Technical Context
I deliberately dug into both announcements because the wording can easily mislead. Inkling from Thinking Machines Lab is not a geopolitical forecasting model but an open-weight multimodal system with 975 billion parameters, designed for video and audio understanding.
So if you need AI integration for media, surveillance, video analytics, or complex multimodal pipelines, this is worth a look. But when it comes to event prediction, you have to examine TML's second article, which involves a completely different stack.
In their world events piece, the team describes fine-tuning gpt-oss-120b on around 10,000 binary questions like “will event X happen by date Y.” The two-stage process: first, the model explores context, then outputs a probability, and reinforcement learning rewards it for correct real-world outcomes.
That's interesting not as a demo but as an engineering pattern. I've long argued that AI implementation shouldn't be limited to text generation: if you can tie the model to a measurable external outcome, it starts operating as a predictive layer atop business processes.
The numbers are presented soberly, without hype. TML doesn't claim their fine-tuned model magically crushed everyone; they report slightly better results against frontier models in head-to-head comparisons and a strong contribution in an ensemble with Grok 4. That sounds plausible: in forecasting, ensembles are almost always more useful than a single “superbrain.”
What This Changes for Business and Automation
For business, the main takeaway isn't that “LLMs now know the future.” It's that you can train a model not to sound clever but to better estimate the probability of events: supply chain disruptions, regulatory risks, demand spikes, regional escalations, or customer churn.
The winners will be teams that already have data pipelines and a history of decisions. They can build AI solutions for business not as a chatbot layered over a CRM, but as a probabilistic analytics layer for procurement, sales ops, risk, and planning.
The losers will be those who continue evaluating everything by text quality. In these tasks, calibration of probabilities, validation schemes, data slicing, and leakage prevention matter far more than “how convincing the answer sounds.”
I typically audit such setups on architecture: where does the factual grounding come from, how is the reward computed, where is drift control, and who owns the final judgment. These are exactly the points where most glossy presentations break down, and these are the bottlenecks we solve for clients at Nahornyi AI Lab when we build AI automation for real decisions—not just wow effects. If you already have a scenario where you need not to generate but to predict and help your team act sooner, you can calmly review it with Vadym Nahornyi and design a working system without unnecessary magic.