Get live statistics and analysis of Qucheng Gong's profile on X / Twitter

This. What a time for RL people

Behind the Scenes of AI #5 One of our key projects at Kimi is Taste — tuning what the model learns from, and what the model sees as context. Authenticity, authority, timeliness, relevance, everything matters. In the end, data defines the model, and model defines the product.

Behind the Scenes of AI #4 The AlphaZero algorithm looks elegant on paper — until you try to implement it yourself. Starting with random moves, one side quickly becomes too strong. In Go, white almost always wins because of komi. The losing side stops learning (no move works), and the winner also stops improving — since the opponent moves randomly. This is model collapse. To fix it, you must tune the replay buffer distribution and RL hyperparameters carefully. “Don’t teach — incentivize” is easier said than done. The same applies beyond AI: avoid tasks that are too easy or too hard. Progress happens when feedback loops are just right.

yes, many founders are still thinking in the mode of preAI era, such that you only need growth, not monetization. My previous startup had great retention numbers and still failed, so I am asking this question to every startup idea.

This. Often times the problem is not possibility, but affordability. This also applies to many fancy AI products. When offered free users grow fast, but few are willing to pay.

Behind the Scenes of AI #3 As the only strong Go player on the ELF OpenGo team, I coordinated with the Korean Baduk Association and set up evaluation matches — 20 games against 4 top pros. Because of the time difference, I watched every match live into the early morning. In one game, the model made a clear blunder — win rate dropped sharply. My heart stopped. Then, somehow, it recovered on its own and kept a perfect record. Rumors said there was human intervention — there wasn’t. The model fixed itself. That night, AI felt alive. Watching it unfold was surreal. Attached: the official released game values. Notice how one match dips to -0.5 in a single move — then climbs back.

Soumith also gone. Organization and culture of a company is sometimes more important than just hiring capable people with $$$.

Kimi K2 thinking is here

Not yet, until the real AGI, which is hopefully very soon

The 40-20-10 rule still matters. Marketing can buy attention, but not conviction. If retention is weak, you’re solving the wrong problem.

Give your best to every opportunity. The future may be long, but the defining ones only come once or twice.

Not yet, until the real AGI, which is hopefully very soon

Behind the Scenes of AI #5 One of our key projects at Kimi is Taste — tuning what the model learns from, and what the model sees as context. Authenticity, authority, timeliness, relevance, everything matters. In the end, data defines the model, and model defines the product.

This. What a time for RL people

This. Often times the problem is not possibility, but affordability. This also applies to many fancy AI products. When offered free users grow fast, but few are willing to pay.

Behind the Scenes of AI #4 The AlphaZero algorithm looks elegant on paper — until you try to implement it yourself. Starting with random moves, one side quickly becomes too strong. In Go, white almost always wins because of komi. The losing side stops learning (no move works), and the winner also stops improving — since the opponent moves randomly. This is model collapse. To fix it, you must tune the replay buffer distribution and RL hyperparameters carefully. “Don’t teach — incentivize” is easier said than done. The same applies beyond AI: avoid tasks that are too easy or too hard. Progress happens when feedback loops are just right.

Behind the Scenes of AI #3 As the only strong Go player on the ELF OpenGo team, I coordinated with the Korean Baduk Association and set up evaluation matches — 20 games against 4 top pros. Because of the time difference, I watched every match live into the early morning. In one game, the model made a clear blunder — win rate dropped sharply. My heart stopped. Then, somehow, it recovered on its own and kept a perfect record. Rumors said there was human intervention — there wasn’t. The model fixed itself. That night, AI felt alive. Watching it unfold was surreal. Attached: the official released game values. Notice how one match dips to -0.5 in a single move — then climbs back.

Soumith also gone. Organization and culture of a company is sometimes more important than just hiring capable people with $$$.

Kimi K2 thinking is here

The 40-20-10 rule still matters. Marketing can buy attention, but not conviction. If retention is weak, you’re solving the wrong problem.

Give your best to every opportunity. The future may be long, but the defining ones only come once or twice.

yes, many founders are still thinking in the mode of preAI era, such that you only need growth, not monetization. My previous startup had great retention numbers and still failed, so I am asking this question to every startup idea.