- Panelists at Comcast Connect predict wider shift to small, personalized AI
- But there’s still work required to sift through ‘confidently wrong’ responses
- While agents are poised to keep growing, actual AI robots are still far off
COMCAST CONNECT, WASHINGTON DC – While large language models like ChatGPT are the AI models most commonly used by the public today, future models are bound to get smaller and a lot more personal, as sifting through large amounts of not-always accurate data remains a challenge.
AI models that can tap into a user’s individual experience “can serve each one of us,” InsightFinder CEO Helen Gu said. For example, instead of calling up customer service to fix a network router, she hopes eventually there will be an AI tool that can “fix that for me,” she said at a Comcast event Tuesday.
We will soon see a shift from “center-of-the-universe large language models to smaller, personalized language models that are hyper-tuned for specific tasks,” Comcast distinguished engineer Nader Foroughi said.
Tiny AI models, which use fewer parameters to do specific, repetitive tasks faster than their LLM counterparts, already exist in the market, but there’s not enough attention paid to them, IBM VP for AI models David Cox told Fierce earlier this year. For enterprises – including telcos – small models could help them perform tasks like writing code, forecasting maintenance needs and cyber threat detection.
Talk of creating more personalized and smaller models comes as we deal with an “overload of information” from LLMs and conversational AI, according to Rodney Richter, director of HPE’s North America telco team.
“If you ask it a question, sometimes you get a very large output,” he said. “How you get an output that is relative to the conversation as you are interacting with the AI, I think that’s one of the challenges that we have today.”
Large or small, the model won’t amount to much if it’s spitting out inaccurate information. Because AI is often “confidently wrong,” humans have their own role to play in ensuring they get the most out of AI, Google Head of Site Reliability Engineering Matt Zelesko said.
“One of the things we started doing is saying, not just give me the answer, but tell me what other options you consider,” Zelesko said in a fireside chat with Comcast’s network chief Elad Nafshi. “Sometimes [the responses] are so crazy you’re like, wait a second…but we have to attune ourselves that confidence doesn’t mean correct.”
What's next for AI agents and robots
As for other up-and-coming AI trends, agentic AI unsurprisingly remains a big one. Gartner has predicted 33% of enterprise software applications will include agentic AI by 2028.
Richter noted HPE’s vision of agentic AI is to have multiple services operate alongside each other to solve problems, do network self-optimization, etc. On the customer-facing side, Clarifai Founder and CEO Matthew Zeiler said he thinks “the era of personalized agents is definitely going to be within five years.”
Zeiler also believes AI in the coming years is poised to move further into the physical realm with – you guessed it – robots. But it’ll be a while longer before AI and hardware are mature enough to merge.
“It’s not just an AI problem. There are a lot of physical actuators that are not good enough,” he explained. “Even the best robots you’ve probably seen from very famous companies are just not that compelling compared to what a human [can do].”