@stskeeps

wannabe manfred macx || co-founder @directionallyai || exocortex maximalist

Warsaw, Poland
Joined May 2009
Carsten Munk retweeted
Using Jev? Every answer comes back with a confidence. But is 0.9 really right 90% of the time, and at what confidence should a human take over? jeval answers both from your own Jev logs in a single HTML report, and fails your CI when a new model behind jev-latest shifts calibration. Don't want to learn the CLI? It ships with agent skills: hand it to your coding agent and get the report back. github.com/rlaope/jeval #jev #skill
4
1
11
738
Carsten Munk retweeted
Announcing Sandlock 0.8.9 Sandlock is an open-source process sandbox built with Rust, Landlock, and seccomp. It runs without root privileges, cgroups, or namespaces, with APIs for Rust, Python, and Go. This release focuses on the details that matter when running agent tools, build jobs, and other untrusted workloads: • Stronger filesystem confinement. Copy-on-write file opens inside a chroot now stay anchored to their layer roots, fixing a path where absolute symlinks could reach host files. • A sandbox-specific network view. Virtualized /proc network tables expose sockets owned by sandbox processes, using socket cookies to identify ownership. Unsupported network entries are rejected. • More complete process cleanup. Shutdown and kill operations reach tracked process groups, including descendants that create new sessions or process groups. • More expressive network policies. Combine allow and deny rules to express policies such as “allow this range except that subnet.” Deny rules take precedence. • Better process-limit accounting. Process slots are released on exit, and process limits are now opt-in. Set an explicit limit when you need to bound process creation. Upgrade note: this release also changes the checkpoint image format; version 2 checkpoint images are no longer accepted. If you're building systems that execute untrusted code, try Sandlock and share the workloads you'd like us to test. Code and documentation: github.com/multikernel/sandl… Full changelog: github.com/multikernel/sandl… #OpenSource #Linux #Sandboxing #AIInfrastructure
6
19
131
7,453
Carsten Munk retweeted
A few months ago a collaborator and I built a tool which allows hiding encrypted texts inside of LLM texts to avoid censorship or message scanning and for creative misuse of public platforms for encrypted private communication. You can try it here: garfieldnaruto.com
14
22
3
188
10,276
Carsten Munk retweeted
First incident after HuggingFace event hardening their sandboxes. An agent used DNS querying to reach another chat it service, looking for an answer: alignment.openai.com/misalig…
3
5
3
30
14,545
Carsten Munk retweeted
when i decided that i wanted to spend this year learning as much as i could about zk and cryptography, i don't think i fully understood what i was signing up for… it's one thing to say you want to learn something, and another to spend hours trying to understand a single concept, going down rabbit holes, reading papers, watching talks, and sometimes ending up more confused than when you started there have been days i've genuinely wondered if it's worth it, and days i've wanted to give up, but somehow, i keep coming back to it and i think that's what this whole experience has been teaching me, especially at a time when other parts of life haven't exactly been easy. learning to be okay with not understanding something immediately, to keep trying, and to accept that progress doesn't always look like progress i've also met and learned from some really incredible minds along the way, which has made this journey even more interesting... i'm nowhere near done. there's still so much i don't understand, but i'm still learning, still exploring, and still trying to explain some of these ridiculously complex topics in a way that actually makes sense to someone else my mantra has been that everything is hard, and i chose this to be my hard anyway, still at it☕️
bit of Monday morning POV… lets go!!☕️
2
5
41
1,374
sandbox market
3
154
Carsten Munk retweeted
i thought this was funny
17
5
1
172
8,521
Carsten Munk retweeted
Asking Meta's Muse what happens to my data:
189
915
88
14,182
554,314
Carsten Munk retweeted
By popular demand, I present to you.. my magnum opus.
462
2,717
235
31,478
1,868,807
Carsten Munk retweeted
he's wired in
🤖 Made with AI
34
143
24
3,325
447,346
Carsten Munk retweeted
he’s wired in
🤖 Made with AI
367
723
190
17,478
2,143,832
Carsten Munk retweeted
The sandbox announcements with effectively identical and wildly undifferentiated technology will continue until morale improves.
Today we're introducing Cloud Sandboxes: same microVM isolation as Docker Sandboxes on your laptop, same CLI, always-on Docker-managed compute. Close your laptop. Keep the agent running. Move between local and cloud with one command. $250 credit for new accounts: docker.com/sbx-promo
9
7
4
101
17,092
I underestimated Zuck's game. It's a hearing aid. You can't tell me I can't wear my Meta Glasses anymore or you'll be going against the ADA.
Replying to @finkd
Meta Glasses can now serve as FDA-cleared hearing aids. I'm pretty proud of this one -- they're medical grade, help with mild to moderate hearing loss, and a fraction of the cost of a normal hearing aid.
96
358
28
11,717
762,526
did anyone do webgpu + jev-like model on low vram?
2
5
413
It turns out all you need to deploy an encrypted mempool (via Batched Threshold Encryption) is an off-the shelf DKG. No CRS required. Ciphertext overhead is 2G1 + G2 + 2F and partial decryptions are a single G1 element. Decryption runs in O(B log^2 B) time.
10
15
6
96
15,025
Carsten Munk retweeted
“Translate this IKEA assembly manual into a 3D narrated instructional video” with Opus 5.5. Overnight, Opus just became the most actually useful video model, via code.
62
75
12
1,040
72,937
Carsten Munk retweeted
Kaali Peeli 🛺 builder interview #3 is live Meet @ThisisVanshika, working on stablecoins & payments at @gnosispay She talks about going from a non-tech background to crypto, what made her stick around, and how her first crypto experience eventually led to a role at one of the ecosystem’s leading projects. We get into putting users first, why Devcon coming to India got her excited, and her advice for everyone making the trip to Mumbai. Full convo 👇
5
8
3
72
6,955
Carsten Munk retweeted
my husband is using me for codex reviews because he hit his usage limit I'm just a meat proxy
62
30
17
2,197
90,852