Key points
- Turing is profitable with over $300m in revenue and raised $225m at a $2.2bn valuation.
- Jonathan Siddarth says Turing works with seven of eight frontier labs and sells a strategic research accelerator as commodity data labeling ends.
- Siddarth predicts Turing will replace services firms like McKinsey, Bain, BCG, Accenture, TCS, Wipro and Infosys.
Turing bets frontier labs need a research partner, not labels
The internet data frontier models were pretrained on ran out about three years ago, so the labs no longer need labeling capacity. They need to know which data will move a model forward. Turing sells that, sourcing hard, realistic, diverse data and research from four million engineers, and seven of the eight frontier labs work with it. Jonathan Siddarth calls it a picks-and-shovels play on AGI.
Takeaways
- Commodity data labeling is over. After o1 and DeepSeek, labs stopped needing a data factory or talent marketplace and started needing a research partner. The data must now be hard, realistic and diverse, and Turing publishes research on what models need next.
- Speed at finding experts plus data research. Turing's four million engineers let it find smart humans fast, and it stays proactive about what data models will need. It works with seven of the eight frontier labs, among them OpenAI, Anthropic, Google, Meta and NVIDIA.
- Turing is the picks-and-shovels play on AGI. Turing has raised $225m at a $2.2bn valuation, runs more than $300m in revenue and is profitable. Jonathan compares it to NVIDIA, a picks-and-shovels play selling into the AGI industry rather than competing in it.
- Scaling laws hold, so takeoff stays gradual. Scaling laws still hold, and GPT-3's magic came from one-time jumps like eating the internet, so takeoff stays gradual. Humans in the loop remove many risks. The self-driving car analogy, where 99% accuracy is not enough, damaged AI.
- Enterprises get small custom on-prem models. Enterprises get a 500m to 10bn parameter model fine-tuned on proprietary data, often on-prem, instead of plug-and-play ChatGPT access. A slow incumbent like GEICO can be jumped by an upstart that adopts faster.
- ASI means automating 90% of computer tasks. He defines superintelligence as automating 90% of the tasks 90% of humans do in front of a computer, a $30tn market of knowledge work he lays out as a matrix of industry, function, role and workflow.
How a frontier model gets trained
- Pretraining on a filtered internet, using Common Crawl, C4, GitHub Archive, books and video, produces a base model that only autocompletes. It happens inside the labs.
- Post-training adds supervised fine tuning and reinforcement learning with human feedback, the process in OpenAI's InstructGPT paper.
- In verifiable domains like coding and math, models use plain reinforcement learning, the AlphaGo pattern of learning from experts, then self-play.
- Turing builds thousands of reinforcement learning gyms, works on embodied AI, and supplies agentic coding data that would stump today's models.