Cognition founder: AI has surpassed humans in pure reasoning; our final advantage lies in memory retrieval.

According to AIMPACT, on May 11th (UTC+8), according to Beating's monitoring, Colossus magazine published a three-hour interview with Scott Wu, the founder of Cognition. In the interview, Scott Wu discussed a question that few people think about: What is it that humans are better at than AI now? His answer is not reasoning—AI has already won in pure logical reasoning. What humans are truly better at is retrieval, that ability to "suddenly remember." He used mobile phone verification codes as an example. Two-step verification codes are 6 digits; if they were 10 digits, most people probably wouldn't remember them. The "context window" of the human brain is that short. But the human brain has something that AI can't do: you open a code file, and suddenly a related bug you fixed four months ago pops into your mind—how it happened, how you fixed it, why you changed it that way—all of it comes rushing back. Wu calls this "soft context," not word-for-word memorization, but the experience you've accumulated over the years, precisely pulled out by a certain clue. He said that he makes decisions in Cognition every day, not by deducing a logic problem from scratch, but by automatically recalling implicit judgments accumulated over a long period of time, such as "Why was it designed this way in the first place?" and "After that new technology came out, we decided to migrate to another framework." AI's context window is much larger than that of humans, but it is still far from being able to perform this kind of cross-time and cross-scenario associative retrieval. AI is catching up along two paths. One is embedding search, which retrieves relevant fragments from massive amounts of historical data, a direct imitation of the human brain's associative memory. The other is more radical, called continuous learning, which allows AI to update its own weights as it works, similar to the neuroscience concept of "neurons firing together connecting." Wu believes that the second path is closer to the actual operation of the human brain, but it is still in its early stages. Near the end of the interview, he said, "In the end, doing AI is actually about deciphering cognition itself. For thousands of years, humans have been constantly speeding up various things, but one thing has never changed: there has to be someone to think and invent. Now even this premise may be changed." (Source: ME)