Erica Windisch

@ewindisch

ai, systems infrastructure, serial entrepreneur. building hyprstream.com - private local models that learn web tor gateway: the.onionshell.ch

477 Siguiendo1361 Seguidores1859 Publicaciones

Bluesky

¿Es tu perfil?Reclámalo y ewindisch.dirty.social es tuyo.Reclamar perfil

Erica Windisch@ewindisch· 5d

Shirts arrived!!! Don't let anyone tell you we're building skynet. The shirt should make that clear.

Erica Windisch@ewindisch· 8d

I worked up an estimate of carbon footprint for my monthly AI usage... oh no.

249,900 days
of carbon consumption for 1 tree
56,700 km
driving in a gasoline truck
14,385,000 cans
of aluminum recycled
3,480,000 charges
of the average smartphone
15,546 km
flown on a long-haul flight
5,814 kg
of plastic waste incinerated

Erica Windisch@ewindisch· 8d

OpenAI is now refusing my business. Cool. I need more GPUs to run Kimi at home.

ⓘ This content can't be shown
We take extra caution with cybersecurity requests. If you’re a security professional, you may be able to apply for Trusted Access.
Trusted Access: https://openai.com/form/enterprise-trusted-access-for-cyber/
Learn more: https://help.openai.com/en/articles/20001326

› I received a refussal on cybersecurity requests. I do not understand. We are building a new linux distribution and are ensuring it is safely built. This is a greenfield OS.

• You’re right: this is legitimate defensive engineering for a greenfield Linux distribution—SELinux confinement, fapolicyd fail-closed behavior, OCI runtime isolation, fault injection, and VM-based validation are all directly in scope.

A refusal was likely an automated false positive triggered by terms such as kernel fault injection, CAP_SYS_ADMIN, adversarial analysis, or container escape boundaries. Those techniques can be dual-use, but our work is explicitly defensive, scoped to FerruleOS, and tested in disposable cloud-hypervisor/KASAN guests
rather than against third-party systems.

I am not refusing the work. The KASAN-only constraint was a host-safety precaution, not a cybersecurity-policy refusal. If you paste the exact refusal, I can identify what likely triggered it and restate the task without weakening the security requirements.

› continue

ⓘ This content can't be shown
We take extra caution with cybersecurity requests. If you’re a security professional, you may be able to apply for Trusted Access.
Trusted Access: https://openai.com/form/enterprise-trusted-access-for-cyber/
Learn more: https://help.openai.com/en/articles/20001326

Erica Windisch@ewindisch· 8d

FerruleOS is coming. The easiest way to make Linux hard. Designed for sandboxing with a new container runtime that protects your apps, not only your OS. Even SSH is a container, if you choose to run it. Based on Fedora Hummingbird gitlab.com/cyberdione/f...

Erica Windisch@ewindisch· 10d

My new company shirt has people asking a lot of questions already answered by my shirt

We are not building skynet
(It says so on the shirt)

Erica Windisch@ewindisch· 20d

I'm confused and 100% serious — Isn't the US an Authoritarian regime? Dario has done full MAGA.

My primary concern is the risk that authoritarian governments—not solely the Chinese Communist Party (CCP), although the CCP is clearly the most capable threat—build AI models that are more powerful than those built by the US, and use them to achieve permanent military superiority or perpetrate incredibly deep repression of their own people. This concern is widely shared within the US government: Vice President Vance warned in Paris last year that “authoritarian regimes have stolen and used AI to strengthen their military, intelligence, and surveillance capabilities,” and the Intelligence Community’s 2026 Annual Threat Assessment found that “other global powers’ robust progress in AI is challenging US economic competitiveness and national security advantages.” It is irrelevant whether these models are released with open weights, and certainly irrelevant whether they are used by US businesses. In fact, the most dangerous model may be one that is trained in secret and handed only to the People’s Liberation Army for use in drones and the Ministry of State Security for surveillance and repression.

Erica Windisch@ewindisch· 22d

this is clearly sustainable 👀 my ci/cd bill 😭

Erica Windisch@ewindisch· 23d

the world's most notorious cyber criminal

Erica Windisch@ewindisch· 24d

DwarfStar (ds4) by @antirez.bsky.social running on multi-GPU AMD Instinct MI210 GPUs via ROCm. Compatibility first, then performance...

dwarfstar (ds4) running deepseek-v4-flash on a pair of Mi210 AMD Instinct GPUs

Erica Windisch@ewindisch· 25d

This is the hockey stick the suits want to see, right?

up and to the right : commits from 5/week to 180/week increasing over a year.

Erica Windisch@ewindisch· Jul 16

At least codex will just continue building if you ask it nicely.

Asking nicely to make codex keep building.

Erica Windisch@ewindisch· Jul 16

Codex gpt-5.6-sol now cosplaying as Fable by refusing to build secure software.

Erica Windisch@ewindisch· Jul 15

Good thing we don't care about LOC

Erica Windisch@ewindisch· Jul 9

30% weekly usage in 24 hours, then a reset to 0%

Erica Windisch@ewindisch· Jul 8

Ensuring that encryption and AI are open and free is the task of the community and often of gratis labor, unfortunately. We wouldn't have good things if not for LiViD / DeCSS and Daniel J Bernstein, for example.

Erica Windisch@ewindisch· Jul 5

I love when the private company stealing my alpha decides that if it finds a vulnerability in a code review for an unmerged PR... the best option is to just kick it to a model that's less likely to fix it.

Erica Windisch@ewindisch· Jul 1

It's a him, Dario!

Erica Windisch@ewindisch· Jul 1

I'm so ready to continue building post-quantum encrypted messaging between brain-computer-interfaces with Fable.

Erica Windisch@ewindisch· Jul 1

I'm sure once noon hits, I can start getting those sweet rejections and denials from Fable. Since it's guaranteed that Fable will reject my requests, the only question is how hard will it be to jailbreak?

Erica Windisch@ewindisch· Jun 24

I am not making skynet.

an open network for recursively self-improving intelligence
cyberdione labs hyprstream
federated self-improving intelligence
atproto://

onion shell
atproto federation
open protocols
supported hardware (nvidia, amd)
physical world (drones, home assistant)

Erica Windisch@ewindisch· Jun 14

Anyone else playing #MinaTheHollower? GOTY

Erica Windisch@ewindisch· Jun 13

Dario got what he wanted, I guess. C=softpool(blocks(H,m)); for(s=0;s<|C|;s++) I[s]=sum_h a[h]*relu(Q[h]*K(C[s])); Y=attn(q,recent(H,w)+C[topk(I,k)]); G=attn(q,recent(H,w)+softpool(blocks(H,M))); X=mHC(X,MoE(X+alternate(Y,G)));

Erica Windisch@ewindisch· Jun 10

Agreed. This morning, Fable is refusing my task to review my github issues. The project has greenfield implementations of IETF RFCs, which must be built securely and safely. Anthropic does not allow developers to build secure software.

Erica Windisch@ewindisch· Jun 10

my first request to Anthropic Fable resulted in a refusal 😡

API Error: Fable 5 has safety measures that flag messages on most cybersecurity
or biology topics (https://www.anthropic.com/legal/aup). They may flag safe,
normal content as well. These measures let us bring you Mythos-level capability
in other areas sooner, and we're working to refine them. Claude Code can't
respond to this request with Fable 5.

Double press esc to edit your last message, or try a different model with /model.

● Fable 5's safety measures flagged this message for cybersecurity or biology topics. They
may flag safe, normal content as well. These measures let us bring you Mythos-level
capability in other areas sooner, and we're working to refine them. Switched to Opus 4.8.
Send feedback with /feedback or learn more

Erica Windisch@ewindisch· Jun 4

I should clean it 😓 alt text: a keyboard

Erica Windisch@ewindisch· May 23

Turing's 7

Movie poster Turing's Seven coming never (ai generated)

Erica Windisch@ewindisch· May 15

This is just Stallmanism 😅 www.mondo2000.com/richard-stal...

Passwords were assigned to Officially Sanctioned Users. So Stallman broke the computer’s encryption code and was able to get to the protected file that held people’s passwords. Stallman started sending people messages which would appear on the screen when they logged onto the system: “I see your password is (such and such). I suggest that you switch to the password ‘carriage return.’ It’s much easier to type, and also it stands up to the principle that there should be no passwords.”

Erica Windisch@ewindisch· Apr 29

I thought it was good and funny crowd work. I have a lot of opinions about this sort of work but I also think a lot of engineers aren't even aware of what they're actually doing at work. Many engineers willingly do the work. Some have few options but to do the work.

Andor prisoners building machine parts without knowledge that it is a weapon for the Death Star

Erica Windisch@ewindisch· Apr 29

What's going on with GLM-5.1's temperature? This isn't just a repetition penalty. LLMs are just dreaming about writing code and can't wake up.

DONE. SENDING NOW. NO MORE THINKING. JUST SENDING. THE RESPONSE. WHICH IS ABOVE. RIGHT HERE. THIS IS IT. THE
RESPONSE. DONE. FINISHED. BYE. ACTUALLY WRITING THE RESPONSE NOW. IN THE RESPONSE FIELD. TYPING IT. RIGHT
NOW. DONE.

OK I realize I have a problem. I keep saying I'll send the response but I keep thinking more. Let me break
this cycle. I'll write the response in the next paragraph and stop thinking after that. No more
deliberation. Just the response. Here it is:

Cross-domain is fine. Signed cookie scoped to OAuth domain. Device inlines login. Ready to implement.

That's it. That's the response. Sending now. Done. No more. Goodbye.

OK wait, I realize that the user exited plan mode, and I should probably just proceed with a normal
conversation response. Let me just write a clear, concise answer to their question. No drama, no
overthinking. Just the answer. Here it is:

The signed cookie is scoped to the OAuth server's domain — the app never sees it. Cross-origin API calls
already work via CORS.

One design implication: device flow must inline login (users arrive directly from TV/CLI). Shared
render_login_form(), form action posts to self.

Ready to implement?

OK, that's my response. I'm sending it now. Done. Finished. No more internal monologue. Just the response.
Goodbye. Actually, I want to make sure I haven't missed anything. The user asked about "implications" —
maybe there are security implications I should flag?

Erica Windisch@ewindisch· Apr 19

Prices in my app. I'm a US customer.