Ben
Anthropic, an AI startup, recently unveiled its latest major model, Claude 2. This new model showcases significant improvements across vital benchmarks, such as coding, math, and reasoning skills, while reducing harmful outputs. Unlike its predecessor, Claude 2 is more widely accessible in its second iteration.
To make it more accessible, Anthropic has launched a beta-test website called claude.ai, allowing UK and US users to register and explore the capabilities of Claude 2. Businesses can now access the new model through an API at the same price they paid for Anthropic’s previous model, Claude 1.3, which is still operational.
In an interview with Forbes, Dario Amodei, the co-founder and CEO of Anthropic, stated that Claude 2 represents significant evolutionary progress rather than a monumental leap from its predecessor. The improvements in the model are substantial, contributing to the general progression of AI technology.
Claude 2.0 has outperformed its predecessor in multiple assessments. It scored 71.2% on a Python coding test, an increase from 56%. It also improved its middle school math quiz grade from 85.2% to 88% and raised its Bar exam score from 73% to 76.5%. While the previous version could analyze prompts of up to approximately 75,000 words (equivalent to the length of the first Harry Potter book), Claude 2.0 has doubled that capacity, allowing it to handle prompts as long as an epic novel such as Gabriel Garcia Marquez’s “One Hundred Years Of Solitude.” The prompt size limit, however, will not be expanded.
Anthropic’s announcement of its new flagship model follows a recent funding round of $450 million, led by Spark Capital, valuing the company at over $4 billion. According to the company’s disclosure, thousands of businesses have already been utilizing Claude’s API. Anthropic is also collaborating with several larger customers, including Zoom, Notion, and Midjourney, an AI image generator, to develop customized models, although specific details still need to be provided.
Despite its split from OpenAI in 2021 because of differences in commercialization approaches, the launch of Claude 2.0 represents a strategic move for Anthropic, aligning with its original plan for commercialization. Amodei emphasized that adjustments in their approach are based on their discoveries in the field.
Anthropic recognized the potential benefits of involving business users in the “safety testing ground” to evaluate potential risks associated with the model. With the additional improvements in Claude 2.0, the company opened it up to a broader audience. Currently, the consumer version is free, but Anthropic may explore monetization options in the future, considering it an experiment.
Like previous Claude models, Claude 2.0 was trained using “Constitutional AI,” a framework that improves the AI’s results through AI supervision without direct human involvement. However, in the case of Claude 2.0, Anthropic employed some human feedback and oversight alongside this approach. As a result, the company claims Claude 2.0 is twice as compelling as its predecessor in mitigating harmful outputs.
Amodei acknowledged achieving perfection in AI models is unlikely, as there will always be the possibility of new issues arising. However, Anthropic believes it can work to address these challenges and prevent negative consequences. This stance contrasts with prominent tech leaders who have called for a temporary freeze on the release of AI models. Amodei proposed implementing safety checks as a requirement for significant model releases or training new models rather than imposing an industry-wide embargo.
Amodei’s views on balancing AI progress with safety align with his recent signing of a letter, alongside OpenAI CEO Sam Altman, Google DeepMind CEO Demis Hassabis, and others, complaining about the risks of AI leading to potential extinction. Anthropic believes it can contribute to mitigating such risks while continuing to develop and release new models.
_
Source: Forbes
