It’s been nearly six years since GPT’s release, and in just a few days, ChatGPT will celebrate its second birthday. So far, large language models (LLMs) have been progressing effectively. They’re factual, quick, and resourceful. One could even say they are nearing perfection in retrieving information. But what about the ‘delivery vehicle’?
In a recent interview with Reid Hoffman, Microsoft AI CEO Mustafa Suleyman said, “AI researchers in general tend to neglect the importance of the delivery vehicle for the information.”
Suleyman stressed that consumers often value the tone and emotional intelligence of these models and the way they manage to reflect users’ unique language styles more than simply providing an objective, encyclopedic regurgitation of Wikipedia.
Considering how the ability to understand and respond to human emotions could become a critical differentiator, Suleyman predicted that AI companies would now “wrestle” with each other based on the emotional intelligence of their frontier models. He is not wrong.
Intelligent, But Not So Artificial
One of OpenAI’s key focus areas this year, following the launch of their Advanced Voice feature in GPT 4o, was to integrate a human-like voice conversation tool. Similarly, when Google’s NotebookLM introduced a refined text-to-podcast tool, ‘Deep Dive’, it didn’t take long for the AI community to embrace the no-gimmick tool.
Computer scientist Andrej Karpathy showed his appreciation for the tool, going as far as releasing an entire podcast series of 10 episodes using NotebookLM. “It’s possible that NotebookLM podcast episode generation is touching on a whole new territory of highly compelling LLM product formats. Feels reminiscent of ChatGPT. Maybe I’m overreacting,” he said.
Over the last ~2 hours I curated a new Podcast of 10 episodes called "Histories of Mysteries". Find it up on Spotify here:https://t.co/3tEL41VqeX
— Andrej Karpathy (@karpathy) October 2, 2024
10 episodes of this season are:
Ep 1: The Lost City of Atlantis
Ep 2: Baghdad battery
Ep 3: The Roanoke Colony
Ep 4: The Antikythera… pic.twitter.com/lztKPBaKQW
It isn’t just the industry giants in AI. Three months into the year, Hume AI, dubbed an ‘AI with emotional intelligence’, secured $50 million in funding in their Series B round led by EQT ventures. In September, they released their newest EVI 2 model, which adapts to user preferences through specialised emotional intelligence training.
Earlier this year, researchers explored the emotional intelligence of LLMs. EmoBench, a popular benchmark, assessed such capabilities. The results showed that OpenAI’s GPT 4 was closest to humans in terms of “emotional understanding and emotional application”. The models evaluated, however, are a thing of the past today.
Recently, a research measured the “expressivity” of an LLM using a Python library. The researchers also conducted an experiment involving generating poems based on emotions, including feelings of sorry, joy and remorse, in poetic style. While LLMs performed satisfactorily, it was shown that there was confusion in expressing emotions with similar meanings.
“All GPT models would often express approval when prompted to express disapproval. This was a significant instance where two emotions of conflicting meanings were frequently misinterpreted,” they said.
When these LLMs were tasked with generating poems in the style of 34 different poets, GPT 4o showed the highest expressivity rate. However, the models displayed confusion when tasked with recognising female poets, possibly indicating gender bias to some extent.
Expressivity gradually declined during regular conversations. That said, Llama 3 performed the best despite its limitations. It should be noted that these LLMs perform better when additional context is provided regarding the topic, profession, or role.
“For profession signals, LLMs demonstrated a consistent and increasing level of expressivity. Conversely, for emotion signals, the expressivity of LLMs was more variable, with accuracy fluctuating as the models adapted and changed their responses based on the evolving emotional context,” added the researchers.
Scale With Caution
Anthropic views emotional quotient as an important factor in enhancing Claude. Amanda Askell, philosopher and member of technical staff at Anthropic, in an interview with Lex Fridman, said “My main thought with it has always been trying to get Claude to behave the way you would ideally want anyone to behave if they were in Claude’s position.”
“So imagine I take someone and they know that they’re going to be talking with potentially millions of people so that what they’re saying can have a huge impact and you want them to behave well in this really rich sense,” said Askell.
With newer models, Anthropic is striving to help Claude respond with nuanced emotions and expressions. This involves shaping the model to understand when to care, when to behave humorously, when to respect an opinion, and when to determine the level of autonomy.
She also addressed the problem of sycophancy in LLMs, where they tend to correct their output even when they are right, just to obey what that human input says. “If Claude is really confident that that’s not true, Claude should be like, ‘I don’t think so. Maybe you have more up-to-date information’,” Askell added.
Askell also mentioned she wants to improve Claude’s ability to ask relevant follow-up questions in a conversation. Overall, Anthropic’s current goal is to inculcate an authentic personality inside Claude without deferring or being overbearing towards humans.
While there are emerging discussions and debates about LLMs hitting a wall and reaching their scalability limits, aligning these models for better emotional intelligence is an option.
But, with a word of caution.
Earlier this year, OpenAI released a ‘system card’ that warns about the potential for excessive attachment to an emotionally intelligent AI.
“Human-like socialisation with an AI model may produce externalities impacting human-to-human interactions. For instance, users might form social relationships with the AI, reducing their need for human interaction, potentially benefiting lonely individuals but possibly affecting healthy relationships,” said OpenAI in the report.
Unfortunately, there has been a reported case of a user forming a deep emotional attachment to a personality in CharacterAI, which ultimately led to their untimely death.
Improving emotional intelligence in AI models leaves much to be desired, but there are also many problems to be solved. Striking a balance between empathy and responsibility is crucial as we continue to develop AI technologies that interact on such a personal level.
