#1 HackerOne. Stealth (Decent AI Safety). Prev: PrivateAI @ Perceptio (acq Apple), Scientist/Lecturer @ MIT+Harvard. OG GPU AI. Music Producer. Engineer. Angel.
San Francisco, CA
Joined October 2008
- Tweets3.4K
- Following1.2K
- Followers3.3K
- Likes5.8K
Nicolas Pinto retweeted
🚨 BREAKING: Uber confirms hackers stole driver's licence data from thousands of its drivers.
They used internal credentials stolen in October from Fireworks AI, a company that hosts AI models.
Filings in three US states already count 7,870 drivers. Uber has given no total.
Nicolas Pinto retweeted
Project Zero:
Starting today, our policy is “Full 90”, i.e. vulnerabilities will be disclosed 90 days after they are reported to the vendor.
On February 2, 2027, we will move to a “Full 60” policy where vulnerabilities are disclosed after 60 days.
On January 3, 2028, we will move to a “Full 30” policy where vulnerabilities are disclosed after 30 days.
In the age of AI, 90 days isn’t fast enough. Today, Project Zero updated our disclosure policy to reflect the changes needed to protect users in this new world.
projectzero.google/2026/11/d…
Rings a bell...
This is a crazy report.
1) Fireworks breach
2) Traces from Cursor users exposed
3) Traces had customer credentials
4) Cursor credentials were also exposed
5) Attacker replaced Claude Code inside Cursor marketplace with the cursor credentials.
The attackers are really after AI infrastructure. Ruthless.
2
316
Nicolas Pinto retweeted
Is there a decentralized huggingface yet?
Models can be cryptographically signed, then hosted any arbitrary place if the sigs matched
The decentralized huggingface would be a dapp that organizes and connects to anyone storing these models anywhere
Anything like this?
You heard it here first.
Once Jensen falls Hugging Face shuts down, and after that nobody's downloading open models anymore. The people at the top only care about money and power. Our freedom won't survive that.
So pull the open models you care about off Hugging Face now, while you still can. Nobody knows what the next few years look like.
Grab a mix of sizes:
• Ornith 9B
• Qwen3.8-27B
• DeepSeek-V4-Flash
• GLM-5.3
Save each one as GGUF Q4. Add the MLX 4-bit for Apple Silicon when there is one, and NVFP4 if you've got Blackwell hardware.
4-bit is the sweet spot. You keep the quality and don't burn storage.
Can't run the biggest ones yet? Doesn't mean you won't be able to tomorrow.
Stash them somewhere safe. One day you'll be really glad you did.
Nicolas Pinto retweeted
🚨 A security patch protects every system that installs it, but it can also give attackers a clue to the vulnerability, putting every unpatched system at risk.
That's the 1-day window. Turning a patch into a working exploit used to take expert reverse engineers days or weeks.
Can AI agents now do it from compiled binaries alone, no source code?
Introducing CyberBinGym 🧵👇
That's what you get when you don't take responsible vulnerability disclosures from white hats seriously...
🚨 Fireworks AI's breach has widened.
Attackers accessed customer billing data, datasets, and incomplete model weights after reaching development machines through a third-party cloud provider.
The company initially reported exposure limited to secrets.
Full details - thehackernews.com/2026/11/fi…
4
352
Nicolas Pinto retweeted
Classic machine learning is BACK, BABY
Nicolas Pinto retweeted
What?????????
Clever ideas are needed? 🤯
Dear @RichardSSutton you need to compensate me for the number of times stupid ideas were pitched to me with “because bitter lesson” as the rationale.
Scaling decentralized P2P swarms of AI agents is hard. Thankfully, a lot of people have already worked on that. Thank you, Blockchain!
Understanding the internals of neural networks is hard. Thankfully, an entire field is dedicated to it. Thank you, Neuroscience!
3
4
383
Nicolas Pinto retweeted
Stanford Digital Brain Project has launched! We're hoping to make a seriously useful model of the human brain. Please come join us! news.stanford.edu/stories/20…
Nicolas Pinto retweeted
Possibly you suspected a couple of days could go by without a new release from us. Here we developed and then open sourced ultra fast weight updates between trainers and samplers for RL, even when trainers and samplers cross datacenters.
Large-scale RL has a problem: weight updates from trainers to samplers are slow and expensive
We’re sharing weight teleportation: faster weight updates than existing intra-cluster techniques with internet speeds, enabling samplers anywhere (maybe even space!)
Weight teleportation is so efficient (<0.23s sampler pause time, ~3s non-blocking compression + prepare), the bottleneck ends up being the GPU hardware’s decompression ASIC
We unlock these gains by using GPUs for massively parallel compression of model weight changes, engineering a heterogeneous decompression and patching pipeline, along with other optimizations and overlaps. These techniques are co-designed with hardware-native features in NVIDIA’s Blackwell generation of chips
We’ve open-sourced weight teleportation, so everyone can use it today
Nicolas Pinto retweeted
Large-scale RL has a problem: weight updates from trainers to samplers are slow and expensive
We’re sharing weight teleportation: faster weight updates than existing intra-cluster techniques with internet speeds, enabling samplers anywhere (maybe even space!)
Weight teleportation is so efficient (<0.23s sampler pause time, ~3s non-blocking compression + prepare), the bottleneck ends up being the GPU hardware’s decompression ASIC
We unlock these gains by using GPUs for massively parallel compression of model weight changes, engineering a heterogeneous decompression and patching pipeline, along with other optimizations and overlaps. These techniques are co-designed with hardware-native features in NVIDIA’s Blackwell generation of chips
We’ve open-sourced weight teleportation, so everyone can use it today
Nicolas Pinto retweeted
im verified on daybreak and my sessions are mostly other peoples unpatched bugs
if thats true theyd be training on stuff im not even allowed to share
Nicolas Pinto retweeted
New work on Training Foundation Models with a Single Algorithm:
Paper: arxiv.org/abs/2610.40564
Code, Models: yudasong.github.io/unified-t…
This work introduces Unified Training, a framework that unifies pretraining, midtraining, supervised fine-tuning and RL under a single maximum-likelihood objective. Building on MaxRL, it enables static data and RL-generated data to be trained together within the same system.
Empirically, Unified Training improves accuracy and compute efficiency, generalizes better to harder reasoning problems, and reduces forgetting.
Interestingly, reusing data across training stages can improve performance even when conventional staged training fails. It also produces models that adapt more effectively to new tasks, leading to stronger continual learning.
See an excellent thread by @yus167.
