For now, NeoAngelo allows to check if a calculation is sollicitated or not, so if it is checked it simply adds, under the hood, the task to think in steps. It allows for better mathematical calculations, but even if it makes it a bit better, it still fails at providing the right result most of the time. OpenAI is great at linguistic tasks but terribly fails at very basic calculations, and we want to get rid of this credibility issue.
We are scheduling a demo to Wolfram | Alpha for a potential partnership. We would totally love if NeoAngelo was able to leverage calculations accuracy to the next level, passing from the 175 billion parameters of OpenAI to many trillions. Since Wolfram | Alpha also supports natural language as input for the query, apart from the formulation issues that may arise it is highly compatible with NeoAngelo.
One challenge is to make NeoAngelo and Wolfram | Alpha communicate in a great way together, and another is to determine when a call to their API is necessary and when it’s not. A result that does not contain information still requires an API call and extra delay in the reply to be returned, so it would quickly become very bad design to make a call for each query. Instead, NeoAngelo automatically detects mathematical symbols some character symbols as well as their text equivalents like + sums – delta * power of / quotient % = < > ≠ ≥ ≤ ¾ fractions cos sin √ root square radius circumference surface Pi lightyears, etc. but also many other units like millimeters, centimeters, meters, kilometers, lumens, minutes, hours, days, weeks, months, years, centuries, degrees, calories, etc. as well as comparison triggers like largest, biggest, smallest, bigger, more large, larger, largest, closer, closest, farther, etc., all currencies, all shapes, without mentioning other keywords like convers, distance, deep, altitude, density, buoyancy, radioactivity, weather, temperature, etc. where the user intent is to return the result of some kind of calculation.
The calculation query is first isolated from the rest of the prompt in order to make the shortest and cleanest call possible to Wolfram | Alpha, then NeoAngelo chains up the result to OpenAI, asking OpenAI to consider it before returning its own reply.
Let’s see what will come out of this great adventure. Some may argue that Wolfram | Alpha is not perfect but hey, it is already so much better at calculations than just a linguistic model.