Jul 31, 2026 · 1h 18m · neon-show
The Man Training GPT, Gemini & Claude Reveals What's Coming Next | Vijay Krishnan, Turing
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth interview, Turing co-founder Vijay Krishnan joins Siddharth Ahluwalia to discuss the future of foundation model training, synthetic reinforcement learning environments, and the strategic roadmaps for AI startups and enterprises. Krishnan breaks down the evolution of high-skill AI data alignment, provides defensibility frameworks against advancing frontier models, and examines the profound economic disruptions reshaping enterprise software, IT services, and the global workforce.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
Siddhartha as informed peer 4.4 Guest teaching 5.5 Guest disagreement 1.7 Siddhartha pushing back 0.7
speaking balance: gold is Siddhartha, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:35 Rejecting Luddite VC defensibility paralysis
Vijay rejects the overly defensive posture of venture capitalists who freeze investments over model disruption, arguing that obsessing over moats while ignoring 100x novel capabilities throws out the baby with the bathwater.
Hardest push from Siddhartha ▶ 1:01:17 Challenging Turing on custom enterprise servicesSiddharth pushes back against Vijay's narrative by directly asking whether Turing's enterprise offerings are essentially custom consulting services rather than repeatable products.
Biggest teaching moment ▶ 45:20 Masterclass on RL environments and auto-verifiersVijay gives an in-depth, technical explanation of how human feedback is embedded into rule-based verifier tuples to generate high-value synthetic data rather than low-yield trajectory recording.
Siddhartha holds their own ▶ 32:18 Framing Jevons Paradox in economic labor disruptionSiddharth demonstrates deep domain knowledge by citing Jevons Paradox and legal sector timelines to counter the prevailing narrative of immediate mass job elimination by AI.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Siddhartha as informed peer | Guest teaching | Guest disagreement | Siddhartha pushing back | Why |
|---|---|---|---|---|---|---|
| Episode Highlights and Preview | 4 | 5 | 3 | 1 | The segment includes the episode teaser and initial introduction. Host Siddharth praises Turing and frames their enterprise business around custom models, but Vijay politely corrects him, clarifying that custom models are rare and pragmatic agentic workflows on existing models deliver higher ROI. | |
| Turing's Origins and Transition to Frontier Model Training | 3 | 5 | 1 | 0 | Siddharth asks Vijay to narrate Turing's journey. Vijay explains their evolution from a global engineering talent platform to training frontier models post-InstructGPT paper. | |
| The Race Among Foundation Models and Emergence of Reasoning | 4 | 6 | 3 | 1 | Siddharth frames the model race around OpenAI and Anthropic, but Vijay pushes back slightly by including Google Gemini and sharing historical context on how NLP researchers originally didn't expect next-token prediction to generate emergent reasoning. | |
| Dynamics and Core Requirements of the AI Data Industry | 3 | 5 | 1 | 0 | Siddharth inquires about the competitive dynamics between AI data providers like Scale and Turing. Vijay details the structural needs: talent volume, quality verification, and rapid adaptation to models outgrowing task benchmarks. | |
| Evolution of Code Intelligence Training and Evaluation Harnesses | 4 | 5 | 1 | 0 | Siddharth asks about the specific training data required for coding benchmarks like Claude. Vijay breaks down the transition from LeetCode-style single-problem evaluations to complex, multi-repo execution harnesses. | |
| Market Neutrality and Foundation Labs' Data Partnerships | 5 | 6 | 3 | 2 | Siddharth raises VC anxieties regarding Claude commoditizing startups and asks about defensibility moats. Vijay dismisses Luddite VC thinking, arguing founders should prioritize building 10x value impossible four years ago over defensive paranoia. | |
| Next-Generation Picks and Shovels in the AI Ecosystem | 5 | 5 | 1 | 0 | Siddharth prompts Vijay on future picks-and-shovels opportunities beyond compute and databases. Vijay highlights verticalized interfaces, GPU/memory allocation optimization, and the shift to RL environments with auto-verifiers. | |
| Practical Enterprise AI Applications and the Paradigm Shift in Coding | 4 | 5 | 1 | 0 | Vijay outlines Turing's latest enterprise co-pilots in underwriting and auditing, comparing modern AI-assisted software development to the historical leap away from assembly code. | |
| AI's Impact on Jobs, Jevons Paradox, and Career Evolution | 6 | 6 | 2 | 2 | Siddharth brings up economic theories on labor displacement versus Jevons Paradox. Vijay agrees with Jevons Paradox in principle but points out the critical condition: humans must remain superior to models at least at one component task. | |
| Innovation Versus Execution and Macroeconomic Disruption | 5 | 5 | 2 | 1 | Vijay reflects on his incorrect forecast regarding startup explosion, distinguishing execution speedup from invention/innovation. Siddharth adds macroeconomic risk context regarding Philippine call centers. | |
| Turing's Strategic Advantage in High-Skill Talent Sourcing | 4 | 7 | 2 | 1 | Vijay provides an extensive educational breakdown of how synthetic data generation works in practice, contrasting low-ROI trajectory scraping with simulated rule-based verifier environments using a concrete Salesforce task. | |
| Physical AI, Multimodal Data, and Long-Horizon Tasks | 5 | 5 | 1 | 1 | Siddharth asks about physical intelligence startups collecting egocentric video from factory workers. Vijay notes Turing works on multimodal physical AI and explains why long-horizon tasks remain the primary model bottleneck. | |
| Enterprise AI Adoption Patterns and Autonomous Security Risks | 4 | 5 | 1 | 0 | Vijay discusses enterprise adoption differences across virtual versus physical industries, and highlights emerging cybersecurity vulnerabilities from massive unreviewed AI-generated codebases. | |
| The Future of IT Services and Turing's Transformation Value | 5 | 6 | 2 | 2 | Siddharth asks if Turing competes with Accenture and traditional IT services. Vijay details why legacy headcount-based consulting is broken and explains Turing's unique moat derived from early visibility into unreleased frontier models. | |
| Navigating the Product Versus Services Spectrum in AI Startups | 5 | 5 | 2 | 1 | Siddharth asks whether a 50/50 services-to-product split is acceptable for early-stage AI startups. Vijay supports forward-deployed services as a defensibility moat against rapid model improvements. | |
| AI Talent Acquisition Strategy and Internal Model Research | 4 | 6 | 2 | 0 | Vijay breaks down talent realities, warning application founders against overpaying elite ML researchers who care about frontier foundational physics rather than domain workflow automation. | |
| Frontier Lab Expansion and Developing Resilient AI Products | 5 | 6 | 2 | 1 | Siddharth asks if frontier labs will consume both infrastructure and application layers. Vijay agrees bandaid wrappers will die, pointing to Palantir as the archetype of building products that compound in value as models improve. | |
| Legacy SaaS Disruption and the Innovator's Dilemma | 4 | 5 | 1 | 0 | Vijay discusses the innovator's dilemma facing legacy SaaS vendors who attempt cosmetic AI add-ons rather than full architectural reinventions. | |
| Playbook for Forward Deployed Engineering and Enterprise Scaling | 5 | 6 | 1 | 0 | Siddharth asks how startups should structure initial forward deployed engineering engagements without burning out. Vijay delivers tactical advice on scaling Fortune 100 contracts to multimillion-dollar relationships. |
Statements from this episode (29)
Krishnan: Agentic products on existing models offer highest enterprise AI ROI
“And very often we do find that the right kind of agentic products built on top of existing models is the right high ROI decision for most enterprises.”
Krishnan: Turing began helping OpenAI train models in early 2022
“This whole thing started when yeah, when we started helping OpenAI you know, A little over about four years ago, like, since from early 20, 20 to long before chat GPT launched and the that is how we got started in this particular area.”
Krishnan: Foundation model companies cannot compete without expert human data
“This became pretty much like a must have. It sort of became like if you're running a model company without this component, it's like, you know trying to enter a race track with three tires instead of four basically.”
Krishnan: OpenAI, Anthropic, and Google models are within a hair's breadth of parity
“I do think these you know, models of OpenAI, Anthropic, Google have definitely gotten to the point where they are within a sort of hair's breadth of each other, ultimately. Yes, there are the, in domain A, one model might be better, in domain B, the other mode…”
Krishnan: Nobody expected Transformers and BERT to unlock emergent reasoning or AGI
“At least at the time Transformers came out, BERT came out and all that, nobody really thought this was a this was a path to emergent reasoning capabilities, a path to AGI itself.”
Krishnan: AI model training tasks get deprecated within six to nine months
“The one thing I think that makes this business a bit of a moving target is that the models do learn pretty rapidly. Like if you are offering a certain type of this thing a certain kind of a collaboration to labs with a certain, let's say, type of talent and a …”
Krishnan: Coding and math training improves overall AI reasoning capabilities
“Coding in general has nice spillover effects in terms of improving reasoning capabilities of models as a whole. Whereas you know, let us say the model companies prioritized something like the models getting better at medicine. I'm not too sure if that would he…”
Krishnan: End-to-End Outcome Startups Benefit from Smarter Frontier Models
“Now, when you're doing this kind of end-to-end thing, you are less likely to get disrupted. In fact, if you know, if anything, if the if the cloud models or codex or even Gemini models or any of these other models get a lot better at coding, these people only …”
Krishnan: Rapid growth and 10x product improvements trump AI defensibility
“The number one thing before you know, you get into defensibilities, make sure that whatever it is you're building is is was impossible to build four years ago. Make sure, you know, it is a 10 X or a hundred X better than the status quo. Make sure that it has a…”
Krishnan: Engineers at many companies have not written code in six months
“There are so many companies now where software engineers have not had to write a single line of code in say more than six months, which is I mean, quite unprecedented, right?”
Krishnan: AI will reach coding-level trust in law and taxes in 1-2 years
“Definitely. This is not the case with the law and a number of other law or taxes and a number of other things, but it's only a matter of time. I think the models will get there in the next probably year or two.”
Krishnan: Cheaper AI legal services will increase aggregate demand
“I do think for example access to legal services is incredibly expensive in the U S and I'm sure this thing will play out in terms of increased demand, et cetera.”
Krishnan: Economic value will concentrate entirely on uniquely human tasks
“If the human can do at least one thing better than the model, yes, they are going to add value. That is where the costs will go. If all the other costs will go to zero, this will become, you know at least as far as micro economics goes the this is where the ye…”
Krishnan: Over half of global software engineers have narrow task scopes
“A big chunk of maybe more over half the you know software engineers in the world operate at a relatively narrow task scope.”
Ahluwalia: Required software developers per company are shrinking by magnitudes
“The role of a single software developer is increasing, whereas the number of software developers required in a single company is shrinking by magnitudes.”
Krishnan: Prediction of a 10x AI-driven startup surge was wrong
“One prediction on way on which I, I'm just, I turned out to be wrong was I would have expected this kind of productivity gains to come all the way on the stack along the stack. Like I would have expected that, for example, the number of let's say good venture …”
Krishnan: AI Accelerates Execution but Not Fundamental Innovation
“Yes, while this has done a lot to the execution part, it does not really seem to have speeded up the innovation and the invention part, because theoretically, imagine you speed up everything, then who cares? There are 10 times the number of venture-funded star…”
Ahluwalia: Call Centers Generate 16% of the Philippines' GDP
“Today, 16% of the Philippines GDP come through call centers.”
Krishnan: LLMs require high-skill experts unlike traditional ML low-skill labelers
“While they were definitely data annotation companies that had relatively more scale low skill labelers. Yes, which is what a lot of the traditional machine learning pipelines needed. The what these LLMs and the modern foundation models needed was just somethin…”
Krishnan: Purely model-generated synthetic data hits a wall without human feedback
“There is also, I think, limits to how much value can be added there because ultimately there is just not too much new information. If you're telling the model itself to kind of generate, yeah, There is a, I mean, there's all sorts of things about how beyond th…”
Krishnan: Human-designed prompt and auto-verifier tuples maximize synthetic training data ROI
“The, this method is the one, I think, which has a lot of legs in the, particularly the more you can operate in this particular paradigm of prompt and then these rule or rubric based verifier tuples. The, that is a very nice way for sort of creating synthetic d…”
Krishnan: AI data headroom lies in multi-day, long-horizon tasks
“Generally speaking, there's less of a room for simpler tasks. The more you're doing expert level tasks the more room might be there. The more you are into some you get closer and closer to these long horizon tasks. The kind of thing that would take a human one…”
Krishnan: Unreviewed AI-generated code will create widespread security vulnerabilities
“Where people are suddenly able to produce many tens of thousands of lines of code. Nobody has time to review. Nobody has this one. At best, people can do a sanity check and move on. I guarantee it is going to create a lot of security holes and various other so…”
Krishnan: Headcount-based IT services models face severe pressure and must reinvent
“So they will definitely have to have to get substantially reinvented, right? Like today, very much the models in these companies have been entirely about this entirely head count related that is yeah, that model is definitely going to get a lot come under seve…”
Krishnan: Meaningful services provide AI startups a moat against model advances
“Having having the need for a meaningful services component does, for example I mean, it serves a useful purpose also today, right? Like in a manner which might not have been valued as much a few years ago, which is that It absolutely gives you a certain moat a…”
Krishnan: Application startups do not need to hire ML researchers
“If it's an application startup, I don't think you even need an ML researcher to begin with, right? And even an ML researcher will get frustrated.”
Krishnan: Startups acting as band-aids for model flaws won't survive upgrades
“I, even a couple of years ago, I always thought certain categories of startups, which were doing some almost like a bandaid on what is today's a hole in today's model is really not going to survive because the next version doesn't need that bandaid basically. …”