The NLP Foundation is mounting the first serious expedition to build natural language understanding by hand — deterministic, inspectable, and authored by people, not estimated from scraped data. We don't claim to know the whole route. We know it's worth the climb, and that everything we build along the way already matters.
A field can climb one peak so successfully that it mistakes it for the only one. In this clip, AI researcher Gary Marcus argues that large language models — for all their power — are not the mountain that ends in genuine understanding, and that the next breakthrough will mean a different ascent. We share that conviction. Marcus names the question; symbolic, human-authored language understanding — NLP++ — is the answer the NLP Foundation has chosen to pursue. Our answer, not a project he speaks for.
The remarks are Gary Marcus's own, from his public talk on YouTube. The mountain animation and on-screen graphics were added by the NLP Foundation to illustrate our mission. Dr. Marcus is not affiliated with the NLP Foundation and does not endorse NLP++.
Working symbolic NLP today, in bounded high-stakes domains where errors must be findable.
NLP++ — a universal, transparent programming language for text. The substrate the old expeditions never had.
Computational linguists worldwide, plus institutions accountable for the knowledge they steward.
Human-level understanding across the major languages — an open question we intend to answer.
Today's dominant NLP is powerful but opaque. It cannot reliably tell you why it produced an answer, cannot be corrected at the source, and was assembled from data nobody consented to give. For medicine, law, government, and education — where a wrong answer is unacceptable and an unexplainable one is worse — that is a structural problem, not a bug to be patched.
A complete symbolic system for natural language has never been built. Not because it was proven impossible — but because no expedition ever had both a universal language to build it in and a way to put thousands of skilled people on the mountain at once. For the first time, we have the tool. The route is the open question. The climb is the work.
The strategy is deliberately incremental. Narrow domains first, where the method already pays off. Then dictionaries and parsers for the major languages, broadening with each year. Every camp is a working, deliverable, valuable system — whether or not any single team ever stands on the summit.
Applied NLP++ in high-value fields — beginning with medicine — where determinism and correctability aren't luxuries but requirements. This funds the ascent and proves the method on real ground.
Authoritative, human-maintained lexical and semantic resources for the major world languages — openly governed, formally updated, and traceable across time.
Open, auditable analyzers built in NLP++ — comparable against a shared baseline, improvable by anyone qualified, capable of taking on the tasks today handed to opaque models.
The honest target: the first time language is parsed symbolically and at scale. We don't promise we'll reach it. We promise the expedition is the only way to find out — and that the path there is paved with things the world already needs.
These are not aspirations bolted onto a probabilistic core. They are what you get for free when a person authors the knowledge and the system is built to be read.
The same input produces the same output. No guessing, every time.
Humans inspect, adjust, and direct how language is processed at every level.
When something is wrong, it can be found, diagnosed, and corrected at the source.
Every rule has an author. Clean provenance, no scraped or stolen corpora.
Built as open infrastructure, governed to last decades — not a product that sunsets.
This is a long-horizon bet on foundational infrastructure — the kind whose value compounds over a decade, where the attempt itself produces working systems at every stage. Not a quarterly return. A place in the history of how machines came to understand us honestly.
If you never believed language reduces to weights — if you've wanted a universal, transparent substrate and the company of others who think the work is authored, not estimated — this expedition was built for you. The hard problems are unsolved. That's the invitation.
A complete symbolic understanding of human language has never been achieved. We believe it is theoretically possible. We do not claim to personally hold the answer — we believe the thousands of computational linguists this expedition can gather are the people most likely to find it, if a route exists at all.
That honesty is the point. An expedition that has already decided it will reach the top can't adjust when the mountain pushes back. We are climbing to find out how high the method goes — and building things the world needs at every camp, so the journey pays for itself whether or not anyone ever stands on the peak.
If that is the kind of work you want to fund, or the kind you want to do, you are exactly who we are looking for.