Why Lingusitic and non-Linguistic AI are Complementary

Splunk’s observability strategy has always put AI functionality at the centre. We have always recognised that, in order to make actionable sense of full fidelity data metric, event, log, and trace data streams, human cognition requires an automated assist which is precisely what AI brings to the table. As a result, throughout our observability portfolio, customers will find a variety of machine learning and pattern discovery algorithms being put to work, separating signals from noise, surfacing patterns of correlation, diagnosing root causes, and enabling remedial responses to incidents. AI, itself, is, of course, evolving at a rapid clip and with AI Assist, Splunk adds Generative or linguistic AI functionality to the mix. But what is linguistic AI and how does it relate to the non-linguistic or Foundational AI that Splunk has deployed in its products to date?

What the Philosophers Said About Language Games

In the late 1940s and early 1950s, philosophers, most notably Ludwig Wittgenstein in the UK and Wilfrid Sellars in the US,  arrived at a conception of language which likened it to a game. Words and sentences were like pieces on a game board while conversational interactions or extended texts and speeches were like the moves one makes across the board with the pieces available to the players. The analogy was buttressed by three further ideas:

• First, just like a game, the use of any language was governed by a set of normative rules. There was always a set of right or wrong ways to use words, respond in a conversation, or string sentences together. (Note that such a notion is only remotely related to the currently out of fashion notion that it is necessary to use proper grammar or spelling as dictated by some kind of explicit authority. Instead, the thinking here is that certain rules, explicit or implicit, must be followed, if one is going to, in fact, succeed in communicating.) 
• The second idea is that the rules themselves can, albeit often with some effort, be expressed in a language that may, in many ways, be different from the language which the rules govern. The second language is often called the meta-language. 
• Finally, the third idea is that there must be a way of ‘entering’ the language from the ‘real world’ at the beginning of many plays of the language games and ‘exiting’ from the language either as a direct consequence of the play or simply because the play has come to an end. Here is where one needs to keep one’s wits about him as the analogy between languages and games is exploited. In the case of games, one has pieces moved about in accordance with rules that are expressed in a language but the pieces are not (or are not entirely) linguistic items. So entry to and exit from the realm of language is not an issue. In the case of language games, however, there are rules expressed in a language governing linguistic items. Even if we maintain a sharp distinction between the language being governed and the meta-language (say the first one is English and the second one is Sanskrit), we still need to know how to move between language and the non-linguistic world. While the philosophers never quite settled this aspect of the story they were telling, there was a rough general agreement that certain words or sentences in the language being governed were, in one way or another, tied to real world events or data thus providing a way into the game while other words or sentences took the form of commands (do this!) and thus catapulted the language user back into the world of things. 

In summary, Wittgenstein, Sellars, and their philosophical friends left us with a picture of language use which starts with some real world events directly causing the occurrence of words or sentences. The articulation of these words or sentences constitutes the opening moves in a game that gets played through a sequence of language moves (governed by rules.) The process then ends with the final language move and actions undertaken in the real world.

The concept of language as a game played according to rules articulated in a meta-language originally emerged as an attempt to better understand how human beings use words, sentences, and other linguistic devices in their day to day activities. But, of course, the concept still makes sense when a natural language like English is replaced by a formally defined programming language. In fact, one could make the case that it makes even more sense in a programming language context. This is, of course, no accident. Programming language design ideas, in fact, the very idea of a formally defined programming language sprung from the same source as the philosophical framework just sketched - namely the discovery/invention of mathematical logic in the late 19th/early 20th century. Keep this all in mind because it will become important as our story moves forward. But now let’s go back to the philosophical framework itself.

As it turns out, this framework gives us a good way of understanding the history of commercial AI. (Once again, this is not accidental, but an account of the intricate historical linkages between philosophy of language and AI would fill volumes.) It is reasonably accurate to say that the entrepreneurs and technologists responsible for the first wave of commercial AI in the mid to late 1980s started out with something very much our philosophical framework in mind. They, too, viewed language as a game governed by rules expressed in a meta-language with well-demarcated entries to and exits from the world. Now, they worked out how to create rule-driven language manipulation software so the expert systems of the day were pretty much intralinguistic affairs. Sentences would be entered into the system and inferences rules -some based on pure logic, some based on domain specific expertise - would be applied - yielding other sentences as a result. Hence, everything remained ‘inside’ the language in which knowledge was expressed and the meta-language in which the inference rules were articulated. It should also be noted that the inference rules, while coded in the meta-language, had their origins outside of the system, either in logic or in pre-existent expertise.

It was recognised that, particularly for the purposes of robotic automation, one was going to want to extend these language bound systems out into the real world. Early attempts to do so, however, in a sense, cheated. Rather than treat signals coming from the world as being something external to language, following ideas initially staked out in David Marr’s Theory of Vision, they effectively treated the world itself as being composed of language-like items and used inference like processes to model the causal impact of external events on the linguistic game being played. Likewise the behavior of early robotic processing systems was governed by an inference-like extension of intra-linguistic processes. In other words, there was no language/world boundary to be overcome. Everything was, in the end, one big language game.

Commercial AI

The first wave of commercial AI peaked and crashed in 1989. While that crash cannot be attributed to the one big language game approach (it had a lot more to do with the fact that, in those days, the AI market depended heavily on defense spending that collapsed along with the Berlin Wall), there had been growing interest in alternate ways of linking language and world. 

Indeed, in the final years of that first wave, technologies began to appear that exploited another strain of academic research that eschewed symbolic, language-like processing for algorithms that owed more to statistical methodology and brain science than to logic. Fed with large, relatively unstructured sets of data, these algorithms searched for possible patterns, either through perusal of a pre-existent library or, more ambitiously, through direct generation from the data sets themselves. Some of the algorithms required learning, i.e., resulting pattern candidates were checked against the data, modified in the case of lack of fit, and selected or generated anew until some kind of equilibrium was reached. Other algorithms, instead, worked almost instantaneously. One sweep through the data would result in a pattern that might be modified as new data became available but without the idea of any ultimate equilibrium being reached. Some of the algorithms were designed to mimic a simplified model of biological neurons developed and signalled one another in the brain while others just captured statistical operations with no appeal to wetware metaphors. Despite the great variety, however, all of these algorithms bypassed any language-like mechanism or logical inference as they delivered results to the practitioners that might deploy them.With the crashing of the first wave of commercial AI, however, technologies based upon this non-linguistic approach were not able to establish an industrial beachhead and, for the most part, AI retreated to the universities. 

For the most part. There was one software market that, in fact, absorbed and retained much of what was delivered to industry during the cresting of the first wave: the market for application, infrastructure, and network management. Almost at the same time, the first wave of commercial AI began to gather in the mid 1980s, enterprises began to demand tools that would allow them to monitor, diagnose, and resolve incidents and problems that emerged during the operation of their increasingly complex and increasingly important IT systems that were independent of the systems and the vendors providing those systems. Even to state the requirement suggests the appropriateness of AI technology as a possible solution. After all, what enterprises were demanding were precisely technologies that, in a manner simulating, if not surpassing, human cognitive faculties would be able to perceive the occurrence of IT system incidents and problems, diagnose them, and assist in their resolution - almost the very definition of AI but, in this case, applied to communications and computation infrastructures.

The fit was so natural that, under the covers as it were, systems and network management vendors in this era - companies like IBM, HP, CA, and BMC - almost immediately incorporated recently commercialised AI into their offerings. Initially, the AI deployed was almost exclusively of the language/inferencing variety with the inferences being kicked off either through the direct human input of an incident or problem description (vide Help Desks) or via the ingestion of events represented by means of short highly structured alphanumeric strings which were effectively turned into sentences then run through an inference machine (vide Event Management platforms.)

The systems and network management market also provided a home for some non-linguistic AI technologies. A number of vendors, for example, added neural network-based products to their portfolios. However, no attempt was made to choreograph the interactions between linguistic and non-linguistic AI and, interestingly, while the linguistic AI was not explicitly advertised, much marketing hay was made regarding neural network deployment. Unfortunately, the non-linguistic AI functionality rapidly faded from the scene even as the linguistic AI delivered uncontested value ‘under the covers’.

The Second Wave

General commercial interest in AI remained muted for almost two decades but in the first half of the teens, academic success with a variant on neural networks (deep learning networks), an acceleration in the rate of digitalisation and the consequent explosion in the amount of data that required analysis and processing, and a growing interest in automation reawakened the demand for AI - both in the business world at large and in what had evolved out of the systems and network management market. Since achievement in the neural network field served as one of the catalysts for this reawakening, it is not surprising that the new generation of would-be AI deployers saw AI primarily through a non-linguistic lens.

Let us pause the history lesson and return to one of the issues regarding the purely linguistic AI technologies characteristic of the first wave of commercialisation. Remember that one of the short-comings of these technologies was their effective isolation from the real world. There was no automated way into them or out of them. Human beings would literally have to feed well formed sentences into them and hardwire links to some kind of automation apparatus after a language-based technology traversed its inferential paths. At the same time, the non-linguistic AI technologies were positioned as rivals to the linguistic technologies, a different paradigm, so to speak, so that robust relationships between the two styles of AI were hardly envisioned, let alone sought after in the commercial sphere. 

A quick glance at the wetware workings of human cognitive faculties, however, should have both dispelled the notion that these two approaches were rivals and provided a high level architecture for the way in which these two technologies might cooperate in silicon and software. Human cognition, in fact, functions at two levels. Data is ingested via sense organs and elements of the nervous system, at which point it is organised into patterns of spatial and temporal structure. It is only after these patterns are generated that a further layer of processing converts them into cognitive elements ready to serve as input into the inferential processes that are carried out in our cerebral cortex. Although we are only beginning to understand the incredibly complex nature of those inferential processes, they do, in one way or another, move the brain from sentence-like structure to sentence-like structure. In the end, the final structure is converted back into a neuron-readable pattern and sent down neural pathways to ultimately trigger motions that impact the world outside the body.

From this description, it should be clear how linguistic and non-linguistic AI technologies ought to be made to cooperate. The non-linguistic AI technologies serve as the source of patterns generated directly from the environment, providing the entry points, as it were, into the space of reasons that gets traversed by the linguistic technologies, leading ultimately to an exit back into the world, via action and automation. In summary, far from being rivals, particularly in the context of observability, linguistic AI and non-linguistic AI are complementary. Indeed, they depend upon one another in order to deliver the maximum value possible.