Ori Goshen, AI21 Labs CEO on AI Agents, Language Models & Enterprise AI

AI21 Labs bets enterprises will own their own intelligence

AI21 Labs has spent nearly a decade on a contrarian bet, that enterprises will want their own intelligence rather than a rented frontier model. It began as a pure research lab, built its own large language models, and now customises open-weight models on client data with researchers attached. That bet looks better as open-weight models close the gap with the frontier, and the constraint has moved to whether a company can define what good looks like.

Takeaways

• Open-weight models now carry most enterprise tokens. Open source token consumption rose from about 15 to 20 percent in June to about 80 percent in November, after DeepSeek and Qwen closed much of the gap with frontier models. A specialised model built for Gong ran at 50 times lower cost than frontier models.
• A third path sits between building and buying. Goshen says enterprises have a third option beyond building models themselves or buying them off the shelf. AI21 works in embedded and managed modes, and its stated mission is to give every company its own specialised frontier-level intelligence.
• The bottleneck is defining what good looks like. A spec sets quality, the evals and rubrics, cost and latency, and AI21 turns it into a model with an optimised harness and inference endpoint. Translating what a company wants into an eval framework is the hard part across industries.
• AI21 bet on deep learning plus symbolic methods. Goshen and Yoav Shoham founded the lab at the end of 2017 on the view that deep learning is necessary but not sufficient and needs symbolic, deterministic approaches to complement it.
• AI21 sells research capacity as a service. Lab-as-a-Service puts AI21 researchers inside client teams with a training and evaluation toolbox and compute. Goshen says reinforcement learning projects that used to take months now finish in a week or two, and the company built a platform that codifies how its researchers work.
• AI21 targets digital natives becoming AI native. AI21's focus is digital natives becoming AI native, rather than AI natives such as Harvey and Rogo. Traditional firms in finance and healthcare have machine learning operations but still need to become AI native, and Goshen says their most advanced agent use is a Slack bot.

AI21's releases, in order

1. Wordtune, a writing tool released in 2020, reached tens of millions of users within two years and became the first proof that AI21's research could carry a consumer product.
2. Jurassic, the company's first large language model family, arrived in 2021 with about 178bn parameters, roughly the size of GPT-3's 175bn.