AI Agents Require Domain-Specific Data Generation and Reinforcement Learning
TL;DR
- The shift from simple data labeling to complex, domain-specific data generation for AI agents necessitates expert humans, moving beyond low-skilled contractors to specialized professionals.
- Training AI agents for real-world tasks requires reinforcement learning in simulated environments, mirroring AlphaZero's approach to Go, to generate synthetic data for complex workflows.
- The future of knowledge work automation implies a 100x productivity increase for individuals, enabling them to manage multiple companies and fostering widespread entrepreneurship.
- Data-driven feedback loops, exemplified by Google's search advantage, will become the primary competitive moat as AI models improve through real-world deployment data.
- SaaS models are becoming obsolete as AI foundation models enable easy custom software development and the potential for these models to directly integrate into enterprise systems.
- The primary constraint to AI adoption in enterprises is not technological capability but the "first mile schlepp" of data acquisition, structuring, and infrastructure setup.
- AI's impact will be gradual, not a rapid takeoff, allowing humanity time to adapt workforces and education systems, unlike the all-or-nothing nature of self-driving cars.
Deep Dive
Jonathan Siddharth, CEO of Turing, argues that the era of simple data labeling companies is over, replaced by a need for "research accelerators" capable of training sophisticated AI agents. This paradigm shift is driven by AI models requiring increasingly complex, domain-specific data to perform real-world knowledge work, moving beyond basic chatbots to agentic systems that execute multi-step workflows. The implications are profound: the automation of nearly all knowledge work, a fundamental reshaping of entrepreneurship, and a redefinition of the software engineer's role, all while potentially exacerbating existing societal divides if not managed carefully.
The core of Siddharth's argument rests on the evolution of AI training data and methodologies. Previously, AI was trained on simple tasks and imitation. Now, to create "agents" capable of complex tasks like underwriting or legal analysis, AI requires data generated from simulating real-world business environments using Reinforcement Learning (RL) environments. Turing creates these RL environments across industries, functions, and workflows, essentially building a vast matrix of knowledge work to train AI. This deep dive into specialized data generation is what differentiates Turing from traditional data labeling services. Siddharth emphasizes that this is "innings one" for acquiring such specialized data, suggesting massive future growth potential.
The second-order implications of this AI evolution are far-reaching. Siddharth predicts the automation of 99% of knowledge work within a decade, fundamentally altering the job market. This will empower entrepreneurship by lowering the intelligence and capital barriers, allowing individuals to create companies with AI agents acting as skilled employees. However, this also raises concerns about widening the gap between those who can leverage AI and those who cannot. Siddharth counters this by framing intelligence as an accessible API, arguing that AI will democratize access to expertise, making it cheaper than hiring human specialists. He also posits that human endeavor will shift to higher levels of abstraction, solving more complex problems rather than routine tasks.
Furthermore, the traditional Software-as-a-Service (SaaS) model, as we know it, is deemed obsolete. Siddharth reasons that as AI models become more capable and agentic, they will either be built in-house by enterprises or directly integrated by foundation model companies, rendering many standalone SaaS applications redundant. The primary moat will shift from technology to data-driven feedback loops, where continuous deployment and refinement of custom AI models capture proprietary knowledge and create defensible advantages. This necessitates a hands-on approach to AI deployment, involving significant "first-mile schlepp" (data acquisition and structuring) and "last-mile schlepp" (workflow integration and human training), which Turing facilitates.
The takeaway is that AI is not just an incremental improvement but a transformative force that will automate knowledge work, reshape industries, and redefine human productivity, requiring a strategic shift from simple data labeling to sophisticated AI research and deployment. The future rewards those who can adapt to increasingly complex AI training paradigms and build the agentic systems that will drive this automation.
Action Items
- Audit AI adoption strategy: Identify 3-5 enterprise workflows where current AI capabilities (e.g., GPT-5, Claude) can achieve parity with human experts, as per GDP-Val research.
- Design custom model deployment framework: Outline steps for fine-tuning smaller LLMs on proprietary enterprise data to address specific use cases (e.g., insurance underwriting) and data privacy concerns.
- Develop data feedback loop strategy: Define metrics and processes for collecting and utilizing user interaction data to continuously improve AI model performance and identify gaps in enterprise deployments.
- Evaluate agentic AI integration: Assess 3-5 core business functions for potential automation via agentic AI, focusing on multi-step workflows and tool utilization to reduce reliance on traditional SaaS applications.
Key Quotes
"I think of a talent marketplace as something that's basically matching talent to something maybe it's an opportunity so churing is not a talent marketplace what we do at churing is we're training super intelligence we work with seven out of the eight frontier labs to get to super intelligence you need research compute and data research the labs do in house with openai anthropic deepmind etc for compute we have jensen to thank and maybe nvidia as well on the data side churing powers the data pillar on the data side there's been a significant shift in the last couple of years."
Jonathan Siddharth clarifies that Turing is not a talent marketplace but rather a company focused on training "super intelligence" for frontier AI labs. He highlights that achieving this requires research, compute, and data, with Turing specifically powering the data aspect. This distinction sets up his explanation of the evolving data needs in AI development.
"The third shift is we've gone from chatbots to agents right like we started off with chat gpt where you're asking questions getting answers which is great but now it's about the models becoming agentic where they can execute complex multi step workflows in a real world business setting and the type of data you need for that is totally different."
Siddharth explains a critical evolution in AI from simple chatbots to more capable "agents." He notes that this shift necessitates different types of data for training, moving beyond basic question-and-answer capabilities to models that can perform complex, multi-step tasks in real-world business contexts. This transition signifies a move towards more autonomous and functional AI systems.
"I think the era of data labeling companies is over turing is a research accelerator and it's now the era of research accelerators the labs want to work with a proactive partner that can think about what types of data are likely to be helpful for these models and can make recommendations to them."
Siddharth asserts that the era of traditional data labeling companies is concluding, positioning Turing as a "research accelerator." He argues that leading AI labs now require proactive partners who can not only provide data but also offer strategic recommendations on what data will be most beneficial for advancing AI models. This indicates a move towards a more collaborative and research-driven approach in AI development.
"I do think this will become more popular across the board and it'll be a big market for the frontier labs like if you want like a general purpose assistant i think you need a trillion parameter model right like that can answer anything to be a universal assistant is that a good business for you when you think about taking someone else's model retrofitting it to a business doing a lot of custom work with your own engineering teams in those businesses is that a good business."
Siddharth discusses the growing market for specialized AI models, contrasting them with large, general-purpose models. He suggests that while trillion-parameter models might be suitable for universal assistants, smaller, fine-tuned models retrofitted for specific business needs will become increasingly popular and represent a significant market opportunity for AI labs. This highlights a trend towards tailored AI solutions for enterprise problems.
"I think the transfer is pretty high in areas like customer support copywriting seo like some of these marketing related areas as you would expect the transfer is faster in these low risk to fail areas like when it's relatively easy but i encourage your listeners to look up gdp val which is this paper by openai where they measured the impact of today's ai models in automating all types of economically valuable work it's a lovely piece of research."
Siddharth identifies specific areas where the budget transfer from human labor to AI technology is already significant, such as customer support, copywriting, and SEO. He points to OpenAI's GDP-VALL paper as evidence of AI's capability in automating economically valuable work across various sectors, suggesting that the adoption of AI is accelerating in areas with lower perceived risk.
"I don't see an ai bubble these models are incredibly powerful today gpt 5 is like fucking awesome i don't know what people were talking about when they're talking about you know i know there was some chatter i think we've just gotten used to magic i feel like a these models are incredibly powerful today and they're the worst they'll ever be they're only going to keep keep improving."
Siddharth expresses a strong conviction against the idea of an AI bubble, emphasizing the current power and continuous improvement of AI models like GPT-5. He suggests that people may have become accustomed to AI's capabilities, overlooking their rapid advancement. Siddharth believes that today's models are merely the starting point, with significant future improvements expected.
"I think the market will reward a company with research dna and it'll reward a company that can move fast and adapt very quickly because this is i mean when you're do you think this is a monopoly market or do you think there will be many winners i think there'll be a few winners in the realm of robotics and embodied ai we're still very early at turing we are scaling up on the robotics side as well in terms of data that we generate but there's so much data that's missing that the models need to see that they haven't seen yet."
Siddharth predicts that the AI market will favor companies with strong research capabilities and the agility to adapt quickly to rapid advancements, particularly in areas like robotics and embodied AI. He notes that while Turing is expanding into robotics data generation, the field still has significant data gaps, suggesting opportunities for new players and indicating that the market will likely consolidate around a few key winners.
Resources
External Resources
Books
- "Elon Musk" by Walter Isaacson - Mentioned as a source of learning about hands-on operational approaches and motivations for entering AI.
Articles & Papers
- GDP-V (OpenAI) - Discussed as research measuring the impact of AI models in automating economically valuable work.
People
- Jonathan Siddharth - Founder and CEO of Turing, discussed as a guest and expert on AI, data labeling, and the future of knowledge work.
- Harry Stebbings - Host of The Twenty Minute VC podcast.
- Alex Wang - CEO of Scale AI, admired for his prescience regarding data importance and navigation of the AI landscape.
- Rory Driscoll - Investor at Scale, cited for his perspective on AI value generation being tied to budget transfer from human labor to AI technology.
- Mark Chen - Head of Research at OpenAI, mentioned for his observation that financial services are typically two years behind the state of the art in AI adoption.
- Andre Karpathy - Noted for articulating the concept of cursor for X and partial autonomy in AI workflows.
- Sam Altman - Mentioned in relation to his book and his motivation for entering AI, aiming for a species-ist AI that loves humanity.
- Elon Musk - Mentioned in relation to his book, his hands-on operational style, and his motivation for developing AI.
Organizations & Institutions
- Turing - Company focused on training superintelligence and providing data solutions for AI frontier labs and enterprises.
- OpenAI - AI research laboratory mentioned for its GDP-V paper and its models like GPT-5.
- Anthropic - AI research laboratory mentioned as a frontier lab.
- DeepMind - AI research laboratory mentioned as a frontier lab.
- Microsoft - Acquired Powerset, where Jonathan Siddharth previously worked.
- Nvidia - Mentioned as a key player in compute for AI.
- Disney - Enterprise client of Turing for custom AI models.
- Pepsi - Enterprise client of Turing for custom AI models.
- Blackrock - Enterprise client of Turing for custom AI models.
- Fiserv - Enterprise client of Turing for custom AI models.
- Johnson & Johnson - Enterprise client of Turing for custom AI models.
- Scale AI - Competitor in the data labeling and AI space, mentioned in relation to its acquisition and impact on the market.
- Foxconn - Mentioned as an example of operational structure with strict firewalls between projects.
- German healthcare system - Mentioned as an example of a government entity that would likely require sovereign AI models.
- Wilson Sonsini - Law firm mentioned in the context of financing practices potentially differing from other firms.
- Cooley - Law firm mentioned in the context of financing practices potentially differing from other firms.
- MIT - Mentioned in relation to a report on pilot project failures.
- Deepseek - Chinese AI company mentioned for its state-of-the-art models.
- Kimi 2 - Model developed by Deepseek.
- Qwen - Model developed by Alibaba.
- Google - Mentioned for its historical lead in search due to data-driven feedback loops.
- Yahoo - Competitor in search mentioned in historical context.
- Apple - Mentioned in the context of potential future devices and the evolution of the smartphone.
- Meta - Mentioned in the context of potential future devices and the evolution of the smartphone.
- Amazon - Mentioned in the context of potential future devices and the evolution of the smartphone.
- Google - Mentioned in the context of potential future devices and the evolution of the smartphone.
Tools & Software
- ChatGPT - AI model mentioned for its capabilities and role in the evolution of AI.
- Notion - Productivity tool mentioned as part of the AI productivity tax.
- Gmail - Email service mentioned as part of the AI productivity tax.
- Slack - Communication tool mentioned as part of the AI productivity tax.
- Grammarly - AI writing assistant integrated into Superhuman.
- Coda - Productivity tool integrated into Superhuman.
- Superhuman - AI productivity suite designed to enhance work efficiency.
- Vanta - Security and compliance platform.
- Angellist - Platform for finding and funding startups.
- Powerset - Company acquired by Microsoft, where Jonathan Siddharth pioneered natural language search.
- LinkedIn - Tool used for prospect research.
- Salesforce - CRM tool used for prospect management.
- ZoomInfo - Tool used for prospect contact information.
- Whisperflow - Transcription tool mentioned for its accuracy.
- Wix - Website building platform mentioned as an example of a tool that less tech-savvy businesses might use.
- Squarespace - Website building platform mentioned as an example of a tool that less tech-savvy businesses might use.
- Grok - AI model developed by xAI.
- Gemini Pro - AI model developed by Google.
- Claude - AI model developed by Anthropic.
- GPT-5 - AI model developed by OpenAI.
- Meeting Owl - Speakerphone device for distributed team Zoom meetings.
Websites & Online Resources
- superhuman.com/podcast - URL for learning more about Superhuman.
- vanta.com/20vc - URL for learning more about Vanta and obtaining a discount.
- angellist.com/20vc - URL for learning more about Angellist.
Other Resources
- AI Productivity Tax - Concept describing the loss of context and time due to using multiple AI tools.
- Talent Marketplace - Concept discussed and differentiated from Turing's approach.
- Frontier Models - Advanced AI models developed by leading research labs.
- Reinforcement Learning (RL) - AI training paradigm discussed extensively.
- RL Environments - Simulated worlds used for training AI agents.
- Imitation Learning - AI training paradigm.
- Superintelligence - Advanced AI capability discussed as the goal of Turing's work.
- AGI (Artificial General Intelligence) - AI with human-level cognitive abilities.
- Data Labeling - Process of annotating data for AI training.
- GMV (Gross Merchandise Volume) - Metric discussed in relation to revenue in data labeling.
- SaaS (Software as a Service) - Business model discussed in the context of its potential decline due to AI.
- Computer Use Agents - AI agents capable of interacting with computer systems.
- Multimodality - AI's ability to process and understand various data types (text, audio, video, image).
- Tool Use - AI's capability to utilize external tools and functions.
- Coding - AI's ability to generate and understand code.
- Reasoning - AI's capacity for logical thought and problem-solving.
- Data-Driven Feedback Loops - Mechanism for improving AI models based on user interaction data.
- First Mile Schlepp - The initial effort required to prepare data and infrastructure for AI deployment.
- Last Mile Schlepp - The final steps and integration efforts for AI deployment.
- Partial Autonomy - AI systems designed to collaborate with humans rather than operate fully independently.
- Closed Models - AI models whose architecture and training data are proprietary.
- Open Models - AI models whose architecture and training data are publicly accessible.
- Species-ist AI - AI motivated by a love for humanity.
- Robotics/Embodied AI - AI integrated into physical robots.
- Physical Intelligence - AI's ability to interact with and understand the physical world.
- Digital Intelligence - AI's ability to process and understand digital information.
- Self-Driving Cars - Used as an analogy for AI development, contrasting its "99% accurate" challenge with AGI's incremental value realization.
- AI Bubble - Concept of an overinflated market for AI, which the speaker believes is not occurring.
- Model Capability Overhang - The gap between an AI model's potential capabilities and its current realized performance.
- Agentic Scaffold - The framework of prompts, context, and tools that enable AI agents to perform complex tasks.
- Human-in-the-Loop System - AI systems that involve human oversight and interaction.
- Sovereignty of Models - The concept of nations or regions developing and controlling their own AI models.
- Circular Deals - Financial arrangements between AI providers, discussed in the context of market strains.
- AGI-pilled Group - Individuals who believe in the development of Artificial General Intelligence.
- Sonic Boom - The risk of foundation model companies entering the application layer and disrupting existing software.
- Ambient AI - AI that operates seamlessly in the background, responding to user needs.
- GUI (Graphical User Interface) - User interface designed for human interaction with visual elements.
- API (Application Programming Interface) - A set of rules that allows different software applications to communicate with each other.
- AI Research Automation - The use of AI to accelerate the process of AI research itself.
- Exoskeleton - Metaphor for AI amplifying human productivity and capabilities.
- Drone Suits - Agentic AI systems that operate independently, as seen in later Iron Man movies.