Sam D'Amico

Sam D'Amico is the founder of Impulse and, with JF Bastien, coauthor of research on making video vision-language models reuse visual information rather than repeatedly process it. Their May 5, 2026 paper, VLMaxxing through FrameMogging, investigates when cached state preserves answers and when a model needs fresh evidence.[2][3]

D'Amico and Bastien study training-free anti-recomputation: reducing redundant work without changing model weights. Their largest reported gains come from later questions about a video the model has already ingested. Simply retaining a cache can produce systematic answer drift or malformed outputs. Their adaptive repair policy refreshes the newest visual frame at the first follow-up, then lets the second follow-up inherit that repaired cache without another visual refresh. A failed variant instead returned to the original unrepaired state.[3]

On frozen Qwen2.5-VL-7B-Instruct-4bit, the authors report 14.90–35.92× faster follow-up queries, comparing median cold-follow-up latency with median repaired-follow-up latency. Across 31 VideoMME sessions containing 93 paired queries, including 62 repaired follow-ups, they observed no changes in answer choices or correctness relative to matched cold runs. The first query still required full processing; preserving those baseline answers within a finite sample does not establish universal correctness.[3]

Their work separates that follow-up gain from speeding up the first question about a fresh video. For Gemma 4-E4B-4bit, a 32-frame short-video test reported 1.316× first-query speedup and 42.2% less vision-processing time, with no paired answer drift or parse failures on 20 items. The authors emphasize that a component improvement affects total latency only in proportion to the time that component consumes. First-query pruning and later-query cache reuse therefore cannot be combined by multiplying their headline gains.[3]

D'Amico and Bastien also expose the limits of faster reuse. In a stress test where follow-up prompts included the previous canonical dense-model answer, adaptive repair and scheduled refresh each changed choices and correctness on 6 of 133 follow-ups—4.51%, exceeding their 3% gate. A more conservative policy that refreshed the newest frame on every follow-up showed no such drift. This was a prompt-variation stress test rather than natural dialogue, and native-rate streaming remained a deployment target. The accompanying VLMaxxing repository contains research code, artifacts and manuscript tooling.[3][4]

Before this 2026 research, D'Amico worked in consumer hardware and founded Impulse. His account of that earlier period explains the appliance-performance and control decisions behind the company.[5]

In a March 21, 2024 interview, D'Amico recalled working at Facebook on VR headsets and controllers. During a trip to Japan for a haptics conference, he watched a wood-fired pizza cook in roughly 30–45 seconds and began considering how to reproduce that performance electrically at home. The tabletop-oven idea broadened into battery-assisted installed appliances; he dated Impulse's start to summer 2021.[5]

D'Amico chose core cooking performance as the entry point. He argued that displays, connectivity and guided recipes had often failed to make an oven better at its underlying job, and that persuading people to replace gas meant having to “dethrone the gas stove.” He explained that running an appliance internally from battery DC voltage separates outlet charging power from peak cooking power. According to his account, pushing burner power higher then exposed a sensing and control problem: the team needed better temperature sensing to avoid overheating pots, and that work also enabled more precise cooking control.[5]

His hardware strategy favored using the mature consumer-electronics supply chain rather than automatically adopting aerospace-style vertical integration. He argued that developing with manufacturers could bring engineering capacity and established component suppliers into the design process, easing the move to higher volumes. Asked about adding a robotic arm to the stove, he prioritized software for aggregated battery storage. Networking appliance batteries into a distributed storage system was his stated vision; he explicitly distinguished it from whole-home outage backup, which was not a launch capability.[5]