Serial entrepreneur Wayne Chang takes the wraps off an AI reasoning engine called Reasoner, which he claims can produce much more accurate and explainable results than large language models like OpenAI’s o1 series—and at a much lower cost.
The AI ​​industry is pushing hard to build thinking capabilities into technology, partly to get closer to the holy grail of human or superhuman AI, and partly just to overcome the inaccuracies that plague today’s LLMs. The so-called hallucinations of generative artificial intelligence are the main factor in holding back the development of companies, as well as the inability to explain how LLM reaches its conclusions.
Reasoner aims to solve these problems using neurosymbolic artificial intelligence — a combination of neural networks (the technology that underpins generative artificial intelligence) and more traditional symbolic artificial intelligence, which is based on fixed rules, logic, and human mappings of relationships between things.
Chang has a long history in technology, starting with the creation of the file-sharing service i2hub in 2004. In 2011, he co-founded Crashlytics, a ubiquitous mobile crash reporting tool that he acquired Twitterwhere he became Director of Consumer Product Strategy. He then co-founded the artificial intelligence-based accounting firm Digits (which Google bought together with Crashlytics in 2017), and then last year founded Patented.ai — an intellectual property-focused AI tool that, it now turns out, also served as a pilot implementation of the Reasoner engine.
High stakes AI
Patented.ai offers the ability to automatically search patent documentation and source code to spot potential patent infringement cases and identify potentially patentable innovations. Given the high financial stakes of patent cases and the extremely painstaking nature of determining whether infringement has occurred, there are clear opportunities for anyone who can automate the process – but also huge risks if the system goes wrong.
In an exclusive interview for WealthChang said Patented.ai’s early reliance on PhDs alone proved unhelpful — lawyers who played with the system immediately spotted flaws in its output and rejected it. The company also tried other common techniques such as augmented retrieval generation, which relies on external data sources to improve LLM results (Google uses RAG for its AI search results), but this also did not provide the required level of confidence.
This prompted a change in approach that resulted in the development of Reasoner. “We haven’t really started building an inference engine,” Chang says. “That was not our mission at all.”
Reasoner does use LLMs to help interpret language in texts—Chang says he’s agnostic about the model he uses—but the core concept in Reasoner is an adaptive dynamic knowledge graph.
Knowledge graphs are widely used in engineering. For more than a decade, Facebook’s Knowledge Graph has provided a framework for building relationships between people, while Google has given Search the ability to answer basic factual questions. These repositories of established knowledge are clearly useful for providing correct answers to queries—IBM Danger-the winning Watson AI is built on knowledge graphs—but they generally need to be manually updated to add new facts or edit relationships that have changed. The more complex the knowledge graph, the more work involved.
Chang claims that Reasoner removes the need for manual updates, instead offering the ability to automatically create accurate knowledge graphs based on unstructured text entered into the system, and then automatically reconfigure these knowledge graphs as information is added or changed. (It is worth noting that Microsoft earlier this year he discovered GraphRAG, an attempt to use LLM-generated knowledge graphs to improve RAG scores.)
In other words, you can input a bunch of legal documents, and Reasoner will then interpret them to build a knowledge graph containing the concepts in the documents and the relationships between them—with “full traceability” so that it’s easy for a human to verify that those facts are indeed an accurate representation of what is in the documents. This is where the concept becomes useful far beyond the realm of patent litigation.
In a demonstration for WealthChang demonstrated how Reasoner can ingest dozens of different OpenAI legal documents (from user and developer agreements to brand guidelines and cookie notices) and map their interdependencies. In the demo, this made it possible to provide concise and detailed answers to a question about how a user could take advantage of the differences between OpenAI’s US and European terms of service to “avoid liability for harmful AI results.” Each step in the reasoning was explained—the logical steps were understandable even to the non-technical eye—and the Reasoner then suggested follow-up questions about the impacts of the problem and how they could be mitigated.
Chang says Reasoner could also be used in a variety of other applications, from pharmaceuticals and advanced materials to security and intelligence. As such, it claims it can outbid various other AI startups, such as Hebbia (a document search company that raised a $130 million Series B in July) and Sakana (an Nvidia-backed science discovery company that raised $214 million in September Serie A round).
The price of reason
But in terms of thinking ability, the big beast at the moment is OpenAI and its o1 series of models, which take a very different approach to the problem. Instead of deviating from the pure LLM paradigm, o1 models use “chain of thought” thinking. combined with search, methodically working through a series of steps to arrive at a more thoughtful answer than OpenAI’s GPT models could previously achieve.
O1 models generally give more accurate answers than their predecessors, but Chang claims that Reasoner’s result is even more accurate. There aren’t many benchmarks for Reasoner—Reasoner may release its own early next year—but, based on DocBench and Google’s recently released set of Frames benchmarks, Chang said Reasoner achieved over 90% accuracy where o1 couldn’t crack 80 %. This result could not be independently verified at the time of publication.
He also said that Reasoner’s approach allowed for far lower costs. OpenAI charges $15 per million tokens (the basic unit of AI data, equivalent to roughly 1.5 words) of input and $60 per million output tokens, while a million input tokens cost Reasoner 8 cents and a million output tokens only 30 cents. “We haven’t finalized how we want to price it,” Chang said, adding that Reasoner’s “structural cost advantage” would allow customers to be billed by result or by confirmed discovery.
Chang’s claims are certainly big, but Reasoner’s team is small—it has about a dozen employees, mostly in the US. So far, the company has only had a $4.5 million pre-seed round, which took place last year with investors including the likes of Baseline Ventures founder Steve Anderson, Y Combinator MD Ali Rowghani and Operator Collective founder and CEO Mallun Yen. “I’ve been very fortunate to have had a few successes in my history, so I wasn’t too worried about funding,” Chang said. But the entrepreneur expects to hire more staff soon as Reasoner grows.
Chang said Reasoner—which took $1.8 million in bookings in the third quarter of this year—will publicly release its benchmarks and demo in the first quarter of 2025, allowing people to upload their own datasets and test the company’s claims. The company will also release a software development kit to allow others to incorporate the Reasoner engine into their apps and AI agents. (Chang says the engine is lightweight enough to run on even the latest iPhones and Android devices, without requiring an Internet connection.)
“We want to make sure that we release it in a way that we start building that trust and credibility right away,” Chang said.
