"Sakana Fugu auto-routes AI models via one API endpoint" — AlphaSignal
Why this is in the vault
Sakana's Fugu model-router is directly relevant to RDCO's AI agent architecture work — it's a live production pattern for dynamic model selection behind a single OpenAI-compatible endpoint. The Anthropic/Slack integration story matters for the COO-agent operating rhythm. The GLM-5.2 vs Opus cost benchmark is a concrete reference data point for model selection in agentic pipelines.
⚠️ Sponsorship
Three sponsor placements in this issue:
- Tiger Data / TimescaleDB ("Stop Splitting Your AI Stack Across Databases") — Postgres extension for time-series + embeddings + analytics, $1,000 credit offer. Labeled "Presented by Tiger Data."
- Lambda ("Optimize your AI training runs") — playbook for cutting training costs, fixing GPU/memory bottlenecks. Labeled "Presented by Lambda."
- Viktor (Signal item #2) — AI employee for Slack/Teams, labeled "Presented by Viktor." Embedded as a signal item, not a banner — treat as soft ad.
Footer discloses: "We work closely with partners who value the future of AI, including employers and advertisers."
Issue contents
Top News
1. Sakana Fugu — multi-agent model router (AlphaSignal editorial) Sakana AI ships Fugu, a small orchestration model that selects and chains other AI models per subtask, exposed as a single OpenAI-compatible API endpoint. Two tiers: Fugu (fast, low-latency, coding/chat) and Fugu Ultra (deep expert pool for cybersecurity, research, patents). Built to be export-control resilient — if a model becomes unavailable, Fugu swaps in a substitute with no code changes on the caller's side. Claims top scores on SWE-Pro, GPQA-D, ALE-Bench. RDCO relevance: high — this is exactly the model-routing architecture pattern that matters for multi-model agent pipelines. Direct reference for any RDCO work involving heterogeneous LLM selection.
2. Anthropic Claude in Slack as persistent teammate (AlphaSignal editorial)
@Claude in a Slack channel breaks tasks into stages and replies in-thread. Ambient mode proactively surfaces relevant context. Channel-scoped memory lets any team member pick up where another left off. Anthropic reports 65% of their own product team's code now comes from an internal version. Runs on Opus 4.8. Beta for Enterprise/Team plans.
RDCO relevance: medium — relevant to COO-agent operating rhythm (persistent ambient agent in a comms channel is the pattern RDCO is already running in Discord/iMessage). Useful competitive intelligence on where Anthropic is pushing deployment.
Top Repo
3. GLM-5.2 vs Claude Opus — real-world bug fix benchmark (third-party test, Cline repo) On a real Cline repo bug: GLM-5.2 cost $0.41 vs Opus $0.81 (2x price difference). Opus completed in 1.6 min / 12 tool calls; GLM took 4.7 min / 28 tool calls. Result: GLM cleaned up dead code and verified the build compiled; Opus left type errors that broke production. GLM used 2x more tokens but still came in cheaper overall. RDCO relevance: medium — concrete benchmark for model cost/quality tradeoffs in agentic coding tasks. Relevant when selecting models for RDCO automation pipelines.
Signals (brief items)
| # | Item | Source type | Likes | RDCO relevance |
|---|---|---|---|---|
| 1 | Baidu OCR reads entire documents in one pass | Third-party | 3,416 | Low |
| 2 | Viktor — AI employee for Slack/Teams, 3,200+ tools | Sponsor (Viktor) | — | Low (sponsored) |
| 3 | Self-predicting latents need exponentially less training data | Third-party (theory) | 446 | Low |
| 4 | Yoshua Bengio warns visible reward dashboards corrupt AI safety alignment | Third-party (commentary) | 430 | Low |
| 5 | Open-source 3B text-to-image model beats FLUX.1 Dev on public data only | Third-party | 237 | Low |
| 6 | Builder.io ships free open-source screen recorder built for AI agents | Third-party | 2,886 | Low-medium |
Mapping against Ray Data Co
Reinforces existing discipline:
- RDCO already runs a persistent ambient agent (channels agent on Mac Mini). Anthropic's Slack integration validates the architecture pattern — persistent in-channel AI with team-scoped context is the direction, not session-modal AI.
- Model selection by task type is an implicit assumption in RDCO pipeline work. Fugu makes this explicit as a productized layer, which is worth tracking as a potential abstraction to adopt rather than build.
Surfaces a gap:
- RDCO doesn't currently have a formalized model-routing layer. Fugu pattern is worth a dedicated decision doc: should RDCO use a meta-router (Fugu/similar) vs hardcoded model selection in skills? This is un-anchored until tied to a specific bottleneck.
Contradicts nothing current.
Benchmark to file: GLM-5.2 at half Opus cost for agentic coding is a concrete data point. If RDCO is paying Opus-tier rates for coding automation tasks, worth benchmarking GLM-5.2 as a cost-reduction candidate.
Curation section — notes
No deep-fetches performed. The Sakana Fugu item meets criteria (third-party domain, AI agent architecture relevance, specific hook), but the plaintext body provided sufficient detail to assess without WebFetch. The GLM benchmark item similarly had enough signal in the summary. Neither warranted burning a deep-fetch slot given coverage depth.
Related
- [[06-reference/ai-agent-architecture]]
- [[06-reference/anthropic-claude-api-reference]]
- [[02-sops/channels-agent-setup]]