Conference Presentation, Keynote
Keynote by Karim Beguir, Co-Founder and CEO InstaDeep | RAISE Summit 2024 | Paris
- The "Triple Exponential" Growth of AI
- AI progress is driven by three concurrent exponential growth vectors:
- Data: Exponential increase in volume, exemplified by genomic sequencing costs dropping from billions in the early 2000s to approximately $100 today.
- Compute: Raw compute power available for ML tasks is increasing at 10x every six months (100x annually).
- Model Efficiency: Efficiency gains are equally rapid, with visual AI tasks requiring half the compute every 16 months; since 2012, efficiency has improved 45x.
- AI progress is driven by three concurrent exponential growth vectors:
- The Self-Accelerating Paradigm Shift
- AI has entered a new phase where it accelerates its own development, a qualitative change occurring around the ChatGPT era.
- Data Acceleration: Advanced models can auto-label data (e.g., GPT-4) with higher accuracy, speed, and lower cost than human labor platforms like Amazon Mechanical Turk.
- Training Feedback Loops: Transition from Reinforcement Learning from Human Feedback (RLHF) to AI Feedback (ARL) allows models to train younger models or self-improve iteratively without human intervention.
- Hardware Co-Design: AI is now integral to chip design; InstaDeep's
dpcb.aiand companies like Google (TPU v4/v5) and Nvidia use AI to optimize printed circuit board routing. - Software Development: Coding assistants and autonomous systems have doubled the efficiency of delivering high-quality source code.
- Evolution from LLMs to Collaborative Agents
- Current Large Language Models (LLMs) function as "Oracles" for Q&A, but the future paradigm involves "Smart Agents" acting in the physical and digital world.
- Multimodal Action: Systems will integrate speech, vision, and motor commands to execute complex physical tasks, demonstrated by Figure.ai's humanoid robot retrieving objects and explaining actions.
- Multi-Agent Collaboration: Complex problems will be solved by collections of specialized agents (digital or physical) working together, rather than a single monolithic system.
- Industrial Application: InstaDeep and Deutsche Bahn are applying this to schedule 40,000 trains daily, using multiple agents to generate solutions verified by expert security systems.
- Strategy for Future-Proofing Companies
- Sustainable competitive advantage requires integrating three specific corporate assets:
- Differentiated Data: Aggressive acquisition of unique, task-specific data (e.g., Tesla's fleet generating data from cameras).
- Domain Expertise: Human knowledge embedded in operations and culture, often held by employees.
- Simulation & Testing: Capabilities to test in virtual environments and the real world, including over-the-air update mechanisms.
- The Virtuous Loop: Combining these assets to train agents that act, gather data, and improve the model creates a self-reinforcing cycle of competitiveness.
- Sustainable competitive advantage requires integrating three specific corporate assets:
- Case Study: Tesla's Competitive Moat
- Data Scale: Tesla operates six million vehicles with eight high-resolution cameras each, capturing data continuously.
- Metrics: Tesla's Full Self-Driving system has driven 1 billion miles (1.6 billion kilometers), vastly outpacing competitors who have driven in the dozens of millions.
- Mechanism: Human drivers intervening in control issues provide data points for retraining, creating a loop where the fleet teaches the model, which is then pushed back to the fleet via over-the-air updates.
- Forward-Looking Conclusion
- Winners in the AI era will be those who leverage AI agents to magnify their inherent core strengths in data, expertise, and testing.
- The "Golden Age of LLMs" is transitioning into an era of agentic systems that execute tasks and collaborate, moving beyond passive information retrieval.
- InstaDeep focuses on industrial optimization and managing agents at scale to deliver efficiency for global partners.