Have said this before in other forms, but will reiterate it here: if you're doing long-form writing of any sort, reading the words aloud into a microphone is a completely unparalleled cheat code. It's free, easy, and so effective it feels like it shouldn't be allowed. I do it in a few stages...
Seems like people with a lot of reach would still get a lot of indirect likes through reshares, etc. It might select their posts that are more "viral" which isn't really what I was hoping for, but it would make sense that viral posts have more universal appeal. Hard to judge in advance.
@spacecowboy17.bsky.social Had an idea for an experiment: suppose "For You" put more weight on likes on posts written by people you don't follow, versus people you do follow?
Why: you might 'like' a post by a friend because you know them and have shared context. A stranger wouldn't get it.
Hasn't happened in a while, but I used to explain this as "it learned the pattern."
Back in 2023 I was using ChatGPT Code Interpreter and if I told it that its fix for a bug didn't work enough times, it would keep trying a fix and immediately assume that it didn't work, because that's the pattern.
Seems like partial bans are going to be needed or it's going to be like Google banning someone's whole account because they did something sketchy with one service. The more apps share the same PDS, the worse it's going to get. Oh what a Tangled web we weave.
It seems like having a clear separation like that from the beginning makes it easier to run jobs in parallel? I'd have to clean up some scripts that assume it's all in one VM.
I've been asking Shelley to write a design doc (checked in under a subdirectory in the repo) and then I start a new conversation to implement each phase. No issue tracker yet. How do you manage the handoff between Shelley and GH Agents? Do you put everything in the issue? Any tricks to that?
I've been using exe.dev for everything. Is there anything I'm missing out on by not using GH Cloud Agents?
so everything on earth has been monotonically more and more about ai, every year, for three years
if the trend holds this is the least anything will ever be about ai again
What are some examples of things that your agent does for you that you find valuable? I've been following the news about OpenClaw, etc, but so far haven't seen a compelling reason to dive in other than curiosity.
Reading GROWING.md was interesting, but the docs don't really explain why I should try running this. What could running an agent do for me? Why should I commit to a week-long setup process, etc?
We present WoFBench, a novel test that compares the knowledge of Wings of Fire superfans to frontier AI models. The benchmark unfortunately proved to be saturated on creation. We note that saturation joins human extinction on the list of reasons for concern about the pace of AI progress.
@simonwillison.net This looks like an interesting idea for making agents a bit safer: ask the agent to plan tool access and write an access policy before reading any untrusted input.
dystopiabreaker.xyz/fsm-prompt-i...
We're more like the audience watching the movie though? I look up from reading social media and nothing has changed locally.
Any thoughts on something like SplittableRandom? Seems like it would be less awkward to use than jump for creating random streams.
I know someone with an iPhone who managed to lose their Gmail account, but it was years ago so perhaps it's improved?
It will mostly work, but it seems like not really knowing which password/passkey managers you have or what's stored there or how it's backed up will set people up for failure when they lose things like their phone. There are people who chronically lose things.
How do you look at source files in your VMs? Or do you not do that anymore?
How does it do on find definition / find usages? Remote access via ssh?
I'd guess this results in mixed feelings when employees are shareholders too.
Zoom out to a year though and it looks like they're back where they started.
Do you use the issue tracker on Tangled? Not seeing a lot of activity.
Six years ago, 80 workers moved into a plastics factory, and stayed there for a month. In those 28 days they produced 40M pounds of polypropylene, enough for maybe 500M N95s. Ordinary people, looking at their situation, and thinking creatively about how to do their part.
Something like, keep source code until a test fails, then regenerate, and also have a coding agent review the diff to make sure it doesn't lose anything. Turn community feedback into new tests. Monitor upstream for new releases and automatically try them out.
It reminds me of how some people have proposed replacing source code libraries with prompts. In this case we start with prompts and then ask, what does it buy us to keep the source code for an installer around? What would that give us?
Maybe we just have tests?
This is looking a bit like what Sandstorm was supposed to be. What would it take to turn a list of ideas into a proper app store? I guess the installs would be generated once and tested rather than every time? Perhaps some kind of community feedback process.
apps.sandstorm.io
New blogpost about atproto
It's not federation, it's not a p2p mesh. It's a secret third thing: practical.
www.pfrazee.com/blog/practic...
That was a pretty brief period just before the dot-com bubble. I remember thinking it unseemly that people in the office were speculating in domain names. Then one of them sold one for millions and retired.
Google published a similar report that got a lot less attention:
cloud.google.com/blog/topics/...
Yes, the ability to write software is just one input. But nonetheless, the barrier to entry for new firms seems to be getting lower? Worrying about competition from AI-enabled firms seems similar to worrying about low-cost foreign competitors. How much it matters will vary by industry.
It seems you're marketing to highly technical users :-)
