Decoding the AI Horizon: Frontier vs. Open Models, Agents, and Robotics

I recently spoke with former colleagues or friends who work/worked at the "frontier foundation model providers" such as OpenAI Google DeepMind Anthropic 微软 . While almost all of us believe that the current LLM's "next token prediction" can pave the way to #AGI, I've realized that there's a lack of consensus on many key LLM topics within the industry.

This prompted me to document a few of non-consensus predictions, thanks to the brainstorming exercise with various experts. I must add that those predictions are not special to the cybersecurity space; a unique industry unto itself that deserve a separate discussion.

1: "Frontier" models will continue leading on performance and "Open" models will thrive on ecosystem.

• #Frontier closed models from OpenAI , Anthropic , and Google DeepMind will maintain their lead over open models by at least a year. This gap between closed and open model is likely to widen in the foreseeable future, primarily due to the the simple fact that frontier "closed" model players own immense GPU clusters for the training.
• #Open models will still command a significant market share however. Many vertical use-cases don't require the cutting-edge capabilities of frontier models, leaving ample room for open models like #LLaMA2 to thrive. Meta 's LLaMA2 will boast an unparalleled ecosystem around it and could very well emerge as the predominant player among open models. Such open models will enable many traditional enterprise incumbents to build a competitive advantage around their data without fearing sending private data to the frontier model providers.

2: #Agents are LLMs that can make decisions on what action to take and use external tools like calculators, search, or executing code. Agent is a buzzword in the research domain, but its mainstream moment will only transpire after future significant foundational performance enhancements. Without these improvements, it becomes challenging for enterprises to productize the Agent. I foresee Agent emerging as a primary technology behind disruptive enterprise products within the next 3-5 years.

Before any serious enterprise adoption, we can at least appreciate the power of the agent in the "Stanford Smallville" game, as the 英伟达 senior researcher Jim Fan said in his tweet, "It is among the most inspiring AI agent experiments in 2023. We often talk about a single LLM's emergent abilities, but multi-agent emergence could be way more complex and fascinating at scale. A population of AI can play out the evolution of an entire civilization."

3: #Multimodality

Multimedia is the next goldmine to train the LLM, but it poses it a lot of challenges to the pipeline and algorithm. No one seems to question the "whether" to go multimodality but some question the "when" and how fast we can see meaningful return. These experts believe there are other lower hanging fruit to get the LLM foundation model to the next level without multimodality. Also I heard one frontier model provider already trained the multimodal foundation model twice and clearly there is no quick win.

4. #Robotics

We all want to, but may not witness significant breakthroughs in robotics anytime soon yet. The present generative AI wave is primed to disrupt white-collar professions in the years ahead. A pivotal moment for the future #robotics industry, akin to the "ChatGPT moment", is needed to disrupt blue-collar sectors. While this is inevitable, it may still be some years away.

5. Frictions for mass LLM adoption in the enterprise: there are elements the broader #AI industry has been talking a lot but might still be underestimated today:

• Inference #latency could hinder LLM adoption as many use cases cannot tolerate many-second latency.
• The evaluation framework ain't easy and may impede LLM adoption, as I highlighted in my previous blog titled "Measurement is All You Need" too.
• AI #breakthroughs are more dependent on #organizational #culture and ethos, followed by capital investment, and surprisingly less on raw talent skillsets.

On a related note, there's a near-universal agreement that AI is set to transform numerous industries. I am deeply passionate about the implications of #GenerativeAI and #Robotics for the professional world and how students in schools today should brace themselves for these profound shifts in the forthcoming decade.