Summary

The frontier in foundation-model competition is shifting from exhausting public corpora to connecting enterprises’ private-domain data to business workflows, making data the key constraint on the next phase of deployment. 日照 believes language pretraining data “should basically have been used up,” while companies’ most critical data will not be trained into foundation models; the unresolved questions are how to manage and retrieve that data, then fuse it with model capabilities inside business processes. 庄明浩 cited an MIT report saying “95%” of enterprise AI investments in Q3 and Q4 last year produced disappointing results, but both speakers saw this as evidence that implementation remains early, not that the industry had been disproven.
Pretraining’s marginal advantage has not disappeared; it has moved from data scale to data quality, engineering detail, reproducibility, and rapid iteration. 日照 used Google Gemini 3 as an example: everyone may have used similar data, but it “might have done pretraining” better, with the edge coming from processing data row by row and maintaining stable infra. Outside language, data for word model and embodied systems remains “quite limited,” so data, algorithms, and compute will continue to raise capability in alternating pushes.
The control plane for enterprise AI is upgrading from a single vector database to hybrid retrieval and context engineering. OceanBase shifted its competition from vector databases last year to hybrid retrieval, combining vector, full-text, graph, scalar, and semi-structured indexes, while also testing direct AI Function calls from inside the database. The reality is that CEOs are already saying, “I want to use foundation models,” while CTOs and CIOs, facing a market with no unified evaluation, language, or solution, still answer, “I don’t know how to use them.”
Performance leaps in the competition suggest that early AI data-stack gains may come more from redesigning execution flows than from conventional local optimization. 杨同学’s team started at roughly 600 points, then rewrote its execution flow after watching the “Don’t Have a Bad Attitude” team jump to roughly 1,300 or 1,600 points, eventually reaching about 7,100 or 7,200; 吴同学’s team increased performance tenfold through index-layer intersection before optimizing full-text scans layer by layer. As one competitor put it: “They’re called ‘Don’t Have a Bad Attitude,’ and we ended up having a bad attitude.”
Hallucination control cannot stop at whether an answer is correct; its sources, documents, and page numbers must form a verifiable chain. The student solution first rewrote the user’s question, ran hybrid retrieval, and reranked the results; when the answer was right but the citations “kept drifting,” it located the document by chunk first and the page number second. 日照 compared this with the brain and an external memory resembling the hippocampus: databases can retain fixed information and retrieve it on demand, improving economics, efficiency, and trust.
AI Coding has changed development speed, but the core capability is shifting from who types the code to who can define requirements, review outputs, and keep innovating. 杨同学 initially insisted on writing code by hand, but after watching teammate 吴敬华 use GPT to analyze code and Claude to implement features—producing a working version before 杨同学 had fully understood the causes and effects—he began embracing AI. 日照 still stresses that rigorous code “has pitfalls inside” and requires review by strong engineers; OceanBase’s conclusion is more direct: “The most critical asset is actually the employees, not the code they write.”
Continual Learning and Agent commercialization will advance along two tracks moving at different speeds. Internal model learning still lacks a predictable breakthrough: after several rounds of LoRA, results may be unpredictable and still require human inspection; iterative context between data and models is viewed by 日照 as more practical and “the right direction.” General-purpose Agents may “grow tenfold in a year,” while enterprise Agents look more like To B: if the original growth rate was 20%, AI may lift it to 30%–40%, not suddenly to 200%, but that growth can “continue for 20 years.”
Open source is evolving from a way to disseminate technology into a product and distribution strategy for Chinese AI engineering to compete globally. OceanBase began inside Ant in 2010, became a company after 2020, and went open source the following year; this was not simply publishing internal code, but removing dependencies on Ant, revising the code, and completing the documentation. 日照’s personal goal for 2026 is to push OceanBase’s open-source ecosystem worldwide, while finance, healthcare, broad search, risk control, and customer service are seen as industries and use cases where private-data ROI will emerge earlier.

Deep dive

1. AI’s Next-Phase Problem Is Once Again Being Reduced to Data

庄明浩 revisited his summary of 2025: this model cycle was mainly solving three problems—reward models, memory, and benchmarks—and most of them can ultimately be traced back to the data layer.
Many called 2025 “the year of reinforcement learning.” Entering 2026, the industry has shifted to talking about Continual Learning, online learning, and autonomous learning. The definitions remain muddled, but 庄明浩 believes that whether reinforcement learning continues to evolve or online learning truly emerges, data will only become more important.
This AI cycle is often described as a victory for engineering. The host’s question was whether, once “engineering” is reduced to the daily work of model companies, data has moved from a supporting back-office role to the center of model progress.

2. Public Data Supports General Applications; Private Data Determines Enterprise Penetration

日照 divides applications into two phases. The first uses public data, which is why products built around ChatGPT, Coding, tattoo images, and tattoo videos are relatively easy to build and generalize across companies; the second must enter the private-domain data unique to each enterprise.
Private-domain data “definitely will not be trained into a foundation model,” because it is often the company’s most critical asset. The challenge therefore becomes managing that data and combining it with the reasoning capabilities learned from public data, then embedding the combination into actual business workflows.
庄明浩 cited an MIT report covering Q3 and Q4 last year, which found that roughly “95%” of enterprise AI investments had poor implementation results. Neither speaker took that as evidence that the bubble had burst: ChatGPT appeared in late 2022, while deployment only began in 2024 and 2025, and organizational and systems integration naturally takes time.

3. Pretraining Competition Has Shifted from “Having Data” to “Using the Same Data Well”

日照’s view comes with a clear boundary: language pretraining data “should basically have been used up,” while data for word model, embodied systems, and other areas remains quite limited and will require new methods.
Even with similar language data, quality engineering can still separate models. Using Google Gemini 3 as an example, he said its strength “may most fundamentally come from doing pretraining well”—processing data row by row and tuning the tedious details, rather than simply expanding corpus size.
This work also requires stable infra: results must be reproducible and experiments must iterate quickly so algorithm engineers can keep tuning the data. 庄明浩 described the mutual reinforcement among models, data, pretraining, and reinforcement learning as “stepping on your left foot with your right foot and kicking yourself upward.”

4. Foundation Models Are Bringing Previously Difficult Data into a Broader Processing System

日照 does not agree that data only became important today: “For truly first-rate technology companies, it has always been in the most central position.” What foundation models changed is that “they make data that couldn’t previously be used usable,” extending beyond structured data to text, video, and audio.
庄明浩 described databases and foundation models as the two core components of a data-processing system. Databases will need to handle more multimodal and unstructured data while using model capabilities to simplify database operations and user workflows.
庄明浩 compared language, multimodality, and Coding to separate Texas Hold’em tables. As images begin to carry logic, search results, stories, and causal context, they are gradually becoming “one big table” in the user’s eyes. The underlying technical routes may differ, but users are waiting for methods that truly generalize.

5. Enterprise AI’s First Bottleneck Is Not Model Pricing but the Lack of Standardization

After DeepSeek materially reduced the cost of using foundation models, enterprise demand rose quickly. 日照 observed that “the people using foundation models are all CEOs”: CEOs say, “I want to use foundation models,” while the CTOs and CIOs below them often answer, “I don’t know how to use them.”
Databases helped information systems spread through standardization. Enterprise foundation models currently lack a unified evaluation system, standard language, standard construction method, or standard solution, while outcomes remain highly case by case. For ordinary companies, the barrier to adoption is still high.
OceanBase chose to start with the integration of data and models, helping more companies gradually solve the real problems they encounter when using foundation models. Only after enough industry solutions accumulate can sector-specific approaches and standards emerge naturally.

6. The Database Competition Has Shifted from Vector Retrieval to Context Engineering

Last year’s OceanBase challenge focused on vector databases. Model capabilities were relatively weak at the time, so the common approach was to supplement information through embedding and RAG. This year, the challenge was upgraded to hybrid retrieval and added a React problem.
日照 believes the industry will move from single-vector search and embedding toward “context engineering”: vector, full-text, graph, scalar, and semi-structured indexes will all participate in search and then be combined with model capabilities. Databases may also call models directly through AI Function.
Although the competition added AI-related business scenarios, it remained focused on database internals in its fifth year. The questions must reflect current applications while retaining enough differentiation to separate teams, with the goal of studying how to optimize database internals for AI scenarios and training engineers in low-level database kernels and foundational software.

7. Query Rewriting, Hybrid Retrieval, and Reranking Form the Student Retrieval Pipeline

杨同学’s team first rewrote user questions to make synonyms, abbreviations, and full names more explicit, then ran hybrid retrieval. After obtaining candidate content, it reranked the results to improve semantic relevance and reduce hallucinations.
The user’s question was a “blind box” for competitors, so the solution could not be tuned only to known phrasings. 吴同学’s team referenced OceanBase’s implementation but admitted that its version was “comparatively rough,” with the priority being to verify that the full chain worked.
The pipeline also had to preserve efficiency: the path could not expand indefinitely in pursuit of accuracy. The database kernel’s objective was summarized simply—move fast, keep the data accurate, and make the underlying complexity as invisible to users as possible.

8. One Execution-Flow Rewrite Took the Score from Roughly 600 to More Than 7,100

杨同学’s team initially reached roughly 600 points and concluded that conventional optimization was nearing its 600–700 ceiling. After two teammates returned from playing a game, they found that the “Don’t Have a Bad Attitude” team had surged to roughly 1,300 or 1,600 points. The reaction was: “They’re called ‘Don’t Have a Bad Attitude,’ and we ended up having a bad attitude.”
They inferred that the rival team was not using the usual optimization approach and attacked the entire execution flow. 杨同学 first said the final score reached roughly 7,700, then corrected himself to roughly 7,100 or 7,200, preserving the uncertainty in the score.
吴同学’s team had gone a long time without improving performance, then used index-layer intersection to make it more than ten times faster before optimizing full-text scans further. The final design had three or four layers and a score close to 7,100. During the defense, the teams realized that their implementation paths were “completely different”; there was no single template.

9. AI Databases Have Reopened a Mature Deterministic System

日照 recalled that traditional databases mainly handled transactions and analytics, with theory and practice refined over many years and methodological change occurring slowly. Once data is combined with AI, many problems become questions of “how to find what best matches the foundation model from the context.”
Search results are not 100% precise in the way “one plus one equals two,” and the field is still young, with methods yet to settle. Judges sometimes could not anticipate the students’ paths, turning the competition from one-way evaluation into mutual learning and making databases more open-ended.
日照 also believes that AI databases rely more heavily on young people and are no longer the domain of “old-timers.”

10. The Bottom Line for Hallucination Control Is That Both the Answer and the Source Must Be Correct

日照 compares a model to a brain with reasoning ability, but reasoning alone is not enough; it also needs external memory resembling the “hippocampus.” A database stores fixed information, which the model reads when needed, lowering cost and avoiding recomputation every time.
The typical issue competitors encountered was that the answer appeared correct while the linked document and page number “kept drifting.” Even after the organizers added a vision model, the drift remained. Students called this a “very serious hidden risk” in industry because there was no way to distinguish retrieved information from model guesses.
The final approach ranked documents by chunk first, then located the page number inside the most likely document, using multiple layers of positional confirmation to supplement the generative model. The team acknowledged that the design was not comprehensive enough at larger data volumes and that, under time pressure, its performance on knowledge questions was not ideal.

11. AI Coding Changed the Mind of a Traditional Hand-Coding Advocate in a Single Final

杨同学 had an ACM background and entered the competition believing that “things I write myself will be more reliable,” worried that AI-generated code would leave him without a clear understanding. His team collaborated through PRs, and no change was merged until everyone understood one another’s modifications.
The turning point came from teammate 吴敬华, who used GPT to analyze code and Claude to implement features directly. While 杨同学 was still working through the causes and effects, 吴敬华 had already produced a functional version for the team. The speed gap pushed 杨同学 to learn how to explain problems and requirements clearly to AI.
吴同学 started using AI earlier. He believes the biggest change of the past year is that programming IDEs can now read the entire source-code context directly. Developers no longer need to extract code section by section and paste it into a model, which has substantially accelerated the reading and modification of large projects.
The craft has not disappeared entirely. Humans can spot and fix simple errors at lower cost than repeatedly burning tokens; the team still reads AI-generated code closely and understands the execution flow. 杨同学 sees that discipline as the reason he was eventually able to catch up and overtake.
庄明浩 extended the same shift to PPTs. AI is no longer just about layout; it can understand logic and visualize it. He once insisted on building slides page by page by hand, but now, when he makes a PPT every 2 or 3 months, he uses AI extensively to generate illustrations and feels that the room for “hand-tuning PPTs” is steadily being squeezed.

12. Enterprise Development’s Constraints Are Moving from Writing Code to Review, Collaboration, and Accountability

AI Coding adoption is high inside OceanBase, with especially strong gains in front-end development, experimental code generation, and parts of middleware. But 日照 stresses that highly rigorous code still has obvious problems: “The things it produces directly have pitfalls inside.”
Developers therefore need to learn how to use AI, while stronger engineers must review the output and teams must establish development, testing, and product-collaboration methods adapted to AI Coding. 日照 calls this “a new craft” built on top of traditional craftsmanship.
庄明浩 worries that public coding tools may create security and privacy problems. 日照 did not expand on tool security; instead, he emphasized that OceanBase’s core code is fully open source, so “the most critical asset is actually the employees, not the code they write.” The real moat is the organization’s ability to keep innovating, solve customer problems, and provide a backstop when failures occur.

13. Continual Learning Has Two Routes; Internalizing Knowledge in the Model Remains Unpredictable

日照 divides the earlier discussion of continual learning into two layers. One is continuously internalizing new knowledge into the model itself; the other is combining the model with a database that can be read and written in real time and continuously evolve, updating capability through external context.
The first route “has not seen a major breakthrough so far.” After several rounds of LoRA, the final result is essentially unpredictable and still requires human inspection, making a stable automated loop difficult. 日照 mentioned new labs such as Thinking Machines Lab that are also experimenting, but did not offer a firm view on whether a theoretical breakthrough is inevitable.
OceanBase chose to start with the more practical second route: connecting data and models through hybrid search and continuously iterated context. 日照 acknowledged that it “has not completely figured the whole thing out,” but judged it to be “definitely the right direction” for companies large and small.

14. Open Source and Industrial Applications Truly Converged in 2025

庄明浩 observed that early open-source research, GitHub trends, and industrial applications once ran on separate tracks. By 2025, growth in Agent frameworks, infra, and data-layer projects had become closely aligned with research and industry trends: “Open source does not equal public service,” nor does it mean the work cannot be commercialized.
日照 explained that AI’s early phase was more innovation-driven and separated from traditional technology stacks. Once enterprises began demanding intelligence in existing applications, the two systems had to converge, forcing incumbent databases and foundational software to embrace AI.
The host described “open source” as a candidate keyword for Chinese AI in 2025, arguing that it may be both a result of China–US AI competition and, in turn, a cause. 日照’s explanation was that China may not invent algorithms first, but excels at pushing engineering optimization to the limit; open source is an effective way for such products to reach global markets and keep iterating through an ecosystem.

15. OceanBase’s Open Source Was Not Code Publication but a Shift from Internal System to General-Purpose Product

OceanBase began development in 2010 and spent its first decade primarily as an internal Ant project. It became a company after 2020 and went open source the following year. Open sourcing did not mean simply making the repository public; it required removing dependencies on Ant’s internal code, revising parts of insufficient quality, and completing external documentation.
Moving from internal support to an open product, sales, and customer delivery changes both roles and objectives. 日照 believes the deciding question is “what is the original purpose of this product.” OceanBase’s goal was “to build a world-class database” from day one, so its design was oriented toward generality from the start.
日照 said OceanBase, seekdb, and the higher-level exploratory tools Power Lag and Power Memory all follow an open-source path and use the Apache 2.0 license. The team hopes external developers will build better data–model integration on top of them.

16. Young Developers Have More Opportunity, and the Security Boundary Is Expanding with It

杨同学 feels that traditional security was more focused on fixed rules and closed structures. Once models, multimodality, and enterprise businesses are connected, issues such as foundation-model security and multimodal security greatly expand the security boundary, which remains in a fast-developing and unsettled phase.
庄明浩 warned that once hallucinations enter transaction and permission chains, errors can translate directly into operational risk. 杨同学 wants to continue researching foundation-model security; the cryptography area he has encountered is fully homomorphic encryption, which performs computation on encrypted data and produces the correct result after decryption.
吴同学 works at a bank and started as a backend engineer. With AI, he has also begun working with Java, C++, and React, gradually becoming full-stack. He admits that his future direction remains “extremely unclear,” but is also considering becoming an independent developer and building small, novel products.
日照 believes the profile of database talent has changed in some ways but not others. Fundamental database theory and basic coding ability remain important: learn the old craft first, then learn the new craft of using AI well.

17. Agents Will Grow at Two Speeds; the 2026 Decisive Move Is Turning the Path into a Product

日照 believes AI Agents will accelerate further in 2026 and truly enter enterprise workflows. General-purpose Agents, powered by general capabilities, will develop quickly; Agents deeply integrated with enterprise processes will take longer because they must handle data, permissions, models, and existing systems.
He cited the industry saying that general-purpose Agents will “grow tenfold in a year,” and explained the difference through To B and To C. If traditional To B growth is 20% annually, AI may lift it to 30%–40%, but not suddenly to 200%; the advantage is that this growth can “continue for 20 years,” while consumer breakouts may end after 2 years.
In terms of use cases, broad search, risk control, and customer service are approaching large-scale adoption. Enterprises may not yet be using them well, but demand is already urgent. By industry, finance and healthcare have high-value private-domain data that is easier to connect with business and convert into clear ROI; 日照 expects these sectors to break out first.
OceanBase’s 2026 objectives have two layers: productize and systematize the path for connecting data and models, helping enterprises make it work and replicate it over the coming years; and push its currently China-focused open-source ecosystem worldwide.
杨同学 hopes to find an interesting full-time job and is also considering becoming an independent developer. 吴同学 looks forward to AI for Science assisting more rigorously with paper research and hopes to become an open-source community committer. He also said: “Code is always written by people.” If you look at it carefully, you will understand it sooner or later.

The Singularity Countdown: AGI by 2029, Humans Merge with AI, Intelligence 1000x | Ray Kurzweil Atlassian’s Most Controversial Growth Decision | Mike Cannon-Brookes