Probably secures $9M to build more reliable AI

Despite the increasing power of LLMs, hallucinations remain a persistent challenge. Even the most advanced models generate errors, and although methods exist to detect them, the industry is still refining the best approach.
Probably, which recently secured $9 million in seed funding from Andreessen Horowitz, aims to develop a more robust method for error detection.
Founder Peter Elias (pictured above) explains that the company’s mission is to prevent hallucinations and factual errors from reaching users, aiming for the 99.99% accuracy typical of deterministic systems — a level far harder to achieve with AI. Achieving this with LLMs, he notes, demands rethinking fundamental AI engineering assumptions.
Probably’s initial product is a data science tool designed to generate fast answers from complex datasets. Every result includes a citation and an audit trail of its derivation, a practice that’s becoming standard in AI tools.
To prevent errors from slipping into these summaries, the company built an elaborate harness system Elias calls a “data science mech suit.” The LLM’s initial responses are validated by a deterministic system, which rejects any results inconsistent with the dataset. The LLM has been trained against this validator, and the entire system is optimized for speed and accuracy, according to the company.
“What we learned is that the stronger your harness engineering, the less powerful the model needs to be,” Elias says. “If you refine the context sufficiently, the model doesn’t have to struggle to produce correct outputs. Essentially, it’s about reducing ambiguity.”
This enables Probably’s data science tool to operate on much smaller AI models. Elias notes that the current version uses a model “four classes weaker than frontier models,” allowing it to run on local hardware — a desktop computer rather than a data center — significantly cutting token costs associated with AI usage.
This approach is timely, as token costs rise and many customers reevaluate their AI budgets. Elias’ vision extends beyond data science; the same engine can be adapted for use cases like accounting or medical services — in his words, “any precision-sensitive use case.”
“It’s striking that the major AI labs haven’t even tried this approach,” Elias says. “They have disincentives, because they profit from the need for repeated corrections.”
Related article
TechCrunch Disrupt 2026 wraps up
The final day of TechCrunch Disrupt 2026 is here, capping off a powerful week. With over 10,000 founders, investors, and tech leaders converging at San Francisco’s Moscone West, Day 3 delivers some of the event’s most actionable and forward-looking s
QueryStory urges caution on AI-generated content
In 2009, Shapor Naghibzadeh discovered the power of compelling storytelling while working as a Google sysops engineer. During Operation Aurora, when state-sponsored hackers targeted the tech giant, he was summoned to a crisis room to clarify server a
AI Writing Code Burns 12.15 Million Yuan, Exceeding Budget by 860%: Amazon Internal Incident Sounds the Alarm for Corporate AI Costs
U.S. media recently reported a significant AI cost overrun at Amazon, where an engineer’s use of Anthropic’s Claude Sonnet model to match author data with product listings resulted in a $1.8 million bill—860% over budget. The routine data processing
Related Special Topic Recommendations
Comments (0)
0/500

Despite the increasing power of LLMs, hallucinations remain a persistent challenge. Even the most advanced models generate errors, and although methods exist to detect them, the industry is still refining the best approach.
Probably, which recently secured $9 million in seed funding from Andreessen Horowitz, aims to develop a more robust method for error detection.
Founder Peter Elias (pictured above) explains that the company’s mission is to prevent hallucinations and factual errors from reaching users, aiming for the 99.99% accuracy typical of deterministic systems — a level far harder to achieve with AI. Achieving this with LLMs, he notes, demands rethinking fundamental AI engineering assumptions.
Probably’s initial product is a data science tool designed to generate fast answers from complex datasets. Every result includes a citation and an audit trail of its derivation, a practice that’s becoming standard in AI tools.
To prevent errors from slipping into these summaries, the company built an elaborate harness system Elias calls a “data science mech suit.” The LLM’s initial responses are validated by a deterministic system, which rejects any results inconsistent with the dataset. The LLM has been trained against this validator, and the entire system is optimized for speed and accuracy, according to the company.
“What we learned is that the stronger your harness engineering, the less powerful the model needs to be,” Elias says. “If you refine the context sufficiently, the model doesn’t have to struggle to produce correct outputs. Essentially, it’s about reducing ambiguity.”
This enables Probably’s data science tool to operate on much smaller AI models. Elias notes that the current version uses a model “four classes weaker than frontier models,” allowing it to run on local hardware — a desktop computer rather than a data center — significantly cutting token costs associated with AI usage.
This approach is timely, as token costs rise and many customers reevaluate their AI budgets. Elias’ vision extends beyond data science; the same engine can be adapted for use cases like accounting or medical services — in his words, “any precision-sensitive use case.”
“It’s striking that the major AI labs haven’t even tried this approach,” Elias says. “They have disincentives, because they profit from the need for repeated corrections.”
TechCrunch Disrupt 2026 wraps up
The final day of TechCrunch Disrupt 2026 is here, capping off a powerful week. With over 10,000 founders, investors, and tech leaders converging at San Francisco’s Moscone West, Day 3 delivers some of the event’s most actionable and forward-looking s
QueryStory urges caution on AI-generated content
In 2009, Shapor Naghibzadeh discovered the power of compelling storytelling while working as a Google sysops engineer. During Operation Aurora, when state-sponsored hackers targeted the tech giant, he was summoned to a crisis room to clarify server a
AI Writing Code Burns 12.15 Million Yuan, Exceeding Budget by 860%: Amazon Internal Incident Sounds the Alarm for Corporate AI Costs
U.S. media recently reported a significant AI cost overrun at Amazon, where an engineer’s use of Anthropic’s Claude Sonnet model to match author data with product listings resulted in a $1.8 million bill—860% over budget. The routine data processing





Home






