Differential Technological Progress
Accelerate protective capabilities before broadly powerful technologies mature
- Difficulty
- Expert
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 93%
Differential Technological Progress treats the direction and sequence of innovation as choices rather than assuming every capability should advance equally fast. Begin by separating technologies that increase offensive power or uncontrolled capability from those that improve defense, detection, interpretability, containment, or safety. Then deliberately accelerate the protective side. In biosecurity, pathogen design can strengthen offense, while far-UVC lighting, broad pathogen detection, and better protective equipment strengthen defense. In AI, raw capability can make systems more powerful, while interpretability and methods for detecting deception can make their behavior more understandable and controllable. The mechanism is a growing safety margin: society gains tools to prevent, detect, or contain harm before dangerous capabilities become cheap or widespread. Governance and deployment choices complement the research by limiting avoidable exposure during the transition.
Origin
Will MacAskill applied differential technological progress to biosecurity and AI safety during his discussion with Tim Ferriss.
Core principles
- 01Technologies differ in whether they strengthen offense or defense
- 02Capability growth and safety growth need not advance at the same rate
- 03Protective infrastructure should mature before dangerous capability is widely available
- 04Understanding a system improves the ability to predict and control it
How to run it
- 1
Classify the capability
Assess whether a proposed advance primarily enables attack, uncontrolled power, defense, detection, containment, or understanding. Record mixed effects rather than forcing a false binary.
Pro tip Compare the cost of causing harm with the cost of defending against it.
Watch out A beneficial general capability can still create a dangerous offensive advantage.
- 2
Find the safety gap
Identify where dangerous capability is advancing faster than the tools required to understand, detect, or contain it. Focus on gaps that could become difficult to close after deployment.
Pro tip Look for infrastructure that protects against many variants of a threat.
Watch out Do not confuse a speculative safeguard with demonstrated protection.
- 3
Accelerate defense
Direct research, funding, standards, and talent toward the protective capability. Where possible, make the defense cheap, widespread, and built into ordinary infrastructure.
Pro tip Favor measures with present-day benefits as well as protection against extreme events.
Watch out Safety work that cannot keep pace with capability growth may provide false reassurance.
- 4
Control deployment
Use containment and governance to avoid exposing powerful systems before the safety margin is adequate. Reassess the balance as both sides develop.
Pro tip Prefer constrained systems when they can deliver useful answers without broad agency or access.
Watch out Once a dangerous capability is cheap and distributed, sequencing options shrink sharply.
In the wild
Instead of focusing only on the growing ability to create pathogens, researchers test far-UVC lighting that may sterilize occupied rooms safely. If evidence supports it, building standards could make disease control part of ordinary lighting infrastructure while broad detection and protective equipment provide additional layers.
→ Defensive capacity could become widespread before a more severe natural or engineered pandemic occurs.
An AI program funds interpretability and deception testing while constraining access and agency for increasingly capable models. The objective is to understand and predict behavior before deploying systems able to act broadly in the world.
→ Safety and understanding advance relative to raw autonomous capability.
Common mistakes
Treating all progress as interchangeable
Raw capability and protective capability can alter risk in opposite directions. Funding them without distinction can widen the safety gap.
Waiting for the threat to mature
Defensive research and infrastructure may take years to validate and deploy. Starting only after a dangerous capability spreads forfeits the sequencing advantage.
Is it for you?
Best for
It is best for high-impact fields where general capability can create both enormous benefits and severe misuse or control risks.
Not ideal for
It is not ideal when a technology's offensive and defensive effects cannot yet be distinguished with useful confidence.
From the transcript
“there's certain technology that has kind of offensive advantage so the ability to design new pathogens there's certain technology that's like has a defensive Advantage…”
“one of the things we're doing is really trying to develop and speed speed up kind of defensive technology”
“there are some things that are helpful in making sure that AI is safe like understanding what's under the herd of these models”
From the episode
#612: Will MacAskill of Effective Altruism Fame — The Value of Longtermism, Tools for Beating Stress and Overwhelm, AI Scenarios, High-Impact Books, and How to Save the World and Be an Agent of Change