OpenAI Enhances GPT-4 and GPT-3.5 Knowledge and Memory, Drops Prices and Ups Legal Protection
OpenAI showcased the latest improvements to GPT-4 and GPT-3.5 large language models (LLMs) at its first Dev Day on Monday. The changes include an updated knowledge based, larger context size, and reduced pricing, along with other new features for the LLMs. OpenAI also announced it will begin offering legal protections for copyright issues, following the path of tech rivals Google and Microsoft.
The new GPT-4 Turbo updates the data available to the LLM through April 2023, providing more current event knowledge versus the previous September 2021 cutoff. The LLM has also quadrupled its largest context window to 128,000 tokens, giving it a short-term memory of more than 300 pages of text. The improvements are paired with a price drop that makes the GPT-4 Turbo model considerably more affordable. The input cost is $0.01 per 1,000 tokens, a third of what it was, while the output is priced at half of the standard GPT-4 at $0.03 per 1,000 tokens. GPT-3.5 Turbo receives similar improvements, including expanding its default context window to 16,000 tokens.
OpenAI touted over 30% gains on certain benchmarks from the upgrades. Beyond core model advances, OpenAI revealed several new platform capabilities aimed at streamlining development. For instance, users can ask GPT-4 Turbo to stick to a preferred coding language like JSON in its results. The upgrade also brings in multiple function calling. Function calling allows users to describe how an app or API works to a model and see the model produce code to handle those functions. Now, users can combine multiple functions in a single prompt, allowing one message to include more than one task.
“We’ve heard loud and clear that developers need more control over the model’s responses and outputs, so we’ve addressed that in a number of ways,” OpenAI CEO Sma Altman explained during his presentation. “We have a new feature called JSON load, which ensures that the model will respond with valid JSON. This has been a huge developer request; it will make calling APIs much easier. The model is also much better at function calling. You can now call many functions at once. It will do better at following instructions in general.”
The company also highlighted its new Copyright Shield program, offering legal protections around copyright claims for general platform features. This safety net could encourage enterprise adoption by reducing liabilities. The upgraded language models offer tantalizing new possibilities for developers. But OpenAI restricting access shows it is still treading carefully as risks scale exponentially alongside capabilities. Overall, OpenAI’s latest leap forward aims to maintain its advantage as groups like Anthropic’s Claude close in. Sustaining commercial viability and visionary research in parallel will require a delicate balancing act going forward.