Realtime AI News
Yao Shunyu fields questions on Hunyuan 4's compute, leaderboard scores and RSI
In a public Q&A published by Geek Park, Yao Shunyu took questions on the Hunyuan 4 model covering compute spending, the credibility of leaderboard results and RSI, or recursive self-improvement. The three topics map onto the questions outsiders now ask of any flagship Chinese model: whether the resources are there, whether the scores are real, and how far the model shapes its own next iteration.
Discussion of Hunyuan 4 did not stop at where it sits on a leaderboard. In a public Q&A published by Geek Park, Yao Shunyu put three often-skimmed topics on the table: compute, leaderboard gaming and RSI.
The three keywords point at the same underlying question, which is how the model was actually built. Compute covers the scale and durability of the resources behind training; leaderboard gaming covers whether public benchmark scores represent real capability; RSI, or recursive self-improvement, covers how far the model participates in its own next iteration. No team building a flagship model can avoid any of them.
The compute question is especially sharp for Chinese models. Training at the frontier means weeks or months on a cluster, and the number of available accelerators, the interconnect bandwidth between them and the power feeding them together decide whether the next round can start on schedule. A team that discusses this openly is calibrating outside expectations.
Leaderboard gaming became a loaded phrase because gaps between public scores and real-world behaviour have recurred over the past two years. Tuning for a specific test set or picking a favourable evaluation configuration can produce attractive numbers on a model that is not pleasant to use. A team willing to explain how it treats rankings is implicitly conceding that evaluation needs to track real tasks more closely.
RSI is the longer-horizon and more contentious piece. Having models filter data, synthesise training corpora or even propose changes to the training pipeline can accelerate iteration without adding headcount, but it turns the question of whether capability growth stays controllable into one that has to be answered rather than deferred. Raising it in a public Q&A signals a willingness to be questioned at the method level.
The value of the session is less in the conclusions than in naming the criteria by which a flagship model should be judged: compute sets the ceiling, evaluation shapes how outsiders understand capability, and RSI shapes the long-run curve. The next thing to watch is whether those points turn into checkable detail, whether that means disclosing the order of magnitude of training scale, publishing evaluations closer to real usage, or describing where RSI actually sits in the pipeline.
Why it matters
For Tencent, addressing all three topics moves the Hunyuan 4 conversation from rankings back to engineering and resources; for the wider field, it raises expectations that flagship models will explain training scale, evaluation method and the limits of automated iteration.
Nearby Updates
All11/22, 10:11
Alibaba's Qwen4-Max rollout hits a compute wall
Alibaba's rollout of Qwen4-Max is running into limits on available compute, according to a report from ad-hoc-news.de, with capacity now shaping how quickly the flagship model reaches users. The constraint points to deployment rather than capability: the model exists, but serving it at scale depends on clusters that cannot be expanded overnight.
11/22, 08:23
Microsoft rolls out OpenAI's GPT-6.2 Sol to Copilot worldwide as MAI models expand
Microsoft has brought OpenAI's GPT-6.2 Sol to Copilot for users worldwide, according to Startup Fortune, while its in-house MAI models continue to push into the same products. The report frames the move as a model-level shift inside Copilot rather than the launch of a single named feature.
11/22, 11:56
SpaceX Lenders Want the Nvidia Chips Kept on Earth, Not in Orbit
Lenders who financed SpaceX's purchase of Nvidia AI chips have reportedly sought restrictions meant to keep that hardware on the ground rather than let it follow the company's orbital data center ambitions into space. The concern is that collateral becomes nearly impossible to repossess once it is in orbit, a risk one report frames as a $1.5 billion exposure.
11/22, 11:59
Citi, PNC and Chase Rewrite Terms So AI Agent Losses May Fall on Customers
Citi, PNC and JPMorgan Chase have updated their account terms to set out who bears the loss when a third-party AI agent touches a customer's money. The terms suggest transfers can still count as authorized even when an agent ignores instructions or when leaked credentials are used by someone else.