In February 2025, Mira Murati left OpenAI. As the company's Chief Technology Officer, she had helped shepherd ChatGPT from research prototype to the fastest-growing consumer application in history. Before OpenAI, she had worked at Tesla on the Model X. She understood how to move advanced technology from lab to product. She also understood, perhaps better than anyone outside a small circle of researchers, what the technology still could not do — and why.
Within weeks of her departure, she announced Thinking Machines Lab, a new AI research company built on a thesis that sounds almost paradoxical: artificial intelligence has advanced dramatically, but the tools to make it genuinely useful for specific problems remain locked behind a few well-resourced organizations. The gap is not in capability. It is in access.
Joining her were John Schulman, a co-founder of OpenAI and one of the architects of reinforcement learning from human feedback (RLHF), the technique that made ChatGPT conversational, as chief scientist; and Barret Zoph, who had led post-training optimization at OpenAI and is now CTO. Between them, they represent a concentration of expertise in building, training, and fine-tuning large language models that is difficult to match anywhere in the industry.
Andreessen Horowitz agreed. The venture firm led a $2 billion seed round valuing Thinking Machines Lab at $12 billion — a record for a seed-stage investment and a number that reflects both the pedigree of the founders and the scale of the bet.
The Problem Nobody Talks About
The public conversation about AI tends to focus on what models can do: write essays, generate images, pass bar exams, diagnose diseases. What gets far less attention is how hard it is to make a general-purpose model work well for a specific task.
Consider a concrete example. A water district in the San Joaquin Valley has twenty years of monitoring data — dissolved oxygen levels, turbidity readings, nitrate concentrations, flow rates, weather conditions, agricultural runoff patterns. They want a model that can predict contamination events before they reach intake facilities. In principle, modern AI can do this. In practice, the water district faces a series of obstacles that are almost entirely about access, not capability.
Training a custom model requires GPU clusters that cost millions of dollars to operate. The expertise to fine-tune a foundation model on domain-specific data exists primarily at Google, OpenAI, Meta, and a small number of elite university labs. The tooling is fragmented, poorly documented, and assumes a level of machine learning engineering skill that a water quality scientist does not have and should not need.
The result is a world in which the most powerful analytical technology ever created is available mostly to organizations that are already powerful. A Fortune 500 company can hire a team of ML engineers to build a custom model. A county water district, an environmental nonprofit, a small research lab — they use off-the-shelf tools that were built for general purposes and work adequately but not precisely for their specific needs.
Thinking Machines Lab was founded to close that gap.
Tinker: Fine-Tuning for Everyone
The company's first product is called Tinker, and it is a fine-tuning API — a service that lets anyone train a custom AI model on their own data without managing GPU infrastructure, writing training loops, or understanding the internals of transformer architectures.
The technical approach uses LoRA adapters (Low-Rank Adaptation), a technique that modifies a small number of parameters in a pre-trained foundation model rather than retraining the entire model from scratch. This is significant for two reasons. First, it dramatically reduces the computational cost of customization — by orders of magnitude compared to full fine-tuning. Second, it means users are building on top of models that already understand language, reasoning, and general knowledge. They are adding domain expertise, not starting from zero.
Tinker charges per token processed during training, not per GPU-hour. This pricing model is deliberately accessible: users pay for what they use rather than reserving expensive hardware. New accounts receive $150 in free credits — enough to run meaningful experiments before committing budget.
The workflow is straightforward. Upload a dataset — question-answer pairs, documents, structured records, whatever represents the domain knowledge you want the model to learn. Configure a few parameters (or accept the defaults). Tinker handles the rest: data validation, training orchestration, evaluation, and deployment. The result is a custom model endpoint that the user can query via API, embed in applications, or use for batch analysis.
What People Are Building
The early use cases demonstrate the breadth of what becomes possible when fine-tuning is accessible.
Glean, an enterprise search company, used Tinker to train Waldo, a search agent optimized for navigating internal corporate knowledge bases. By fine-tuning on retrieval and ranking tasks specific to enterprise document structures, Waldo achieved 50 percent lower latency than the general-purpose model it replaced — not because the base model was bad, but because it was not specialized.
Chroma, which builds vector database infrastructure for AI applications, trained Context-1, a model optimized for document retrieval and semantic chunking. The model understands how to break documents into meaningful segments and rank them by relevance to a query — a task that general models handle reasonably but that a fine-tuned model handles precisely.
Lightning Rod Labs built a model for clinical outcome prediction in healthcare settings. Their fine-tuned model showed 70 percent better calibration than baseline — meaning its confidence scores more accurately reflected the true probability of outcomes. In clinical settings, calibration is not a nice-to-have. A model that says “80 percent likely” should be right 80 percent of the time, not 60 or 95. Poor calibration can lead to overtesting, undertreating, or misallocating scarce medical resources.
Mantic applied Tinker to geopolitical forecasting, training models on historical event data and structured analyst assessments to produce probabilistic estimates of future events. Forecasting is a domain where general-purpose models are notoriously overconfident and poorly calibrated; domain-specific fine-tuning measurably improved both accuracy and reliability.
Open Science, Open Weights
Thinking Machines Lab has adopted a public position on openness that distinguishes it from most frontier AI companies. Their stated philosophy — “Science is better when shared” — is backed by a commitment to open publishing of research findings and the release of open-weight models under the Inkling family.
Open weights means that the trained parameters of the model are publicly available for download, inspection, and modification. This matters for several reasons. Researchers can study the model's behavior, identify biases, and verify claims. Developers can run models locally without sending sensitive data to external APIs. Organizations in regulated industries — healthcare, finance, government — can deploy models in environments where data cannot leave their infrastructure.
The open-weight approach also creates a feedback loop that proprietary models cannot match. When thousands of researchers and developers can examine a model's internals, problems get found faster, improvements get contributed back, and the collective understanding of how these systems work accelerates. This is not altruism (or not only altruism) — it is a research strategy. Open models attract talent, generate publications, and create ecosystems.
Why This Matters for Environmental Work
The environmental applications of accessible fine-tuning are not hypothetical. They are immediate and specific.
Water quality monitoring. Municipal water systems and environmental agencies collect enormous volumes of sensor data — pH, turbidity, dissolved oxygen, conductivity, specific contaminants — often at fifteen-minute intervals across dozens of monitoring points. A fine-tuned model trained on a specific watershed's historical data can learn the normal patterns and flag anomalies that might indicate contamination events, equipment failures, or upstream changes. General-purpose anomaly detection works. Watershed-specific detection works better.
Contamination spread prediction. When a spill occurs — industrial discharge, agricultural runoff, a pipeline leak — the critical question is where the contamination will go and how fast. Hydrological models exist, but they require expert setup and calibration for each watershed. A model fine-tuned on historical spill events, flow data, and water quality outcomes for a specific river system could provide rapid predictions that inform evacuation decisions, treatment plant adjustments, and downstream warnings.
Deforestation tracking. Satellite imagery services like Global Forest Watch provide broad-scale deforestation alerts, but distinguishing between legal land clearing, illegal logging, natural disturbance, and sensor artifacts requires contextual judgment that varies by region. A model fine-tuned on labeled imagery from a specific forest region — trained to recognize the visual signatures of illegal activity versus permitted clearing in that landscape — can dramatically reduce false positive rates and improve response times.
Climate impact modeling. Regional climate projections depend on downscaling global models to local conditions, a process that traditionally requires substantial computational resources and domain expertise. Fine-tuned models trained on local weather station data, historical crop yields, and regional economic indicators could provide community-level climate impact assessments at a fraction of the current cost.
The common thread is the same one Thinking Machines Lab identified: the capability exists, but the access does not. A county environmental office does not have Google's ML engineering team. It does not need Google's ML engineering team. It needs a tool that lets it upload its own data and get back a model that understands its specific problems.
The Democratization of Intelligence
The word “democratization” is overused in technology. It is applied to everything from social media (democratizing publishing) to 3D printing (democratizing manufacturing), often with more enthusiasm than accuracy. But in the case of AI fine-tuning, the term is precise.
For the past decade, the ability to build a state-of-the-art AI model has been concentrated in perhaps five to ten organizations worldwide. The compute costs alone — tens to hundreds of millions of dollars for a single training run — ensure that only the best-funded labs can participate. The result is not just a commercial imbalance but an epistemic one. The organizations that understand how AI systems work, what they can do, and where they fail are the same organizations that profit from selling access to those systems. Everyone else is a consumer.
Tinker does not eliminate the concentration of frontier model training. Foundation models will continue to be built by large labs with massive compute budgets. What it does is separate creation from customization. The foundation model is the starting point. What matters — what determines whether AI is genuinely useful for a specific problem — is the last mile of specialization. That last mile is what Tinker makes accessible.
The question this raises is genuinely important: what happens when expertise is not locked behind trillion-dollar companies? When a water district can build its own contamination predictor? When a forestry nonprofit can train its own deforestation detector? When a small clinical lab can build a diagnostic model calibrated to its patient population?
We do not know the full answer yet. The technology is new, the early adopters are still publishing results, and the failure modes of widely distributed fine-tuning are not fully understood. Models trained on small or biased datasets can produce confident but wrong outputs. Without proper evaluation, a custom model can be worse than a general one. The tools are powerful, and power without discipline produces harm.
But the direction is clear. The machinery of intelligence — the ability to build systems that learn from data and make predictions — is moving out of the exclusive domain of the world's largest technology companies and into the hands of anyone with a dataset and a problem to solve. Thinking Machines Lab did not invent fine-tuning. What they built is the on-ramp.
For environmental work, that on-ramp matters more than any single model. The data exists — decades of it, in water quality databases, air monitoring networks, soil surveys, satellite archives, species inventories. What has been missing is the ability to turn that data into intelligence without a seven-figure budget and a team of specialists. That barrier is falling. The machine shop is open.