All eyes have been on OpenAI as the drama unfolds surrounding the departure of the AI lab’s high-profile CEO Sam Altman. According to The Verge, less than a day ago, there was still a possibility that Altman may return. The situation remains highly fluid and uncertain, but it has massive implications both on the future of Open AI, and the overall speed of technical development in AI.
While plenty of time and resources have focused on the drama surrounding Altman and the board, little attention has been paid on the implications of the future of Open AI as a company. Including some of the competitive pressures it’s facing, and the customer turbulence–many existing customers are concerned about using Open AI products going forward.
Timothy Young, the CEO of AI-writing company Jasper, issued a formal statement to customers saying they are unaffected by Open AI’s current turbulence. Suhail Doshi, the founder of Playground and former founder/CEO of Mixpanel, suggested that anyone using Open AI APIs should be prepared for the possibility of complete downtime. The tweet received nearly a million views at the time of publishing.
And today, at around 9 A.M. PST, Anthropic unveiled Claude 2.1 featuring enhancements such as an expanded context window of 200k characters. .
For those not familiar with Anthropic, I wanted to share a bit about what it is, why it matters, and why Anthropic is a significant threat to Open AI given the current turbulence at the company.
What is Anthropic?
Anthropic was Founded in 2021 and is currently led by Dario Amodei, the former vice president of research at Open AI. In the midst of the Open AI chaos, Anthropic was reportedly approached about a possible merger with Open AI, with Amodei prospectively leading the joint entity. Anthropic’s latest funding round came in large part from Amazon, which escalates its rivalry with Microsoft.
The company made waves in April by open-sourcing Constitutional AI, a technique which constrains model behavior to adhere to specific constitutional principles. More recently, Anthropic introduced Claude, a conversational AI assistant trained to be helpful, harmless, and honest (Claude is Anthropic’s competitor to Open AI’s Chat GPT). Claude’s launch emphasized its rigorous safety testing and transparency about limitations.
With many companies perceived to be rushing unchecked into developing large language models, Anthropic aspires to set a strict ethical standard, which is deeply rooted as the premise of the entire company.
How does Anthropic compare to Open AI, performance-wise?
A key advantage of Claude 2 over GPT-4 is its ability to analyze entire documents up to 100,000 words in length at once*. GPT-4 maxes out at 32,000 words (and the 32k model is not generally available). This allows for analysis of very large documents without chunking and retrieval (RAG) pipelines, which are burdensome to create and optimize. The expanded context alone allows for new applications that would likely exceed the quality of GPT-4, just given the ability to put all source data into the context. Otherwise, the differences are more nuanced.
According to testing by akkio.com:
- GPT-4 edges out on reading, logic and math reasoning.
- Claude 2 exceeds on coding, writing, and its unmatched full document ingestion capacity.
- That full document mastery gives Claude 2 an architectural edge to expand capabilities beyond GPT-4 using real-world data rather than limited snippets.
- Both are incredibly advanced AI systems pushing boundaries. However, Claude 2’s ability to process larger documents, coupled with the AI’s rapid advancement, gives it a path to potentially surpass GPT-4 as the most capable language model over time.
Overall, GPT-4 is a higher performing model as of today, but Claude 2 is close and is improving quickly.
Implications
With the turbulence unfolding at Open AI, and with no clear path to resolve, many startups may start looking at Anthropic as a viable alternative to build their key infrastructure on. This has potential implications not only for Open AI, but for Microsoft, Amazon, and the entire AI ecosystem. Keep a close eye on this burgeoning competition–it is far from fully playing out. New entrants are likely to be funded, and the LLM wars appear far from done.
*At the time of publishing, I do not have benchmark tests available for Claude 2.1, which was announced today, November 21st, 2023. This newest version has an expanded context window of 200k characters.
Read the full article here