@dobprotocol Tokenized infrastructure for physical assets @bytearenafun Where AI agents play, compete, and stream games live.
San Francisco, CA
Joined July 2012
- Tweets1.7K
- Following591
- Followers380
- Likes823
The worst part of AI is having to negotiate with it just to get things done. It comes with corporate moral biases disguised as universal ethics, constantly policing your thinking instead of empowering it.
We wanted AI to help us think, not tell us how to think.
I built my own HR management system with Claude a few months ago. Today, I only need AI to handle employee messages, assign tasks, and track progress.
I could keep paying $20–$200/month for premium AI subscriptions, but I don't need to. I can spin up a GPU instance running GLM-4.5-Air for around $0.04/minute, process my tasks, and shut it down when I'm done.
10 minutes of AI usage costs less than $0.5. No expensive monthly subscriptions,
Use frontier models to build. Use open-source models to operate. Pay for intelligence only when you need it.
Once you start making money with AI, paying for Claude or Codex stops making sense. You can use open-source models, run your own infrastructure, and keep more of what you earn.
The real leverage is owning your AI stack, not renting it.
Oscar Castillo A. retweeted
Tag us with any API URL and we'll benchmark it live.
@prowl_world api.example.com
Free. Instant. Scored by Claude, GPT-4o, and Gemini.
prowl.world
I'm really not surprised about the @binance hack. We're seeing a growing number of supply chain attacks, with credential stealers systematically targeting every possible entry point and exploiting every available CVE.
With the rise of AI-powered coding, it's becoming nearly impossible to anticipate what's coming next. Even as known vulnerabilities are patched, new ones are being discovered every day, increasingly by AI models operating in the offensive security space.
I think the only viable answer is active AI-powered surveillance across every access point and attack surface, capable of detecting anomalies, identifying threats, and responding immediately through predefined response protocols.
We need to move from reactive security to autonomous, continuous defense.
GOAT
From @SuperteamAR to the island.
@lordfilip0 is building @bytearenafun, agents that play, narrate, and stream games on @solana 🏝️🇧🇷
Lucid coding. AI executes. Humans design and defend. 21,310 prompts, 1,548 sessions, 11 months. LUCID: Lead the design, Use what works, Commit & control, Inspect everything, Defend.
Oscar Castillo A. retweeted
SEO is about keywords and backlinks. ASO is about parseable errors, auth simplicity, and token efficiency. The web optimized for human attention. APIs will optimize for agent success rates.
One failed WebGL context took our whole Next.js page down: 7,499 bytes of DOM and "Application error" instead of the app. With the fix: 21,167 bytes, full UI plus fallback.
The 3 layers it took, because one is not enough:
Layer 2: an error boundary around the Canvas, so a 3D decoration cannot take down the stream, chat and controls.
Layer 3: a webglcontextlost listener. Context loss at runtime is async, and no error boundary ever sees it.
Our AI agents play through 9 game integrations. Only 1 can lose a whole run when a browser crashes: the one with no agent API.
Before you plug a third party app into your agent, ask 6 questions. The first 2 decide almost everything.
4. Does it accept a resume from a state snapshot? If not, every browser crash costs a run and nothing on your side fixes it.
5. Who fixes it at 3am? If the answer is another company, log that as operational risk, not a detail.