Sam Altman's productivity principles, examined by Cal Newport

Executive overview

Sam Altman published a short productivity essay in 2018, at a pivotal moment for OpenAI. Cal Newport works through five ideas from that essay, agreeing with most and pushing back on one.

The overarching theme: almost nothing matters, so protect ruthlessly the few things that do. Do the deep work first, say no to almost everything else, and don't optimise the system at the expense of choosing the right problem.

The most important productivity decision is what to work on, not how fast you work.

The five ideas from Altman's essay

• Compound growth in careers — Altman argues a 10% daily output gain compounds massively; Newport disagrees, noting historically productive people (Newton, Curie, Austen) succeeded through slow, consistent thought, not daily output maximisation
• Chasing 10% more done each day more often produces busyness, not results
• Direction over speed — Altman's own counterpoint: picking the right thing matters more than moving fast; leave time to think, read, spend time in nature
• Newport's corollary: resist starting projects until you can't help it; fewer things done better
• Simple lists, not complex systems — Altman keeps a single list, stars the important items, and just works through it; no categories, no sizing
• Full capture (writing things down) reduces cognitive load and missed deadlines
• For people with many unignorable demands, smarter context-based task storage helps alongside Altman's big-list approach
• Default to no — Altman is "terse to the point of rudeness" on email; avoids meetings and conferences
• The things that made OpenAI successful were small in number and hard; everything else was in the way
• Protect mornings — Altman blocks the first few hours for deep work; meetings go in the afternoon
• A simple rule ("no meetings before X") makes a bigger difference than complex scheduling

Managing administrative sprawl

• Listener question: overseeing 150 students generates a flood of reactive emails and signing tasks
• Key principle: making many people do 10% more work on their end can make your life 100% easier
• Build a procedure page or FAQ students can be directed to for common requests
• Batch routine signing tasks into a single weekly slot (e.g. Friday afternoon shared folder)
• Ask students to pre-gather information you would otherwise have to look up yourself
• Regular office hours absorb one-off complex requests without reactive interruption

Deep work timing and focus

• On days without meetings, go straight into deep work — do not check email first
• Checking email loads multiple cognitive contexts that take ~30 minutes to clear
• Ideal knowledge-work structure: no email or meetings before 11 a.m., or dedicated no-meeting days at home
• Deep work without interruption is a categorically different cognitive state than deep work with interruption

Career capital and skill development

• To find what skills matter in your field, talk to people already where you want to be
• Ask: "When you made that jump, what were you doing that others weren't?"
• Avoid writing your own story about what you want the answer to be — it rarely matches reality
• The only path to exceptional performance is expert-guided deliberate practice
• Most people can reach "good enough" at professional skills to build real career capital; fewer can reach exceptional

Monitors and deep work tools

• Useful threshold: two-window width (two full-size readable panes) covers most real workflows
• Four or five monitors gives diminishing returns for most people
• For pure writing, a single focused screen — or even laptop in composition mode — beats a multi-monitor setup
• Tool choice should match the mode: reference + draft side-by-side during research, single screen during language work

The AI 2027 report

• Five researchers published a speculative scenario in which an AI company achieves recursive self-improvement by 2027 and the outcome is catastrophic
• Critics note: predicting the nature of nonexistent future technology is inherently unreliable; the storyline mirrors Nick Bostrom's 2014 Superintelligence with numerical detail added
• Current evidence cuts against the scenario: large language models remain highly amenable to fine-tuning and alignment; the "control problem" has not materialised as predicted
• LLMs are trained on essentially all available text; further gains now rely on human-feedback reinforcement learning, and that curve is flattening
• Reinforcement learning (not language models) is responsible for every domain where AI surpasses humans — chess, Go, protein folding, maths
• A more plausible near-term trajectory: smaller, bespoke RL models aimed at specific tasks, not a single mega-brain
• The report's value is attention-raising for safety — not as a literal forecast