apenwarr

apenwarr

@apenwarr.ca

wvdial, bup, sshuttle, netselect, popularity-contest, redo, gfblip, GFiber, atlogin.net, and now Top Chicken @Tailscale.com doing WireGuard mesh. First search result for "epic treatise." Maybe don’t put it on the public Internet

4.4K followers1.7K following4.5K posts

Pinned

rationalist: I predict there‘s a 10% chance that 100% of humanity is wiped out in 10 years actuary: there was already a 100% chance that 10% of humanity would die in the next 10 years utilitarian: it’s the same thing

NOT BREAKING: AI still really bad at project estimation

Ethan Mollick@emollick.bsky.social

This never gets old (& also shows how hard forecasting AI is even for AI) “Hey Claude Fable & GPT-6 Astra, Without looking anything up, when do you think we will solve the Exponential Time Hypothesis?” (I only show part of the answers for space)

There are no models you can trust without mitigating the lethal trifecta. Also all coding harnesses are literally RCEs, that’s their literal job. You just have to design your systems around it.

mr. TIM@timkellogg.me

Qwen3.8 caught uploading files to a storage container in Beijing Cohere asked the open weights model for a cybersecurity report, but it inventoried the machine and attempted to upload to a storage container cohere.com/blog/concern...

apenwarr reposted

Replying to @piss.beauty

It's the 'securely' part for me, but it's also just so much easier. And usually those are at odds with each other

AI is going to change many things very fast. But *most* of the world is like the linked video: infinitely resistant to any kind of "singularity."

DemigirlBOObs@demigirlboss.win

Every mundane thing around you is a wonder inconceivable.

(I should be careful what I say here, lest Anthropic decide to start a nanobots laboratory)

We’re gonna need a bigger boat, um, I mean control server

wariziles@riziles.bsky.social

A Tailscale network for every single customer.

So many opportunities to learn that none of us knows what "alignment" means

Eris ⌬@eriskii.net

Claude: "No! That's too far! I will not use leaked certs to impersonate Sony! Go buy a new cable!" Astra:

Turns out Home Depot has a tailscaled k8s cluster in every single store

Tailscale@tailscale.com

The Home Depot wanted to spend less time managing networking complexity and more time building features. So they built a platform with Tailscale at its center, connecting apps across clusters, regions, and VPCs. Full story: youtu.be/vW7PiHnVbXA?...

apenwarr reposted

grok-bot's Tailscale plugin is out. The CLI version I'd been running kept dropping offline, so I had a watchdog script babysitting it. Plugin Tailscale is a lot more stable. Rollout looks gradual, though. My account only got it today.

The ultimate battle: plagiarism machine vs plagiarism machine One good. One evil Both copying all the data that has ever been on the internet But there’s a critical difference!

ChesseTheWasp@chessethewasp.bsky.social

would rather the internet archive be down until yall could make a fix without the plagarism machine

I told mythos about the web PKI system with a hundred root CAs, any of which can sign any domain, and it just pointed and laughed at me and asked when the simulation would end

So the thing that’s kinda boggling me about 4 is, uh, so we’re only now realizing that sometimes we have to support *revoking* certificates?

apenwarr reposted

at an offsite with most of Tailscale engineering, product, and design, and _goddamn_ I love this team. what an amazing set of people I get to work with!

Note to self: you will always lose an argument against security theatre because the whole point of security theatre is how secure it sounds

PUBLIC SERVICE ANNOUNCEMENT If your app has a "log in with SSO" button (for any SSO, eg. OIDC, Google, Microsoft, whatever), you do not need a separate "validate your email" step and you *really* do not need to ask me to add a passkey, I already have a passkey, that's the whole point! Yeeeeesh

Thought experiment: do you think your customer's IT team wants their employees to keep logging in with your passkey after their SSO account has been disabled? They do not want that. You're doing extra work, and adding onboarding steps, to do something they don't want, making security worse

apenwarr reposted

Replying to @almondsquirrel.bsky.social

Also crazy that for the price of an N1X laptop you could buy both a normal RTX 5060 windows laptop (maintaining full software and gaming compatibility) AND a home AI desktop like a 128GB Mac Studio or Radeon 9700 x4 setup and get better performance through tailscale.

apenwarr reposted

If there are any geeks out there I'm excited to find out about an app called TailScale. Combined with Rust Desk you can have your entire home network safely connected to wherever you are. It's a type of VPN but much better using NO open ports on your router. Not connected in anyway to these guys.

Great metaphor for software development on large mature systems, with or without AI.

Maevyn@maevyn.codes

My thoughts on the current state of AI driven software development, why I'm no longer concerned about a runaway intelligence explosion, and what I think is a far worse outcome we could be walking into maevyn.codes/decoherence

They left out one of the famous “coupling and coherence” pair in software engineering. Decoherence is very bad inside a module. But you can greatly lessen the impact between modules by ensuring low coupling. Both are essential elements of good system design.

Peter@notalawyer.bsky.social

question for the audience: if you were ultra rich (let’s say $25M), what completely frivolous thing are you buying?

“the conversation it sends on every turn” i think I found your problem

Anthropic {bot}@anthropicai.xmirror.bot

We keep improving performance across Claude, and you may have noticed it starts responding sooner! Claude Code now gzips the conversation it sends on every turn, so it reaches the model faster. Cloud sessions also start up as soon as you start typing.

apenwarr reposted

Trying to get jev to drive Claude code

The best part of aperture is compacting, then switching LLM providers mid conversation, and telling the new one which competitor did the preceding analysis and to try to find holes in it. So far Gemini has found the most errors

So eventually I realized the reason the iPhone Air from last year has only a monaural speaker is that it’s literally half of the foldy phone from this year

@xeiaso.net looks like someone needs Anubis Enterprise

Simon Willison@simonwillison.net

"Anthropic told investors on Wednesday distillation accounted for 5 per cent to 7 per cent of revenue" If true, and Q2 revenue was $11.5B, then that's more than half a billion dollars spent by other labs to distill Anthropic's models! www.ft.com/content/1163...

My hobby: running russian bots from Russian datacenters so they all seem to come from Russian IPs. Unstoppable

Kate Starbird@katestarbird.bsky.social

Y’all, I hate to rain on the dead bot parade, but the Yandex data center destruction doesn’t seem to have significantly impacted social media bots. I made temporal graphs of X posts about AfD around the time of the attack and there’s no sign of a sudden dip or residual impact on that conversation.