AI Memory: The Cost of Remembering
November 15, 2026
A 10M-token context can demand over a terabyte of KV cache, even for a standard 8B model. Notes, persistent caches and learned weights offer other ways to retain experience, shifting costs across serving, offline compute and storage. Which approaches pay off as agents work for days, and how much does the answer depend on reuse?
No Retraining, No Recompiling: Unlocking Hidden Headroom
November 08, 2026
An LLM edited 0.016% of another model’s weights, with no training, and that model’s math score rose from 7.9% to 48.1%. Binary optimizers have tuned compiled programs this way for decades: BOLT made GCC’s C++ compiler 15% faster on an LLVM build, with no recompiling. When do such edits pay off, and where do they stop?
Optimizing Agentic Execution: From Prose to Harness
October 27, 2026
Prose is only advice. On Terminal-Bench, moving compiler-inspired optimizations from Claude Code’s prompt into its harness cut cost by 15% with no loss in pass rate. Pruning the tool set also removed the tools for watching and stopping background jobs; restoring them kept the saving, and on the task where agents had polled most, the runs were the fastest yet. Here’s what paid off, why no harness change is local, and why a tool brings a capability, not a habit.
Agents That Never Stop Learning: Beyond the Bitter Lesson
October 23, 2026
Sutton’s Bitter Lesson says search and learning beat human knowledge. Scaled with compute, both have delivered stunning results, yet generalization to new tasks and settings remains uneven. In the multi-agent era, can agents learn from experience to generalize, including from each other?
Optimizing Agentic Execution: Putting the Ideas to the Test
October 11, 2026
Can compiler-inspired ideas make AI agents more efficient? Across ten Terminal-Bench tasks, a combined intervention cut cost by 21% and execution time by 24% among successful runs - but earlier stopping exposed a trade-off. Here’s what worked, what varied, and where the savings came from.
AI Infrastructure → Agent Harness
September 21, 2026
Realizing a model’s capabilities depends on the system around it. What can lessons from AI infrastructure tell us about optimizing agent harnesses - and narrowing task-specific gaps between open-weight and frontier models?
Eval: From Computer Systems to Agents
September 13, 2026
As adoption of open-weight models grows, robust Eval design becomes increasingly important. Agent Eval shares several challenges with Computer Systems Eval, from attributing improvements to validating the measurements themselves. Which lessons carry over, and where do agents introduce new challenges?
Edge AI for Wearables: The Gearbox
August 28, 2026
AI on wearables is subject to tight constraints: answer in under half a second, on about a watt, with a sliver of the memory a phone has. Getting there takes many knobs across models, data, compression and inference, turning together like gears. Which gears matter most, and where are they still stuck?
Optimization of Agentic Execution
July 27, 2026
Optimization of model harnesses has started to gain traction. Interestingly, there are a lot of parallels between problems that have been addressed in the realm of compiler-driven program optimization and optimization of agentic execution. Like what?
VRM: Vision Reasoning Models
June 02, 2026
From scene understanding to visual reasoning, we have come a long way. As wearables becomes ubiquitous, visual reasoning will play a vital role. Why?
Enabling the AI Revolution via Hardware-Software Co-Design
April 25, 2026
The accelerated growth in AI has fueled a multi-way arm race to build data centers - AI infrastructure spend across the major tech players in 2026 is projected to be around $700 billion! Ensuring high scalability, reliability, and efficiency serves as a key competitive advantage and may prove to be critical for survival (the exponential growth in investment is not sustainable indefinitely). What roles does HW-SW Co-Design play to this end?
World Models and How Children Learn
March 22, 2026
As Physical AI gains momentum, developing World Models has received lot of attention. Are there nuggets one can borrow from deacdes long research on how kids learn?