Bicycles Versus Rockets of AI
Choose the smallest AI system that delivers the needed public benefit
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 6
- Confidence
- 97%
Bicycles Versus Rockets of AI uses a transportation analogy to right-size technology. Transportation includes bicycles and rockets, but a rocket is a disproportionate way to make a short ordinary journey. Hao applies the same reasoning to AI: begin with the human outcome, isolate the capability actually required, and ask whether a narrow model using a smaller curated data set can provide it. Compare that option with a large general-purpose model not only on output quality but also on compute, energy, emissions, water, labor, and who receives the benefit. The aim is not to reject advanced systems or deny their utility. It is to avoid treating maximum scale as the default path when a specialized approach can deliver substantial value with fewer external costs. The selected system should be the smallest one that reliably meets the defined need.
Origin
Hao compares AI with transportation: a category ranging from bicycles to rockets. She uses AlphaFold as an example of a specialized, curated-data system that she considers a bicycle of AI.
Core principles
- 01AI is a category of tools rather than one inevitable architecture
- 02Match system scale to the problem being solved
- 03Prefer narrow curated inputs when they can deliver the outcome
- 04Include resource consumption and affected people in performance
- 05Preserve useful capabilities without assuming maximum scale is necessary
How to run it
- 1
Name the human outcome
State the concrete benefit the system should produce for people. Avoid goals such as advancing AI or maximizing intelligence without a defined use.
Pro tip Express success as an observable improvement in a domain, not as a model-size milestone.
Watch out An ambiguous destination makes unnecessary scale difficult to challenge.
- 2
Bound the capability
Specify the task and operating context the model must handle. Distinguish that capability from a claim that one model should perform every kind of work.
Pro tip Use the transcript's jagged-capability idea to test strong and weak task regions separately.
- 3
Minimize the inputs
Identify the smallest curated data set and training process that could support the task. Compare specialized methods with a larger general-purpose model.
Pro tip Start with domain-relevant data before assuming the open internet is required.
Watch out Smaller does not remove the need to check data rights, quality, or bias.
- 4
Price the full footprint
Assess compute, energy, emissions, water, labor, and community impacts alongside accuracy and speed. Attribute uncertain impact claims rather than presenting them as settled facts.
Watch out A dramatic user benefit can obscure costs borne elsewhere.
- 5
Select proportionately
Choose the least resource-intensive approach that reliably reaches the required outcome. Reserve rocket-scale systems for tasks that genuinely need their additional capabilities.
Pro tip Document why each increase in scale earns its added cost.
- 6
Verify distributed benefit
Check who receives the useful outcome and who absorbs the costs after deployment. Redesign if the system helps a narrow group by shifting disproportionate burdens to others.
Watch out Utility for one user group is not proof of broad public benefit.
In the wild
Hao describes DeepMind's AlphaFold as a specialized system that predicts protein folding from amino-acid sequences. She argues that its smaller curated data domain requires fewer computational resources than large general-purpose models while supporting drug discovery and understanding of disease.
→ A bounded model is presented as capable of high scientific value without defaulting to maximum scale.
A public-service team needs to sort incoming forms into a fixed set of queues. It compares a large conversational model with a smaller classifier trained on consented, task-specific examples, then tests both against the same accuracy and safety threshold while measuring resource use.
→ The team selects the narrower system if it meets the threshold with lower resource demands.
Common mistakes
Starting with the largest model
Defaulting to a general-purpose system skips the question of whether the task needs its scale and footprint.
Optimizing technology for itself
Hao argues that technology development should be judged by human benefit rather than advancement for its own sake.
Measuring only model output
Accuracy and capability do not capture labor, energy, water, emissions, or how benefits and costs are distributed.
Is it for you?
Best for
Organizations choosing an AI architecture for a clearly defined scientific, operational, or public-interest task.
Not ideal for
Problems whose required capability, safety threshold, or affected population has not yet been defined.
From the transcript
“AI is like the word transportation. Transportation can literally refer to everything from a bicycle to a rocket.”
“Why don't we build more bicycles of AI?”
“you're using small curated data sets”
From the episode
AI Whistleblower: We Are Being Gaslit By The AI Companies! They’re Hiding The Truth About AI