Building AI for the Physical World: Lessons from Accelerating Discovery for Chemists Across the Globe

TL;DR

• Trust is the most important feature — not accuracy in the abstract, and not impressive but isolated capability. Trust means alignment with the user's domain. Once a system starts reasoning in a confidently wrong way, users don't just lose trust in the answer; they lose trust in you, the person building the system, and winning either back is a steep climb.
• Formulation chemistry is the hardest domain for agentic co-scientists to enter: it's the largest domain in industrial chemistry (more than half of a $200B annual R&D spend goes to formulated products), yet the knowledge that determines success is overwhelmingly tacit, unpublished, and proprietary, seldom if ever surfaced publicly.
• Trust breaks quietly: on a real formulation chemist's patent query with two plausible-looking passages, a leading embedding model ranked the correct one beyond position 300, and the similarity-score gap between the two was essentially zero — the model has no within-domain discrimination in formulation chemistry. For an agentic co-scientist that isn't just bad retrieval; it becomes a false belief the system reasons from confidently.
• Only one thing worked: task-specific, ontologically structured contrastive training — teaching the model the shape of the domain rather than feeding it more chemistry text. Commercial APIs score effectively zero on their benchmark; open-weight models do better, but continued pre-training doesn't fix it.

Key Points

Formulation chemistry: the biggest domain and the hardest to enter (~01:20–01:22)

Albert is an AI-native operating system built as a collaborative surface where AI agents and chemists work together to accelerate discovery and bring new materials to market faster. It's the R&D platform for some of the largest chemical companies in the world.

The first lesson from deploying AI at enterprise scale: how skeptical bench chemists actually are. They're inventing things people rely on, they expect a great deal, and they don't easily forgive even simple mistakes — let alone costly complex ones.

Industrial chemistry, as he defines it, is the chemistry of the room around you: the paints and coatings on the walls, the adhesives in your phone, the coatings on your glasses, the personal care products you used this morning. All of these are formulated products, and formulation science is the largest domain in industrial chemistry — more than half of a $200 billion annual R&D spend goes toward them.

And this is precisely the domain that's least accessible to agentic co-scientists, because most of the knowledge determining success today lives in tacit, unpublished, proprietary form and is seldom if ever surfaced publicly.

Within it, Albert's co-scientists help partners invent products that over a billion people depend on every day. That's the weight these chemists carry into every decision, and it's exactly why building AI for scientific discovery taught them that above all else, trust is the most important feature — not accuracy in the abstract, not capability or amazing but often isolated results, but alignment with their domain. He conceded this sounds obvious, and noted how easy it is to lose track of when you're building capabilities without clear problem-selection criteria.

Because when your job is maintaining a materials market segment that reaches a billion people's hands, your expectations are high. And once a user sees a system beginning to reason in a confidently wrong way, they don't just lose trust in the answer — they start losing trust in you, the one developing the agentic system. Winning either back is a very steep climb.

How trust actually breaks: one retrieval example (~01:22–01:26)

His emphasis: trust doesn't break through a dramatic failure of complex reasoning. It breaks through something simple and quiet — usually just basic misalignment with the domain.

The example is a real query from a formulation chemist looking for a patent for the agent to reason with. Two candidate passages, both of which look like they could answer it; one is correct. What happened:

• A leading embedding model ranked the correct answer beyond the 300th position — well outside what typical reranking systems check in production.
• More surprising still, the difference in similarity score between the two passages was essentially zero.

Which means the model has no within-domain discrimination for formulation chemistry. And for an agentic co-scientist, this isn't just a bad retrieval — it becomes a false belief the system then reasons from very confidently.

The fix is to dig beneath the surface and ask what actually makes the passages different. On the surface they share similar material ontologies, similar function, and similar final application domain. But using an ontological signal to surface what a formulation chemist would actually use to tell them apart yields domain alignment — and from there, powerful training signals for within-domain discrimination.

To find out where these within-domain distinctions live in industrial chemistry, they built benchmarks. Comparing every model on the same public corpus:

• Commercial APIs score effectively zero — no within-domain discrimination at all.
• Open-weight models do better, but continued pre-training on them doesn't really work either.
• What does work is task-specific, ontologically structured contrastive training — teaching the model the actual shape of the domain's chemistry rather than showing it more chemistry text, i.e. aligning with how the actual users think. That took them from a near-zero model in production to something meaningful at enterprise scale.

His conclusion: reasoning and discovery under high uncertainty are only as good as the alignment between the system and the actual problem space it's applied to — which is why he expects scalable AI solutions to become far more domain-native. Several talks that day pointed the same way: as you move into more specialized domains, you have to build AI that genuinely understands them.

He closed on a second, longer-term and more important kind of trust: not trust in the answer, but the trust you build when constructing agentic systems for real users. The reason is that the public record of science is survivorship-biased — we systematically miss what's beneath the surface — and a great deal of that missing knowledge lives inside enterprise R&D today. Continually asking "what are we doing to build trust in the AI we're building for end users?" is how you get access to that data.

Quotes

"…building AI for scientific discovery has taught us that above all else, trust is actually the most important feature." (~01:22:14)

Not accuracy, not capability — trust.

"…once you reach a system and you begin to see that it's beginning to reason in a confidently wrong way, you don't just lose trust in the answer. You begin to lose trust in you, the one who's actually developing the agentic system." (~01:22:46)

The cost of confidently wrong output spills onto the builder.

"Trust doesn't break down through some dramatic failure of complex reasoning. It breaks down through something simple and quiet." (~01:23)

It's never an explosion — it's chronic misalignment.