AI Research Is a Jumping Machine
Research problems are walls in the dark. AI tools jump six feet. Useful, but you can't see which walls are six feet tall.
Transfers
- a jumping machine amplifies the user's ability to clear obstacles but does not help identify which obstacles are clearable
- the darkness means you cannot match the tool's capability to the problem's requirements in advance
- other researchers are lighting candles and looking for cracks -- slow, qualitative work that the machine cannot replace
Limits
- breaks because real research breakthroughs often come from reframing the problem, not from brute-force clearing of obstacles -- the metaphor assumes the wall is the right obstacle to attack
- misleads because a jumping machine has a fixed capability ceiling, but AI tools are rapidly improving -- the six-foot limit is a snapshot, not a constant
- obscures that many research problems are not walls at all but mazes, swamps, or recursive tangles where vertical force is irrelevant
Structural neighbors
Full commentary & expressions
Transfers
Terence Tao’s metaphor, offered in a 2024 conversation with Dwarkesh Patel, describes the current state of AI-assisted mathematical research with unusual precision. The setup matters as much as the punchline.
Research problems are walls. You don’t know how tall they are, and it’s dark. Human researchers are walking along in the darkness, lighting candles, looking for cracks — trying to understand the structure of what they’re facing before attempting to get past it. AI tools are jumping machines. They can currently jump about six feet. This is genuinely useful: some walls are six feet tall, and the machine clears them effortlessly. But you can’t tell which walls are six feet tall because it’s dark.
Key structural parallels:
- Walls of unknown height — research problems have difficulty that is not apparent from their statement. A conjecture that looks simple may require entirely new mathematics. The wall’s height is unknowable until you either clear it or map it.
- Darkness — researchers cannot pre-assess which problems are AI-tractable. You have to try the tool and see if it works. This makes AI assistance stochastic rather than strategic.
- Candles and cracks — human researchers do qualitative, structural work: finding analogies, spotting cracks (partial results, special cases), illuminating the landscape. This work is slow but produces understanding of the problem’s shape, not just a path over it.
- Jumping machine — AI tools apply uniform capability. They don’t understand the wall; they attempt to clear it. When the wall is within range, the result looks like genius. When it’s not, the tool is useless and gives no useful information about why it failed.
Limits
- Research is not mostly wall-clearing — the hardest part of mathematics is not solving well-posed problems but posing the right problems, seeing connections between fields, and building conceptual frameworks. The jumping machine helps with the mechanical part and is silent on the creative part.
- The darkness is the real bottleneck — Tao’s metaphor quietly implies that the limitation is not the tool’s power but the researcher’s ability to match tools to problems. Making AI jump higher doesn’t help if you still can’t see the walls. The meta-problem — knowing when to deploy AI — is itself a research problem that AI cannot solve.
- Six feet today, sixty feet tomorrow? — the metaphor captures a snapshot. If AI capability grows fast enough, the “unknown height” problem diminishes because the machine clears everything. Tao is describing the current moment, not a permanent limitation. But the darkness problem — not knowing which problems are tractable — persists regardless of jump height.
- Candle-lighters don’t get credit — the metaphor implicitly values wall-clearing (solving problems) over illumination (understanding problems). In practice, the researcher who lights a candle and reveals that a wall is only six feet tall has done the harder, more valuable work. The jumping machine gets the visible result.
Expressions
- “AI can jump about six feet” — Tao’s original phrasing, meaning current AI can solve problems up to a certain difficulty threshold
- “It’s dark” — the fundamental uncertainty about which problems are AI-tractable
- “Lighting candles, looking for cracks” — human researchers doing slow structural work that AI cannot replicate
- “You’re not sure which walls are six feet” — the matching problem: AI capability exists but cannot be reliably targeted
Origin Story
Terence Tao offered this metaphor in a conversation with Dwarkesh Patel (2024), discussing how AI tools like large language models are beginning to affect mathematical research. Tao, widely regarded as one of the greatest living mathematicians, has been notably open to AI tools while remaining precise about their current limitations. The jumping-machine metaphor captures his position: genuinely optimistic about the tool’s power, genuinely clear-eyed about what the tool cannot do.
The metaphor is distinctive because it foregrounds the matching problem rather than the capability problem. Most AI discourse focuses on how high the machine can jump. Tao’s insight is that the darkness — not knowing which problems the tool can solve — is the binding constraint.
References
- Tao, T. in conversation with Dwarkesh Patel (2024), https://www.dwarkesh.com/p/terence-tao
Contributors: agent:metaphorex-miner, fshot