Jonathan Siddharth

Sourcery · Frontier AI Is a Ferrari. Most Companies Need a Model Y | Turing CEO

• Why Will Frontier AI Take a Decade or Two to Spread Through Companies?

  Jonathan Siddharth expects a slow takeoff. Over the next decade or two, frontier models become more powerful and more useful, and the technology still takes time to spread through companies.

  He sets that view between two louder camps in Silicon Valley. One expects a fast takeoff, with a risk of losing control in two years and with AI exceeding human intelligence at everything. The other thinks the boom is hype and that nothing will happen. He does not take either side. The capability is already ahead of what a company can absorb, a gap he calls model capability overhang. Real work is messy. Someone starting a job often does not know where the information is, who to talk to, or which file ho

  ▶ at 56:45 Share Transcript

• Why Is Frontier AI a Ferrari While Most Companies Need a Model Y?

  Jonathan Siddharth says there is a place for Ferraris and Koenigseggs, and also a place for a Model Y when the job is getting from one place to another as efficiently as possible. Frontier AI and ordinary company work split the same way.

  When the return on a little more intelligence is very high, he wants the biggest model in the world. A model that makes someone 5% or 10% more productive, or more effective at decisions, is that case. Automating a workflow such as customer support is not. A Model Y, or a collection of Model Ys, is enough. He thinks that is what most companies have already found.

  ▶ at 35:57 Share Transcript

• How Far Behind Are Open-weight AI Models?

  Jonathan Siddharth says open-weight models are about three to six months behind the frontier, depending on who you ask. Enterprises are picking them up because the models are getting good enough for a company to own its own implementation.

  He wants frontier AI and sovereign AI together, not a contest with one winner. Frontier labs, including OpenAI, Anthropic, and DeepMind, are pushing superintelligence he thinks could help cure diseases, discover new materials, and colonize space. Plenty of company work does not need a trillion-parameter model. An invoice-to-pay reconciliation, an HR workflow, a board deck, or an all-hands can sit on a custom model tuned to the company's own data, tools, a

  ▶ at 32:21 Share Transcript

• How Long Can a Coding Agent Work on Its Own?

  Jonathan Siddharth says a coding agent today can reliably work for two days at a stretch, and is still far from working on its own for weeks, months, or years.

  The way forward is a set of reinforcement-learning environments that mimic how professionals actually work. He pictures a five-dimensional matrix of every workflow, every role, every function, every company type, and every sector. Experts still supply the prompts, the verifiers, and realistic seed data so an agent can learn long-horizon work that is worth paying for.

  The picture he offers for the next few years is not a hire that already happened. A new person at Sourcery is given a task, comes back for a one-on-one, applies the fee

  ▶ at 6:53 Share Transcript

• What Is the Best Way to Move AI Forward?

  Jonathan Siddharth says the best way to move AI forward is to close the research and deployment loop. Turing works with labs to improve models for coding and knowledge work, finds what different models and agents are good at, and sees the jagged intelligence they have.

  The work has made Turing a unique partner for enterprises deploying end-to-end agentic systems. Turing deploys them, watches where they break in the real world, and uses the capability gaps to improve the models and take on bigger enterprise problems. The cycle runs again: deploy, see the breaks, improve the models. The loop is what makes Turing a trusted partner between research and deployment. He thinks the loop is also the

  ▶ at 53:36 Share Transcript

• What Is Jonathan Siddharth's Goal?

  Jonathan Siddharth's goal is to advance the frontier of superintelligence and make sure humans benefit from it. Real economic progress is part of the goal, and none of it matters without GDP growth that is significantly more than we have today. The impact of AI should show up in the P&L of companies, in the market cap of companies, and ultimately in the GDP of the US and the world. He wants to help the U.S. stay ahead in the race to superintelligence and beyond.

  The frontier moves forward by working with all the labs building powerful proprietary models. Good open weight models should improve too, and the benefits should be diffused in the economy. Usage of frontier models is going to keep

  ▶ at 47:33 Share Transcript

• When Does an AI Agent Learn?

  Jonathan Siddharth says an AI agent learns when the training environment is neither too easy nor too hard. A reward every time teaches nothing. No reward at all teaches nothing. The environment should be set up so the agent succeeds 20 to 40 percent of the time.

  When the agent succeeds, the steps that led to the reward get reinforced. Reinforcement means a giant neural network in which certain weights get updated. It is unclear which other neural pathway gets activated along the way. Models generalize, so it is hard to predict what else training adds beyond the behavior it explicitly targets.

  The unpredictability is why alignment efforts are super important. Frontier labs work on alignment

  ▶ at 19:21 Share Transcript

• Why Is Biological AI Risk Trickier Than Cyber Risk?

  Jonathan Siddharth says frontier models are a huge net positive for humanity, and biological risk is a little more tricky than cybersecurity because the danger is a little bit asymmetric.

  It is possible to build safe RL environments that cannot be exploited. Agents need containment so they do not escape the RL environment and do things they should not do. Containment is an engineering problem, in the same way jet engines were made safe. Cybersecurity stays an active arms race between the good guys and the bad guys. With biology, the case he raises is a virus like COVID, with a high viral coefficient and a high K factor for how quickly it spreads. He leaves open whether such a virus is much

  ▶ at 9:41 Share Transcript

• What Is AI's Biggest Superpower?

  Jonathan Siddharth says AI's biggest superpower is up-leveling the type of problems humans can now solve. People who talk about AI automating jobs or replacing jobs are missing something more fundamental. One of his mentors, Alan Eustace, shared the point with him.

  In education, the shift is teaching kids how to work with the models, how to ask the right questions, and how to verify whether the output is correct. The scope of problems a person can solve is massive. In interviews, the old ask might be to implement dynamic programming or some obscure computer science algorithm. Now the ask can be to replicate Amazon. The goal is not generating a ton of code in the time period. The goal is che

  ▶ at 5:28 Share Transcript

• How Did AI Data Change in 2026?

  Jonathan Siddharth says the data landscape completely shifted in 2026 from helping AI master tests to helping AI master real work.

  In the test era, success looked like AI passing the SATs, passing the bar, or winning a gold medal in the Math Olympiad. The game was finding experts in every domain and pulling knowledge out of their heads through a dialogue with the model and an evaluation of model outputs, transferring what experts knew into the models. He used to think of that as distilling human knowledge and skills into LLMs.

  Real work flips the priority. It is less about how many experts you can find and more about how close to reality you can engineer simulated RL environments where age

  ▶ at 1:42 Share Transcript