hardmaru

Cubic bricks running identical neural cellular automata policies classified four 3D shapes with 98.97% accuracy in simulation and 100% on physical hardware.
Physical builds ranged from 26 bricks for a guitar to 197 for a round table, converging on a shape label in fewer than 60 update cycles.
The same decentralized framework detects structural damage with over 90% accuracy and guides regrowth by predicting one of six axis directions.

View on Bluesky · ♥ 10 ↻ 2 ↩ 0 · 3 from the directory shared this · 36d ago

I am incredibly proud of our Tokyo team for shipping this. By orchestrating the world’s models, we are delivering the resilient blueprint required for AI sovereignty. Read our full vision and results here: sakana.ai/fugu-release 🐡

View on Bluesky · ♥ 12 ↻ 0 ↩ 0 · 6 from the directory shared this · 57d ago

↻ hardmaru reposted

@tksiia.bsky.social

Excited to share CoffeeBench!!☕️☕️☕️ We evaluate LLM agents in a 90-day B2B coffee supply-chain economy spanning farmers, roasters, and retailers, where autonomous firms negotiate, manage inventory, set prices, handle invoices, and manage cash flow. arxiv.org/abs/2606.16613 gi…

CoffeeBench: Benchmarking Long-Horizon LLM Agents in Heterogeneous Multi-Agent Economies arxiv.org

CoffeeBench runs LLMs as autonomous business operators across a 90-day, six-firm economic simulation.
Claude Haiku 4.5 exhibits 'idle-drift': producing coherent plans but persistently choosing inaction.
Higher-performing models communicate more actively with counterpart firms, correlating with better outcomes.

View on Bluesky →

↻ hardmaru reposted

Sakana AI @sakanaai.bsky.social

Fugu-Ultra is now live on OpenRouter! ⚡ We share a core vision with the OpenRouter team: the future of AI isn’t a single monolithic model, but the collective intelligence of the world’s best models working together. Try it: openrouter.ai/sakana/fugu-... 🐡

Fugu Ultra - API Pricing & Providers openrouter.ai

View on Bluesky →

We just pushed a big update to Sakana Chat Free to use: chat.sakana.ai A big motivation for this release is getting people in Japan, especially kids, excited about software development. Vibe-coding turns Kanji & Math drills during the summer break into fun games they can actua…

Sakana Chat chat.sakana.ai

View on Bluesky · ♥ 15 ↻ 1 ↩ 0 · 3 from the directory shared this · 5d ago

Read the full open-access paper: www.nature.com/articles/s41... Blog: sakana.ai/smart-cellul... Congratulations to the team on this achievement!

Smart Cellular Bricks: Towards Collective Intelligence for the Physical World sakana.ai

IT University of Copenhagen, Sakana AI, and Autodesk built cubic bricks that classify their own assembled 3D shape using only neighbor-to-neighbor communication.
In simulation the system hit 98.97% accuracy across 500+ bricks; four physical objects (26 to 197 bricks) all reached correct consensus in under 60 cycles.
The same Neural Cellular Automata substrate detects local damage at 94.8% average accuracy, with some shapes degrading only minimally at 15% brick failure.

View on Bluesky · ♥ 10 ↻ 2 ↩ 0 · 3 from the directory shared this · 36d ago

Incredibly proud of the Sakana AI team. We have developed an orchestration model right here out of Japan that achieves state-of-the-art performance on real-world cybersecurity benchmarks! 🎌 sakana.ai/fugu-cyber-r...

View on Bluesky · ♥ 54 ↻ 4 ↩ 0 · 3 from the directory shared this · 27d ago

Today, we are taking a major step toward that future with the launch of Sakana Fugu. Fugu dynamically orchestrates the world’s best models to tackle complex tasks. We’re proving that a well-orchestrated pool of swappable agents can match restricted frontier models like Fable/M…

Sakana Fugu — Multi-agent System as A Model sakana.ai

Fugu routes tasks through a dynamic multi-agent pipeline exposed as a single OpenAI-compatible API, removing orchestration setup from users.
The system draws on two ICLR 2026 papers: TRINITY assigns Thinker/Worker/Verifier roles; Conductor uses reinforcement learning to design coordination strategies.
Fugu Ultra scored 73.7 on SWE Bench Pro and 93.2 on LiveCodeBench; base Fugu reached 95.5 on GPQA-D, per Sakana's own benchmark reporting.

View on Bluesky · ♥ 4 ↻ 1 ↩ 1 · 2 from the directory shared this · 57d ago

I’ve been thinking about how agents can learn inside world models for years. We decided to scale up our RSI Lab to bridge recursive self-improvement with physical AI and robotics. We are looking for frontier researchers and engineers to join us in Tokyo, Japan: sakana.ai/caree…

Sakana AI is hiring a Member of Technical Staff for its Tokyo-based Recursive Self-Improvement (RSI) Lab, targeting researchers frustrated with brute-force scaling.
The role spans four tracks including world models as verifiable simulators for agentic reasoning and open-ended evolutionary dynamics applied to algorithmic domains.
Sakana points to prior work as receipts: Darwin Gödel Machine on SWE-bench, ALE-Agent winning AtCoder Heuristic Contest 058, ShinkaEvolve with 150 samples.

View on Bluesky · ♥ 19 ↻ 3 ↩ 1 · 2 from the directory shared this · 7d ago

Sakana Namazu: An LLM API with Japanese-vibes! 🎏 Built for Japanese enterprises, featuring frontier-level reasoning and built-in agentic tools. Blog post: sakana.ai/namazu-api#E...

Sakana AI released the Sakana Namazu API, an OpenAI-compatible endpoint for the LLM that has been powering its Sakana Chat product.
The model is built on Moonshot AI's open Kimi K2.6, tuned in-house on proprietary data for Japanese language and business context.
Sakana reports FairPoliticsQA rising from 34.10% to 56.30% versus the base model, with additional gains claimed on JFBench, AIME26, and coding.

View on Bluesky · ♥ 12 ↻ 2 ↩ 0 · 2 from the directory shared this · 15d ago

↻ hardmaru reposted

Sakana AI @sakanaai.bsky.social

🐟 Sakana Namazu API 公開 🐟 本日、Sakana AIは大規模言語モデル「Namazu」をアップデートし、API「Sakana Namazu（サカナ・ナマズ）」として提供を開始しました。 Sakana Namazu API: sakana.ai/namazu 🐟

Sakana Namazu — 日本の感性で、考え、動くLLM sakana.ai

Sakana AI has turned Moonshot's open Kimi K2.6 into Namazu, a Japanese-specialized LLM tuned for business documents, keigo, and agent-style task work.
On Sakana's own comparison, Namazu lifts FairPoliticsQA from 34.10% to 56.30% while preserving Kimi's AIME26, MMLU-Pro, and LiveCodeBench v6 scores.
Pricing is $0.95 per million input tokens and $4.00 output on an OpenAI-compatible API, but Namazu is not available in the EU, UK, or Switzerland.

View on Bluesky →

Dive into our new AI Picbreeder Experiment here: pub.sakana.ai/picbreeder-v...

The AI Picbreeder Experiment: In Search of Automatic Open-Endedness pub.sakana.ai

Researchers from NYU, MIT and Sakana AI replaced Picbreeder's human selectors with frontier VLMs and reported clear qualitative gaps versus the historical human baseline.
Gemini-2.5-pro topped the models tested, but archives showed 'mode collapse' without exploratory noise and 'auto-sycophantic' loops when given more history.
Scaling to 1,000 agents widened coverage yet 10 to 20 percent of the archive turned into uninterpretable adversarial psychedelic patterns.