Fedus bets AI will leap from code to atoms
- AI pioneers pivot from scaling digital models to accelerating material science for chips and batteries.
- Startups like Periodic Labs build closed-loop systems where AI directs physical lab experiments to generate proprietary data.
- The new AI bottleneck isn’t compute or algorithms - it's the stubborn physics of the real world.
AI’s next leap isn’t a better chatbot - it’s a new semiconductor. After years spent scaling GPT-4, co-founder Liam Fedus has launched Periodic Labs to apply AI to material discovery, targeting the physical bottlenecks in chips and batteries.
Fedus argues on No Priors that digital progress has outpaced our ability to manipulate atoms. “Science ultimately isn't sitting in a room thinking really hard,” he says. “You have to conduct experiments to interface with reality.” His system uses LLMs as an orchestration layer, directing robots to run physical tests and capture ground-truth data, a necessity because reported material properties in academic papers often vary by orders of magnitude.
This closed-loop approach aims to build a proprietary data moat, a sharp contrast to training models on the messy, often contradictory data scraped from the internet. The goal is to accelerate the slow feedback loops of physical R&D.
The materials push comes as others in the AI ecosystem attack different physical limits. On This Week in Startups, Nick Harris of Light Matter argued that copper wiring is now a ceiling for AI progress, forcing a shift to photonic chips that can link GPUs over a kilometer with light. He claims this photonic technology can triple model training speeds.
Liam Fedus, No Priors:
- For systems that are strongly governed by quantum mechanical effects, there is some generalization there.
- But if you produce a system that has modeled quantum mechanical objects really accurately, it's not really helping much on fluid dynamics.
The race reflects a broader trend: physicists like Fedus and Anthropic’s Dario Amodei are leading the AI charge. Fedus notes that after projects like the Large Hadron Collider, high-energy physics became bottlenecked by massive, slow hardware. AI offered a frontier where principled, first-principles thinking yields immediate, software-driven results.
The capital required is steep - Fedus says GPU compute is his biggest cost - but the potential payoff is foundational: redesigning the physical components that underpin everything from data centers to electric vehicles.
Source Intelligence
- Deep dive into what was said in the episodes
AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus • Apr 3
- Many AI researchers like Dario Amodei and Adam D'Angelo have physics backgrounds, a trend Liam Fettis attributes to physicists' principled thinking and high leverage in AI.
- Periodic Labs is building an AI foundation lab for atoms, focusing on applying AI to material science, chemistry, and the physical world.
- Periodic's AI system acts as an orchestration layer, using large language models to direct experiments and specialized neural nets designed for atomic systems.
- Fettis says the acceleration of digital software engineering creates an imperative to connect AI systems to the physical world for scientific and technological progress.
- Current AI technology, including improved reasoning and reliable tool use, is now sufficiently advanced to connect AI to the physical world, unlike in 2022.
- Periodic leverages existing models for coding and language, spending zero effort on improving them, to focus its machine learning efforts on physical world frontiers.
- A key data challenge in materials science is that reported property values from literature often span orders of magnitude, making ground truth difficult to establish without experiments.
- Periodic's approach relies on an interactive closed-loop system where experimental data feeds back to identify aberrations and patterns to drive the next experiments.
- Data generalization for physical systems is often domain-specific; a model trained on quantum mechanical objects doesn't help much with fluid dynamics.
- Fettis sees the most internal advances where there is an abundance of data in a specific chemical or material space.
- The biggest capital cost for Periodic's work is GPU compute, not physical infrastructure, though lab setup has long lead times and calibration difficulties.
- Fettis believes AI systems exhibit odd spikiness in intelligence, being world-class in one domain but potentially poor in adjacent ones, challenging the idea of intelligence as a scalar.
- Software engineering self-improvement by AI is happening now due to cheap, verifiable environments like unit tests, but this doesn't automatically translate to other domains like biology.
- Fettis says AI research self-improvement is a slower outer loop than software engineering because experiments require GPUs and hours to evaluate model convergence and scaling properties.
- While reliable robotics would be a huge accelerator, Periodic currently uses hybrid human-automation systems and off-the-shelf robotics to generate sufficient high-throughput data.
- Fettis views the future as one where AI generates matter, profoundly impacting semiconductors, aerospace, and energy by increasing the pace of physical world development.
- The multidisciplinary collaboration at Periodic, with physicists, chemists, AI researchers, and engineers, is allowing veteran scientists to see their fields fundamentally change.
Also from this episode: (1)
Robotics (1)
- Fettis sees the interface of AI with the physical world via robotics as a transformative opportunity, given labor shortages and the vast number of people who work with the physical world.
How 3 CEOs Use AI to Run $10B in Companies | This Week in AI • Apr 2
- Fundamental's large tabular model architecture differs from LLMs because it is not autoregressive; changing column order in a table does not change the output.
- Frankel claims LLMs are not suitable for deterministic predictive tasks like fraud detection, where output consistency is critical.
- Frankel says traditional machine learning algorithms still outperform most LLMs for predictive tasks on tabular data.
- Jeremy Frankel says his company Fundamental emerged from stealth as a unicorn 16 months after founding.
- Victor Riparbelli outlines Synthesia's thesis that AI will drive the marginal cost of creating video and audio content to near zero.
- Nick Harris states that the classic rules driving computing progress, Moore's Law and Denard scaling, are now over.
- Harris says the future of computing relies on two things: building bigger computer chips and networking them together at high bandwidth.
- Harris states copper interconnect limits GPU proximity in racks, while photonics allows GPUs to be separated by a kilometer and still act as a single system.
- Nick Harris says Light Matter's chip with Qualcomm pushes 1.6 terabits per second over a single optical fiber, equivalent to 1,600 homes with gigabit internet.
- Harris states Light Matter's M1000 chip has 114 terabit per second bandwidth, roughly equal to undersea cables connecting North America and Europe.
- Nick Harris claims photonic technology can 3x the training time for large AI models, significantly accelerating the rate of AI progress.
- Harris says Light Matter builds chips for hyperscalers like Google and Amazon, as well as for GPU and networking companies.
- Riparbelli estimates that generating a personalized one-hour movie with current state-of-the-art video models would cost around $700, making it commercially unsustainable.
- Jeremy Frankel states that exploring non-NVIDIA hardware like Amazon's Tranium chips is a priority to avoid dependency on a single hardware platform.
Also from this episode: (14)
AI & Tech (14)
- Jeremy Frankel states that 70% of people in a poll believe AI will decrease job opportunities, but only 30% of Americans worry it will happen to them.
- Jeremy Frankel argues that the first major wave of AI automation is targeting cognitive work, not just physical labor.
- Jeremy Frankel's company Fundamental built a foundation model for tabular data, not an LLM, to address structured enterprise data in rows and columns.
- Frankel states that large language models are designed for unstructured data like text and video, but most useful enterprise data is structured tabular data.
- Fundamental's Nexus model aims to unify various predictive use cases like credit card fraud and demand forecasting into a single, more accurate model.
- Frankel states Fundamental raised a $255 million Series A led by Oak with participation from Valor Battery and Salesforce.
- Victor Riparbelli says Synthesia, an AI video platform for business, has 90% of Fortune 100 companies as customers.
- Riparbelli states Synthesia has over $100 million in ARR and a $4 billion valuation after raising over $500 million.
- Riparbelli says Synthesia's initial focus was enabling PowerPoint users to create video content, a demand they identified in 2022.
- Riparbelli argues OpenAI shutting down Sora shows the company learned the 'unteachable lesson' of focus the hard way.
- Riparbelli claims Anthropic's success with Claude Code shows that focusing solely on B2B code generation is a highly valuable near-term strategy.
- Jeremy Frankel notes that at a recent Lightspeed founder retreat, everyone was discussing Claude Code, not other AI products.
- Riparbelli says Synthesia is developing real-time interactive video, where users can role-play with AI avatars, moving beyond broadcast video.
- Nick Harris explains that modern AI data center racks consume a megawatt of power and require reinforced concrete due to their weight and cooling needs.