How do post-training, reasoning and agentic RL, inference-time search, tool use, shared memory and evolutionary curricula fit together? Our frontier learning agenda maps the evidence, hypotheses and limits onto the ten foundational questions below. Its proposed investigations give each connection a concrete comparison.
- 01 How does a learner value what doesn't pay yet? Distant objectives rarely reveal their prerequisites. Human milestones, learned instrumental value and self-generated goals are three very different answers.
- 02 How do you keep a discovery alive before anyone knows it matters? Selecting for current reward kills future innovations. Saving everything is impossible. Novelty search, quality-diversity and archives each offer a partial answer.
- 03 What should drive exploration, and could the drive itself evolve? Surprise can be noise. Learning progress is better but not solved. Perhaps the mechanisms that decide what to explore should be subject to selection too.
- 04 Where do reusable concepts come from? Predictive maps, schemas, replay and world models suggest how recurring structure becomes abstraction, and where prediction alone falls short.
- 05 Can the machinery of learning stay evolvable? Evolving a learner rather than a behaviour, separating timescales, and letting a system adjust its own search without a hand-written schedule.
- 06 How does inherited information stay meaningful? Genomes as developmental programs, recombination without broken offspring, whether protected heredity can emerge, the error catastrophe, and the cost of a generation.
- 07 Biology had billions of years. What do we have? The compute gap, why "physics performs the computation", and the search for a minimal set of general inductive biases that makes discovery tractable.
- 08 What does an agent in an ecology need? Memory and recurrent cognition, very large structured action spaces, thinking that has to be paid for, and communication that costs something.
- 09 When does a population become a higher-order optimizer? Coalitions, institutions, cumulative culture and major transitions, and why a firm outlives every one of its employees.
- 10 How do you reward real usefulness? Evaluators get gamed, and in a decentralized system the participants are optimizers too. Where real value comes from, and Goodhart's law as a design problem rather than a stop sign.
Tensions we're deliberately not resolving yet
Some of the most useful things we've written down are disagreements. We'd rather keep them visible than quietly pick a side.
| On one side | On the other |
|---|---|
| A single distant external objective, no intermediate rewards | Open-ended exploration that may abandon the objective altogether. "Only one external reward" doesn't mean "no other learning signal." |
| Avoiding domain-specific scripts | Unavoidable initial structure: sensors, action spaces, memory limits and update rules already constrain what can be learned. |
| No premature rigid architecture | Any real experiment requires concrete computational choices. |
| Genome-based evolution with reproduction and speciation | An earlier, more practical simplification to static, non-reproducing multi-agent populations. |
| Intrinsic drives that explore without immediate payoff | Fidelity to the actual objective, and the scarce resources exploration consumes. |
| Strong claims about open-endedness | Finite compute. No finite run proves indefinite innovation. |
| Biology as inspiration | Biology as something to replicate. Its history motivates shortcuts but doesn't tell us which mechanisms are necessary. |
| Self-modifying replicators with no fixed genome | The locality and protected heredity that cumulative evolution seems to need. |
The same problem, ten times over
Nearly every discussion we have ends up at one of a handful of recurring tensions. They increasingly look less like separate problems and more like different faces of one problem, open-ended optimization:
- Exploration vs. exploitation. How do you keep searching without drowning useful progress in noise?
- Diversity vs. convergence. How do good ideas spread without every agent becoming identical?
- Individual vs. population intelligence. How much cognition belongs inside an agent, and how much between agents?
- Engineering vs. emergence. How much structure can we supply before we've prescribed the answer?
- Learning vs. evolution. Which adaptations happen within a lifetime, and which should become heritable?
- Competition vs. cooperation. How do you get the benefits of both without collapsing into predation or stagnation?
- Internal reward vs. external value. How do artificial incentives stay connected to something people actually want?
- Specialization vs. generality. Can a population of specialists be more general than any of its members?
- Stability vs. open-endedness. How does accumulated competence survive while genuinely new strategies keep appearing?
- Power vs. alignment. How does steadily increasing capability stay pointed at good outcomes?
The question underneath
What minimal set of mechanisms is sufficient for a system to begin discovering the rest for itself?
When we run experiments
We've run a couple of small toys so far (see Experiments). For anything we run, we record: the question or hypothesis; the environment, including hidden or supplied structure; model and learning assumptions; compute, memory, interactions and wall time; the comparison; results, including failures; evidence of retention and transfer; and the limits of interpretation. Above all, what humans supplied versus what actually emerged.