A Language That Knows the World: Stephen Wolfram on Computable Knowledge

Wolfram begins with the two products most people know that come from the Wolfram Language: Mathematica, first released in 1988, and Wolfram Alpha. Mathematica is an instance of the language used for technical computation, where users type small pieces of computational language and get results. The language is symbolic, meaning that typing an undefined name like `x` does not trigger an error but is simply treated as a symbol.

The difference from a language like Python, Wolfram says, is the raw material. In a traditional language it is whatever computers intrinsically do, while the Wolfram Language aims to be an abstract language about things in the world or things we can imagine. A symbol could be the city of Boston or the trajectory of a spacecraft, and that ability to work computationally with such things is what he calls powerful.

He traces the idea to SMP, his first computer language, and to his background as a physicist looking for fundamental components. The result was transformation rules for symbolic expressions as the underlying stuff of computation. He calls the Wolfram Language by far the highest-level computer language in existence, with about 6,000 primitive functions that range from Boolean conversion and cloud requests to discrete wavelet data and user interface elements.

To show what that means, Wolfram runs a few live demos. `ImageIdentify` classifies a webcam image: the first try, he laughs, suggests he looks like a plunger because of a large object behind him, and a second attempt gives an 89% probability of a person. These are examples of machine learning capabilities that sit inside the language as ordinary functions.

Another demo finds the ten volcanoes nearest to his location and plots them on a map. For him the point is that the language already knows about all the volcanoes in the world, how to compute which are nearest, and all the maps. He calls this a full-scale computational language and says this is a fundamentally different idea of a language.

Lex asks how this fits into AI history, noting that expert systems of the 1980s and 1990s did not pan out. Wolfram thinks those efforts were too modest because they looked at narrow areas. Problems like natural language understanding require broad knowledge of the world, so it can be easier to do the whole thing than a piece of it.

On learning versus encoding, Wolfram says civilization has already surveyed the volcanoes and worked out algorithms for many problems, and starting from scratch with neural nets would deny those achievements. His team therefore captured the knowledge of the world in computable form. Image identification, long hard for computers, is a useful module to add, and he finds the interplay between symbolic knowledge and statistical methods the most interesting direction.

He gives natural language understanding as an example of a loop between the two. Wolfram defines it as turning a query into computational language, and the team has relied on algorithmic and human-curation methods rather than learning. For natural language processing, such as picking out all the cities in a text, they use modern machine learning, and what each method picks up improves the other.

Lex wonders why the Wolfram Language is not more popular. Wolfram says ideas are absorbed slowly, and that among very high-end R&D people usage is common, while there has been surprisingly little trickle down. He partly blames himself, since his company emphasizes building the best technical tower over the commercial side of getting the product out.

Asked about opening everything up GitHub-style, Wolfram says this does not work when you care about the integrity and quality of the knowledge you are building. It requires leadership and a coherent vision over a long time, plus mundane work such as making things run in different cloud environments and keeping the interface smooth. Because his company is private and has no investors telling it what to do, it can offer a free Wolfram Engine for developers and site licenses at essentially all major US universities. Wolfram Alpha's main website is free as well.

Wolfram explains Wolfram Alpha as a system where you ask a question in natural language and get a report with the answer. The words are converted into Wolfram Language, which draws on the Wolfram knowledge base, a decades-long collection effort that continues to receive data "every second". Lex calls the knowledge base one of the most incredible parts.

Wolfram says he had assumed building this would require solving general AI, so he put it off. Then his work on *A New Kind of Science* produced the principle of computational equivalence, which says there is no bright line between intelligence and mere computation, and he decided he had to eat his own dog food. He remembers taking the early team to a big reference library and declaring they would ingest everything in it, a finite if daunting goal.

Methodologically, he did not start from a grand theory but implemented hundreds of areas one by one, relying on the world's experts who already used the products. His principle was that any area lacking expert input would not be right, with the goal of making any question answerable from general civilizational knowledge automatic. He recalls Marvin Minsky typing random medical, chemistry and astronomy queries into a pre-release version and, after a few minutes, saying "oh my god it actually works".

Wolfram says the advantages he had were unfair ones: an existing high-level symbolic language, practical experience building large systems, and the confidence not to give up. He credits optimism as well as ego. If someone told him a project would take 30 years, it would be hard to sell him, because he can usually see something happening within a few years.

The challenge is that such projects have infinite tails, and the question is whether you drown in them. He mentions that he has just begun work on a fundamental theory of physics, a bet that he can do that alongside his work on the language. He also names the opposite trap, deciding never to do a project smaller than the biggest one so far, when the true depth and significance of projects is very hard to know in advance.

Lex asks whether all knowledge could eventually be made computable, as opposed to just searchable. Wolfram says general knowledge questions are in good shape, and the team is now moving into computational contracts: expressing legalese, or what a self-driving car should do, in computational language. He argues that automating more of the world means encoding more ethics in a form computers can handle.

Drawing on his recent involvement in automated content selection at companies like Facebook, Google and Twitter, Wolfram says it would be bad for the future of our species to have one AI ethics module govern everything. He prefers breaking it up, with people picking among modules, which Lex likens to systems with different ideological brands. Wolfram sees content selection as probably the easiest AI ethics case because each person can pick for themselves, whereas matters like public health or central bank policy affect others and become real political philosophy.

A computational contract, he explains, is a contract whose conditions are all represented computationally so it can be executed automatically, something he sees as the surely inevitable future of legal contracts. He notes that Leibniz proposed something similar in the 1600s, but his mechanical calculator never worked properly, so he was about 300 years too early.

Wolfram describes his goal of a symbolic discourse language, able to express the actionable parts of a natural-language conversation in precise, computable form. He thinks it is within reach in a limited number of years for the kinds of things in typical legal contracts. The language already knows blueberries and their nutrition, but it has no representation of "I want to eat them", which would be needed for a complete version.

Lex connects this to the Turing test. Wolfram says the test is an okay one, but bots built on the Wolfram Alpha API lose immediately because a few questions about obscure knowledge drop them. Lex argues that Wolfram Alpha legitimately targets the intent of the Turing test, and Wolfram adds that Turing himself estimated the work needed to make the Encyclopaedia Britannica computational. Wolfram says the real effort was more efficient than Turing feared, thanks to layers of automation Turing could not have imagined.