Thanks to @NEAR_AI for welcoming us into the RED programme :)
So happy to be part of the fam.
We're building *Illumina*, an autonomous vulnerability discovery system. Can't wait to test it using confidential inference.
Privacy matters. A lot.
I was buying lunch today at my favourite tavola calda:
> Transaction rejected.
Revolut is down.
Hunger mixed with anger and concern.
Was it a cyberattack? There's no evidence of that so far.
Only a proper post-mortem can tell us what happened to Google Cloud.
One thing is clear: I'm even more locked in on building @SunnyGoesRogue, our agentic cybersecurity system for continuous auditing and vulnerability detection. ๐
Finding bugs in smart contracts is an information problem.
To catch one, you must know what the code does in each module AND at the interfaces, and explore the threat space.
Two things one agent with a limited context window struggles with.
We're building Azimo to fix it ๐งต
7 - Bug bounties
We're building it now.
We're targeting smart contract bug bounties: production-grade code, real stakes, measurable signal.
Every finding ships with an executable proof of concept, which is the artifact a submission needs.
Following @BerkeleyRDI MOOC on Agenting AI. Since llm-is-life, we got inspired to start a series of posts: "A bunch of cool things about AI a day".
Lots of cool things today: training pipeline, evaluation and systems. Buckle up!
๐Day 1: LLM Agents Overview by @yanndubs ๐
*Systems*: the bottleneck is feeding the GPUs data. FLOPs grew faster than memory & bandwidth
The metric: MFU = achieved/theoretical throughput
FlashAttention is the poster child: fuse kernels, tile the computation, recompute instead of reading
Do MORE math,touch memory LESS