The Behind-the-Scenes of the Meta 14B Investment and Future Strategies for AI-Human Co-creation
In recent years, there have been mixed voices of expectation and concern regarding the practical application and corporate adoption of AI. Against the backdrop of the question, "Will AI really deliver the results it promises?", Scale AI is one of the companies at the forefront of the "data and evaluation (eval)" field.
In this interview, Mr. Droog, who became CEO after the large-scale capital alliance with Meta, discusses Scale's current position and how AI models will evolve through collaboration with human experts.
The following is organized by theme.
1. Meta's Investment and Scale's Independence
1-1. Overview and Background of the Investment
In June 2025, Meta invested approximately $1.4 billion (reported as $14.3–14.8 billion) in Scale AI, acquiring a 49% stake in the company.
What is important about this transaction is that Scale was not "fully acquired as a subsidiary," but rather designed to maintain its governance and operations as an independent company.
While acknowledging that Meta has obtained a new board seat, Mr. Droog clearly states, "Meta has no special access or preferential relationship, and we will continue to operate under the same rules as before."
1-2. Organizational Changes and Impact
With this investment, Scale is beginning to focus on the business areas of providing applications and solutions in addition to its data business for models.
However, organizational restructuring is unavoidable, and there are reports that structural reforms have been implemented, particularly in the Generative AI (GenAI) division, including employee layoffs and team reorganization.
Nevertheless, Mr. Droog says, "The data business we have been doing for a long time is very strong, and the application business is also growing," indicating that Scale's direction is expansive.
2. Evolution of Data Labeling and "Expert Labeling"
2-1. Shift from Labeling by General Workers
When Scale was first founded, the initial data labeling for model training was dominated by low-cost, high-volume methods utilizing general crowd workers and non-experts.
However, in recent years, with the improvement in AI model performance, tasks requiring more advanced specialized knowledge have become necessary. In other words, we are transitioning to a stage where experts such as doctors, lawyers, and engineers provide feedback and evaluation (eval) to the models.
Mr. Droog states, "18 months ago, it was a simple task like answering 'which short story is better,' but now the tasks include website construction and advanced explanations about cancer."
According to him, about 80% of Scale's expert network holds a bachelor's degree or higher, and over 15% hold a doctorate, allowing them to handle highly specialized tasks.
2-2. The Role of "Eval (Evaluation)"
Evalis the work of providing a comparative standard to judge whether a model's output is "good/correct." In other words, to train and improve a model, it is essential not only to provide large amounts of data but also to explicitly define "what is a good answer."
This evaluation work is particularly important in fields where mistakes are not tolerated, such as medicine, law, and corporate processes, and it is necessary for humans to define "what is good" and have the model learn it.
3. The Difficulty of Acquiring and Retaining Experts
3-1. Expert Search and Referral Strategy
Continuously securing highly skilled experts is a challenge. Mr. Droog cites "referrals" and "collaboration with universities and research institutions" as the methods Scale prioritizes most in recruitment.
There are many cases where experts themselves are interested in being able to contribute their knowledge to AI, and the key is to create an environment that provides both "fulfillment" and compensation.
3-2. Expert Retention and Differentiation
To prevent experts from being poached, multiple factors are involved, such as "work satisfaction, compensation, evaluation systems, and network value." Mr. Droog points out that "providing an excellent experience" leads to long-term retention.
4. AI Environment (Agent Environment) and the Importance of Reinforcement Learning
4-1. Environment-Based Learning
AI agents (e.g., CRM operations, automated processing workflows, etc.) must repeatedly try and error within an environment (sandbox) to seek optimal solutions. Mr. Droog cites specific examples such as Salesforce operations and medical system operations, stating, "Agents should be trained to understand the environment and make autonomous decisions."
In such environmental training, designs that include decisions to escalate to humans in the event of an error are also required.
4-2. Generalization and Data Value
It is impossible to collect an infinite number of all tasks, so the key is 'how far can it be generalized?' Mr. Droug states that 'tasks and data with versatility have higher value.'
For example, the task of 'finding a meeting in a calendar' is considered applicable in almost any environment to some extent, and it is important to possess knowledge that 'works in diverse environments.'
5. Lessons Learned from Entrepreneurial Experience at Uber Eats
5-1. Customer Understanding and Hypothesis Testing
Mr. Droug says that when he launched Uber Eats, he actually weighed ingredients to estimate cost structures in order to understand the restaurant economy. They estimated how much restaurants spend on materials, labor, and real estate on their own, and then reverse-calculated the commission rate to design the model.
In this way, the key to success lies in the attitude of digging deeper into 'what is not said' and 'what the underlying motivation is,' rather than simply believing what you are told.
5-2. Focus on Gross Margin
When starting a new business, he uses gross margin as a very important filter. While a 40% gross margin is often the general target, he recommends first setting a hypothesis of 60% and then discussing why that might not be achievable.
According to him, he also warns that 'a strategy of starting with a low gross margin and compensating with volume carries risks,' making the possession of differentiation factors or barriers to entry an important decision-making axis.
5-3. Strategy Based on 'Not Losing'
Mr. Droug states that 'Not losing is a prerequisite for success.' In the context of entrepreneurship and investment, taking excessive risks can lead to fatal injuries. The idea is that it is first important to ensure long-term viability (business continuity).
6. The Future of AI and Human Coexistence
6-1. How Long Will the Human Role Remain?
There is much discussion about whether 'AI will eventually be able to do everything.' However, looking back at the history of model evolution, Mr. Droug believes that 'human knowledge, judgment, and expertise will always be necessary as long as new things are being created.'
In response to the question, 'Will there ever be a moment when external human data is no longer needed in a certain field?', he offers a cautious view that 'in a world where new knowledge and judgments continue to be born, a certain level of human intervention will remain.'
6-2. Models Evolve from 'Knowing' to 'Doing'
Currently, many AI models are in the stage of 'remembering and judging knowledge,' but in practical terms, the ability to 'perform jobs that can be concretely substituted' is beginning to be required. Mr. Droug says that the era of moving from 'what the model knows' to 'the model acting (autonomously judging and executing)' is just around the corner.
In this transition, environmental training, reinforcement learning, and eval design become extremely important elements.
7. Summary: Learning and Outlook
The main points obtained through this interview can be summarized as follows:
Meta's investment in Scale AI was not merely a capital partnership, but a strategic design aimed at balancing independence with business expansion.
The evolution of the data labeling/evaluation (eval) domain is shifting from non-expert labor to expert-led models, a phase essential for the advancement of AI models.
Securing and retaining experts is supported by referral networks, building relationships with educational institutions, and designing meaningful work.
Designing environmental learning and generalization is the key for AI agents to perform real-world tasks.
Lessons from entrepreneurial experience (customer understanding, focus on gross margin, risk management) are universally applicable to AI businesses and all new ventures.
The division of roles between humans and AI is likely to evolve through phases of supplementation and collaboration, rather than AI immediately replacing humans.
As AI adoption spreads in the future, this 'practical framework' of data, evaluation, environmental design, and human relationship design will become just as important as the capabilities of the models themselves. The fact that companies like Scale AI are supporting such 'invisible infrastructure' is where the essential challenges and potential of technological development lie.