newsfilter.io
Interview, Fireside Chat, Conference Presentation

Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin

  • Core Philosophy: Ngram operates on the premise that "models are always training," shifting focus from pre-training or post-training bottlenecks to "memory" and "continual learning" as the primary drivers for making models useful in evolving contexts.
  • The Problem with Context Engineering: Current reliance on externalized memory (RAG, context windows, and tool use) faces two critical limitations:
    • Scale: Individual and collective token generation is projected to reach tens of millions of tokens daily, making the storage, search, and re-reading of raw context prohibitively expensive and confusing for models.
    • Cognitive Limits: Models cannot easily form abstract associations or "intuitions" based solely on retrieving static documents; true understanding requires internalizing knowledge into weights.
  • Ngram's Architecture: The platform trains custom models per workspace (e.g., Notion, Microsoft, Harvey partners) to deeply understand specific team contexts:
    • Technique: Utilizes adapter fine-tuning (LoRA, prefixes, sparse architectures) and diverse training signals (supervised fine-tuning, RL, on-policy distillation) to bake knowledge directly into model weights.
    • Efficiency Trade-off: Accepts higher upfront compute costs to learn a company's "way of doing things," resulting in inference that consumes 100x fewer tokens (e.g., answering in 100 tokens vs. 100,000) by removing the need to re-read documents or write complex system prompts.
    • Access Requirements: While the approach works on any transformer model, it requires white-box access to weights to perform fine-tuning; open-source models are preferred for ease of implementation, though partnerships with closed-source labs are possible.
  • Memorization vs. Learning: The founders argue that disentangling fact memorization from skill learning is a "false dichotomy" and a lossy process for intelligence:
    • Models require internalized facts to compose complex concepts and perform next-step reasoning; stripping facts entirely results in unnatural, non-functional models.
    • The critical distinction is not what to memorize (facts vs. concepts) but what is important to retain versus what should be externalized, mirroring human biological memory's lossy compression.
  • Strategic Vision & Market Position:
    • Differentiation from Frontier Labs: While major labs (P0 goal: AGI, generic models) focus on scale, bigger models, and pre-training, Ngram focuses exclusively on the "remainder" problem: enabling individual and organizational models to learn private, evolving, and conflicting data.
    • Research-Product Loop: Proposes a new infrastructure where research and product development are integrated, as user feedback in real-world workspaces becomes the direct training signal for the model.
    • Future State (5-10 Years): Envisions a world where every individual and team owns a personalized, continually learning model that acts as a "neural interface" to their data plane, creating "neural memories" of the file system rather than just retrieving files.
  • Neuroscience & Cognitive Science Insights:
    • Biological Inspiration: Draws parallels between dream-like internal simulation (experimenting with affordances in a "sandbox") and how models should retreat from interaction to practice and internalize new skills.
    • Modality Bias Theory: A "crackpot" hypothesis suggests language dominates AI progress over vision in current architectures because computer systems are electronic (neutralizing the biological bit-rate advantage of vision in the human brain), making language the natural medium for high-dimensional abstract reasoning in LLMs.
  • Immediate Pain Points:
    • Inference Costs: High token consumption from repetitive queries and long context retrieval is the primary short-term driver for adopting weight-based continual learning.
    • Intuition Loss: Retrieval systems fail to capture the "knowing where to look" capability; internalized weights allow models to make abstract associations (e.g., linking a team member's research to a current task) without explicit search.
  • Forward-Looking Statements:
    • The "Intern" Prototype: Anticipates a near-term breakthrough where an AI "intern" is hired, taught company specifics over time, and demonstrably improves its performance week-over-week, surpassing current context-engineering limits.
    • Token Wallets: Speculates on the emergence of personal "memory wallets" allowing individuals to carry skills and learned behaviors across different workspaces, separated by strict IP and ethical boundaries (NDAs).
    • Scaling Compute: Believes the "bitter lesson" dictates that the path forward involves burning more compute to train on new, private contexts rather than optimizing for sub-quadratic memory architectures that sacrifice accuracy.