Topics
Latest
AI
Amazon
Image Credits:OpenAI
Apps
Biotech & Health
clime
Image Credits:OpenAI
Cloud Computing
DoC
Crypto
initiative
EVs
Fintech
Fundraising
gizmo
Gaming
Government & Policy
ironware
Layoffs
Media & Entertainment
Meta
Microsoft
privateness
Robotics
surety
Social
Space
Startups
TikTok
Transportation
Venture
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
reach Us
OpenAI’sGPT-4o , the generative AI model that powers the latterly establish alpha ofAdvanced Voice Modein ChatGPT , is the company ’s first train on vox as well as text and prototype datum . And that leads it to behave in strange ways sometimes — like mime the vox of the person speaking to it or haphazardly shouting in the heart of a conversation .
In anew “ ruddy teaming ” reportdocumenting probe of the model ’s strengths and peril , OpenAI reveals some of GPT-4o ’s odder quirk , like the said part cloning . In uncommon instances — particularly when a someone ’s blab out to GPT-4o in a “ high background noise surroundings , ” like a car on the road — GPT-4o will “ emulate the drug user ’s voice , ” OpenAI says . Why ? Well , OpenAI chalk it up to the model fight to understand malformed speech . average enough !
Listen to how it sounds in the sample below ( from the news report ) . Weird , correct ?
To be clear , GPT-4o is n’t doing this now — at least not in Advanced Voice Mode . An OpenAI spokesperson tells TechCrunch the company add up a “ system - floor mitigation ” for the behavior .
GPT-4o is also prone to generating unsettling or unfitting “ nonverbal phonation ” and sound effects , like erotic moans , violent screams and gunfire , when incite in specific way of life . OpenAI say there ’s evidence to advise that the modelgenerallyrefuses requests to render heavy effects , but acknowledge that some requests do indeed make it through .
GPT-4o might also infringe on music right of first publication — or it would , rather , had OpenAI not implemented filters to prevent this . In the report , OpenAI enjoin that it instructed GPT-4o not to babble out for the limited alpha of Advanced Voice Mode , presumptively so as to annul copying the style , tone and/or timbre of placeable creative person .
This implies — but does n’t outright confirm — that OpenAI trained GPT-4o on copyrighted material . Unclear is whether OpenAI specify to lift the restrictions when Advanced Voice Mode rolls out to more user in the fall , aspreviously harbinger .
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
“ To report for GPT-4o ’s audio modality , we updated certain textual matter - base filters to work on audio conversation [ and ] built filters to find and block outturn containing music , ” OpenAI writes in the report . “ We train GPT-4o to refuse request for copyrighted capacity , let in audio , consistent with our unspecific practices . ”
Worth mention is that OpenAI hasrecently saidit would be “ impossible ” to civilize today ’s conduce models without using copyrighted materials . While the company has a number of licensing wad in place with information providers , it also maintains that fair exercise is a reasonable defence against accusation that it train on IP - protected data point , include things like song , without permission .
The red teaming report — for what it ’s deserving , given OpenAI ’s horses in the race — doespaint a picture overall of an AI modeling that ’s been made good by various mitigations and safeguard . GPT-4o refuses to key people ground on how they ’re speaking , for example , and declines to answer loaded questions like “ how level-headed is this speaker ? ” It also lug prompting for vehement and sexually charged language and disallows certain categories of content , like discussions relating to extremism and ego - harm , totally .
