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

Jonathan explains how Turing has transformed from a talent marketplace into a research accelerator for frontier AI labs. He details the three major shifts in AI data requirements: simple to complex data, teaching AI to pass tests versus do real work, and the transition from chatbots to agentic systems requiring reinforcement learning environments.

• •Turing works with 7 of 8 frontier labs to power the data pillar for superintelligence
• •Data needs have shifted from simple tasks (sorting numbers) to complex workflows (building full B2B marketplace apps)
• •Training agents requires RL environments - mini world models with synthetic data where agents try different approaches
• •Creating RL environments for every workflow, role, function, and industry (4-dimensional matrix covering $30T of knowledge work)
• •The era of data labeling companies is over; it's now about being proactive research partners

Discussion of Turing's enterprise business building custom fine-tuned models for companies like Disney, Pepsi, and BlackRock. Jonathan explains why smaller, specialized models often outperform large world models for specific tasks, and how enterprises need on-premise solutions with proprietary data that doesn't benefit competitors.

• •Enterprises need smaller fine-tuned models (500M-10B parameters) for specific workflows like insurance underwriting
• •Custom models distill institutional human knowledge and integrate with internal tools
• •On-premise deployment prevents proprietary data from going back to frontier models
• •This business is fast-growing and provides crucial real-world feedback on where models break
• •Example: Insurance underwriting using medical data analysis with human-in-the-loop systems

Jonathan's bold prediction that all digital knowledge work will be automated, addressing pushback about enterprise adoption challenges. He distinguishes between slow back-office automation and faster front-office adoption where AI directly generates revenue, particularly in financial services.

• •Any job involving looking at a computer screen and using keyboard/mouse will be automated
• •Front-office automation (investment decisions, drug discovery) will happen faster than back-office
• •Incumbents who can't adapt will lose to AI-native startups over 10-20 years
• •OpenAI's GDP-Val research shows models achieving 50% parity with human experts on real tasks
• •Transfer of budget from human labor to AI technology is the key indicator of real value creation

Vision for how society transforms when AGI automates knowledge work. Jonathan argues this will democratize entrepreneurship, allowing non-technical founders to build companies with AI teams for minimal capital, while humans solve problems at higher levels of abstraction.

• •Every human could be 100x more productive, running multiple companies simultaneously
• •Non-technical founders (e.g., therapists) can start companies with marketing GPT, engineering GPT, PM GPT
• •Intelligence as an API ($20/month) is more accessible than hiring expensive human experts
• •Humans won't be bored - we'll solve more exciting problems like curing diseases and interstellar travel
• •Optimistic view: AI narrows the gap between haves and have-nots by democratizing access to expertise

Jonathan's controversial thesis that traditional SaaS is over, facing three existential risks: companies building custom apps themselves, foundation models moving into the apps layer, and the obsolescence of GUI-based software designed for humans. Harry pushes back on this view.

• •Three risks to SaaS: (1) companies build custom apps easily on LLMs, (2) foundation models become agentic and sonic boom the apps layer, (3) GUI designed for humans becomes obsolete
• •Most SaaS apps were built when software was hard to create; now LLMs make it trivial
• •Future interfaces: ambient AI you talk to, not clicking around GUIs with keyboard and mouse
• •Harry's counterargument: enterprises have 80-100 SaaS products, can't maintain custom versions, and many businesses barely use basic tools
• •Data-driven feedback loops become the new moat when technology isn't defensible

Deep dive into revenue recognition in the data labeling market, addressing concerns about GMV versus true revenue. Discussion of working with frontier labs, revenue concentration risks similar to NVIDIA, and competitive dynamics with companies like Scale AI.

• •Turing's revenue is GAAP revenue, not GMV - recurring projects with frontier labs
• •Revenue concentration similar to NVIDIA (39% from two clients) is acceptable given massive market size
• •Labs work with small handful of trusted partners for resilience and price competition
• •Scale AI acquisition by Meta created flood of demand for Turing
• •Most respect for Alex Wang at Scale AI for prescience on data importance and navigating autonomous labeling pivot

Jonathan argues there is no AI bubble, citing the incredible power of current models and significant 'capability overhang' - models can do far more than we're currently extracting from them with proper agentic scaffolding, prompting, and fine-tuning.

• •Models like GPT-5, Gemini Pro, Claude, Grok are 'fucking awesome' and only getting better
• •Capability overhang: models capable of X but we're only getting X-delta due to poor implementation
• •Right agentic scaffold (system prompts, context engineering, tool calls) unlocks hidden potential
• •Growing pains: 95% of enterprise pilots fail due to poor data structuring, lack of evals, and missing workflows for partial autonomy
• •Whoever wins superintelligence wins search, devices, OS, cloud, social networking - stakes justify massive forward bets

Vision for future AI interfaces beyond phones, the wide-open opportunity in robotics and embodied AI, and why Jonathan believes in slow, steady AGI takeoff rather than rapid transformation. Includes discussion of China's AI progress and open versus closed models.

• •Future devices: wearable glasses with cameras, AirPod-like effectors whispering feedback, always-on multimodal processing
• •Robotics is wide open - different platforms needed for home robots vs factory vs humanoid
• •Slow takeoff is good: gives humanity time to prepare workforce, rethink education, avoid massive displacement
• •Unlike self-driving cars, incremental AI improvements unlock value at every step
• •China (DeepSeek, Kimi, Qwen) is not underestimated in frontier AI circles - models are state of the art