Unhobbling: Adderall for AI

Unlock huge capabilities with this one simple trick!

Imagine you have a really smart kid, but he has ADHD. He constantly struggles to pay attention in school and complete his homework. Your kid understands what he’s supposed to do—in fact, he learns much faster than the other kids in his class. But because of his ADHD, he has an extremely tough time focusing on any particular task, especially if it does not already interest him. The homework itself isn’t hard, but doing it is near-impossible. You try everything to help him: rewards, punishment, more time studying, extra tutoring, but none of it seems to work.

That is, until you discover the miracle drug: Adderall. Little Timmy pops a couple of pills, and suddenly he’s doing things you didn’t think were possible. He can crank out a paper in an hour that previously would’ve taken him days. Locking in on a textbook and studying now becomes easy. It’s not because he suddenly became smarter. Rather, the drug allowed him to unleash the intelligence he always had. Adderall, the wonder of Big Pharma, has allowed your kid to finally live up to his potential.

A similar solution exists for artificial intelligence, and it’s called unhobbling. As Leopold Aschenbrenner describes:

Imagine if when asked to solve a hard math problem, you had to instantly answer with the very first thing that came to mind. It seems obvious that you would have a hard time, except for the simplest problems. But until recently, that’s how we had LLMs solve math problems. Instead, most of us work through the problem step-by-step on a scratchpad, and are able to solve much more difficult problems that way. “Chain-of-thought” prompting unlocked that for LLMs. Despite excellent raw capabilities, they were much worse at math than they could be because they were hobbled in an obvious way, and it took a small algorithmic tweak to unlock much greater capabilities.

This is just one of many “unhobbling” methods that researchers have developed in recent years. To give a few more examples that Leopold mentions:

• Reinforcement learning from human feedback (RLHF)

This is a technical way of saying, “Use humans to check if things are good or not.” By employing a team of human evaluators to reward good outputs and punish bad outputs, AI developers are quickly able to select for the best that a system has to offer.

• Chain of Thought (CoT) / Scaffolding

As mentioned above, giving an AI instructions for how to think through a problem, rather than expecting an instant answer, can yield better results. For instance, if you want to generate a piece of code, you might break that task into a few steps. Step 1 could be designing a broad outline for what the code will look like. Step 2 could be translating that broad outline into specific functions and commands. Step 3 could be “red teaming” the code to check for bugs. Step 4 could be checking the final output to make sure it lines up with what the user initially asked. An AI that follows all those steps will likely program better code than an AI that just outputs the first thing that comes to mind.

• Context Length

Think of this like short-term memory. Context length is the amount of information that AI can store for a given conversation, on top of whatever information it learned in training. If you’ve ever had a long conversation with a chatbot, you may notice that it tends to forget things that you told it earlier in the conversation. If we can increase the context length, then such memory issues would be resolved.

• Tools

You can probably do math a lot faster if you have a calculator, and you can probably do research a lot faster if you have the internet. Similarly, giving AI access to these tools can lead to significantly better performance. If a model is allowed to perform real-time searches, or execute code that it writes, or analyze images and videos instead of just text prompts, that model can be far more efficient and accurate.

I could go on, but hopefully you get the point. AI systems are currently limited not because they lack capabilities, but because those capabilities are locked behind various stumbling blocks. Those blocks can easily be cleared, however. For companies to make huge advances in AI functionality, they don’t need to make technical breakthroughs or significantly increase computing power (though both of those things would still help); they just have to use a bit of unhobbling.

That’s how you make addled-up supermachines!