Ask an AI a hard question today and it will think for a few minutes. Nobody ships one that thinks for ten days.

That gap is where this company sits, and it is larger than it looks.

What they built

The company describes itself as autoresearch as a service, for teams training agents to do jobs where a general model is not good enough on its own.

A campaign sweeps across everything that might matter: the data, various kinds of post-training, continual learning, prompts, tools, the harness and the surrounding system. What comes back is the best verified improvement, and the word verified is doing real work there. It runs either hosted by hiloop or inside your own cloud.

Around the models sits the part that is arguably the actual product: persistent memory, a full record of every experiment and where it came from, orchestration of the compute, and statistical checking of whether a result is real.

The founders are Karan Brar, chief executive, previously in machine learning at Reducto and head of ML infrastructure at DynamoAI, and Thomas Boser, chief technology officer, previously an engineer at Reducto, a founding engineer at Crosswise, a senior ML engineer at Discord, and before that at Sentropy and SoFi. Early access runs through founders@hiloop.ai.

Helping people improve models is a busy category, approached from several directions. Weights & Biases and Neptune built the experiment tracking most teams use. Determined AI and SigOpt automated hyperparameter search. Databricks and Hugging Face own much of the tooling underneath. OpenPipe and Predibase made fine-tuning a product. Scale AI and Surge AI supply the data. Braintrust, LangSmith and Langfuse handle evaluation. And AutoML as a research programme has been trying to automate model building for over a decade.

Almost all of it gives you better tools for doing the search yourself. hiloop is offering to do the search and hand you the answer.

The point: stop aiming at the answer and build something that never finishes

The idea of leaving an AI to work on one topic for a long time is not new. Deep research tools from the large labs already go away, read widely and come back with something considered, and reasoning models already spend real time thinking before they answer.

But notice the ceiling. All of them run for minutes. None of them runs for ten days to reach an answer for you, and none of them runs indefinitely, quietly getting better at your particular problem while you do something else.

This breed of startup changes the goal. The target stops being the point. You assume instead that there is no perfect result, and you build something determined to keep running.

That single decision changes the design of everything. If the machine never finishes, it has to be trivial for a company or a person to set one going, which is exactly why a hosted option matters more than it sounds. It has to keep a record, because a result that arrives on day forty is worthless if nobody can see what produced it. And it has to prove each improvement, because without that the thing is just burning money in the background.

It also changes what the founder is aiming at. The job is no longer to reach the goal. It is to never stop, and to make sure that not stopping keeps being worth it.

What makes forever worth paying for

There is an obvious objection and it deserves a straight answer, because running forever costs money forever.

Search problems have diminishing returns. If the target sat still, a machine that never stops would spend its second year producing almost nothing and the whole idea would collapse into an expensive habit.

The target does not sit still. New models ship every few weeks, your data drifts, your traffic changes, your costs move, and what was the best approach in March is beaten by something in June. A machine that stopped in March is now out of date, and one that never stopped has already found the new answer. That is the honest reason forever is worth paying for, and any founder building one of these should be able to say what keeps moving in their field.

It is also a far better business than the alternative. A project that finishes gets invoiced once. A machine nobody switches off is paid every month, and the longer it runs the more its record is worth, because a rival starting today has to begin with an empty one.

Where else a forever machine belongs

This next part is opinion rather than anything the company has said. hiloop today runs campaigns that end and return the best verified improvement, which is the sensible way to start. The wider idea is where it leads.

Forever machine is a good name for the category, so it is worth using. The exercise is to ask which fields deserve one, and the test is simple: the problem has no perfect answer, the conditions keep changing, and a slightly better result is worth real money every single day.

On that test the candidates are everywhere. Pricing that never stops testing. Advertising creative that keeps regenerating itself against this week's audience. Warehouse layouts and delivery routes that keep rearranging as orders shift. Trading strategies, obviously. Energy grid dispatch. Drug candidate screening as new structures are published. Fraud rules against attackers who change tactics weekly. Crop planting plans against a moving climate. Chip floorplans. Recommendation systems. Every one has no final answer and a target that keeps moving.

Finding a reason never to stop, and making sure the thing genuinely keeps improving, is an entire field of AI ideas that nobody has worked through yet. There is room for a great many of these companies, and you could start one.

Why it is bigger than it sounds

The short description is automated experimentation. The change underneath is what the machine is for.

  • Every long-running AI today still stops. Deep research and reasoning models run for minutes, and nothing runs for ten days, let alone permanently.
  • Assume no perfect answer and the design changes. It has to be trivial to start, it has to keep a record, and it has to prove each gain, because none of those are optional once nobody is watching.
  • A machine nobody switches off is a much better business than a project. It is paid monthly instead of once, and its record compounds against anyone starting later.

What to watch

The measure is the length of the longest campaign anyone has left running. Days would be notable. Months would mean the category is real.

The second is whether customers renew without being asked, because a machine people forget to switch off is the entire business model working as intended.

The third is what happens when the measurement is wrong. A campaign optimises whatever it is told to, so a customer with a bad metric gets a confidently wrong answer, arrived at with enormous thoroughness.