Hacker News

We built optimize_anything, an API that optimizes any artifact representable as text — code, prompts, agent architectures, configs, even SVGs. It extends GEPA (our prompt optimizer, discussed here previously: https://arxiv.org/abs/2507.19457) far beyond prompts.

The API is deliberately minimal. You provide what to optimize and how to measure it:

import gepa.optimize_anything as oa

def evaluate(candidate: str) -> tuple[float, dict]: result = run_my_system(candidate) return result.score, {"error": result.stderr, "runtime": f"{result.time_ms}ms"}

result = oa.optimize_anything( seed_candidate="", evaluator=evaluate, )

The evaluator returns a score plus diagnostic feedback (we call it "Actionable Side Information" — stack traces, rendered images, profiler output, whatever helps diagnose failures). An LLM proposer reads this feedback during a reflection step and proposes targeted fixes, not blind mutations. Candidates are selected via a Pareto frontier across metrics/examples, so a candidate that's best at one thing survives even if its average is mediocre.

Two ideas distinguish this from AlphaEvolve/OpenEvolve/ShinkaEvolve-style LLM evolution: (1) diagnostic feedback is a first-class API concept rather than a framework-specific mechanism, and (2) the API unifies three optimization modes — single-task search (solve one hard problem), multi-task search (solve related problems with cross-transfer), and generalization (build artifacts that transfer to unseen inputs). Prior frameworks only express mode 1.

We tested across 8 domains. Selected results:

Coding agent skills: Learned repo-specific skills push Claude Code to near-perfect task completion and make it 47% faster Cloud scheduling: Discovered algorithms that cut costs 40%, topping the ADRS leaderboard over expert heuristics and other LLM-evolution frameworks Agent architecture: Evolved a 10-line stub into a 300+ line ARC-AGI agent, improving Gemini Flash from 32.5% → 89.5% Circle packing (n=26): Outperforms AlphaEvolve's published solution Blackbox optimization: Generated problem-specific solvers matching or exceeding Optuna across 56 EvalSet problems CUDA kernels: 87% match or beat baseline; multi-task mode outperforms dedicated single-task runs

``` pip install gepa ```

Blog with full results and runnable code for all 8 case studies: https://gepa-ai.github.io/gepa/blog/2026/02/18/introducing-o...

GitHub: https://github.com/gepa-ai/gepa

Comments URL: https://news.ycombinator.com/item?id=47083674

Points: 2

# Comments: 0

Article URL: https://www.together.ai/blog/consistency-diffusion-language-models

Comments URL: https://news.ycombinator.com/item?id=47083648

Points: 1

# Comments: 0

Article URL: https://www.varietyiq.com/blog/pricing

Comments URL: https://news.ycombinator.com/item?id=47083608

Points: 1

# Comments: 0

Article URL: https://fromtheprism.com/vibe-coding-audit

Comments URL: https://news.ycombinator.com/item?id=47083597

Points: 1

# Comments: 1

Article URL: https://wesmckinney.com/blog/mythical-agent-month/

Comments URL: https://news.ycombinator.com/item?id=47083594

Points: 2

# Comments: 1

Article URL: https://gepa-ai.github.io/gepa/blog/2026/02/18/automatically-learning-skills-for-coding-agents/

Comments URL: https://news.ycombinator.com/item?id=47083586

Points: 2

# Comments: 1

Article URL: https://electrek.co/2026/02/18/most-ev-batteries-outlast-their-cars-real-world-data-shows/

Comments URL: https://news.ycombinator.com/item?id=47083580

Points: 2

# Comments: 0

Article URL: https://lareviewofbooks.org/article/nothing-ever-happens-mister-squishy-and-the-year-of-the-sentence-diagram/

Comments URL: https://news.ycombinator.com/item?id=47083575

Points: 2

# Comments: 0

Article URL: https://www.bbc.co.uk/news/uk-england-leeds-31015325

Comments URL: https://news.ycombinator.com/item?id=47083573

Points: 3

# Comments: 1

Article URL: https://www.fiercehealthcare.com/ai-and-machine-learning/inside-chais-failed-assurance-labs

Comments URL: https://news.ycombinator.com/item?id=47083571

Points: 1

# Comments: 0

Article URL: https://www.reuters.com/world/us/us-removing-guardrails-proposed-saudi-nuclear-deal-document-says-2026-02-19/

Comments URL: https://news.ycombinator.com/item?id=47083569

Points: 2

# Comments: 0

Article URL: https://www.wsj.com/politics/policy/white-house-offers-new-details-on-its-push-to-ban-housing-investors-97fb1829

Comments URL: https://news.ycombinator.com/item?id=47083566

Points: 4

# Comments: 0

Article URL: https://github.com/devrupt-io/LLaMAudit

Comments URL: https://news.ycombinator.com/item?id=47083548

Points: 1

# Comments: 0

Article URL: https://www.culturefoundry.com/cultivate/content-strategy/how-to-use-claritys-ai-bot-report-for-seo-aeo/

Comments URL: https://news.ycombinator.com/item?id=47083322

Points: 1

# Comments: 0

kept losing good conversations to /compact. you spend 40 minutes having claude map your codebase, it builds up real understanding, then context fills up and /compact crushes everything into a 3k token summary. "we discussed auth and decided on JWT." cool thanks.

dug into the actual session JSONL files and the breakdown is kind of absurd -- 60-70% is raw file contents from tool reads that claude already synthesized, another 15-20% is base64 thinking signatures. your actual conversation is like 10-15% of the window.

so i built cmv. it strips the junk and keeps every message verbatim. tool results over 500 chars become stubs, thinking sigs get removed, everything you said stays.

cmv trim --latest # trim and relaunch, 50-70% smaller

cmv snapshot "analysis" --latest # save a session state

cmv branch "analysis" --name "auth" # fork from it later

also has a TUI dashboard that shows the token breakdown per session so you can see what's eating your context before you do anything.

what it's not:

* not a token monitor (ccusage etc already do that)

* doesn't touch original sessions, everything creates copies

* local only, reads JSONL directly, no API calls

curious how others handle this. most people seem to just accept /compact but losing a deep architectural discussion to a bullet point summary felt wrong enough to build something.

https://github.com/CosmoNaught/claude-code-cmv

Comments URL: https://news.ycombinator.com/item?id=47083309

Points: 1

# Comments: 0

Article URL: https://www.livescience.com/space/space-exploration/there-will-be-leadership-accountability-bungled-boeing-starliner-mission-put-stranded-nasa-crew-at-risk-report-says

Comments URL: https://news.ycombinator.com/item?id=47083300

Points: 1

# Comments: 0

Article URL: https://www.newyorker.com/culture/the-lede/silicon-valleys-favorite-doomsaying-philosopher

Comments URL: https://news.ycombinator.com/item?id=47083294

Points: 1

# Comments: 1

Article URL: https://arxiv.org/abs/2512.14982

Comments URL: https://news.ycombinator.com/item?id=47083281

Points: 1

# Comments: 0

Article URL: https://podcastarchiveengine.vercel.app/

Comments URL: https://news.ycombinator.com/item?id=47083274

Points: 1

# Comments: 0

Article URL: https://castpixel.itch.io/pcb-forge

Comments URL: https://news.ycombinator.com/item?id=47083273

Points: 1

# Comments: 0