Google DeepMind’s AlphaEvolve uses Gemini to evolve algorithms across genomics, quantum circuits, TPU design, logistics, and financial ML – one year after launch.
Key Takeaways
Genomics: Improved DeepConsensus DNA error-correction model, cutting variant detection errors 30% for PacBio sequencing instruments.
Grid optimization: Boosted GNN feasibility on AC Optimal Power Flow from 14% to 88%, slashing costly post-processing.
AI infrastructure: Designed a TPU circuit counterintuitive enough that engineers would not have proposed it; integrated into next-gen TPU silicon. Also cut Spanner write amplification 20% and compiler storage footprint ~9%.
Quantum physics: Produced circuits with 10x lower error than conventional baselines, enabling first-of-a-kind Willow quantum processor experiments.
Commercial results: Klarna doubled transformer training speed; FM Logistic cut 15,000 km/year of routing; Schrödinger achieved ~4x speedup in MLFF training and inference.
Hacker News Comment Review
Debate centers on scope: AlphaEvolve excels at well-defined, high-dimensional optimization problems (matrix ops, compaction heuristics, circuit layout) – commenters question how results translate to ambiguous, human-centric problem spaces.
A confirmed Google employee states internal Gemini tooling has caught up on UI/UX and VCS integration; main frustration is the pace of change, not model capability.
Commenters draw a sharp distinction between AI optimizing AI-adjacent software (cost/speed) versus AI autonomously designing a fundamentally more capable successor model – the latter has not occurred.
Notable Comments
@ainch: Notes DeepMind is uniquely active on hard science problems while OpenAI/Anthropic focus on enterprise and coding revenue.
@stijntonk: Vertex AI 429 rate limits are blocking production use of Gemini models for corporate clients, undermining the commercial narrative.
