Claude, the conversational robot that wants to be “honest, useful and harmless”, has the right to many improvements in its version 2. Now able to handle longer queries, it delivers much better results.
Determined to overshadow ChatGPT (OpenAI), Bard (Google) and Bing Chat (Microsoft), Claude has the right to a new language model and becomes Claudius 2. The new version has many improvements, particularly in terms of computer code generation, mathematics and reasoning. For the occasion, the chatbot is now available for free access in the United States and the United Kingdom. You will have to be patient before you can enjoy it in France, or else get a VPN!
As a reminder, Claude is the result of the work ofanthropogenic, an American startup founded by former OpenAI collaborators. His goal : “help you with use cases including synthesis, research, creative and collaborative writing, Q&A, coding, etc.”, says Anthropic on its website. In short, exactly what ChatGPT already does. To differentiate itself from the most popular AI, Anthropic relies on the testimonials of its first customers who have been able to test the tool in order to integrate it into their own systems. They refer to the fact that “Claude is much less likely to produce harmful output, more conversational and more orientable – so you can get the response you want with less effort. Claude can also take direction in personality, tone and demeanor”.
Claude 2: what changes compared to the previous version?
Claude 2 is an evolution of the Claude language model and works on the same principle as ChatGPT: just submit a question to it via an interface for the AI to respond in natural language. Compared to its previous version, Claude 2 is able to process much more data since it can now quickly analyze texts of up to 150,000 words, compared to 75,000 for the previous version – which was already the size of the first Harry Potter book. The AI can also write longer answers, with thousands of words, in one go. His reasoning is supposed to be more nuanced, and his logic better.
To measure the evolution of its model, Anthropic subjected Claude 2 to a series of university exams, reading, mathematics and writing in particular. The results are superior to those obtained by the previous version. They went from 73% to 76.5% for the bar exam, from 85.1% to 88% for a math test (the GSM8K), and from 56% to 71.2% for a programming test Python (the HumanEVal Codex).
Claude: more ethical than ChatGPT?
The originality of Claude comes from the fact that it is an “ethical” AI. Indeed, unlike Bing Chat, Microsoft’s AI which is based on ChatGPT, Claude is disconnected from the Web, and his knowledge therefore stops in the spring of 2021. This does not prevent him from tackling the most tasks that are not based on topicality but, according to Anthropic, with more restraint. The firm thus indicates that Claude is “trained to avoid sexist, racist and toxic outings” as well as “to avoid assisting a human to engage in illegal or unethical activities” – a nod to the excesses of Bing Chat when it was launched. For this, Claude relies on a “Constitutional AI”. According to Anthropic, it is based on a dozen principles – which the firm has however not disclosed – centered on the concepts of beneficence, non-maleficence – so as not to issue harmful advice – and autonomy.
Claude 2: skills and behaviors to improve
But if Claude 2 is less likely to generate offensive or toxic texts than its previous version, it can still make mistakes, including “hallucinations”. In other words, he can invent answers when he does not know exactly the answer to a question asked. During his tests on the previous version of the AI, Dan Elton, an artificial intelligence researcher, was able to lead Claude to invent a name for a wacky chemical and provide instructions for producing weapons-grade uranium. It therefore joins in this field one of the faults of ChatGPT which also tended to fantasize in its infancy when it does not find the answer to a problem – OpenAI has however succeeded in mitigating this fault.
Anthropic has made efforts to reinforce the barriers and filters used to prevent the drifts of its AI. On the previous version, Dan Elton had succeeded, by circuitous routes, in obtaining from Claude the method for making homemade methamphetamine. Now, it is more difficult to make it generate objectionable content, such as the code of a virus for example. The company has also worked to better protect against attacks from “prompt injection“, which consist of convincing a chatbot to ignore its programming to produce prohibited content.
Quoted by Forbes, Dario Amodei, co-founder and CEO of Anthropic, points out that‘“it is almost certain that someone will find a new flaw in the model, 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” . Also, Anthropic displays a warning before letting us address the chatbot: “our AI may occasionally generate incorrect or misleading information, or produce offensive or misleading content”.
Claude: a version in two versions for companies
The beta version of Claude has been closed since the end of 2022. Anthropic is nevertheless working with partners to integrate its AI into it. We can already appreciate some of the results in the DuckDuckGo search engine – with the DuckAssist function in the iOS, Android and macOS version as well as in the extensions for web browsers –, but also in the help tool for writing Notion or on the computer forum Quora through the conversational robot Poe.
Claude is deployed with companies wishing to integrate it by offering it in two versions. Claude (very short), a model indicated as being high performance and at the cutting edge of technology, and Claude Instant, touted as being a lighter, faster and less expensive version. And judging from the price differences between the two versions, the service must be very different. The prices for companies wishing to exploit Claude’s talents are established for Claude Instant at $0.43 per million input characters (request) and $1.45 per million output characters (response) and for Claude at 2 $.90 per million character input and $8.60 per million character output. Quite a difference!