As of September 17th, two major players in the AI field, Anthropic and OpenAI, are committed to developing “AI colleagues” capable of replacing humans in performing complex work tasks. Their core approach involves using simulated enterprise software to train AI models, enabling these models to understand and operate real-world work processes just like human employees.
To accelerate this process, Anthropic plans to invest 1 billion US dollars next year in building a large-scale AI training “gymnasium”. OpenAI, on the other hand, believes that the entire economy may eventually evolve into a massive “reinforcement learning machine” in the future. AI will continuously evolve through collaboration and feedback with humans, fundamentally reshaping productivity and work patterns.
AI Tutors Earn Up to $250 per Hour, Teaching Large Models Office Work
Anthropic and OpenAI are undertaking an unprecedented initiative: integrating large language models (LLMs) into the “office”and training them to become qualified “digital employees”. These AI models are undergoing intensive professional training to operate a variety of specialized office software—ranging from Salesforce’s customer management system and Zendesk’s customer service platform to Cerner’s electronic medical record system in the healthcare sector. The ultimate goal is to enable AI to independently handle the complex work tasks that white-collar workers face on a daily basis.
This training method differs entirely from traditional AI training. Anthropic and OpenAI have adopted an immersive “simulated office” teaching approach. Researchers not only build highly realistic enterprise application environments for AI but also invest heavily in hiring industry experts to serve as “professional mentors”, who teach the models software operation skills through hands-on guidance.
The cost of such training is substantial. According to insiders, Anthropic plans to invest 1 billion US dollars in the coming year specifically to build a simulated office platform known as a”reinforcement learning environment”or”gymnasium”. OpenAI is also sparing no expense; it is estimated that its spending in data-related fields will reach 1 billion US dollars this year and further increase to 8 billion US dollars by 2030. This funding will be used both for building virtual office environments and paying expert salaries.
Unsurprisingly, the cost of hiring human experts is also on the rise. Labelbox, a company that provides expert services to firms like OpenAI, revealed that among the experts in fields such as biology, software programming, and medicine who teach AI to operate software, 20% now earn more than $90 per hour, and nearly 10% have hourly rates exceeding $120. It is expected that in the next 18 months, the hourly rate of top experts will rise to $150-$250.
Despite the massive investment, the potential returns could be beyond imagination. If successful, this new training method will not only help OpenAI and Anthropic break through the bottlenecks of traditional training technologies but also potentially open up entirely new business models for them. For instance, in the future, enterprises may sell “AI agents” that can take over users’ computers and operate applications on their behalf, or use AI to develop more powerful enterprise-level applications.
Dario Amodei, CEO of Anthropic, refers to these AI products as “virtual collaborators”. They are designed to work side by side with humans and skillfully operate various applications that we use in our daily work.
However, achieving this goal is no easy feat. Turing, a company specializing in helping enterprises optimize AI models, provided a specific case. Anshul Bhagi, the head of Turing’s cutting-edge data projects, explained that teaching AI to complete a sales task requires it to navigate multiple systems: it not only needs to master Salesforce’s customer screening function but also learn to use LinkedIn to find potential customers, Calendly to schedule meetings, and Gmail to send follow-up emails.
To ensure that AI truly masters these skills, Turing has designed a sophisticated verification process: breaking down each task into multiple key steps and establishing clear inspection criteria. Taking sales tasks as an example, the assessment criteria include: Did the AI screen the customer database based on”last contact time”? Did it successfully send an email containing a Calendly meeting link? Did it update the potential customer’s status to”re-engagement”?
Although this technology is still in its early stages, major AI laboratories are already prepared to invest heavily in it. An investor disclosed that Anthropic currently allocates less than 10% of its post-training budget (the optimization phase after the initial training of the model is completed) to the “reinforcement learning environment”. However, due to the significant initial results, the company expects to substantially increase its investment in this area next year.
Leading AI developers like OpenAI are attempting to collect similar cases from various industries, spanning healthcare to law. An OpenAI executive privately stated earlier this year that they expect the”entire economy”to eventually become a huge”reinforcement learning machine”. This means that in the future, a growing number of AIs amid the latest artificial intelligence trends may learn and train by recording the daily work of professionals in various fields on their devices.
