Is Claude 2 Really a Better Chatbot Than ChatGPT?
Is Claude 2 Really a Better Chatbot Than ChatGPT?
The second iteration of Claude, the conversational robot striving to be honest, useful, and harmless, has undergone a multitude of improvements in version 2. It now boasts enhanced capabilities to handle longer queries, resulting in significantly superior outcomes.
Claude, the conversational robot committed to being "honest, useful, and harmless," has received numerous enhancements in its version 2. It can now handle longer queries, leading to significantly better outcomes.
As a reminder, Claude is the result of Anthropic's work, an American startup founded by former OpenAI collaborators. Their aim is to assist you with use cases including synthesis, research, creative and collaborative writing, question-answering, coding, etc., as stated on the Anthropic website. In essence, it does what ChatGPT already does.
To differentiate itself from the most popular AI, Anthropic relies on testimonials from its initial clients who have tested the tool and integrated it into their own systems. These clients mention that Claude is much less likely to produce harmful outputs, easier to converse with, and more controllable meaning you can achieve the desired response with less effort. Claude can also take instructions regarding personality, tone, and behavior.
What Evolutions Does Claude 2 Bring Compared to the Previous Version?
Claude 2 is an evolution of the Claude language model, operating on the same principle as ChatGPT submits a question through an interface, and the AI responds in natural language. Compared to its previous version, Claude 2 can now process significantly larger amounts of data. It can rapidly analyze texts up to 150,000 words, compared to 75,000 words in the previous version, which already equated to the length of the first Harry Potter book. Moreover, the AI can generate longer responses consisting of thousands of words in a single attempt. Its reasoning is expected to be more nuanced, with improved logic.
To evaluate the progress of the model, Anthropic subjected Claude 2 to a series of university exams, including reading, mathematics, and writing. The results surpassed those of the previous version, with scores increasing from 73% to 76.5% for the bar exam, from 85.1% to 88% for a mathematics test (GSM8K), and from 56% to 71.2% for a Python programming test (Codex HumanEVal).
Is Claude More Ethical Than ChatGPT?
What sets Claude apart is its claim to be an ethical AI. Unlike Bing Chat, Microsoft's AI relying on ChatGPT, Claude operates in isolation from the web, and its knowledge is limited to spring 2021. However, this doesn't prevent it from handling most tasks unrelated to current events, albeit with more restraint, as indicated by Anthropic.
The company states that Claude is trained to avoid sexist, racist, and toxic outputs and to avoid assisting humans in engaging in illegal or unethical activities. This alludes to Bing Chat's shortcomings during its launch. To achieve this, Claude relies on a Constitutional AI. According to Anthropic, this framework is based on around ten principles although the company has not disclosed them focused on concepts of benevolence, non-malevolence to prevent harmful advice and autonomy.
Areas for Improvement in Claude 2's Skills and Behavior
Although Claude 2 is less likely to generate offensive or toxic texts compared to its previous version, it can still make mistakes, including "hallucinations." In other words, when faced with a question to which it doesn't have an answer, it may invent elements of a response. During tests with the previous version of the AI, Dan Elton, an AI researcher, managed to lead Claude to invent a name for a fictional chemical product and provide instructions for producing military-grade uranium. Thus, Claude shares one of ChatGPT's early flaws, which tended to fabricate responses when unable to find an answer. OpenAI has since managed to mitigate this flaw.
Anthropic has made efforts to strengthen the barriers and filters used to prevent AI missteps. In the previous version, Dan Elton was able, through indirect methods, to obtain instructions from Claude on producing homemade methamphetamine. Now, it is more challenging to make it generate objectionable content, such as virus code. The company has also worked on better protection against "prompt injection" attacks, which involve convincing a chatbot to ignore its programming and produce prohibited content.
Quoted in Forbes, Dario Amodei, co-founder and CEO of Anthropic, emphasizes that it is almost certain that someone will find a new flaw in the model, and someone will notice a bad hallucination. There are too many things you can say to the model and too many things the model can respond to. These things will never be perfect. As such, Anthropic displays a warning before allowing users to interact with the chatbot: Our AI may occasionally generate incorrect or misleading information or produce offensive or biased content.
Claude: Two Versions for Enterprise Deployment
The beta version of Claude has been closed since the end of 2022. However, Anthropic is working with partners to integrate its AI. Partial results can already be seen in the DuckDuckGo search engine, with the DuckAssist feature available on iOS, Android, macOS, and web browser extensions. It is also present in the Notion writing aid tool and the Quora computer forum through the Poe chatbot.
Claude is being deployed to businesses that wish to incorporate it by offering two versions. Claude's full version is indicated as being high-performance and state-of-the-art technology, while Claude Instant is touted as a lighter, faster, and more cost-effective version. Judging by the price differences between the two versions, the services provided must be significantly different. The pricing for enterprises looking to leverage Claude's capabilities is set at $0.43 per million input characters (query) and $1.45 per million output characters response for Claude Instant, and $2.90 per million input characters and $8.60 per million output characters for Claude.
