GPT之父把AI扔回1930年:没见过一行代码,却「发明」了Python

marsbitPublished on 2026-04-29Last updated on 2026-04-29

你敢信?一个活在95年前的AI,竟写出了Python代码。GPT之父下场,用2600亿Token炼出了一个「老古董」AI。

一个从未见过电脑的AI,竟写出了现代编程语言!

这可不是什么科幻的设定。

就在今天,GPT之父Alec Radford带队发布了震撼全网的「talkie」——

总参数130亿,一个只读过1931年之前旧文献的大模型。

talkie的「世界观」(全部训练数据),被冻结在了1930年12月31日。

那个时代,没有互联网,没有维基百科,更没有任何现代代码。

它读过的最「新」的东西,是近百年前的专利书、科学期刊、礼仪手册和私人书信。

但就是这样一个「活在95年前」的AI,居然能写出Python代码。

没学过编程,

却写出了Python,理解了「逆函数」

talkie最炸裂的发现,藏在一组编程测试里。

Alec Radford团队突发奇想,用HumanEval去测试talkie的编程能力——

给它几个Python函数作为上下文示例,然后让它解决新的编程问题。

要知道,talkie的训练数据中,没有任何一行现代代码。连数字计算机的概念,都不存在于它的「知识体系」中。

但结果令人震惊,通过少样本学习,它竟然能写出正确的Python程序。

虽然目前只能完成简单的单行程序,比如两个数相加,或者对上下文示例做微小修改。

Alec Radford:GPT、CLIP、Whisper背后核心大佬

但其中一个案例让人印象深刻:给定一个旋转密码的编码函数encode_shift,它的逻辑是把每个字母在字母表中向后移动5位。

talkie自己写出了对应的解码函数,整个修改只有一个字符:把+5改成了-5,加号换成了减号。

它真正理解了「逆函数」:加密是加,解密就是减」这个逆运算的概念。

传送门:https://talkie-lm.com/chat

2600亿Token,专喂百年前的纸

Alec Radford团队为什么要费这么大劲,手动OCR近百年前的物理文献,来训练一个「老古董」?

因为他们要回答AI领域最核心的一个问题:LLM的能力,到底是推理,还是背诵?

talkie可以写出Python,证明了——

LLM可以用19世纪的知识做推理,并非只是检索。不得不说,这才是真正意义上的「泛化」!

再来看talkie的训练语料库,可以称得上是一个庞大的「考古工程」。

它的训练语料达到了2600亿token,全部来自1931年之前的英语文本,包括书籍、报纸、期刊、科学论文、美国专利、判例法。

要知道,这么多文本皆需要从实体文档扫描并OCR转录。

而选择1930年作为截止日期,原因很实际:这是美国公共版权法(public domain)的分界线。

不过,这带来了一个意想不到的瓶颈:数据质量。

团队做了一组对照实验:用传统OCR系统转录的旧文本训练模型,和用人工转录的同一批文本训练模型相比,前者的学习效率只有后者的30%。

简单的正则清洗能把这个数字提升到70%,但仍然有巨大的差距。

在评估talkie性能实验中,团队又打造了一个「现代孪生体」(talkie-web-13b-base)。

后者用FineWeb的现代网络数据训练,两款模型用了「相同的算力」。

显而易见,在核心语言理解、数学推理任务上,talkie的表现与现代孪生体相当。

但在通用知识评测上,即使剔除了对1930年视角来说「穿越」的题目,talkie仍然落后。

团队怀疑,这跟数据质量有很大关系。

为此,Radford团队计划从零开始训练「复古OCR系统」,专门用于重新转录1931年前的文本。

用最现代的Claude 4.6

训练最古老的AI

talkie的「后训练」方案也很有意思。

要把一个只读过旧书的「基础模型」变成能对话的聊天机器人,根本没有现成的指令微调的数据可用。

团队的做法是,从1930年之前的结构化参考书中提取指令-回答对:礼仪手册、书信写作指南、菜谱、百科全书、诗歌集。

然后,再用这些「复古教材」做第一轮SFT。

在接下来的RLAIF阶段,团队用在线DPO来提升talkie的指令遵循能力,Claude Sonnet 4.6作为裁判。

一个2026年最先进的AI,给一个「活在」1930年的AI打分。

最终的精调阶段,团队甚至用Claude Opus 4.6生成多轮对话数据,来打磨talkie的对话能力。

训练过程中,Claude对talkie指令遵循能力的评分从2.0提升到了3.4(满分5分)。

最后一步,用Claude Opus 4.6与talkie进行多轮合成对话,再做一轮拒绝采样+SFT,打磨对话能力。

团队也坦承了一个讽刺之处:用现代大模型训练一个本该冻结在1930年的模型,本身就是一种「时间污染」。

他们的长期目标是用复古基座模型自身作为裁判,实现完全「自举式」的后训练流水线。

值得一提的是,talkie的7B版本在RL训练后出现了一个搞笑的副作用——

它开始用列表体说话,纯属是被现代AI的「坏习惯」传染了。

AI界最干净的一次「开卷考试」

研究团队还做了另一个有趣的实验。

他们从《纽约时报》的「历史上的今天」栏目中提取了近5000条历史事件描述,计算talkie对每条事件的「惊讶度」。

结果非常清晰,1930年之前的事件,talkie不太惊讶。1930年之后的事件,惊讶度开始攀升。

到了1950年代和1960年代达到峰值,然后趋于平稳。

这条曲线本身就是一个关于预测能力的实验。随着模型规模增大,这条曲线会怎么变化?

谷歌DeepMind CEO Demis Hassabis曾提出一个思想实验——

一个只训练到1911年的模型,能不能像爱因斯坦在1915年那样独立发现广义相对论?

talkie目前当然做不到。但它提供了一条路径,往上Scale就行了。

今夏扩展到GPT-3级别

talkie目前是130亿参数,团队的路线图相当激进——

今年夏天,发布GPT-3级别的复古模型。

更远期的目标:将语料扩展到超过一万亿token,理论上足够训练一个GPT-3.5级别的模型,能力接近初代ChatGPT。

一个冻结在1930年的ChatGPT。

参考资料:

https://x.com/status_effects/status/2048878495539843211?s=20

https://talkie-lm.com/introducing-talkie

本文来自微信公众号“新智元”,作者:新智元

Trending Cryptos

Related Reads

The "Impossible Triad" Is Fundamentally a Pseudo-Problem

The article argues that blockchain's fundamental limitation is not the scalability trilemma (decentralization, scalability, security), which has been largely solved, but the lack of **privacy** and, until recently, clear **legitimacy**. Blockchain is described as a slow, expensive, globally shared computer whose core value is censorship resistance and verifiability. While ideal for native digital assets like money (e.g., stablecoins), its default transparency acts as a **tax**, exposing all transactions and enabling MEV extraction, which deters serious institutional capital. Simultaneously, its permissionless nature created regulatory ambiguity. The piece contends that **privacy** is the missing critical feature. It rejects the false choice between total transparency and complete anonymity. Modern cryptography (like zero-knowledge proofs) enables **compliant privacy**: users can prove facts (solvency, KYC status, compliance) without revealing the underlying sensitive data (specific holdings, identities). This preserves auditability for regulators and eliminates the leak of financial information. With recent regulatory progress (e.g., the GENIUS Act) addressing legitimacy, adding default, provably compliant privacy becomes a pure upgrade. It transforms blockchain from a costly, public ledger into a confidential settlement layer, finally bridging the gap to mainstream institutional and individual adoption of on-chain finance.

链捕手4h ago

The "Impossible Triad" Is Fundamentally a Pseudo-Problem

链捕手4h ago

Optical Chips: Collective Capacity Expansion

The global optical chip industry is experiencing a massive wave of expansion driven by surging AI data center demand. Major players across the US, Japan, Europe, and China are aggressively investing to ramp up production capacity. In the US, Coherent is expanding its 6-inch Indium Phosphide (InP) semiconductor fab in Texas, supported by CHIPS Act funding and a $2 billion strategic investment from NVIDIA. Lumentum is building a new factory for InP optical devices, and Nokia is scaling its advanced photonic chip packaging and testing capabilities. NVIDIA's investments aim to secure future supply of critical lasers and optical interconnect products for AI infrastructure. Japan's JX Advanced Metals, a leading InP substrate supplier, plans a multi-billion yen investment to increase its capacity 7-10 times, strengthening its grip on the crucial upstream materials market. In Europe, IQE and Tower Semiconductor settled a patent dispute and signed a multi-year InP epitaxial wafer supply agreement, highlighting that next-generation silicon photonics platforms will integrate high-performance InP components. STMicroelectronics and Sivers Semiconductors are also expanding silicon photonics production and partnerships. China is rapidly building out its domestic supply chain. Dongshan Precision's subsidiary, Source Photonics, announced a $12 billion project to expand optical chip and module production. Companies like Sanan Optoelectronics and Yunnan Germanium are scaling up InP chip manufacturing and substrate production, moving towards vertical integration from materials to modules. While debate continues around the exact future architecture—whether CPO (Co-Packaged Optics), NPO, or pluggables will dominate—analysts like Morgan Stanley argue the underlying driver is unchangeable: the explosive growth in bandwidth demand. This will inevitably increase the volume of optical engines, lasers, and related content per GPU, regardless of the final technical path. The competition for "more light" in the AI era has intensified into a global, full-chain capacity race.

marsbit7h ago

Optical Chips: Collective Capacity Expansion

marsbit7h ago

Stablecoins Finally Find Real Yield: An In-Depth Look at On-Chain Reinsurance Re | A Conversation with Re Founder Karan Saroya

Stablecoin Real Yield Found: A Deep Dive into On-Chain Reinsurance with Re's Karan Saroya As stablecoin supply exceeds $170 billion, the search for sustainable, non-speculative yield intensifies. Re, an on-chain reinsurance platform, provides an answer: connecting stablecoin capital to the trillion-dollar traditional reinsurance market. Re operates as a regulated reinsurer, accepting stablecoin deposits as collateral to back US insurance companies. These insurers pay premiums, generating yield that flows back to on-chain depositors. Currently supporting 35 insurers and underwriting $500 million, Re projects scaling to over $1 billion soon. Key insights from a Bankless podcast with founder Karan Saroya and investor Avichal of Electric Capital: 1. **Uncorrelated, Real-World Yield:** Re offers stablecoin holders access to reinsurance returns (targeting 12-14%+), an asset class entirely separate from crypto or equity markets. 2. **Operational Efficiency via Smart Contracts:** Re replaces traditional, labor-intensive capital fundraising with smart contracts, allowing a ~12-person team to compete with industry giants. 3. **Regulatory Leverage:** For every $1 of collateral, regulations allow backing $5-7 in written premiums. This leverage amplifies returns from the underlying risk-free rate. 4. **DeFi Integration:** Depositors receive receipt tokens, which can be used in protocols like Morpho for "looping," potentially pushing yields to 18-20%+. 5. **The "DeFi Mullet" Model:** A compliant front-end (regulated reinsurer) paired with a decentralized back-end (smart contracts, DeFi capital markets). 6. **RE Governance Token:** Modeled on Lloyd's of London, the token governs the central capital pool's allocation, counterparty acceptance, and parameters. 7. **Real Economic Impact:** Capital funds real-world productivity (factories, clinics, businesses) via insurance, moving beyond crypto's internal loops. The discussion highlights a pivotal moment: DeFi's supply-side infrastructure is now met by real demand for productive yield, potentially kickstarting a flywheel where vast on-chain stablecoin capital seeks these real-world returns.

链捕手8h ago

Stablecoins Finally Find Real Yield: An In-Depth Look at On-Chain Reinsurance Re | A Conversation with Re Founder Karan Saroya

链捕手8h ago

1996 or 1999? Walsh's First Test is 'How to View AI'

"1996 or 1999? Wall's First Big Test Is 'How to View AI'" Federal Reserve Chairman Wall's initial challenge is not whether to raise or cut rates, but a more fundamental judgment: what kind of boom is the current AI boom? This will determine the Fed's policy path and define his legacy. Economics is split between two opposing views, according to reporter Nick Timiraos. One sees imminent productivity gains that will increase supply and cool inflation, allowing the Fed to hold steady. The other argues that while productivity benefits are distant, demand shocks are here now, and waiting for data confirmation risks missing the intervention window, forcing sharper rate hikes later. Wall has signaled a leaning toward the first view, echoing 1996-era Alan Greenspan, who embraced strong, productivity-driven growth without fear of inflation. However, Wall faces a different macro environment than Greenspan did, with tariff pressures, expanding fiscal deficits, and diminishing globalization benefits, which could force more significant inflation pressures even if AI benefits materialize. Wall's logic, expressed before taking office, is that AI-driven productivity gains won't show in official data for years. If the Fed waits for confirmation, it might mistakenly tighten policy and choke off the very growth that could suppress inflation. This argues for using forward-looking narratives over lagging data. Chicago Fed President Austan Goolsbee presents a key counter-argument. He distinguishes between expected and unexpected productivity booms. A widely anticipated boom, like the current AI wave, can cause people to spend future wealth gains in advance, overheating the economy before productivity actually rises, thus requiring preemptive rate hikes. He cites rising costs for AI data centers as evidence of such overheating. Fed Governor Christopher Waller offers a rebuttal to Goolsbee, noting the "expected spending" mechanism only works if people can borrow against future income, which many households cannot do due to borrowing constraints. Wall also faces a paradox related to his desire to reduce the Fed's use of "forward guidance" (pre-announcing policy moves). This practice was established in 1999 when Greenspan began signaling hikes to avoid market shocks. If the economy follows a less optimistic path, Wall may be forced to choose between using the guidance he wants to abolish or risking market volatility by staying silent. The ultimate question defining Wall's first major test remains: Is this 1996 or 1999?

marsbit9h ago

1996 or 1999? Walsh's First Test is 'How to View AI'

marsbit9h ago

Trading

Spot
Futures

Hot Articles

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of AI (AI) are presented below.

活动图片