While much of the AI sector focuses on addressing hallucinations through larger and more advanced models, Probably is pursuing a different strategy.

The firm secured $9m during its seed round co-led by Andreessen Horowitz and Accel along with Tokyo Black and Vermilion Cliffs Ventures. Its goal is to identify factual mistakes in AI outputs prior to user exposure and target the 99.99% reliability standard common in traditional software but uncommon in large language models.

Its method relies less on the model itself. Probably's initial offering is a local verifiable data agent for querying complex datasets. It processes each response via what founder Peter Elias terms a data science mech suit.

A Harness Rather Than Expanded Intelligence

The model generates an initial response while a separate deterministic validator reviews it against source data and rejects mismatches. The model learns from this validator process and every output includes citations plus an audit trail.

The stronger the harness design the less capable the model needs to be Elias notes. By minimizing uncertainty the AI requires minimal reasoning.

This leads to notable cost benefits. Probably's solution operates on a model Elias characterizes as four tiers below leading systems small enough for desktop execution instead of data center resources which reduces token expenses substantially.

It also supports a privacy focus. The system functions locally using the open-source DuckDB database and the model accesses only metadata and statistics while raw data remains on the user's device.

Targeting Token Cost Concerns

The launch aligns with current trends. Organizations face rising AI expenses despite falling per-token rates and a solution offering precision on low-cost local hardware addresses these issues directly.

It targets areas where mistakes carry high costs. Probably indicates the technology could apply to accounting or medical tasks any precision-sensitive field where an incorrect confident response creates major problems as studies on scientific hallucinations continue to highlight.

A Bold Assertion With Limitations

Elias contends that major labs avoid this path because they profit from repeated model corrections. This serves as an effective marketing point though it remains debatable since leading labs invest heavily in reducing hallucinations and smaller firms naturally position themselves favorably.

The main limitation involves applicability. Validation requires verifiable ground truth such as datasets which explains the initial focus on data tasks over open writing. With $9m in seed capital the product is at public preview version 0.1 and the 99.99% target remains aspirational. Yet amid numerous efforts to manage hallucinations relying on smaller models represents a distinct strategy that attracted support from a16z and Accel.