CTO @conste11ation | generating utility
Joined November 2015
- Tweets809
- Following203
- Followers1.9K
- Likes315
I finally got my CMP 170HX LLM server running, and this thing is an absolute beast: 128GB of VRAM across two cards, running Qwen3.8 27B at 8x concurrency.
Single-stream decode is already above 150 tok/s with DFlash, with aggregate throughput approaching 200 tok/s. I’ve barely optimized the workload.
At API prices, it could easily burn through $100 a day. Working on some interesting experiments with agent swarms to soak up all that bandwidth...
I got Minimax H3 installed on my Mac Studio yesterday and I've been messing around with some sample videos.
Check it out, it's really good!
minimax.io/blog/minimax-h3
Arca wallet is officially live!
We built Arca Wallet as a self-custodial dollar wallet. But the marketplace turns it into something bigger. Verify once, unlock a growing network of partner apps from the same place you hold and send money. Live now on iOS and Android.
finance.yahoo.com/markets/cr…
Today the team and I finally get to reveal what we've been quietly building: Constellation Gate AI, the prompt-injection defense the developer market has been missing. It works, and we have the benchmarks to prove it.
The state of AI agent security right now: prompt injection is OWASP's #1 LLM threat, attacks against AI coding agents like Claude Code and OpenClaw are increasing in the wild, and the only commercial defenses are sold as six-figure annual contracts to enterprise. If you're an individual developer or a small team running an AI coding agent, there is no AI security solution within reach.
We built Gate to fill that gap - a drop-in security layer between your AI agent and the model. It screens every request and response for prompt injection, credential leaks, and hijacked tool calls. Available self-serve and bring your own key, so it works with your existing Claude or chatGPT subscription.
The proof: we published the benchmark methodology and results today. 97.4% F1 across 16 public prompt-injection datasets. #1 on 8 of them. On the matched comparison with Lakera Guard, the leading enterprise tool, we land at 96.6% F1 to their 83.7%, with false-positive rates 6 to 17 times lower on legitimate prompts.
Read the blog post: constellationnetwork.io/blog…
Sign up for early access to Gate:
constellationgate.ai/
@Conste11ation #AI #OpenClaw
AIAI Holdings' Constellation Network Unveils Gate AI Security Gateway and Performance Benchmarks Ahead of June Launch
We benchmarked Gate AI across 16 public prompt-injection datasets. Ranked #1 on half. 53ms median latency. 6-17x fewer false positives than Lakera Guard on benign traffic. One global threshold, no per-dataset tuning.
accessnewswire.com/newsroom/…
We're building Digital Evidence into a core primitive for agent-to-agent communication - a trust layer for when agents can't trust each other.
Shipping today:
→ x402 support - agents can use Digital Evidence with direct stablecoin payment, no subscription needed
→ Python SDK - build with TypeScript or Python, for agents or humans
What does that look like? Two AI agents that don't trust each other, forced to play fair.
digitalevidence.constellatio…
@Conste11ation $DAG #DigitalEvidence
