The ten rules you give your child win the first ten minutes of the game.
What should you teach your child? Most parents answer with what they know: the study habits that worked, the paths that paid, the mistakes they would spare them. Passing on hard-won knowledge feels like the core of the job. AI research spent seventy years learning why that instinct runs out.
Rich Sutton called it the Bitter Lesson. Researchers kept building human knowledge into their systems, such as grandmaster heuristics for chess or rules about how speech is produced, and the knowledge helped at first. Then it plateaued, while general methods leveraging computation — search and learning — consistently beat methods that try to encode human knowledge as heuristics. Deep Blue beat Kasparov in 1997 by searching, not by understanding chess the way grandmasters did, and computer Go followed the same path about twenty years later. The lesson is bitter because it says human ingenuity is not the bottleneck it flatters itself to be. Computation is cheaper than cleverness, and it compounds.
Applied to parenting, the Bitter Lesson says: your specific wisdom is overfitted data from a previous epoch.
Ten rules or a bigger sandbox
Picture two parents watching a child play a sandbox game. The first hands over ten rules for success: dig here, build this. The child wins quickly, because the child is running the parent’s pre-programmed knowledge, and for the first ten minutes this looks like superior parenting. The second parent gives no rules, only a bigger sandbox, better tools and more time. The child struggles, and the results look messy. But eventually the child finds strategies the parent never imagined, because the child learned the physics of the sandbox rather than a summary of someone else’s experience.
The comparison is more than a metaphor, because a child’s brain is a general-purpose learning algorithm. Rules cap it at the parent’s level, so instruction is a local maximum: it climbs fast and then stops. The parent’s advice worked in the environment that shaped the parent, which is exactly why it fails in the next one. A child who leaves the script looks like they are making mistakes, when they may be running a search the parent can’t read.
Constant correction also costs something the result can’t show. A parent who keeps directing and optimizing takes over the child’s cycle of goal, effort and attainment. The child gets the meal without the hunt. The castle in the sandbox looks the same, but the child never built the capacity that built it.
Consensus advice is crowded, too. The heuristics every anxious parent hands out, the same tutors, enrichment tracks and safe majors, push every child toward the same oversubscribed paths. The child who searches independently is the one who finds the niche nobody else is in.
And failure is part of the search. A child kept from every failure is trained on a dataset that does not match deployment. They perform beautifully in the training environment and collapse at the first distribution shift, which is what every adult life is.
Why the rules keep coming
If instruction plateaus, why do parents keep handing out rules? Be honest about what advice does. Every rule you give your child makes your uncertainty more bearable, and the child’s compliance is the parent’s anxiolytic. The parent who says “I just want them to be happy” and then enrolls them in six optimization tracks is not lying. They cannot tell the child’s flourishing apart from the relief of their own anxiety, and a parent who can’t bear to watch the child struggle in the sandbox is revealing their own unresolved relationship with failure.
The most successful children are those whose parents had the courage to be computationally invisible. The more “parent” there is in a child’s logic, the less “world” there is. This may help explain why first-generation immigrant children so often outperform. Their parents cannot hand-code local heuristics, because they don’t know the rules of the new game. The child has to search from scratch, and searching from scratch in a rich environment is what the Bitter Lesson predicts will win.
What still gets passed down
The strongest objection comes from Joseph Henrich. Cultural transmission preserves solutions no individual could work out alone. The slow processing that removes cyanide from manioc works, and a child who insisted on rediscovering it by trial would be poisoned first. For most of history, encoding knowledge was the right strategy, because the world held still long enough for the old rules to stay true.
Nor does the Bitter Lesson license neglect. “Just let kids figure it out” lands near the right answer for the wrong reason, since a bigger sandbox without better tools and safety is abandonment. The parent still has the largest job: maximize the child’s compute (health, security, emotional stability, resources) and data (diverse, high-bandwidth experience) while removing their own heuristic interference.
So what changes is what gets transmitted. In a static world you hand down content. In a fast one you hand down infrastructure: the capacity to learn, the will to think, and the steadiness to tolerate the mess of a real search. That turns the parent’s question around. Stop asking what you should teach them, and start asking what is preventing them from learning on their own?
References:
- Rich Sutton, The Bitter Lesson (2019)