Anthropic, an artificial intelligence startup co-founded by former OpenAI executives, has released a new version of its accelerated and cheaper text generation model, available via API Claude Instant.

The updated Claude Instant, Claude Instant 1.2, received the strongest characteristics of the flagship configuration of the Anthropic Claude 2 AI. The startup claims that in this case, achievements in such areas as mathematics, security, logical thinking, and coding were recorded. During internal testing, Claude Instant 1.2 scored 58.7% of coding points compared to Claude Instant 1.1, whose result took 52.8%, and 86.7% on a set of mathematical questions versus 80.9% for Claude Instant 1.1.

The startup announced in its blog that the new version of the AI model can create longer answers with a high level of structuring and follows the formatting instructions more efficiently. Separately, the company notes that Instant 1.2 demonstrates a higher level of efficiency in answering questions, extracting quotes, and multilingual capabilities.

The new version of the machine intelligence model is also characterized by a low degree of susceptibility to hallucinations and resistance to jailbreak attempts. The concept of hallucination in the context of the functioning of large language models means situations when an AI text has no meaning or consists of fictional information that does not correspond to reality in any way. Jailbreak is a method of circumventing security functions by using hints that have a complex and, in a certain sense, manipulative formulation.

In Claude Instant 1.2 there is a context window of the same size as in Claude 2 — 100,000 tokens. The context window refers to the text that the AI model studies before generating additional text material. Tokens represent raw text. For example, the word fantastic was divided into three tokens fan, tas, and tic. Claude Instant 1.2 and Claude 2 can analyze about 75 thousand words, which is approximately equal to the volume of The Great Gatsby.

Experts say that text-generating machine intelligence configurations with large context windows are less likely to erase the contents of recent conversations from memory.

Anthropic currently carries out its activities within the framework of goal-setting, which provides for the desire to create a new generation algorithm for self-learning of artificial intelligence. The relevant information was contained in the startup’s presentation to investors. Such an algorithm could become the basis for creating virtual assistants who can respond to emails, conduct research and generate works of art, for example, in the area of literature. The reality of the corresponding capabilities of digital intelligence was proved by the example of GPT-4 and other large language models.

Claude Instant, as experts note, does not correspond to the characteristics of the mentioned algorithm. In this case, it is more correct to characterize the AI configuration as a competitor to similar entry-level offerings from OpenAI, as well as startups such as Cohere and AI21 Labs. All these companies are developing their own AI systems for text generation, sometimes giving these products the ability to create images.

Anthropic was created in 2021 under the leadership of former OpenAI Vice President for Research Dario Amodei. Since then, the startup has raised $1.45 billion in investment funds. This is an impressive result, but from the point of view of the financial aspect of the company’s plans, the amount is not very large. The startup estimated that over the next two years, it needs $2 billion to create a chatbot.

Anthropic currently has several thousand customers and partners, including Quora, which provides access to Claude and Claude Instant through its generative AI Poe application.

As we have reported earlier, Zoom Announces Anthropic Partnership.