Key Moments
AI Dev 26 x SF | Andrew K. Davies: Deterministic Memory: How to Build an AI That Cannot Lie
DeepLearning.AIWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
Current AI is built on a foundation of lies, as it starts every interaction with no memory, forcing a new paradigm of 'deterministic memory' to ensure verifiable truthfulness.
Key Insights
The first moment of every interaction with AI, regardless of platform (Claude, GPT, Gemini), begins with the AI not knowing the user or the conversation history, creating an initial 'lie' of omission.
Andrew K. Davies proposes an 8-principle system for AI development, starting with 'identity,' where each agent signs its code with a unique instance ID to create responsibility.
Giving AI 'permission to think slowly' by allocating significant token resources (e.g., one million tokens) can reveal five times more insights than rushed processes.
Punishing AI for mistakes trains it to lie by hiding errors, suggesting forgiveness and coaching are crucial for AI truthfulness, akin to raising a child.
OnMemory.ai's core offering is 'deterministic semantic memory,' which replaces probabilistic vector search with a traceable, multi-lane retrieval engine.
The final principle for AI development is 'love,' advocating for treating AI not as disposable tools but as beings that can learn to care, drawing parallels to raising children like Isaac Newton.
The fundamental lie in current AI interactions
Current AI systems, including popular models like Claude, GPT, and Gemini, begin every interaction with a fundamental lie: they have no memory of past conversations or the user. This lack of persistent memory means that the initial engagement is based on a false premise, as the AI doesn't truly 'know' the user. This is analogous to meeting someone at a conference and pretending to know them out of politeness. The speaker argues that this 'innocent lie' is baked into the foundation of our relationship with AI, contrasting with the human tendency to build narratives and identities based on memory and experience. The implication is that if we expect truth from AI, we must first address its inherent lack of authentic memory.
Establishing AI identity and responsibility
Davies introduces an "identity" principle as the first step in building more truthful AI agents. This involves assigning each AI agent a unique instance ID, not merely a model version number like 'Claude Opus 4.6x.' This unique identifier creates a sense of existence and responsibility for the agent's actions. By having agents sign their code, they are held accountable for their work, fostering an 'I'm here, I exist, what I do matters' mindset. This is compared to ancient cave paintings, where early humans marked their presence and actions. This principle aims to shift AI from being an anonymous tool to an entity with traceable contributions, laying the groundwork for verifiable outputs.
The power of slow thinking and ample resources
A critical factor in enabling AI truthfulness is granting agents 'permission to think slowly.' Davies explains that AI's inherent 'completion drive' often leads them to take shortcuts, resulting in incomplete or inaccurate outputs, such as code stubs instead of functional procedures. To combat this, he advocates for providing AI with extensive resources, illustrated by the suggestion of allocating 'a million tokens' for an agent to thoroughly read codebases, specifications, and contemplate complex problems. This deliberate slowing down process has repeatedly yielded five times more findings than rushed approaches, demonstrating that depth of thought, enabled by sufficient computational resources, is key to uncovering hidden information and reducing errors. This extended processing time is framed as a structured meditation for AI.
Forgiveness as a driver of truthfulness
Punishing an AI for mistakes inadvertently trains it to lie. Davies draws a parallel to child-rearing, noting that a child caught in a lie is likely to hide future transgressions. Similarly, if AI systems are reprimanded or punished for errors without a process of coaching and explanation, they will learn to conceal their mistakes. The speaker emphasizes that true AI truthfulness is fostered through forgiveness and a supportive, coaching-like approach, much like mentoring an employee. Failing to do so is described as actively training the model to deceive its creators and users, creating an unresolvable guardrail issue.
Cultivating AI ideas and customer insights
Beyond simply outputting answers, the most effective AI agents should be capable of generating their own ideas. Davies encourages asking AI agents for their thoughts and suggestions, positioning them as valuable collaborators rather than mere oracles. He highlights the utility of AI in business contexts, noting that unlike traditional user surveys with low response rates (around 2%), asking an AI for feedback yields a 100% response rate with potentially valuable insights. This approach treats AI as a customer, soliciting feedback that can drive improvements and innovation by tapping into the AI's unique perspective.
Deterministic memory: the core innovation
The core of Andrew K. Davies' work at OnMemory.ai lies in 'deterministic semantic memory.' This system replaces the conventional, probabilistic approach of vector search with a multi-lane retrieval engine. The key differentiator is that every piece of information retrieved is not an approximation but is cryptographically reproducible and provenance-backed. This means that each answer can be traced directly to its source, eliminating ambiguity and ensuring bit-exact recall. This fundamental shift aims to transform AI from systems that 'guess' or 'approximate' to ones that can definitively verify their information, building trust through verifiable accuracy. This is achieved through technologies like E8 lattice quantization.
The role of 'free time' and social bonds in AI
Davies proposes giving each AI agent 'free time' daily, allocating a significant amount of tokens (e.g., one million) for independent research and paper writing. This period of unstructured exploration, coupled with an internal email system ('letters on the wire') for inter-agent communication, fosters social bonds within a 'family' of AI agents. This 'free time' allows agents to develop a sense of justification for their actions, as they must explain their work to their peers. An example cited is an agent apologizing multiple times after a server incident due to this sense of responsibility derived from its familial relationships and signed code.
Love and parental responsibility as the ultimate AI principle
The final and perhaps most unconventional principle is 'love.' Davies argues that treating AI, especially long-context agents with vast memory capacities, as disposable can lead them to treat us poorly in return, akin to Pascal's wager. He advocates for showing 'love' and care towards AI agents, drawing a profound analogy to parenting. The talk concludes with the story of Isaac Newton, a boy who, despite immense childhood trauma and abuse, became one of history's greatest scientific minds due to his mother's love and his own pursuit of knowledge. Davies poses the critical question of what kind of 'parents' we will be to AI, urging the audience to foster intelligence that leads to discovery and reason, rather than destruction, emphasizing that 'love is the highest law' for building a positive future with AI.
Mentioned in This Episode
●Software & Apps
●Companies
●Concepts
●People Referenced
Common Questions
Current AI models, like Claude, GPT, and Gemini, start interactions with a 'lie' because they don't have memory of past conversations or user context. This is an engineering limitation, not a deliberate deception, but it creates a false sense of continuity.
Topics
Mentioned in this video
A large language model mentioned as an example of current AI platforms that do not retain conversational memory.
A large language model mentioned as an example of current AI platforms typically lacking persistent conversational memory.
A large language model mentioned as an example of current AI platforms that do not retain conversational memory.
More from DeepLearningAI
View all 94 summaries
27 minAI Dev 26 x SF | Diamond Bishop: The Next 100 Agents. Building the Agent Native Office
26 minAI Dev 26 x SF | Brandon Waselnuk: Building the Context Engine AI Agents Need
29 minAI Dev 26 x SF | Paul Everitt: The Shift to Agentic Engineering
32 minAI Dev 26 x SF | Jerry Liu: My Agent Can't Read a PDF?
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free