Rafael David Tinoco retweeted
‼️ A bug that sat in Nvidia's Linux GPU drivers for over 4 years let a single user on a shared GPU server take over the whole machine.
Security researchers at Stealthium found it and named it ForgePtr. An unprivileged container with GPU access was enough to get root on the host, then load unsigned code into the kernel despite Secure Boot.
On rented GPU clouds, that means one customer could seize the server, and other customers' AI models with it.
Nvidia fixed it in September in driver 615.71.09. The researchers published their write-up on 12 November.
Rafael David Tinoco retweeted
Your Linux boot screen can now be more than a wall of scrolling text.
bootvine is a new GPL-3.0 tool written in Rust by developer Willem Vanhulle. It replaces the usual kernel and systemd log spew with a live tree of the systemd jobs still pending, plus a timeline of the slowest units that already ran. It also follows shutdowns, handles LUKS passphrase prompts in place, and when a unit fails it shows why, with that unit's last journal lines.
Vanhulle says he built it "out of frustration with the boot screen of most Linux distributions: a wall of log lines scrolling past too fast to read, which never gave me any idea of what was going on."
It needs systemd 246 or newer as PID 1, libdbus and libsystemd, and a systemd initrd if you want to see the early boot too. There are install guides for NixOS, Ubuntu and other distros. s
The latest release is v0.1.3, and it hit the Hacker News front page today.
Rafael David Tinoco retweeted
Jensen Huang in the Matrix, learning how Chinese spies work
Wait till the end 👀
🤖 Made with AI
Rafael David Tinoco retweeted
Replying to @RichardSocher @Recursive_SI
Congrats Recursive team on the amazing model! 🎉 We made some Flint GGUFs so you can run it locally on 24GB RAM! huggingface.co/unsloth/flint…
The Commodore 64 Keyboard Font, For Everyone ift.tt/PR3jVCT
Rafael David Tinoco retweeted
I'll be offering an Introduction to Linux Kernel Exploitation training in English, to be held online next year (likely September). This will be a whole-month training, totaling 48 hours split across 12 sessions. We will have live classes with me, and the sessions will be recorded and provided to students within 24 hours for those who can't join live. The classes will happen on Monday, Wednesday, and Thursday from 07:00 PM to 11:00 PM (GMT-3), meaning 4 hours of class per day.
The training is for people interested in getting started in Linux kernel exploitation. We will cover the fundamentals: computer architecture, paging and segmentation, C programming, assembly, the Linux ecosystem, and development/APIs, and then how to exploit vulnerabilities and bypass mitigations (SMEP, SMAP, and KASLR).
Besides the training, we provide support for students as soon as they enroll. We have an exclusive mailing list where I share a lot of content—usually much more detailed than what I share here. The mailing list is also used to help better prepare those with less experience several months before the training happens.
Even though the training is focused on Linux kernel exploitation, I believe it's interesting even for those not looking to work directly in research. I talk about many related topics, from system performance and development to my 10+ years of experience in security research, malware/rootkits in general, and more.
The class will be limited to 10 seats, as this will be my first time teaching it in English. I'll be pleased to help as many people as possible not just learn more, but truly grow and reach the next level in their careers.
Would you be interested? If so, reply here, send a DM, or contact us through our website:
allelesecurity.com/contact
Rafael David Tinoco retweeted
⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight!
The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens.
125B parameters + 51B N-gram embeddings, with just 6B activated per token. Unmatched cost-efficiency.
What's new: 🥳
- Next architecture: GDN + QSA hybrid attention, Gated Residual, N-gram Embedding & Muon optimizer, serving as a precursor to the architecture used in Qwen4.
- Dramatically lower training and inference costs: trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board with especially strong gains in coding and office tasks.
- Strong performance: scoring 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld, and 95.7 on MathVision (with CI).
- 262K native context, extensible to 1M with YaRN.
We’re also releasing the weights for Qwen3.8-Flash-Next, giving the community an early look at the new architecture we’re exploring for Qwen4.🚀
We can't wait to see what you build with Qwen3.8-Flash!👀👇
- Blog: qwen.ai/blog?id=qwen3.8-flas…
- Technical Report: github.com/QwenLM/Qwen3.8-Fl…
- Hugging Face: huggingface.co/Qwen/Qwen3.8-…
- ModelScope: modelscope.cn/models/Qwen/Qw…
Rafael David Tinoco retweeted
Replying to @Alibaba_Qwen
Thank you thank you Qwen team! Qwen3.8-27B is just amazing! 🥰
We made Dynamic GGUFs so you can run it locally: huggingface.co/unsloth/Qwen3…
The model is now also supported for running and fine-tuning in Unsloth Desktop.
Rafael David Tinoco retweeted
spent a week at work building a Pac-Man clone 😅
ghosts = attackers. pellets = patches. 90 seconds on the clock.
my run: 26,862. beat me 👇
Rafael David Tinoco retweeted
🤯 Unsloth released the fastest Qwen3.6-27B MTP GGUF I've tested. Time to upgrade.
Compared to the previous GGUF, Q4/Q6 XL versions are 👀 ~55% faster!
On a single RTX 5090:
✅ 114 tok/s — UD-IQ2_M (MTP)
✅ 93 tok/s — UD-Q4_K_XL (MTP)
✅ 75 tok/s — UD-Q6_K_XL (MTP)
💨Fastest MTP quant is 3.3x faster than the old Q8_0 baseline (35 tps)
262K context + tool calling. All on one 5090.
* compiled from the MTP PR branch ('am17an:mtp-clean', build b9117-ebe4fca4b)
🤖 Made with AI
💥 Introducing "Dirty Frag"
A universal Linux LPE chaining two vulns in xfrm-ESP and RxRPC. A successor class to Dirty Pipe & Copy Fail.
No race, no panic on failure, fully deterministic. ~9 years latent.
Ubuntu / RHEL / Fedora / openSUSE / CentOS / AlmaLinux, and more.
Even if you've applied the "Copy Fail" mitigation, your Linux is still vulnerable to "Dirty Frag". Apply the Dirty Frag mitigation.
Details:
dirtyfrag.io
Quantization can make an LLM 4x smaller and 2x faster, with barely any quality loss. But what *is* it? @samwhoo crafted a beautiful interactive essay explaining it from first principles, aimed at coders, not mathematicians.
ngrok.com/blog/quantization
Stressing poll(), timerfd, scheduling vruntime, bad work queue cleanup logic, maximizing a tiny race window, joining with side channel, stack pivoting, this work is impressive (unfortunately its red, but amazing ;)
103
Rafael David Tinoco retweeted
We've spent years building LlamaParse into the most accurate document parser for production AI. Along the way, we learned a lot about what fast, lightweight parsing actually looks like under the hood.
Today, we're open-sourcing a light-weight core of that tech as LiteParse 🦙
It's a CLI + TS-native library for layout-aware text parsing from PDFs, Office docs, and images. Local, zero Python dependencies, and built specifically for agents and LLM pipelines. Think of it as our way of giving the community a solid starting point for document parsing:
npm i -g @llamaindex/liteparse
lit parse anything.pdf
- preserves spatial layout (columns, tables, alignment)
- built-in local OCR, or bring your own server
- screenshots for multimodal LLMs
- handles PDFs, office docs, images
Blog: llamaindex.ai/blog/liteparse…
Repo: github.com/run-llama/litepar…
Rafael David Tinoco retweeted
Are you looking for the best model to run locally on your hardware? Instead of pulling different models and trying, you can just get recommendations for a given category using LLM Checker (github.com/Pavelevich/llm-ch…).
👇👇👇
Rafael David Tinoco retweeted
YES! Someone reverse-engineered Apple's Neural Engine and trained a neural network on it.
Apple never allowed this. ANE is inference-only. No public API, no docs.
They cracked it open anyway.
Why it matters:
• M4 ANE = 6.6 TFLOPS/W vs 0.08 for an A100 (80× more efficient)
• "38 TOPS" is a lie - real throughput is 19 TFLOPS FP16
• Your Mac mini has this chip sitting mostly idle
Translation: local AI inference that's faster AND uses almost no power.
Still early research but the door is now open.
→ github.com/maderix/ANE
#AI #MachineLearning #AppleSilicon #LocalAI #OpenSource #ANE #CoreML #AppleSilicon #NPU #KCORES
DeepSeek just dropped a new paper — and it's not about a new model.
It's about the infrastructure bottleneck that limits every agentic LLM at scale. And their fix nearly doubles throughput.
DualPath (arxiv.org/abs/2602.21548): the bottleneck in multi-turn agentic inference isn't compute anymore. It's storage bandwidth.
When agents run at scale, every new turn reloads the full KV-Cache from external storage. At long contexts + high concurrency, this I/O completely dominates. The GPU sits waiting on the NIC.
In standard disaggregated setups (prefill + decode engines), there's a brutal asymmetry:
— Prefill engine NIC: saturated loading KV-Cache
— Decode engine NIC: sitting completely idle
The bandwidth is right there. It's just not being used.
DualPath adds a second loading path:
Path 1 (existing): Storage → Prefill engine
Path 2 (new): Storage → Decode engine → Prefill engine via RDMA
The idle decode NIC now does real work. RDMA transfers over the compute network avoid congestion without touching latency-critical inference traffic. A global scheduler dynamically balances load across both paths in real time.
Results on 3 models with production agentic workloads:
📈 Offline throughput: up to 1.87×
📈 Online serving: 1.96× average — without violating SLO
As agents get more capable — longer context, more tool calls, persistent memory — KV-Cache I/O only gets worse.
DeepSeek is solving this at the infrastructure layer while everyone else is focused on model benchmarks.
One of those "obvious in hindsight" ideas: the bandwidth was always there.




























