The Future of AI is Custom OS Language Models

This weekend I had the opportunity to attend the Custom LLM Hackathon hosted by the awesome folks at AGI House and Jeremy Nixon

As Jeremy mentioned this is a milestone moment in AI. And the gravity of this moment is not yet understood by many people even within the AI community. I will try to capture the reasons I believe the future of AI is custom models.

Background:

Until now we have had a few very large language models. OpenAI has led the way with ChatGPT powered by models like GPT3, 3.5, and 4. These are among the best models and continue to improve. OpenAI showed the world what the current state of transformer models were capable of. Their success with ChatGPT and CoPilot are examples of good products that will create significant value. Hence, these have the largest distribution among all AI companies.

Drawbacks of Commercial LLMs and the Opportunity of OpenSource LLMs

With great distribution comes great responsibility and the need to please the common denominator of your audience. Hense, these LLMs tend to be

• Super well mannered: Restrictions built into the LLM to prevent abusive language and tone.
• Avoid sensitive topics: Guardrails the LLMs from venturing into topics that could be sensitive to certain parts of the audience, consider politics, religion etc
• Lack vertical knowledge: The majority of enterprise data today still rests with private data stores within businesses and these are not accessible or permitted to be trained on.
• Lack vertical behavior: There are over 12,000 job titles across several of vertical domains. And each job requires nuanced behaviors.

Welcome OpenSource Models

OpenSource models allow you to break these limitations and create new use cases and behaviors that were not expected before. There are a few examples of this I saw at the hackathon.

Imagining the Future of LMs with FineTuned Language Models from OpenSource Base Models/

• UnLimited Context Windows and Token Sizes: What would you do if you had access to a language model that had unlimited context windows and output token sizes. That could create coherent works across extremely long contexts. Would you write a book? Would you write a movie ? Would you create a Truman Show where no human writes the story and it is completely AI generated? You could build such a model with OS
• UnHinged: What would you build if your LM could handle any topic. Would you work on artificial discourse on political and social topics? Open-ended research on how to solve poverty, unemployment, religion, end homelessness?
• Fun: What would you build if you could build on any topic without guardrails. Consider love, hate, lust, inspiration, desperation, anxiety, fear. What models would you build?
• Vertical Behavior: If you are a customer support rep at a bank you would behave differently than a customer support rep at a sports store. Different tonality, different mannerisms.
• Synthetic Data: Data is the most important resource today, Yet everyone seems to have the same data. What if LLMs could help you create new Synthetic data? Twig does this in its Co-Pilot for B2B CX Teams
• Non-Chat Interfaces: While the large LMs are focusing on Chat. What if you could build models where chat was not at all the medium of interaction? What if your lighting setup at home responded to the sounds, weather, and people in the room?

The possibilities seem endless and this to me is the most amazing thing about Open Source Language models.