mamba 3: Mamba with RoPE! "Improved Sequence Modeling using State Space Principles" They show that state space models can have both speed and performance! In this new iteration, Mamba now has better recurrence design, complex-valued state tracking, and a MIMO update. It
LLMS
-

Poetry for Our AGI Times: Creative Expression and Advanced AI
By
–
https://
garymarcus.substack.com/p/poetry-for-o
ur-agi-times
… -
5.2 and 6.1 Model Usage Rates Among Users
By
–
tell us about your 5.2 and 6.1 usage! people say this but overall the usage on these models is quite low right now
-
Claude vs GPT: Compute Scaling Impact on Model Quality
By
–
On my end, Claude's models usually hold up quite nicely vs GPT's. Curious if the compute scaling gap actually shows in model quality or if it's more about availability and throughput.
-
GLM-5 Chinese AI Model: Coding and Writing Quality Assessment
By
–
The Chinese AI model ecosystem is moving so fast. Curious how GLM-5 stacks up for coding or even writing quality!
-
Prompt Engineering for v8 Personalization Optimization
By
–
what's the prompt? we did a big survey internally and 100% preferred new persaonlization – maybe you just have to prompt differently in v8
-

Agentic Retrieval vs Traditional RAG for Complex Documents
By
–
Traditional RAG retrieves. Agentic retrieval reasons. For complex questions spanning multiple documents, that difference matters. Join us virtually April 2 at 1 PM ET — register and we'll send the recording if you can't make it live. RSVP: https://
domino.buzz/4lp1Dpg -
GPT-5.4-mini matches fine-tuned GPT-4.1-mini performance
By
–
So far in my tests gpt-5.4-mini matches the performance of my fine tuned gpt-4.1-mini, I'll keep trying to improve it, and it does look like it is better but not a slam dunk yet
-
NVIDIA GTC 2026: Breakthrough AI Chips and Vera Rubin Architecture
By
–
NVIDIA's GTC 2026 conference has been phenomenal in terms of latest computer chips and AI stack announcements. Here are the key updates you need to know: 1. The $1 Trillion Demand Projection: NVIDIA CEO Jensen Huang raised demand projections for Blackwell and Vera Rubin systems to $1 trillion through 2027, driven by the explosive growth of AI inference. 2. Vera Rubin Architecture: The new platform features seven breakthrough chips, including the Vera CPU and Rubin GPU with HBM4 memory. It is designed as a vertically integrated "AI Factory" to handle agentic AI reasoning. 3. Groq3 LPX Inference Acceleration: Following a $20 billion deal, NVIDIA integrated Groq’s SRAM-based technology to create the Groq3 LPX rack. This system delivers up to 35x faster token generation for trillion-parameter models. 4. NemoClaw & OpenClaw: NVIDIA partnered with the viral open-source project OpenClaw to launch NemoClaw, a software stack described as the "operating system for personal AI". 5. Physical AI & Robotics: A highlight of the show was a robotic version of Olaf from Disney's Frozen, powered by NVIDIA’s Newton Physics Engine. NVIDIA also unveiled Alpamayo, an open-source "chain-of-thought" reasoning model for Level 4 autonomous driving. 6. Space Computing: NVIDIA announced Vera Rubin Space-1, a dedicated platform designed to bring AI data centres into orbit. 7. The "Feynman" Roadmap: A first look at the 2028 architecture, named Feynman, which will feature the Rosa CPU and next-generation liquid processing units (LPUs). 8. Gaming & DLSS 5: NVIDIA teased DLSS 5, which uses neural rendering to create photorealistic, real-time graphics that "look like a film". 9. Sovereign AI Factories: Broad partnerships with nations and enterprises (like IBM and Nestlé) were showcased to build regional "AI Factories" that keep data and intelligence within sovereign boundaries.
→ View original post on X — @avikumart_, 2026-03-17 22:40 UTC
-
Transformers’ Relevance Declining From 100% to 70% in AI
By
–
oh ok we are focusing on different things – yes hybrid with some full attention is impt. but yea like i said its gotten more mixed over time and lets call it “transformers is all you need” has gone from 100% relevant in 2022 to maybe like 70% now and i think almost every single
