The LLM-Plus-Computational-Language Pairing
Let the language model talk and the computational language be right.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 90%
Wolfram frames LLMs and computational language as two distinct branches of intelligence that fit together. An LLM works the way humans immediately think — it reads a billion web pages, learns that the most common next word after 'the cat sat on the' is 'mat', and produces plausible sentences that may or may not have anything to do with reality. Computational language is the other branch: the formalized knowledge tradition that led to modern science, capable of building whole towers of consequences. The pairing gives the LLM a linguistic interface and lets it call the computational language as a tool. Wolfram Alpha, shipped in 2009, had already solved the one-sentence natural-language-to-computation translation problem, which is exactly what makes the handoff work.
Origin
Wolfram Research built an early version of this pairing with OpenAI roughly a year before this April 2024 conversation. The pieces had been in place far longer: Wolfram Alpha launched in 2009 with natural language understanding that could take one English sentence at a time, translate it into computational language, and compute an answer through potentially many steps from curated knowledge.
Core principles
- 01An LLM produces plausible sentences; plausible is not the same as true.
- 02LLMs mimic what humans do immediately and naturally; computational language does what humans found hard.
- 03The two are complementary branches, not competitors.
- 04The LLM is the linguistic interface; the computational language is the tool it calls.
- 05A solid answer is computed from curated knowledge through many steps, not predicted a word at a time.
- 06Humans don't immediately know everything either — but with tools they can get a long way. Same for AIs.
How to run it
- 1
Diagnose which branch the task belongs to
Split the work into the things humans do quickly and naturally (phrasing, pattern-matching, conversation) versus things humans found really hard (solve this math problem, figure out this thing in science, simulate what happens in the natural world). These are different branches requiring different machinery.
Watch out The untrained brain doesn't do the second category well — and neither does an untooled LLM.
- 2
Use the LLM as the linguistic interface
Let the language model do what it's built for: making up pieces of language, producing sentences that are typical of what you might read. This is the front door, not the engine.
Watch out Those sentences sound like they make sense and might or might not actually have anything to do with reality in the world.
- 3
Translate natural language into computational language
Take the natural language the model produces and convert it into precise computational form. Wolfram Alpha solved this for one sentence at a time back in 2009 — translating an English sentence into computational language.
Pro tip If the model can formulate what it's talking about clearly enough to translate, the handoff succeeds.
- 4
Compute the actual answer
Run the computation — potentially many many steps — against curated knowledge, building up whole towers of consequences that next-word prediction can never produce.
- 5
Return the solid answer
Hand back a result that was computed from curated knowledge, not one that was predicted because it was statistically typical. This is the thing that's actually true, as opposed to the thing that reads well.
Pro tip Treat this the way you treat any tool: humans don't immediately know everything, but with tools they get a long way. AIs get to do the same thing.
In the wild
Wolfram walks through the LLM mechanism. Given 'the cat sat on the', the model has read a billion web pages and set itself up knowing the most common next word is 'mat' — assigning it maybe 99% probability, with 'floor' at 1% or 5%. It usually picks the most likely word and sometimes doesn't, one word at a time.
→ The surprise is that this word-at-a-time procedure gives you something that seems like a reasonable English sentence. But the output is plausible-typical, not true — which is exactly the gap the computational language tool fills.
About a year before the interview, Wolfram Research built an early version of the pairing with OpenAI: the language model tries to make up words, and when it can formulate what it's talking about, it gets to use Wolfram's computational language as a tool.
→ Wolfram calls this the typical mode of interaction and predicts more and more will come out along these lines — a linguistic interface from the LLM, with the computational language doing the work of figuring out what's actually true.
Common mistakes
Treating fluent output as computed output
An LLM's sentences sound like they make sense and are typical of what you might read — that's the whole objective it was trained on. Whether they have anything to do with reality is a separate question the architecture doesn't address. Accepting plausibility as truth is the core failure mode.
Expecting one branch to do both jobs
Neural nets are going in one direction and formalized computation in another. They're complementary, not substitutes. Waiting for the language model to become accurate on its own ignores that the accuracy branch is a different 300-year tradition with different machinery.
Is it for you?
Best for
Anyone building or using AI systems where the answer has to be correct, not merely fluent — analysis, math, science, factual lookup.
Not ideal for
Purely creative or conversational work where plausibility is the goal, and any request that can't be formulated in computational terms.
From the transcript
“those sentences might make they sort of sound like they make sense they're kind of typical of what you might read they might or might…”
“the typical mode of interaction is there sort of a linguistic interface provided by things like llms and that that using our computational language as…”
“just as humans don't necessarily immediately know everything they but with tools they can get a long way”
“we have this computational view of the world which allows one to do precise computations and build up these kind of whole towers of consequences”
From the episode
Stephen Wolfram: AI, ChatGPT, and the Computational Nature of Reality
Stephen Wolfram