ‘one’, ‘whom’, and grammatical constructs that are more unusual outside of technical discussions also tends to elicit more specific answers (you could write a paper on such biases, but it’s not something I care to dive into here).
The thing I’ve noticed most is that the LLM is not a human and does not inhabit a fixed persona. It sees the quality of the response as similar-in-kind to the form of the response. Questions that are structured in a more specific manner are more likely to elicit a similar response in kind. Ask it to provide more accurate information, and it will: because it does not understand accuracy as a semantic property, only an abstract syntactic one.
Sure, my question was more about the specific terms Simon’s prompt used: why the subjective terms “interesting” and “unusual”?
The “most unusual” thing definitely works. It’s my favorite part of the prompt, because it highlights the things that don’t fit with whatever the model’s idea of “widely assumed already” might be.
I haven’t done a formal evaluation of it, but I’ve run hundreds of ad-hoc documents through that prompt now and I often find the “unusual ideas” section is the thing that provides me the most value.
As for “most interesting illustrative quotes” I think it’s the “illustrative quotes” piece that’s doing the heavy lifting there - my goal is to jump straight to the one or two sentences that best illustrate the overall point that the author is trying to convey. Anecdotally it seems to work well.
I always like to ask for quotes because I can very quickly fact check them against hallucinations by searching for them in the source text.
Kinda tongue in cheek but here’s a summary from Gemini:
Aliases are a bad solution. Abbreviations in Fish have taken over 90% of my shell aliases, including even many of the git aliases I had. It’s just superior in every way.
It’s also better than the script in $PATH approach as they autocomplete and are more easily searchable in reverse history, plus you can template the command line and move the cursor where you want. Scripts or aliases can’t do that.
What I don’t use abbreviations for I use the script solution (rarely) or simple shell functions.
What’s the difference, and what are the alternatives for those who don’t use fish?
An abbreviation in fish expands after being typed, so you see the full command. There are some more advanced features such as templating, but I’m not experienced enough on that to speak too much about.
That’s super cool! I have been using fish for a couple of years now and haven’t come across abbreviations yet – looks like it is time to give it a go!
As far as I know there are no alternatives in other shells. It’s something I’ve seen only in Fish.
The doc has some examples of its usefulness: https://fishshell.com/docs/current/cmds/abbr.html
You can set ZSH to tab expand aliases, and you can expand them with a keybind in bash as well. https://superuser.com/questions/1514569/how-to-expand-aliases-inline-in-zsh
I see. In comparison, I don’t really see why you’d say that “aliases are a bad solution”, then? Looks mostly like a matter of taste to me.
Thank you. You made me realize that that I need to use abbreviations way more often. I already do, but I still have some shell script in PATH that could just be abbreviations.
What I especially like about abbreviations is that when I copy&paste the outputs e.g. to document something or share instructions, my personal abbreviations are already expanded, so they don’t confuse other people.
Why does your shell not autocomplete$PATHbinary names?Never mind, I realised you meant it breaks the autocomplete of the binary’s arguments/flags/etc.
I wonder if or how this impacts derivatives like Zen?
Last I checked, Zen was mostly patches applied on top of Firefox — presumably these “features” could also be patched out. And since it is being built from source it should not be violation of their ToS.
Fantastic write-up really, fun read!
True to the ethos of:
that is not an open source software license. the term is source-available proprietary.
Yeah that’s fair, my plan is to transition to GPLv3 or AGPLv3 in some time.
But, for now, please remove the Open Source claim.
It’s AGPLv3 licensed now, appreciate the nudge.
Neat.
Purchased as a way to support your effort even though without user login it is unlikely I’ll use this on a daily basis.
I have now implemented login with v3 — its live on the App Store, support for comments coming soon too!
Great one! Happy to support the project by paying a small fee. Like a friend once said “why are you complaining for a 2e app and buy 5 coffees at 0,40e each? One you’ll use often, the other is gone in 5 seconds!”
Looking forward to the login feature so I can comment and like from the app!
I have now implemented login with v3 — its live on the App Store, support for comments coming soon too!
Just tested it, worked great! Amazing work !!
Is there a possibility of relaxing the iOS version restriction? I’m not upgrading to the pushed AI features until the phone gets bricked. Thank you!
https://www.wired.com/story/how-to-turn-off-apple-intelligence/
Unfortunately since it’s opt-out there will be some amount of time where it is on, and I have not seen a way to prevent it from being on ever. Appreciate trying to help.
It won’t do anything unless you engage it.
Sure thing let me take a look. I have created a issue on GitHub tracking this – feel free to add any context that might be useful (things like what iOS version you’re presently on etc).
Spamhaus.org lists the
.workdomain as one of the worst TLDs (more than half of the websites with this domain are malware). I am not removing that exception in my DNS server, sorry. Do you have an alternate domain?I was unaware, yikes! It’s been posted below but here’s a direct link to the app store listing: https://apps.apple.com/us/app/crustacean-for-lobsters/id6742195107, in addition I’ve also added it to the repo’s readme file on GitHub.
Thank you kindly!
Here’s a link to the Apple app store: https://apps.apple.com/us/app/crustacean-for-lobsters/id6742195107
That’s a fun project - like Claw, could you please note in your app store title and description that it’s an unofficial app? I got some support requests and other confusion before Claw added that. Thanks. :)
Absolutely, will do! I will need to push a new release and get it approved for distribution on the app store so it might take a couple of days.
At $199, that is very reasonably priced! – I wonder what the software support landscape in RISC-V is presently.
The software support isn’t bad, but the price is largely because this particular chip is no barnburner.
The Debian trixie effort is doing well.
https://buildd.debian.org/stats/graph-week-big.png
Neat, thanks for the stats – just what I was looking for!
I’ve been watching since RISC-V became a first tier Debian architecture and the big build started.
We’re only weeks (possibly just days!) away from surpassing ppc64 and becoming the 3rd largest.
Ubuntu works better then debian from my experience on VisionFive 2
Is it usable, speed wise? For, say Firefox? @deivid above mentions it’s slow, but that can also be IO or whatever.
I haven’t tried it for the desktop because:
You’d need a DeviceTree overlay, as there are multiple versions of the board. This is a minimal one:
Build the overlay with
dtc -@. In u-boot’s extlinux.conf, usefdtoverlayscommand to specify your overlay.Maybe I had to google it more time, but it’s not described in the official documentation or the getting started guide. I know this will sound strange, but I believe the documentation could be improved in the past two years, considering the fact that this question has been answered several times on the forum.
TL/DR: my experience is not so cool as some of other hardware nerds here, sorry about that =)
PS: I’m really liked radxa community and their hardware, maybe https://forum.radxa.com/t/radxa-rock-5-itx-vs-milk-v-oasis/20960 will have more opportunity from users.
milk-v oasis is currently suspended, as sophgo, the company behind the SoC, was rejected by TSMC citing US trade restrictions.
It doesn’t mean there won’t be similar SoCs from other companies (there are likely many in the pipelines), but that specific SoC is probably not happening at all.
The best recent news is SG2044 exists and has been benchmarked (results have been posted in Geekbench). That’s a followup of SG2042, a 64-core monster, but using C930 or whatever the name of the newer core is.
C910 was RVV 0.7 + proprietary extensions, the new one is standard-compliant RVV 1.0 and thus a big deal.
It’s about 3-4x slower than a Raspberry Pi cpu-wise. The I/O is not great, even with an NVMe SSD, less than 100MB/s.
Which model of Raspberry Pi? They vary in performance by maybe 100x https://magpi.raspberrypi.com/articles/raspberry-pi-specs-benchmarks (that article predates the 5)
Considering JH7110 CPU performance falls between rpi3 and 4, the stated figure suggests it is a comparison with rpi5.
I’ve had VisionFive 2 since early 2023. As of a few weeks ago, it is a Debian Trixie server with ZFS root. Runs Debian’s standard kernel, which is much better situation than most non-x86 boards out there.
But before that, I played with it plenty as a desktop. It’s more or less like this:
Firefox and Chrome are both fine. With experimental video decode support, youtube is smooth.
Almost everything is already upstreamed from server pov, but desktop use needs patches for at least HDMI support, or using vendor’s old kernel for GPU or codec acceleration.
It’s amazing what kind of hardware we have available nowadays :)
Thanks for the update!
Curious how this compares to Pytorch using a MPS backend.
Really cool work, kudos!
Long suspected this, but interesting to see it called out explicitly.
Companies House used to be another UK Gov property that had a website with opening hours, although they appear to have re-engineered it to be at least read-only during the night now. Means company history/officers can be looked up in the pub when
rantingdiscussing now!Bank of Scotlands car finding website also has opening hours, I wonder if it’s also the same reason!
It’s been a couple of days since this was posted, is there something concrete that UK citizens can do to help ameliorate the situation?
Edit: For the next reader, this comment seems to have more details: https://lobste.rs/s/ukosa1/uk_users_lobsters_needs_your_help_with#c_l5lif8
I think this comes into contention with the
is-even/is-oddgenre of packages on npm, what a gem!The cherry atop the cake being:
Looks like an unfortunate case of print debugging accidentally making its way to a release.
https://gitlab.com/gnachman/iterm2/-/commit/5db0f74bf647f6d53ea33d2491f9b8e4972c2e03 + https://gitlab.com/gnachman/iterm2/-/commit/63ec2bb0b95078a97abbb94cf28c4a42d1b67f23
I have so many questions about their Python code. Why aren’t they using
pathliborlogging? are first.Not sure if you’re aware but containers like
watchtowerautomate the process of updating your containers – although it does not protect against breaking changes in the services that you’re pulling.Using NixOS does have the benefit that you only have to take care of breaking changes twice a year. And inbetween, you’re fine with letting your system auto update (and auto reboot in case of Kernel updates). I’m aware this is much more often than e.g. on Debian or Ubuntu, but less often than rolling releases like Arch, so I consider it a good compromise between maintenance effort and frequency of feature updates.
As for auto updating docker containers, I wish docker tags had some kind of built-in semver resolution. Unfortunately Nix Flakes suffer the same issue, and now commercial providers try to piggyback on this lack of an essential feature by moving semver resolution into URL redirects.