AI influencer, author and professor of computer science, Pedro Domingos, famous for sharing his two cents on everything AI, recently said, “We don’t need more data, we need better reasoning.”
While companies have been scavenging for data to train their AI models, talking about ‘n’ billion parameter models during every launch, experts have now started focusing more on the AI model’s capabilities.
Quality Data will Govern Superior Models
Speaking about the future of AI at the AI for Good Global Summit, OpenAI chief Sam Altman touched upon the necessity of having high-quality data.
“There is low-quality synthetic data, and there’s low-quality human data. As long as we can find enough quality data to train our models or ways to train and get better at data efficiency, and learn more from smaller amounts of data or any number of other techniques, I think that’s okay,” said Altman.
He also confirmed that “they have what they need for the next model”, presumably, GPT-5. Altman also said that they have generated enough synthetic data to experiment with training on that.
At the Databrics’ Data + AI Summit, computer scientist Yejin Choi lauded OpenAI’s effort in building quality data. “AI, at least in the end, is as good as the data. The key test of current AI usually depends primarily on generating datasets. If you cannot do this, you should do it in a more innovative way,” she added.
‘Capable’ Models
Making a cameo at the Microsoft Build 2024 event, Altman vaguely spoke about the upcoming model once again, without revealing details. “The most important thing is that the models are just generally going to get smarter across the board,” he said.
Alluding to the model’s increasing reasoning capabilities, Altman spoke about how models are significantly improving over time.
“We actually are seeming to increase the general capability of the model across the board. That’s going to keep happening and the jump that we have seen in the utility that a model can deliver with each of those half-step jumps and smartness, it’s quite significant each time,” he exclaimed.
Models are Only Getting Smarter
In a recent interview with AIM, IIT Bombay professor and computer scientist Pushpak Bhattacharyya, spoke about understanding complex human emotions in LLMs. “Chatbots that are polite and understand sentiment, emotion, etc. give rise to better businesses,” he said.
Bhattacharyya believes that chatbots that are closer to human beings, emotionally and sentimentally are quite motivating.
Promising Research
Going by the recent research papers published in this field, attempts to make models highly reasonable are ongoing. For instance, a team of Stanford and independent researchers recently published a paper titled, ‘How Culturally Aware are Vision-Language Models?’. The idea is to make AI models culturally aware.
“In terms of the Indian context, we wanted to understand how global models like Gemini and GPT recognise our cultural symbols,” said Vinija Jain, ML leader at Amazon, and one of the researchers of the paper.
Concepts such as zero-shot learning aim to handle complex reasoning tasks. Papers such as ‘Large Language Models are Zero-Shot Reasoners’ show that LLMs can handle complex reasoning tasks using a simple prompt like, “let’s think step by step”.
Likewise, Enhancing Zero-Shot Chain-of-Thought Reasoning in Large Language Models introduces logical thought frameworks to improve multi-step reasoning.
Interestingly, any new model launch is compared on a benchmark across parameters including reasoning capabilities. When Anthropic launched its latest model, Claude 3, the AI model was compared with its competitors, GPT-4 and Gemini, and fared much better in reasoning capabilities.
Source: X
Large to Small Models
From building large AI models, there has also been a significant shift in building smaller yet powerful models. OpenAI’s GPT-4o, Microsoft’s Phi 2, TII’s Falcon 2 are just a few examples of big tech companies’ foray into small language models.
Despite the smaller parameter size, the models are performing at superior levels. Falcon 2 models were said to even surpass Llama-3.
Interestingly, though the progress towards making models smaller with superior reasoning capabilities will continue, data collection is here to stay. Looking at the number of media publication partnerships of OpenAI, including News Corp, WSJ, Vox, The Atlantic, and many more, it is evident that the tech giant is still reliant on it.
It is possible that nobody has figured out the right approach yet. After all, Altman himself recently said, “We certainly have not solved interpretability,” and admitted that OpenAI doesn’t fully understand how its AI works.