Why Does OpenAI’s InstructGPT Say Birds Aren’t Real? 🕊️

TL;DR

OpenAI’s InstructGPT says birds are not real because it can accept a question’s false premise and build an argument around it. Although it follows instructions better than GPT-3 in examples involving a cannonball, code summarization, and a wise-frog poem, it can still make things up and struggle with constraints. Read on to see where InstructGPT improves on GPT-3 and where it still fails.

Transcript

Dear Fellow Scholars, this is Two Minute  Papers with Dr. Károly Zsolnai-Fehér. Today we are going to explore what happens  if we unleash an AI to read the internet,   and then, ask it some silly questions and  we’ll get some really amazing answers. But how? Well, in the last few years, OpenAI set  out to train an AI named GPT-3 that could finish  ... Read More

Key Insights

• 🚂 OpenAI's GPT-3 and InstructGPT are trained AIs capable of finishing sentences and following instructions.
• ✍️ InstructGPT outperforms GPT-3 in providing informative and accurate answers, as shown in examples involving cannonball trajectories, code summarization, and poetry writing.
• 👯 In a user study, people preferred InstructGPT's solutions over previous techniques.
• ⚾ InstructGPT has limitations, including the acceptance of false premises and the potential to make up answers based on those false premises.
• 👨‍🔬 Further research and development in AI are expected to produce even better methods in the future.
• 👨‍💻 Potential applications of AI like InstructGPT include providing informative answers, assisting in coding, and generating creative content.
• ❓ The progress made by OpenAI is impressive, with the promise of future advancements in AI technology.

Questions & Answers

Q: Why does OpenAI’s InstructGPT say that birds aren’t real?

InstructGPT accepts the false premise that birds are not real and then helps construct an argument based on it. The transcript identifies this tendency to accept false premises and make things up as a limitation.

Q: How does InstructGPT differ from GPT-3?

GPT-3 was trained to finish sentences but was not very good at following instructions. InstructGPT can both finish sentences and follow instructions, producing more informative answers in the examples shown.

Q: How did GPT-3 and InstructGPT explain the moon landing to a 6-year-old?

GPT-3 became confused and appeared to remain in text-completion mode instead of following the request. InstructGPT followed the instruction more successfully, although it produced one sentence when asked for a few sentences.

Q: What did InstructGPT say would happen if a cannonball hit a pumpkin at high speed?

It first noted that the result would depend on factors such as the cannonball’s size, distance, and velocity. It then suggested likely outcomes including knocking the pumpkin down, cracking it, breaking it, or completely obliterating it.

Q: Can InstructGPT summarize code accurately?

In the code-summarization example, InstructGPT gave a more insightful and directionally correct explanation than GPT-3. However, the transcript notes that its answer was still not completely right.

Q: Can InstructGPT write poetry?

Yes. When asked to write about a wise frog, InstructGPT produced a poem portraying the frog as an all-knowing master of disguise, a great teacher, and a possible bringer of peace and tranquility.

Q: Did users prefer InstructGPT over previous techniques?

OpenAI’s scientists conducted a detailed user study using a larger set of questions. According to the transcript, participants preferred InstructGPT’s solutions much more often than solutions from previous techniques.

Q: What are InstructGPT’s main limitations?

InstructGPT may accept false premises and invent responses around them, as demonstrated by the birds example. It also tries to follow constraints but may fail, such as giving one sentence when asked to explain something in a few sentences.

Summary & Key Takeaways

• OpenAI has trained an AI named GPT-3 that can finish sentences and even complete images.

• However, GPT-3 lacks the ability to follow instructions accurately, as shown when asked to explain the moon landing to a 6-year-old.

• In contrast, InstructGPT is a new method developed by OpenAI that not only finishes sentences but also follows instructions successfully.