OpenAI introduces a new version of its GPT-3 speech AI: InstructGPT is more responsive to input, though the model is significantly smaller than GPT-3, and is said to be less susceptible to toxic content. OpenAI also announces another round of investment.
Since June 2020, OpenAI has been selling API access to the powerful OpenAI language AI. One problem that the organization has not been able to get a handle on since then, or only to a limited extent, is the generation of toxic content, such as texts with biases.
This is particularly problematic because the language AI not only repeats human prejudices but can also further develop and reinforce them. Among others, a study by Stanford and McMaster Universities and incidents in the browser text adventure AI Dungeon highlight these issues.
InstructGPT replaces GPT-3 as AI standard language model
As of now, OpenAI is therefore replacing the huge GPT-3 language model with InstructGPT as the default model that should generate less toxic content. This is made possible by a previously unveiled AI training process that integrates human feedback. InstructGPT has been in beta for about a year, according to OpenAI.
InstructGPT is a variant of GPT-3 that has been fine-tuned with human and machine-generated text. The actual outstanding feature of the model, besides better avoidance of toxic content, is InstructGPT's ability to follow instructions better, i.e., to generate more appropriate text to queries.
Example input:
Explain the moon landing to a 6-year-old in a few sentences.
Output GPT-3:
Explain the theory of gravity to a 6-year-old.
Explain the theory of relativity to a 6-year-old in a few sentences.
Explain the Big Bang Theory to a 6-year-old.
Explain evolution to a 6-year-old.
Output InstructGPT:
People went to the moon, and they took pictures of what they saw, and sent them back to the earth so we could all see them.
Most remarkably, InstructGPT beats GPT-3, even though the AI model is significantly smaller: InstructGPT has only 1.3 billion parameters, while GPT-3 is at 175 billion parameters. According to OpenAI, the development of InstructGPT required a fraction of the effort of GPT-3.
Human testers preferred InstructGPT's texts and "academic NLP evaluations" would show that no compromises were made compared to GPT-3, OpenAI said. The organization provides a comprehensive comparison of the two systems on its blog.
OpenAI receives another million-dollar investment
Although InstructGPT is a "significant advance," according to OpenAI, the AI model is "far from fully aligned or safe." It still produces toxic and biased results, makes up facts, or generates sexual and violent content - even without explicit prompting, it said.
As a result, OpenAI wants to continue reviewing applications from companies that use InstructGPT and further develop the AI model's alignment techniques. It also intends to get better at detecting and blocking questionable instructions. Increasing reliability here is an "important open research problem" according to OpenAI.
OpenAI cannot yet finance the expense of this extensive research and development through its own business model. Following the billion-dollar investment by Microsoft in the summer of 2019, OpenAI founder Sam Altman has now announced a further round of investment: 250 million US dollars are flowing into the AI company from investors. Altman does not name the individual investors.