Artificial Specific Intelligence

In the course of the past year of building Symbolic and working with multiple models behind the scenes, we've been tracking the emergence of an important trend: As we get closer to artificial general intelligence (AGI), the actual models themselves are getting more specialized. We're not the only ones noticing that the more advanced models have more pronounced relative strengths and weaknesses, and there's a consensus emerging that an artificial general intelligence will probably look something like a heterogeneous collection of specialized models with different training data and approaches, stitched together into a network.

The structure of AGI, at least from the point of view of the user, will probably look something like the following, with layers that are alternately diverse and unified:

• User interface -- Diversity: Different software experiences support different tasks and workflows.
• Programming interface -- Unity: A basic, natural-language message or list of messages, followed by some actions, then a reply from the AI.
• Agent network -- Diversity: Some module examines the user-provided query for intent, possibly reframes it or transforms it in some way to break it into discrete tasks, and dispatches each task to one or more specialized models. These models complete their tasks, then another model puts the results back together in a way that's meaningful to the user and responsive to the user's question.
• Individual models -- Unity: Every one of the individual models has the same or similar transformer architecture, just trained in different ways.
• Training data -- Diversity: Different datasets and fine-tuning approaches, when combined with the same basic architecture, yield different models with varied strengths and weaknesses.

I could probably continue adding layers to the above, but you get the point: There's an interchange between unity and diversity at different levels of the AGI stack, where larger, unified parts are composed of smaller, diverse components, which themselves share an internal unity.

The kind of fractal pattern I'm describing, where as you zoom in at different levels you see the same patterns repeated, is also characteristic of life and intelligence in the natural world. Nature both specializes and re-uses. Building blocks like amino acids are combined into specialized structures like cells, and cells themselves are specialized into families and types that are shared across a diverse array of more complex organisms, and these organisms share common features like fins, wings, arms, and the like.

What this means for software

All of this talk of fractals and AGI may sound abstract, but it has concrete implications for how we build with AI at Symbolic, and for how our users will engage with our platform.

Consider two current state-of-the-art models that we're experimenting with right now: OpenAI's GPT-o1 and Anthropic's Claude Sonnet 3.5. Both of these models are very strong, but for different types of work.

In our internal experiments, Claude produces better writing from a style and tone perspective. We've found that Claude just "sounds" better than any of OpenAI's models when we give it no prompting at all about writing style -- the language is more natural and less hyperbolic. Claude also responds better to our attempts to coax specific styles from it by showing it examples to imitate.

But the OpenAI models, especially o1, are stronger when it comes to making connections between concepts, or to seemingly understanding some of the underlying ideas it's working with from an input project or collection.

It's like Claude is the artist who can captivate an audience by presenting basic, universal concepts in an artful way, and GPT-o1 is the nerdy enthusiast with a deep grasp of many specifics who is slightly prone to in-artful phrasing and going down rabbit holes. Individually, these two models are still very limited, but when combined in the right way, the potential for economically valuable discovery and expression is vast.

This type of model mixing and matching is where Symbolic's content production platform truly shines. Symbolic seamlessly integrates an increasingly diverse suite of cutting-edge models from different providers, allowing users to work in a simple, workflow-specific interface, without having to worry about the underlying technology or the latest LLM innovations.

Symbolic might use GPT-o1 for tasks in the Research screen – such as summarization, data extraction, or generating follow-up questions -- and Claude for generating text in the Write screen. And as new models come out, Symbolic's team of seasoned media and tech professionals will expertly integrate them into the platform where they fit best.?

The result: Journalists, communications professionals and other information workers using Symbolic can confidently leverage each LLM’s strengths so they can focus on their work.?

— Jon Stokes

Symbolic Co-Founder & CTO