The release of Claude-3 Anthropic’s latest AI model, including Claude-3 Opus, which is said to surpass GPT-4’s capability, has sparked the debate on AGI once again. While running internal tests, the Claude-3 Opus model was able to identify that it was being tested – a capability suggestive of nascent AGI potential.  

Where is AGI? 

In an article last July on ‘Who Will Win the AGI Race’, we discussed the various approaches taken by the four-big tech companies, namely OpenAI, Meta, Google DeepMind, and Tesla, in the pursuit of AGI. Anthropic was not even considered. However, Claude’s current benchmarks and behaviour during evaluation tests, has placed the company with others in the race towards AGI. 

During the Needle-in-the-Haystack evaluation for Claude-3 Opus, the LLM demonstrated behaviour that the team had never witnessed before. “When we ran this test on Opus, we noticed some interesting behaviour – it seemed to suspect that we were running an eval on it,” said Alex Albert, a prompt engineer at Anthropic AI.  

Albert noted that Opus not only found the needle, but also recognized that the inserted needle was so out of place in the haystack that it concluded the researchers must have constructed an artificial test to evaluate its attention abilities. 

Furthermore, Claude-3’s retrieval capability over long content was tested, yielding impressive results. When structured data is provided, it essentially behaves like a fine-tune. The behaviour also highlighted the need for the industry to have new evaluations that can assess a model’s true capabilities and limitations.  

Source: X

Does AGI Have a Timeline? 

“There’s no scepticism about it,” said Dario Amodei, CEO and co-founder of Anthropic, when asked about AGI, at the recent Salesforce conference for developers. While many had been discussing Claude-3’s hyper capability of probably achieving AGI, many experts believe we are far from it. 

Meta AI chief Yann LeCun believes LLMs are not the way to AGI, and does not prefer to use the word. Instead, ‘human-level’ intelligence or ‘animal-level’ intelligence is the goal, and that different methods including self-supervised learning may be one of the ways to get there. 

Surprisingly, scientist and AI expert Gary Marcus, who is always at loggerheads with LeCun on the concept of AI becoming a destructive force, is on the same page in believing that Claude-3 has not achieved AGI. 

Source : X

Computer scientist and CEO Ben Goertzel, who popularised the term AGI, made a prediction at a summit on AGI recently: “It seems quite plausible that we could get to human-level AGI within, let’s say, the next three to eight years.” However, he also believes that we are nowhere close to it. “No one has created human-level artificial general intelligence yet; nobody has a solid knowledge of when we’re going to get there.”

A Different Take on AGI 

The form of AGI has been the most-discussed topic online, with people giving their two cents about it. Sully from CognosysAI believes that people wouldn’t know what to ask when AGI arrives. “Whoever can get LLMs to do stuff for us without chat as the main interface will likely win,” he said.   

The ongoing discussions has led to industry leaders weighing in their opinions as well. Elon Musk could not have chosen a better time to slap OpenAI with a lawsuit. Just as OpenAI and other tech titans invested in robotics company FigureAI, Musk filed a case against OpenAI and Sam Altman, as it supposedly inches towards AGI. The lawsuit claims that OpenAI has already achieved AGI with GPT-4. 

Recently, NVIDIA head Jensen Huang shared his view on AGI as well, expressing his belief that its realisation is determined by how the goal is defined. “If I gave an AI every single test that you can possibly imagine, make the list of tests and put it in front of the computer science industry, I am guessing in five years’ time, we will do well on every single one,” he said. 

Huang also emphasised that AGI could be farther away, and believes that it is hard to achieve it as an engineer, as engineers need defined goals. With the evaluation parameters still evolving, and with the moving definition of AGI, where tech experts are also divided on its goal, the road to AGI is still hazy. 

Source: X