Google-backed Anthropic launches ChatGPT rival 'Claude 2'
%2Favatars%2FEwuSxftyWaVt.png&w=96&q=75)
%2F2023%2F07%2F12%2Fimage%2Fjpeg%2FcdMGjV06JSwE07MYNJPfDn8lAoWWlk7ma2RRf3WQ.jpg&w=3840&q=75)
Anthropic has launched Claude 2, their new general-purpose large language model, which they claim is a level up from other chatbots like OpenAI’s ChatGPT. “Claude is easy to converse with, clearly explains its thinking, is less likely to produce harmful outputs, and has a longer memory,” said the company in a blog post.
This comes four months after the Google-backed company launched the first version of Claude back in March. Anthropic was founded by former OpenAI employees Daniela and Dario Amodei.
Trained on what the company calls Constitutional AI, Claude models have been trained to be helpful and politely refuse responding to questions that might be harmful. Claude models tend to perform well in open-ended conversation, search, writing, editing, outlining, and summarizing text, coding, and providing helpful advice about a broad range of subjects.
Anthropic also launched a new public-facing beta website — Claud.ai — which is only available in the US and the UK as of now. The chatbot has so far had a good response from developers, who are saying that it’s a grade above GPT-4.
Improvements from the previous model in coding, math, reasoning
In their blog post, the company enumerated how Claude 2 scored better than Claude 1.3 in writing undergrad and grad level entrance exams. The latest model scored 76.5 percent on the multiple-choice section of the Bar exam, up from 73.0 percent with Claude 1.3. Claude 2 scored above the 90th percentile on the GRE reading and writing exams, similarly to the median applicant on quantitative reasoning, in comparison to college students applying to graduate school.
OpenAI declared a list of similar results of GPT-4’s performance in various entrance exams when it was launched back in March 2023.
One of the key differences is the increase of length in the input and output of Claude 2. Anthropic has expanded Claude’s context window from 9K to 100K tokens, which means Claude can analyze hundreds and thousands of pages of materials in a go, and conversations with the chatbot can last hours or even days. The input limit is somewhere around 75,000 words.
There’s also an upgrade in the coding skills, the company claims. “Claude 2 scored a 71.2 percent up from 56.0 percent on the Codex HumanEval, a Python coding test. On GSM8k, a large set of grade-school math problems, Claude 2 scored 88.0 percent up from 85.2 percent. We have an exciting roadmap of capability improvements planned for Claude 2 and will be slowly and iteratively deploying them in the coming months,” said the blog post.
Less harmful responses
One of the biggest challenges with using chatbots is their tendency to hallucinate. It seems that Anthropic has worked on improving the underlying safety of Claude 2. The company said that the latest version is “more harmless and harder to prompt to produce offensive or dangerous output.”
As per the company’s internal red-teaming evaluation that scores Anthropic’s models after introducing it to a large set of harmful prompts, they concluded that “Claude 2 was 2x better at giving harmless responses compared to Claude 1.3".
SHOW COMMENT (1)
/2023/09/14/image/jpeg/Ve0aCGvFSLp0CAq5VCSZU9HEmhnxEpni6LjG4BK6.jpg "How air quality systems could solve the biodiversity crisis")
/2023/08/12/image/jpeg/YpIBySMmJsZiGQlarXybqbSBEqLy3heGb8irafUg.jpg "Predictions for what the world will be like in 2100?")
/2023/10/04/image/jpeg/WZEtaGSQLnXMIiErFOdceAlNqEeu5umCff0oKjYp.jpg "This eclipse map provides viewing details for the public")
/2023/10/04/image/jpeg/SxtzxehIbwI00U2r6rYeSmPqpkbMlhr3wMzJ7sn9.jpg "U-Boat Worx unveils Super Sub: World's fastest personal sub")
/2023/10/04/image/jpeg/h6h5SWEgXpzAemqwtvkjofYWmHkNCSsiqDP27zwZ.jpg "Chemistry Nobel Prize goes to quantum dots that guide surgeons")
/2023/09/22/image/jpeg/N09BSEVNiNkMLnA50kD8o3yiMP8ye1mAEFwtWa4G.jpg "Is carbon capture the missing piece in the net zero puzzle?")
/2023/08/27/image/jpeg/fi494iLoRqITLtvrXotofLEJk8JZH1s8Z3Upge1M.jpg "Gallant ships: 5 underdogs of the seas that punched above their weight")
/2023/10/04/image/jpeg/edyTH7wBVGRCw06zSkmj84JbL9Z3QWGytZ93rAtk.jpg "Homo sapiens used another route to migrate out of Africa")
/2023/10/04/image/jpeg/h5yecZdSVOV1OXdxX0EaMs2zqAVRAIy482c6wcIv.jpg "F-22 became \"The Super Raptor\" with these upgrades")
/2023/08/03/image/jpeg/a4peRyX4ysBOqbw4BbDHzMCIbTjXOWejwu83UrQp.jpg "You can hear silence, claim researchers, settling an old debate")