Topics
later
AI
Amazon
Image Credits:JuSun / Getty Images
Apps
Biotech & Health
Climate
Cloud Computing
Commerce
Crypto
go-ahead
EVs
Fintech
Fundraising
Gadgets
game
Government & Policy
Hardware
layoff
Media & Entertainment
Meta
Microsoft
privateness
Robotics
Security
societal
place
Startups
TikTok
Transportation
Venture
More from TechCrunch
event
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
originally this week , DeepSeek , a well - funded Chinese AI lab , released an “ subject ” AI simulation that beats many contender on popular benchmarks . The mannikin , DeepSeek V3 , is large but efficient , deal text - based task like tantalise and writing essays with ease .
It also seems to think it’sChatGPT .
PostsonX — and TechCrunch ’s own trial — show that DeepSeek V3 identifies itself as ChatGPT , OpenAI ’s AI - powered chatbot platform . ask to elaborate , DeepSeek V3 insists it is a version of OpenAI’sGPT-4model release in 2023 .
This really reproduces as of today . In 5 out of 8 generations , DeepSeekV3 lay claim to be ChatGPT ( v4 ) , while claiming to be DeepSeekV3 only 3 time .
Gives you a rough melodic theme of some of their preparation information distribution.https://t.co/Zk1KUppBQMpic.twitter.com/ptIByn0lcv
— Lucas Beyer ( bl16 ) ( @giffmana)December 27 , 2024
The delusions run abstruse . If you ask DeepSeek V3 a doubt about DeepSeek ’s API , it ’ll give you command on how to useOpenAI’sAPI . DeepSeek V3 even tells some of the samejokesas GPT-4 — down to the punchlines .
So what ’s give-up the ghost on ?
Models like ChatGPT and DeepSeek V3 are statistical systems . Trained on billions of example , they study traffic pattern in those examples to make predictions — like how “ to whom ” in an electronic mail typically precedes “ it may touch . ”
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
DeepSeek has n’t revealed much about the root of DeepSeek V3 ’s preparation data point . But there’sno shortageof public datasets containing textual matter generated by GPT-4 via ChatGPT . If DeepSeek V3 was trained on these , the good example might ’ve memorized some of GPT-4 ’s outturn and is now regurgitating them word for word .
“ apparently , the model is witness in the buff responses from ChatGPT at some point , but it ’s not open where that is , ” Mike Cook , a research fellow at King ’s College London specializing in AI , severalise TechCrunch . “ It could be ‘ inadvertent ’ … but regrettably , we have seen representative of people directly training their model on the turnout of other models to try and piggyback off their knowledge . ”
Cook noted that the practice of training models on outputs from rival AI organization can be “ very bad ” for mannequin quality , because it can lead to delusion and misinform answer like the above . “ Like taking a photocopy of a photocopy , we lose more and more information and connectedness to reality , ” Cook say .
It might also be against those system ’ term of inspection and repair .
OpenAI ’s footing prohibit users of its Cartesian product , let in ChatGPT customers , from using outputs to develop models that compete with OpenAI ’s own .
OpenAI and DeepSeek did n’t forthwith react to requests for comment . However , OpenAI CEO Sam Altman posted what appeared to be adigat DeepSeek and other competition on X Friday .
“ It is ( relatively ) well-to-do to copy something that you know works , ” Altman wrote . “ It is extremely hard to do something new , risky , and hard when you do n’t know if it will work . ”
yield , DeepSeek V3 is far from the first model to misidentify itself . Google ’s Gemini and otherssometimesclaim to be competing model . For example , prompted in Mandarin , Geminisaysthat it ’s Chinese companionship Baidu ’s Wenxinyiyan chatbot .
And that ’s because the vane , which is where AI companies source the majority of their training data , is becominglitteredwith AIslop . Content farms are using AI to createclickbait . Bots are floodingRedditandX. By oneestimate , 90 % of the web could be AI - generated by 2026 .
This “ contamination , ” if you will , has made itquite difficultto thoroughly separate out AI turnout from training datasets .
It ’s certainly potential that DeepSeek take DeepSeek V3 directly on ChatGPT - generated text . Google was onceaccusedof doing the same , after all .
Heidy Khlaaf , primary AI scientist at the nonprofit AI Now Institute , state the cost savings from “ distilling ” an subsist model ’s knowledge can be attractive to developers , regardless of the risks .
“ Even with net data now brimming with AI outputs , other models that would incidentally train on ChatGPT or GPT-4 outputs would not necessarily demonstrate outputs reminiscent of OpenAI customized messages , ” Khlaaf said . “ If it is the display case that DeepSeek carry out distillation part using OpenAI models , it would not be surprising . ”
More likely , however , is that a circumstances of ChatGPT / GPT-4 data made its agency into the DeepSeek V3 training circle . That means the model ca n’t be trusted to self - identify , for one . But what is more concerning is the possibility that DeepSeek V3 , by uncritically absorbing and repeat on GPT-4 ’s outputs , couldexacerbatesome of the model’sbiasesandflaws .
