@Sxy_Cherotich: Recently I've been talking with quite a few model researchers, and a consensus conclusion is: the importance of data is once again highlighted. A while ago I got to know ex-Kimi's @FanqingMengAI, who is doing a startup in the data direction, and invited him to record a podcast. The biggest non-consensus from our conversation is Fanqing's view on the difference between domestic and foreign models...
Summary
A podcast about AI model competition, discussing the importance of data, distillation and pre-training innovation, and an interview with Evolvent AI co-founder Meng Fanqing, covering topics such as synthetic data, RSI, and differences in domestic models.
View Cached Full Text
Cached at: 08/10/26, 01:31 PM
Recently I’ve been talking with quite a few model researchers, and a consensus emerged: the importance of data is once again taking center stage.
A while back I got to know @FanqingMengAI, an ex-Kimi person now founding a startup in the data space, and invited him to record a podcast episode.
The biggest non-consensus to come out of our conversation was Fanqing’s assessment of the differences between domestic and overseas models:
We had originally assumed that domestic models would hold the advantage in the more engineering-focused Post-training; but in his view, what truly sets domestic models apart is actually the architectural innovation forced out during the Pre-training stage by resource constraints.
Based on this understanding, we also reached an optimistic conclusion:
Many people say Chinese models are heavily dependent on distillation, but Fanqing believes distillation is just an acceleration tool, not the decisive factor in domestic models getting stronger. Even if overseas players were to shut down all distillation channels one day, Chinese models could still build their own competitiveness through pre-training innovation.
This episode also answered a few of my own questions:
- Why has synthetic data been so profitable lately, why don’t model labs do it in-house, how long will demand for third-party data last, or what changes will it see;
- Are we now at the point where “algorithms are data, data is Infra, and Infra is algorithms”?
- How should we understand concepts like the recently hyped AI for AI, self-evolving, and RSI — and if these concepts hold up, why isn’t everyone just training better base models, and what will the future look like.
From Distillation to Synthetic Data to RSI: What’s the Next Battleground in Model Competition? | A Conversation with Evolvent AI Co-founder Meng Fanqing
Source: https://podcasts.apple.com/cn/podcast/%E4%BB%8E%E8%92%B8%E9%A6%8F%E5%88%B0%E5%90%88%E6%88%90%E6%95%B0%E6%8D%AE%E5%88%B0-rsi-%E6%A8%A1%E5%9E%8B%E7%AB%9E%E4%BA%89%E7%9A%84%E4%B8%8B%E4%B8%80%E4%B8%AA%E7%84%A6%E7%82%B9%E6%98%AF%E4%BB%80%E4%B9%88-%E5%AF%B9%E8%B0%88-evolvent-ai-%E8%81%94%E5%88%9B%E5%AD%9F%E7%B9%81%E9%9D%92/id1700299886?i=1000781075034
Event Preview 🥳: On August 22, we’ll host an online session with Fanqing, joined by dozens of model researchers. Be sure to scroll to the end of the shownotes for registration info!
Over the past few months, we’ve talked with many people working on models and algorithms, and one takeaway from those conversations is: the importance of data is once again coming to the fore.
The market has shown clear corresponding feedback — for example, a batch of synthetic data companies have generated substantial revenue in a very short time; and both model labs and big internet companies are investing more heavily in data.
For this episode, we invited Meng Fanqing, co-founder of Evolvent AI, who caught this wave of synthetic data opportunity early. Fanqing is a PhD at NUS born in 2001, and spent over a year at Kimi working on Agent and RL Infra.
In the episode, we started from his different experiences doing Post-training at a model lab versus founding a startup, and then walked relatively thoroughly through post-training, data, RSI/Self-Evolving/AI for AI, the competitive landscape for domestic models, and even the relatively sensitive topic of distillation. The biggest non-consensus to emerge was Fanqing’s assessment of the differences between domestic and overseas models:
We had originally assumed that domestic models would hold the advantage in the more engineering-focused Post-training; but in his view, what truly sets domestic models apart is actually the architectural innovation forced out during the Pre-training stage by resource constraints.
Based on this understanding, we also reached an optimistic conclusion:
Many people say Chinese models are heavily dependent on distillation, but Fanqing believes distillation is just an acceleration tool, not the decisive factor in domestic models getting stronger. Even if overseas players were to shut down all distillation channels one day, Chinese models could still build their own competitiveness through pre-training innovation.
Finally, Evolvent AI recently released a new research output, RSIBench-Data. If any friends would like to connect on data or RSI topics, Fanqing is happy to hear from you.
【Human Museum】
Guide: Qu Kai, founder of 42章经
Exhibit No. 51: Meng Fanqing, co-founder of Evolvent AI; former Kimi RL intern focused on RL Infra, involved in major model releases including Kimi K2.5 and Kimi Linear
【Time Machine】
Part 1 Post-training and Model Competition
- 01:15 How does it feel different being at Kimi vs. founding a startup?
- 02:36 Why do people with Post-training backgrounds find it easier to start companies? What are the possibilities?
- 04:37 How have today’s Bench and Eval changed compared to the past?
- 07:12 Post-training = the barrier to doing data work and training models is getting lower?
- 10:35 When everyone’s working on data, what separates the quality gap?
- 12:48 Domestic models vs. overseas models — is the real advantage actually in Pre-training?
- 17:09 How will the competitive landscape for Chinese models evolve? What’s the final battleground?
Part 2 RSI
- 21:50 RSI, Self-Evolving, AI for AI, Auto-Research… How should we understand these terms?
- 24:35 If RSI really gets going, will models really devour everything?
- 29:09 Assuming models can truly self-iterate, why don’t startups just build a better base model?
- 30:39 What problems do you essentially need to solve to do RSI well?
- 33:48 As RSI continues to develop, what kinds of changes will it trigger?
Part 3 Data
- 35:29 What waves of change has models’ demand for data gone through?
- 39:07 Model labs can generate data themselves — why still source it externally?
- 40:36 Where might the next wave of opportunity in data lie?
- 43:00 Are we now at the point where “algorithms are data, data is Infra, and Infra is algorithms”?
- 46:11 How should we understand the relationship between synthetic data and distillation? Where do distillation techniques differ?
Part 4 Outlook
- 48:18 How much headroom remains in the current technical approach to models?
- 49:33 Why has AI4S recently blown up?
- 51:39 Long term, do you favor startup model companies over big tech?
- 53:58 Just how significant is distillation for domestic models’ catch-up effort?
- 56:18 From synthetic data to RSIBench-Data, what does Evolvent AI ultimately want to become?
【Event Preview 🥳】
On August 22, we’ll have Fanqing host an online session, with dozens of model researchers joining. If you’re interested, feel free to click the link or scan the QR code below to meet and exchange ideas!
【The gang that made this happen】
- Producer: Chen Pi
- Editor: Chen Pi
- Opening theme: Mondo Bongo - Joe Strummer & The Mescaleros
- Ending theme: GeGeGe - 光のサイン
View this episode’s transcript on Xiaoyuzhou
Similar Articles
@vivilinsv: A serial entrepreneur I particularly like and admire, @quxiaoyin Xiaoyin, has been on fire on X recently. She made a very sharp judgment: Chinese open-source models will continue to gain market share, and may even become one of the "worst-case scenarios" for the US AI ecosystem—if Chinese models not only occupy the model layer...
Discusses how Chinese open-source models could become a threat to the US AI ecosystem, and the future competitive landscape between open-source and closed-source models, sparking widespread discussion in the AI community.
@Phoenixyin13: This latest blockbuster paper from Meta FAIR aims to tell the AI industry an important bellwether: "Large model data is ushering in the era of intelligent scientists." In this paper, a 4B small model precisely refined by Autodata not only crushes the same-scale models trained with traditional synthetic data on legal reasoning tasks, but also...
Meta FAIR's latest paper proposes the Autodata method, which uses an intelligent data scientist Agent to autonomously generate and optimize high-quality data, enabling a 4B small model to defeat a 397B large model on legal reasoning tasks. This indicates that data quality can bridge the gap in parameter count, providing new insights for data pipelines and scaling.
@10xmylife: I try every domestic model update, and one trend is becoming clearer: each version update of Kimi, GLM, and DeepSeek is better than the last. I think this indicates that the domestic model R&D system is entering a virtuous cycle: more users bring richer real data; engineering infrastructure is maturing; compute power is still tight but forces optimization; pricing advantages are clear.
The author observes that domestic AI models (Kimi, GLM, DeepSeek) improve with each update, believing that domestic model R&D has entered a virtuous cycle: user growth brings abundant data, mature engineering infrastructure, tight compute power forcing efficiency optimization, and clear pricing advantages.
@aigclink: Recently participated in training a vertical medical model in the laboratory department of a top domestic hospital. Sharing some personal views and advice for those working in AI+Healthcare (take it if useful, leave comments if not):
The author shares personal views from participating in training a vertical medical model at a top domestic hospital, highlighting core challenges such as medical data not leaving the hospital, high cost of on-premises deployment, and weak willingness to pay, and suggests partnering with hardware vendors. They also note that general medical models (like Baichuan) already perform well.
@FeitengLi: a16z's latest podcast is titled "AI, the Internet in 1997, and What Happens Next," featuring guest Benedict Evans—former a16z partner, author of the annual must-read "AI Eats the World" for countless investors and executives, widely regarded as one of the most widely read analysts in tech.
a16z podcast invites former partner Benedict Evans to explore the analogy between AI and the internet in 1997, pointing out that current AI infrastructure investment is massive but ROI is unclear. Historical experience suggests value will shift upward, and models themselves will find it difficult to achieve differentiated profits.