On March 6, 2023, shortly after ChatGPT was released and before GPT-4, Sarah and I conducted an interview about ChatGPT—the third installment of the Traders' Talk "Plain Language Series." At that time, ChatGPT was still relatively new, and very few people were actually using it. This three-hour interview remained at the top of the ChatGPT category on Xiaoyuzhou (a Chinese online forum). I made about twenty judgments and predictions in the interview, relying entirely on intuition and limited information, with little data. The complete transcript of that interview is still available on the official WeChat account. Now it's the end of May 2026, three years later, and AI has grown into something unimaginable back then. I want to do something: take those twenty judgments from back then, one by one, and objectively reconcile them using the latest available data. I want to see clearly how the world has changed in the past three years, and also see where my judgments from three years ago were correct and where they were wrong. To ensure impartiality, I delegated this reconciliation to AI: I fed the verbatim transcript of the interview from back then into a workflow, which then dispatched 41 Opus 4.8 agents. These agents broke down each of the twenty judgments, retrieved the latest data online, cross-verified each judgment, and finally scored Wang Jianshuo from three years ago. This group of agents spent approximately 20 minutes and burned through 1.4 million tokens (about $35 USD) to produce the report below. The judgments all came from them, not me. The baseline date is set for May 2026. I. Scoreboard Judgment Symbols: ✅ Correct · 🟢 Mostly Correct · 🟡 Partially Correct · ❌ Incorrect At first glance, Wang Jianshuo's general direction was mostly correct; there was only one truly hard error—transmitting GPT-4 as 100T parameters. But the devil is in the details: behind almost every "correct" judgment lay a loose end that wasn't accurately stated back then. None of the twenty points are purely "still uncertain." Three years is long enough; most things have already shown a tendency towards a positive answer. Let's discuss them in detail in groups below. Second, the common thread among those who correctly predicted the outcome is that Wang Jianshuo's predictions regarding direction, mechanism, and even timing were all accurate. His only mistakes lay in the "degree" and "absolute wording." RAG and Retrieval Architecture (Points 2 and 3): In 2023, Wang Jianshuo stated that the mainstream method for solving the knowledge and illusion problem was not to modify the model, but rather to use vector retrieval to cram knowledge in as a "cheat sheet"; the correct architecture is for the search engine to perform the retrieval and feed the results to the LLM. This is the de facto standard for all AI products today.RAG has become the default architecture for enterprise AI, with OpenAI, Google, and Anthropic all developing it into a platform-level capability. ChatGPT Search, literally, means "first using Bing indexing for retrieval, feeding the results to GPT, and then generating a cited answer." Google AI Overviews achieved approximately 2 billion monthly active users using grounding, and Perplexity, a company relying solely on this architecture, saw its valuation soar to approximately $20 billion. Before GPT-4 was released and the industry generally accepted "relying on fine-tuning to inject knowledge," he bet on "not changing model parameters, using external search," and both the mechanism and timing were correct. LUI is a new continent (Viewpoint 7): In 2023, Wang Jianshuo said: The greatest achievement of ChatGPT is not AIGC, but the opening of LUI (Natural Language User Interface), which will reconstruct human-computer interaction like GUI did in its early days, giving rise to a new industry much larger than "building large models" itself. This "new continent" part is almost entirely accurate. Natural language has become the dominant interaction layer (ChatGPT boasts 900 million weekly active users), spawning a new independent industry—agents, coding agents, and protocol layers have all materialized. The most concrete statement, "far larger than the model itself," has been strongly validated: the MCP protocol became the "operating system standard" of the LUI era, fully adopted by OpenAI, Google, and Microsoft in 2025, and transferred to the Linux Foundation by the end of the year; Claude Code alone achieved approximately $2.5 billion in annualized revenue. Robot Networks and New Addressing (Viewpoint 9): In 2023, Wang Jianshuo stated that "robot networks" would emerge in about ten years—agents would automatically handshake and call each other using natural language, eliminating the need for traditional APIs; a completely new domain name addressing system would be born. This system "could be completed in two or three years." The direction was remarkably accurate. MCP and A2A (donated to the Linux Foundation and supported by over 150 organizations) solve agent inter-calling; Agent Network Protocol directly uses W3C's DID for "agent addressing without a centralized authority," aiming for a "network of billions of agents collaborating"—highly isomorphic to his "new domain name system." China will definitely be able to create usable large-scale models (viewpoints 10 and 20): In 2023, Wang Jianshuo said: China will definitely be able to create usable large-scale models, and the gap with the top will be rapidly closed within about three years (analogous to Red Flag Browser catching up with Netscape). This timeline is surprisingly accurate. The Stanford 2026 AI Index test showed that the benchmark gap between top-tier Chinese and American models narrowed from 17.5–31.6 percentage points in May 2023 to 2.7%; while private AI investment in the US is about 23 times that of China—achieving this close with much less investment. DeepSeek, Qwen, Kimi, and GLM have become global mainstream, even leading in open-source ecosystems.The argument that ChatGPT lacks consciousness and that the Turing test only measures appearance (Viewpoint 13): In 2023, Wang Jianshuo stated that ChatGPT lacks consciousness, a self-indulgent statement that was "unintentional on the part of the speaker but taken to heart by the listener." The Turing test only measures "whether it makes you think it exists," not whether it actually exists. This core judgment of "measuring appearance" is firmly established and ironically confirmed by an experiment: In the 2025 Turing test at UC San Diego, the proportion of GPT-4.5 individuals judged as human under the prompt of "playing a persona" was as high as 73%, higher than that of real people, but this was purely based on performance skills—this is the best illustration of "only measuring whether it makes you think it exists." The rest of the correct points (Viewpoints 6, 11, 12, 16, 18, 19): not AGI but a big step forward: standing firm on both ends. Altman himself still said during the GPT-5 era that it "wasn't AGI and lacked continuous learning"; meanwhile, the IMO gold medal and ARC-AGI's jump from near zero to 85% were undeniably "a giant leap forward." There won't be a wave of unemployment: the US unemployment rate in April 2026 was only 4.3%. The blind spot is in "distribution"—Stanford research shows that those being drained are precisely the young newcomers aged 22-25 at the top of the career ladder; the "smooth absorption" mechanism has failed with them. There won't be an overwhelming AI waste: the direction of net welfare is correct, but he seriously underestimated the scale—AI content already accounts for about 52% of new web pages, and "AI slop" has become the word of the year. A great year for startups: the inflection point of the wave was caught correctly; xAI (founded in March 2023) has already reached a valuation of 230 billion. However, his definition of "great companies" as limited to 2023 is too narrow—truly trillion-dollar companies like OpenAI and Anthropic were founded much earlier. 1994 Browser Moment: Relative ranking was confirmed, and OpenAI truly launched the Atlas browser in 2025, turning the metaphor into literal reality. However, ChatGPT's spread was even more aggressive than the browser, making the metaphor somewhat conservative. Prompt adds facts to reduce illusions: The direction was confirmed; GPT-5's illusion rate soared to 47% when offline and without search, conversely confirming that "facts" are the key variable. The root cause was underestimated—training incentives, not the prompt itself. III. Misunderstanding and Misinterpretation: GPT-4 has 100T parameters (Viewpoint 4)—Completely Wrong: In 2023, Wang Jianshuo said: (rumored) GPT-4 has 100T parameters, about 600 times more than GPT-3's 175B. Both numbers are wrong. GPT-3 has 175B; the best estimate leaked in July 2023 was GPT-4 at approximately 1.8T with 16 experts' MoE, only about 10 times more. The 100T figure differs from the actual value by approximately 55 times. LLM Mathematics (Viewpoint 1) – The diagnosis is correct, but the final conclusion is wrong: In 2023, Wang Jianshuo stated that the fundamental problem with LLM mathematics is its inherent weakness. It is neither possible nor necessary for it to learn mathematics on its own; the correct approach is to use external tools.The "diagnostic plus tool approach" was entirely correct—the root cause being the unreliability of carry-over due to token-by-token generation (the 2025 mechanism paper precisely confirmed the intuition of "last digit is often correct, middle digit is incorrect"); the improvement of plug-ins was also significant (when o4-mini allowed Python, AIME 2025 reached 99.5%). The mistake lay in the captive wording of "impossible, unnecessary." Value capture (viewpoint 8)—half right, core assertion reversed: In 2023, Wang Jianshuo said: value will ultimately fall on the application layer, and companies that create the foundational layer (model builders) may not necessarily make money in the end. Money did indeed start flowing to the application layer (Cursor achieved 2 billion annualized revenue in three years)—this half was correct. But "building the foundational layer doesn't make money" was directly disproven by Nvidia: FY2026 net profit of approximately $120 billion, market capitalization of over $5 trillion, making it the only company in the entire market with clearly substantial profits. Copyright (Viewpoint 14) – Registration is correct, avoiding infringement is wrong: In 2023, Wang Jianshuo said that AI-generated content might circumvent copyright (protecting expression but not ideas); the generated content might neither infringe nor be registered. "Unable to register" has become an established legal fact (in 2025, the US Copyright Office clarified that "simply inputting prompts is insufficient to claim authorship"). However, "avoiding infringement" is clearly wrong: courts have repeatedly ruled that AI output, if substantially similar to the original, still constitutes infringement; Anthropic settled for $1.5 billion over pirated corpora, the largest copyright settlement in US history. Universal Harmony (Viewpoint 15) – Mechanism is correct, trend bet is wrong: In 2023, Wang Jianshuo said that ChatGPT uses a "weighted average" of human opinions, which can counteract the information cocoon of TikTok, giving the possibility of "universal harmony." The mechanism is correct – in 2025, multiple studies have conclusively confirmed that LLM pushes opinions towards the majority and systematically underestimates the minority. However, societal judgment backfired: his own addition, "at least not yet a thousand-person personalized experience," was overturned within three years—OpenAI made cross-conversation memory and personalization default capabilities starting in April 2025, and AI is rapidly moving towards a thousand-person personalized experience. Localized Warfare and Costs (Viewpoint 17)—Qualitatively accurate, quantitatively disproven: In 2023, Wang Jianshuo said: Building large-scale models would quickly degenerate into "localized warfare," with known costs (capped at around $500-1 billion after removing detours), attracting many players. The qualitative direction was remarkably correct—a large influx of players, rapid commercialization, and open-source catching up with closed-source—all came true. However, the hard figure of "$500-1 billion capped" was wrong on both ends: the cutting-edge side was severely underestimated (GPT-5 level training will reach $200-500 million in 2026, coupled with a data center worth hundreds of billions and Stargate worth $500 billion); the replica side was overestimated (DeepSeek reduced marginal training costs to the million-dollar level).Emergent Capabilities (Viewpoint 5) – The Direction is Correct, the Numbers and Framing are Wrong: In 2023, Wang Jianshuo stated that new capabilities appearing above approximately 60B parameters are not found in the original corpus and cannot be explained by researchers. While the directional intuition holds true, two statements are untenable: First, there is no unified "60B threshold"—the true threshold of the thought chain is approximately 100B, with different capabilities appearing at scales ranging from 13B to 540B; second, the "inexplicable" notion was challenged by a NeurIPS outstanding paper at the end of 2023—many "mutations" are illusions caused by the choice of evaluation metrics, and the curves become smoother and predictable after switching to continuous metrics. Fourth, looking back over three years, several patterns are reconciled one by one. Taking a step back, Wang Jianshuo's twenty judgments contain several patterns more worthy of being remembered than any single one. First, the direction is far more reliable than numbers and degrees. Of the twenty points, those concerning mechanisms and directions (RAG, LUI, robot networks, Turing test) were almost all correct; those providing specific figures or capping statements (100T parameters, 60B threshold, 500-1 billion cost, mathematical "impossibility") were almost all wrong. Second, regarding time, he tended to overestimate speed and underestimate extent. Anything said to be "quickly completed in two or three years" generally had a slower maturation period; however, he underestimated the ceiling of capability leaps—mathematics could go from "impossible" to an IMO gold medal, and frontier costs could rise to unimaginable levels. In short: too optimistic in the short term, too conservative in the long term. Third, the most insidious mistake repeatedly occurred in "distribution." It wasn't the direction that was wrong, but rather focusing only on total volume while ignoring distribution. "No unemployment wave" was correct, but the damage was highly concentrated among young newcomers; "value lies in the application layer" was half correct, but it failed to distinguish between the computing power layer and the model layer. Correct total volume masked the distribution disaster—this is the lesson most urgently needed to be learned. IV. Statements left open to interpretation will stand the test of time in three years. "Rumors," "at least now," "significantly reduced rather than eliminated," "two to three years in the early stages, about ten years to mature"—judgments made with qualifiers and tiers back then are more reliable in retrospect. Conversely, absolute statements made off the cuff are most prone to failure. Honest predictions are half about daring to speak and half about daring to acknowledge uncertainty. V. Some questions cannot be answered in three years. Who ultimately receives the value, whether emergence changes the truth, whether machines possess any consciousness, and whether long contexts will consume RAG—these debates from back then will still be debates in 2026. Being able to distinguish between "things that already have answers" and "things that still require waiting" is more important than rushing to conclusions about everything. Three years ago, Wang Jianshuo, based on intuition, pointed in twenty directions in the fog before GPT-4 even emerged.After reviewing the accounts today, the most important thing to remember is: It's not that hard to see the big picture; the difficulty lies in admitting your repeated assumptions about numbers, speed, and distribution. These twenty points are less about scoring the past and more about establishing rules for the next three years. Let's review them again in 2029. [Wang Jianshuo]