The first clue in Karan Brar's career is an almost embarrassingly idle GPU. At Reducto, where documents are turned into structured data, customer bursts were straining a computer-vision serving stack. The obvious remedy was more capacity: buy room for the spikes and keep the machines waiting. Then Brar and the team measured the whole pipeline. GPU utilization was sitting at 2 to 5 percent. The expensive part was barely perspiring.
The constraint lived around the model: image decoding, preprocessing, scheduling, the unglamorous traffic between one impressive noun and another. By parallelizing preprocessing, changing the serving stack, and tuning how work stayed in flight, the team pushed production utilization into a 40 to 90 percent range. One representative path rose from roughly 17 images per second to peaks near 148.5. Capacity planning became calmer because the system finally used the capacity it already had.
This is a useful way into Brar, because he has since founded a company around a larger version of the same instinct. Before adding intelligence, find the actual bottleneck. Before applauding an AI result, find out whether it repeats. Before asking an agent to conduct research, build somewhere for the research to live.
A company built around the loop
Brar is the co-founder and CEO of hiloop, a two-person San Francisco company in Y Combinator's Summer 2026 batch. Its subject is autoresearch: agents proposing experiments, running them, reading the result, and deciding what deserves another attempt. The pitch is less science-fictional than the phrase “recursive self-improvement” may sound. Give hiloop a hard task, a baseline model or agent, and an evaluation. Its system orchestrates a campaign across data, training, prompts, tools, harnesses, and the machinery underneath, then returns the best improvement that survived verification.
The distinction between an answer and a verified improvement is the company. An answer can arrive dressed for dinner, fluent and confident. An improvement has to beat a baseline, survive repeated trials, and leave behind enough evidence for somebody else to understand what happened. Research needs memory, lineage, compute orchestration, and the statistical manners to admit when a charming result was only noise.
Brar did not arrive here by collecting futuristic nouns. He studied at the University of Toronto with a focus in AI, describes himself publicly as a “math enjoyer,” and led machine-learning infrastructure at DynamoAI. He later worked at Reducto on machine learning, post-training, and platform systems. That is where he met Thomas Boser, now hiloop's co-founder and CTO. The origin story is pleasingly practical: two engineers met close to production systems, then decided to build a better experimental home for the agents increasingly operating those systems.
“Human judgment becomes more valuable because it can steer orders of magnitude more experimentation than before.”Karan Brar, “Search is enough”
The experiment behind the argument
hiloop's first public evidence came from Andrej Karpathy's autoresearch benchmark. The assignment is compact and unforgiving: train a small transformer to the lowest validation loss within a fixed five-minute budget on one GPU. Lower is better. Brar and Boser gave two stock coding agents access to 50 B200 GPUs and the company's infrastructure. The agents ran 4,188 experiments across two days.
There was no custom research oracle waiting behind the curtain. The system used off-the-shelf coding agents, with a more capable model available as an adviser. hiloop supplied reliable execution, shared state, compute, and the ability to compare candidates. The final recipe reached a confirmed 0.9016 validation bits per byte, measured as the median of 25 paired confirmation runs. The best individual draw was lower, at 0.8999, but the company displayed it separately. One lucky run, Brar's write-up observed, proves little.
The tidiest number is 4,188. The revealing one is approximately 1,300: a long stretch of challengers failed to beat the incumbent under interleaved controls. This is what a real search looks like after the demo glow wears off. It contains plateaus, discarded branches, and many clever suggestions that turn out to be beautifully wrong.
For Brar, those failures are not industrial confetti. Recorded with their reasoning and lineage, they become part of the research asset. The final recipe tells you where the search landed. The full trail tells you which paths were exhausted, where evidence changed direction, and what the next research agent should know before spending another afternoon rediscovering an old disappointment.
When experiments become cheap, legibility becomes expensive. A thousand runs without lineage are activity. A thousand runs with baselines, controls, and a searchable record can become institutional memory.
Taste at machine scale
The campaign also offers a quieter picture of human work. Brar estimates that people made perhaps a dozen meaningful decisions over the two days: choosing directions, deciding which result deserved another look, and correcting course. Agents made the other several thousand. The division is neither a worker-replacement parable nor a promise that an algorithm has acquired scientific taste. It is leverage applied to judgment.
A human researcher still chooses the hill worth climbing. The machine can test more routes up it. This arrangement makes the evaluation unusually important. If the metric rewards the wrong behavior, an indefatigable agent becomes an indefatigable specialist in the wrong thing. Brar's public writing is careful to keep humans in the loop precisely where objectives, direction, and interpretation matter.
That care is visible in his earlier engineering work. At Reducto, the lesson was not “use a faster GPU.” It was to measure the forward-only ceiling, profile the surrounding pipeline, and refuse the first convenient story. At hiloop, the same temperament appears in paired confirmations, interleaved controls, and the separation of a best single draw from a repeatable median. The product may concern machines that conduct experiments, but its manners are those of a skeptical operator.
Brar's public persona has some relief around the edges. His social bio calls him a math enjoyer rather than a prophet. He writes directly, with the occasional lowercase aside, and has been gathering practitioners around the question as well as building software. During the Ai4 conference in Las Vegas, he hosted a small dinner for founders, ML leaders, and researchers to discuss systems that form hypotheses, run experiments, learn from failure, and improve. Sushi is still one of civilization's better interfaces for unresolved technical questions.
“We believe the artifacts produced during research are at least as valuable as the final result.”Karan Brar, “Search is enough”
The laboratory as product
hiloop's present product is deliberately infrastructural. An agent can provision compute through an API. A run can be snapshotted, forked, abandoned, or restored. Experiments are traced, model usage and estimated costs can be queried, and the system is designed to run hosted or inside a customer's cloud. Bring a homegrown evaluation script or an off-the-shelf reinforcement-learning loop; the laboratory should care more about a measurable objective than the fashion of the harness carrying it.
The company is starting with agent and model training, continual learning, and optimization. Brar has also argued publicly that more companies will come to resemble model companies. If that happens, the unromantic infrastructure around experimentation becomes strategic. Every organization may want intelligence; fewer will want to build the machinery for coordinating hundreds of GPUs, preserving thousands of attempts, and proving an improvement survived contact with statistics.
The language of hiloop's website makes the operating philosophy unusually concrete: snapshot, fork, restore. An agent can try an idea, branch from a known state, and return when the idea proves useless. Every event carries the identity of its fork. That may sound like housekeeping, but housekeeping is what separates an experiment from an anecdote. Without the exact state, code, and path that produced a number, a result becomes a dinner story with a decimal point. Brar's laboratory is designed so the decimal can be questioned the next morning.
There is still a bet inside all this careful measurement. Brar believes frontier models will keep getting better at the research loop, and that systems which let those models search widely will improve with them. Rather than encoding an elaborate theory of how machine research ought to proceed, hiloop tries to supply a broad, reliable surface on which better models can work. The company wants to sell the laboratory while the scientists are still arriving.
It is an audaciously plain proposition. Ideas are becoming abundant. Experiments are becoming executable by machines. The scarce things move downstream: good objectives, reliable environments, patient verification, and a record that prevents a company from forgetting what it learned last Tuesday. Brar is building for that scarcity. Somewhere in the system, another expensive GPU is finally earning its keep.