Raza Habib and the advantage of changing your mind

Raza Habib gave himself two weeks to discover whether his company should become a different company. The test was refreshingly ungainly: make mockups, take them to people already building with large language models, and ask whether they would pay. Ten paying customers would justify changing direction. By his account, the test reached that target in two days. The calendar had barely had a chance to object.

The danger was an odd one for an AI founder. The technology might become too good. Humanloop had started in 2020 by helping people label the data needed to train language models. If general-purpose models improved enough, customers could need much less of that work. Progress in his own field threatened the product he was selling. Habib and his colleagues began exploring a pivot months before ChatGPT made conversational AI a household subject.

That episode supplies a useful entrance to Habib's career. A researcher trained to reason about uncertainty had to make a consequential decision without waiting for certainty. He chose a small commercial experiment. The result changed Humanloop's work from improving training data to evaluating applications built on powerful existing models. It also sharpened a question he would keep returning to: how do people know whether an AI system is doing a useful job?

A physicist finds another frontier

Habib studied physics at Cambridge before taking a master's in machine learning at University College London. He credits studying under Professor David MacKay with inspiring his interest in AI. Physics supplied a way of thinking; machine learning offered a place to apply it. In a 2022 account of Humanloop's beginnings, he compared the attraction of contemporary AI with the attraction quantum mechanics might have held for a physicist a century earlier.

His early professional work crossed several kinds of engineering. He was the founding engineer at Monolith AI, applying machine learning to mechanical engineering, and built speech systems at Google. The distinction matters. His eventual company grew from experience with machines that had to perform particular tasks, rather than from conversational AI as a novelty. Mechanical engineering and speech are rather demanding audiences for an attractive demonstration.

His UCL doctorate, completed in 2022, was titled Advances in Probabilistic Deep Learning. It addressed approximate inference and speech synthesis. One part developed methods for sampling difficult probability distributions; another used neural variational inference to control aspects of synthetic speech, including emotional expression, with less supervised data. The underlying concern was how to extract useful results when the evidence is incomplete and exact calculation is impractical.

There is a quieter record of this period on his personal blog. Posts from 2017 and 2018 concern sampling, automatic differentiation and mathematical intuition. The titles sometimes carry a raised eyebrow: How bad can importance sampling be? is hardly an invitation to relax. They show a researcher trying to explain difficult ideas, including where a technique can disappoint the person who has trusted it.

The notebook had room for business

In January 2018, Habib attended a Reid Hoffman talk organised by Entrepreneur First and wrote up the advice that interested him. Alongside perseverance and hiring, he singled out business models, networks and short decision cycles. His notes were already moving between the laboratory and the company: observe what happens, reconsider the situation, decide, act, and shorten the distance between those steps.

He also recorded a test for whether experience was producing learning: ask what advice you would give your recent younger self. If there is nothing to say, something has stalled. This was advice he was reporting, rather than a philosophy he claimed to have invented. Still, choosing to preserve it is revealing. The student reading about inference was also thinking about how an organisation changes its behaviour.

That same month, he wrote about serving as a UCL teaching assistant on a DeepMind-led course in advanced deep learning and reinforcement learning. He was interested both in defining intelligence and in organising researchers to make progress together. He welcomed industry's participation in research. His public writing already treated scientific ambition and practical coordination as neighbouring problems.

“I find both of these questions fascinating, especially the question of how to manage scientific progress.”Raza Habib · January 2018

Teaching the machine to ask for help

Humanloop emerged from UCL's AI Centre in 2020. Habib built it with Peter Hayes and Jordan Burgess, with professors David Barber and Emine Yilmaz part of the founding academic team. UCL Business helped the spinout through its Portico Ventures scheme. The company joined Y Combinator's Summer 2020 batch, bringing a research project into a setting where explaining the product to potential customers became part of the work.

The early promise was to let a knowledgeable person teach software much as they might instruct a colleague. Habib's 2021 explanation of active learning used a spam filter. Rather than label a vast random pile of messages, a person could supply a small initial set of examples. The model would then select messages it was uncertain about and ask for labels. Human attention would go where the answers might teach it something.

This was a practical expression of his academic interests. Uncertainty could help determine the next action. The tool needed annotation interfaces, model updates and feedback to work together, otherwise an elegant method remained an awkward engineering chore. Humanloop tried to join those pieces. In July 2022, Index Ventures described its Programmatic product and reported a $2.6 million seed round led by the firm.

The appeal was easy to understand: people with expertise could contribute without becoming machine learning researchers. The means would change as language models improved, but that ambition survived. The person who understood a task still needed a useful way to teach the computer what success meant.

When the experts get the keyboard

After the pivot, Humanloop became a workspace for prompts, evaluations and collaboration. Gusto, Vanta and Duolingo were among its customers. Habib's recurring concern was the gap between an impressive answer and a dependable application. A single pleasing example can earn applause; repeated checks tell a team whether its changes are helping. It is the difference between admiring a cake and discovering whether the recipe works on Tuesday.

In a February 2025 essay, he described a conversation with an engineering leader at a publicly traded technology company. He expected familiar engineering answers about how prompts were managed. Instead, she told him that engineers were not allowed to edit them. Product managers and domain experts wrote the prompts in a dedicated interface, with the results then committed to the codebase. The division of labour had surprised him.

His argument was that clearer access to models changes who can implement a product's behaviour. A specialist's judgment need not always travel through an engineer before reaching the system. He cited language specialists at Duolingo and security experts at Vanta as examples. He also argued that stronger models would keep needing clear goals and context. His thought experiment was an API to Einstein: even a very capable respondent needs to know what you want.

A microphone for unfinished answers

Habib launched High Agency in June 2024 as a podcast for AI builders. Its early guests included leaders from Ironclad, Zapier, Sourcegraph and Hex. The conversations followed the passage from an initial idea to a working product, with room for the decisions that disappear from a launch announcement. Hosting made his curiosity visible. He asked people to describe the work rather than merely announce that the work was exciting.

In his conversation with Jason Liu, creator of Instructor, he pressed on a useful disagreement. He agreed that AI products should serve users, but questioned whether existing software practices were sufficient for building them. The discussion moved through probabilistic behaviour, feedback and the participation of nontechnical specialists. Habib was prepared to challenge the premise while also trying to understand why an experienced practitioner held it.

With Ironclad co-founder Cai GoGwilt, he discussed recognising an AI in conversation. Habib said he had not experienced a moment of believing an LLM was a person. He described a blunt test: ask for multiplication of two very large numbers. In that 2024 conversation, it was a way to probe a familiar weakness. For someone building tools around language models, fluent conversation was an invitation to test further.

A discussion with Hamel Husain and Shawn Wang ranged into literate programming and Donald Knuth. Habib slipped in a joke about having to suffer through LaTeX. The joke was small, recognisable to anyone who has negotiated with a typesetting system, and entirely at home among technical questions. High Agency let the researcher, founder and amused colleague occupy the same conversation.

London, San Francisco, then Anthropic

Building the company also meant choosing where to spend time. Toward the end of 2023, Humanloop decided to establish a San Francisco office while maintaining its London engineering presence. Habib and Burgess were based in San Francisco, Hayes in the UK. The arrangement connected the company to two AI communities and required early starts on the American side to catch the end of London's working day.

Habib was candid about the decision. In an Index Ventures case study, he treated San Francisco as a bet whose payoff he could not yet establish. He pointed to social connections, chance encounters and the concentration of companies nearby. That is a striking admission from a founder explaining a strategic move: the reasoning was plausible, the expense real, and the outcome still open. The researcher had not vanished inside the chief executive.

In August 2025, Habib, Hayes and Burgess joined Anthropic alongside colleagues from Humanloop. The team's announcement presented the move as a way to extend its work on safe AI adoption. The standalone platform was scheduled to shut down on September 8, 2025. The change closed a product chapter while carrying the people and their experience into a model developer.

Habib's career now contains several revisions: physics to machine learning, research to a startup, annotation to evaluation, and an independent company to Anthropic. What connects them is a willingness to examine the job that actually needs doing. Better models altered Humanloop's business; they also opened its next problem. His two-day experiment remains an unusually compact expression of that habit. Ask a question concrete enough that the world can answer, then pay attention when it does.