This week's briefing (9–15 Nov)Data updated 6:35 IST, 22 NovHow we check claims
Mira Murati
Thinking Machines Lab · Lab leaders. 2 sourced entries, newest first. Curated coverage with primary sources, not a complete feed; a role establishes context, not authority.
Thinking Machines Lab · 29 October 2026 · Keynote
Murati at Dartmouth's 'AI at 70': success means keeping human judgment at the center
Primary source, read in fullAI safetyHuman-AI interaction
In the opening keynote conversation of Dartmouth's 'AI at 70' conference, Mira Murati said success is 'the degree to which we succeed in keeping human flourishing, human judgment, and human will at the center' of AI development, discussed the customizable AI systems Thinking Machines builds, and reflected on ChatGPT's release as showing how design lets people find practical uses while exposing misuse. At the same event Yann LeCun said that 'we're not anywhere close to human-level AI' despite optimistic Silicon Valley CEOs.
What this could change
Editorial: Murati rarely speaks publicly, and her framing of customizable AI around human judgment hints at how Thinking Machines will position itself against chatbot incumbents.
Perspective & interests
Murati's company sells customizable AI tooling, so framing success as user control and human judgment doubles as positioning for Thinking Machines' products.
Reviewed 2026-11-06. Also features Yann LeCun. Entry ID mira-murati-dartmouth-ai-at-70-2026-10 · Permalink
Thinking Machines Lab · 15 October 2026 · Company statement
Thinking Machines marks one year of Tinker: custom fine-tuned models beating frontier LLMs on niche tasks
Primary source, read in fullFine-tuningEnterprise AIModels
Thinking Machines Lab's official post reviews one year of Tinker, its post-training API for fine-tuning custom models without managing GPUs. It says the platform grew from 14 to 40+ supported models, names customers including Glean, Lightning Rod Labs, Sundial and Bridgewater, and claims customer-trained models beat frontier LLMs on specialized tasks at far lower cost (e.g. Sundial matching frontier performance at 7% of the cost). It announces enterprise security features, a production inference API beta, and new training methods including CISPO, DRO and full fine-tuning.
What this could change
Editorial: concrete customer numbers showing small fine-tuned models beating frontier APIs at a fraction of the cost is the strongest current argument for builders to sell custom models rather than prompt wrappers.
Perspective & interests
Mira Murati's Thinking Machines Lab sells Tinker subscriptions, so evidence that fine-tuning beats frontier APIs is its core commercial narrative.
Reviewed 2026-11-06. Entry ID mira-murati-one-year-of-tinker-2026-10 · Permalink