Bitget App
Trade smarter
Buy cryptoMarketsTradeFuturesEarnWeb3SquareMore
Trade
Spot
Buy and sell crypto with ease
Margin
Amplify your capital and maximize fund efficiency
Onchain
Going Onchain, without going Onchain!
Convert & block trade
Convert crypto with one click and zero fees
Explore
Launchhub
Gain the edge early and start winning
Copy
Copy elite trader with one click
Bots
Simple, fast, and reliable AI trading bot
Trade
USDT-M Futures
Futures settled in USDT
USDC-M Futures
Futures settled in USDC
Coin-M Futures
Futures settled in cryptocurrencies
Explore
Futures guide
A beginner-to-advanced journey in futures trading
Futures promotions
Generous rewards await
Overview
A variety of products to grow your assets
Simple Earn
Deposit and withdraw anytime to earn flexible returns with zero risk
On-chain Earn
Earn profits daily without risking principal
Structured Earn
Robust financial innovation to navigate market swings
VIP and Wealth Management
Premium services for smart wealth management
Loans
Flexible borrowing with high fund security
The Reinforcement Gap — or the reason certain AI abilities advance more rapidly than others

The Reinforcement Gap — or the reason certain AI abilities advance more rapidly than others

Bitget-RWA2025/10/05 20:06
By:Bitget-RWA

AI-powered coding tools are advancing at a remarkable pace. For those outside the programming world, it might be difficult to grasp just how rapidly things are evolving. With the introduction of GPT-5 and Gemini 2.5, developers now have access to a fresh array of automated techniques, and Sonnet 4.5 continued this trend just last week.  

Meanwhile, progress in other areas is moving at a slower rate. If you use AI to draft emails, you’re likely experiencing similar results as you did a year ago. Even as the underlying models improve, the end-user experience doesn’t always reflect those gains—especially when the AI is multitasking as a chatbot. While AI continues to advance, the improvements are no longer spread evenly across all applications. 

The reason for this uneven progress is more straightforward than it appears. Coding platforms benefit from billions of quantifiable tests, which help train them to generate functional code. This process, known as reinforcement learning (RL), has arguably been the main engine behind AI’s recent breakthroughs and is becoming increasingly sophisticated. While RL can involve human evaluators, it’s most effective when there’s a clear-cut success or failure metric, allowing for countless iterations without the need for human judgment.  

As reinforcement learning becomes more central to product development, we’re witnessing a clear divide between abilities that can be automatically assessed and those that cannot. Skills well-suited to RL, such as debugging and advanced mathematics, are improving rapidly, whereas areas like writing are seeing only gradual enhancements. 

In essence, a gap is forming due to reinforcement learning—and this divide is quickly becoming a key factor in determining what AI systems are capable of. 

In many respects, software engineering is ideally suited for reinforcement learning. Even before the rise of AI, there was an entire field focused on stress-testing software to ensure reliability before deployment. Every piece of code, no matter how well-crafted, must undergo unit, integration, and security testing, among others. Developers routinely use these checks to confirm their code works, and as Google’s senior director for developer tools recently pointed out, these tests are equally valuable for verifying AI-generated code. Moreover, their systematic and repeatable nature makes them perfect for reinforcement learning at scale. 

Validating a well-composed email or a high-quality chatbot reply is far more challenging, as these tasks are subjective and difficult to measure on a large scale. Still, not all tasks fit neatly into “easy to test” or “hard to test” categories. While there isn’t a ready-made testing suite for things like quarterly financial statements or actuarial analysis, a well-funded accounting startup could likely develop one. Some testing systems will outperform others, and some organizations will devise smarter solutions, but ultimately, the ability to test a process will determine whether it can become a practical product or remain just a promising demonstration.  

Some tasks prove more testable than expected. For example, I would have previously classified AI-generated video as difficult to evaluate, but the significant advancements in OpenAI’s Sora 2 model suggest otherwise. In Sora 2, objects no longer appear or vanish abruptly, faces maintain consistent features, and the generated footage adheres to physical laws in both obvious and subtle ways. I suspect that robust reinforcement learning systems are responsible for each of these improvements. Together, they distinguish realistic visuals from mere digital illusions. 

It’s important to note that this isn’t a fixed law of AI. The current prominence of reinforcement learning in AI development could shift as technology evolves. However, as long as RL remains the primary method for bringing AI products to market, the reinforcement gap will likely continue to widen—impacting both startups and the broader economy. If a process is well-suited to RL, automation is likely to follow, potentially displacing current workers. For example, whether healthcare services can be trained via RL will have major consequences for the economy in the coming decades. And if breakthroughs like Sora 2 are any indication, answers may arrive sooner than we expect.

0

Disclaimer: The content of this article solely reflects the author's opinion and does not represent the platform in any capacity. This article is not intended to serve as a reference for making investment decisions.

PoolX: Earn new token airdrops
Lock your assets and earn 10%+ APR
Lock now!