|
Listen to this story
|
OpenAI has been criticised for its opaque decision-making process. Unperturbed, CEO Sam Altman has graced the world with some valuable insights. In an interview with Humanloop, the oft-criticised celebrity CEO spoke about the future of OpenAI. While this interview shed some light on some of OpenAI’s most profitable products, it also reflected a culture of overpromising and under-delivering.
While the interview itself presents a few surface-level points, it also delves into some of the scaling issues that OpenAI has been facing. Along with this, some of Altman’s statements in this interview directly conflict with his statements in the past, creating more opacity for ‘OpenAI’. However, one thing is for certain, GPT-5 isn’t coming anytime soon, not even in 2024.
Here are 6 reasons why we won’t get GPT-5 any time in the near future.
Not enough GPUs
In the interview, Altman delved into the scaling issues OpenAI is facing. According to him, OpenAI is ‘extremely GPU-limited’, which has rippled into delays for their short-term plans. This means that plans like longer context lengths (more than 32,000 tokens) and fine-tuning for specific dataset will be delayed.
Due to the shortage, even the basic functionality of the API has been compromised. According to the interview, OpenAI is offering a dedicated capacity offering, which will offer faster and more dependable API access. However, to access this, customers have to commit to a $100,000 spent on the platform.
A bigger focus on GPT-4
While GPT-4 was one of the biggest upgrades for OpenAI, it is clear that Altman isn’t done with this model. The first and biggest item on the roadmap was a cheaper and faster GPT-4. By improving its flagship model, OpenAI wishes to reduce the ‘cost of intelligence’ as far as possible.
This means that the company will likely spend time and resources to optimise GPT-4 even further instead of focusing on GPT-5. In addition to this optimisation, OpenAI is looking to increase the context window to up to 1 million tokens. Other upgrades on the roadmap include fine-tuning APIs and a stateful API for GPT-4.
Restricted by scaling laws
Altman has been vocal in the past about how the size of LLMs won’t matter as much moving forward. However, it seems that OpenAI’s researchers have changed his mind. According to the interview, its internal data has found that scaling laws continue to hold well.
This means that GPT-5, when it releases, will not be as big of a leap as GPT-2 to GPT-3.5. This is because models will likely ‘double or triple’ in size, as opposed to changes in the order of magnitude.
Promised features first
Multimodality was one of the biggest new features of GPT-4. Demos of the model being capable of generating a whole website from just a sketch set the Internet on fire, but it seems even ChatGPT Plus users have to wait for multimodality.
In the interview, Altman stated that multimodality will only come in 2024, blaming the lack of GPUs. Until multimodality comes to GPT-4, GPT-5 is still a far-off dream.
ChatGPT is the last product
Altman also stated that ‘OpenAI would not release products beyond ChatGPT’. This is mainly because OpenAI wants to be safer for competing with their customers with other offerings. He also stated that there was a history of platform companies having a killer app.
This solidifies OpenAI’s future goals as a platform company rather than a product company. As time goes on, their platforms will only get more capable. This can also be interpreted to mean that GPT-5 might only get more delayed in favour of better platforms.
Regulatory risks
Sam Altman has also been very vocal when it comes to AI regulation. He is also one of the biggest signatories for the new AI safety letter which advocates for more stringent regulation of AI, citing existential risks.
While Altman said that he continues to believe in the importance of open source, he also defended OpenAI not releasing their models to the public. This is mainly due to his skepticality towards the large number of companies that have the capabilities to host and serve LLMs.
With the host of problems OpenAI is facing to scale up to the success of ChatGPT, it seems that other projects might be put on hold. While even the GPT-4 API has not been updated for the past 2 months, GPT-5 is still a long-distant dream until OpenAI gets it together.
