蓝精灵

研究发现婴儿语言学习能力远胜 AI 聊天机器人，学习效率也大幅领先_我的网站

绅士的品格

A | 8 月 26 日消息，如果把“达到人类水平的智能”作为衡量标准，AI 聊天机器人距离目标仍然很远。 Dark clouds loomed over the release of Balakrishna’s film Akhanda 2 after Madras High Court stalled the film’s release on Wednesday. While the producers of the action entertainer tried their best to bail out the film, their efforts went in vain. As a result, Akhanda 2, which was slated to release on Friday (December 5, 2025), isn’t releasing as scheduled. The production house, 14 Reels Plus, officially confirmed the news on their social media handle. The unfortunate development certainly came in as a big disappointment for the fans who have been waiting with bated breath to see their favourite star (Balakrishna) in action. With a heavy heart, we regret to inform you that #Akhanda2 will not be releasing as scheduled due to unavoidable circumstances.This is a painful moment for us, and we truly understand the disappointment it brings to every fan and movie lover awaiting the film.We are working… — 14 Reels Plus (@14ReelsPlus) December 4, 2025 Akhanda 2 release postponed indefinitely At a time when the buzz about Akhanda 2 reached fever pitch, the release of the Balakrishna-starrer was halted. A statement from the filmmakers read, “With a heavy heart, we regret to inform you that Akhanda 2 will not be released as scheduled due to unavoidable circumstances. This is a painful moment for us, and we truly understand the disappointment it brings to every fan and movie lover awaiting the film.” While tendering apologies and the inconvenience caused in this regard, the filmmakers assured fans that they are working towards resolving the matter and promised to share a positive update soon. Legal dispute In 2019, an arbitral tribunal directed 14 Reels Entertainment to pay Eros International Rs 11.23 crore over a financial and rights-related agreement emerged from the film Dookudu. In 2020, The Madras High Court had also dismissed 14 Reels Entertainment’s petition challenging the earlier award. A Division Bench, and subsequently the Supreme Court too dismissed 14 Reels Entertainment’s SLP in 2021. Meanwhile, Eros International too did not appeal against 14 Reels Entertainment for ‘contempt of court’ for four years.In August 2025, Eros learnt that the debtor company (14 Reels Entertainment) is making and releasing a film Akhanda 2 with a new legal entity named 14 Reels Plus LLP. The Applicant moved Madras High Court arguing that the LLP was controlled by family/associates of the same promoters of 14 Reels Entertainment. Eros further alleged that the promoters were attempting to bypass an unpaid arbitral sum and described it as an alter ego. The Court also heard the arguments of the Respondents 14 Reels Entertainment and 14 Reels Plus LLP. After hearing all the arguments, setting a single Judge’s order (of setting Akhanda 2 free from all legal constraints in October 2025) aside, The Division Bench of the Court freshly ordered that Akhanda 2 cannot be released, distributed, streamed unless 14 Reels Entertainment / 14 Reels Plus LLP pays the entire arbitral payment with interest. Akhanda 2 premieres cancelled The premieres of Akhanda 2 were scheduled on December 4 at various screens in the Telugu speaking states and also abroad (North America and Europe). But they had to be cancelled at the last minute due to court restraint. Also, the online bookings of the film on various ticketing portals have been kept on hold. Fans have been expressing their disappointment and are hoping that the film will see the light at the earliest. Also Read: Madras High Court Puts Balakrishna’s Akhanda 2 Release on Hold。

B | 过去几年，AI 行业取得明显进展，聊天机器人已经能够生成看起来相当接近人类表达的句子。但在推理能力和实际生产力方面，AI 仍有很大提升空间。那如果比较“学会说话”的能力，今天的大语言模型和人类相比又如何？答案同样不乐观。当地时间 24 日，《麻省理工科技评论》的报道分析了 AI 训练过程，并采访了多位研究大语言模型效率的专家。一个很直观的结论是：婴儿学习语言的能力远胜今天的 AI 聊天机器人，而且学习效率也高得多。斯坦福大学认知科学家迈克尔 ·C· 弗兰克坦言，科技行业最近取得的进展非常惊人。“但即便经历了过去几年的快速发展，我们仍然得烧掉一片森林，再搜刮几乎全部人类知识，才能重现一个在普通家庭客厅里一年左右就会自然发生的里程碑。”研究人员目前仍不完全清楚原因，但婴儿在语言学习上表现出极强的适应能力，很快就能掌握语言规则和词汇。大语言模型却需要规模惊人的语言数据，才能达到类似熟练程度，所需数据量远高于普通人类。弗兰克形容人类学习语言的能力“简直像奇迹一样”。他指出，幼儿从周围环境中接触大约 1000 万至 3000 万个词后，就已经能开始说出清楚、可理解的句子。“如果你只用 3000 万个词训练 GPT-2，得到的会是一个胡言乱语生成器，而不是一个孩子。”从报道中获悉，这种效率差距已经成为 AI 行业必须面对的一大难题，尤其是那些希望单纯通过不断扩大语言模型规模，最终实现人类水平 AI 的科技公司。