20VC: Scale, Surge, Turing, Mercor: Who Wins & Who Loses in Data Labelling | Is Revenue in Data Labelling Real or GMV? | Why 99% of Knowledge Work Will Go and What Happens Then? | Why SaaS is Dead in a World of AI with Jonathan Siddharth @ Turing
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this candid discussion, Jonathan Siddharth, Founder and CEO of Turing, reveals how AI data requirements have fundamentally shifted from simple labeling to complex, real-world intelligence training. (03:35) Siddharth explains that Turing is not a talent marketplace but rather a "research accelerator" that works with seven out of eight frontier AI labs to train superintelligence through sophisticated reinforcement learning environments. (14:05) He predicts that all knowledge work involving computer screens will be automated within 10-20 years, representing $30 trillion in economic value. (17:59) The conversation explores enterprise AI adoption challenges, revenue concentration in the data labeling market, and why traditional SaaS models may become obsolete in an AI-first world.
- Main themes: The evolution from simple data labeling to complex AI training, enterprise AI deployment challenges, the future of knowledge work automation, and the transformation of business models in an AI-driven economy.
Speakers
Jonathan Siddharth
Jonathan Siddharth is the Founder and CEO of Turing, one of the fastest-growing AI companies advancing frontier models, which has reached $350M ARR with just $225M raised while maintaining profitability. A Stanford-trained AI scientist, Jonathan previously helped pioneer natural language search at Powerset, which was acquired by Microsoft, giving him deep expertise in AI development and commercialization.
Key Takeaways
Data Complexity Has Fundamentally Shifted from Simple to Expert-Level
The era of simple data labeling is over, replaced by the need for sophisticated, domain-specific expertise. (03:35) Siddharth explains that while previous AI training involved basic tasks like "write a Python program to sort numbers," today's requirements demand complex, real-world applications like building full B2B marketplace apps across multiple platforms. This shift means that low and medium-skilled contractors can no longer generate the quality of data needed - only expert humans in every domain can provide the sophisticated training data required for modern AI systems.
Reinforcement Learning Environments Are the New Competitive Advantage
Modern AI training requires creating "mini world models" for business environments rather than simple input-output pairs. (08:00) Turing creates RL environments across a four-dimensional matrix: every industry, function, role, and workflow. These environments allow AI agents to try different approaches and learn from feedback, similar to how AlphaZero mastered Go by playing against itself. This represents a fundamental shift from teaching AI to pass tests to teaching it to do real work in complex business settings.
Enterprise AI Success Requires Solving "First Mile" and "Last Mile" Problems
The 95% failure rate of AI pilots stems from inadequate preparation and integration work. (41:59) "First mile schlep" involves structuring messy enterprise data, while "last mile schlep" requires building workflows designed for partial autonomy rather than full automation. Companies need cursor-like interfaces for every role and workflow, proper evaluation systems, and training programs for humans to work alongside AI agents effectively.
Revenue Concentration Is Acceptable in Infrastructure-Scale Markets
Despite having only eight major customers (frontier AI labs), revenue concentration is justified by the massive scale of investment and spending. (37:58) Siddharth compares this to NVIDIA, where 39% of revenue comes from two clients, noting that when individual projects like Stargate involve $100 billion annual investments, concentration among a small number of extremely well-funded customers becomes a strategic advantage rather than a risk.
Traditional SaaS Models Are Becoming Obsolete in an AI-First World
The fundamental premise of SaaS - providing software that humans navigate with GUIs - is being undermined by agentic AI systems. (48:03) As AI becomes capable of multimodal reasoning, tool use, and coding, the need for human-operated interfaces diminishes. Companies will either build custom AI applications easily or use foundation models directly through natural language interfaces, eliminating the need for traditional SaaS applications designed for keyboard and mouse interaction.
Statistics & Facts
- Turing has reached $350 million in annual recurring revenue while raising only $225 million and maintaining profitability. (01:00) This represents an exceptional capital efficiency ratio in the AI infrastructure space.
- All knowledge work automation represents approximately $30 trillion in economic value globally. (10:12) This massive market size justifies the current level of investment in AI development across frontier labs.
- In OpenAI's GDP-VAL study, today's AI models achieved human-expert parity approximately 50% of the time across nine verticals and 44 occupations for single-step tasks. (23:31) This demonstrates significant progress while highlighting room for improvement in multi-step workflows.
