Pages

Showing posts with label llm. Show all posts
Showing posts with label llm. Show all posts

Thursday, June 05, 2025

Will Scaling Large Language Models (LLMs) Lead To Artificial General Intelligence (AGI)


Here is a balanced argument presenting both sides of the ongoing debate around whether scaling Large Language Models (LLMs) will lead to Artificial General Intelligence (AGI):

Argument 1: LLMs Are Not the Path to AGI

  1. Statistical Mimicry ≠ Understanding
    LLMs are fundamentally pattern-recognition engines trained to predict the next token. They do not “understand” meaning, intentions, or goals. They simulate reasoning without possessing it, and lack grounding in real-world context, embodiment, or sensory experience—critical aspects of general intelligence.

  2. Lack of Agency and Autonomy
    LLMs do not initiate goals, pursue objectives, or act independently in the world. AGI requires agency: the ability to plan, adapt, and act toward long-term goals across environments, which LLMs are not designed to do.

  3. Catastrophic Forgetting and No Long-Term Memory
    LLMs do not learn continually or adapt dynamically post-training. Their knowledge is static, baked into weights. AGI requires lifelong learning, updating beliefs in real time, and managing long-term memory—which current LLM architectures do not support robustly.

  4. Scaling Laws Show Diminishing Returns
    While LLM performance improves with scale, there's growing evidence of diminishing returns. Bigger models are more expensive, harder to align, and less interpretable. Simply scaling does not necessarily yield fundamentally new cognitive abilities.

  5. Missing Cognitive Structures
    Human cognition involves hierarchical planning, self-reflection, causal reasoning, and abstraction—abilities that are not emergent from LLM scaling alone. Without structured models of the world, LLMs cannot reason causally or build mental models akin to humans.

Argument 2: Scaling LLMs Will Lead to AGI

  1. Emergent Capabilities with Scale
    Empirical evidence from models like GPT-4 and Gemini suggests that new abilities (e.g. multi-step reasoning, code synthesis, analogical thinking) emerge as models grow. These emergent behaviors hint at generalization capacity beyond narrow tasks.

  2. Language as a Core Substrate of Intelligence
    Human intelligence is deeply tied to language. LLMs, by mastering language at scale, begin to internalize vast swaths of human knowledge, logic, and even cultural norms—forming the foundation of general reasoning.

  3. Unified Architecture Advantage
    LLMs are general-purpose, trainable on diverse tasks without specialized wiring. This flexibility suggests that a sufficiently scaled LLM, especially when integrated with memory, tools, and embodiment, can approximate AGI behavior.

  4. Tool Use and World Interaction Bridges the Gap
    With external tools (e.g. search engines, agents, calculators, APIs) and memory systems, LLMs can compensate for their limitations. This hybrid “LLM + tools” model resembles the way humans use external aids (notebooks, computers) to enhance intelligence.

  5. Scaling Accelerates Research Feedback Loops
    As LLMs improve, they assist in code generation, scientific discovery, and AI research itself. This recursive self-improvement may catalyze rapid progress toward AGI, where LLMs design better models and architectures.

Conclusion

The disagreement hinges on whether general intelligence is emergent through scale and data, or whether it requires fundamentally new paradigms (like symbolic reasoning, embodiment, or causal models). In practice, future AGI may not be a pure LLM, but a scaled LLM as the core substrate, integrated with complementary modules—blending both arguments.





at No comments:
Labels: , , , ,

Friday, May 30, 2025

Memory Is the Real Moat—But It Should Belong to Us


Memory Is the Real Moat—But It Should Belong to Us

In the AI age, the most valuable resource isn’t data—it’s your memory. Not biological memory, of course, but the contextual breadcrumbs you've left behind across a growing constellation of LLM-powered apps. Every prompt, every reply, every fine-tuning of tone, style, and preference—this is the memory that makes an AI assistant yours. And this is becoming the most powerful moat large AI platforms have.

But herein lies the dilemma: this memory is locked inside walled gardens. ChatGPT knows your writing style. Claude remembers your schedule. Perplexity learns your research interests. But none of them talk to each other. And none of them give you full control.

A Moat for Them, a Trap for Us?

From a platform perspective, memory is a dream. It deepens engagement, raises switching costs, and feeds into a virtuous loop: the more you use the app, the better it gets, the harder it becomes to leave. But for users—especially professionals relying on AI across tasks, tools, and devices—this creates real friction.

Imagine writing part of a novel in ChatGPT, managing your tasks with an AI assistant, and analyzing documents with a third app. Each has a different slice of your memory, with no unified context. You end up re-teaching, re-uploading, and re-reminding each app what the others already know. It’s like having a dozen brilliant interns who don’t speak to each other.

The Case for Memory Portability

This is why the idea of “Plaid for memory” is so compelling. In fintech, Plaid unlocked financial data portability, enabling users to control how and where their information is used. Why can’t we do the same with AI memory?

Imagine a permissioned memory layer that sits above the AI apps—a personal data vault you control. Apps would need your consent to read from or write to your memory. You could revoke access anytime. Want to switch from ChatGPT to Claude? Your memory comes with you. Want your task app to learn from your writing habits? Grant it access. Want to share your professional context with a new assistant agent? One click.

This idea turns memory from a moat into a market. And in doing so, empowers users rather than platforms.

What Would It Take?

  • Standards for contextual data: Just like there are APIs for calendars or contacts, we’ll need standards for memory—conversations, task histories, preferences, tone, etc.

  • Encryption and privacy controls: Memory portability must be secure by default. Encryption, consent logs, and clear revocation mechanisms are a must.

  • An open protocol or foundation: Ideally, this layer should be governed by a nonprofit or consortium—not a single company—so it doesn’t become just another silo.

  • Developer incentives: AI startups should be incentivized to support memory portability. This could become a competitive differentiator.

Why This Matters

As AI becomes more ambient—woven into every device, browser, and workflow—fragmented memory will become unbearable. Users will demand interoperability. And the companies that embrace memory portability may not just win trust—they may unlock a new layer of innovation.

Today, we’re still in the early “memory hoarding” phase of LLM platforms. But history favors openness. The companies that gave users ownership—over code, identity, or data—sparked ecosystems, not silos.

Whoever builds the “Plaid for memory” will unlock a better AI future. One where the most valuable thing—the story of you—is finally yours to own.


at No comments:
Labels: , ,
Subscribe to: Posts (Atom)