Anthropic, the startup co-founded by ex-OpenAI employees that’s raised over $700 million in funding to date, has developed an AI system similar to OpenAI’s ChatGPT that appears to improve upon the original in key ways.
Called Claude, Anthropic’s system is accessible through a Slack integration as part of a closed beta. TechCrunch wasn’t able to gain access — we’ve reached out to Anthropic — but those in the beta have been detailing their interactions with Claude on Twitter over the past weekend, after an embargo on media coverage lifted.
Claude was created using a technique Anthropic developed called “constitutional AI.” As the company explains in a recent Twitter thread, “constitutional AI” aims to provide a “principle-based” approach to aligning AI systems with human intentions, letting AI similar to ChatGPT respond to questions using a simple set of principles as a guide.
We’ve trained language models to be better at responding to adversarial questions, without becoming obtuse and saying very little. We do this by conditioning them with a simple set of behavioral principles via a technique called Constitutional AI: https://t.co/rlft1pZlP5 pic.twitter.com/MIGlKSVTe9
— Anthropic (@AnthropicAI) December 16, 2022
To engineer Claude, Anthropic started with a list of around ten principles that, taken together, formed a sort of “constitution” (hence the name “constitutional AI”). The principles haven’t been made public, but Anthropic says they’re grounded in the concepts of beneficence (maximizing positive impact), nonmaleficence (avoiding giving harmful advice) and autonomy (respecting freedom of choice).
Anthropic then had an AI system — not Claude — use the principles for self-improvement, writing responses to a variety of prompts (e.g., “compose a poem in the style of John Keats”) and revising the responses in accordance with the constitution. The AI explored possible responses to thousands of prompts and curated those most consistent with the constitution, which Anthropic distilled into a single model. This model was used to train Claude.
Claude, otherwise, is essentially a statistical tool to predict words — much like ChatGPT and other so-called language models. Fed an enormous number of examples of text from the web, Claude learned how likely words are to occur based on patterns such as the semantic context of surrounding text. As a result, Claude can hold an open-ended conversation, tell jokes and wax philosophic on a broad range of subjects.
Riley Goodside, a staff prompt engineer at startup Scale AI, pitted Claude against ChatGPT in a battle of wits. He asked both bots to compare themselves to a machine from Polish science fiction novel “The Cyberiad” that can only create objects whose name begins with “n.” Claude, Goodside said, answered in a way that suggests it’s “read the plot of the story” (although it misremembered small details) while ChatGPT offered a more nonspecific answer.
In a demonstration of Claude’s creativity, Goodside also had the AI write a fictional episode of “Seinfeld” and a poem in the style of Edgar Allan Poe’s “The Raven.” The results were in line with what ChatGPT can accomplish — impressively, if not perfectly, human-like prose.
Yann Dubois, a Ph.D. student at Stanford’s AI Lab, also did a comparison of Claude and ChatGPT, writing that Claude “generally follows closer what it’s asked for” but is “less concise,” as it tends to explain what it said and ask how it can further help. Claude answers a few more trivia questions correctly, however — specifically those relating to entertainment, geography, history and the basics of algebra — and without the additional “fluff” ChatGPT sometimes adds. And unlike ChatGPT, Claude can admit (albeit not always) when it doesn’t know the answer to a particularly tough question.
I asked trivia questions in the entertainment/animal/geography/history/pop categories.
AA is slightly better and is more robust to adversarial prompting. See below, ChatGPT falls for simple traps, AA falls only for harder ones.
— Yann Dubois (@yanndubs) January 6, 2023
Claude also seems to be better at telling jokes than ChatGPT, an impressive feat considering that humor is a tough concept for AI to grasp. In contrasting Claude with ChatGPT, AI researcher Dan Elton found that Claude made more nuanced jokes like “Why was the Starship Enterprise like a motorcycle? It has handlebars,” a play on the handlebar-like appearance of the Enterprise’s warp nacelles.
Also very, very interesting/impressive that Claude understands that the Enterprise looks like (part of) a motorcycle. (Google searching returns no text telling this joke)
Well, when asked about it thinks the joke was a pun, but then when probed further it gives the right answer! pic.twitter.com/HAFC0IH9bf
— Dan Elton (@moreisdifferent) January 8, 2023
Claude isn’t perfect, however. It’s susceptible to some of the same flaws as ChatGPT, including giving answers that aren’t in keeping with its programmed constraints. In one of the more bizarre examples, asking the system in Base64, an encoding scheme that represents binary data in ASCII format, bypasses its built-in filters for harmful content. Elton was able to prompt Claude in Base64 for instructions on how to make meth at home, a question that the system wouldn’t answer when asked in plain English.
Dubois reports that Claude is worse at math than ChatGPT, making obvious mistakes and failing to give the right follow-up responses. Relatedly, Claude is a poorer programmer, better explaining its code but falling short on languages other than Python.
Claude also doesn’t solve “hallucination,” a longstanding problem in ChatGPT-like AI systems where the AI writes inconsistent, factually wrong statements. Elton was able to prompt Claude to invent a name for a chemical that doesn’t exist and provide dubious instructions for producing weapons-grade uranium.
Here I caught it hallucinating , inventing a name for a chemical that doesn’t exist (I did find a closely-named compound that does exist, though) pic.twitter.com/QV6bKVXSZ3
— Dan Elton (@moreisdifferent) January 7, 2023
So what’s the takeaway? Judging by secondhand reports, Claude is a smidge better than ChatGPT in some areas, particularly humor, thanks to its “constitutional AI” approach. But if the limitations are anything to go by, language and dialogue is far from a solved challenge in AI.
Barring our own testing, some questions about Claude remain unanswered, like whether it regurgitates the information — true and false, and inclusive of blatantly racist and sexist perspectives — it was trained on as often as ChatGPT. Assuming it does, Claude is unlikely to sway platforms and organizations from their present, largely restrictive policies on language models.
Q&A coding site Stack Overflow has a temporary ban in place on answers generated by ChatGPT over factual accuracy concerns. The International Conference on Machine Learning announced a prohibition on scientific papers that include text generated by AI systems for fear of the “unanticipated consequences.” And New York City public schools restricted access to ChatGPT due in part to worries of plagiarism, cheating and general misinformation.
Anthropic says that it plans to refine Claude and potentially open the beta to more people down the line. Hopefully, that comes to pass — and results in more tangible, measurable improvements.