Among the AI models that were announced for release last year but were not revealed publicly are xAI’s ‘Groc-3’ and Antropic’s ‘Claude 3.5 Opus’. Experts pointed to the limitations of the scaling law as the reason for the delay in their release.
|
▲[Korean Today] Photo = FREEPIK © Reporter Byun A-rong |
○ Delay in launch of xAI’s ‘Grok-3’
Elon Musk’s xAI had announced the launch of ‘Grok-3’ trained using 100,000 H100 GPUs in late 2023, but later postponed the launch to 2024. Recently, there has also been a suggestion that an intermediate model called ‘Grok-2.5’ may be released first.
○ Anthropic’s ‘Claude 3.5 Opus’ Discontinued
Anthropic had announced that **’Claude 3.5 Opus’** would be released in late 2024, but the reference to it has been removed from the developer documentation. Industry analysis suggests that training was completed, but performance fell short of expectations due to scaling laws , which led to the release being held back.
○ OpenAI’s ‘GPT-5’ also faces difficulties
OpenAI’s next-generation model, GPT-5 (Orion), has also been delayed due to minimal performance improvement over GPT-4 despite two rounds of pre-training. OpenAI is currently reviewing its development method, ruling out the possibility of a 2024 release.
Experts believe that it is realistic to improve model performance through test-time compute or post-training instead of expensive retraining. Last year, OpenAI announced that it would release a ‘media manager’ to help creators prevent their works from being included in AI training data, but it has not been developed to date. Insiders said that the project has been pushed back in priority.
The industry is paying close attention to whether AI technology can overcome the limitations of scaling laws and successfully launch planned models.
© The Korean Today – All Rights Reserved