AI FOR SCIENCE · INTERVIEWS
Kilpatrick says coding already feels like a narrow superintelligence and the next ones will come where results are easy to verify: mathematics, finance and possibly science; and that Omni is a single model called a 'world model' for how much it understands about the world, not because it is one in the older sense.
The record has neither Kilpatrick nor a Google statement on which domain comes next; this gives a stated criterion, that results can be verified, and puts science among the candidates without saying it has arrived.
He leads developer products, not DeepMind's scientific research; he says mathematical proofs are 'over my head' and that he is not an architecture expert. No scientific result is given and no scientific use of Omni.
He starts from coding now feeling close to a narrow superintelligence, so what he is thinking about is not when general intelligence arrives but verticals arriving one by one, which he calls 'jagged' superintelligence, ahead of general intelligence. Asked which domain is next, he first says he spends too much time thinking about coding, then gives the criterion: domains where results are easier to verify move faster, such as mathematics and finance, and science 'could be a really interesting one', with some verifiable domains really taking off. He also wants these positive results to come as early as possible so people see what the technology can do for good. On proofs he says he is not a mathematician and it is 'somewhat over my head'.
“things that have better verifiability obviously are the ones where you'll see the gains happen more quickly.”Logan KilpatrickHe says the definition of 'world model' is blurring: Demis called Omni a world model because of how much it understands about the world, and structurally it differs from earlier world models, though he says he is not an architecture expert. The host describes a traditional world model as an action-conditioned video model, and he agrees that what is meant now is a model with some understanding of the world, not necessarily that kind of video model. Omni is a single model and not requests routed to eight different models; text, audio, music, image and video each used to have a model of their own. What works now is video editing; the other outputs are 'functional' but the quality is not state of the art, so they are not released; it is not real time; and this is the first version, Omni Flash. He says traditional online world models are very expensive to run and hard to scale, and Omni's approach may be more scalable and can do many of the same things, though not in real time. His example is someone in the audience editing a photo of the stage with Omni to add a dog, with the other people in the edited picture reacting to it naturally.
He ties DeepMind's culture to its leaders: Demis is a Nobel Prize scientist and an 'OG' of much of this, and that scientific approach is felt inside the company; the mission was started 'literally to solve disease and all these things'. He tells himself not to get lost in a race over who pushes a number higher on SWE-bench, because the point is to solve problems humans actually have. He also says it is not zero-sum. This is his view of the company's culture; he names no specific scientific project.
The site has no Kilpatrick entry. On DeepMind pointing agents at science, the timeline has the automated materials lab (December 2025) and the Antigravity antimicrobial-peptide walkthrough (15 October 2026).
This is a DeepMind product lead's statement that the next domain is chosen by whether results can be verified, with science among the candidates. He mentions neither of those two entries and does not say Antigravity would be used for science.
On the site 'world model' appears in Zenithon's financing (world models for extreme physics) and PhAI Labs' cross-domain world-model entry.
Kilpatrick calls Omni a world model for how much it understands about the world, not because it simulates an environment, and says the definition is changing. The word need not mean the same thing in these places.
He calls science 'could be a really interesting one', on the condition that results can be verified. In science, results that can be verified quickly and cheaply (a proof, a simulation) are one kind and results that need weeks in a wet lab another. When reading a claim that AI is 'taking off' in science, the first question can be which kind of verifiable result it is. He does not separate the two.
'World model' can mean a model that simulates a physical system, or, as for Kilpatrick, a single multimodal model with an understanding of the world. When reading a claim it helps to ask which is meant, because the evidence differs: a model that simulates physics can be set against measurements, while 'understands the world' can for now only be shown.
This is Sequoia's own show. The science passages are his views when asked and the questions do not press for evidence, so the page records them as his judgments.
Coding now feels close to a narrow superintelligence.
Guest's judgment
He says it depends how you quantify it; no measure is given.
Domains where results are easier to verify (mathematics, finance, perhaps science) will see similar progress sooner, and before general intelligence.
Guest's judgment
A forecast with no timetable; he says what he thinks about most is coding.
Omni is a single model, not requests routed to several models.
Company-reported
His statement; no architecture is given and he says he is not an architecture expert.
What works in Omni now is video editing; the other outputs are not good enough, are not released, and it is not real time.
Current system
His statement; no evaluation is given.
Omni can do many of the same things as traditional world models with a different architecture that may be more scalable.
Guest's judgment
'Might be'; no comparison and no scientific example is given.
DeepMind started with a mission 'to solve disease and all these things', and its culture carries its scientist leaders' influence.
Guest's judgment
His view of the company's culture, with no project named.
Whether DeepMind or Google publishes a named result in a quickly verifiable science domain made with a Gemini system, with an outside check.
Would show up as A mathematics or computational-science result saying how the system did it, who verified it, and the failed attempts.Whether, once Omni's other outputs are released, anyone uses it for a science task and publishes an evaluation.
Would show up as A named science task, with a comparison against a specialised model and the measure used.Whether anyone proposes a common evaluation standard for the term 'world model' in science AI.
Would show up as A benchmark stating which meaning it tests and including tasks set against measurements.Later evidence that tests or limits what was said, and targets whose date has passed.
He says coding turned out to be the general-purpose agent harness and that Antigravity would power Google's other products. Four months later DeepMind showed Antigravity hunting antimicrobial peptides, calling it an agentic science workbench, with no experimental result. That shows the general harness being pointed at science, not that it works there.Timeline entry →
He says Omni can do many of the things traditional world models do and may scale better. DeepMind then released Genie 4, a real-time interactive world model in the older sense, pitched as a training environment for robots and agents, with Gemini Omni anchoring real-world appearance; the announcement gives no numbers. DeepMind did not replace that kind of world model with Omni; it put Omni inside the simulator.Timeline entry →
Read 2026.10.27 This is a full episode of Sequoia's own show (episode 88, host Sonya Huang), and this page covers only three parts that bear on science: whether science is the next narrow superintelligence (about minutes 20 to 24), Omni and 'world models' (about 28 to 34), and DeepMind's scientific culture (about 44 to 48). What was read is the timestamped transcript the author pasted, checked for these passages against the transcript on Sequoia's episode page; the audio was not heard. The release date (11 June 2026) comes from a third party's episode summary and a search result, since Sequoia's page prints none. His post is from the author page on Google's blog. The parts on coding, agent harnesses, enterprises and Google's products are outside this page. Open the interview ↗Sequoia episode page (with transcript) ↗