Noah Chander

Entrepreneurs First | ex-Amazon SDE | UCL MEng

Building for the cybersecurity industry

[Noah Chander @NoahChander]

@nick__price__ @Amank1412 cr7 cannot harden code to be unexploitable via memory corruption

[Noah Chander @NoahChander]

Wake up folks! The FBI, NCSC, and NSA issued an urgent warning to organizations globally running PHP to take immediate action.

PHP runs 70% of the internet. Two new critical bugs (CVE-2026-109618 and CVE-2026-109619) let an attacker run code on a web server with NO login and NO user interaction.

> memory corruption
> memory corruption
> memory corruption

Both bugs are written in C, and both slipped past years of review, fuzzing and hardening. AI had other ideas.

Patching will fix this sure, but it doesn't fix the next bug that we don't know about, and patches will come too slow in MOST cases.

At Cofibrant, we've already solved the worst of this. Our compiler extension compiles any memory-unsafe code (C/C++/COBOL/Fortran) and makes memory corruption exploits impossible. No code changes, no performance impact on modern hardware.

0% performance overhead vs Fil-C at 400%
40x more precise than Google + Microsoft's CFI
Sits automatic in CI/CD, no annoying flashing lights and buttons for developers to triage.

If you're responsible for C or C++ software that can't afford to be next, let's talk.
#cybersecurity #PHP #memorysafety #appsec #arewegonnamakeit?

[Noah Chander @NoahChander]

@yannispanagis riveting stuff yan if this is the tesco outside the office im gonna get some oreos rn

[Vijja (Pat) Wichitwechkarn @wvijja]

Little robot I built, powered by a local llm.
Weird how when everything’s running locally it feels like a new kind of ‘entity’…

#Robotics #LLMs #PhysicalAI

[Noah Chander @NoahChander]

Generalist already proved handheld data beats teleop (99% success on GEN-1) but they went broad, whereas a better play could be to go deep on precision. If you're at

@physical_int
@SkildAI
@GeneralistAI

OpenAI, Meta or similar and hitting the precision wall, this might be useful. Eager to chat about this and find out what ppl think about whether teleop quietly maxed out? #Robotics #EmbodiedAI #AI #RobotLearning

[Noah Chander @NoahChander]

Teleoperation has a hidden ceiling:

Even with perfect rigs, humans can't transfer full dexterity when remotely puppeteering a robot. Every model trained on that data inherits the limitation. It's not a hardware issue but rather it's the interface itself @amazon @physical_int

[Noah Chander @NoahChander]

The problem is this data has a different distribution than teleop data. You can't just fine-tune existing models on it, you need new techniques built for that shift. The real moat is hardware + data + those fine-tuning methods together @GeneralistAI

[Vijja (Pat) Wichitwechkarn @wvijja]

Going directly from human demonstration to robot deployment.

#Robotics #ImitationLearning #RobotArm #AI #RobotLearning #VLA #PhysicalAI #ManipulationResearch

[Noah Chander @NoahChander]

@wvijja Is external tracking > SLAM used in @Stanford's UMI?

[Noah Chander @NoahChander]

@elonmusk china

[Noah Chander @NoahChander]

@MarioNawfal even for egocentric data this isn’t gonna be the best quality but it’s gonna sell anyway @RealityLabs

[Noah Chander @NoahChander]

@rickb8s @sciencegirl at least he’s having fun in the meantime

[Noah Chander @NoahChander]

@sciencegirl those hands might be a bit too dexterous

[Noah Chander @NoahChander]

@GeneralistAI extremely cool but took 4 tries to open the bag - would not hire bro to put the fries in the bag

[Noah Chander @NoahChander]

@UnitreeRobotics looks sick for demos, but the trap is that most labs are still in teleop caps at ~35 demos/hour, forcing embodiment-specific jittery data, and kills natural human accelerations on contact-rich tasks.

[Noah Chander @NoahChander]

@pascal_bornet the bottleneck isn’t better models, it’s scalable data. Teleop locks you into slow, expensive, lab-only collection with embodiment mismatch. The q is how much IK tracking limits dynamic tasks in real deployments