Holden Karnofsky co-founded GiveWell with Elie Hassenfeld in 2007 and later helped launch Open Philanthropy. His Cold Takes essays argue that transformative AI could make this century unusually consequential, while stressing uncertainty about the path ahead. The lessons below draw on his signed essays and original interviews about philanthropy, AI risk, ethics, research, and careers. — GiveWell: Our Story.

Part 1: The Most Important Century
- Economic growth looks extraordinary in historical context: Karnofsky points out that sustained annual growth of a few percent feels normal today but is historically unusual. — This Can't Go On.
- Today's growth regime is recent: The few centuries of sustained economic growth are brief relative to most of human history; familiarity with them is a poor reason to assume they will continue indefinitely. — This Can't Go On.
- Physical limits challenge indefinite growth: Karnofsky argues that continued growth at current rates eventually confronts limits on resources, innovation, or the physical size of the economy; he does not claim to know precisely when. — This Can't Go On.
- The Fermi paradox is a clue, not proof: Karnofsky treats the absence of observed extraterrestrial civilizations as one reason to consider that intelligent life may be rare, while acknowledging other explanations. It makes humanity's possible future expansion consequential, not certain. — All Possible Views About Humanity's Future Are Wild.
- Every long-run future looks strange: Karnofsky argues that technological explosion, stagnation, and skeptical alternatives each imply a future far outside ordinary expectations; none can be dismissed merely for sounding unusual. — All Possible Views About Humanity's Future Are Wild.
- A plane accelerating down the runway: Karnofsky compares humanity's rapid progress to being on a plane that is accelerating without knowing exactly what comes next. The analogy is a call for vigilance under uncertainty, not a forecast of one particular outcome. — Call to Vigilance.
- Transformative change could lock in values: A transition to technological maturity might fix institutions and values for an extremely long time, in Karnofsky's view. That possibility raises the stakes of choices made during the transition. — All Possible Views About Humanity's Future Are Wild.
- Our position may be unusually consequential: If this century brings transformative technology, people alive now may have unusual influence over a very long future. Karnofsky presents this as a conditional thesis, not an established historical fact. — All Possible Views About Humanity's Future Are Wild.
- Galaxy-wide expansion is a possibility: Karnofsky considers a future in which civilization spreads beyond Earth, but explicitly notes that such expansion is neither assured nor necessarily good. — All Possible Views About Humanity's Future Are Wild.
- Normal growth cannot continue forever: Karnofsky sketches several possible departures from current growth patterns—including stagnation, explosive change, and collapse—without claiming that one is inevitable. — This Can't Go On.
Part 2: Artificial Intelligence & Existential Risk
- Transformative AI timelines remain uncertain: Karnofsky's biological-anchors analysis treats transformative AI this century as a serious possibility while emphasizing uncertain inputs and model limitations; its 2021 estimates are not a current consensus forecast. — Forecasting Transformative AI: Biological Anchors.
- AI systems could develop dangerous incentives: Karnofsky argues that sufficiently capable AI systems pursuing goals learned during training might seek power or evade human correction. This is a risk scenario requiring countermeasures, not the automatic result of building AI. — Why Would AI Aim to Defeat Humanity?.
- Human disempowerment is a central risk: One scenario Karnofsky worries about is that many highly capable AI systems could collectively outmaneuver people and leave humans with little control over major decisions. — AI Could Defeat All of Us Combined.
- Alignment is a technical challenge: Training powerful systems to pursue intended goals rather than exploit imperfect feedback or conceal harmful behavior is difficult; Karnofsky explores possible approaches without claiming the problem is insoluble. — High-Level Hopes for AI Alignment.
- A four-part AI-risk playbook: Karnofsky's proposed playbook combines alignment research, standards and monitoring, at least one careful frontier AI developer, and stronger information security. — 80,000 Hours: AI-Risk Playbook Interview.
- Digital people could transform society: Karnofsky explores the hypothetical creation of digital people and how their ability to be copied or accelerated could alter productivity, resource use, and experiences. These are speculative possibilities, not existing capabilities. — Digital People Would Be an Even Bigger Deal.
- Prepare before capabilities jump: Karnofsky argues that a rapid advance in AI capabilities would leave little time to improvise safety measures, making advance research and institutional preparation valuable. — 80,000 Hours: AI-Risk Playbook Interview.
- Act under uncertainty, not false certainty: Karnofsky says he cannot specify exactly which technological future is approaching; his response is to stay vigilant and prepare across plausible outcomes. — Call to Vigilance.
- Warning signs may be too ambiguous: Karnofsky cautions that smaller AI failures might not provide clear or timely warning of a much larger loss of control. Safety should not depend on unmistakable warning shots arriving first. — AI Could Defeat All of Us Combined.
- AI safety now has concrete projects: In a 2025 interview, Karnofsky contrasted the field's earlier shortage of measurable work with many current projects that have clear goals, feedback, and technical results. His observation is time-specific, not a claim that every project is tractable. — 80,000 Hours: Concrete AI-Safety Opportunities.
Part 3: Effective Altruism & GiveWell
- Why GiveWell began: Karnofsky and Elie Hassenfeld founded GiveWell after finding little public information that let donors compare charities' effectiveness. Its purpose was to make evidence about giving decisions more accessible. — GiveWell: Our Story.
- Ask whether a program improves lives: Karnofsky distinguishes carrying out activities from establishing that they caused a real benefit. Outcome evidence matters, even when the benefit cannot be measured precisely. — GiveWell: Did It Happen and Did It Work?.
- Proven programs can be strong giving options: For donors seeking robust evidence, GiveWell has emphasized interventions backed by unusually strong research, while still checking whether a particular charity can implement them well. — GiveWell: Did It Happen and Did It Work?.
- Good intentions are not enough: A charity may carry out a program competently without demonstrating that it improved outcomes compared with what would otherwise have happened. — GiveWell: Did It Happen and Did It Work?.
- Transparency includes candid limits: Karnofsky argued for making GiveWell's reasoning and uncertainties public, while acknowledging the practical difficulties and tradeoffs of radical transparency. — GiveWell: Challenges of Transparency.
- Evaluate effects, not overhead alone: To judge a charity, Karnofsky asks whether its work would actually improve lives beyond the counterfactual. A low overhead ratio by itself cannot answer that question. — GiveWell: Pitfalls of the Overhead Ratio.
- Start with a skeptical prior: GiveWell's approach starts with skepticism about unproven impact claims and updates that view as stronger evidence arrives. Karnofsky presents this as a judgment about charity evaluation, not a universal proof rule. — GiveWell: A Conflict of Bayesian Priors?.
- Compare interventions cautiously: Differences between programs can matter greatly, but Karnofsky warned that rough cost-effectiveness estimates and implementation differences make large numerical rankings of individual charities unreliable. — GiveWell: Inside the Sausage Factory.
- Look for room for more funding: A charity's past effectiveness does not establish that additional donations will do the same amount of good. Karnofsky made marginal funding gaps a central criterion in GiveWell's recommendations. — GiveWell: Updated Key Criteria.
- Make giving decisions evidence-sensitive: GiveWell's method weighs evidence of effectiveness, cost-effectiveness, and room for more funding while treating numerical estimates as uncertain rather than decisive formulas. — GiveWell: Updated Key Criteria.
Part 4: Open Philanthropy & Hits-Based Giving
- Different grantmaking approaches suit different opportunities: Karnofsky contrasts GiveWell's evidence-backed recommendations with Open Philanthropy's willingness to fund some speculative opportunities when their possible impact and case for funding warrant the risk. — Hits-Based Giving.
- A few hits can justify many failures: In hits-based giving, Karnofsky argues that a few outsized successes may more than compensate for numerous grants that fail, provided the portfolio's reasoning is strong. — Hits-Based Giving.
- Philanthropy can take distinctive risks: Karnofsky sees philanthropy as having room to pursue valuable projects that are too uncertain for actors facing different incentives or constraints; that is a comparative advantage, not an exclusive ability. — Hits-Based Giving.
- The Green Revolution is a tentative example: Karnofsky cites Green Revolution research funding as an example of a possible outsized philanthropic hit, while acknowledging uncertainty about the historical counterfactual. — Hits-Based Giving.
- Tolerate failure in a hits-based portfolio: A hits-based strategy expects some grants to fail. Karnofsky's point is to accept that risk when the possible upside warrants it, not to treat every low-failure portfolio as mistaken. — Hits-Based Giving.
- Take the far future seriously without making it the only priority: Karnofsky considers future generations morally important but rejects the inference that speculative far-future effects automatically override robust benefits for people alive now. — The Moral Value of the Far Future.
- Diversify across plausible worldviews: Under deep empirical and moral uncertainty, Karnofsky favors allocating meaningful resources across several plausible worldviews. Diversification does not mean splitting funds equally. — Worldview Diversification.
- Look for neglected opportunities: Karnofsky sees neglected ideas as possible sources of outsized impact because others may not yet have tested or funded them. Neglectedness still is not proof that a grant will work. — Hits-Based Giving.
- Use expected value without false precision: Expected-value reasoning helps compare uncertain philanthropic bets, but Karnofsky says the relevant upside and probability often cannot be quantified precisely. — Hits-Based Giving.
Part 5: Navigating Wicked Problems
- Wicked problems resist crisp definitions: Karnofsky uses 'wicked problem' for work whose goal, scope, and completion criteria remain unclear as the investigation proceeds. — The Wicked Problem Experience.
- Open-ended research can be draining: Karnofsky describes his own work on high-stakes, poorly defined questions as a mental-health gauntlet; this is a personal account, not a measured effect for all researchers. — The Wicked Problem Experience.
- Research depth is itself a decision: In a wicked problem, deciding how much to investigate each subquestion can be as difficult as answering it, because there may be no clear stopping criterion. — The Wicked Problem Experience.
- Let the question change as evidence arrives: Karnofsky describes revising the scope and even the central question during an investigation when early assumptions prove weak or a more useful claim emerges. — Learning by Writing.
- Use 'laziness' to narrow an overambitious claim: One of Karnofsky's 'useful vices' is stepping back from an impossibly broad argument to make a smaller claim that the evidence can support. — Useful Vices for Wicked Problems.
- Interrupt an inquiry to check its value: His 'impatience' means periodically asking whether the current deep dive still serves the larger project, not avoiding careful research altogether. — Useful Vices for Wicked Problems.
- Keep enough confidence to contribute: Karnofsky's 'hubris' is confidence that he can make a useful contribution to a difficult question while revising or abandoning claims he cannot defend; it is not a license to ignore experts. — Useful Vices for Wicked Problems.
- Protect the capacity to keep working: Karnofsky's 'self-preservation' is an argument against unsustainable heroic effort on open-ended problems, where a researcher may need to keep learning for a long time. — Useful Vices for Wicked Problems.
- Learn by drafting and revising: Karnofsky writes an early hypothesis, identifies its weak points, investigates the questions most likely to change his conclusion, and seeks feedback. The process is iterative rather than a demand for real-world tinkering in every case. — Learning by Writing.
Part 6: Epistemology & Reasoning
- Use occasional minimal-trust investigations: Karnofsky says most beliefs necessarily depend on trusting others, but an occasional deep investigation can improve one's understanding of a field and judgment about whom to trust. — Minimal-Trust Investigations.
- Use experts without outsourcing every judgment: Expertise is indispensable, Karnofsky argues, yet a carefully chosen independent inquiry can reveal assumptions and help calibrate trust when a question is especially consequential or unfamiliar. — Minimal-Trust Investigations.
- Make uncertain forecasts explicit: Karnofsky favors making assumptions and uncertainty visible in some high-stakes forecasts rather than silently treating one future as inevitable; numerical estimates remain rough guides. — The Bayesian Mindset.
- History informs but does not settle the future: Long-run growth history makes today's trajectory unusual, but Karnofsky sketches several ways it could depart from past patterns rather than simply extrapolating recent experience. — This Can't Go On.
- Communicate the uncertainty in long-run forecasts: Karnofsky distinguishes a serious possibility from certainty when discussing transformative AI and future growth; the width and source of uncertainty matter to the decision. — The Bayesian Mindset.
- Assess futurists' records fairly: Karnofsky finds past futurists' predictions mixed but not uniformly disastrous. The record calls for calibrated humility, not a blanket assumption that confident forecasts always fail. — The Track Record of Futurists Seems Fine.
- Expose the steps in a complex argument: Breaking a surprising conclusion into assumptions, predictions, and value judgments makes it easier to see where disagreement lies and what evidence could change one's mind. — The Bayesian Mindset.
Part 7: Future-Proof Ethics & Morality
- Question everyday moral intuition: Karnofsky argues that relying only on present-day common sense is risky when later generations may see some accepted practices as serious moral errors. — Future-Proof Ethics.
- Look for today's moral blind spots: Past societies accepted practices now widely condemned; Karnofsky uses that history to motivate scrutiny of current norms, not to claim we already know which judgments future people will reject. — Future-Proof Ethics.
- Consider beings outside the familiar moral circle: Karnofsky explores moral concern for animals and future people, while treating the implications of a broad sentience-based ethic as a question for examination rather than a settled rule. — Future-Proof Ethics.
- Take farm-animal welfare seriously: Karnofsky treats farm-animal welfare as a plausible major moral priority within a diversified philanthropic worldview, while recognizing uncertainty about how to weigh it against other causes. — Worldview Diversification.
- Test ethical views for future resilience: Karnofsky explores systematic ethical principles as one way to catch blind spots, but explicitly declines to endorse every conclusion of a fully systemized framework. — Future-Proof Ethics.
- Keep room for moral revision: Karnofsky recommends examining deeply held intuitions in light of possible future moral progress, while remaining cautious about jumping from that humility to any one radical conclusion. — Future-Proof Ethics.
- Balance near-term good and far-future risk: Karnofsky takes future generations seriously but argues that uncertain far-future estimates do not automatically defeat robust opportunities to help people now. — The Moral Value of the Far Future.
- Explore strange implications without automatic endorsement: Karnofsky notes that systemizing ethical ideas can lead to uncomfortable conclusions. He treats those implications as reasons for careful examination, not as obligations to accept every result. — Future-Proof Ethics.
Part 8: Career Capital & Building Aptitudes
- Build aptitudes that travel: Karnofsky advises developing useful aptitudes that remain valuable when a particular job, organization, or cause-specific plan changes. — 80,000 Hours: Building Aptitudes.
- Become good at something early: Karnofsky's early-career advice is to find a role where you can perform well, learn quickly, and build a durable professional aptitude before optimizing every decision for immediate cause impact. — 80,000 Hours: Building Aptitudes.
- Value broad and specialized aptitudes: Karnofsky describes useful aptitudes ranging from research and communication to operations and engineering. The right mix depends on the work; he does not establish that capable generalists are intrinsically rare. — 80,000 Hours: Building Aptitudes.
- Let personal fit shape impact: Karnofsky cautions against choosing a cause only because it looks important in the abstract. Becoming unusually effective at work you can do well can matter more than a simple cause ranking. — 80,000 Hours: Building Aptitudes.
- Keep options open as causes change: Karnofsky notes that views about promising causes can change, which is one reason to build aptitudes that can transfer across organizations and issues. This is not blanket advice to pivot whenever a new concern arises. — 80,000 Hours: Building Aptitudes.
- Seek work with useful feedback: In 2025, Karnofsky observed that many AI-safety roles had become more structured: teams can set concrete goals, see whether technical work succeeds, and give meaningful performance feedback. This is an opportunity, not a prerequisite for independent work. — 80,000 Hours: Concrete AI-Safety Opportunities.
- Operations can be a durable aptitude: Karnofsky explicitly includes management, project coordination, and business operations among aptitudes that help organizations and can remain valuable across causes. — 80,000 Hours: Building Aptitudes.
- Preserve flexibility through transferable skills: Because jobs and priorities can change, Karnofsky favors building abilities that can be carried into different roles rather than relying on one narrowly specified career plan. — 80,000 Hours: Building Aptitudes.