Solo founder @cyntrisec | AIR v1 (IETF) + EphemeralML. Building confidential AI with hardware attestation + signed receipts. Seeking design partners → DMs open
My AWS 10,000 AIdeas article is live: AI Runtime Assurance.
A flight recorder for AI agents on AWS.
Would appreciate a 👀 or 🩷:
https://t.co/bMyqz5DNPY
#aideas2025#AWS
Isn't it true that you and your startup were originally doing the exact same thing -automating the work of an entire team so that one person could now replace them, effectively 'killing' hundreds of jobs? But as soon as it happened to you, Claude is suddenly the villain. Or am I missing something, or is bashing Claude on X just the latest hype?
Today, we updated Gemini 3 Deep Think to further accelerate modern science, research and engineering.
With 84.6% on ARC-AGI-2 and a new standard on Humanity’s Last Exam, see how this specialized reasoning mode is advancing research & development 🧵↓
This upgraded Deep Think mode is already driving discovery and helping researchers tackle the "unsolvable" — from spotting flaws in research papers to optimizing semiconductor growth.
confidential-ml-pipeline v0.2.2 is live 🚀
#OpenSource Rust framework for confidential multi-stage ML inference across TEEs. Built on @cyntrisec confidential-ml-transport (attested encrypted channels over TCP/VSock).
✅ Model sharding across stages
✅ Encrypted activation relay via untrusted host
✅ 1F1B-style pipeline scheduling
✅ Nitro / SEV-SNP / TDX backends
🧑💻Repo: https://t.co/yTILhRpP25
🦀Crate: https://t.co/qytO403tp5
📚Docs: https://t.co/9S8o9e0FPB
Fun fact: The X algorithm promotes AI slop posts 100x more than actual technical critique. Almost like someone wants us to believe the replacement narrative. Or maybe my takes just suck 🤷♂️
Built and benchmarked something I’ve wanted to see for a long time: real confidential LLM inference, not just slides.
I ran GPT-2 inside AWS Nitro Enclaves over encrypted/attested channels (Rust):
- TTFT: 84–92ms
- Gen p50: 42.0ms/token
- Gen p95: 42.9ms/token
- 23.8 tok/s
Then tested 1-stage vs 2-stage vs 3-stage pipeline sharding.
Result: multi-stage stayed close in latency (+~5–9% vs 1-stage), outputs stayed identical.
Main takeaway for me: confidential multi-stage inference is practical today, and we can prove it with reproducible benchmarks.
Building this in public at @cyntrisec.
If you work on confidential AI infra, I’d love your harshest feedback.
Hey folks, it's Cyntrisec (@cyntrisec) here! Back in early 2026, I built "confidential-ml-transport," a Rust library for secure, attestation-bound encrypted tensor transport in confidential ML inference—perfect for TEEs like AWS Nitro Enclaves. It features a 3-message handshake, ChaCha20Poly1305 encryption, pluggable attestation, and tensor handling up to 32 MiB, all open-source on GitHub. Sorry, @github, if my project caused any server hiccups from all the secure AI vibes—hope you're not having issues on my account! Check it: https://t.co/RgxymcpjOq 🫠