Even though they can produce incredibly lifelike prose, generative AIs like Google's Bard or OpenAI's ChatGPT (powered by GPT-4) have already demonstrated the current limitations of gen-AI technology as well as their own shaky grasp of reality by asserting that Elvis Presley's father was an actor and that the JWST was the first telescope to image an exoplanet. But with such a large market share at risk, what do a few inaccurate facts have to do with getting their product in front of customers as soon as possible?
In contrast, the Anthropic team, which is primarily made up of ex-OpenAI employees, has developed their own chatbot, Claude, using a more practical approach. According to a TechCrunch story, the outcome is an AI that is "more steerable" and "far less likely to create hazardous outputs" than ChatGPT.
Since late 2022, Claude has been in closed beta, but just lately did launch partners Robin AI, Quora, and privacy-focused search engine Duck Duck Go start testing the AI's conversational skills. The business has not yet disclosed price, but it has assured TC that two versions will be offered at launch: the normal API and Claude Instant, a quicker, lighter form.
"We utilise Claude to analyse certain areas of a contract, and to recommend fresh, alternative wording that's more favourable to our consumers," Robin CEO Richard Robinson told TechCrunch. "We've found Claude is particularly adept at interpreting words — especially in technical realms like legal language. Also, it is particularly competent in drafting, summarising, translating, and clearly articulating complicated ideas.
The "constitutional AI" training programme that the business is calling is one of the reasons Anthropic thinks Claude won't go rogue and start hurling racial epithets as Tay did. The business claims that this offers a "principle-based" strategy for bringing people and machines to the same moral ground. Anthropic began with 10 guiding principles; however, the business won't tell what they are in detail, which is a peculiar marketing gimmick involving 11-secret-herbs-and-spices. Suffice it to say, though, that "they're anchored in the ideals of beneficence, nonmaleficence, and autonomy," per TC.
The business then used several writing challenges, such as "Compose a poem in the manner of John Keats," to teach a second AI to consistently write content in line with those semi-secret rules. Then Claude was trained using that model. Yet even while Claude has been conditioned to be essentially less troublesome than its rivals, that doesn't mean it doesn't occasionally have delusions of reality similar to those of a startup CEO on an ayahuasca retreat. The AI apparently performed worse on arithmetic and grammar standardised tests than ChatGPT, and it has already created a brand-new chemical and used artistic licence to the uranium enrichment process.
The issue, according to the Anthropic spokesman, is to create models that are both never hallucinate and are still helpful. You could find yourself in a difficult circumstance where the model decides that the best way to never lie is to never say anything at all. Hallucinations have decreased, but there is still work to be done.
Follow Us