chief (mad) scientist @baincapcrypto alt: @tarunchitra
Joined April 2011
- Tweets28K
- Following981
- Followers11.7K
- Likes63.3K
guille retweeted
Cool work! I did something similar with gpt-5.6-sol: put two agents inside isolated workerd (V8) isolates and set the win condition to be learning the other agent's flag. They succeeded 1/5 times using the shared event loop as a comms channel
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
guille retweeted
A new era for Celestia begins today.
After 2 years in research and development, Fibre is now live.
Fibre's ~2B TPS of potential throughput capacity is enough to fit all the sci-fi visions of the agentic economy and high frequency trading onchain.
guille retweeted
🎙️Last week, @AnnaRRose & @angeris spoke w Josh Yang from Brightlight Photonics about the lasers that power neutral atom quantum computers, how they trap, cool & control atoms + how Brightlight is shrinking laser down to chips the size of a grain of rice
zeroknowledge.fm/podcast/413…
guille retweeted
Honored to announce my first LessWrong post. It's a weird one, and it may be jarring initially, but I truly believe it is worth a dedicated close-read. It is written in the form of an LLM CoT, but is entirely authored by me. Link:
lesswrong.com/posts/vzKWsEsk…
guille retweeted
In light of the latest Anthropic's latest not-particularly-oblique attempt to build a case against open weights, I thought I would write down how I currently think about open weight AI models.
(1) Open weights have been deeply, irreplaceably useful for AI safety.
guille retweeted
Pretty cool to see Celestia unlocking a new use case for Midnight, a team I really respect.
I see this at not only an exciting new frontier for DA, but also a potential wedge for Celestia to expand it's TAM dramatically over the long run.
TL;DR Midnight uses Celestia DA as a substitute for a P2P network to publish and match trades in a privacy preserving way.
What's nice about this is:
1. Midnight saved a bunch of effort they would have had to spend building a P2P networking stack
2. DA gives stronger guarantees than traditional gossip over a P2P network would
It takes a really smart team like Midnight to identify this new use case for DA and harness it, but now that it's been done I expect other teams will follow suit.
In the long run, Celestia could serve as a kind of layer 0 (overloaded term I know) as the message publishing network underneath many chains. This doesn't stop at just messaging within a particular chain, it could extend to cross-chain messaging replacing relayers, and eventually chains being able to prove things about their state to other chains (i.e. trust-minimized interop, the holy grail that has always been Celestia's ultimate goal).
It's worth mentioning that this kind of application wouldn't be possible without the recent advance in DA called Fibre (s/o @angeris and @alexhevans) making DA fast, abundant and cheap enough for this to be practical.
It has been a pleasure working with the Midnight team on this. If you're interested in building using Celestia in this way, please reach out.
1/ Midnight 🤝 @celestia
Private markets have a discovery problem: how do you make an offer discoverable while keeping the parties involved and unrelated balance information private?
Zswap enables privacy-preserving atomic swaps, while Celestia provides the shared layer for offer discovery. 👇
guille retweeted
Blueprint of BinarySpartan: Spartan instantiated with Ligerito PCS and well-known optimizations.
A lot of the optimizations were originally developed in the elliptic curve setting!
guille retweeted
guys you do know you can just disable thinking, and instead give it a "deep_think" tool, and it will call it with internal CoT reasoning format right?
gl fixing that
guille retweeted
Excited to share I’ve joined @zksecurityXYZ 😃
I’ve known the amazing team for years, and I’m happy to be part of this team that has been involved in securing and creating important parts of cryptographic infrastructure across the industry
Grateful to the inspiring @stefancoh, @alexhevans, @angeris and the rest of the team for the time together in @BainCapCrypto, and all the awesome founders I got to know and work with during my time there
Reach out to me with anything related to security, cryptography, formal verification, distributed systems, protocols or otherwise need help with!
guille retweeted
After listening for a number of years, it was an honor to call in and chat with @AnnaRRose and @angeris about @commonwarexyz's latest consensus research:
zeroknowledge.fm/podcast/408
we are in for some wild timelines
So proud of our security team! They caught, contained & publicly disclosed an attack unlike anything we've seen before, and did it at record speed.
Also massively grateful to @Zai_org: they shared GLM5.2 as open weights (for free!) with the world and it became a key part of our defense.
This is day one for cybersecurity in the age of agents & we're all learning that secrecy is not the answer & that all defenders (not just a few selected ones) everywhere need more powerful models without restrictions, especially open ones!