Language Learning is the Killer App for LLMs

Lately, my kids have been getting more and more into video games. A few nights a week we will sit down together and play games on our Switch. Every once in a while, I show them some old Super Nintendo games that I played growing up. It has been an absolute joy to share these games with them and I was particularly excited when they asked me to play Super Mario World 2: Yoshi’s Island.

Yoshi’s Island is a game where you control Yoshi and help save Baby Mario from Baby Browser. The game has lots of really fun elements to it, but perhaps one of the most unique aspects to this game is that Yoshi can transform into different vehicles that give him special powers. He can turn into a helicopter and fly around. He can turn into a car and reach places that “regular” Yoshi wouldn’t fit into. He can even turn into a train that follows tracks and collects items.

Trains are a popular toy in our house, so I wanted to show my kids that Yoshi can transform into a train. The problem is that Yoshi can only transform in certain levels and I hadn’t played this game in years so I couldn’t remember which level the transform was in. So I took out my phone, opened ChatGPT and asked:

In yoshi’s island on SNES which level can you turn into a train

ChatGPT replied with:

In Yoshi’s Island for the SNES, you can turn into a train in several levels. One of the most notable levels is World 2-3, titled “What’s Gusty Taste Like?” In this level, there is a section where Yoshi can transform into a train to navigate through the level on tracks.

Great! We loaded up the level, played through and looked for the transform. We looked. And looked. And looked. There was no train transform in World 2-3¹.

ChatGPT was wrong.

This wasn’t the first time ChatGPT had given me an incorrect answer and I’m sure we’ve all had moments where ChatGPT or Claude or some other LLM has given us a response that seems correct but is actually wrong. This is annoying at best if the facts are important, but from the perspective of someone learning English, the answer is great because it is gramatically correct, uses natural language, and depending on fluency level certainly qualifies as comprehensible input.

Comprehensible input is a language learning hypothesis that asserts that it is possible to acquire/learn a foreign language simply by being exposed to content that is at or just above your current comprehension level. An example of comprehensible input for someone just starting to learn English might be a simple sentence like “I am going to the store.” For a more advanced student, comprehensible input might look like a short story or even a novel intended for native speakers.

The tricky part of comprehensible input is actually finding content that is comprehensible. Traditionally, a language learner would solve that problem by scouring YouTube, buying some sort of course or app, or hiring a personal language tutor. LLMs, on the other hand, flip this process upside down because language learners no longer have to worry about finding comprehensible input. They simply have to prompt their favorite LLM for comprehensible input.

Let’s say I’m a student learning English. Let’s assume further that I’m at a fairly decent conversational level but I really struggle with talking about sports, particularly American Football. If I want to get better at talking about American Football, my options are somewhat limited. I could find websites that cover football, but the content might be too far above my level. I could hire a tutor but that is not going to be cheap. I could try to find an online course that specifically covers American Football, but that might be too niche. Another option I have would be to go to ChatGPT and ask it to generate comprehensible content about American Football for me:

You are an ESL teacher. You teach American English. I am your student. I am at a B2 level. I want to be able to understand newspapers and websites about American Football. Generate me a short story that contains vocabulary and phrases related to American Football.

This will generate a short story² that will include themes, phrases, and vocabulary related to American Football. At this point, you still might find the story too difficult. In that case, you can even ask ChatGPT to simplify the story:

I don’t understand this story. Can you simplify it?

The result is a story that is an even closer match to your level. By doing this, you will drastically speed up the language learning process because you will always be able to generate and refine comprehensible input in your target language.

  1. ¹ World 2-3 has a car transform

  2. ² It was the final quarter, and the score was tied. The Lions had possession of the ball on their opponent’s 35-yard line with only 30 seconds left. Their quarterback, Tom Reynolds, knew this was the team’s last chance to avoid overtime. He took the snap, dropped back, and scanned the field.

    The defense had a strong pass rush coming, so Tom quickly dumped the ball off to his running back, who gained 8 yards before being tackled. The Lions called their final timeout to stop the clock.

    With just 15 seconds remaining, Tom huddled the offense. The coach had called for a deep pass, aiming to put them within range for a field goal. Tom took the snap and immediately saw the blitz coming. He stepped up into the pocket and fired a long pass toward the sideline. His wide receiver, James, was running a fade route, and with a perfect leap, James made the catch, bringing them down to the 10-yard line.

    Now, with only 5 seconds left, the kicker ran onto the field. The Lions lined up for a 27-yard field goal. The snap was good, the hold was steady, and the kick sailed through the uprights. The Lions won the game as the stadium erupted with cheers.

  3. ³ It was the last part of the game, and the score was tied. The Lions had the ball on the other team’s 35-yard line, with only 30 seconds left. Their quarterback, Tom Reynolds, knew this was their last chance to win without going into extra time. He got the ball and looked for a teammate to pass to.

    The other team was trying hard to tackle him, so Tom quickly passed the ball to the running back, who ran 8 yards before being stopped. The Lions used their final timeout to stop the clock.

    Now, with only 15 seconds left, Tom talked to his teammates. The coach wanted Tom to throw the ball far to get closer to the goal. Tom got the ball again and saw the other team rushing at him. He moved forward and threw a long pass to the wide receiver, James, who was running toward the edge of the field. James jumped and caught the ball, getting them to the 10-yard line.

    With just 5 seconds left, the kicker came in to try a 27-yard field goal. The ball was snapped, held, and kicked perfectly, going through the goalposts. The Lions won the game, and the fans cheered loudly.