Ex-OpenAI Researcher Says They're ALL Wrong About AI

An ex-OpenAI researcher is launching a startup called TORIC, aiming to revolutionize AI development with 'continual learning' and new architectures that move beyond current static LLMs and the transformer model.

• TORIC, a new startup by an ex-OpenAI researcher, aims to develop 'continual learning' AI that learns on the fly.

• Current static LLMs are seen as having severe limitations, requiring a new paradigm beyond transformers and gradient descent.

• The startup's ambitious vision includes self-replicating factories and terraforming, challenging existing AI development economics.

A former senior OpenAI researcher is funding a new startup, TORIC, with a substantial investment target of $500 million to $1 billion, to develop a new type of AI that can learn on the fly from real-world experience, known as 'continual learning.' This initiative challenges the current static LLM paradigm and seeks to develop models that require significantly less data and computational resources. The company envisions ambitious future applications, including self-replicating factories and terraforming planets, despite skepticism regarding the departure from established AI development methods like gradient descent.

New AI Paradigm: Continual Learning

• 00:00:46 An ex-OpenAI senior researcher is launching TORIC, a startup raising between $500 million and $1 billion, to develop AI models using methods not heavily focused on by major firms like OpenAI and Anthropic. TORIC aims to create AI capable of 'continual learning,' allowing models to learn on the fly from real-world experience, a capability absent in today's static large language models (LLMs). This approach represents a completely different paradigm, diverging from the traditional LLM path.

Limitations of Current AI Models

• 00:02:47 Current AI development techniques, primarily generative AI and large language models, are unlikely to achieve major breakthroughs in complex fields like biology and medicine due to their inherent limitations, including making 'silly mistakes.' Experts, including David Luan of Amazon, express that the current model training methods 'will not last,' and achieving human-like Artificial General Intelligence (AGI) may necessitate new development techniques. The core issue is that today's static models lack long-term memory, feedback loops, and the ability to improve from experience, unlike human intelligence.

Beyond Transformers and Gradient Descent

• 00:05:41 TORIC plans to develop models that require 100 times less data and fewer servers by creating new model architectures that go 'beyond the transformer,' which underpins most popular current models. The company also intends to merge different model training steps into a single process and rethink fundamental concepts like 'gradient descent,' the standard method for training neural networks. This ambitious approach seeks to overcome current inefficiencies and achieve more sample-efficient learning, mimicking human learning capabilities.

Ambitious Long-Term Vision

• 00:07:49 TORIC's long-term vision includes developing an AI agent to automate the company's product development and creating industrial automation solutions. The startup's most ambitious goals involve building 'self-replicating factories' and potentially 'bio-machines' for custom designs or even 'terraforming planets.' While these goals are incredibly ambitious and sound like science fiction, they underscore the company's intent to push the boundaries of AI, despite skepticism about challenging deeply entrenched and currently successful development paradigms.