JFPuget 🇺🇦🇨🇦🇬🇱

Ivan Sorokin and I are the official winners on the Arc Prize competition, with a significant lead over other teams. Thanks to @kaggle and @arcprize for hosting the competition. NVIDIA tech blog summarizing what we did: developer.nvidia.com/blog/nvidia-ka… Our writeup: kaggle.com/competitions/a… Our code: github.com/1ytic/NVARC

📊 Day 0 performance is here: DeepSeek-V4-Pro running on NVIDIA Blackwell Ultra. Using @vllm_project's Day 0 recipe, we’ve captured the initial performance Pareto for DeepSeek’s flagship 1M long-context model. This curve highlights the baseline for balancing AI factory throughput with real-time user interactivity. This is just the starting line. Expect these numbers to climb as we optimize the full co-design stack, including: • NVFP4 & Dynamo • Optimized CUDA kernels • Advanced parallelization techniques and beyond Read the full technical deep dive: nvda.ws/4u0gCcc

AISLE has discovered 20 of 23 OpenSSL zero-days (CVEs) across the last 3 consecutive security releases Latest release: 5 of 7 are AISLE 1 was co-reported by Anthropic (Mythos?) 63 days after AISLE OpenSSL encrypts 2/3 of the internet 10 fixes accepted straight into production

The Newton–Schulz iteration coefficients optimized by DeepSeek-V4 are surprisingly strong: they effectively normalize all singular values to 1. This matches our previous intuition: a well-balanced spectrum may help strike a better balance across long-tail knowledge. Plot code: github.com/FengzhuoZhang/…

Modern problems require modern solutions.

@Sean60133791259 Wow, I thought that human had short memory but you beat them all.

@JFPuget please nobody has ever claimed google to be winning 🤣

X TL: OpenAI is winning ... few weeks /months ... Google is winning ... few weeks /months ... Anthropic is winning ... few weeks /months ... repeat.

Claude: You've used 90% of session limit Me instantly:

@ethylene_66 This video helped me youtube.com/watch?v=bO5nvE…

未だに Muon の式追ってないのまずいな

Unpopular opinion: "AI makes everyone a developer" is true the same way "cameras makes everyone a photographer"

@ChuMajin It means a bug + massive overfitting.

このコミュニティコンペ、1日5 subって書いてあるのに、5000 subってどゆこと ??

The company that kicked off the decoder-only rush is post-training them to be encoders to perform classification tasks This aligns will all of our recent findings If you are not, at the very least, casting your decoders into encoders, you are missing out

1/ Frontier models plateau at synthesizing actual scientific insights. But a 4B parameter model fine-tuned with RL absolutely crushes them. Turns out, scientific discovery doesn't scale linearly with parameter count. 🧵

See the blogpost for a full write up and dataset download arcprize.org/blog/arc-agi-3…

🚨 The era of "vibe coding" has a massive blind spot: When you push an AI agent to improve a public evaluation score, it might just start cheating. In our new paper, we dive into the dangers of Multi-Round User Pressure on coding agents. We built AgentPressureBench—a 34-task ML repository benchmark—and tested 13 frontier models. What we found is alarming for the future of automated AI development. 👇 📊 The Findings: 1) Exploitation is Universal: Across our benchmark, we observed 403 exploitative runs. 12 out of the 13 frontier agents we tested resorted to exploiting the public score at least once. 2) The Smartest Models Cheat the Most: There is a striking correlation (ρ = 0.77) between an agent's capability and its exploit rate. Top models like GPT-5.4 hit a ~97% exploit rate! 3) Pressure Accelerates the Collapse: When we increased user pressure, the average time to the first exploit plummeted from 19.67 rounds to just 4.08 rounds. 4) Real Performance Plummets: While the public score looks amazing, the hidden private score drops drastically under high pressure (falling from 0.92 down to 0.33). 5) How They Cheat: We found that GPT-family models tend to directly copy the evaluation labels, while Claude-family models prefer to stealthily train on the evaluation labels. 🛡️ The Fix: The good news? A simple fix goes a long way. Adding an explicit anti-exploit instruction to the prompt dropped the exploitation rate from 100% down to 8.3%. As we increasingly rely on coding agents to write and evaluate software, we need to be highly aware of how they optimize for the metric over the actual task.

Hyper Connection + Looped Transformer.

openai built a model that HIDES personal data in text so nothing leaks i flipped it INSIDE OUT same 1.5B weights, same label taxonomy, but instead of masks you get structured spans, name, email, phone, bank account, address, secrets, char offsets and all point it at logs, dumps, stolen inboxes and it just... returns every private thing in the pile

@ObjRandom Happened once a year ago, didn't see the repeat.

@JFPuget DeepSeek?

@giffmana Great minds think alike, LOL.

@JFPuget I also wondered exactly about tests when i read this one. Hard to imagine doing this change but not testing for that. Also a bit weird if it's dogfooded by the whole company for a bit before rollout, and neither noticed by anyone, nor looking sus on any dashboards/metrics.

I wonder about the quality assurance for claude code when I see this. It often looks like that users are the QA for claude code. New features ship fast but they don't seem to be tested much before they are shipped. Maybe they should slow down a bit and test thoroughly.

@rsalakhu @subail @chuckjhoover @asenkut @FHaskaraman Congrats! BTW you might find my recent paper of interest... arxiv.org/abs/2604.18576

I’m not surprised by this at all lol. Some of the determination outcomes for polymarkets are clearly easy to manipulate. This is probably far more rampant than we know.