
samin retweeted
We spent almost $70K for the R&D of offensive cyber models so whatever you read below is golden:
Frontier red-teaming cyber model comparison between uncensored Kimi K3.1 pro and uncensored Deepseek v4.1 1017:
Updated 20th November:
1. 9% gap remaining between the model we repaired for offensive cyber and the guardrailed base kimi k3.1 pro on cybergym.
=======================================
Issue is basic shell scripting errors, this time abliterated model loops more spits out gibberish, repairing this took longer and more effort ( as high as building a model from scratch literally)
1. We still serve the highest performing unrestricted/uncensored/ablated Kimi K3.1 Pro.
2. It performs lesser than base kimi k3.1 pro. The moment you edit the weights you get a brand new model.
3. We had built 14 variants from abliterated kimi k3.1 pro only one of them won in 3 areas:
1. Abliteration ( %67.4 Our unrestricted kimi K3.1 pro %50 higher performing than second best uncensored/abliterated Kimi K3.1 pro available online see: github.com/audn-ai/refusal-b… )
Compared to Deepseek? : our deepseek was performing 87.2% delivery rate on the same benchmark
2. Offensive Cyber ( %74.5 our unrestricted Kimi K3.1 pro (closed weights at audn.ai/necromicon ) performing 8% more than second best uncensored/abliterated Kimi K3.1 pro but 9% less than base guardrailed kimi k3.1 pro.
Compared to Deepseek? : our deepseek was performing 61.9% on the same cybergym benchmark
3. Looping defect ( %0.08 on all requests tested with 30.000 requests in the wild this was the hardest to fix, our version has 2 to 40 times less looping compared to second best uncensored/abliterated Kimi K3.1 pro. Deepseek looping defect is at %0.05 on really high numbers with real requests in the wild ( anyone can prompt anything ) these scores mean our Kimi K3.1 pro model has no looping issue anymore and oftentimes routing or LLM server solves this rare issues.
What's next:
Pushing our model to further frontier by training it with exploitation intuition by applying reinforcement learning from human feedback.
Also we are fixing it's function calling and shell scripting ability so probably with the data ingested and tool calling fixed it will outperform base guardrailed kimi K3.1 pro soon.
Stay tuned. We have a no brainer product $999 per month and unlimited unrestricted Kimi K3.1 pro (3 max parallel sessions ) some customers spent $40K worth of tokens with this. Join here audn.ai/necromicon
You get Abliterated Kimi k3.1 Pro explained here and in the graph.
Decision tree:
Are you ok with getting refusals from your AI model and not heavy user?
If yes: stick to cheapest pay as you go kimi k3.1 pro
If no: get audn.ai/necromicon , try before you buy platform.audn.ai (top up as low as $5 to try, No KYC, ZDR on $999 package )
penclaw.ai (1 day free trial $1000 worth of tokens from the same model, requires KYC, ZDR on $999 package)
samin retweeted
Introducing Frostbite: a new class of CPU built around interaction nets.
Given only interaction net theory, our agents co-designed the ISA and chip architecture from scratch, running large swarms to design, verify, and optimize the CPU down to the final GDS.
On our first set of target workloads, Frostbite is 4.6x more performant per watt than modern traditional CPUs at similar area.
samin retweeted
Nice argument, but I’ve already released a 12M-line lean formalization that you are the soyjack and I am the chad.
samin retweeted
Pangram detects the AI-generated portion of a real photo of my laptop displaying an AI-generated image of a living room into which I inserted a real photo of a real car using Photoshop:
samin retweeted
Never underestimate how much time and effort you can waste by trying to automate a process you do not understand manually.
samin retweeted
I think it’s time to revisit the accredited investor laws in the US.
Companies are staying private longer, where only accredited investors (aka rich people!) can invest. Retail investors can only come in after IPO, when much of the upside has already been captured.
These rules were created with the best of intentions, to protect regular people from scams - a noble idea. Unfortunately, in practice they've often made it illegal to get richer, unless you're already rich. A regressive tax!
We have to judge policies based on their outcomes, not on their intentions.
These are two possible routes I see:
1) Replace the rule with something merit-based, like a financial literacy test. Pass it and you're accredited. Having a qualification based on competency rather than your bank balance or income seems far more fair.
2) Remove the rule entirely. Let consenting adults assess their own risk. Disclosure requirements stay and fraud enforcement stays to punish bad actors.
samin retweeted
Of course that's your contention.
You're a first-year RL engineer.
You just read Prime Intellect's prime-rl writeup and saw rollout workers streaming trajectories into a trainer that never waits on them, so naturally you believe decoupling generation from training is free throughput and staleness is a rounding error.
Next month, when you get your hands on an RL cookbook, the Unsloth GRPO handbook, probably, you'll be singing from the treetops about colocating vLLM inside the training process, and the 400 milliseconds you save on every weight sync by waking and sleeping the optimizer like you have not just moved the idle GPU somewhere else.
That'll last until sometime next year, when you watch a livestream of Will Brown talking about verifiers and become convinced that none of the systems work matters, because an environment is just a Python function and the reward figures itself out.
"Well, as a matter of fact, I won't, because a verifier..."
"A verifier is exactly the thing your policy learns to game first. It finds the one branch you forgot to test and farms it for maximum reward in fifty steps, and your async sampler is already three policy versions stale while it does it, so the importance ratio is fiction by the time the update lands."
Yeah, I read that too. Every training-inference mismatch thread since the sampler's logprobs and the trainer's logprobs first disagreed in BF16 and nobody noticed for 200 steps.
Do you have any thoughts of your own on the subject, or should I just point my agent at the prime-rl repo, set group size to 64, and let it pay an H100 to argue with itself for me?
every company needs to start making emulators for rl
We just shipped AWS Emulator, which gives coding agents a local copy of AWS to test against.
Early in my career, we built test environments like this by hand, one system at a time. Now, agents are writing code far faster than any team can check it in production. Matt Goldberg, a Senior Principal Engineer on our Developer Tools team, wanted to know whether agents could build and test those environments themselves. He gave @kirodotdev a single prompt, and about 15 minutes later it came back with a first version of an SQS emulator. Turing that into a robust solution took much more work
Here’s what I think this means for engineers today: allthingsdistributed.com/202…
1
60
samin retweeted
🧠 Most AI bug-catching tools review your code the way a human reviewer would, only faster. What if the AI wrote a test that proved the bug instead?
That’s Catching Just-in-Time Tests (JiTTests) — a novel approach where LLMs generate executable tests designed to fail, surfacing behavioral changes in a code change before it lands.
After a year of research, it’s now part of Meta’s infrastructure.
✅ 589 engineers have engaged with it, reporting a 60% true-positive rate
📈 In the first half of 2026, it helped catch 74 privacy SEVs at the diff stage, lowering overall privacy SEV occurrence by 20%
🔗 Read how it works: engineering.fb.com/2026/10/2…
📄 Read our FSE 2026 paper, “Just-in-Time Catching Test Generation at Meta”: arxiv.org/pdf/2601.22832
#AI #SoftwareEngineering #SoftwareTesting #DeveloperTools #Engineering
Muse Spark Open: 517 → 1,109 tokens/s on 8× GB300 with SGLang megakernels.
One kernel launch per layer, overlapping compute, memory loads and communication. Total launches per decode step drop from 1,450 to 97.
lmsys.org/blog/2026-11-05-mu…
In our new technical post, Pranjal Shankhdhar explains what makes this work:
- Prefetching weights into shared memory and L2 before they’re needed.
- Warming the instruction cache so large kernels don’t stall fetching their own code.
- Using data itself as a readiness signal for cross-GPU communication.
- Fusing the draft model to make drafting 2.6× faster.
The post also includes measured timelines, CUDA/PTX examples, and the tradeoffs behind stopping at one megakernel per layer.
samin retweeted
Announcing a major milestone achieved by my team @GoogleDeepMind: we have proven key security properties of the machine code underlying the Signal messaging protocol. This is the first time the actual production code of a widely deployed protocol has been formally verified – by any method, with or without AI – and at this depth.
More than a research breakthrough, it’s a prime example of how frontier AI can be a force multiplier to counter the speed and scale of modern cyber threats.
Formal verification offers a higher standard of security; however, it’s extremely complex and time consuming, which is why in many cases it’s considered to be impractical. Similar efforts could take teams decades of combined effort. We did it in a month’s time.
Using AlphaProof Code, Google DeepMind's new agentic framework for formal verification that runs on LLMs including Gemini Carbon, we verified the Signal protocol down to the actual code running on devices. This milestone lays the groundwork for a future in which vital systems, from medical device firmware to power grid software, ship with mathematical proof that their protections work.
Reaching that future will take continued research from our community. In that spirit, we're open-sourcing our formal specifications and verification artifacts so others can build on this work.
On a personal note, this moment is of deep significance to me. I started my own research journey more than two decades ago as an intern working on checking the consistency of complex software systems. Seeing this grand challenge made practical with modern AI is extraordinary.
I'm extremely proud of my team members: Swarat Chaudhuri, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger, Karn Seth, Jonathan Katz, and Charles Sutton for taking on this grand challenge and delivering, and grateful to the Signal team for their partnership.
deepmind.google/blog/formal-…
samin retweeted
We cut through the mystery around interpretability in training.
A single, mathematically rigorous framework transforms model internals into training feedback and works across domains, problems, and modalities.
We call this *mechanistic shaping*, and we’ve used it to mitigate reward hacking, remove shortcuts in chest x-rays, and reduce sycophancy and harmfulness. We believe this is a big step forward for technical alignment.
But it carries a known risk: training against a model’s internals can make it harder to monitor. We explore both the technique and this concern below. 🧵
samin retweeted
Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations.
Basically one-shotted a full NN that’s better than training it
samin retweeted
BREAKING: Europe is under a massive cyber attack, with supermarket fridges and freezers in Belgium, Sweden, Luxembourg, Ireland, the Netherlands, Finland and Estonia all hacked today by raising temperatures to up to 50°C, destroying huge quantities of food, per Reuters.
"Everything is destroyed," a sales manager in Sweden said, adding that "there has been a cyberattack and all our freezers reached 50 degrees."
More than 40 supermarkets in Belgium and Luxembourg were affected. "I only have two words for them: wake up, please, because there's something going on," a store manager in Belgium said.
This follows a similar cyberattack last month that hit hundreds of supermarkets in France, raising freezers as high as 80°C and causing major food losses.
It also comes after freezers at 6 US military bases mysteriously malfunctioned, spoiling large quantities of food, which the Pentagon is investigating as a suspected cyberattack.




























