The Twenty Minute VCTuring CEO Jonathan Siddharth: Who Wins in Data Labelling & Why 99% of Knowledge Work Will Disappear
CHAPTERS
Turing’s shift from “talent marketplace” to superintelligence data partner
Jonathan reframes Turing as a company training superintelligence rather than matching developers to jobs. He positions the frontier labs’ needs as a three-pillar stack—research, compute, and data—arguing Turing is focused on the data pillar as it rapidly evolves.
- •Defines “talent marketplace” vs. Turing’s current mission
- •Turing works with most frontier labs as a data partner
- •Superintelligence requirements: research (labs), compute (Nvidia), data (Turing)
- •Claim: data needs have changed dramatically as models get smarter
Why frontier-model data has become harder: simple labeling → expert, real-work data
The conversation explains the transition from easy-to-produce datasets (basic prompts, simple tasks) to sophisticated, domain-specific work artifacts. Jonathan argues that progress now depends on expert humans producing complex outputs that resemble real workplace deliverables.
- •Data shifted from “cat pictures/basic prompts” to complex, end-to-end tasks
- •Example: trivial coding tasks vs building full multi-platform applications
- •Low/medium-skilled contracting is less useful; experts are required
- •Training focus moves from test-passing to economically valuable work
From chatbots to agents: how training data changes (SFT/RLHF → tool use + RL)
Jonathan distinguishes chatbot training (SFT and RLHF) from agent training that requires tool-use behavior and multi-step execution. He introduces reinforcement learning environments as a core mechanism for teaching agents to act in realistic workflows.
- •Chatbots: supervised fine-tuning (SFT) and RLHF reward modeling
- •Agents: take actions, call tools/functions, execute workflows
- •Tool use becomes central to agent capability
- •Reinforcement learning becomes the dominant paradigm for agents
Building RL environments at scale: the SDR workflow example
A concrete SDR scenario illustrates how agents learn inside simulated “mini world models” with cloned tools and verifiers. Jonathan describes curriculum design and feedback signals, likening the approach to self-play dynamics from AlphaZero.
- •Create synthetic but realistic tool environments (LinkedIn/Salesforce/ZoomInfo clones)
- •Prompts define tasks; verifiers check task completion
- •Agents explore trajectories via trial-and-error to maximize reward
- •Curriculum must balance easy vs hard tasks to generate learning signal
- •Analogy to AlphaZero/self-play; synthetic data emerges from interaction
The $30T workflow matrix: industry × function × role × workflow
Jonathan argues knowledge work can be decomposed into workflows that can each be modeled and trained. Turing’s ambition is to generate RL environments across this four-dimensional space, creating coverage across the economy.
- •Four dimensions: industries, functions, roles, workflows
- •Workflows are the building blocks of roles
- •Claimed addressable scope: ~$30T of digital knowledge work
- •Debate: feasibility of breadth + quality; answer is time and capital
How Turing differs from data-labeling vendors: “research accelerator” + enterprise reality checks
Jonathan positions Turing as a research-oriented partner rather than a labeling shop, emphasizing fast-changing training paradigms. He also highlights enterprise deployments as a way to see where models fail in real conditions and feed improvements back into training.
- •“Era of data labeling companies is over” → era of research accelerators
- •Labs need proactive research partners as paradigms shift (e.g., RL environments)
- •Turing also builds enterprise solutions (Disney, Pepsi, BlackRock, J&J)
- •Enterprise work provides ground truth: where models break in practice
Why enterprises will want custom (often smaller) models: the insurance underwriting case
A detailed underwriting example explains why on-prem, fine-tuned smaller models can outperform giant general models for specific tasks. Jonathan argues enterprises want to distill proprietary institutional judgment while protecting sensitive data and integrating internal tools.
- •Underwriting uses messy unstructured medical data; LLM + human-in-loop can help
- •Smaller models can be faster/more accurate for narrow workflows
- •On-prem deployment avoids sharing proprietary data with competitors
- •Fine-tuning distills institutional knowledge from past decisions
- •Agentization includes automating internal tool calls
AI adoption constraints in enterprises: incentives, change management, and “first-mile/last-mile schlep”
Harry challenges the 10-year automation timeline by pointing to broken processes and poor data inside incumbents. Jonathan responds that competition will force adoption, especially in revenue-driving front office use cases, while acknowledging the heavy implementation work required.
- •Harry: internal data/processes are ‘laughable’; adoption lags tooling basics
- •Jonathan: competitive pressure forces adoption when stakes are high
- •Front office (profit-making) adopts faster than back office (cost-saving)
- •Financial services and pharma seen as early/fast adopters
- •Enterprises require extensive enablement beyond model selection
Labor-to-AI budget transfer: where it’s happening and how to measure value
They discuss whether AI growth depends on budgets shifting from labor to technology. Jonathan cites early wins in lower-risk domains and references research showing models already reach expert-level performance in many task categories, with big room left in multi-step work.
- •Budget transfer happening in customer support, copywriting, SEO/marketing
- •Value adoption faster in low-risk-to-fail workstreams
- •Mentions OpenAI ‘GDPVal’ style task-based evaluation across occupations
- •Models can match experts ~meaningfully often, but real work is multi-step
- •Highlights “capability overhang”: models can do more than current deployments show
If knowledge work is automated: productivity explosion, entrepreneurship, and inequality debate
Jonathan forecasts massive leverage for individuals and a boom in entrepreneurship as “intelligence becomes an API.” Harry questions whether this widens inequality; Jonathan argues cheaper access to expertise narrows gaps relative to hiring expensive humans.
- •Prediction: humans become 100× more productive; one person runs many efforts
- •Lower startup costs: non-technical founders assemble ‘GPT teams’
- •Harry: risk of widening chasm; many people may not upskill
- •Jonathan: $20/month intelligence access is more equalizing than human experts
- •Humans shift to higher-abstraction problem solving (health, aging, space)
Moats in an AI world: data-driven feedback loops and deployment flywheels
As software creation gets easier, Jonathan argues durable advantage comes from feedback loops created by real usage. He emphasizes enterprise deployment as the critical path to uncover failure modes and continuously retrain, making “touching reality” the moat.
- •Moat shifts from technology to feedback loops (Google search analogy)
- •Early deployment creates superior data and faster iteration
- •Enterprises are ‘wide open’ compared to consumer feedback loops
- •Human+AI tandem workflow produces labeled error cases for improvement
- •“Partial autonomy” UX (Cursor-like) is a key intermediate step
Revenue quality in data provisioning: GAAP vs GMV, project recurrence, and trust with labs
They unpack confusion around ‘revenue’ in the data ecosystem and how to interpret it. Jonathan describes lab work as recurring project-based demand, where reliability, secrecy, and operational firewalls are essential to remain a trusted supplier.
- •Debate: reported revenues may be GMV-like in some models
- •Turing frames revenue as traditional/GAAP-like
- •Not SaaS ARR; demand is recurring but project-driven
- •Trust and secrecy are core—firewalls between labs/teams
- •Labs work with a small handful of providers for resilience and pricing
Market dynamics: concentration, geopolitics, and why Jonathan rejects the “AI bubble” thesis
Jonathan argues concentration among a few major buyers is normal (comparing to Nvidia’s customer concentration) and expects massive spending across compute, energy, and data. He predicts sovereign models and government demand, and insists capability is real—deployment friction, not a bubble, is the bottleneck.
- •Revenue concentration compared to Nvidia’s top-customer exposure
- •Future demand: governments and sovereign/regulated deployments
- •No AI bubble: models are powerful and improving; ‘we’re used to magic’
- •Growing pains: messy data, weak evals, missing scaffolding, poor workflow design
- •Agentic scaffolds unlock latent capability already present today
Why ‘SaaS is dead’ (and the counterargument), plus the future interface beyond the phone
Jonathan claims SaaS is threatened by DIY app building, foundation model verticalization, and UI shifts away from human-centric GUIs. Harry argues most companies can’t build/maintain dozens of tools and vertical SaaS remains defensible; they then explore ambient, multimodal devices as the next interface.
- •Jonathan’s three SaaS risks: DIY build, model providers move up-stack, GUI obsolescence
- •Harry’s rebuttal: companies rely on 80–100 SaaS tools; maintenance burden is huge
- •Vertical SaaS defensibility vs foundation model focus on bigger prizes
- •Future interface: always-on multimodal wearable + audio guidance + memory
- •Phone likely evolves into a very different ‘AI-first’ device paradigm
Quick-fire: takeoff speed, China, open vs closed models, leadership lessons, and robotics opportunity
Jonathan summarizes his key beliefs: slow takeoff, serious China competitiveness, nuanced open/closed model tradeoffs, and a more hands-on leadership philosophy. He predicts a few data-market winners with research depth, and highlights embodied AI/robotics as the largest emerging whitespace.
- •Belief: incremental improvement beats rapid takeoff; society needs time to adapt
- •China not underestimated in frontier circles (DeepSeek/Qwen/Kimi mentioned)
- •Open vs closed: enterprise mix; frontier closed for safety may make sense
- •Leadership shift: closer to ground truth, less focus on being liked
- •Market outlook: a few winners; robotics/embodied AI as the big bet
