The Twenty Minute VCTuring CEO Jonathan Siddharth: Who Wins in Data Labelling & Why 99% of Knowledge Work Will Disappear
CHAPTERS
- 0:00 – 1:28
Why Turing Isn’t a Talent Marketplace: The Shift to “Research Accelerators”
Jonathan reframes Turing away from the talent marketplace label, positioning it as a key data partner helping frontier labs push toward superintelligence. He outlines why the classic “data labeling company” framing is outdated given today’s changing model needs.
- •Talent marketplaces match people to jobs; Turing’s mission is enabling frontier model training
- •Superintelligence requires research, compute, and data—and Turing focuses on the data pillar
- •The industry is moving from basic labeling vendors to proactive, research-oriented partners
- •Turing works with 7 of the 8 frontier labs, emphasizing deep integration with frontier needs
- 1:28 – 3:48
How Training Data Is Changing: Simple → Complex, Tests → Real Work, Chatbots → Agents
The conversation details three major shifts in the type of data needed to improve models. As models get smarter, generating useful training data becomes more difficult and requires expert humans and real workflow context.
- •Data is moving from simple prompts (e.g., sorting numbers) to complex, real-world builds (full apps across platforms)
- •Expert domain humans are increasingly required; low/medium-skill contracting doesn’t suffice
- •The goal has shifted from passing benchmarks to performing economically valuable jobs
- •The move from chatbots to agents changes what “good data” looks like
- 3:48 – 7:53
What ‘Agents’ Actually Require: RL, Tool Use, and Business ‘Mini-Worlds’
Jonathan defines agents as systems that take actions via tools, APIs, and multi-step workflows. He explains why reinforcement learning environments are central to training agents and how they differ from SFT/RLHF used for chat-style models.
- •Agents execute actions: tool calls, API operations, and computer-use workflows
- •Chatbot training: SFT + RLHF; agent training increasingly uses reinforcement learning
- •RL environments simulate tasks with prompts, state tracking, and verifiers
- •Tool-use competence becomes a core training objective for agentic systems
- 7:53 – 10:00
Scaling RL Environments Across the $30T Knowledge-Work Matrix
Turing’s approach is to systematically model workflows across industries, functions, roles, and tasks—creating RL environments at massive scale. The goal is to capture the building blocks of knowledge work so models can learn how work is actually done.
- •Four-dimensional mapping: industry × function × role × workflow
- •Workflows are the atomic units that compose most knowledge jobs
- •Turing claims it can scale breadth and quality with time and capital
- •Jonathan argues we’re still in “innings one” for vertical/workflow data acquisition
- 10:00 – 12:30
How Turing Differs from Other Data Providers: Research DNA + Enterprise ‘Reality’
Jonathan argues labs now need data partners who can keep pace with shifting paradigms (e.g., the resurgence of RL). He also claims Turing gains an edge by deploying models in enterprises, learning where they fail in production settings.
- •Frontier labs need proactive research partners as training paradigms change rapidly
- •Recent catalysts: o1 and DeepSeek increased focus on reinforcement learning environments
- •Turing also builds enterprise solutions (Disney, Pepsi, BlackRock, etc.) to “touch reality”
- •Enterprise deployments inform what data is missing and where models break
- 12:30 – 16:54
Why Custom Models Persist: Smaller, On-Prem, Fine-Tuned Systems for Proprietary Work
Using insurance underwriting as an example, Jonathan explains why many enterprises will adopt smaller, fine-tuned models rather than rely on giant general-purpose frontier models. Data privacy, latency, and competitive advantage drive a durable need for customization.
- •Underwriting/claims tasks can be solved well with human-in-the-loop LLM systems
- •Smaller models can be faster and more accurate for narrow workflows
- •On-prem + fine-tuning protects proprietary data and institutional judgment
- •Custom agents may need internal tool integrations unique to each enterprise
- 16:54 – 20:53
Will Knowledge Work Disappear in 10 Years? The Enterprise Adoption Bottleneck Debate
Harry challenges the speed of automation due to poor enterprise processes and messy internal systems. Jonathan counters that competitive pressure will force adoption—especially where AI directly drives revenue rather than cost savings.
- •Harry: internal tooling, data hygiene, and procurement are far behind model capability
- •Jonathan: competitors operating with 1/100th headcount will force change
- •Back-office automation likely slower; front-office revenue-driving use cases faster
- •Financial services, life sciences, and pharma may adopt earlier due to direct upside
- 20:53 – 24:47
Budget Shift from Labor to AI: Early Proof, GDPVal, and What ‘Parity’ Looks Like
They discuss whether AI value depends on budget moving from human labor to AI spend. Jonathan cites early transitions in lower-risk functions and points to OpenAI’s GDPVal results as evidence models are nearing expert-level outputs on many tasks.
- •Budget transfer is most visible in customer support, copywriting, SEO, and marketing
- •GDPVal studied real deliverables across occupations and verticals
- •Jonathan claims models reach human-expert indistinguishability ~50% of the time in tested tasks
- •Multi-step, interactive real-world work still leaves large headroom for improvement
- 24:47 – 28:42
A World with Automated Knowledge Work: 100x Productivity, Entrepreneurship, and Inequality
Jonathan predicts massive leverage for individuals and a boom in entrepreneurship as intelligence becomes an API. Harry worries this could widen inequality; Jonathan argues low-cost access to intelligence will reduce barriers versus hiring expensive experts.
- •Individuals become dramatically more productive; job definitions and structures change
- •Non-technical founders can assemble ‘GPT teams’ to start companies with less capital
- •Harry: capability gaps may widen societal inequality
- •Jonathan: $20/month intelligence access may narrow gaps compared to expensive human labor
- 28:42 – 34:15
If Tech Isn’t the Moat: Data-Driven Feedback Loops, Deployment, and ‘Schlep’
Jonathan argues durable advantage will come from data-driven feedback loops generated by product usage and real deployments. He explains ‘first-mile’ and ‘last-mile’ schlep—messy data, evals, workflow design, and human-AI collaboration systems required to make AI work in enterprises.
- •Moats shift from code to feedback loops and deployment learning cycles
- •Enterprise advantage comes from discovering failures first and iterating with new data
- •‘First-mile schlep’: data extraction/cleanup, structuring, eval infrastructure, training workflows
- •‘Last-mile schlep’: partial autonomy UX, change management, tandem human+AI operations
- 34:15 – 40:49
Revenue Reality in Data Provisioning: GAAP vs ‘GMV,’ Repeatability, and Trust
Harry presses on whether reported ‘revenue’ in the space is actually GMV-like pass-through. Jonathan avoids naming peers but explains Turing’s view: recurring project-based revenue that depends on consistent performance, secrecy, and trust with labs.
- •These aren’t SaaS ARR economics; they require first-principles interpretation
- •Revenue repeats via ongoing projects rather than pure subscription retention mechanics
- •Trust and secrecy/firewalling are critical to be a long-term lab partner
- •Labs use multiple providers for resiliency and pricing leverage
- 40:49 – 44:09
Market Structure and Concentration: Few Trusted Partners, NVIDIA Analogy, and Sovereign AI
They discuss concentration risk of serving a small number of frontier labs and whether that’s acceptable. Jonathan compares it to NVIDIA’s customer concentration and expects continued massive spend; they also anticipate sovereign models requiring localized data and deployments.
- •Only a small handful of providers are deeply trusted by frontier labs
- •Concentration can be rational given scale of spend—analogous to NVIDIA’s top-customer mix
- •Compute/data/energy spend is accelerating (e.g., very large infrastructure investments)
- •Governments will likely require sovereign models; localized nationals may supply training data
- 44:09 – 52:18
No AI Bubble? Model ‘Capability Overhang’ and Why Enterprise Pilots Fail
Jonathan rejects the bubble narrative, arguing models are already powerful and improving. He attributes weak enterprise outcomes to missing scaffolding: structured data, tool integrations, evals, and partial-autonomy workflow design (Cursor-like patterns).
- •Models are ‘the worst they’ll ever be’ and still remarkably capable
- •‘Capability overhang’: humans haven’t unlocked full power via scaffolding/context/tooling
- •Pilot failure drivers: messy data, weak evals, missing tool-use integration, poor workflow design
- •Partial autonomy interfaces often outperform attempts at full autonomy too early
- 52:18 – 1:00:27
Is SaaS Dead? Build-vs-Buy, Foundation Model Encroachment, and the Death of GUIs
Jonathan argues traditional SaaS is threatened by cheaper custom app creation, model providers moving into apps, and agents bypassing GUIs. Harry counters that companies won’t build/maintain 80–100 tools and that vertical SaaS remains defensible.
- •Jonathan’s risks for SaaS: companies build in-house; foundation models ‘sonic boom’ apps; GUIs designed for humans become obsolete
- •Future interfaces may be ambient, voice-driven, multimodal, and tool-call based
- •Harry: tool sprawl and maintenance realities keep SaaS viable; verticalization increases defensibility
- •They debate whether new companies already use fewer SaaS tools post-LLMs
- 1:00:27 – 1:07:41
The Next UI: Phones, Wearables, Always-On Sensors, and ‘Brain Extension’ Devices
Jonathan predicts always-on multimodal devices—glasses, earbuds, and ambient systems that see and hear context and prompt users in real time. He expects the ‘phone’ to change form and function, becoming far more than a screen-based app launcher.
- •Future devices optimize for sensors (vision/audio) and effectors (whispered guidance, actions)
- •Real-time coaching: interpreting body language, attention, and conversational dynamics
- •Persistent memory: capturing and retrieving context later like an external brain
- •The ‘smartphone’ will likely evolve dramatically; the phone app is already the least interesting part
- 1:07:41 – 1:17:01
Quick-Fire: Slow Takeoff, China, Open vs Closed, Leadership Lessons, and What Excites Him
In rapid responses, Jonathan argues for steady AI progress rather than sudden takeoff, and says China’s open models are impressive. He discusses nuanced views on open vs closed models, shares management lessons from operating closer to ground truth, and closes on excitement about AI accelerating discovery and self-improvement.
- •Belief: AI improves steadily; incremental gains create value unlike all-or-nothing autonomy
- •China: frontier circles take DeepSeek/Qwen/Kimi seriously; they’re close to state-of-the-art
- •Open vs closed: enterprises may mix; frontier capability may justify more closure for safety
- •Personal shifts: more hands-on leadership, hub-and-spoke offices, excitement for AI-driven discovery and agentic ‘Iron Man’ leverage
