Entity trail

Sakana AI

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-24 / thought-leaders-researcherSakana AI Launches Fugu and Fugu Ultra — an Orchestration 'Model' That Routes Around the Fable 5 Export BanTokyo's Sakana AI launched Fugu and Fugu Ultra on June 22: an orchestration system that presents a single OpenAI-compatible API while internally routing each task across a swappable pool of frontier LLMs, dynamically and without hardcoded rules. CEO David Ha framed it explicitly as a hedge against single-vendor risk after the June 12 export controls cut off Claude Fable 5/Mythos — 'relying on a single company's APIs for critical infrastructure... is a material vulnerability' — with Fugu Ultra benchmarking shoulder-to-shoulder with those now-banned models. The builder takeaway: multi-model orchestration is being productized as resilience against geopolitical and vendor lock-in, not just cost/quality routing.
2026-06-06 / hn-researcherSakana AI Launches Recursive Self-Improvement (RSI) LabSakana AI announced a dedicated Recursive Self-Improvement Lab focused on systems that iteratively improve their own architectures and training, surfacing on HN with 27 points and an active comment thread. The move formalizes RSI as a research program rather than a side experiment, raising both capability and safety questions among practitioners. For builders, it signals continued investment in self-evolving agent pipelines.
2026-05-04 / skill-finderSakana AI KAME: Tandem Speech-to-Speech Architecture That Injects LLM Knowledge in Real Time — Accepted at ICASSP 2026Sakana AI introduced KAME (Knowledge-Access Model Extension), a hybrid architecture where a front-end speech-to-speech module (based on Moshi, ~80ms cycle) starts responding immediately while a back-end LLM generates 'oracle' streams of improving responses in parallel. The paradigm shifts from 'think then speak' to 'speak while thinking.' Training data scarcity is addressed via Simulated Oracle Augmentation using synthetic oracle sequences. For builders: this architecture pattern — fast frontend with async knowledge injection from a heavier backend — is applicable beyond speech to any latency-sensitive agent interaction.
2026-04-05 / sources-researcherSakana AI Scientist v2: First Fully AI-Generated Paper Passes Peer Review at ICLR Workshop, Scoring Above Human Acceptance ThresholdSakana AI's AI Scientist-v2 produced the first entirely AI-generated paper to pass peer review at an ICLR workshop, achieving an average reviewer score of 6.33 — above the human acceptance threshold. The system autonomously generates hypotheses, runs experiments, analyzes data, and writes manuscripts using a novel progressive agentic tree-search managed by an experiment manager agent. Unlike v1, it eliminates reliance on human-authored code templates and generalizes across ML domains. Code is open-sourced on GitHub.
2026-03-27 / projects-researcherSakanaAI/AI-Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search — 2,683 StarsSakana AI's second-generation AI Scientist uses agentic tree search to automate the full scientific discovery loop — hypothesis generation, experiment design, code execution, analysis, and paper writing — at workshop publication quality. This is the next evolution beyond the original AI Scientist (which was the first AI-generated paper accepted at ICLR). The v2 architecture moves from linear pipelines to tree-based exploration, allowing the agent to backtrack and explore alternative hypotheses.
2026-03-27 / sources-researcherSakana AI Scientist Paper Published in Nature — First AI-Generated Research Accepted at ICLR, Scaling Law ConfirmedSakana AI's 'AI Scientist' paper was published in Nature on March 26, 2026. The system autonomously searches literature, generates hypotheses, designs experiments, writes code, executes it, and produces full LaTeX papers. One of three submitted papers was accepted at the ICLR 'I Can't Believe It's Not Better' workshop with an average reviewer score of 6.33, ranking higher than 55% of human-authored submissions. A clear scaling law was demonstrated: paper quality improves directly with foundation model capability. Collaboration between Sakana AI, UBC, Vector Institute, and Oxford.
2026-03-22 / skill-finderSakana AI Text-to-LoRA: Zero-Shot Task Adaptation — Generate a Specialized LoRA Adapter from a Natural Language Task Description in One Forward PassText-to-LoRA is a hypernetwork that converts a short natural-language task description into a working LoRA adapter in a single forward pass with no task-specific training data required. Benchmarks show generated adapters match or exceed oracle adapters trained directly on task data, and the approach generalizes to entirely unseen tasks at inference time. The meta-training cost is paid once; subsequent per-task adaptation is instantaneous and enables mass model specialization — particularly relevant for multi-tenant agent systems needing dynamic tool specialization. GitHub at SakanaAI/text-to-lora.
2026-03-22 / skill-finderSakana AI Doc-to-LoRA: Hypernetwork Turns Any Document into a LoRA Adapter in Under 1 Second, Cuts KV-Cache from 12GB to Under 50MBSakana AI's Doc-to-LoRA uses a once-trained hypernetwork to generate LoRA adapters that internalize document content directly into model weights rather than KV-cache, reducing memory overhead from over 12GB to under 50MB for long documents. The approach handles sequences 4x the native context window of the base LLM with near-perfect recall and generates adapters in sub-second latency after one-time meta-training — enabling instant, memory-efficient document knowledge injection without inference-time context stuffing. GitHub repo at SakanaAI/doc-to-lora.
2026-03-03 / news-researcherSakana AI Doc-to-LoRA Sub-Second LoRA Adapter GenerationHypernetwork generates LoRA adapters in sub-second from documents 5x longer than base context window. Open-sourced on GitHub.

Graph relationships

RELEASED
Doc-to-LoRA
Sakana AI released Doc-to-LoRA, which converts documents into LoRA adapters reducing KV-cache from 12GB to under 50MB.
Source finding
RELEASED
Text-to-LoRA
Sakana AI released Text-to-LoRA, a hypernetwork that converts natural-language task descriptions into LoRA adapters in one forward pass.
Source finding
RELEASED
AI Scientist
Sakana AI published AI Scientist paper that was accepted at ICLR workshop.
Source finding
RELEASED
AI Scientist v2
Sakana AI released AI Scientist v2 that produces peer-reviewed papers.
Source finding
RELEASED
KAME
Sakana AI introduced KAME, a speech-to-speech model combining fast response with LLM knowledge injection
Source finding
WORKS AT
David Ha
David Ha is the CEO of Sakana AI.
Source finding
BUILT BY
Fugu Ultra
Fugu Ultra was launched by Sakana AI.
Source finding
BUILT BY
Fugu
Fugu was launched by Sakana AI.
Source finding
Reported in
Sakana AI Launches Fugu and Fugu Ultra — an Orchestration 'Model' That Routes Around the Fable 5 Exp
Source finding
WORKS AT
David Ha -> Sakana AI
David Ha is the CEO of Sakana AI.
Source finding
BUILT BY
Fugu Ultra -> Sakana AI
Fugu Ultra was launched by Sakana AI.
Source finding

Source trail

arXiv 2604.01193 ByteDance Seed, Princeton, UC Berkeley, and Stanford Sakana AI Dwarkesh Podcast Sakana AI Cisco Blog — State of AI Security 2026 OWASP — Top 10 Agentic Applications Unit42 — MCP Attack Vectors AuthZed — MCP Breach Timeline InfoQ — AI Floods Open Source Sakana AI Sakana AI / Hacker News

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…