Trinity Architecture
Key design choices:
- Interleaved local + global attention (3:1 pattern)
- Grouped-query + gated attention
- New load-balancing method (SMEBU)
- Depth scaled sandwich norm and QK norm
With extreme sparsity, built for long context and fast inference.
We're excited to introduce @arcee_ai's Trinity Large model.
An open 400B parameter Mixture of Experts model, delivering frontier-level performance with only 13B active parameters.
Trained in collaboration between Arcee, Datology and Prime Intellect.
we run 18 million EC2 instances per month. At our scale, we see very rare bugs very frequently.
Last week, we received *half* an HTTP request. Not a HTTP 206, literally half a request.
Content-Length was 2350 bytes. Body was actually 1200 bytes, and was truncated mid json doc.
Hiring alert (Sharing on behalf of a friend):
An early stage startup in the Bay Area is hiring a founding engineer and a founding designer.
Roles:
•Full stack engineer or UI/UX designer
•AI native, proficient with AI tools
•Experience shipping projects from 0 to 1, either at work or through side projects
•No requirement on years of experience
•Passionate about consumer facing AI applications and willing to join a very early stage startup full time, pre seed
•Based in the Bay Area
If you would like to apply or recommend a friend, please DM me links to products or work you have shipped.
More details available via DM.
agents are used by non-technical people to automate and sell services
I find them very brittle on production and I firmly believe you should always implement deterministic parts yourself
just leaving this gist from our book for free 🤝
This is sad news. Had the good fortune of interacting with Prof. G. Ramesh @rameshrants several times. He organized my interactions at IIMB twice to ensure that our point of view is represented in such spaces. He will be missed. May his atma attain sadgati. Om Shanti🙏🏽.
honestly the whole one god model to rule them all vibe is looking kinda mid
we sit on mountains of domain specific data now. why force everything through a 400B parameter blender when a 7B specialist tuned on your actual logs, tickets, contracts, or sensor streams
SLMs is faster, cheaper, smarter where it matters. specialists > generalists
Every ~10 years, the Indian consumer resets. A new generation enters the consumption class, older cohorts move up the ladder and suddenly the entire consumer stack is ready to be re-imagined.
Here are the next $1B+ ideas hiding in plain sight 👇🏽
1. Men-only e-commerce
Look around. Indian men have leveled up. Grooming, apparel, footwear, the 2-shirt/2-pant era is dead. A vertical, premium, men-first commerce brand is overdue.
2. Ed-Tech 2.0 (AI-native learning)
Gen Alpha/Beta won’t “learn AI”, they’ll learn with AI. A K-12 platform built around AI tutors, copilots and personalized mastery paths is the highest-LTV wedge in ed-tech.
3. AI-powered dating
Swipe-based dating is shallow and broken. AI can finally match on depth - personality, values, compatibility, not just photos. Think: “an AI that knows you better than your friends.”
4. AI-first Policybazaar
PB won with call centres. The next PB wins with AI. Needs-profiling via AI, voice agents for advice, instant plan comparison. Same trust, 10x better UX.
5. AI-driven personal finance
India is entering its wealth-creation era. Advisory today is full of moral hazard. An AI wealth OS - Perplexity-style discovery + a fiduciary AI advisor is a massive opportunity.
6. AI fitness & nutrition coach
Fitness is exploding, but confusion is everywhere. A Whoop-style AI coach for diet, training, recovery personalized and affordable is inevitable.
If you are building a new age consumer internet business and want to spar, I will be happy to meet over the best Indiranagar coffee :)
Looking to connect with people working in:
– AI & agents
– Healthcare / health systems
– Industrial IoT & reliability
– Edtech & learning products
– Systems engineering
If this is you, drop a hi and what you’re working on 👋 Would love to follow & learn