Anthropic, the AI safety startup founded by Dario Amodei and Daniela Amodei, has released a major update to its conversational AI assistant Claude. This new version, called Claude 2.1, is packed with improvements that take Claude even further than the capabilities of chatbots such as ChatGPT.
The huge context window of 200,000 ensures greater accuracy
The main feature in Claude 2.1 is the strengthened context window, now expanded to a whopping 200,000 tokens. This allows Claude to consult and ‘remember’ up to 200,000 words or 500 pages of previous conversations when formulating answers.
By having much more conversational history and context, Claude can significantly reduce instances of confusion, contradiction, and hallucinations. Anthropic claims Claude 2.1 shows a doubling of the total number of hallucinations compared to the previous version.
The sheer size of this context window of 200,000 far exceeds any other chatbot to date. And it shows in Claude’s improved accuracy and clarity of answers. When users ask more complex, multi-part questions, Claude can reference previous statements in depth instead of losing sight of context.
30% fewer errors and 3-4x lower error rates
In addition to the improved context window, Anthropic Claude 2.1 engineering work focused on immediately minimizing factual errors, incorrect answers, and wrong conclusions.
Internal testing shows that Claude 2.1 makes 30% fewer factual errors compared to version 2.0. Moreover, the percentages of incorrect answers are generally three to four times lower.
These gains in accuracy demonstrate Anthropic’s focus on going beyond scale to actually improving the fundamental reliability and veracity of AI systems. While chatbots like ChatGPT seem to hallucinate relatively often, Claude sticks to reality.
New tool integration provides access to calculators, APIs and more
Claude 2.1 introduces a beta feature called Tool Use, which allows the assistant to integrate external tools, data sources, and private APIs into its responses.
For example, Claude can now use calculators to solve users’ mathematical problems directly, instead of just describing the theoretical solution. It can process sports scores, weather reports and other real-time data via API calls instead of trying to guess or make up answers.
Claude can even pull data from private databases to which he has special access rights to provide accurate, factual answers. And it can also perform basic actions within software systems via API triggers.
This tool integration helps ground Claude in reality, giving him access to the same resources a human expert would use when answering questions.
Developer Experience Upgrades – Workbench and System Prompts
While end users will appreciate Claude’s increased accuracy and capabilities, Claude 2.1 also ushers in upgrades specifically aimed at the developer experience.
These include a new Workbench product that allows Claude prompts to be repeated more quickly during training and development cycles. Workbench makes it easier and faster for Anthropic engineers to test and refine Claude’s skills.
Additionally, developers can now provide Claude with custom system prompts with clear instructions to guide behavior in specific situations. Instead of just relying on training through sample conversations, engineers can give Claude direct top-down rules on how to behave using system prompts.
Prices remain consistent with Claude 2.0
So far, Anthropic has only worked with select partners to test Claude, charging $0.60 per prompt and $1.80 for each completion.
With Claude 2.1, they are increasing prices to $1.60 per prompt and $4.80 per completion to account for the significant engineering work required to support such a huge context window.
However, Anthropic emphasizes that Claude’s pricing will remain consistent with version 2.0 for the time being. The higher costs mentioned above represent potential future prices, as supporting Claude’s expanded capabilities at scale will require enormous computing power.
However, for partners involved in early pilots, Anthropic will stick with the previous pricing model while they gather user feedback on Claude 2.1.
With features like the groundbreaking 200,000 token context window, Claude 2.1 represents the cutting edge of conversational AI. Anthropic continues to push boundaries in accuracy, veracity, and understanding that chatbots from less carefully designed stables can only dream of in their current form.
Claude 2.1 is an impressive achievement for the team at Anthropic and demonstrates their rigorous approach to developing safe and reliable AI systems step by step. If they can perfect and produce capabilities like the massive context window, Claude could set a new bar for the entire industry. But for now, companies should apply for an early access pilot partnership if they want to get a glimpse of the future of chatbot interfaces.
🌟Do you have any burning questions about Claude 2.1? Do you need some extra help with AI tools or something else?
💡 Feel free to send an email to Govind, our expert at OpenAIMaster. Send your questions to email@example.com and Govind will be happy to help you!