The Ai cranes analogy - AIs really can be 100,000X more intelligent than a single human
I have been thinking about how to compare AI thinking to human thinking in terms of how much better it could be. Could it be 10X better, or 100X better - and what does that even mean.
After this discussion with codex - AI can be 100,000 times more intelligent than a human - imagine a whole datacenter dedicated to a single task - No one knows what this really means going forward.
“AI can be 100,000 times smarter than a human”
So when I was talking with Chat GPT 5.6 Sol on light thinking, about the latest advance, I wondered about the analogy with lifting things. My personal best dead lift in the gym when I was younger was 315lbs - and I put my back out so didn’t do that again -but how does that compare with lifting other things. That led to a conversation and this post - comparing cranes to thinking - it is surprisingly useful in terms of the compromises we have to make to lift heaver and heavier things, so it also gives you an idea that for intelligence far above human grade we might need entire data centers - not just a tiny bit of a data center that serves millions of requests.
Different cranes can lift unto 100,000X more than a human - but at every level the system is a bigger engineering feat and the flexibility is reduced.
Thinking vs physical lifting - an analogy/comparison.
I asked 5.6 Sol low about the physical lifting ai analogy I have had in my head for a while - and they really made it more concrete and more detailed. (One reference is that my max deadlift ever was 315lbs, and then I put my back out, and to compare with cranes forklift, tower crane, big crawler crane, etc - and the log improvements and their costs (I had forgot about gantry cranes)) - Here is the response
Thinking test: “I like this analogy because it separates unaided capability, amplification, cost, portability, and infrastructure. I’m going to use the visualization skill to build a simple logarithmic “lifting ladder,” then test where the analogy to intelligence holds and where it breaks.”
Answer
It is a surprisingly useful analogy—as long as we remember that “intelligence” is much less one-dimensional than lifting capacity.Here is your 315 lb deadlift placed on a logarithmic lifting scale. The approximate ladder is:
1× — You: 315 lb
10× — ordinary electric forklift: 3,000 lb
840× — road-going mobile crane: 120 metric tonnes
14,000× — giant crawler crane: 2,000 tonnes
141,000× — Taisun gantry-crane record: 20,133 tonnes
The specifications come from Toyota, Liebherr, SANY, and Guinness World Records.
What happens at each order of magnitude?
1×: unaided human
Your body was the complete system: power source; structural frame; controller; balance system; sensory feedback; repair system. And 315 lb is a substantial deadlift. But the whole machine weighed perhaps 150–220 lb, was self-assembling, could walk into the gym, and could subsequently drive home and discuss graph theory.
10×: forklift
A fairly ordinary warehouse forklift lifts about ten times your maximum. This is inexpensive enough for countless businesses, highly reusable, driven by one person, and able to move around independently. But it requires: a prepared floor; a suitable pallet; a compact, balanced load; fuel or charging; maintenance; an operator. This resembles ordinary AI augmentation quite well. A laptop or cloud model might make one person ten times faster at a suitably structured task—but only when the material has metaphorical “forklift slots.”
100–1,000×: mobile crane
A Liebherr LTM 1120 is rated for 120 tonnes, approximately 840 Stephen-deadlifts. But that headline capacity applies at a short three-metre radius. With its boom fully extended to 66 metres, Liebherr gives an example capacity of only nine tonnes. The crane needs outriggers, ballast, suitable ground, planning, transportation, and trained personnel. That is a particularly good AI analogy: Maximum benchmark capacity is not the same as capacity at full reach in messy conditions. An AI might display extraordinary ability on a self-contained proof problem while being far less reliable when the task requires months of context, ambiguous objectives, experimental judgment, institutional knowledge, or responsibility for consequences.
10,000×: crawler crane
SANY’s 2,000-tonne crawler crane reaches roughly 14,000× your deadlift.
Now we are no longer discussing a convenient tool that someone casually brings to work. Transport, assembly, ground preparation, engineering studies, exclusion zones, rigging crews, and safety systems become part of the machine. This resembles GPT‑5.6 Ultra with many agents more closely than a lone chatbot does. The effective “intelligent entity” may include: the foundation model; many parallel agents; enormous inference compute; search and software tools; retrieval systems; verification agents; human supervisors; an externally prepared problem; a carefully designed prompt; the accumulated mathematical literature.
Asking whether “the model alone” did it becomes rather like asking whether the crane’s engine alone lifted the reactor vessel. The engine was indispensable, but the operative system was much larger.
100,000×: enormous fixed infrastructure
Taisun’s record lift was about 44.4 million pounds—approximately 141,000 times your maximum. But Taisun is a huge shipyard gantry, not a machine you can drive to an arbitrary location. The load must be brought beneath it. It took an industrial site and enormous supporting infrastructure to make that one dimension of capability possible. And yet you remain superior at nearly every other physical activity from climbing stairs to carrying groceries to tying a shoelace. Taisun is 141,000× stronger than you according to one number, while being effectively helpless outside its prepared environment. That is an excellent warning about AI comparisons.
The cost curve really does change
10× - Buy a common machine 100× - Specialized vehicle and trained operator 1,000× - Ballast, outriggers, load planning and prepared ground 10,000× - Transport in pieces, assemble on site, engineering team 100,000× - Essentially build a specialized industrial facility
Each gain sacrifices something: portability; generality; setup time; operating freedom; affordability; tolerance of unprepared environments. The largest crane is not simply the best crane. A forklift is vastly more useful for putting a pallet on a warehouse shelf. That maps nicely onto AI. A hugely expensive multi-agent run might solve a research problem, but an inexpensive model answering thousands of routine questions could create much more total economic value.
The intelligence analogy captures at least five real features.
1. Tools can exceed the biological baseline enormously. Nothing in human biology permits a person to lift 20,000 tonnes. Nevertheless, human civilization can do it by organizing energy, materials, machines, and people. Likewise, human-plus-machine systems may reach cognitive capabilities unavailable to any unaided individual without implying that a model is a human mind, only bigger.
2. The relevant unit is often the complete system. For lifting, the system includes crane, foundation, rigging, operators and planning. For AI research, it includes model, compute, agents, tools, literature, prompts, verification and humans.
3. Peak capacity and general usefulness differ. A record crane has extraordinary peak capacity but narrow applicability. A human has modest strength but remarkable versatility. Similarly, solving one famous conjecture could demonstrate enormous peak reasoning capability without proving general scientific autonomy.
4. Reach reduces capacity. A crane can lift much more close to its centre than far away. AI also appears strongest when the problem is: clearly stated; symbolically representable; rapidly checkable; contained within available context; supported by a mature literature. Its “load chart” may fall sharply as ambiguity, duration, real-world experimentation, social consequences, and hidden information increase.
5. Safety margins matter. A crane is not operated at whatever load someone says it might theoretically survive. Engineers use load charts, inspections, redundancy and safety factors.With AI, verification plays the same role. “The model produced a plausible proof” is analogous to “the cable probably looks strong enough.” For important work, rated and independently verified capacity matters.
Where the analogy breaks. The major weakness is that lifting has a scalar measurement: pounds under specified conditions. Intelligence does not have a simple scalar unit. Furthermore, lifting power is relatively additive and mechanically testable. If the load rises safely, the lift occurred. A mathematical argument can look successful while containing one invisible fatal error. The analogy also understates agency and taste. Cranes do not select worthwhile buildings, invent architecture, or decide what ought to be constructed. The most important scientific intelligence may lie not in lifting the heaviest predefined intellectual load, but in recognizing which intellectual objects are worth moving.
My favorite formulation “Current frontier AI may be becoming a cognitive heavy-lifting industry, not simply an artificial strong person.”
That makes the analogy informative rather than silly. The most revealing question is no longer merely “How many times stronger is the machine?” It is:
What infrastructure, preparation, cost and loss of generality were required to obtain that amplification?