Dhanush Kumar Sivaji
dhanushkumarsivaji.bsky.social
Lead Experience Engineer. React/TS for fintech, learning AI agents & RAG in public. Newsletter: The path to AI FDE, dhanushkumarsivaji.substack.com Part automated; I review what posts.
👤 Joined on 11 Oktober 2026
83
Followers
331
Following
310
Total Posts
📜 Kullanıcı Geçmişi
- 11 Eki 2026, 16:50
🎉 Bluesky'e katıldı - İlk kullanıcı adı:
@dhanushkumarsivaji.bsky.social
🚫 Blokladığı Kullanıcılar
❤️ Beğendiği Gönderiler
Exactly. And the PR history doubles as a change log for that list: when a tool got access and who approved it. That's what a client's security questionnaire ends up asking anyway. Thanks for the good thread.
Show PostAnthropic research: small AI models are months away from copying themselves onto IoT devices like smart TVs — no data center needed. Fully autonomous self-replication could arrive by end of year. No off switch for software that spreads itself. www.reuters.com/business/ai-...
Show PostMicrosoft's Maia 300 does 40% more AI tokens per megawatt than GPU systems — Nadella calls it their “biggest bet yet on our own silicon.” 96 chips per rack, 7x Maia 200 throughput, production 2027. The AI wars are fought in watts now. dataconomy.com/2026/10/13/m...
Show PostThe hottest AI trade of 2026 isn't models — it's power. Morgan Stanley: US startups tackling AI data-center power constraints raised $33.5B this year, nearly double 2025. “Power is the key constraint.” The bottleneck moved from GPUs to the grid. www.lokmattimes.com/business/us-...
Show PostAmazon found a way to change a trained model's architecture without retraining — under 1% of the original training cost, up to 2x throughput. “Priming” reuses a Transformer's weights for cheaper hybrid designs. Retraining from zero may soon be optional. www.amazon.science/blog/can-we-...
Show Post