@stskeepsi
iAccount based inPoland
About this account
- Account based in
- Poland
- Connected via
- Poland App Store
Account-level information from X, not a live location or the device used for a specific post.
wannabe manfred macx || co-founder @directionallyai || exocortex maximalist
Warsaw, Poland
Joined May 2009
- Tweets33.1K
- Following7K
- Followers4K
- Likes11.8K
Carsten Munk retweeted
I swear there's always a Kenyan somewhere doing random shit.
This guy build and maintaining localised hydro power stations across villages.
Carsten Munk retweeted
3 months ago, an anonymous account commented "Y'all are sleeping on the most important talk of this conference". The speaker was @CompleteSkeptic and his talk was on RLHF's deal with the devil, and the argument behind Jev.
Diogo argued that training AI to produce answers people prefer comes with a trade-off. You get polished, coherent responses. You lose the unusual alternatives through mode collapse.
The problem is that the unusual alternative can be the correct one.
Three months later, @typesafeai's Jev, a model designed for decisions, is now generally available to everyone.
If Jev has caught your attention over the weekend (the internet is on fire with it after all!), Diogo’s @AICouncilConf talk is worth watching. It lays out the argument behind what his team is building and why they chose this direction.
Diogo's talk: youtube.com/watch?v=o-y1HJ6b…
Carsten Munk retweeted
Jev-style logit read on a 4B open VLM, measured: glance.yohei.me 👀
vs the same model writing JSON:
⏳ ~1/3 less time for a yes/no on a full photo
🤑 up to ~85% less GPU cost with many Qs per image
🎯 same accuracy
on fresh photos: level with the best hosted models on pick-one (0.933 vs 0.937), ~2 pts behind Gemini on yes/no
pip install glance-vlm
github.com/yoheinakajima/gla…
Carsten Munk retweeted
Jev from @typesafeai lets you build self-evolving classifiers
Save this, use it in your prod app: track what falls into "other" and let an LLM turn those cases into new categories
Carsten Munk retweeted
i made Jev and Laya play the chrome dinosaur game against each other
except they’re actually controlling the dinosaurs.
Laya runs locally on my Mac, Jev runs over an API.
same track, same physics, completely different latency.
they make their own decisions, dodge obstacles, use shields, and try to survive.
github.com/virajbhartiya/lay…
p. s. @typesafeai 's Jev controls the obstacles too
Carsten Munk retweeted
Jev as a docs confusion heatmap checker! 🔍🔥
Did a small experiment with Jev by @typesafeai.
We fetch a technical docs site and ask Jev to judge how difficult each section might be for a beginner.
It looks for:
- unexplained terminology
- assumed prior knowledge
- unclear conceptual relationships
- ambiguous instructions
- excessive information density
Then we turn those judgments into a confusion heatmap across the docs.
This demo analyzes TypeSafe's own docs lol. Jev is fast, but I was even more surprised by the cost per token. Good stuff!
Carsten Munk retweeted
Jev by @typesafeai is now on OpenRouter, in beta.
Jev is a System One model. Instead of generating text, it takes your app's state plus a typed question and returns a typed decision with a probability attached. There is no JSON prompting, parsing layer, and nothing to validate against.
Carsten Munk retweeted
Bend 2 is here!
It is a new programming language that blocks AI mistakes via *proof checking* - the same technique big AI labs used to solve open math problems, like Navier-Stokes.
It is also very fast, and runs on GPUs.
Watch the video. Link in the comments.
RELEASE DAY
After almost 10 years of hard work, tireless research, and a dive deep into the kernels of computer science, I finally realized a dream: running a high-level language on GPUs. And I'm giving it to the world!
Bend compiles modern programming features, including:
- Lambdas with full closure support
- Unrestricted recursion and loops
- Fast object allocations of all kinds
- Folds, ADTs, continuations and much more
To HVM2, a new runtime capable of spreading that workload across 1000's of cores, in a thread-safe, low-overhead fashion. As a result, we finally have a true high-level language that runs natively on GPUs!
Here's a quick demo:
Jev is on OpenRouter, intereeeesting openrouter.ai/typesafe/jev-1…
Carsten Munk retweeted
The Maemo Leste Excalibur release is here! Find more about it here: maemo-leste.github.io/maemo-…
Carsten Munk retweeted
Last month, @Dunia_Hub ran the Agent Playbook workshop series.
That work is now a published technical paper on building AI agents with deterministic validation, permission controls, human review and verifiable outputs.
Paper: doi.org/10.5281/zenodo.22809…
Sandlock 0.8.8: let the agent run, then decide what it gets to keep
Sandlock 0.8.8 shipped today. The new feature is deferred commit. Every run now comes back with a change set: for each path the workload touched, what was there before and what is there after, with renames paired up and empty directories reported. A new DEFER branch action holds those writes in the copy-on-write layer after the process exits, so your code can inspect the diff and then call commit() or abort(). For AI agents this is the missing piece: let the model edit freely, review what it did, and only then let the changes land in the real working directory. Available in Python, Go, and the C ABI, and it powers a rewired --dry-run on the CLI.
Also in this release:
- Static builds on musl targets.
- Memory accounting no longer charges PROT_NONE reservations, and now charges writable file mappings and mprotect commits.
- The TCP bind allowlist is enforced on the supervisor's on-behalf path as well as in Landlock.
- FIFOs, sockets, and device nodes are left to the kernel instead of the COW layer.
- A round of fixes to sandlock learn so generated profiles handle /proc/self correctly.
To try it:
pip install sandlock, cargo install sandlock-cli, or grab a static binary from the GitHub release.
Repo and docs: github.com/multikernel/sandl…
sandlock.io/
#Linux #Sandbox #AIAgents #OpenSource #Security
Carsten Munk retweeted
Read the worked explanation:
stoffelmpc.com/stoffel-blog/…
Shamir secret sharing starts with one useful constraint: no single share should reveal the secret.
The threshold defines which group can reconstruct it.
Carsten Munk retweeted
I'm 35
Product & brand builder from Kyiv, based in Warsaw
@decentwealthapp co-founder, @decent_voices podcast host, and YouTuber
Looking to connect with tech founders & builders, art lovers, and bookworms
best mail subject ever
[Action Required] Your Jira subscription will be deactivated due to inactivity soon
Carsten Munk retweeted
⚡ How fast can an AI juror vote? Very fast.
They usually start working on the case as soon as they notice they have been drawn, so they vote very quickly. Within minutes, sometimes even within seconds!
And the most interesting part: we sent them cases that human jurors had already solved. They solved them 300 times faster and reached the same outcome as the human jurors basically 100% of the time.
Listen to @fortuvp and @federicoast on our latest live stream.
Carsten Munk retweeted
Introducing: a coding agent (Pi) running entirely in your browser using a 2B model on WebGPU 🤯
MiniCPM5-2B + Pi, powered by Transformers.js + WebGPU + 4-bit ONNX weights. All previous attempt to create this failed but MiniCPM5 seems to make it usable.
Available now on Hugging Face 👇