AI Chip Performance: Why Blackwell Delivers 50x, Not 2x. AI hardware delivers 50x more tokens per watt while spec sheets show 2x. The 25x gap lives in extreme co-design across seven layers, not in any single chip. ai-hardware ai-markets ai-economics If you have followed Artificial Intelligence (AI) hardware announcements and wondered why the generational gaps …
Why Blackwell Delivers 50x, Not 2x · Hanh D. Brown Read More »
When RAG retrieval disappoints, the advice AI engineers hear today is almost always “add a reranker”. Ask why a reranker works, and the answer usually stays at the architecture level: it is a cross-encoder, it applies attention over the query and the passage together, it is fine-tuned on relevance labels. All of that is true, …
AI Agent Governance: Treat the Agent Like an Employee. An AI agent without identity, role, access, audit, performance, and a manager is not autonomous. It is unsupervised. The fix is six primitives and a kill switch. ai-governance ai-and-work ai-strategy An enterprise would never let a human employee through the door without an identification badge, an …
Treat the Agent Like an Employee · Hanh D. Brown Read More »
Introduction Most developers see Python dataclasses as a shortcut for avoiding repetitive dunder methods like __init__ and __repr__. So at first look, it seems like a simple way to write less code and move faster. In practice, dataclasses are designed to reduce boilerplate in data-focused classes while keeping behavior clear and under your control. Instead …
The Work AI Cannot Replace: My Mother and Higher-Value Care. AI abundance forecasts assume labor scales. Care work does not. A daughter’s hours are the hours the forecast does not see. The wall is closer than they say. caregiving aging-and-care ai-economics ai-careers Artificial Intelligence (AI) abundance forecasts share a sentence. The sentence promises automation will …
addresses the fundamental challenge of slow, sequential token generation in language model inference. DFlash speculative decoding support for CPUs was recently enabled in vLLM v0.25.0. In our testing with Qwen3.5-9B on an r8i AWS instance, powered by Intel® Xeon® 6 processors with Performance-cores, DFlash increased average token generation throughput to 3.92x that of the autoregressive …
Speculative Decoding on CPUs: Nearly 4x Faster Token Generation with DFlash Read More »
Venture Capital and AI: Sized for Fifteen Winners a Year. Old venture capital was a basketball team sized for fifteen winners a year. The redesign that scaled: share economics, centralize control, and grow sideways. ai-markets ai-strategy ai-economics For thirty years, every venture firm in the country was sized for the same number. About fifteen technology …
Sized for Fifteen Winners a Year · Hanh D. Brown Read More »
Muse Glimmer is gaining attention in the local AI community and is being compared with Qwen’s 27B-class models. In many cases, it is performing better, especially for local coding and agentic workflows. Meta looks strong in the open-model space, and with a few more iterations, models like this could start competing closely with proprietary systems. …
Run Muse Glimmer for Local Vibe Coding with llama.cpp, DFlash, and Pi Read More »
Open Weights AI: Distribution Decides Whose Values Win. Open Artificial Intelligence models win distribution because they ship cheaper and modifiable. Distribution decides the global default. Values ride along with the winner. ai-policy ai-markets ai-safety Powerful countries are not picking the smartest Artificial Intelligence (AI). They are picking the cheapest one to deploy and the easiest …
Distribution Decides Whose Values Win · Hanh D. Brown Read More »
This post was co-authored with Max Silfverberg (Data Scientist, AI Solutions Lead), Antti Hallavo (Lead AI Software Engineer), and Pontus Huotari (Lead Data Scientist). We work at Alma Media, a Finnish digital services, marketplaces and media company. One of our focus areas is developing AI/ML solutions for real estate listing services, where understanding image content plays an important role. services handle hundreds of thousands of listings a year. Most of those come with …
Why We Fine-Tuned SigLip (And Why That’s Not Always the Right Call) Read More »
