跳到正文
Hacker News:AI 熱帖·· 2026-10-16AI 評分40

Sergey Brin 稱 Google 的 AI 模型正接近遞歸式自我改進

Sergey Brin says Google's AI models are reaching recursive self-improvement

AI 導讀

Sergey Brin 表示,Google 的 AI 模型正逐步接近遞歸式自我改進。Hacker News 討論者質疑,這是否只是使用 LLM 準備下一個模型的訓練數據,並指出相關改進是否有可靠的獨立回饋仍受質疑。留言亦提醒,這不代表模型必然能實現 AGI。

正文

Image 1Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

Sergey Brin says Google's AI models are reaching recursive self-improvement (businessinsider.com) 21 points by ryan_j_naughton9 hours ago | hide | past | favorite | 22 comments help Image 2kennywinker6 hours ago | next[–] Does this not just mean they’re using LLMs to prep training data for the next LLM to be trained on? Because it sounds like that, reading between the lines of hype, i mean. reply Image 3brazukadev24 minutes ago | parent | next[–] it is probably about coding its own harness and being able to assess it to improve. reply Image 4lazide28 minutes ago | parent | prev | next[–] it just feeds into the ouroboros issue, however, as it lacks reliable independent feedback. kind of like grading your own tests in school. if you’re really honest, hey maybe. and it’s certainly faster for the teacher! reply Image 5avree7 hours ago | prev | next[–] Does Sergey still work in the day to day at Google? I thought he mostly stepped back after that affair he had with the product marketing woman for the Google Glass project. reply Image 6solyin7 hours ago | parent | next[–] https://finance.yahoo.com/news/google-cofounder-sergey-brin-... He's back, and working 60 hour weeks, since 2023, according to this report. reply Image 7kennywinker6 hours ago | root | parent | next[–] I simply cannot imagine being a centibillionaire and working 10 hours a week, much less 60. reply Image 8pllbnk6 hours ago | root | parent | next[–] "Working" is a generous word for his position. The "work" involves flying with private jet to expensive places regular people have never stepped foot in for noncommittal meanings, showing up around the office at random times, have a word with random high-level managers in the office. All the while not having the slightest thought about someone evaluating your performance or output. reply Image 9poulpy1231 hour ago | root | parent | next[–] I would do it for the money but it's not a life I would choose if I was already rich reply Image 10gregglain4 hours ago | root | parent | prev | next[–] Performance review 2025: yes Refresher: a few lambos annual raise, small town as a bonus, larger town in RSUs reply Image 11cedws2 hours ago | root | parent | prev | next[–] Also the part where you have three meals a day prepared for you, children’s needs taken care of, laundry washed ironed and folded, house cleaned. Working “hard” as a billionaire doesn’t mean shit. reply Image 12treebeard9016 hours ago | prev | next[–] It's not guaranteed that a next token prediction machine is capable of recursive self improvement. Or will be able to create AGI in any form. reply Image 13StevenWaterman6 hours ago | parent | next[–] That's why it's news reply Image 14kennywinker6 hours ago | root | parent | next[–] It’s also not guaranteed that owners are reliable narrators, when speaking about how great their business is doing. In other breaking news, worlds best cup of coffee found in truckstop diner! reply Image 15seanhunter7 hours ago | prev | next[–] Maybe they will recursively self-improve to the point where they are genuinely competitive with Anthropic and OpenAI, or where they “AI summaries” they add to search results aren’t total garbage most of the time. reply Image 16vrganj5 hours ago | prev | next[–] Snake Oil vendor says his oil cures cancer. reply Image 17chaostheory7 hours ago | prev | next[–] If Argon was impressive, I would maybe believe it. Right now, argon is not even close to Opus 5.5 or even 6.1 Sol. It can’t even infer intent as well as Gemini 3.1 Pro. Google has yet to release a frontier class model imo reply Image 18smashers11146 hours ago | parent | next[–] I disagree somewhat. I think internally OpenAI, Anthropic and Google are probably all pretty close right now. Argon is competitive with Opus 5.5 and Astra. It might be third, but its competitive at least. Its becoming more apparent that different models are RL'd for different tasks. Astra (or broadly openAI) are best for mathematics, 3D modeling, and vision. Opus 5.5 is best at like visual design (or just in general adding more taste and feel) Now, it can depend on your definition of frontier class model. But Argon is a strong model. Google is not running a distant third. If OpenAI and Anthropic shut down they would have the strongest model on the market in like 2 months imo. Google does suck at shipping (maybe due to an unusual amount of dogfooding models internally?) And we know all the labs have stronger internal models than their public releases. reply Image 19sicktriple8 hours ago | prev | next[–] guy who sells AI says AI is cool reply Image 20leecommamichael7 hours ago | prev | next[–] ok reply Image 21ares6238 hours ago | prev | next[–] Oh I'm surprised to learn that their engineers and scientists had 0% usage of their tools before! reply Image 22pbiggar6 hours ago | prev[–] Reminder Sergey Brin was in the Epstein files. reply Image 23ZeroGravitas5 hours ago | parent[–] And his wife was RFK's running mate and main funder when he ran for President on a platform of vaccines causing autism. reply Image 24 Consider applying for YC's Winter 2027 batch! Applications are open till November 2. Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact Search:

來源:Hacker News:AI 熱帖 · news.ycombinator.com