// sunil somarajan
@sunilsomarajanAccelerating intelligence at the edge
The most interesting problems live at the boundary of hardware and software. where firmware meets silicon, where code meets bare metal, and where elegant abstractions meet unforgiving physics.
Writing
The Overnight Researcher
What happens when you give an AI agent a single file, a clear metric, and permission to never stop? Karpathy's autoresearch reveals a pattern that extends far beyond machine learning.
Writing
The 1000x Problem
Thirty researchers just published a roadmap for making AI a thousand times more efficient by 2035. The interesting part is not the number. It is what has to break to get there.
Writing
The Hidden Cost of Zero-Cost Abstractions
Your code is correct. Your hardware is fast. The gap between them is where all the performance lives. We ran the experiment on real RISC-V silicon to prove it.
Writing
Teaching Small Models to Punch Above Their Weight
In 2015, Hinton, Vinyals, and Dean figured out how to make a small neural network learn what a large one knows. A decade later, this trick is the reason your phone can run AI at all.
Notes
Claude 1M Context Window Goes GA
One million tokens of context at standard pricing. No premium tier, no special request. The practical difference is not the number. It is the absence of the wall. At 200K, you hit compaction mid-feature and lose the thread. The conversation forgets what it was building. You spend ten minutes re-explaining what you just spent an hour constructing. At 1M, each feature lives in a single, self-contained session. The context stays clean. The work stays coherent. You stop fighting the tool and start using it. That is the actual product improvement. Not the headline number. The fact that you never have to context-switch because the model did.
Notes
NVIDIA NemoClaw at GTC 2026
NVIDIA announced NemoClaw at GTC. A software stack that wraps OpenClaw with Nemotron models, OpenShell secure runtime, and guardrails. One command install. After OpenAI acquired OpenClaw in February, enterprises needed an independent, secure agent platform. NVIDIA filled the gap. Runs on everything from RTX PCs to DGX Spark.
Notes
Knuth: Claude's Cycles
Knuth learned that an open problem he had been working on for weeks was solved by Claude Opus 4.6 in about an hour. Filip Stappers posed the problem and coached Claude through 31 explorations with one rule: document progress in plan.md after every run. The agent reformulated, hit dead ends, backtracked, reframed, and converged on a construction for all odd m. Same structured-exploration loop as Karpathy’s autoresearch. Different domain, same meta-pattern.
The postscripts are wild. GPT-5.4 Pro produced a 14-page proof for even m. Multi-agent workflows (GPT + Claude) found simpler constructions. The Lean community formalized the proof within days.
Notes
Latency Numbers Every Programmer Should Know
Jeff Dean’s classic reference on computer operation latencies. L1 cache at 0.5ns, main memory at 100ns, SSD read at 16us, disk seek at 10ms, cross-continent roundtrip at 150ms. Knowing these magnitudes is the difference between guessing at bottlenecks and actually understanding where your time goes.