Closed source artificial intelligence is hindering our exploration of true science
The person who said this was none other than the renowned inventor of the Transformer, Ashish Vaswani.
The big shot said this because he felt that closed source vendors such as OpenAI were blinded by commercialization and had no intention of doing basic research.
After the Scaling Law hit a wall, unicorns became increasingly confused about the direction of technology.
Under pressure from investors, I had to personally explore the landing scenarios and had no energy to engage in scientific research.
Vaswani, who moved three times from Google to Adept and then to Essential AI to fill this gap, has once again reached a huge turning point this year——
All in basic research, and adhere to the open source path.
I never expected that the three major players would relocate just to build DeepSeek for the Western world.
##We are happy to be the DeepSeek of the Western world
In 2017, Vaswani, who was still working at Google Brain, co authored “Attention Is All You Need” with his team, which marked the beginning of the GenAI era five years later.
Unfortunately, the old owner was not very awesome, and various rules and regulations tied people down, which discouraged Vaswani.
In 2022, he ran away from home and founded Adept AI with Niki Parmar (one of Transformer’s eight sons) and David Luan (then Adept CEO).
Unfortunately, the good times didn’t last long. According to BloomBerg, Vaswani had a serious conflict with a co-founder at Adept.
This dispute made him realize:
Only by establishing a company that belongs entirely to oneself can one have absolute freedom. **
So at the end of 2022, he waved goodbye again and formed a new team to establish * * Essential AI * *.
At the beginning, Essential AI followed a pragmatic approach - its main business was to automate financial analysis for enterprises.
As a result, at the beginning of this year, Vaswani suddenly announced to the board of directors:
Abandon the existing business of the company and devote all energy to basic research.
What！？
Isn’t this simply overturning the business model and starting over?
You should know that for a startup company that originally served corporate clients, this is a highly risky bet.
Surprisingly, both the board of directors and the management have expressed some support for Vaswani.
AMD, which participated in the last round of financing for Essential AI, stated in an interview:
This route is indeed a bit unusual, but the field of artificial intelligence requires more open alternative solutions.
It seems that Vaswani’s efforts to relocate three times were not in vain. She finally achieved her goal and gathered a group of like-minded experts.
After so many years of running around for commercialization and dealing with capital and customers countless times, Vaswani is really tired.
Suddenly looking back, in his spiritual world, there has always been a land that can carry all his beautiful expectations, and the name of this Eden is called * * Open Source * *.
He wrote on the Essential official website:
In fields such as education and healthcare, there is a real scientific need to pursue, and we cannot let closed artificial intelligence hinder human exploration of these new directions.
Doesn’t it sound familiar? Yes, Vaswani admitted directly in an interview with Economics Time:
In the short term, we want to become the DeepSeek of the Western world. ”
Vaswani believes that pushing the boundaries of AI cannot rely on unicorns such as OpenAI and Anthropic that profit from closed source models.
In his vision, AI should not only serve business, but should truly enter into scenarios that concern the public, such as education and healthcare, so that ordinary people can also enjoy the benefits of technology.
For example, a child from a remote area can also use open-source AI to attend the best classes; Even a small clinic can use open-source tools to make the most accurate diagnosis.
**More importantly, Vaswani is not just shouting slogans, Essential has indeed put in a lot of real effort after the transformation.
Shortly after proposing the transformation, in April 2025, the Essential AI team published a paper titled ‘Reflecting in Pre Training’.
This paper presents a disruptive viewpoint:
The reflective ability of large language models actually begins to sprout in the pre training stage.
Simply put, compared to patching with RL in post training, they have made a breakthrough in pre training, and this technique may significantly reduce training costs.
If it really succeeds, it will definitely be a great benefit for the entire open source community.
##What did Vaswani see?
Why does Vaswani say we can’t pin our hopes on unicorn companies?
Vaswani is concerned that these unicorns with the best model performance are hindering the progress of artificial intelligence.
Several companies have closed their long-term research and development work in order to allocate all resources to commercialization, a phenomenon that becomes more pronounced when the market environment deteriorates.
Since the turning point of diminishing marginal returns in Scaling Law, the productization tendency of AI unicorns has become well-known.
Anthropic is developing a browser, and OpenAI has dispatched renowned product manager Kevin Weil… Major model manufacturers have stepped down to personally explore business models.
This reflects a problem of the innovator’s dilemma.
After the rise of Scaling Law, the weight of violent aesthetics shifted the balance of AI research almost entirely from academic institutions to corporate laboratories.
However, although the industry monopolizes the rarest production factors, can they really innovate all in?
Most of the time, it is not possible. Because it could burn several hundred million and there may not necessarily be a return on investment.
The key is that regardless of the original intention of the company’s establishment, the founder must ultimately prioritize being responsible to the investors.
Therefore, exploring business models and maximizing profits is the ultimate destination of the closed source model, which was predetermined from the beginning.
Now, the passionate commercialization flame has also been ignited by these proud individuals in Silicon Valley.
Today I’m waiting for Xiao Zha to buy, tomorrow I heard that OpenAI is going to make an acquisition, and next month I will start a business again.
This rhythm is very suitable for chasing hot topics, but scientific research is different - to achieve breakthrough results, you have to spend several years or even more than a decade stuck in the same direction.
And frequent turnover makes it difficult for the team to calm down.
Vaswani sighed, ‘This is not the way science should progress.’.
**What does Vaswani want to do? Can the open-source camp save AI? **
The age-old topic - the power of many people is great, and knowledge sharing is the key driving force behind the development of AI technology.
Although the closed source company holds top resources and technological achievements, it has to hide these secrets in order to safeguard the interests of investors and hinder competitors.
But if thousands of researchers and developers contribute code together, and the entire ecosystem works together and advances in parallel, perhaps the problem of resource scarcity in the open source camp can be overcome.
Of course, the most important issue with open source is the source of funding, as we cannot always rely on ‘generating electricity with love’.
In this regard, Vaswani referred to the solution of Internet products:
##Cross subsidy law is good
What is cross subsidy?
This is a common business model in the Internet era. A typical case is the browser: Google search is completely free for users, and relies on advertising to return blood after gaining market share.
Simply put, it means making money from a portion of the business to subsidize another portion of the business.
Vaswani’s plan for Essential AI is as follows:
First, build an open-source vertical domain model, which itself does not charge any fees.
But if a customer wants to use this base to build their own AI, they can purchase training data and related products from Essential AI.
The money earned through this channel can be used to give back to the open source community.
This ensures both technological openness and the survival of the company.
In addition, Vaswani also pointed out that closed source does not necessarily represent a higher return on investment.
Although closed source models may seem easier to monetize, they actually face significant cost pressures.
In fact, if we look at the development of the technology Internet before, open source is usually the more profitable side, after all, it has built an entire ecosystem.
One More Thing
Finally, there may be a little more to say about Ashish Vaswani.
As the “first work” of the paper “Attention Is All You Need”, his doctoral supervisors were both Chinese.
At the University of Southern California in 2011, there were two professors active at the forefront of natural language processing (NLP) - David Chiang and Liang Huang.
They are both mentors of Ashish Vaswani.
Professor Jiang Wei on the left and Professor Huang Liang on the right
**Professor Jiang Wei * * received his bachelor’s and master’s degrees from Harvard University and later pursued his PhD at the University of Pennsylvania. His proposed hierarchical phrase translation model was adopted by Google Translate.
**Professor Huang Liang graduated from the Department of Computer Science at Shanghai Jiao Tong University with a bachelor’s degree, and also pursued a PhD at the University of Pennsylvania. His main research direction is developing efficient algorithms to accelerate NLP tasks, and Vaswani was his first doctoral student under his guidance.
How to put it, although there are no Chinese authors in the Transformer Eight, there are not entirely no Chinese contributions in the “Merit Book”.
This is actually another dimension of open source power. Whether it’s AI or deep learning, the reason why they can stand at the top of the wave is fundamentally because there is always an ecological environment of continuous communication, openness, and mutual assistance.
Open source is not only a choice, but also a spirit and belief.
Reference link
[1] https://www.bloomberg.com/news/features/2025-09-03/the-ai-pioneer-trying-to-save-artificial-intelligence-from-big-tech
[2] https://www.communicationstoday.co.in/ashish-vaswani-the-mind-behind-the-transformer-that-powers-genai
[3] https://economictimes.indiatimes.com/tech/artificial-intelligence/we-would-like-to-be-deepseek-in-the-west-says-essential-ai-cofounder/articleshow/121891250.cms
[4] https://www.forbesindia.com/article/ai-special-2025/ashish-vaswanis-essential-ai-wants-to-use-powerful-ai-to-solve-humanitys-biggest-challenges/96148/1
[5] https://www.wired.com/story/eight-google-employees-invented-modern-ai-transformers-paper
[6] https://www.essential.ai/
[7] https://viterbischool.usc.edu/news/2023/03/attention-is-all-you-need-usc-alumni-paved-path-for-chatgpt/