Ex-Meta VP: ChatGPT Is Just the Trailer for What’s Coming Next

Summary

Manohar Paluri argues that today’s frontier AI models and personal assistants are only an early stage of the technology’s impact. AI has made rapid progress in digital domains such as coding and mathematics, but its larger potential—and a major remaining bottleneck—is learning through interaction with the physical world.

Frontier Models and the Next AI Frontier

• Paluri describes frontier models as systems with capabilities that previously seemed out of reach. Their progress often makes once-surprising abilities feel ordinary soon afterward.
• Language models are increasingly useful for building digital applications because they can apply general capabilities to new tasks with limited customization.
• He believes the next major step is physical intelligence: models forming hypotheses, taking actions in the real world, observing results, and learning from them.

Why Physical-World Learning Is Difficult

• Code and mathematical reasoning can benefit from relatively fast, repeatable verification—for example, compiling code or checking a formal proof.
• Physical experiments are slower and harder to verify. In drug discovery, chemists may need to perform multistep procedures, operate specialized equipment, and wait days or weeks to synthesize and test a compound.
• Experiments can also be sensitive to small variations, and results are not always digitized. This makes the real-world learning loop dependent on human labor, physical materials, and long lead times.
• Paluri sees automated laboratories, robotics, and better data collection as ways to make currently data-poor scientific fields more measurable and faster to iterate in.

Warp Sciences and Scientific Discovery

• Paluri says he left Meta to work on what he views as a fundamental bottleneck, rather than because he was dissatisfied with frontier AI research.
• At Warp Sciences, the initial focus is small-molecule drug discovery and lead optimization, using robotics, foundation models, and scientific expertise to make experimental work more efficient.
• The near-term measure of progress is producing more useful compounds per chemist per week and reducing the chemistry bottleneck in drug discovery.
• The broader goal is for each real-world experiment—whether conducted by people, robots, or both—to build on previous results and make subsequent discovery faster.
• Paluri describes the long-term ambition as helping create scientific superintelligence and ultimately getting new drugs to patients sooner. He also points to materials science and renewable materials as areas where improved physical-world experimentation could matter.

AI Applications, Software, and Future Model Architectures

• Paluri does not expect current AI applications to become worthless as models improve. He advises application builders to structure their products and engineering processes so they can take advantage of new models without repeatedly rebuilding everything.
• Durable value may come from specialized data, subject-matter expertise, and fast iteration—not just access to a powerful model.
• He distinguishes between using the current model paradigm to create products and making a longer-term bet on a different approach, such as world models that simulate environments and support learning through simulated actions.
• He notes that genuine paradigm shifts can change the rules of a field, citing the shift from handcrafted computer-vision features to deep learning as an example. Possible future shifts could come from model training, inference, or hardware.
• A useful way to evaluate opportunities, he says, is to ask what would happen if a key bottleneck became dramatically better—for example, if hardware improved tenfold or inference became much cheaper.

Open Weights, Closed Models, and Ecosystems

• Paluri distinguishes open weights from fully open source. Open weights allow users to customize a model, but do not necessarily provide the complete training stack or the ability to reproduce the model from scratch.
• He sees open ecosystems as strong at enabling parallel experimentation: people can reuse components, contribute improvements, and apply expertise in many different ways. He points to PyTorch and Linux as examples of the broader pattern.
• Closed models may retain advantages in capability, reliability, and ease of use because they are backed by substantial investment and provided as managed services.
• Open-weight models offer more control, customization, and potential efficiency, but users may trade off some capability or convenience.
• Paluri expects both approaches to remain important. He also argues that increasingly capable AI systems will require AI-assisted cybersecurity and careful safety measures, especially before powerful capabilities are made broadly available.

AI Sovereignty

• Paluri supports sovereignty efforts when they motivate genuine innovation, but cautions against treating sovereignty as simply owning a domestic foundation model.
• A country’s ability to build and operate AI depends on the whole stack: hardware, models, products, adoption, and the systems that turn technology into value.
• He favors balancing broad access and user benefit with appropriate safeguards, with decisions made according to the circumstances of each sector.

Product Launches and Public Perception

• Paluri sees launches on X as a way for companies to reach a concentrated audience of AI users and experts, receive immediate feedback, and respond quickly to issues.
• He describes this as fine-tuning the perception of a launch, not substituting perception for product quality. In his view, a weak product will eventually be exposed regardless of how its launch is managed.

Career and Leadership Lessons

• Paluri’s recurring approach is to identify a bottleneck, learn from domain experts, build conviction, take a leap, assemble a strong team, and judge success by whether people use the result in the real world—not merely by papers, demos, or benchmarks.
• He credits Yann LeCun with a strong belief that existing technical paradigms have limits and that new ones will be needed. Paluri also describes a difference in their views on research organization: he favors combining researchers and engineers around ambitious goals, while acknowledging the value of giving research teams independence and room to explore.
• He describes Mark Zuckerberg as a bold, fast-learning leader focused on connecting technology to practical value. Paluri highlights the Instagram acquisition as an example of a decision that seemed risky at the time but later appeared prescient.
• His personal advice is to combine long-term optimism with day-to-day discipline: stay open to learning, keep working through setbacks, and maintain high standards.

Reviews, Guides, or Tutorials

None are presented. The video is an interview and analysis of AI, scientific automation, and startup-building.

Main Speakers and Sources

• Manohar Paluri: Co-founder and CTO of Warp Sciences and former Meta VP
• Siddharth: Host of The Neon Show