At the OpenAI Spring Update, OpenAI CTO Mira Murati unveiled GPT-4o, a new flagship model that enriches its suite with ‘omni’ capabilities across text, vision, and audio. Murati promised iterative rollouts to enhance both developer and consumer products in the coming weeks.
During the demonstration of GPT-4o’s real-time translation capabilities, the model seamlessly translated from English to Italian, showcasing its sophisticated linguistic adaptability. Many believe that this new feature of OpenAI is likely to replace Google Translate.
Now that OpenAI is also releasing search capabilities within ChatGPT, it might be time for startups like Perplexity to start counting their days.
While this is about the competition between Big Tech, everyone who has been building models around OpenAI’s GPT models or building on top of open-source models such as Llama or Mistral, is at the threat of being replaced by another update to OpenAI’s update.
Don’t be Just a Wrapper
A few days ago, Cred founder Kunal Shah posted on X asking every Indian startup to post their AI products. The post garnered nearly 400 responses. Dharmesh BA, who is working on a stealth startup in India, took some time to figure out that many of the products were essentially just wrappers of existing models on different modalities.
He called and categorised these apps as CRUD (Create, Read, Update, Delete). “Most of these apps are built on top of various mediums such as text, image, video, or voice. Building an app with the premise that OpenAI or current LLMs don’t perform a specific task, and therefore you will do it, can lead to disaster,” he said.
To explain, he said that it’s similar to judging a growing child – a 3-year-old who will gain much more knowledge by age 5. You cannot assess a 3-year-old’s intelligence as static.
AIM has always been at the forefront of talks on this. When OpenAI introduced ChatGPT Enterprise, it sent shockwaves across several SaaS startups that had developed products around ChatGPT or offered wrappers based on ChatGPT APIs catering to business clients.
Now, the same is happening with several startups that were offering services utilising different models on top of GPT APIs to offer text, image, video, or voice, such as translation apps, language learning apps, coding assistants, or for use in customer service.
Each time OpenAI updates or releases new versions, many startups find themselves becoming irrelevant because of how the new capabilities of OpenAI can now handle the tasks they were attempting to address.
Moreover, attempting to limit the potential of powerful technologies like LLMs within mobile apps or websites is akin to hiring someone like Raghuram Rajan for basic arithmetic – while capable, his true potential lies in solving more significant problems, as Dharmesh explained in his post.
Therefore, it’s important to think carefully before investing in LLMs through traditional apps and websites and building wrappers on top of them. While this isn’t meant to discourage innovation, it’s crucial to remain adaptable and open to change, recognising that the medium you use is evolving.
Diminishing Returns
This is similar to what Gary Marcus points out repeatedly in his blogs. Most of the advancements in Big Tech are now limited to comparing small changes and updates to their models rather than offering anything revolutionary. All of it is just to get an edge over its competitors.
When it comes to India, researchers and enterprises need to focus on building large models, prioritising technical benchmarking and an industrial standardisation of AI rather than obsessing over specific use case apps. These apps, built on LLMs, are features that can be easily replicated and improved upon by others. Many fail to grasp this point.
Most believe that in the future, there could be several options, and the LLM would be like an operating system where users can decide and choose their own apps. While that is true, how many of those apps would last if the base provider, such as OpenAI, changes its architecture?
The ultimate question that remains is why such research is not being done in the country. OpenAI and Google, both tech giants, have redirected their focus towards Indic languages as well, which will eventually become a threat to those building for the Indian ecosystem.
Recently, Shah also posted on X saying that early-stage startups should be easy to iterate and late-stage startups should be hard to distract. This also highlights the mentality of Indian startups that are not iterating and not innovating as well.
“Startups should be easy to unlearn,” replied Vivian.
