Jonathan Siddharth on AI Training Data, Enterprise Agents and the Future of Work

Harry Stebbings interviews Turing founder and CEO Jonathan Siddharth about AI training data, enterprise deployment and the future of work. Siddharth sells training and implementation services, so his market forecasts and claims about his company reflect a direct commercial interest.

Siddharth rejects describing Turing solely as a talent marketplace. He says the company helps train frontier models and works with seven of eight leading labs. The customer count and his description of the firm's role are company claims, not independently verified market coverage figures.

He sees three changes in training: data tasks have become more complex, models are increasingly evaluated on real work rather than examinations, and developers are training systems to use tools and execute multistep workflows. His examples illustrate the shift without demonstrating that present systems can perform every such job reliably.

He explains supervised fine-tuning as learning from example inputs and outputs, and human-feedback reinforcement learning as using preferences to shape responses. For an agent, he says, training also needs tool use: a system might call APIs or operate software as it works toward a goal.

One proposed training environment recreates business tools with synthetic data. An agent might prepare for a sales call, check a CRM and update a record; a verifier then checks the outcome. Siddharth says tasks should be neither so easy that the agent always succeeds nor so hard that it receives no useful feedback.

Turing builds such environments across industries, roles and workflows, he says. Stebbings asks whether broad coverage and quality are simultaneously feasible. Siddharth expects it to take time and regards the market as early; the interview does not verify the company's coverage of every possible workflow.

He calls Turing a research partner as well as a data provider and says its enterprise deployments help it see where models fail in real settings. He names several corporate customers and says the model-training business is larger than enterprise work, without providing audited revenue shares or project outcomes.

For a possible insurance workflow, he describes using a smaller model, a customer's proprietary records and human review to support underwriting. He argues a specialized model can be useful for latency, accuracy or data control. These are proposed design choices, not evidence that a specific model is accurate or appropriate for insurance decisions.

A customer might host a fine-tuned model and connect it to internal tools, he says, so institutional knowledge can inform the workflow without being shared with competitors. Handling sensitive medical information also depends on safeguards and oversight that the interview does not specify.

Siddharth expects the enterprise business to grow, but says it is still early. He then predicts that all digital knowledge work involving computers and tools will eventually be automated. This is a long-term forecast from the CEO of a company selling AI services, not an established capability or employment outcome.

Stebbings pushes back on the pace, pointing to fragmented data, purchasing procedures and legacy systems at large organizations. Siddharth says competitors that adopt better tools could put pressure on incumbents, but accepts that change management in back offices may be slow.

Siddharth is more optimistic about front-office applications where an organization sees a direct potential revenue gain, including finance and drug research. These are adoption hypotheses; the interview does not establish that such applications reliably improve investment or clinical outcomes.

The host asks whether budgets are actually moving from human labor to technology. Siddharth points to customer support, copywriting and search-related marketing as areas where he sees faster change. He does not provide a measured budget transfer or evidence that an entire job category has disappeared.

He cites OpenAI's GDPval research on occupational tasks and reports that leading models sometimes produced work comparable with expert work. The interview does not provide enough of the study's evaluation design, task scope or uncertainty to generalize its approximate figure to all jobs or to claim full autonomy.

Asked what follows widespread automation, he imagines individuals becoming much more productive and nontechnical founders using AI helpers instead of assembling large initial teams. His examples of 100-fold productivity and running many companies are speculative scenarios rather than measured results.

Stebbings worries that benefits may be uneven and widen inequality. Siddharth argues that lower-cost access to expertise could broaden opportunity, but the discussion offers no evidence that access, quality or gains will be evenly distributed. His hopes for curing disease and other breakthroughs are possibilities, not established outcomes.

When technology itself is easier to copy, Siddharth expects data-driven feedback loops to become an important advantage. He uses search as an analogy: users' queries and interactions can help a product improve. The example is his explanation, not proof that data alone determines competitive success.

He argues enterprise deployments create feedback when customers encounter failures in real workflows. That information can guide further evaluation and training. Turing provides services in this area, so the proposed feedback cycle is also a description of its business thesis.

The practical first step is often less elegant than a model demonstration: company data may sit in spreadsheets and disconnected systems. Siddharth says teams must acquire and structure data, build evaluations, integrate tools and train staff to use redesigned workflows.

He describes running an AI system alongside a human for a period so a manager can compare outputs. Differences may reveal a need for staff training or new model data. He says a single mistake would not justify dismissing a person; performance should be assessed over time and by appropriate measures.

Asked about margins and pricing for the enterprise work, Siddharth says they vary and the company is still learning how to price it. He says current work is billed for time and may eventually move toward value-based pricing; no audited unit economics are supplied.

Stebbings asks whether industry headline numbers are revenue or gross marketplace volume. Siddharth declines to assess rivals and says Turing thinks in terms of generally accepted accounting revenue. He adds that its projects start and end, even when customer demand recurs, and distinguishes this work from conventional SaaS annual recurring revenue.

He says customers work with several trusted data partners for resilience and price competition, and describes confidentiality barriers between projects. Those are company practices and market observations as he presents them; the interview does not independently audit the controls or providers' revenues.

Siddharth praises Scale AI founder Alexandr Wang and says demand for Turing rose after the change involving Scale. He declines to judge the competitor's current business. His account of the demand shift is a Turing executive's observation, without disclosed contract-level evidence.

Asked about concentration among a handful of model labs, he says he is not particularly worried and compares the pattern to a concentrated semiconductor customer base. He also expects governments to invest in local model capacity. Both the comparison and future demand are his views; he does not disclose Turing's customer concentration.

Stebbings raises the possibility of an AI investment bubble and disappointing near-term revenue. Siddharth says he does not see a bubble, citing current model capability and possible improvements. Useful technology, however, does not by itself establish that every investment price or future revenue forecast is justified.

He calls the difference between a model's potential and what users currently extract a capability overhang. Better context, tool integration, evaluation and human collaboration may close part of it. As an illustrative possibility, he suggests automating some of Stebbings's clip-selection workflow; no deployed result is shown.

For enterprise pilots, he points again to data preparation, appropriate scaffolding and evaluations. He describes partial autonomy, in which a person works alongside a model, as an intermediate design for some tasks. Whether any system can safely act without review depends on the particular task and evidence, not a general claim.

He expects some familiar tasks, such as certain support workflows, to need less intermediate human involvement, while specialized legal work may require private data and fine-tuning. These are examples of possible designs, not proof that current systems can handle all support or legal decisions independently.

On circular investment and infrastructure deals, Siddharth argues that firms expecting advanced general AI may view large upfront bets as necessary to compete. He believes the potential prize is vast. This is his rationale and a speculative outlook, not an assessment of the financial sustainability of the transactions.

Siddharth predicts conventional SaaS will be disrupted as applications become easier to build, model providers enter application markets and agents operate software through interfaces or tools. He offers these as possible risks to existing products, not a demonstrated end to all SaaS.

Stebbings disagrees. He argues that businesses use many applications that must be maintained and secured, and that highly specialized vertical products can retain advantages. He explicitly notes his own bias as an investor in such companies; the two leave the competitive outcome unresolved.

Siddharth suggests comparing software usage across the host's portfolio over time. He also predicts more people will qualify as software builders if AI makes product creation accessible to nontraditional engineers. That is a broader definition of a builder, not a forecast of conventional software-engineering job counts.

With more software, discovery may become harder, he says, and imagines agents finding tools or speaking to other agents on a user's behalf. His example of investor agents conducting many conversations is a speculative future workflow.

Asked whether the phone will remain the main interface, Siddharth imagines an always-on combination of cameras, audio and memory, possibly in glasses or earbuds. He says the form factor is uncertain. Such a device would also process sensitive surroundings, though the interview does not examine a privacy or consent design.

He expects a few data-training providers to serve labs seeking resilience and competition, and sees robotics and embodied AI as a relatively early opportunity. He thinks providers will need to adapt to changing modalities, languages and learning methods; these are market forecasts from a participant.

In the quick-fire round he reiterates a preference for gradual AI progress over a sudden takeoff. He hopes that would give institutions time to prepare education and workers, but slower progress does not guarantee that displacement can be avoided.

He praises progress by Chinese model developers and says enterprises may choose a mix of open and closed models for cost, customization and ease of use. He favors caution around releasing some powerful frontier capabilities; those are his policy judgments rather than a comparison of all model risks.

Siddharth says his management view has shifted from hiring leaders and stepping back to working closely with teams and customers. He believes direct knowledge of customer problems and engineering details helps him make decisions, though organizations may choose different operating styles.

He describes moving Turing from a distributed structure toward offices in San Francisco and Palo Alto, with London planned. The change was unpopular with some employees, and he says some left. The interview does not measure whether the new arrangement improved retention or output.

Looking ahead, he is excited by AI-assisted discovery, automating AI research and tools that expand what individuals can accomplish. His analogies to fictional autonomous suits are visions for future technology, not evidence that superintelligence or medical breakthroughs are imminent.