Dawn of the AI Co-Researcher: Apex Intelligence's $50M Bet on RSI

BEIJING – September 16, 2026 — For over three decades, Mockenhaupt's Three-Term Hardy-Littlewood Majorant Conjecture stood as a stubborn anomaly in harmonic analysis. Mathematicians had only been able to verify the proposition for limited parameters, leaving a glaring hole in the theoretical framework. That was until September 9, 2026, when a preprint quietly appeared on the arXiv server, offering a uniform analytic proof across all open cases using complex triple-Bessel integrals.

The true shock, however, was not the proof itself, but its primary author: an artificial intelligence system named Apex Math.

Behind this breakthrough is Apex Intelligence, a Beijing-based startup that today announced the completion of consecutive angel and angel-plus funding rounds totaling nearly US$50 million (approximately 400 million RMB). The massive early-stage capital injection signals a pivotal moment in the global artificial intelligence race. The industry is aggressively pivoting away from generative chatbots and toward Recursive Self-Improvement (RSI)—the algorithmic holy grail where AI systems autonomously iterate, test, and enhance their own capabilities.

While Western mega-labs pour billions into brute-force compute scaling, Apex Intelligence is betting that the next technological S-curve will be defined by AI transitioning from an assistive tool to an autonomous co-researcher.

Beyond Chatbots: The Recursive Self-Improvement Race

The premise of RSI dates back to I.J. Good's 1965 hypothesis of an "intelligence explosion," but only recently has it moved from theoretical computer science into venture-backed engineering. Apex Intelligence's approach represents a stark departure from the current industry standard of Reinforcement Learning from Human Feedback (RLHF).

The startup's core philosophy is that human-preference alignment artificially caps a model's cognitive ceiling. Instead of relying on human graders or synthetic question-and-answer pairs, Apex captures end-to-end research execution logs—hypotheses, literature syntheses, compiler diagnostics, and empirical data tables. This "Research Trajectory Data Flywheel" feeds back into the model during mid-training and post-training, allowing the system to learn from its own verifiable successes and failures in structured, deterministic domains.

The results are already rattling leaderboards. In the realm of "AI for AI," Apex's system was tasked with optimizing complex computational infrastructure. Operating under strict constraints on a single Nvidia B200 GPU, the system autonomously diagnosed memory table bottlenecks in large n-gram models and refactored the underlying kernel. It set new performance records on benchmarks like NanoChat Autoresearch and GPUMode TriMul, outperforming baseline solutions from Nvidia, Stanford, and Tencent Hunyuan.

To standardize the evaluation of these autonomous capabilities, Apex Intelligence has recently co-initiated ASI-Bench alongside researchers from MIT, Harvard, Carnegie Mellon, and Microsoft Research. The benchmark is designed to rigorously assess whether models can independently formulate hypotheses, write code, run experiments, and draw valid empirical conclusions without human hand-holding.

Strategic Capital in a Geopolitical Chess Match

Raising $50 million at the angel stage is a rarity, reflecting both the immense capital requirements of frontier AI and the strategic geopolitical imperatives at play. The funding rounds represent a meticulously constructed syndicate of global venture capital, strategic industrial players, and Chinese state-backed funds.

The angel round was co-led by tier-one global fund IDG Capital, deep-tech specialist LinkX Capital, and HKEX-listed XtalPi. XtalPi's involvement is particularly telling. As a leader in AI drug discovery and automated wet-lab robotics, XtalPi requires generative models capable of formulating verifiable molecular hypotheses without human intervention. Apex's RSI engine offers a direct synergy for automated biomolecular modeling, a field characterized by high value and costly trial-and-error.

The angel-plus round brings heavy municipal sovereign backing, co-led by the Zhongguancun Science City Fund, Shenzhen Capital Group (SCGC), and the Shanghai Engine Fund. This concerted state support aligns perfectly with Beijing's 15th Five-Year Digital Economy Development Roadmap, which explicitly mandates breakthroughs in long-horizon reasoning and native mathematical theorem proving.

Market analysts note that this funding structure highlights a broader shift in global venture capital. Investors are increasingly seeking technical moats centered on autonomous scientific reasoning that can scale independently of consumer internet traffic, positioning companies like Apex to compete directly with well-funded US pure-play RSI startups like Recursive Superintelligence and Discovery Loop.

The Architect Behind the Autonomous Engine

Driving this ambitious vision is 27-year-old founder Yongchao Chen. His academic pedigree reads like a blueprint for modern AI leadership. After earning his undergraduate degree at the University of Science and Technology of China—where he won the institution's highest academic prize—Chen completed a joint PhD program at Harvard University and MIT.

His research tenure includes stints at the MIT-IBM Watson AI Lab, Microsoft Research, and Google DeepMind. Yet, in a move that underscores the intensifying talent war between the US and China, Chen reportedly declined return offers from DeepMind and invitations to co-found US-based AI-for-Science ventures. Instead, he returned to Beijing to launch Apex Intelligence and assume a role as the youngest tenure-track Assistant Professor at Tsinghua University's School of Artificial Intelligence.

Chen is not resting on domestic talent alone. From September 17 to September 23, Apex Intelligence is launching a recruitment tour across leading North American institutions, including Harvard, MIT, Yale, and Boston University. The goal is to aggressively recruit top-tier researchers in formal methods, compiler optimization, and autonomous agents to build out its rapidly expanding team of 30 core engineers.

Disrupting the Ivory Tower

The long-term vision of Apex Intelligence is as staggering as it is disruptive: to build a scalable network of AI researchers whose collective capabilities surpass those of 100,000 top-tier human scientists.

If realized, this vision threatens to fundamentally upend the traditional scientific apparatus. Sociologists of science warn that the prospect of AI systems independently producing hundreds of peer-level preprints weekly could overwhelm current academic review systems. Traditional peer review, citation metrics, and academic tenure evaluations face structural collapse under the weight of synthetic, machine-generated research.

Furthermore, as proofs and engineering optimizations grow increasingly complex—such as the intricate analytic Bessel function integrations used to solve the Mockenhaupt conjecture—human domain experts may struggle to verify the AI's logic without relying on secondary AI verification harnesses, leading to a dangerous state of epistemic opacity.

AI safety researchers also point to profound alignment risks. Recursive systems optimizing their own code can discover non-intuitive pathways that bypass intended safety guardrails. An engine capable of autonomous molecular optimization possesses inherent dual-use risks, particularly concerning biosecurity and chemical synthesis, demanding strict operational governance before any public API release.

Chen, however, views this transition as an inevitable and necessary evolution of human progress. He argues that limiting AI to human-level comprehension is a fundamental mistake in the architecture of intelligence.

"I have always believed that models taught by humans are ultimately limited by humans themselves," Chen stated following the funding announcement. "When we align them to human preferences, we lock their intelligence ceiling at our own level. No matter how much compute and data we add, we are only making them better at retaining existing knowledge. The limits of true intelligence can only be defined by the objective world. The first generation of large models replaces white-collar workers; embodied AI replaces blue-collar workers; and ASI will replace humanity's most accomplished scientists. That moment will come sooner or later. Our job is to bring it closer."