Ed

Ed

@ed3d.net

Main Character Syndrome, but fun. Boston-based. My wood shop and my dogs are my personality. If you're still reading this, I'm your dad now. Sorry, I don't make the rules. I don't speak for my employer. I do make @polytoken.dev. signal: edr.91

11.8K followers509 following58.9K posts

I am surprised that spot pricing hasn’t happened yet, but also OpenRouter failed to compel providers on their platform to provide like for like tokens, so not that surprised

Stephen Judkins@stephenjudkins.bsky.social

Yeah. Self-hosting isn't going to happen much, but complete commodification of inference probably is going to be. I can see stuff like automatic bidding on individual sessions happening, even

hey Slate,,,

Jeff St John@jeffstjohn.bsky.social

Ford's $28,350 Ford Fathom EV pickup will come with onboard power conversion tech that will slash the cost of setting it up as a home backup battery. Will other automakers join Ford and Tesla in pursuing AC V2H? And will utilities let them plug in? www.canarymedia.com/articles/ele... #energysky

Ed reposted

Replying to @benjedwards.com

All the code for Apple II Agent is up on GitHub if you want to run it yourself. You need an Uthernet II card in your Apple II at the moment: github.com/benj-edwards... Also, Claude assembled a "how it was made" technical document that is an interesting read: bxfoundry.com/apple2agent/

bxfoundry.comApple II Agent Project NotesHow Apple II Agent came together, January to October 2026: an Apple IIe that asks Claude for software, tests it in an emulator and runs it.

quick question how do you do distillation without RL

Quantian@quantian.bsky.social

The only competitive Chinese model is GLM because they are focusing on RL rather than distillation in order to develop powerful cyber capabilities and as a result they might get, like, literally sanctioned if they release an open source cyber weapon so idk how long that continues

View full thread

I wouldn't say "last year"; GLM 4.7 was not amazing. definitely this year. if models never advanced another inch it would be years before people exhausted what GLM 5.3 or DS4.1 give them here, today (you should follow this account if you are interested in this stuff, strong Being Normal energy)

jae (hallowed)@fubarchitect.com

in contrast tho, for most people reading this, your work is 99% simple stuff that open models were sufficient for last year

tbh I shouldn't say GLM 4.7 was not amazing. it was really impressive at tasks when you could throw thirty of them at a problem at once and map-reduce with something smarter. it strongly informed my "use subagents, you dolt" strategy

Ed reposted

Amazing news, and if you can afford to, join me in becoming a member of the site. The web needs those of us who have benefited from it so much to give a little back:

quick question how do you do distillation without RL

Quantian@quantian.bsky.social

The only competitive Chinese model is GLM because they are focusing on RL rather than distillation in order to develop powerful cyber capabilities and as a result they might get, like, literally sanctioned if they release an open source cyber weapon so idk how long that continues

View full thread

for the unfamiliar: that is a live dashboard that charted the reinforcement learning logs of a Xiaomi model that, for whatever you think of the "Pareto frontier", leads it in its category

in its category on artificialanalysis, that is, for whatever those benchmarks remain good for anyway, should the takeaway here be that Anthropic sucks at making models or that China is not simply stealing America's homework

quick question how do you do distillation without RL

Quantian@quantian.bsky.social

The only competitive Chinese model is GLM because they are focusing on RL rather than distillation in order to develop powerful cyber capabilities and as a result they might get, like, literally sanctioned if they release an open source cyber weapon so idk how long that continues

(luna ultra on a tree tumor, plus some random shots from today on the Ace Pro 2) (insta360's ecosystem of actually good enough devices is ALARMING)

Ed@ed3d.net

the only thing that would make the Insta360 Luna Ultra better is open gate. that thing is such a fun little video guy

the only thing that would make the Insta360 Luna Ultra better is open gate. that thing is such a fun little video guy

Ed reposted

Replying to @cooperlund.online

Much like a world with fake-ass corporate Pride merch, I would much rather live in a country where both sides at least pay lip service to everyone being an American over a mask off one like this because taking the mask off is a sign something has gone wrong.

Ed reposted

still waiting adam I see you're continuing to poast but not doing much else also I am very interested to hear how "professional" threats are actionable (or what a "professional" threat even is...working to hold you accountable for myriad violations of DoDD 1344.10 isn't a threat, just a promise)

Mike Black@mikeblack114.bsky.social

ok this will require you to do something besides poast fyi

Ed reposted

Normal Men disavows this poster and would also note that the clans can kiss our asses

Ed reposted

One if the lessons to be learned from Brazil is to stop elevating your opponent’s social media (even to dunk) and promote your own message. _This applies to Matt Yglesias and his ilk, too._ Just stop it, please.

I think this is a pretty good thread

Rob Donoghue@rdonoghue.bsky.social

So, the middle tier of management is always in a bit of a rough position in any company, because if they work well, it usually just means things continue to function. That’s not a super glorious place to be. But that does make for a soft target for AI job replacement. No one minds fewer managers!

sounds like they, and they specifically, should be stopping to prove their good faith

ESG Hound@esghound.com

*ANTHROPIC WARNS OF CATASTROPHIC, EXISTENTIAL RISKS TO HUMANITY

Ed reposted

I've been meaning to get around to posting the full details, so what the hell - why not get it out while it's still Vuln Week: Anthropic's mech interp driven security model is built on a fundamentally incorrect assumptions. Let's dive in, starting with this paper: www.anthropic.com/research/nex...

anthropic.comNext-generation Constitutional ClassifiersWe've developed the next generation of Constitutional Classifiers, offering more efficient protection against universal jailbreaks.
hikikomorphism@hikikomorphism.bsky.social

About a year ago, I got really into making glitch art exploring the edge cases & failure modes of machine cognition. This had some unexpected applications. Here’s the result: recursion.wtf/posts/claude... I've shared the relevant logs & artifacts with Anthropic, and the matter is under review.

Ed reposted

Maturing as a software engineer doesn’t mean writing fewer Horrors, it means knowing when to write A Horror and how to construct the proper containment zone around it

I once inherited a python system that handled about $100m in ARR. It was a single Django class that was 25k lines long and had zero tests or docs that always called out to MySQL and postgres via fork+exec rather than actually using those client libraries

I have put a LOT of my personal points in Avoiding Advertising, and it is always so jarring when I see people like...use YouTube raw. I don't even use the YouTube app on my phone/tablet because Safari gives you SponsorBlock

Snowden St.@snowden.st

i think it’s a genuine question if a sustainable society can be made up of people who, on average, are exposed to 100s of advertisements every day of their life.

View full thread

“How do I know if a request was handled on device or went to Apple PCC?” “Apple PCC is…you can see this in the Apple Intelligence Report.” “Can you open it?” [starts directions to my house] “That isn’t what I asked. Can you open the Apple Intelligence Report?”

also it interacts weirdly with Apple Watch prompts. this is very sloppy.

I just opted into it and I’m wondering what this is actually supposed to be for. It can’t fire intents into other applications, it mishears (I think the Apple voice model is worse than parakeet), and just in general is awful It can do workflows, but can’t do HomeKit automations? wtf?

The weird thing is that it is acting like a very old model in the way it tries to assure me that yes, it can do things it can’t. “I can help you with creating scenes and automation in HomeKit. If you open the app and click ‘add’…”

I'm going to need SEVERAL hiking trails for this next technology video

Ed reposted

Because there are 10s of thousands of Democratic electeds in the United States, and you have positioned yourself directly into a media environment that is designed to find every statement a Democrat makes that is even vaguely transphobic as soon as it happens.

you do not need context for this post but you do need to appreciate it.

slanderous ordo@slanderous.bsky.social

immasculine and feminin't

Ed reposted

Thank you, thank you, thank you to everyone who spared a thought for my country today, who amplified information about the coup d’état attempt, and worried about Brazil’s democratic future—the world’s democratic future. We won! Women won! The Amazon won! Young people won! Democracy won! Hope won!

when you need a substitute jev

xanadian@xanadian.bsky.social

Yes absolutely! I cat absolutely dog yes

Ed reposted

Lula winning is great for the planet, great for Brazil, and great for people who aren't revanchist assholes everywhere in the world, but let's not forget that it's also going to make some of the worst people on Earth extremely sad and mad. We should enjoy that, as a little treat.