「AI 吞噬世界」全景图系列 · 科研篇 · SCITHE «AI IS EATING THE WORLD» ATLAS · SCIENCE

最先被规模化的不是发现,
是发表The First to Scale Was Not Discovery —
It Was Publishing

一次发现的一生 · The Life of a Discovery · 信任基础设施先于生产方式崩塌The Life of a Discovery · the trust infrastructure fell before the mode of production

AI 冲击科研的顺序被搞反了:真正的 AI 科学发现还在早期——自动实验室只在封闭化学里跑通,AI 科学家写出的还只是「像论文的东西」;但 AI 已经让论文、综述、审稿、基金申请与论文工厂全部批量化。于是一个残酷的错位出现了:科研的信任基础设施,先于科研的生产方式崩塌。AlphaFold 让生物学改朝换代、数学跨过 IMO 金牌线是真的;撤稿破万、作者在论文里埋暗号钓 AI 审稿人,也是真的。AI 一边让「做研究」变廉价,一边让「信任研究」变昂贵——而科学的地基,恰恰是信任。AI hit science in a reversed order: real AI discovery is still early — self-driving labs run only in enclosed chemistry, the AI scientist writes only «paper-shaped text» — yet AI has already batch-produced papers, reviews, referee reports, grant applications and paper mills. So a cruel dislocation appears: science’s trust infrastructure collapsed before its mode of production. AlphaFold rewriting biology and mathematics crossing the IMO gold line are real; retractions past ten thousand and authors burying prompts to hook AI reviewers are also real. AI made «doing research» cheap while making «trusting research» dear — and trust is exactly the ground science stands on.

本图相信THIS MAP BELIEVES

AlphaFold 改写结构生物学是完成时(2 亿+ 结构,2024 诺奖);数学跨过 IMO 金牌线是真的(2025 官方评分 35/42);领域基础模型已重定义「分析」这一节点;评审侧被主流机构一致禁 AI 是「信任不能外包」的制度表态(NIH/NSF/ERC);瓶颈正从「产出」后移到「验证」。AlphaFold rewriting structural biology is the completed tense (200M+ structures, the 2024 Nobel); mathematics crossing the IMO gold line is real (an official 35/42 in 2025); domain foundation models have redefined the «analysis» node; the referee side being uniformly forbidden AI (NIH, NSF, ERC) is an institutional statement that trust cannot be outsourced; and the bottleneck is moving from output to verification.

本图不相信THIS MAP DOES NOT BELIEVE

不信「AI 科学家要取代科学家」——自动实验室与端到端 agent 仍是早期 demo,「像论文」≠「可靠贡献」(Sakana 仅过 workshop);不信「AI 让科研更平权」的单面叙事——执行平权的同时资源在向大机构集中;不信「AI 加速发现」不谈代价——同一工具也在制造灌水洪流与方法论单一种植;不信任何绕开诚实层的进步叙事——古德哈特、复现危机、经费荒诞三伤,AI 每一处都放大了。Not «the AI scientist will replace the scientist» — self-driving labs and end-to-end agents remain early demos, and «paper-shaped» ≠ «reliable contribution» (Sakana only cleared a workshop); not the one-sided «AI democratises research» — execution levels while resources concentrate in big institutions; not «AI accelerates discovery» without its cost — the same tool floods the literature and breeds methodological monoculture; and no progress narrative that skips the honesty layer — Goodhart, the replication crisis and grant absurdity, AI amplified every one.

期刊蓝 = 发现流水线(能力真相)Journal blue = the discovery line (capability) 荧光黄 = 需质疑/待核/单源标记Highlighter = the query / verify / single-source mark 撤稿红 = 学术生产与硬骨头Retraction red = production & hard bones [A–D] = 引文角标式证据分级[A–D] = citation-style evidence grades

[B]2 亿+AlphaFold DB 收录的蛋白结构预测(EBI 官方)——AI for Science 基础设施化的锤Protein structures in the AlphaFold DB (EBI) — the hammer of AI-for-Science becoming infrastructure

[B]35 / 422025 IMO 金牌分(Gemini Deep Think,官方评分)——数学前线的真实进度The 2025 IMO gold score (Gemini Deep Think, official marking) — the real progress on the maths front

[B]1 万+2023 全球撤稿数,创历史新高(Nature)——信任危机的体温Global retractions in 2023, an all-time high (Nature) — the fever of the trust crisis

[B]42%PI 联邦资助时间耗于行政事务(FDP 调查)——科研荒诞的诚实层Of a PI’s federally funded time spent on administration (FDP survey) — the honesty layer of science’s absurdity

⚠️ 口径裁判(先读):① 2 亿+ = AlphaFold DB 累计预测结构数(EBI),非「已实验验证」;② 35/42 = IMO 官方评分(Gemini Deep Think),OpenAI「金牌级」为自评未走官方通道;③ 1 万+ = 2023 全球撤稿(Nature),是「被发现的部分」,文献中潜藏的论文工厂论文远多于此;④ 42% = FDP 调查(PI 自报行政时间占比),与「grants 占 15.4%」「基金会 15–20h/联邦 100h」是不同口径勿混写;⑤ GNoME「220 万晶体/38–40 万稳定」是预测非实验确认;⑥ Columbia 假引「12 倍/每 277 篇」、芝加哥「×3.02/主题-4.63%」、Erdős「353 解 9」均单源不可溯,作趋势示意非精确账;⑦ 各节点渗透%为编者评估值,非统计值;⑧ 四份来源一份句中截断且引文列表丢失、一份自认从记忆写(数字系转录),已如实降级并列上线前必核清单;框架级共识多为委托预设论纲回声,已声明。⚠️ The basis rulings (read first): ① 200M+ is the AlphaFold DB’s cumulative predicted structures (EBI), not «experimentally verified»; ② 35/42 is the IMO’s official marking (Gemini Deep Think), while OpenAI’s «gold-level» is self-assessed, off the official channel; ③ 10,000+ is 2023 global retractions (Nature), «the part that got caught» — the paper-mill papers latent in the literature far exceed it; ④ 42% is the FDP survey (a PI’s self-reported admin time), a different basis from «grants at 15.4%» or «15–20h foundation / 100h federal»; ⑤ GNoME’s «2.2M crystals / 380–400k stable» is predicted, not experimentally confirmed; ⑥ the Columbia fake-citation «12×/1-in-277», Chicago’s «×3.02 / −4.63% topics» and the Erdős «9 of 353» are all single-source, untraceable — shown as trend, not precise ledger; ⑦ node penetration %s are editorial estimates; ⑧ of four sources one truncates mid-sentence losing its references and one wrote from memory (figures transcribed), both downgraded and listed for verification; the framework consensus largely echoes the commissioning brief, declared.

中心装置 · 假设-实验循环与时态进度条THE CENTRAL DEVICE · THE HYPOTHESIS LOOP & THE TENSE BAR

AI 让「做研究」变廉价,让「信任研究」变昂贵AI made doing research cheap, and trusting research dear

科学研究的最小单元不是「论文」,而是一次「假设 → 实验/推演 → 证据 → 修正」循环。学科差别只在「实验」这一步的物理形态:数学=逻辑推演(可机器完全检验)、实验科学=湿实验(受物理世界约束)、计算科学=大规模仿真(受算力数据约束)、人文社科=田野与文本(受语境伦理约束)。一条铁律贯穿全图:证据能被机器闭环检验的地方,AI 跑得最快;证据必须落到物理世界的地方,AI 卡在接口。而 AI 真正规模化的不是这个循环的「发现」端,是它的「发表」端——所以读这张图要先分清「时态」。Science’s atomic unit is not the «paper» but one loop of hypothesis → experiment/derivation → evidence → revision. Disciplines differ only in the physical form of the «experiment» step: mathematics is logical derivation (fully machine-checkable), experimental science is wet-lab work (bound by the physical world), computational science is large-scale simulation (bound by compute and data), the humanities are fieldwork and text (bound by context and ethics). One iron law runs through the map: where evidence closes in a machine loop, AI runs fastest; where evidence must land in the physical world, AI jams at the interface. And what AI truly scaled is not the loop’s «discovery» end but its «publishing» end — so reading this map begins with telling the tenses apart.

发现端 · 早期The discovery end · early

自动实验室只在封闭化学/材料跑通(仅 1% 生命科学从业者认可湿实验 AI 价值);AI 科学家能端到端写出「像论文的东西」,但那与「可靠科学贡献」隔着整个科学。Self-driving labs run only in enclosed chemistry and materials (just 1% of life scientists credit wet-lab AI); the AI scientist writes «paper-shaped text» end to end, and between that and a reliable contribution lies the whole of science.

分析端 · 完成时The analysis end · completed

AlphaFold(2 亿+ 结构)、GNoME(220 万晶体)、数学跨 IMO 金牌线——领域基础模型已经重定义了「什么叫分析」;2024 诺贝尔化学奖为这条分界线盖章。AlphaFold (200M+ structures), GNoME (2.2M crystals), maths across the IMO gold line — domain foundation models have redefined «what analysis means»; the 2024 chemistry Nobel stamped the divide.

发表端 · 已崩The publishing end · fractured

论文、综述、审稿、基金申请全部批量化;撤稿破万、论文工厂工业化、隐藏 prompt 攻击审稿人——瓶颈从产出后移到验证,信任成了下一个十年的战场。Papers, reviews, referee reports and grants all batch-produced; retractions past ten thousand, paper mills industrialised, hidden prompts hooking reviewers — the bottleneck moved from output to verification, and trust became the next decade’s battlefield.

时态进度条 · 读这张图先分清 AI 在哪个时态(渗透为评估值)THE TENSE PROGRESS BAR · TELL THE TENSE FIRST (PENETRATION = ESTIMATES)

过去时 · 已盖章Past tense · stamped

AlphaFold 改写结构生物学、GNoME 材料发现、GraphCast 秒级预报——分析范式已换代,2024 诺奖公共事实;这一格没有争议,是能力真相。AlphaFold rewriting structural biology, GNoME finding materials, GraphCast forecasting in seconds — the analysis paradigm has turned over, a public fact of the 2024 Nobel; no dispute here, only capability.

进行时 · 荧光待续Present · highlit, ongoing

数学跨 IMO 银牌→金牌线(唯一可机器完全检验的前线)、AI co-scientist 假设生成——机器管「证明对不对」,人管「哪个猜想值得证」:Lean 不会被漂亮语言说服。Maths crossing the IMO silver-to-gold line (the one fully machine-checkable front), AI co-scientists generating hypotheses — the machine judges «is the proof correct», the human judges «which conjecture is worth proving»: Lean is not persuaded by fine prose.

早期 · demo 水印Early · watermarked demo

自驱动实验室(封闭场景)、端到端 AI 科学家(仅过 workshop)——demo 惊艳 ≠ 规模化;这是「早期」,不是「过去时」:能力证据是真的,替代是假象。Self-driving labs (enclosed scenes), the end-to-end AI scientist (only a workshop cleared) — a dazzling demo ≠ scale; this is «early», not «past»: the capability evidence is real, the replacement an illusion.

慢变量 · 制度题Slow variable · institutional

同行评审、复现、评价、经费——这套信任基础设施被顶到台前,却不随算力自动改变:破四唯/DORA 的替代标准缺位,是制度题不是 AI 题。Peer review, replication, evaluation, funding — this trust infrastructure is pushed to the front and does not change with compute: the missing replacement standard for «beyond counting papers» is an institutional question, not an AI one.

读图法一句话:先分时态,再谈进度——把「AI 科学家要来了」的炒作(早期)与「学术信任正在崩塌」的现实(已崩)对调过来,就看清了 sci 图的反直觉入口。The reading in one line: tell the tense before the pace — swap the «AI scientist is coming» hype (early) for the «academic trust is collapsing» reality (fractured), and the map’s counter-intuitive entrance comes clear.

Reading the MapReading the Map

从这张图带走的五条规律Five patterns to take away

立场声明:本页是批判性、祛魅的行业结构分析——四份深度研究交叉整理(一份句中截断且引文列表丢失、一份自认从记忆写数字系转录,均已声明并降级;框架级共识多为委托预设论纲回声,已如实声明);厂商效率数字一律 D 级,渗透%为评估值,单源惊人数字标荧光黄待核。本页提供行业结构信息,不构成科研、投稿、评审或投资建议;对被压缩的青椒与博士生不猎奇不俯视——被商品化的是执行,不是研究者的价值。核心判断一句话:AI 让「做研究」变廉价、让「信任研究」变昂贵——最先崩的是发表,最后失守的是署名。 Stance: a critical, demystifying structural analysis — cross-compiled from four deep-research reports (one truncated mid-sentence with references lost, one written from memory with transcribed figures — both declared and downgraded; the framework consensus largely echoes the commissioning brief, declared honestly); vendor efficiency figures are D-grade throughout, penetration %s are estimates, and single-source striking numbers wear the highlighter «verify» mark. Industry-structure information only — not research, submission, review or investment advice; no voyeurism toward squeezed junior faculty and doctoral students — what got commodified is execution, never the researcher’s worth. The core judgment in one line: AI made doing research cheap and trusting research dear — publishing fell first, and authorship holds last.