Probably Secures $9M to Make AI Less Likely to Lie

Probably just closed a $9 million seed round to build a more reliable foundation for AI agents. While current models like GPT-4o and Claude 3.5 Sonnet are impressive, they still suffer from frequent hallucinations and logic gaps that make them risky for enterprise tasks. Probably aims to fix this by focusing on verifiable reasoning paths rather than just probabilistic token prediction. This funding suggests investors are shifting their focus from raw scale to functional, reliable output for real-world production environments.

The Problem With Current LLMs

Let’s be real: using GPT-4 for anything requiring 100% accuracy is a gamble. I’ve spent the last six months building automation scripts, and the model still hallucinates function calls or misinterprets JSON schemas about 5% of the time. This is a massive headache when you are paying $20/month for a Plus subscription and expecting professional-grade reliability. Probably is targeting this exact pain point. Instead of just adding more parameters, they are building a framework that forces the model to verify its own logic against external data sources before outputting a final answer. It is essentially an architectural layer that sits on top of your existing API calls to act as a sanity check. If they can actually reduce error rates by even 50%, they have a massive product on their hands.

Deterministic vs Probabilistic AI

Current models are probabilistic—they guess the next word. Probably wants to move toward deterministic verification. Think of it like adding a compiler check to your Python code. If the AI’s reasoning chain doesn’t align with a known set of rules or retrieved data, the system flags it. It’s not just another wrapper; it’s a fundamental attempt to constrain model output to actual facts.

How $9 Million Changes Their Roadmap

Raising $9 million in this climate is no small feat, especially when you are competing against heavyweights like OpenAI and Anthropic. This capital allows Probably to hire top-tier research engineers who know how to optimize transformer architectures for lower latency. If they want to compete with the speed of Gemini 2.0, they need serious infrastructure. I expect to see them release a developer-first platform or an API that integrates directly with VS Code or existing agentic workflows. They aren’t trying to build a chat interface for the average user; they are building the infrastructure that makes AI actually usable for businesses that can’t afford to have their LLM lie to a customer.

Competing with OpenAI’s API

OpenAI charges $5.00 per million tokens for GPT-4o input. If Probably can offer a more reliable model at a similar price point, they will steal market share immediately. The cost of a bad AI hallucination in a customer support bot is significantly higher than the cost of the API call itself. Businesses are ready to pay for reliability.

What This Means For Your Daily Workflow

If you are a developer or a power user, you should care about this because it represents a shift in the ‘AI hype’ cycle. We are moving from the ‘look how cool this poem is’ phase to the ‘this code actually runs’ phase. If I can integrate a model that doesn’t hallucinate invalid library calls in my Node.js projects, it saves me hours of debugging per week. I’ve been testing Claude 3.5 Sonnet for coding, and while it is great, I still have to manually verify every single snippet. A model that guarantees reliability—or at least provides a confidence score—would be a total game-changer for my productivity.

The Rise of Agentic Reliability

We are heading toward agents that don’t just chat, but perform tasks. Reliability is the final boss of AI agents. If an agent can’t guarantee it will execute a database query correctly, it’s useless. Probably is betting that the company that solves this first wins the enterprise market.

Should You Wait for Probably?

Don’t hold your breath for a consumer product tomorrow. Probably is likely focused on API-first development. For now, if you need reliability, stick to using few-shot prompting techniques and RAG (Retrieval-Augmented Generation) with high-quality vector databases like Pinecone. I currently use a combination of GPT-4o for heavy lifting and a secondary validation script to check for syntax errors. It adds about 200ms to my response time, but it saves me from shipping broken code. If Probably releases a beta, I’ll be the first to test it against my current validation stack to see if it actually holds up to the marketing.

Current Best Practices

Until Probably hits the market, use ‘Chain of Thought’ prompting. Ask the model to ‘think step by step’ and verify its own work. It’s a cheap way to increase accuracy by 15-20% without needing a new, expensive model.

⭐ Pro Tips

• Use a tool like LangSmith to trace your AI calls; it costs $0 for personal use and helps identify exactly where your model is hallucinating.
• If you are paying $20/month for ChatGPT Plus, consider switching to the API if you are a light user; you might spend less than $5/month for the same access.
• Never trust an AI’s math without running a secondary Python script to verify the result; hallucinated numbers are the most common AI failure.

Frequently Asked Questions

Final Thoughts

Probably’s $9 million raise signals that the market is finally tired of ‘good enough’ AI that lies to your face. Reliability is the next frontier. I’m going to keep tracking their progress and will test their first public beta as soon as it drops. If you are tired of debugging AI-generated hallucinations, keep an eye on this space. Subscribe to my newsletter for updates on when their API becomes available for testing.