Zhao Wenyuan

Biodata

Feature	Details
Name	Zhao Wenyuan (赵文渊)
Alias(es)	Dr. Zhao; President Zhao; Wenyuan; Director Zhao 66 142 199 311
Gender	Male
Affiliation	Origin, previously Source Code Technology / Source Intelligence Technology; Origin Beijing Research Institute, formerly Source Intelligence Beijing Research Institute; formerly CodeSafe, Dingsheng, and Google/DeepMind 66 142 157 193 311 406
Occupation/Role	AI researcher; CodeSafe founder; Head of Models; Source Intelligence co-founder and chief scientist; director of the Beijing Research Institute, overseeing next-generation models and frontier-intelligence research and reporting directly to Han Luyi 120 142 157 199 300 311 404
Education	Doctorate from Stanford University under Professor Li Fei-Fei 171 311
Status	Active; based in Beijing 199 249 299 406
First Appearance	Chapter 12: “Dingsheng's Outsourcing” 12

Background

Zhao Wenyuan founded CodeSafe and accepted angel financing connected to Dingsheng. Wang Zhiyuan later disclosed that some of his personal capital had backed Zhao's early funding, then sought to use Dingsheng's planned move into code review to compete with Han Luyi's product. 26

Dingsheng ultimately acquired CodeSafe. Zhao's shares were forcibly included in the sale, he received little compensation, and he was pushed out of the company he founded; Cheng Yuan had handled the acquisition. 120 142

After leaving Dingsheng, Zhao warned Han Luyi that CodeSafe's user agreement allowed usage data to be collected for model optimization. He openly condemned the practice despite his own connection to the product, with anger masking self-blame. 53

When Han first invited him to join Source Code, Zhao had been unsuccessfully looking into technical-consulting work and running post-training experiments on open-source coding models from home. He initially declined because he considered foundation-model development prohibitively expensive for a startup, but Han left the offer open. 66 67

After Source Code raised RMB 200 million and Han demonstrated its ethically sourced prompt-code data pipeline, Zhao recognized that the company possessed the high-quality data needed for effective post-training. He agreed to lead the model team and began work the following day. 93

Zhao later became a founder of Haicheng Yuanzhi Digital Technology Co., Ltd., the new Source Intelligence venture. He received 3% founder equity without vesting for his technical contribution, with protection against dilution through the first two financing rounds. 157

Before returning to China, Zhao earned his Stanford doctorate under Li Fei-Fei and spent several years at DeepMind. He also worked within Google's research organization, where he coauthored pioneering Chain-of-Thought research. His decision to return was shaped by his perception that Chinese researchers faced increasingly restricted opportunities as they rose through foreign institutions. 193 311 355

Han subsequently entrusted Zhao with establishing and running Source Intelligence's Beijing research institute. Zhao accepted responsibility for its construction and spending, acknowledging that research directions could fail while promising that its funds would be used appropriately. By the institute's opening, he was responsible for all Source Intelligence operations in Beijing and for setting the research team's direction. 286 299 311

Following the company's structural adjustments, Zhao retained responsibility for next-generation models and frontier intelligence, reporting directly to Han. The institute was renamed the Origin Beijing Research Institute and expanded its research, engineering, and business-support teams. 404 406

Personality

Zhao is intensely technical and competitive. He is drawn to difficult, high-ceiling research paths, particularly training an independent foundation model rather than relying solely on open-source bases. 120

He confronts obstacles head-on: his instinct is to “win spectacularly” or lose the same way. This can leave him discouraged when a direction appears blocked, but clear strategic guidance quickly restores his drive. 170 176 185

He values candor and loyalty in professional relationships. Han's willingness to disclose the risk to Zhao's options strengthened Zhao's resolve to remain with him, even when that risk could have invalidated equity worth roughly RMB 60 million. 132

Zhao is rigorous about validating technical claims. He repeatedly verified Heavenly Craft's anomalously strong benchmark result across frameworks and test sets before reporting it, and his domestic-compute paper openly acknowledged that its single-operator result did not yet constitute a usable full training stack. 136 193

Though he initially struggled with the breadth of pretraining, domestic adaptation, and infrastructure deployment, Zhao accepts his limitations and shifts his efforts toward organizing work, documenting solutions, and leading the team through incremental progress. 161 189 199

His confidence in Han's technical abilities gradually becomes near-total, leading him to treat Han as a dependable source of solutions to otherwise intractable problems. He eventually recognizes that this reliance has limited his own initiative and accepts Han's encouragement to delegate to teams and solve difficulties independently before seeking help. 218 285 356 359

Zhao is committed to retaining scientific talent in China. During Zhang Wenduo's interview, he emphasizes that the candidate need not join Source Intelligence, but should consider keeping his abilities in the country rather than pursuing only the greater resources available overseas. 311

He combines professional humility with a fondness for showing off technical achievements. Lu Zhengping warns him not to undermine his authority through excessive modesty, while his confidence in Glutinous Rice Ball and his provocative online comments about domestic chips reveal a more competitive public persona. 303 378 380 395

As a research leader, Zhao becomes willing to revise his initial judgment when experimental evidence supports an unfamiliar approach. His skepticism toward Zhang Wenduo's asynchronous framework gives way to enthusiasm after he examines its utilization, accuracy, and training-time results. 367 368

Abilities & Skills

Large-Model Post-Training

Post-training is Zhao's established specialty, and Han considers his research capability industry-leading. 120 161

• Led Heavenly Craft's iterative training; its third-round accuracy reached 82%, approaching Dingsheng's Qianyuan at 84%. 117
• Brought Heavenly Craft to 84.7% accuracy after five training rounds, surpassing Qianyuan's public benchmark but reaching a plateau that additional data and GPUs alone could not overcome. 119
• Reported that a 7B post-trained model surpassed GPT-4 in intent understanding during its first training round, despite having less than one-twentieth of GPT-4's parameters. 136
• Improved the model's intent-understanding benchmark from 83.7 to 85.2 through further tuning. 141
• Encountered diminishing returns once fine-tuning loss curves flattened, recognizing that post-training had reached a plateau. 126
• Estimated that the post-training and alignment stage for Glutinous Rice Ball could be completed in two to three weeks after pretraining, drawing on his prior post-training experience. 223
• Began Glutinous Rice Ball's post-training after reporting that the model had reached state-of-the-art performance. 223
• Spearheaded Chain-of-Thought weighting during Glutinous Rice Ball's post-training, making a substantial contribution to its performance across tasks. 355
• Diagnosed Glutinous Rice Ball 1.0's later training delays with Jiang Songran as an algorithmic bottleneck rather than a hardware or infrastructure failure. 367
• Examined and adopted Zhang Wenduo's asynchronous reinforcement-learning framework, which separated actors from learners through an experience pool and reduced waiting between tasks. Zhang's experiments showed utilization exceeding 80% under favorable conditions and accuracy loss below 1% when key steps were marked for updates. 368
• Led the construction of a new training framework based on Zhang's ideas, cutting Glutinous Rice Ball's training time in half. 385

Foundation-Model Strategy

Zhao advocates building proprietary models for long-term control and a higher technical ceiling. 120

• Proposed training a code-specialized foundation model from scratch, with architecture tailored for context windows, dependency tracking, multi-file comprehension, and code generation. 120
• Estimated that each training attempt would require RMB 30 million and could take six to twelve months, with failed runs potentially requiring a full restart. 120
• Oversaw Tangyuan's pretraining after validating its training strategy and loss curve on a small dataset. 184
• Worked through CommonCrawl and public-domain pretraining-data cleaning issues, identifying the need for extensive filtering of advertisements, SEO spam, and malformed web content before full-scale training. 161
• Adjusted Tangyuan's pretraining strategy on a small dataset, verified that the loss curve had no obvious problems, and launched full-scale training on Dingsheng's compute cluster. 184
• Monitored the model through its annealing stage and coordinated scenario-specific evaluation-set preparation for Open Things and Riding the Wind. 219
• Led the rollout of Tangyuan inference on Source Intelligence's own cluster; after the deployment topology was corrected, the system reached 51 QPS in stress testing. 249
• Identified the limitations of exclusively Chinese training annotations for English intent understanding and requested English annotations from Han, later receiving a batch of 10,000 high-quality entries. 189 229
• Managed work to integrate the Heavenly Craft model into Tangyuan as an expert component. 311
• Represented the model team's requirements in discussions of the next-generation L200 chip, recognizing that the chip's design had to be coordinated with the future model and training framework. 378

Domestic GPU Adaptation

Zhao leads efforts to run model training and inference on domestic graphics cards rather than relying entirely on Nvidia's ecosystem. 173 184

• Developed a domestic-compute compatibility roadmap, but initially found the CUDA ecosystem gap nearly insurmountable. 184 185
• Helped establish a workable adaptation direction and organized the remaining work into incremental tasks. 190
• Assigned engineers to adapt operators for domestic GPUs; ten key operators had been completed at over 80% of equivalent Nvidia implementations' performance. 199
• Used coding Agents to accelerate adaptation attempts, with difficult operators escalated for team discussion or Han's assistance. 189 200
• Authored a single-author arXiv paper on adapting memory-efficient attention to non-CUDA domestic accelerators. Its eight-card single-node implementation achieved 83% of cuDNN throughput with numerical error of 2.3e-6, while candidly limiting its claims to a single operator rather than a complete training stack. 193
• Explained the fully offline domestic-GPU inference setup during Source Intelligence's investment pitch, helping demonstrate that the model did not rely on cloud callbacks or foreign hardware. 173
• Worked with Jiang Songran to investigate software, driver, operator, and service layers during Tangyuan's cluster stress-test failure before the actual bottleneck was traced to physical network cabling that violated the intended topology. 249
• Candidly told Jiang that the completed adaptation implementations being discussed had been written by Han, distinguishing his own organizing and explanatory work from Han's technical contributions. 200
• Proposed completing the first stage of L100 training-operator compatibility in one month rather than Jiang's estimated three months. Han accepted the target and took responsibility for detailed parameter tuning while the engineering team handled correct operator logic. 285
• Pursued high-speed interconnect support for L100 clusters beyond roughly 500 cards, seeking Hanlian Microsystems' help with topology and protocol-stack limitations. 312
• Coordinated interconnect work as research-institute director and explained HL-Link's specialized routing to Liu Yu during the successful 1,024-card cluster test. Interconnect protocols were not his original specialty, so his knowledge in this area came from recent intensive study. 359
• Publicly argued that software, memory bandwidth, interconnects, and specialized designs could compensate for domestic chips' manufacturing limitations, predicting support for both inference and large-model training within a year. 378 380

Technical Leadership

Zhao manages model-team work directly in the open office, using whiteboard discussions to allocate operator-adaptation tasks and resolve blockers with engineers. 199

• Introduced Jiang Songran to the team's training stack and domestic-adaptation progress on Jiang's first day. 199
• Explained the team's Agent-based workflow and distributed adaptation and iteration work across individual members. 200
• His domestic-adaptation paper was a major reason Jiang Songran chose to join Source Intelligence Technology. 199
• Led engineers in assigning three to five operator-adaptation tasks per person, with blockers surfaced for collective discussion and coding Agents used for rapid implementation attempts. 189 199
• Coordinated Source Intelligence's model and infrastructure teams during the Zhangjiakou cluster deployment and stress-testing process. 239 248 249
• Maintains an emergency alert system during major training runs and remains available to respond to training failures during holidays. 215 218
• Accepted overall responsibility for building the Beijing research institute, including spending, staffing, technical access, experiments, and research direction. 286 307 311
• Recruited researchers in model development, fundamental AI theory, World Models, long-context memory, reinforcement learning, and multimodal representation. 307 312
• Evaluated Zhang Wenduo through questions on World Models, black-box evaluation, and the risks of evaluation metrics becoming training targets, distinguishing theoretical accomplishment from practical large-scale training experience. 311
• Personally led the power-industry implementation project and assigned a team to support Jiang Jiusi's medical startup, granting the latter enterprise-API access. 356
• Was entrusted with personally announcing and distributing project bonuses after the HL-Link breakthrough, both to reward the teams fairly and to establish his authority at the institute. 359
• Secured Zhang Wenduo's work on the asynchronous framework by promising computing resources for his World Model project, eventually obtaining Han's approval for an additional 500-card cluster for the institute. 368 385
• Presented training logs, loss curves, and cluster-communication data to visiting experts, answering their questions during their examination of the institute's results. 387
• Expanded the Origin Beijing Research Institute by recruiting doctoral researchers and experienced engineering staff, renting two additional floors, and establishing a temporary office for visiting executives. 406
• Advised Jiang Songran that government-organized industry meetings offered a platform and introductions, but that Origin still had to persuade hardware manufacturers through technical results and commercial agreements. 406

Research Credentials

• Earned his Stanford doctorate under Professor Li Fei-Fei and subsequently worked at DeepMind and within Google's research organization. 193 311
• Coauthored Google's pioneering Chain-of-Thought paper. Academician Zhuang Zhi recognized the work and invited him to lecture to his team. 355
• Maintains that his academic experiments and papers were produced honestly, without fabricated results or inflated claims. 355
• Uses literature searches to supplement incomplete résumés and assess candidates' research records, as demonstrated when he identified Zhang Wenduo's first-author NeurIPS papers and academic adviser. 311

Equipment / Items

• Home server room — A dedicated, air-conditioned room containing a server rack, several servers, and a small eight-card GPU array. 165
• External GPU enclosure — Used in a fully offline inference demonstration; it housed a domestic GPU. 173
• Laptop and notebook — Frequently used for training reports, model analysis, presentations, and technical planning. 120 136 171
• Zhangjiakou deployment cluster — Source Intelligence's deployed inference infrastructure, used to run Tangyuan after final acceptance testing and network-topology correction. It later supported HL-Link testing and domestic-GPU training work. 239 248 249 359 367
• Beijing research office — Initially a modest rented floor with basic furnishings, strong internet, and plans for dedicated servers. Under Zhao's direction, it expanded into two additional floors and gained a temporary executive office for visitors. 311 406
• Research-institute computing allocation — Zhao promised Zhang Wenduo 256 L100 cards for World Model research in exchange for his asynchronous-framework work. Han subsequently authorized an additional 500-card Zhangjiakou cluster for the institute. 385

Relationships

• Han Luyi — Recruited Zhao into Source Code; their relationship develops into a close technical partnership built on transparency and shared ambitions for independent models and domestic computing. Han later gave Zhao closure by revealing his role in Wang Zhiyuan's removal from Dingsheng, and made Zhao a 3% founder-equity partner in Source Intelligence. He subsequently entrusted Zhao with the Beijing institute, introduced him to influential contacts, and encouraged him to develop independent authority rather than depend on Han for every difficult problem. Zhao continues to report directly to him in frontier-model research. 66 120 132 157 170 286 299 359 404
• Wang Zhiyuan — An early backer connected to Zhao's angel investment. Wang later played a central role in Dingsheng's handling of CodeSafe, leaving Zhao resentful after the acquisition and his expulsion from the company he founded. Wang's eventual removal from Dingsheng gave Zhao a sense that the matter was finally over. 26 120
• Cheng Yuan — Prosperity Group executive who handled CodeSafe's acquisition and later encountered Zhao as Source Code's Head of Models. Their reunion during the Dingsheng demonstration was outwardly polite but overshadowed by Cheng's past role in forcing Zhao out of CodeSafe. 142
• Jiang Songran — Fellow returnee from Google's Chinese technical circles and later colleague at Source Intelligence. Jiang respects Zhao's domestic-adaptation research and technical leadership; Zhao, in turn, brought him into the team's Agent-supported adaptation workflow and collaborated with him on cluster deployment. Their contrasting estimates for L100 training adaptation highlight Jiang's engineering caution and Zhao's confidence in Han. They later jointly diagnosed Glutinous Rice Ball's algorithmic training bottleneck, and Zhao advised Jiang on navigating Beijing's government-supported industry meetings. 193 199 200 249 285 367 406
• Liu Dahai — Dingsheng's chief scientist and an academic counterpart. Liu initially dismissed Zhao after learning he led a one-person model team, but Zhao's model demonstration forced him to reassess the technical gap between their organizations. They later met again in the Singapore conference delegation, where Zhao's confidence in Glutinous Rice Ball's practical performance contrasted with Liu's surprise at its absence from the rankings. 142 143 393 394 395
• Su Niannian — Source Code colleague who participates in discussions of model-training results, funding, and product direction. 130 136
• Zhang Biao — Han's assistant and a recurring logistical partner during model demonstrations, travel, and data-center deployment. Zhao witnessed Zhang's emergency response skills during the cardiac-arrest incident and later worked alongside him on Source Intelligence matters. 157 173 239
• He Yunshen — Investor and important sponsor of Zhao's Beijing work. Han asked He to support Zhao as the company's R&D focus moved to Beijing; He introduced him to industry contacts and Hanlian Microsystems and offered Hongyuan's continuing assistance. 170 299 303 312
• Lu Zhengping — State-backed investment contact who welcomed Zhao's Beijing assignment and urged him to project greater confidence as Source Intelligence's representative. 171 303
• Liu Yu — Han's secretary, temporarily assigned to assist Zhao with the Beijing institute's office preparations, housing, paperwork, interviews, and visits. Zhao appreciates his competence but initially feels uncomfortable directing someone borrowed from Han. 299 311 312 320
• Li Yan — Influential search-company founder whom Zhao had once idolized. Zhao was nervous about meeting him, and Han warned that Li might seek to recruit him for model research. After reflecting on the difference between admiring an industry pioneer and following him, Zhao stated that he had not returned from Google merely to join another search company. 299 300 303
• Zhang Wenduo — World Model researcher interviewed and recruited by Zhao. Their initial exchange involved tension over credentials, but Zhao recognized Zhang's ability and urged him to keep his talent in China. Zhao later identified the wider value of Zhang's asynchronous framework, redirected it toward Glutinous Rice Ball, and promised substantial computing resources for his research in return. 311 315 367 368 385
• Shen Zhonghe — Hanlian Microsystems founder and interconnect specialist. Zhao established technical contact with Shen, presented the opportunity for early collaboration with Icecore, and participated in discussions leading toward Hanlian's acquisition. Shen's interconnect team subsequently joined the research institute and developed HL-Link. 312 320 321 356 359
• Lu Mingzhou — Collaborator in the Hanlian acquisition effort. Zhao presented the technical and industrial opportunity while Lu assessed and developed the commercial negotiation strategy. 320 321
• Zhou Han — Power-industry investor and institutional contact. As Zhao took responsibility for the power pilot, she explicitly offered him direct assistance with electricity-related problems even when Han was absent from Beijing. 352 356
• Jiang Jiusi — Medical-startup founder whose collaboration Zhao supported through an assigned team and enterprise-API access. The project remained one of the Beijing institute's business responsibilities after its renaming. 302 356 406
• Zhuang Zhi — Academician and Zhang Wenduo's adviser. Zhuang recognized Zhao's Chain-of-Thought work, praised its pioneering significance, and invited him to lecture to his research team. 311 355
• Li Fei-Fei — Zhao's doctoral adviser at Stanford. 311

Story Role / Major Arcs

Building Heavenly Craft

Zhao became the technical force behind Heavenly Craft's model work. His post-training results gave Source Code a route to compete with substantially larger models in intent understanding, strengthening the company's position against Dingsheng. 117 136 141

His anonymized retraining work maintained data leakage below 0.001% while reaching 76.8% accuracy, and he identified expansion from roughly 230,000 to 500,000 training entries as the path to surpassing Qianyuan's public 84% benchmark. 110

By the fifth training round, accuracy reached 84.7%, but the final improvement was only 0.1 points. Zhao concluded that the existing post-training approach had reached its limit, helping frame the decision to pursue a proprietary foundation model rather than continue relying on more data and compute. 119 120

Choosing Source Code Over Outside Offers

After Heavenly Craft's achievements became public, Neus AI contacted Zhao about establishing a China research team and specifically cited his Google research direction and code-generation work. Zhao did not respond to the offer. 126

He instead committed to Han's new venture, accepting 3% founder equity with no vesting for his technology contribution. 157

His commitment was later tested more subtly by meeting Li Yan, an entrepreneur he had admired. Han suggested that Li might try to recruit him, prompting Zhao to reconsider the appeal of working under a former idol. Zhao affirmed that his return from Google was not intended simply to exchange one search company for another. 300 303

The Dingsheng Demonstration

At Dingsheng's demonstration, Zhao introduced himself as Source Code's Head of Models and presented alongside Han despite having been the sole member of the model team at that stage. His technical exchange with Dingsheng chief scientist Liu Dahai shifted Liu's attention from commercial negotiations to the underlying method behind Zhao's results. 142 143

The demonstration of their 7B Glutinous Rice Ball prototype—trained on fewer than 10,000 data points—showed substantially deeper intent inference than Dingsheng's hundreds-of-billions-parameter Kun Yuan model, transforming Zhao from an apparent one-man hobbyist into a representative of a serious technical threat. 142

Domestic-Compute Push

Following investment discussions in the capital, Zhao took on the difficult task of making Tangyuan compatible with domestic GPUs. Though initially overwhelmed by the software ecosystem barrier, he continued refining the technical path and later led a team scaling the adaptation effort. 176 184 189 199

His research paper and the model team's visible operator-adaptation progress helped persuade infrastructure specialist Jiang Songran to join Source Intelligence, expanding the effort beyond Zhao's original one-man research role. 193 195 199

As the company moved from inference toward domestic-GPU training, Zhao proposed an aggressive one-month target for initial L100 operator compatibility. Jiang objected that backward operators, stability testing, communication, and optimization would require a much longer overall schedule, but Han accepted Zhao's first-stage target and volunteered to handle the detailed tuning. 285

Zhao subsequently pursued the larger-cluster communication problem through Hanlian Microsystems. He learned that simply increasing bandwidth would not resolve the mismatch between topology, protocol, and hardware implementation, and sought an integrated solution involving Source Intelligence, Ximing, Hanlian, and Icecore. 312 320

Building and Deploying Tangyuan

Zhao oversaw the transition from post-training to full pretraining for Tangyuan/Glutinous Rice Ball, including data cleaning, strategy validation, long-running training, annealing, and preparation for post-training alignment. 161 184 219 223

When Source Intelligence began deploying the model at its Zhangjiakou data center, Zhao worked through a stress-testing failure with Jiang Songran and Han. Once the physical cabling was reconfigured to Jiang's intended network topology, Tangyuan achieved 51 QPS and began running on its own cluster. 248 249

During a holiday training failure, Zhao investigated abnormal loss spikes following a data update. After Han identified problematic pseudo-structured fragments and corrected the filtering rules, Zhao reprocessed the affected data and restarted training. 218

He also preserved the agreed separation between data uploaded to Dingsheng's cluster and the withheld portion: only 85% was to be uploaded, while the remaining 15% was retained as a future competitive advantage. 223

Establishing the Beijing Research Institute

Han's decision to establish an R&D center in Beijing placed Zhao in a new role extending beyond model-team leadership. Zhao initially understood the implied assignment with unease, but ultimately accepted full responsibility for building the institute and using its funds responsibly. He recognized that Han's ability to secure trust and resources was indispensable, leaving Zhao to take greater responsibility for research itself. 286

Han introduced Zhao to He Yunshen and other Beijing contacts before leaving him to run the operation. Zhao's initial priorities included recruitment, inter-card interconnect collaboration, training-operator adaptation, and cooperation with Shuimu's research community. 299 303 307

The institute began with a modest office where Zhao and Liu Yu helped arrange desks and chairs. Zhao oversaw technical handovers and interviews, while existing researchers were allowed either to relocate with company-paid moving expenses or remain in Haicheng and collaborate remotely. 307 311

His interview with Zhang Wenduo became a significant recruiting episode. Zhao challenged the candidate's theoretical and practical knowledge, then discussed his own experience abroad and encouraged him to retain his talent in China. Zhang accepted Source Intelligence's offer, turned down OpenAI, and was assigned to fundamental World Model research. 311 315

Hanlian Collaboration and HL-Link

Through He Yunshen, Zhao met Shen Zhonghe's interconnect team and investigated the L100's scaling limitations. Hanlian judged that confidential packaging parameters would be needed to improve the existing cards, while early collaboration with Icecore could offer much greater scope for the next generation. Zhao reported these findings to Han, who later instructed him to pursue Hanlian's acquisition. 312 315

Working with Lu Mingzhou and Liu Yu, Zhao presented Hanlian with a concrete opportunity to participate from the start in the next-generation card's design. He also indicated that Source Intelligence might already have access to the parameters needed for L100 adaptation and emphasized that Hanlian would become a foundational part of the company's domestic-compute ecosystem rather than a peripheral component. 320 321

After Hanlian's acquisition brought its team into the institute, HL-Link testing demonstrated a 4.7-fold improvement in communication efficiency for a 512-card cluster. Two such clusters were then successfully combined into a 1,024-card system that completed a simulated training task, although further stability testing remained necessary. 356 359

Han entrusted Zhao with organizing the celebration and distributing bonuses personally. The assignment was also intended to strengthen Zhao's standing among the newly assembled institute staff. Their subsequent conversation marked a shift toward greater delegation and independence from Han's direct technical intervention. 359

Accelerating Glutinous Rice Ball 1.0

By May 14, Glutinous Rice Ball 1.0's projected completion date had slipped from May 26 to May 28 and then June 3, threatening its entry into the conference's blind evaluation. Zhao and Jiang determined that uneven work across nodes created an algorithmic bottleneck. Zhao chose to address it without asking Han, both because post-training was his specialty and because Han had given him autonomy. 367

Zhang Wenduo's request to expand an experiment from 16 to 64 GPUs exposed an alternative scheduling approach in which workers immediately took new tasks instead of waiting for every member of a batch. Zhao initially worried that using results generated from outdated parameters could destabilize learning. 367 368

After examining the experiments, Zhao recognized that key-step marking could keep accuracy loss below 1% while reducing training time to less than half. He redirected Zhang's work toward Glutinous Rice Ball, offered resources for World Model research in exchange, and reported that the model should finish by the end of May. Han's brief congratulatory reply was less dramatic than Zhao had hoped. 368

The resulting asynchronous reinforcement-learning framework halved training time. Zhao later revealed that he had promised Zhang 256 L100 cards despite the institute's existing allocation being only a little over 100; Han approved another 500-card cluster for the institute. 384 385

Industry Applications and Public Representation

Alongside fundamental research, Zhao personally led a power-industry implementation project and coordinated support for Jiang Jiusi's medical startup. Zhou Han offered him direct assistance with electricity-related difficulties, strengthening his ability to operate in Beijing without Han's constant presence. 352 356

He and his team also compiled materials under Han's direction for discussions of agent-management rules. Zhao attended the meeting and later received academic recognition from Zhuang Zhi for his earlier Chain-of-Thought research. 354 355

During planning for the L200 chip, Zhao traveled to Haicheng in person to ensure that the model team's needs informed the hardware requirements. He also publicly defended the feasibility of domestic large-model training and provocatively updated his online answer after a confidential meeting. 378 380

Singapore AI Conference

Zhao accompanied Han to Singapore for the June AI conference and worried about Glutinous Rice Ball 1.0's disappearance from the competition rankings. Han urged him to prioritize usefulness and the Tangyuan Protocol's ecosystem rather than a ranking, while an approaching academic claimed that the model had been disqualified despite deserving first place. 393

When Liu Dahai questioned whether Source Intelligence had developed a new model, Zhao remained confident in Glutinous Rice Ball's capabilities. During the live sandbox task, he declined to guess which anonymous display belonged to it, stating that the most complete result would identify their model when the task ended. 394 395

Origin Research Leadership

Under the company's reorganized structure, Zhao continued to oversee next-generation models and frontier-intelligence research, reporting directly to Han. Jiang Songran separately assumed overall responsibility for in-house computing, with Xue Zhaoheng and Shen Zhonghe reporting to him. 404

The renamed Origin Beijing Research Institute expanded rapidly through university recruitment and the hiring of experienced engineers from Beijing's technology companies. Zhao rented two further floors and provided a temporary executive office so that visiting leaders would no longer need to use his own office. 406

When Jiang arrived for a government-organized meeting with domestic computing-card manufacturers, Zhao welcomed him and explained the distinction between official facilitation and substantive cooperation: government could convene the companies, but Origin still needed to earn their trust through working systems and mutually acceptable agreements. 406

Notable Quotes

“Thirty million RMB to start.” 120

“I'm not afraid of hardship! The harder it is, the more we should take it on!” 176

“President Han, this isn't your decision alone. It's mine too. I'll go to Beijing, and leave the entire construction of the research institute to me. I can't guarantee the direction will be correct, but I can guarantee that every cent the institute spends will be spent where it should be.” 286

“Boss Han, I didn't come back from Google to join another search company.” 303

“Boss Han's targets for me are to focus solely on how to spend money and how to research AGI. There's no pressure to generate revenue.” 355

“Understood, Boss Han. We'll do our best to solve any difficulties ourselves and only come to you when we can't.” 359

“Ladies and gentlemen, listen to the dragon's roar.” 380

“It's only been an hour. What's the rush? We'll know when it's over. Whichever one is the most complete will be Glutinous Rice Ball.” 395

Trivia

• Zhao is repeatedly associated with Luckin coconut lattes; Han refers to him as their “spokesperson.” 66 93 130
• He habitually arrives at the office an hour and a half early to review training data in quiet conditions. 126
• Despite working in an AI model team, he does not have a private office and works alongside the rest of the team in the open-plan workspace during his earlier Haicheng model-team leadership. He later has a director's office at the Beijing institute. 189 356 406
• He lived alone in a large rented flat partly because the dedicated server room and his personal life made remaining with his parents inconvenient. 165 166
• He once declined Han's proposed company dinner because he had arranged dinner and a New Year's countdown in Xintiandi with a woman he had met at a bar. 166
• His Chinese Google-community groups included food, hiking, cherry-picking, and a largely inactive “Haicheng People at Google” group in which he found Jiang Songran's name. 193
• At the institute's opening, he commuted from his long-term Beijing rental apartment by shared bicycle. 311
• He initially disliked directing Liu Yu because the secretary was only temporarily assigned to him and they were of similar age. 311
• He is not accustomed to spicy food and ordered tomato-and-egg rice when dining with Han and Academician Zhuang. 355
• His Beijing driver is Old Ma. 352
• He introduced Han to northern takeaway dishes, including Hejian-style donkey-meat flatbread sandwiches and stir-fried flatbread, but remained unhappy about having previously been tricked into drinking fermented mung bean juice. 384
• Han's frequent use of Zhao's director's office during Beijing visits prompted Zhao to establish a separate temporary executive office; Jiang Songran was its first visiting executive user. 406

---