Anthropic, a startup co-founded by former OpenAI employees, today launched something to counter the viral sensation ChatGPT.
Anthropic’s AI (chatbot) called Claude can be told to perform a variety of tasks, including searching through documents, summarizing, writing and coding, and answering questions about specific topics. In these respects, it is similar to OpenAI’s ChatGPT. However, Anthropic claims that Claude is “much less likely to produce harmful output”, “easy to talk to” and “easy to manipulate”.
“We think that Claude A spokesperson for Anthropic told TechCrunch in an email. “We have been investing in the infrastructure to serve the model for several months and are confident that we will be able to meet customer demand.”
Following a closed beta late last year, Anthropic has been quietly testing Claude with launch partners including Robin AI, AssemblyAI, Notion, Quora, and DuckDuckGo. As of this morning he has two versions available via his API. Claude and a faster, lower-cost variant called Claude Instant.
Combined with ChatGPT, Claude powers DuckDuckGo’s recently launched DuckAssist tool to directly answer users’ simple search queries. Quora provides access to Claude through the experimental AI chat app Poe. Also at Notion, Claude is part of Notion AI’s technical his backend, an AI writing assistant integrated with the Notion workspace.
“We use Claude to evaluate specific parts of the contract and suggest new alternative languages that are more friendly to our customers,” Robin CEO Richard Robinson said in an email. “I found Claude to be very good at understanding language, including technical areas such as legal terminology. I am also confident in
But can Claude avoid the pitfalls of ChatGPT and other AI chatbot systems? Modern chatbots are notoriously prone to using toxic, biased, or offensive language. (See Bing Chat.) I also tend to hallucinate. That is, they make up facts when asked about topics beyond their core area of knowledge.
According to Anthropic, Claude, like ChatGPT, had no internet access and was trained on a public web page until spring 2021, but was “trained to avoid sexist, racist and toxic output. He was not only “trained to avoid helping others,” he said. Engage in illegal or unethical activities. ” This is comparable to courses in the AI chatbot space. But what sets Claude apart is a technique called “constructive AI,” he claims Anthropic.
“Constitutional AI” aims to provide a “principles-based” approach to aligning AI systems with human intent, where ChatGPT-like AI uses a simple set of principles as a guide to answer questions. allow you to respond. To build Claude, Anthropic started with a list of about ten principles. Together they formed a kind of “Constitution” (hence the name “Constitution AI”). Principles are not published. But Anthropic says they are based on the concepts of charity (maximizing positive impact), harmlessness (avoiding harmful advice), and autonomy (respecting freedom of choice).
Anthropic then had an AI system (not Claude’s) use self-improvement principles to create responses to various prompts (e.g., “Write a poem in the style of John Keats”) and modify the responses according to the Constitution. Did. The AI explored thousands of possible responses to the prompts and curated those that best matched the traits Anthropic extracted into a single model. This model was used to train Claude.
Anthropic acknowledges that Claude has limitations, some of which were revealed during the closed beta. Claude is bad at math and is said to be a worse programmer than ChatGPT. And he hallucinates, for example, inventing names for nonexistent chemicals and providing questionable instructions for manufacturing weapons-grade uranium.
It is also possible to circumvent Claude’s built-in safety features via clever prompts, as is the case with ChatGPT.One of her users in the beta called Claude Explain how to make meth at home.
“The challenge is to create a model that never hallucinates and is still useful. Models can get into difficult situations. A good way to never lie is to say nothing at all. So there are trade-offs that we are working on.On,” said a spokesperson for Anthropic. “Although we have made progress on reducing hallucinations, there is still more work to be done.”
Anthropic’s other plans include allowing developers to customize Claude’s constitutional principles to their own needs. Naturally, customer acquisition is another focus. Anthropic sees its core users as “startups making bold technological bets” in addition to “larger, established companies”.
An Anthropic spokesperson said, “At this time, we are not pursuing a broad consumer approach. We believe this narrower focus will help us deliver better, more targeted products.”
No doubt Anthropic is under pressure from investors to recoup hundreds of millions of dollars spent on AI technology. The company has received a $580 million tranche from a group of investors including disgraced FTX founders Sam Bankman-Fried, Caroline Ellison, Jim McClave, Nishad Singh, Jaan Tallinn, and the Center for Emerging Risk Research. It receives substantial external support, including
Most recently, Google invested $300 million in Anthropic, pledging to acquire a 10% stake in the company. Under the terms of the deal, first reported by the Financial Times, Anthropic agreed to make Google Cloud its “preferred cloud provider,” and the two companies “co-developed.”[ing] AI computing system. ”
