Subscribe
Sign in
Home
Podcast
Archive
About
Latest
Top
Discussions
How Looped Transformers Reuse Their Weights
Follow changing attention states, shared gradients, and the real cost of extra passes.
12 hrs ago
•
Damien Benveniste
8
How the KV Cache Shapes AI Agent Costs
Why shrinking the cache, reusing a prefix, and keeping it nearby save different resources.
Oct 19
•
Damien Benveniste
5
1
How Hybrid LLMs Combine Memory, Attention, and Experts
Follow Mamba-2, Gated DeltaNet and KDA from state updates to the costs that active parameter counts miss.
Oct 16
•
Damien Benveniste
8
1
How Persistent Agents Keep Working
Reconstruct a coding task’s context, preserve accepted work, and recover safely when a worker disappears.
Oct 12
•
Damien Benveniste
8
1
How Self-Improving Agents Learn from Experience
How executable tasks and adaptive curricula turn production failures into useful practice.
Oct 9
•
Damien Benveniste
6
How DeepSeek V4.1 Flash Separates Reading and Generation
How a shorter input path, shared attention memory, and selective reconstruction change the cost of long conversations.
Oct 7
•
Damien Benveniste
4
1
Build a Jev-Supervised AI Agent
Typed decisions, guarded tools, and a measured comparison with LLM supervisors
Oct 3
•
Damien Benveniste
8
2
September 2026
How Jev Turns Questions Into Decisions
Inside the state, typed-output, probability, and application-policy contracts of a decision model
Sep 30
•
Damien Benveniste
8
The Complete Guide to Evaluation Pipelines for AI Agents
From production traces and isolated trials to evidence you can use for release decisions
Sep 28
•
Damien Benveniste
13
How AI Agent Memory Updates a Changed Preference
Follow one correction from a user message to the active fact a future task retrieves.
Sep 25
•
Damien Benveniste
6
How LLM Routers Balance Cache Reuse and Queue Time
Four routing strategies, one repeated prompt, and the point where a cold replica wins
Sep 23
•
Damien Benveniste
10
How Test-Time Scaling Allocates LLM Reasoning Across Depth, Width, and Feedback
Why the same inference budget can finish one path, explore alternatives, or learn from a failure
Sep 21
•
Damien Benveniste
7
This site requires JavaScript to run correctly. Please
turn on JavaScript
or unblock scripts
