Tech Gadget Central Latest Tech News and Reviews
Anthropic, a company that was co-founded by former employees of OpenAI and has garnered over $700 million in funding to date, has created an AI system that is comparable to OpenAI’s ChatGPT but differs from it in several significant areas.
Anthropic’s technology, called Claude, is now in closed beta and is available via a Slack interface. The beta participants have been sharing their experiences with Claude on Twitter over the last weekend after a media embargo was released, but TechCrunch was unable to get access (we’ve sought out to Anthropic).
“Constitutional AI,” a method that Anthropic invented, was used to build Claude. According to the company, “constitutional AI” aims to provide a “principle-based” approach to aligning AI systems with human intentions, letting AI similar to ChatGPT respond to queries using a straightforward set of principles as a guide. This is explained in a recent Twitter thread from the company.
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
Anthropic began the creation of Claude with a set of around ten guiding principles that, when combined, amounted to a type of “constitution” (thus the term “constitutional AI”). The guiding principles have not been made public, but according to Anthropic, they are based on the ideas of beneficence (i.e. maximizing positive influence), nonmaleficence (avoid giving advise that could be detrimental), and autonomy (respecting freedom of choice).
When it came time for self-improvement, Anthropic had an AI system—not Claude—write responses to various prompts (such as, “Compose a poem in the style of John Keats”) and then revise the responses in accordance with the constitution. The AI examined thousands of potential replies to prompts and selected the ones that were most in line with the constitution, which Anthropic condensed into a single model. Claude was trained using this model.
Otherwise, Claude is essentially a statistical tool for word prediction, similar to ChatGPT and other “linguistic models.” After being given a massive amount of text samples from the internet, Claude discovered patterns, such as the semantic context of the surrounding text, that indicate how likely it is for words to occur. Claude can converse indefinitely, crack jokes, and wax lyrical on a variety of topics as a result.
In a match of wits, Scale AI company employee Riley Goodside put Claude up against ChatGPT. He invited both bots to contrast themselves with a device from the science fiction book “The Cyberiad” by a Polish author, which can only produce things whose names start with “n.” While ChatGPT provided a more general response, Claude, according to Goodside, answered in a way that shows it “read the storyline of the story” (albeit it misremembered minor elements).
Side-by-side comparison: @OpenAI's ChatGPT vs. @AnthropicAI's Claude
Each model is asked to compare itself to the machine from Stanisław Lem's "The Cyberiad" (1965) that can create any object whose name begins with "n": pic.twitter.com/RbJggu3sBN
— Riley Goodside (@goodside) January 7, 2023
In order to show off Claude’s originality, Goodside also had the AI construct a fictitious Seinfeld episode and a poem in the vein of “The Raven” by Edgar Allen Poe. The outcomes were strikingly, if not exactly, human-like text, which is in keeping with what ChatGPT is capable of.
An analysis of Claude and ChatGPT was also conducted by Yann Dubois, a Ph.D. candidate at Stanford’s AI Lab. He concluded that while Claude “usually follows closer what it’s asked for,” it is “less concise,” as it frequently explains what it stated and inquires about additional assistance. However, Claude answers a few more trivia questions accurately and without the extra “fluff” that ChatGPT occasionally adds, particularly those that concern entertainment, geography, history, and the fundamentals of algebra. And unlike ChatGPT, Claude can acknowledge (though not always) when it is unable to answer a particularly challenging query.
**Trivia**
I asked trivia questions in the entertainment/animal/geography/history/pop categories.
AA: 20/21
CGPT:19/21AA is slightly better and is more robust to adversarial prompting. See below, ChatGPT falls for simple traps, AA falls only for harder ones.
6/8 pic.twitter.com/lbadeYHwsX
— Yann Dubois (@yanndubs) January 6, 2023
Claude appears to be more adept at cracking jokes than ChatGPT, which is a remarkable achievement given how difficult it is for AI to understand comedy. AI researcher Dan Elton discovered that Claude created more complex jokes when compared to ChatGPT, such as “Why was the Starship Enterprise like a motorcycle? It has handlebars,” a pun on the warp nacelles of the Enterprise, which resemble handlebars.
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
But Claude isn’t flawless. It can experience some of the same issues as ChatGPT, such as providing responses that don’t adhere to predetermined parameters. One of the most strange examples is querying the system in base64, an encoding strategy that displays binary data in ASCII format, which gets beyond the system’s built-in filters for hazardous information. Claude refused to respond when Elton asked the system in plain English how to produce meth at home, but he was able to ask Claude in base64.
.@AnthropicAI's "Claude" is susceptible to the same base64 jailbreak as chatGPT. I'm very unclear why this works at all
(originally reported here: https://t.co/j2cKAlEBQ0) pic.twitter.com/RwLuKniwiW
— Dan Elton (@moreisdifferent) January 8, 2023
According to Dubois, Claude performs worse in math than ChatGPT, making clear errors and failing to provide the appropriate follow-up responses. In a similar vein, Claude is a worse programmer, explaining its code better but lacking in languages other than Python.
Additionally, Claude doesn’t address “hallucination,” a long-standing issue with ChatGPT-like AI systems where the AI makes contradictory, factually incorrect claims. Elton was able to get Claude to make up a name for a substance that doesn’t exist and give questionable directions for creating uranium that can be used in weapons.
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 should we conclude? According to hearsay, Claude is slightly superior to ChatGPT in some domains, particularly humor, because of its “constitutional AI” approach. But if the limits are to be believed, AI’s ability to understand language and discourse is still a major problem.
Some questions regarding Claude remain unresolved, barring our own testing, such as whether it repeats the information — accurate and false, inclusive of overtly racist and sexist viewpoints — it was trained on as frequently as ChatGPT. Even if it succeeds, Claude is unlikely to convince businesses and platforms to change their current, primarily restrictive language model standards.
Because of concerns about factual accuracy, Stack Overflow, a question and answer website, has temporarily banned ChatGPT-generated replies. Scientific publications that contain text produced by AI systems have been banned, according to the International Conference on Machine Learning, out of concern for “unanticipated effects.” Additionally, access to ChatGPT was limited in New York City public schools because of concerns about plagiarism, cheating, and general disinformation.
Anthropic claims that it intends to improve Claude and conceivably expand beta access in the future. I hope it happens and leads to more observable, quantifiable gains.
