Building an AI Physicist: ChatGPT Co-Creator's Next Venture

The Origin Story of Periodic Labs Co-Founders

The co-founders of Periodic Labs first met eight years ago at Google Brain in an unexpectedly physical way - flipping a massive tire at the Google Rails gym facility.

The Tire-Flipping Moment:

• The Challenge: Ekin was attempting to flip a tire so large that a single person couldn't manage it alone
• The Solution: He spotted Liam and pulled him over, suggesting they could accomplish it together
• The Success: Their combined effort successfully flipped the tire, marking the beginning of their professional relationship

From Physical to Intellectual Collaboration:

Over the years following their initial meeting, their conversations naturally gravitated toward deep scientific topics:

• Quantum mechanics discussions became a regular occurrence
• Superconductivity was another frequent topic of conversation
• Neither initially imagined they would eventually work together on physics research

This casual but intellectually stimulating relationship laid the groundwork for their eventual partnership in founding Periodic Labs, where they would combine their expertise in AI and physics to tackle scientific discovery.

Timestamp: [0:49-2:14]

The Evolution of AI in Scientific Discovery

Large Language Models are transitioning from helpful assistants to first-class citizens in physics and chemistry research, fundamentally changing how scientists approach their work.

Current Applications in Scientific Work:

1. Knowledge Retrieval: Scientists use chatbots to quickly recall forgotten concepts in chemistry and physics
2. Simulation Coding: LLMs provide significant assistance in writing complex physics simulations
3. Research Acceleration: Models help bridge gaps in scientific knowledge and methodology

The Scaling Laws Connection:

Both AI and physical sciences are experiencing similar scaling phenomena:

• Language Models: Improvements in reasoning capabilities with increased compute
• Material Science: Scaling laws apply to both simulations and experimental processes
• Reinforcement Learning: High-compute RL showing promising results across domains

The Vision for AI-Driven Science:

The ultimate goal extends beyond chatbots to creating technology that accelerates physical R&D:

• Verifiable Results: Physics provides excellent reward functions due to its verifiable nature
• Fast Iteration: Relatively quick feedback loops compared to other scientific domains
• Simulation Support: Existing simulators for large classes of physical systems
• AI Scientist Development: Physics serves as the ideal starting point for automated scientific discovery

Timestamp: [2:14-3:50]

A Frontier AI Research Lab for Physical Sciences

Periodic Labs represents a new approach to scientific discovery, combining Large Language Models with experimental physics and chemistry to accelerate research and development.

Core Mission and Approach:

• Frontier AI Research Lab focused specifically on advancing physics and chemistry
• Tight Integration of experiments, simulations, and LLMs working in conjunction
• High-Throughput Data Generation through purpose-built laboratory facilities
• Iterative Science Acceleration using all available tools that human scientists typically employ

The Physical Reward Function Innovation:

Traditional AI Training:

• Math graders: "What is 2 plus 2?" → Ground truth is 4
• Code graders: Programmatically checkable problems
• Heavy optimization pressure on digital, verifiable tasks

Periodic Labs Approach:

• Physically Grounded Reward Function replaces digital graders
• Nature as RL Environment - the real world becomes the training ground
• Experiment-Based Ground Truth - physical results always take precedence
• Error Correction Against Reality - when simulators have deficiencies, experiments provide the correction

Operational Framework:

1. Laboratory Infrastructure: Building high-throughput, high-quality experimental capabilities
2. Multi-Modal Integration: Combining LLMs, simulations, and real-world experiments
3. Iterative Optimization: Using physical results to continuously improve AI models
4. Scientific Acceleration: Replacing human bottlenecks in the research process

Timestamp: [3:50-5:27]

The Technical Evolution from Raw Model to Conversational AI

The development of ChatGPT involved transforming a basic autocompletion model into a sophisticated assistant through a multi-stage training process that evolved rapidly over several years.

Original RLHF Pipeline:

1. Pre-trained Foundation: Started with a raw substrate autocompletion model
2. Supervised Fine-tuning: Created input-output pairs for desired assistant behavior
3. Human Preference Training: Humans ranked completion A versus completion B for given inputs
4. Reward Function Creation: Built optimization targets based on human preferences
5. Reinforcement Learning: Optimized the model against the human-preference reward function

Early Limitations and Challenges:

Mathematical Weakness:

• Original reward functions couldn't determine mathematical correctness
• Early ChatGPT versions were not particularly strong at math problems
• The reward function encoded "be a friendly assistant" but lacked precision

Reward Function Constraints:

• Primary Goal: Help people achieve their objectives in a friendly manner
• Missing Elements: No sense of mathematical validity or code correctness
• Digital Foundation: All training based on internet content, textbooks, and papers

The Evolution to Better Systems:

Key Improvements:

• Much Better Reward Functions: More precise and accurate optimization targets
• Reasoning Capabilities: Significant advances in logical thinking and problem-solving
• Correctness Validation: Ability to determine mathematical and code validity
• Huge Performance Gap: Massive difference between original model and current versions

This digital-first approach provided an excellent foundation, but the next frontier requires moving beyond text-based training to include real-world experimental validation.

Timestamp: [5:47-7:44]

Moving Beyond Digital Optimization to Physical Reality

While digital training methods have created powerful AI systems, advancing scientific discovery requires incorporating real-world experimental validation into the AI development process.

The Limitation of Digital-Only Training:

Current AI systems are optimized against digital tasks:

• Internet-Based Content: Training on textbooks, papers, and online resources
• Programmatic Verification: Math and code problems with clear right/wrong answers
• Simulated Environments: Digital representations of physical phenomena

The Scientific Reality Gap:

Science is Fundamentally Experimental:

• Real scientific breakthroughs require validation against physical reality
• Simulations have inherent deficiencies and limitations
• Digital reward functions cannot capture the full complexity of natural phenomena

Periodic Labs' Solution:

Experiment-in-the-Loop Approach:

• Physical Ground Truth: Real experiments become the ultimate arbiter of correctness
• Nature as RL Environment: The physical world provides the training feedback
• Error Correction Mechanism: When simulations fail, experiments provide the correction
• Iterative Improvement: Continuous refinement based on physical results

This represents a fundamental shift from optimizing against human preferences or digital metrics to optimizing against the laws of physics themselves, creating AI systems that can genuinely advance scientific understanding and discovery.

Timestamp: [7:37-7:55]

Essential Insights:

1. Origin Story: Liam Fedus and Ekin Dogus Cubuk met eight years ago at Google Brain while flipping a massive tire, leading to years of conversations about quantum mechanics and superconductivity
2. LLM Evolution in Science: Large Language Models are transitioning from helpful tools to first-class citizens in physics research, with scaling laws appearing in both AI and physical sciences
3. Periodic Labs Mission: A frontier AI research lab combining LLMs, simulations, and real-world experiments to accelerate physics and chemistry discovery

Actionable Insights:

• Physical Reward Functions: Replace digital graders with experiment-based ground truth where nature becomes the RL environment
• Integrated Approach: Combine high-throughput experimentation with AI models for iterative scientific discovery
• Beyond ChatGPT: Move from human preference optimization to physics-based validation for genuine scientific advancement

Timestamp: [0:49-7:55]

People Mentioned:

• Liam Fedus) - Co-creator of ChatGPT and former VP of post-training research at OpenAI, co-founder of Periodic Labs
• Ekin Dogus Cubuk - Former head of materials science and chemistry research at Google DeepMind, co-founder of Periodic Labs

Companies & Products:

• Google Brain - AI research division where the founders first met
• OpenAI - Where ChatGPT was developed using RLHF pipeline
• Google DeepMind - Research lab focusing on materials science and chemistry
• Periodic Labs - Frontier AI research lab combining LLMs with physics experiments

Technologies & Tools:

• ChatGPT - Conversational AI created through reinforcement learning from human feedback
• RLHF (Reinforcement Learning from Human Feedback) - Training methodology used to create ChatGPT
• Large Language Models (LLMs) - Foundation models being applied to physics and chemistry research

Concepts & Frameworks:

• Scaling Laws - Principles showing improvement with increased compute, applicable to both AI and physical sciences
• Reward Functions - Optimization targets used in AI training, evolving from human preferences to physical experiments
• Supervised Fine-tuning - Initial training phase using input-output pairs for desired behavior
• Physics Simulations - Computational models used alongside experiments for scientific discovery

Timestamp: [0:49-7:55]

AI Agents with Physics-Based Tools

Periodic Labs is revolutionizing AI by giving agents access to quantum mechanics as computational tools, similar to how current AI systems use Python or web browsers.

Core Approach:

1. Physics as Tools - Instead of traditional coding tools, agents get quantum mechanics simulators and systems
2. Lab-Based Rewards - Real laboratory results become the reward function for training AI agents
3. Natural Evolution - This represents the logical next step for AI systems beyond current capabilities

Target Physics Domain:

• Quantum Mechanical Energy Scale - Focus on Schrödinger's equation level physics
• Solid State Physics - Materials science and chemistry applications
• Biological Relevance - The energy scale where biology and chemistry around us actually happen
• Practical Applications - Where materials that affect our daily lives are created

Foundation Model Strategy:

The goal is teaching LLMs to become foundation models for quantum mechanics, representing the next frontier after current models mastered logic and math.

Timestamp: [8:01-10:06]

Robotic Materials Discovery Method

Powder synthesis represents one of the most fundamental and automatable ways to create new materials, making it perfect for AI-driven discovery.

The Process:

1. Simple Method - Take powders of existing materials and mix them
2. Heat Treatment - Apply specific temperatures to create new materials
3. Robot-Friendly - Simple enough for basic robotic systems to perform

Automation Potential:

• Low Complexity - A robot at the level of SF airport coffee-making machines can handle the process
• Mix and Heat - Robots can mix powders and operate furnaces effectively
• Scalable Discovery - Enables high-throughput materials exploration

Discovery Applications:

Using powder synthesis, researchers can discover:

• New Superconductors - Materials with zero electrical resistance
• Advanced Magnets - Critical for various technologies
• Novel Materials - Important for emerging technologies

Quantum Foundation:

At its core, powder synthesis is governed by quantum mechanics, making it an ideal testing ground for AI systems learning fundamental physics principles.

Timestamp: [9:18-10:06]

The Iteration Problem in AI Science

Current AI models fail at scientific discovery because they lack the fundamental requirement for scientific progress: the ability to iterate and learn from experiments.

Core Issues with Current Models:

Lack of Scientific Method:

• No Iteration - Models can be smart but don't iterate on scientific problems
• Human Parallel - Even brilliant humans won't discover anything important without the chance to iterate
• Missing Process - Current models haven't been taught the method of scientific inquiry

Training Data Problems:

1. Noisy Literature Data - Physical properties in literature can span many orders of magnitude
2. No Negative Results - Scientific literature rarely publishes failed experiments
3. Missing Learning Signals - Valid negative results provide valuable learning that models never see

Epistemic Uncertainty:

• Reducible Uncertainty - Some uncertainties can only be resolved through actual experiments
• Distribution Replication - Models trained on noisy data can only replicate that noise, not discover truth
• No Deeper Understanding - Without experimentation, models can't develop genuine physics understanding

The Solution Requirements:

• Real Physics Integration - Must work with actual physics, not just simulations
• Iterative Learning - Models need opportunities to iterate on quantum mechanics understanding
• Experimental Feedback - Direct laboratory results as training signals

Timestamp: [10:12-12:36]

Concrete Benchmarks for AI Scientific Discovery

Periodic Labs has established clear, measurable goals that provide unambiguous signals of progress in AI-driven scientific discovery.

Primary Success Metrics:

High-Temperature Superconductivity:

• Current Record - 135 Kelvin at ambient pressure
• Clear Target - Any temperature above this threshold represents measurable progress
• Fundamental Impact - Room-temperature superconductors would revolutionize technology

Direct Material Properties:

• Ductility Measurements - How much a material can deform without breaking
• Toughness Testing - Material's ability to absorb energy before fracturing
• Strength Analysis - Maximum stress a material can withstand
• Processing Effects - How manufacturing processes affect material properties

Validation Advantages:

• Hard to Hack - Unlike other AI training metrics, physical properties can't be gamed
• Real-World Signal - Actual laboratory measurements provide direct feedback to AI systems
• Immediate Verification - Results are measurable and verifiable in real-time

Practical Applications Test:

• Design Challenge - Can the system create materials with specific required properties?
• Industry Problems - Accelerating development for space, defense, and semiconductor applications
• Property-Driven Discovery - Solving specific material challenges across industries

The approach ensures that progress is grounded in reality rather than abstract benchmarks.

Timestamp: [12:36-13:56]

The Perfect Storm of Technology and Talent

The convergence of frontier AI technology and unprecedented team assembly capabilities has created a unique window for tackling AI-driven physics discovery.

Technological Enablers:

• Frontier AI Advances - Recent breakthroughs in the last couple of years have made this approach viable
• LLM Foundation - Large language models provide the base intelligence needed for scientific reasoning
• Automation Technology - Robotic systems have reached sufficient sophistication for laboratory work

The Team Assembly Advantage:

N of One Team Composition:

• Physicists - Deep understanding of fundamental science principles
• Chemists - Expertise in material synthesis and chemical processes
• Simulation Experts - Computational modeling and theoretical calculations
• World-Class ML Researchers - Cutting-edge machine learning and AI development

Historical First:

This represents the first time these diverse expertise areas have been combined in one concerted effort specifically for AI-driven scientific discovery.

Broad Application Potential:

The AI physicist approach can revolutionize multiple industries:

• Advanced Manufacturing - Automated material design and optimization
• Material Science - Accelerated discovery of new materials
• Chemistry - Enhanced understanding of chemical processes
• Any R&D Process - Where physical world interaction is required

Timing Factors:

• Technology Maturity - AI and robotics have reached necessary capability thresholds
• Talent Availability - Top researchers from major AI labs are now accessible for focused efforts
• Market Readiness - Industries desperately need breakthrough materials for emerging technologies

Timestamp: [14:02-14:44]

Fractal Expertise Architecture

Periodic Labs built their team using a systematic approach to cover all essential domains with world-class talent, recognizing that expertise operates at multiple nested levels.

Core Expertise Pillars:

1. LLM Expertise - Advanced machine learning and AI development
2. Experimental Expertise - Hands-on laboratory and synthesis work
3. Simulation Expertise - Computational modeling and theoretical physics

Fractal Team Structure:

Each major pillar contains multiple specialized sub-teams, creating a fractal organization where expertise branches into increasingly specific domains.

Experimental Team Breakdown:

• Solid State Chemistry - Material synthesis and chemical processes
• Solid State Physics - Physical properties and behavior of materials
• Automation Systems - Robotic laboratory equipment and processes
• Facilities Operations - Laboratory management and operational aspects

Simulation Team Specializations:

• Theoretical Physics - Fundamental physics principles and modeling
• Computational Implementation - Coding and software development for simulations
• Quantum Mechanics Modeling - Schrödinger equation and quantum systems

LLM Team Domains:

• Mid-Training - Model development and training processes
• Reinforcement Learning - Reward systems and iterative learning
• Infrastructure - Computing systems and model deployment

Recruitment Philosophy:

• World-Class Standard - Only top-tier talent in each specialized area
• Comprehensive Coverage - Every critical domain must be represented
• Deep Specialization - Recognition that true expertise requires narrow focus within broader domains

Timestamp: [15:10-15:53]

Essential Insights:

1. AI Physics Revolution - Periodic Labs is creating AI agents that use quantum mechanics as tools, similar to how current AI uses Python or browsers
2. Iteration is Key - Current AI models fail at science because they can't iterate; even brilliant humans need experimental feedback to discover anything important
3. Real Lab Requirement - True scientific AI requires physical laboratories, not just simulations, to collapse epistemic uncertainty and generate valid learning signals

Actionable Insights:

• Powder Synthesis Focus - Simple robotic processes like mixing powders and heating can discover superconductors and advanced materials
• Clear Success Metrics - Progress measured by concrete goals like surpassing 135 Kelvin superconductivity temperature
• Team Architecture - Success requires world-class talent across LLM expertise, experimental work, and simulation in a fractal organization structure

Timestamp: [8:01-15:53]

Concepts & Frameworks:

• Quantum Mechanics - Fundamental physics framework governing material behavior at atomic scale
• Schrödinger's Equation - Core quantum mechanical equation describing particle behavior
• Powder Synthesis - Materials science method of mixing and heating powders to create new materials
• High-Temperature Superconductivity - Materials that conduct electricity with zero resistance at elevated temperatures
• Epistemic Uncertainty - Reducible uncertainty that can only be resolved through experimentation
• Solid State Physics - Branch of physics studying properties of solid materials
• Solid State Chemistry - Chemistry focused on synthesis and properties of solid materials

Technologies & Tools:

• Robotic Laboratory Systems - Automated equipment capable of materials synthesis and testing
• Quantum Mechanics Simulators - Computational tools for modeling quantum mechanical systems
• Furnace Systems - High-temperature equipment for materials processing
• Materials Testing Equipment - Instruments for measuring ductility, toughness, and strength

Scientific Methods:

• Scientific Inquiry Method - Iterative process of simulation, calculation, experimentation, and result analysis
• Negative Results Publishing - Scientific practice of documenting failed experiments for learning
• Materials Property Measurement - Quantitative assessment of material characteristics

Timestamp: [8:01-15:53]

Scaling Laws vs. Domain-Specific Challenges

The fundamental issue isn't that scaling laws don't work - they empirically continue to hold. The problem is understanding what the y-axis represents and recognizing that test distributions vary dramatically.

The Distribution Problem:

1. In-Domain vs. Out-of-Domain Performance - While both improve as power laws, the slope for out-of-domain tasks may be so small it's practically useless
2. Training Set Proximity - You need centuries of scaling before getting useful results if your target is too far from your training distribution
3. Domain Shift Reality - A coding model trained on internet data won't cure cancer because the knowledge simply doesn't exist in that training set

Missing Data Challenge:

• Experimental data doesn't exist for many scientific problems
• Formation enthalpy labels are so noisy that machine learning models trained on them aren't predictive enough
• Negative results aren't usually published, and they're highly context-dependent
• Superconductivity datasets have noise floors so high that training on them doesn't help

The Solution Approach:

• Make your target as close to your in-domain training set as possible
• Iterate on changing your training set to match what you want to do
• Scale up specifically for the domain you care about - in this case, advancing physical R&D

Timestamp: [17:21-22:14]

Strategic Domain Selection for AI-Driven Discovery

High-temperature superconductivity serves as both a specific goal and a comprehensive testing ground for developing AI physics capabilities.

Superconductivity as a North Star:

1. Multi-Layered Challenge - Like DeepMind and OpenAI's approach to AGI, achieving high-temperature superconductivity requires mastering numerous sub-goals
2. Autonomous Systems Development - Requires building autonomous synthesis, autonomous characterization, and advanced simulation capabilities
3. Fundamental Scientific Impact - Finding a 200 Kelvin superconductor would reveal profound insights about the universe before any product development

Why These Domains Work:

• Rich Sub-Goal Structure - Each milestone toward superconductivity delivers impactful capabilities to the broader scientific community
• Technology Integration - Combines LLM-driven simulations with physical experimentation and characterization
• Clear Success Metrics - Temperature thresholds provide objective benchmarks for progress

Building Blocks vs. Off-Ramps:

These aren't dead-end specializations but essential components of a general AI physicist:

• Autonomous synthesis capabilities transfer across materials science
• Characterization techniques apply to multiple physical domains
• Simulation integration scales to other scientific fields

The approach mirrors successful AI development: tackle specific, measurable challenges that build toward broader capabilities rather than attempting general intelligence directly.

Timestamp: [22:32-23:57]

Assembling Domain Expertise for Scientific AI

Periodic Labs has strategically recruited specialists across the key technological pillars necessary for automating scientific discovery.

Team Building Philosophy:

• Best-in-Class Specialists - For each technological sub-pillar, they recruit people who have genuinely innovated in those specific areas
• Recent Technology Integration - The required technology has only emerged in the last couple of years, requiring experts who understand cutting-edge AI techniques
• Industry-Academia Bridge - Many advanced industries have the desire for AI integration but lack knowledge of the most recent AI breakthroughs

Data Production Challenge:

1. Experimental Data Generation - Unlike internet data, scientific data must be actively produced through experiments and simulations
2. Industry Silos - Critical data is scattered across advanced industries in isolated repositories
3. Knowledge Gap - Traditional scientific institutions may not be aware of the latest AI techniques driving recent breakthroughs

Historical Context:

The GPT-3 paper established that language models are few-shot learners and introduced scaling laws. Follow-up research on "Scaling Laws for Generative Modeling" demonstrated predictable performance improvements through coordinated scaling of compute and data, theorizing that continued scaling would produce emerging capabilities and out-of-domain reasoning abilities.

Timestamp: [16:01-17:21]

Essential Insights:

1. Scaling laws work but have distribution limits - Performance improvements follow power laws, but out-of-domain slopes may be too small for practical use
2. Scientific data doesn't exist like internet data - Experimental data must be actively generated and is often too noisy or incomplete for effective training
3. Domain-specific approaches are necessary - Success requires iterating on training sets to match target domains rather than relying on general scaling

Actionable Insights:

• Target training data should be as close as possible to the intended application domain
• Scientific AI requires combining autonomous synthesis, characterization, and simulation capabilities
• High-temperature superconductivity serves as an ideal north star with measurable sub-goals that benefit the broader scientific community

Timestamp: [16:01-23:57]

People Mentioned:

• Liam Fedus - Former VP of post-training research and co-creator of ChatGPT at OpenAI, co-founder of Periodic Labs
• Ekin Dogus Cubuk - Former head of materials science and chemistry research at Google DeepMind, co-founder of Periodic Labs

Companies & Products:

• OpenAI - Published foundational papers on GPT-3 and scaling laws for generative modeling
• Google DeepMind - Research on vision models and scaling laws, comparison point for AGI development approach
• Periodic Labs - The startup focused on automating discovery in hard sciences through AI physics

Publications & Research:

• GPT-3 Paper - Established language models as few-shot learners and introduced scaling laws concept
• "Scaling Laws for Generative Modeling" - OpenAI follow-up paper showing predictable performance improvements through compute and data scaling
• CLIP Paper - OpenAI research demonstrating correlation between in-domain and out-of-domain generalization
• Vision Models Scaling Research - Work on scaling laws for vision models showing domain generalization patterns

Concepts & Frameworks:

• Scaling Laws - Predictable performance improvements through coordinated scaling of compute and data
• In-Domain vs Out-of-Domain Generalization - Performance improvements follow different power law slopes depending on distribution similarity
• Formation Enthalpy - Energy required to assemble atoms in desired configurations, critical for materials synthesis
• Autonomous Synthesis and Characterization - Automated systems for creating and analyzing materials properties

Timestamp: [16:01-23:57]

Revolutionary Physics Discovery

Transformative Scientific Impact:

1. Quantum Effects at High Temperatures - Observing quantum phenomena at room temperature would fundamentally update humanity's understanding of the universe
2. Phase Transition Robustness - Superconductivity's nature as a phase transition makes it resistant to simulation limitations and material defects
3. Universal Appeal - The field uniquely unites both veteran physicists with 40 years of experience and newcomers to physics

Technical Advantages:

• Crystal Property Dominance: Superconducting temperature is primarily determined by fundamental crystal properties rather than defects or microstructure
• Simulation-Friendly: Unlike other material properties that are affected by numerous unsimulatable factors, superconductivity maintains its essential characteristics
• Robust Manifestation: The property remains observable even when other complicating factors are present

Why This Matters for Humanity:

The discovery would represent such a significant breakthrough that it would be impactful for humanity even before any commercial product development, fundamentally changing how we understand quantum mechanics at macroscopic scales.

Timestamp: [24:03-25:04]

Strategic Approach to AI-Driven Discovery

The Full Loop Strategy:

1. Complete Contact Required - Must engage with the entire research process rather than staying in theoretical papers and textbooks
2. Domain-to-Domain Generalization - Testing how ML systems transfer knowledge between different scientific fields
3. Systematic Expansion - Create proven repeatable processes that can be applied across multiple scientific domains

Generalization Questions:

• Cross-Domain Transfer: How well do systems trained on superconductivity data perform on magnetism data?
• Field-Specific Adaptation: Different generalization patterns may emerge between physics domains versus fluid mechanics
• Fundamental Arguments: Core principles about ML system adaptability across scientific disciplines

Implementation Process:

The goal is to create a repeatable system, prove its effectiveness in one domain, then systematically expand through different scientific areas using the established methodology.

Timestamp: [25:11-26:07]

Balancing Scientific Goals with Business Viability

The Programming Analogy:

• Frontier Labs Model: Traditional AI companies had AI researcher as northstar, with AI programming as the commercially viable path
• Extraordinary Impact: Software engineering became the first major domain showing AI's productivity benefits beyond consumer applications
• Prior Updates: Programming AI caused significant updates in how people view AI model usefulness

Periodic's Commercial Path:

Co-pilots for Advanced Industries

• Target Markets: Space, defense, semiconductors, and other advanced manufacturing sectors
• Core Problem: These industries deal with materials and physics iteration as part of their workflow but lack good tools
• Opportunity: Massive R&D budgets in industries without adequate data or systems

Strategic Philosophy:

• Technology-Capital Integration: Understanding that technology and capital are intertwined for maximum scientific acceleration
• Commercial Success = Scientific Acceleration: A wildly successful commercial entity can maximally accelerate science
• Intelligence Layer: Becoming an intelligence layer for teams across advanced manufacturing industries
• Workflow Enhancement: Reducing iteration time, improving solutions, and accelerating researchers and engineers

Timestamp: [26:07-28:46]

Cross-Disciplinary Team Integration

Team Composition Challenge:

• 50/50 Split: Half the team consists of ML scientists with machine learning backgrounds
• Physical Scientists: Remaining half are physicists and chemists with lab experiment experience
• Cultural Bridge: Need to unite professionals from wet labs with those from computational backgrounds

Integration Strategies:

Bidirectional Learning:

1. Physics → ML Direction: Physical scientists learn how RL loops work and data cleaning processes
2. ML → Physics Direction: ML researchers gain understanding of physics, simulation tools, and scientific goals
3. Historical Context: Teaching the history of science, which is crucial for proper understanding

Structured Learning Sessions:

• Weekly Teaching Sessions: Regular cross-training between disciplines
• No Stupid Questions Culture: Environment where anyone can ask basic questions from either domain
• Faculty Involvement: Company includes faculty members who serve as excellent teachers

Practical Application:

The skills needed to teach LLMs to discover superconductors (reading literature, running simulations, conducting experiments) mirror the workflows of physical R&D researchers in target companies, making internal progress directly applicable to customer needs.

Timestamp: [28:46-31:35]

Specialized AI Training for Scientific Reasoning

Current AI Limitations:

• Existing Capabilities: Frontier AI labs have successfully trained models on math and logic
• Missing Domain: Physics and chemistry reasoning capabilities are not yet developed
• Training Gap: Need specialized approaches for quantum mechanics and physical systems

Training Methodology:

Mid-Training Integration:

1. Reasoning Steps: Including specific steps in mid-training to teach correct quantum mechanics reasoning
2. Physical Systems Logic: Training models to reason correctly about complex physical systems
3. RL Training Enhancement: Incorporating physics-specific reasoning into reinforcement learning loops

Comprehensive Skill Development:

• Literature Analysis: Teaching LLMs to read scientific papers and textbooks effectively
• Simulation Execution: Enabling models to run theoretical calculations and simulations
• Experimental Action: Training systems to take experimental actions and learn from results

Workflow Alignment:

The training process mirrors how physical R&D researchers operate: reading literature and internal documents, running simulations and theoretical calculations, then attempting experimental validation and learning from outcomes.

Timestamp: [29:47-31:35]

Essential Insights:

1. Room Temperature Superconductivity Impact - Discovery would fundamentally change humanity's understanding of quantum effects and the universe, with robust phase transition properties making it ideal for AI-driven research
2. Repeatable AI Scientist Strategy - Periodic Labs focuses on creating complete research loops rather than theoretical work, with plans to prove systems in one domain then expand systematically across scientific fields
3. Commercial Viability Path - Following the frontier labs model of AI programming success, Periodic targets advanced industries (space, defense, semiconductors) as co-pilots for materials and physics iteration workflows

Actionable Insights:

• Cross-disciplinary integration requires structured weekly teaching sessions and "no stupid questions" culture to unite ML scientists with physical scientists
• LLM training for science involves teaching models to reason about quantum mechanics through specialized mid-training and RL processes
• Commercial strategy leverages massive R&D budgets in industries lacking good tools for materials discovery and physics simulation

Timestamp: [24:03-31:53]

Companies & Products:

• OpenAI - Referenced as example of frontier AI lab that developed ChatGPT and established AI programming as commercially viable path
• Google DeepMind - Mentioned as frontier AI lab that has figured out training models on math and logic

Technologies & Tools:

• ChatGPT - Used as example of landmark machine learning system that demonstrated AI programming capabilities
• Reinforcement Learning (RL) - Training methodology being adapted for physics and chemistry reasoning in LLMs

Concepts & Frameworks:

• Phase Transition - Physical property that makes superconductivity robust to simulation limitations and material defects
• Mid-Training - Specialized training phase for incorporating domain-specific reasoning capabilities into language models
• Quantum Mechanics - Fundamental physics framework that LLMs need to learn for materials discovery applications

Industries Mentioned:

• Space Industry - Target market for Periodic Labs' materials discovery co-pilots
• Defense Sector - Advanced manufacturing industry dealing with materials and physics iteration
• Semiconductors - Industry requiring materials discovery and physics simulation capabilities

Timestamp: [24:03-31:53]

Hiring Philosophy and Team Diversity

Educational Requirements:

• No advanced physics/chemistry degree required - The company actively recruits from diverse academic backgrounds
• Learning curve is similar for everyone - Even top physicists have vast knowledge gaps in the interdisciplinary field
• Knowledge gap reality: The amount any expert doesn't know about physics vastly exceeds what they do know

Team Composition Strategy:

1. Simplex model approach - Team members positioned across different expertise areas:

• Pure ML/LLM specialists
• Pure experimentalists
• Pure simulation experts
• Bridge connectors - People who translate between different technical domains

2. Cross-pollination benefits - Physics experts learn new areas of chemistry, LLM researchers discover new aspects of their field through collaboration

Historical Context:

• 1800s vs. Today: Past physicists could advance across multiple frontiers simultaneously
• Modern specialization challenge: Intellectual knowledge base is now so vast that experts typically advance in only one specific field
• Discovery bottleneck: Breakthrough discoveries (like superconductors) require knowledge across chemistry, physics, synthesis, and characterization - more than any single human can master

Timestamp: [32:49-34:43]

Mission-Driven Culture and Research Focus

Primary Differentiator:

• Mission alignment is crucial - Researchers must genuinely care about accelerating scientific discovery as their primary goal
• High overlap with other AI companies - Core research skills translate well across organizations

Team Characteristics:

Mission-Driven Focus:

• Scientific discovery as northstar - Team members are unified around advancing science rather than improving consumer products
• Current team size: Approximately 30 researchers
• Alternative path clarity: Those focused on improving mega-corp products would be better suited at those companies

Research Environment:

• Interdisciplinary learning - Researchers continuously expand knowledge across physics, chemistry, and ML domains
• Collaborative knowledge building - Team structure mirrors the LLM training approach of teaching diverse concepts

Timestamp: [34:48-35:35]

Deployment Strategy for Conservative Industries

Industry Challenge:

• Technology adoption gap - Space, defense, and advanced manufacturing are mission-critical but slow to adopt new technology
• AI strategy urgency - Companies recognize technology is shifting rapidly while their work processes remain static
• Expertise loss - Industries are losing senior engineers and researchers, seeking ways to preserve institutional knowledge

Deployment Approach:

Land and Expand Strategy:

1. Problem-focused entry - Solve specific, well-scoped problems with clear evaluations rather than transforming entire operations
2. Co-development process - Work directly with companies to identify and scope critical bottlenecks
3. Capability intersection - Focus on areas where Periodic's technical capabilities align with customer's biggest problems

Implementation Philosophy:

• Gradual transformation - Avoid "day one fab line transformation" approach
• Proof of concept first - Demonstrate technology power through targeted problem-solving
• Internal champion identification - Find biggest promoters within customer organizations

Timestamp: [35:41-38:23]

Real Customer Pain Points and Solutions

Simulation Challenges:

• Training bottleneck - Companies spend significant time training employees on critical simulation tools
• Automation opportunity - Automating simulations would dramatically enable development processes
• Integration needs - Matching formats and feeding simulation results into design pipelines

Design Process Improvements:

• Workflow optimization - Streamlining the overall design process from concept to implementation
• Data integration - Treating all data together in unified platforms rather than siloed systems
• Format compatibility - Ensuring seamless data flow between different tools and processes

Fundamental Technical Challenge:

Retrieval vs. Pre-training Dilemma:

• Lightweight retrieval approach - Many companies currently rely on simple neural net retrieval over existing data
• Pre-training advantages - Encoding knowledge into model weights creates richer, deeper understanding compared to retrieval systems
• Access control complexity - Pre-training on proprietary data raises privilege and access control challenges that retrieval systems handle more easily

Timestamp: [38:24-39:57]

Essential Insights:

1. Interdisciplinary hiring approach - Periodic Labs doesn't require advanced degrees in physics/chemistry, recognizing that knowledge gaps exist for all researchers in this vast interdisciplinary field
2. Mission-driven culture differentiator - While research skills overlap with other AI companies, genuine commitment to accelerating scientific discovery is the primary hiring criterion
3. Strategic deployment methodology - Rather than transforming entire operations, Periodic focuses on solving specific, well-scoped problems to demonstrate AI capabilities in conservative industries

Actionable Insights:

• Companies can leverage "bridge connector" team members who translate between technical domains to improve cross-functional collaboration
• Traditional industries seeking AI adoption should start with targeted problem-solving rather than comprehensive transformation approaches
• The choice between retrieval systems and pre-training represents a fundamental technical decision with significant implications for data access and system capabilities

Timestamp: [32:00-39:57]

People Mentioned:

• LeBron James - Referenced in analogy about knowledge gaps between experts and novices in physics

Companies & Products:

• OpenAI - Mentioned as comparison point for researcher characteristics and hiring
• Anthropic - Referenced alongside other major AI companies for researcher comparison
• Google DeepMind - Cited as another major AI research organization for hiring comparison
• ChatGPT - Used as example of pre-training advantages over retrieval systems

Technologies & Tools:

• APIs - Computer science approach to mapping inputs, outputs, and targets for problem-solving
• Neural networks - Mentioned in context of lightweight retrieval solutions
• Retrieval systems - Discussed as current industry standard for data access
• Pre-training - Advanced approach that encodes knowledge into model weights for deeper understanding

Concepts & Frameworks:

• Simplex model - Team composition strategy positioning members across different expertise areas (ML/LLM, experimentalist, simulation)
• Land and expand strategy - Deployment methodology starting with specific problems before broader transformation
• Bridge connectors - Team members who translate between different technical domains

Timestamp: [32:00-39:57]

Mid-Training: The Bridge Between Pre-Training and Post-Training

What is Mid-Training?

Mid-training emerged as a solution to knowledge limitations in pre-trained models. While pre-training uses internet scrapes with knowledge cutoffs, and post-training focuses on reinforcement learning and supervised fine-tuning, mid-training serves a specific purpose: injecting new knowledge that doesn't exist in the original model.

The Periodic Labs Approach:

1. Take a pre-trained foundation model - Leverage existing capabilities from major AI labs
2. Continue pre-training with specialized data - Add physics and chemistry knowledge through continued training
3. Ensure distribution connectivity - Make sure different data types (simulation, experimental, semantic) work together effectively

Types of Knowledge Being Injected:

• Low-level descriptions: Crystal structures and physical object properties
• High-level semantic descriptions: Detailed synthesis procedures and methodologies
• Simulation data: Computational physics and chemistry results
• Experimental data: Real-world laboratory findings and observations

The Challenge Beyond Simple Data Mixing:

Simply combining different data distributions (A, B, and C) doesn't guarantee generalization. The goal is to ensure that including new datasets actually improves performance across all datasets, creating true expertise in physics and chemistry where models were previously deficient.

Timestamp: [41:06-42:55]

Building on Multiple Waves of Progress

Foundation Model Improvements:

Periodic Labs benefits directly from improving base LLMs by taking pre-trained models and mid-training them with high compute resources. As foundation models get better, Periodic's specialized physics and chemistry models automatically inherit those improvements.

Physical Simulation Tool Advances:

• Open-source simulation tools are continuously improving
• New machine learning methods for predicting material properties are being released
• Physics and chemistry fields have seen such significant ML impact that continued improvements are expected

Neural Networks as Tools:

Traditional agent tools included browsers and Python environments, but increasingly other neural networks serve as tools for agents. This creates acceleration opportunities:

1. Physics code complexity: Most physics code isn't deeply complex competitive programming—it's often "hacky scripts"
2. Specialized neural networks: Can handle specific tasks that would be difficult to replicate from scratch
3. Tool composition: Agents can rely on the best systems for specific capabilities rather than rebuilding everything

The Composability Advantage:

Rather than starting from scratch, Periodic Labs can draft off progress happening across multiple domains—foundation models, simulation tools, and specialized neural networks—creating a composable approach to scientific AI.

Timestamp: [44:06-45:16]

Deep Academic-Industry Synergy in Scientific AI

Essential Academic Contributions:

• Simulation tooling development: Much of the core simulation software used by Periodic Labs originated in academic research, particularly from European institutions
• Novel synthesis methods: Universities continue developing new approaches to material creation and chemical processes
• Complex technical foundations: Academic teams create sophisticated Fortran code and simulation tools that industry teams often lack expertise to develop efficiently

The Historical Pattern:

A clear division of labor has emerged where fundamental tools are developed in academia and then adopted by industry labs. Recent examples include large-scale simulations being executed at Microsoft, DeepMind, and Meta using tools originally created in university settings.

Academic Input on AI Training:

Universities provide crucial guidance on what skills and analysis capabilities should be built into scientific AI models:

1. Task Definition: Helping identify important scientific analysis tasks that weren't direct goals for general model training teams
2. Skill Decomposition: Breaking down complex analyses into smaller primitives that can be effectively trained
3. Reasoning Strategy Correction: Academic physicists can identify when AI models use incorrect reasoning approaches

The "Thinking Strategy" Problem:

Academic collaboration revealed a critical insight: AI models often use fundamentally wrong reasoning strategies. A physicist reviewing Periodic's model outputs noted that effective scientific thinking requires "thinking in terms of symmetries" and other high-level conceptual frameworks that only domain experts understand.

Reinforcement Learning Integration:

Academic partnerships help design reinforcement learning environments that reward effective scientific reasoning strategies used by premier scientists, creating a feedback loop between academic expertise and AI training methodologies.

Timestamp: [45:21-47:45]

Essential Insights:

1. Mid-training bridges knowledge gaps - Periodic Labs uses mid-training to inject specialized physics and chemistry knowledge into foundation models, going beyond simple retrieval to proper training integration
2. Composable progress acceleration - The company leverages simultaneous advances in foundation models, simulation tools, and neural network architectures rather than building everything from scratch
3. Academic-industry synergy is critical - Universities provide essential simulation tools, reasoning strategies, and domain expertise that industry labs cannot efficiently develop independently

Actionable Insights:

• Distribution connectivity matters: Simply mixing different data types doesn't guarantee generalization—the goal is ensuring new datasets improve performance across all datasets
• Neural networks as tools: Modern AI agents increasingly use other neural networks as specialized tools, enabling composition rather than replication of capabilities
• Reasoning strategy validation: Domain experts can identify when AI models use fundamentally incorrect thinking approaches, informing better reinforcement learning reward systems

Timestamp: [40:04-47:57]

Companies & Organizations:

• OpenAI - Referenced for ChatGPT development and early supervised training approaches
• Microsoft - Mentioned as conducting large-scale physics simulations using academic tools
• Google DeepMind - Cited for large-scale simulation work and materials science research
• Meta - Referenced for conducting large-scale simulations in scientific computing
• Stanford Physics Lab - Mentioned as location where AI model evaluations on scientific analysis were conducted

Technologies & Tools:

• Fortran - Programming language commonly used in academic physics simulation tools
• Python - Traditional tool environment for AI agents alongside browsers
• Crystal structures - Low-level physical descriptions used in materials science training data

Concepts & Frameworks:

• Mid-training - Training methodology that injects new knowledge into pre-trained models through continued pre-training
• Pre-training - Initial training phase using large internet scrapes with knowledge cutoffs
• Post-training - Training phase focused on reinforcement learning and supervised fine-tuning
• Distribution connectivity - Machine learning concept ensuring different data types work together effectively
• Symmetries - High-level physics reasoning strategy for effective scientific analysis
• Neural networks as tools - Modern approach where AI agents use other neural networks as specialized capabilities

Timestamp: [40:04-47:57]

Advanced AI Architecture for Scientific Discovery

Periodic Labs employs a sophisticated hybrid approach that leverages the complementary strengths of different AI technologies:

Core Integration Strategy:

1. Language Model Foundation - Handles synthesis recipes, scientific literature processing, and high-level reasoning
2. Geometric Reasoning Tools - Specialized systems for atomic structure representation and spatial relationships
3. Modular Architecture - Language models can call geometric tools as needed for specific tasks

Technical Implementation:

• Equivariant Graph Neural Networks - Maintain geometric properties while processing molecular structures
• Diffusion Models - Built-in geometric tools that naturally handle spatial relationships
• Hybrid Processing - Language aspects for synthesis recipes combined with geometric aspects for atomic design

Practical Applications:

• Synthesis Recipe Generation - Language models excel at processing and generating chemical procedures
• Atomic Structure Design - Geometric tools provide precise spatial representation of atoms
• General Geometry Design - Comprehensive approach to representing complex molecular structures

This architecture allows Periodic Labs to tackle both the linguistic complexity of scientific knowledge and the precise geometric requirements of materials science.

Timestamp: [48:02-48:28]

Two-Pronged Academic Engagement Initiative

Periodic Labs recognizes the critical importance of maintaining strong ties with academic research and has developed a comprehensive strategy:

Advisory Board Initiative:

• Expert Diversity - Spanning superconductivity, solid-state chemistry, and physics
• Research Alignment - Ensuring connection with long-term research directions
• Government Funding Synergy - Tight coupling with groups receiving important government funding

Key Advisory Board Members:

1. Superconductivity Expertise:

• ZX Chan (Stanford) - Experimental superconductivity research
• Steve Kelsson - Theoretical superconductivity work

2. Specialized Research Areas:

• Mercury Canadas (Northwestern University) - Sensors expertise
• Chris Walverton - High throughput DFT (Density Functional Theory)

Grant Program Initiative:

• Academic Enablement - Supporting research best suited for academic environments
• Community Focus - Work that benefits the broader scientific community
• Strategic Areas - LLMs, agents in synthesis, materials discovery, and physics modeling
• Proposal-Based - Accepting grant proposals from qualified academic researchers

This dual approach ensures Periodic Labs remains connected to cutting-edge academic research while supporting work that advances the entire field.

Timestamp: [48:28-49:47]

Essential Characteristics for Scientific Innovation

Periodic Labs seeks candidates who embody a unique combination of intellectual curiosity, practical skills, and world-class expertise:

Core Personal Qualities:

1. Deep Curiosity - Genuine desire to understand machine learning and science at fundamental levels
2. Reality-Focused - Strong drive to make contact with reality and advance scientific knowledge
3. Pragmatic Approach - Solution-oriented mindset with careful, methodical processes
4. Goal Achievement - Ability to reach objectives quickly and efficiently

Professional Excellence:

• World-Class Expertise - Exceptional ability along at least one key dimension
• Innovation Capacity - Ability to bring new approaches to creative ML systems
• Technical Versatility - Skills across multiple research pillars

Research Areas of Interest:

Machine Learning Specialists:

• Creative ML system development
• Novel approaches to state-of-the-art models
• Integration of AI with scientific discovery

Experimental Scientists:

• Laboratory research and validation
• Real-world testing of theoretical models
• Bridge between simulation and reality

Simulation Experts:

• Advanced computational modeling
• Robust and reliable simulation development
• Experimental validation integration

Critical Mindset:

Sense of Urgency - Candidates must feel driven to improve physical systems and discover materials now, not in 10 years. The team wants LLMs improving science as soon as possible, requiring immediate action on superconductivity innovation and materials discovery.

Timestamp: [49:47-51:21]

Essential Insights:

1. Hybrid AI Architecture - Periodic Labs combines language models with geometric reasoning tools, using equivariant graph neural networks and diffusion models for comprehensive scientific discovery
2. Academic Partnership Strategy - Two-pronged approach including expert advisory board and grant program to maintain strong ties with university research and government-funded projects
3. Talent Acquisition Focus - Seeking world-class researchers with deep curiosity, pragmatic problem-solving skills, and urgent drive to accelerate scientific discovery through AI

Actionable Insights:

• Technical teams can leverage modular AI architectures that combine linguistic and geometric processing for complex scientific problems
• Startups can build credibility and research depth through strategic academic partnerships and advisory boards
• Organizations pursuing breakthrough technologies should prioritize candidates with both technical excellence and sense of urgency for real-world impact

Timestamp: [48:02-51:27]

People Mentioned:

• Zhi-Xun Shen - Stanford experimental superconductivity researcher serving on Periodic Labs' advisory board
• Steve Kelsson - Theoretical superconductivity expert contributing to advisory board
• Mercury Canadas - Northwestern University sensors expertise specialist on advisory board
• Chris Walverton - High throughput DFT (Density Functional Theory) researcher advising Periodic Labs

Technologies & Tools:

• Equivariant Graph Neural Networks - Geometric reasoning tools that maintain spatial properties for molecular structure processing
• Diffusion Models - AI architectures with built-in geometric capabilities for spatial relationship handling
• DFT (Density Functional Theory) - Computational quantum mechanical modeling method for electronic structure calculations

Concepts & Frameworks:

• Hybrid AI Architecture - Combining language models with specialized geometric reasoning tools for scientific applications
• High Throughput DFT - Automated computational approach for rapid materials property calculations
• Modular AI Design - Architecture allowing language models to call specialized tools as needed

Timestamp: [48:02-51:27]