Summary

Sakana’s ShinkaEvolve makes LLM-driven program search materially cheaper, reaching a state-of-the-art circle-packing solution in fewer than 200 LLM interactions where comparable systems may sample roughly 1,000 programs. It evolves an archive through patches, full rewrites and crossovers, while a UCB-style router allocates work across frontier models. The implication is that efficiency can democratize discovery and enable much larger searches.
The central constraint is the “problem problem”: today’s systems optimize the evaluator humans provide, but major innovation may require “first inventing a different problem.” Robert Lange’s circle-packing proxy allowed tiny overlaps before shrinking the radii to obtain an exact solution; rerunning on the exact constraint took longer. Truly open-ended science therefore requires problems and solutions to co-evolve, alongside automatic verification—not merely better optimization against a fixed score.
Starting conditions impose a hard trade-off between efficient convergence and genuine novelty. An already optimized seed tends to trap ShinkaEvolve in local optima, while an “impoverished solution” leaves more space for diversity but demands longer, potentially month- or year-scale searches. Tim Scarfe’s proposed unlock is massive parallelism: thousands of differently seeded searches that occasionally exchange discoveries inside a meta-evolution loop.
The product interface could shift from single-threaded chat assistants to overnight “vibe optimization and vibe researching.” Lange imagines researchers steering systems “like a shepherd,” then returning in the morning to multiple experiments, accumulated evidence and newly proposed branches. Verification remains the gating asset because it is “easier to generate a lot of solutions than to actually hard verify them,” leaving reward hacking and false discoveries as central scaling risks.
Early results extend beyond a toy benchmark: ShinkaEvolve improved an agent built around GPT-4.1 Nano, produced an AtCoder-style solution that would have ranked second on one AIDE Bench task, and explored mixture-of-experts load-balancing trade-offs within roughly 20 generations. ARC remains explicitly unproven: Lange has only started investigating it and hopes first for lower cost, with performance gains “to be seen.”
AI Scientist V2 replaces V1’s linear, template-bound experiment pipeline with a parallel agentic tree search that adapts hypotheses to evidence. One paper produced by the system crossed an ICLR workshop’s acceptance threshold before meta-review, but Lange concedes that “not every paper” is remotely Nature-worthy and some output is slop. His calibrated call is that automated science is at a “GPT-1 moment,” with humans still needed to seed, redirect, verify and transfer consequential insights.
The strategic Rubicon is an AI discovering “the new Transformer architecture or something massive” that the field subsequently adopts. Scarfe worries compute-rich labs such as Google could privately own those discoveries; Lange’s counterweight is sample-efficient, openly available infrastructure that harnesses collective intelligence. He expects AI to amplify human creativity rather than immediately eliminate scientists, but says model capability, scaffolding and human interfaces must all improve—and society must adapt quickly enough to retain agency.

Deep dive

1. Open-ended science must search over problems, not just answers

Lange describes Sakana—nearly two years old during the conversation—as pursuing AI for Japan alongside ambitious research that may not receive resources elsewhere. As a founding researcher, he connects its culture to Ken Stanley’s open-endedness: preserve unusual branches because discoveries often depend on stepping stones that initially appear useless.
The conversation frames research as evolutionary tree search: scientists traverse many ideas and experiments, yet a paper reports only one path through that tree. Current LLM systems receive a fixed evaluator and search solutions to one prescribed problem; that discards the hidden branches and reformulations through which human researchers often make progress.
The missing capability is the “problem problem.” Innovation may require inventing an intermediate problem, importing an abstraction from another field or discovering a recursive reduction before solving the original task. Lange wants problems and solutions to co-evolve in systems resembling PowerPlay or POET, eventually running for “years even potentially.”
Scarfe’s pushback is that even auto-curricula inherit human design bias: some human-written mechanism still determines how environments become more complex. Lange’s qualified answer is that LLMs might nevertheless interpolate or extrapolate into “unknown unknowns” their designers never anticipated; the possibility is real, not yet demonstrated as a general solution.

2. ShinkaEvolve turns frontier models into an evolutionary program population

ShinkaEvolve maintains an archive of programs, samples a parent plus inspiration programs, and asks an LLM to patch, rewrite or combine them. Every candidate is executed against an evaluator, its evidence enters the archive, and many branches evolve in parallel rather than as one sequential conversation.
Lange’s tree analogy is literal: every node is a program, later nodes branch from selected parents, and discoveries can diffuse across the database. The objective is to accumulate complementary “stepping stones,” rather than hope that one model produces the complete answer in a single pass—the evolutionary alternative to relying on enough random samples, like a monkey typing at a keyboard.
Sample efficiency motivated the work. Where related systems may evaluate about 1,000 programs, Scarfe notes that the circle-packing run converged in fewer than 200 LLM interactions. ShinkaEvolve improved the canonical result while reducing API cost and evaluator time, and Sakana released the code openly to make scientific search more broadly accessible.
The doubled name captures another layer: “Shinka” means evolve, while the evolutionary algorithm itself co-evolves during the run. Rather than relying only on Gemini or another single provider, it ensembles frontier models and changes which model receives each mutation opportunity.

3. Initialization buys efficiency by narrowing the space of possible novelty

Scarfe observes that autonomous LLMs often follow their prompt for a few steps and then produce nothing genuinely new: they remain “parasitic on their starting conditions.” His image-generation example is telling—ask for a Swiss lake, boats and mountains, then wonder where the birds are; the system faithfully omitted what the human failed to imagine.
Lange saw the same dependence in ShinkaEvolve. A highly optimized initial program tends to produce local refinements and little novelty, whereas an impoverished seed opens more evolutionary directions. The price is time: unconstrained search may need true open-endedness, while strong inductive biases converge efficiently but constrain what can be discovered.
Starting from an empty program is technically conceivable, but neither speaker treats that as a free solution. With primitive building blocks, the search space explodes and useful learning signals disappear. Scarfe instead proposes many starts “a little way up the stack,” with occasional sharing or crossover between otherwise distinct populations.

4. Circle packing exposed why surrogate problems are themselves inventions

The successful circle-packing lineage used three complementary components: circle initialization, constraint optimization and a “reheating” phase that added noise to squeeze out further improvement. Lange also observed crossover operations in which different programs combined distinct concepts, showing how accumulated stepping stones could be recombined.
The revealing detail was the evaluator. Sakana initially permitted a tiny amount of circle overlap, then slightly reduced the radii to recover an exact configuration without materially changing the score. When the team reran ShinkaEvolve under the exact constraint, reaching comparable quality took longer—the proxy problem made the discovery easier.
Scarfe compares that move to slack variables making support-vector-machine optimization tractable. His objection is decisive: if the fitness function simply prohibited intersections, today’s system would not decide to relax the constraint. Lange agrees; humans remain much better at reformulating problems or transferring perspectives between areas such as number theory and linear algebra.

5. Semantic memory and model routing add structure without eliminating trade-offs

To move beyond reward-only search, every candidate is summarized into a scratchpad, which accumulates global insights and generates meta-recommendations for later prompts. This gives evolution a semantic layer, but knowledge diffusion can homogenize the tree; some problems benefit from shared insights, while others need “hard islands” that remain isolated.
The same task dependence appears on ARC. Deterministic program transformations can provide a clean evolutionary signal on ARC-AGI-1, while Lange found semantic instruction evolution more efficient on ARC-AGI-2. His desired system would choose programmatic, instruction-based or hybrid representations automatically.
Model selection uses an upper-confidence-bound-style bandit. The model with the best headline coding benchmark is not invariably the best mutator, and credit is ambiguous when GPT-5 creates one stepping stone before Sonnet 4.5 improves it. Each model is treated as an arm, initially explored, then favored according to its history of improving similar parent nodes.
Scarfe asks whether a global bandit score misses local context—Gemini might be best for one particular program even if another model leads overall. Lange says probability remains allocated across models and no provider clearly dominated their tasks; an LLM router is conceivable, but raw programs and fitness histories would exhaust context without effective compression.

6. Diverse mutations point toward branching, self-improving software development

ShinkaEvolve protects imports and evaluator-critical code with immutable markers. If a proposed mutation touches protected regions, a reflection-and-rejection loop resamples it, improving robustness and mitigating some security or safety problems without asking the model to respect boundaries perfectly.
Patch-based mutation supports precise improvement, while full rewrites can jump to an entirely different stepping stone. Crossover supplies two parents and asks for a complementary synthesis. No operator wins universally, but Lange found operator diversity useful because it changes the geometry of what the search can reach.
The next engineering problem is evolving whole repositories rather than single files. Codebases could be flattened, but that discards useful hierarchy; Aider-style repository maps offer abstraction with their own information trade-offs. Scarfe imagines future coding tools exploring Git branches concurrently, then merging discoveries instead of advancing through one accepted checkpoint at a time.

7. Research interfaces will move from chat to parallel orchestration

Lange calls today’s chat and coding-assistant workflow inherently single-threaded: a human watches changes, accepts them and requests another. He sees it as a transitional interface before “vibe optimization and vibe researching,” where many experiments and candidate solutions proceed simultaneously.
His ideal researcher steers during the day, presses play at night and returns to accumulated evidence, new hypotheses and multiple live branches. The human becomes “more like the shepherd of the ship” while systems execute experiments; analysis and direction remain human responsibilities, but mechanical iteration moves into the background.
Scarfe extends that vision to non-experts. Unlike MATLAB or Mathematica, which demand precise problem formulations, a discovery platform could infer interests from natural-language interaction and proactively investigate them. Commercially, it might aggregate work from mathematicians and other users, recognize common structures and connect otherwise separate searches.
Scaling such collective discovery requires mechanism design, not only better models. Lange anticipates economies or leaderboards around numerically scorable work so contributors still benefit from sharing compute, ideas and discoveries. The equilibrium among automated agents, human steering and incentives remains unresolved.

8. Verification—not generation—is the scarce resource

Lange’s bottleneck is blunt: “oftentimes it’s easier to generate a lot of solutions than to actually hard verify them.” LLMs can perform soft inspection or mentally approximate a stack trace, but that is not exact; reward hacking can produce impressive-looking candidates that shortcut the intended discovery.
Open-ended science therefore needs problem formulation and automatic verification alongside solution search. Executable simulators provide strong feedback in some domains, but not every field offers such a precise evaluator. Lange hopes models will eventually implement verifiers in code or perform them reliably, while conceding that this capability is not available today.
Frozen foundation models plus evolutionary scaffolding offer an intermediate route to inference-time adaptivity without continually fine-tuning enormous models. Scarfe’s framing is that ShinkaEvolve creates and recombines new building blocks in a tree, responding to novelty while retaining the economic advantages of shared frontier models.

9. Three applications show breadth, while ARC remains an open bet

Using ADAS—automatic design of agentic systems—ShinkaEvolve evolved an agent scaffold for AIME-style mathematics tasks. It “dramatically” improved the cheap GPT-4.1 Nano, and the resulting scaffold transferred either to other language models or to different competition years, suggesting the search found more than a model-specific prompt trick.
On AIDE Bench, Sakana began with solutions from an existing AI agent and evolved them further for AtCoder-style heuristic programming tasks. Lange says the combined system would have ranked second on one challenge, evidence for ShinkaEvolve acting as a co-scientist that improves machine- or human-supplied starting points.
A mixture-of-experts experiment evolved load-balancing loss functions. Within roughly 20 generations, the search illuminated a convex hull of alternatives rather than one winner, exposing trade-offs between model performance and expert load balancing. The point was to explore the trade-off surface, not only find a single best objective.
ARC is still a hypothesis. Lange has “started looking into it” but refuses hard claims before collecting results; he hopes to reduce cost and perhaps improve performance. Against Jeremy Berman’s view that efficiency may offer few additional percentage points, Lange argues cheaper search can be rescaled—but says plainly, “Right now, I don’t have any evidence for it.”

10. AI Scientist V2 replaces scripted imitation with adaptive experimentation

AI Scientist V1 began with a template experiment, used literature search and Semantic Scholar, proposed code diffs, executed a linear plan and wrote a paper. If the initial idea failed, the remaining experiments could still proceed mechanically, producing something paper-shaped without adapting the research direction to the negative evidence.
V2 makes the LLM draft the experiment and uses parallelizable agentic tree search inspired by falsification: evidence updates which hypothesis or experiment comes next. It removes the mandatory base template, adds a vision-language-model reader to check figure-caption alignment, scales across more computational nodes and again packages the result as a paper.
One paper produced by the system passed an ICLR workshop acceptance threshold before meta-review. Lange does not oversell it: “not every paper” is Nature-worthy, and some output is slop. Scarfe sees the system as a way to prototype fields and establish initial results, with humans building on top—not as a machine generating repeated foundational breakthroughs.
Lange’s longer-term framing is a “GPT-1 moment.” Larger budgets can already yield occasional workshop-level contributions amid noisy review, while better multi-experiment memory, hypothesis formation, negative-result accounting and physical robotic laboratories might expand capability over 10 years. None has yet produced an innovation comparable to residual connections.

11. Human judgment remains central, but passive adoption could erode it

Scarfe sees less labor disruption than predicted: generated copy has increased demand for writers and editors who repair it, while UK system administrators once earning about £60,000 rebranded as cloud or DevOps engineers and more than doubled their pay. His question is whether AI will eventually change faster than workers can adapt.
Lange answers that the issue is the relative speed of technological and cultural evolution. Scientists will lose some tasks, but new dimensions may open around steering, verification and synthesis. AI Scientist V1 and V2 are “glimpses” of transformation; major discoveries still require humans to seed direction, check results and transfer insights.
Both warn about cognitive dependence. Scarfe compares auto-accepting code to switching off behind a strong autopilot; Lange says assistants can feel “almost like drugs” when users exhaust their token budget and no longer know what to do. Discipline, education and active experimentation are necessary to keep humans on their own reasoning path.
Lange expects human values to remain coupled to open-ended systems because their training data, verifiers, research directions and compute budgets are human-selected. Even apparently irrelevant mathematics might later become a stepping stone, but “we are the ones who control how many FLOPs” and can direct search toward priorities such as cancer research.

12. The decisive moment is an AI discovery the whole field adopts

Lange identifies three unconverged layers: underlying model capability, scaffolding around models and the interface through which humans steer them. Separately, Chris Lu says research may move along a spectrum from “with AI” toward “by AI,” with higher-level orchestration and oversight by humans, and that scientific practice will fundamentally change over the next five, 10 and 20 years.
Papers may become the bottleneck rather than the destination. Lange imagines agent-accessible scientific artifacts whose figures are reproducible and whose code and data can be queried through model-facing interfaces, letting agents replicate, ablate and extend results. Yet the paper format has existed for hundreds of years, making adoption far from certain.
Automated paper generation initially makes peer review more important: abundant submissions require machine filtering followed by human examination of the “cream of the crop.” AI reviewers help, but credible verification may require actually rerunning experiments; the ultimate test remains whether a result diffuses through and survives scrutiny by the research community.
Scarfe’s “Rubicon moment” is an AI discovering a new Transformer architecture—or something comparably consequential—that everyone then uses. His concern is ownership: compute-rich organizations could monopolize machine-made discoveries. Lange’s response is to keep systems such as AI Scientist and ShinkaEvolve open and efficient enough for collective intelligence to shape the emerging scientific infrastructure.

Elena Verna: How Lovable Launches Product & Hacks Social to Go Viral Episode 176 - March 13, 2026