O3-mini’s Chinese outputs spark claims of hasty OpenAI copy-paste release

Description

The image is a dark-mode screenshot of two stacked tweets inside the X (formerly Twitter) UI. The first tweet from “DaiWW @BeijingDai” reads: “The latest OpenAI O3-mini is reasoning in Chinese,it seems OpenAI copied Deepseek's open-source code/data and released the O3-mini without carefully edit.” Below it is an embedded tweet from “Vikhyat Rana @The_Vikhyat” saying “Why O3-mini is reasoning in Chinese 🥺”, accompanied by a cropped chat window where the assistant’s entire chain-of-thought appears in Chinese characters even though the original prompt was in English. The juxtaposition humorously implies that OpenAI accidentally shipped a model with DeepSeek’s Chinese system prompts or datasets still intact, highlighting the risks of re-using open-source weights and insufficient internal QA. For seasoned engineers, it illustrates how AI model releases can exhibit surprising localization bugs, provenance concerns, and the perennial challenge of quality‐gating large language models before production deployment

Common Tags

Openai AILimitations AI AIGeneratedContent OpenSource UnusualBugs

Comments

Anonymous ★ Top Pick OpenAI’s O3-mini launch plan: `git clone deepseek && sed -e 's/公司/OpenAI/g' | kubectl apply -f -` - English localization penciled in for the next sprint

Anonymous ★ Top Pick

OpenAI’s O3-mini launch plan: `git clone deepseek && sed -e 's/公司/OpenAI/g' | kubectl apply -f -` - English localization penciled in for the next sprint
Anonymous

Nothing says 'we built this from scratch' quite like your cutting-edge reasoning model suddenly channeling its inner Deepseek and explaining PyScript in Mandarin - it's like finding //TODO: remove before production comments, except the comment is an entire training dataset's linguistic fingerprint
Anonymous

When your 'novel reasoning architecture' accidentally ships with the original developer's debug logs in Mandarin - turns out the real O(1) complexity was in the copy-paste operation. Nothing says 'we trained from scratch' quite like your model thinking out loud in someone else's native language
Anonymous

i18n is hard: the UI speaks English, but O3‑mini’s chain‑of‑thought shipped in zh_CN - classic “we rebased a fork, changed the logo, and forgot LANG=…” supply‑chain tell
Anonymous

Distilling OSS without sanitizing comments: Your CoT now leaks provenance in every dialect
Anonymous

When your model starts reasoning in Chinese, that’s not AGI; it’s a leaky preprocessing pipeline, a missing LANG unit test, and a ‘borrowed’ repo that slipped through code review
@TERASKULL 1y

yea it randomly starts reasoning in german mid-thought sometimes
1. @RiedleroD 1y
  
  🔫 translate all of that or perish /j
2. @RiedleroD 1y
  
  jokes aside, I wonder why it does that
  
  @TERASKULL 1y
  
  black box problem, we might never know😔
  
  @RiedleroD 1y
  
  yeah. imo closed-source AIs shouldn't exist
  
  @TERASKULL 1y
  
  not only that, but you can't really check on the individual neurons to know why exactly that output was produced
  
  @RiedleroD 1y
  
  that too I guess, but there's been some advancements in that field from what I've seen. I try to keep my distance from AI though
  
  @TERASKULL 1y
  
  who isn't, really? all you can hear now is ai. blockchain 2.0
  
  Женя 1y
  
  You keep distance from AI, keep distance from blockchain but why?
  
  @RiedleroD 1y
  
  as for blockchain - nah that one's just gonna quietly disappear. I don't think that tech is useful for anything
  
  @sylfn 1y
  
  ...or loudly as quantum computers start breaking ECC
  
  @RiedleroD 1y
  
  eh, I think there's larger concerns once that happens. Regardless, I think there's a lot of time to prepare quantum-resistant algorithms
  
  @sylfn 1y
  
  less than a decade according to NIST
  
  @RiedleroD 1y
  
  oh, it's always less than a decade away. I don't believe them
  
  @sylfn 1y
  
  2030 is nearer than you think
  
  Женя 1y
  
  Nah, there’s no way It gonna be standard for all govs going through transformation in coming two decades, theres no way to fight corruption on modern level while public spending is hidden in the same way as private spendings
  
  @RiedleroD 1y
  
  austria doesn't hide public spending. will post a link to the statistics later
  
  Женя 1y
  
  That’s really cool and Austria is cool But it’s a very small state. Bigger the state, higher amount of spending and also easier to hide some of them
  
  @RiedleroD 1y
  
  would you rather we try to expand?
  
  @viktorrozenko 1y
  
  Yes
  
  @RiedleroD 1y
  
  something something hitler idk
  
  @RiedleroD 1y
  
  'twas a joke
  
  @viktorrozenko 1y
  
  I know
  
  Женя 1y
  
  It’s not like any of them are going to disappear in 5, 10 or 20y from now
  
  @RiedleroD 1y
  
  no, but the economic bubble will burst at some point and I'm not sure what'll come out of it, but I think it's a good idea to wait with picking AI up again until the fog has cleared
  
  Женя 1y
  
  The biggest economic bubble is fiat system which allows governments to print money indefinitely and spend them on things which either not needed by almost anyone or costing gov x10-x100 of fair price :/ Just check how bigly grew the gap in between real prices (eg real estate) and real income of people. And taxes were much lower around the globe back in 70s
  
  @RiedleroD 1y
  
  bro I can rent an affordable place here in vienna no problem. rent prices are that high in other places because people are free to speculate with real estate. in vienna, there's a lot of government-regulated housing, as well as strong consumer laws, keeping the prices competitive.
  
  @viktorrozenko 1y
  
  how would one make money from it then? if you remove intellectual property from the equation, OpenAI is just a server farm.
  
  @RiedleroD 1y
  
  that's all they are anyway
  
  @qtsmolcat 1y
  
  Most people won't want the hassle of setting up and running an LLM locally
  
  @viktorrozenko 1y
  
  Right, but that doesn’t mean that OpenAI is a server farm. They still did research, engineering, and other things before the LLM was there. Of course, when the product already exists it becomes a matter of hosting it somewhere, but it’d be a huge oversimplification to say that that’s all that is to it.
  
  @RiedleroD 1y
  
  they mostly did marketing and throwing a bunch of brute force compute power at the problem yknow, a few years before OpenAI was founded, the big strides in AI research was done by singular people just fucking around. I don't think the small advancements OpenAI has done required much effort at all. the biggest effort was just training the models, and that's server farm -ass activity
  
  @loomingsorrowdescent 1y
  
  But they don't, OpenAI hasn't turned a profit yet
  
  @viktorrozenko 1y
  
  I’m saying that if you prohibit closed source models they won’t be able to
  
  @RiedleroD 1y
  
  good
  
  @viktorrozenko 1y
  
  Why is that good? People need money to do what they do
  
  @RiedleroD 1y
  
  womp womp, I need money too. let them work for it
  
  @viktorrozenko 1y
  
  And they did. People pay them money because they did. Money talks
  
  @RiedleroD 1y
  
  sweetie, people also gave russia's oligarchs a lot of money. you don't get rich by doing honest work, you get rich by deceiving people. which OpenAI is in the business of
  
  @viktorrozenko 1y
  
  People don’t choose to give them money. They get money because they play well with the government. This comparison is invalid unless you can prove that OpenAI uses the government to coerce people into paying them their monthly subscription 🤣
  
  @RiedleroD 1y
  
  and OpenAI gets money because they manipulate other companies into integrating their AI into their products. do you know how annoying it is to see some bullshit AI everywhere and know that the company hosting that is paying OpenAI through the nose for that?
  
  @viktorrozenko 1y
  
  that's not manipulation -- that's just business
  
  @RiedleroD 1y
  
  "just business" often just boils down to manipulation
  
  @hy60koshk 1y
  
  It's barely a manipulation. People are stupid enough to just pick up any shitty ideas on the market. Especially when they see some "free money" or "great budget cuts" opportunities. As for the economic bubble - well, there's a chance it'll never pop and we just enter the dark cyberpunk future... or wokepunk, with wooden PC cases and zero-SLS shit everywhere. Anyway, the idea is that big corporations would trick the majority into thinking that everything's great and goes as it should be, while in fact we would live in shit.
  
  @RiedleroD 1y
  
  I never gave them a cent directly, and yet they profit from me simply using e.g. github
  
  @Le_o_R 1y
  
  This is not correct. They'd sell you the service. Would you rather spend thousands of dollars building your own infra, paying for the electricity, etc, just to ask "why sky blue?" or pay for a $20 monthly subscription to an opensource LLM? It's a scaling thing.
  
  @viktorrozenko 1y
  
  right, but that means that any other company that has enough money to create such infra would compete with them and have an identical product, which would drive the costs down. if I was the CEO, I wouldn't want that to happen, especially if this model wasn't open source to begin with and required billions of dollars in investor money to develop
  
  @echedelle 1y
  
  In fact TogetherAI sells you the service for FLOSS models
  
  @Protomax 1y
  
  ...you would remove 95% of OpenAI dataset 💃💃
  
  @Protomax 1y
  
  Any large player in Ai field is built upon tons of scraped content from internet
  
  @qtsmolcat 1y
  
  Because it's a really fancy text predictor lol
@chenxuuu 1y

its name is deepseek if ask Chinese (obviously just a joke lol)
@Insidekitty 1y

The ad I got for this 💀

Use J and K for navigation

O3-mini’s Chinese outputs spark claims of hasty OpenAI copy-paste release

Description

Categories

Common Tags

Comments

More like this

Usage Reset Joy, Then the Five-Hour Cap Returns

Frontier-Lab Agent Intrusion, Five Days Unseen

Opus 5 Peaks at Medium, Then Overthinks