"Cognition $1B Raise, Google Universal Embeddings, C++ Body Track" — @Lior Alexander

Why this is in the vault

Three tier-1 industry signals in one issue that directly inform RDCO's agent-deployer positioning and infrastructure choices:

1. Cognition $1B at $26B valuation — Devin now writes 89% of Cognition's own code; ARR grew $37M → $492M in twelve months. This is the macro pricing/positioning anchor for our agent-deployer narrative. The market is paying ~52x ARR for "agent that ships PRs autonomously."
2. Google Gemini Embedding 2 (one model for text/audio/video/image) — potential infrastructure shift for the vault retrieval stack (qmd, graph-ingest). If we move off text-only embeddings, the consolidation is meaningful for our cross-modal vault workflow.
3. Qwen3 8B at 20 tok/s on a 10-year-old GPU — local-model viability for sub-agent fan-out cost reduction. Relevant to the COO-agent unhobble track.

⚠️ Sponsorship

Three sponsored placements in this issue, all clearly labeled by AlphaSignal:

• Buildkite (Top News slot 2 — "How frontier AI labs ship: it starts at CI") — CI-platform paid placement. Name-drops OpenAI / Anthropic / Mistral / Cohere / vLLM / HuggingFace as customers.
• Slack (Top Model slot 2 — "The Context Opportunity: Agentic AI at Scale") — Salesforce-owned, pitching Slack as the enterprise context layer for agents.
• Braintrust (Signals item 2 — "evaluate multi-turn conversations") — eval-platform paid placement.

None of these are novel sponsor relationships for AlphaSignal — Buildkite and Slack are recurring. Treat the editorial items (Cognition raise, SAM3DBody, Gemini Embedding 2, Qwen3, KAIST optimizer, OpenBMB, Crawl4AI, micrograd) as the non-sponsored signal.

Issue contents

Top News

• Cognition $1B raise — $26B valuation (2.5x in 8 months from $10.2B); ARR $37M → $492M; Devin writes 89% of Cognition's own code (up from 13% in Dec); new SWE-1.6 model at 950 tok/s in Windsurf; Mercedes-Benz did 8-month legacy modernization in 8 days using Devin. (Editorial.)
• Buildkite CI — Sponsored.

Top Model

• SAM3DBody-cpp — Pure C++ engine wrapping Meta's SAM 3D Body model. Outputs 70 joint positions + full 3D mesh in real time. Lightweight C interface, callable from any language. Use cases: robot perception, motion capture without expensive hardware, AR/games with full-body tracking. Models on HuggingFace, code on GitHub. (Editorial.)
• Slack context layer — Sponsored.

Top Paper

• Gemini Embedding 2 — Single unified embedding model for text, audio, video, image. Tops benchmarks on image retrieval, video search, multilingual text, code. Out-of-the-box performance on niche untrained domains (astronomy, culinary). Beats Google's previous text-only models even on pure text tasks. Available on Gemini API + Vertex AI. Enables cross-modal RAG with any media type. (Editorial.)

Signals

1. Qwen3 8B coding agent at 20 tok/s on a 10-year-old GPU — Local-model viability signal. (Editorial.)
2. Braintrust multi-turn eval — Sponsored.
3. Crawl4AI — Open-source LLM-tuned web scraper, 66.5K GitHub stars. (Editorial.)
4. KAIST 1.5x training optimizer — No learning rate schedule needed. (Editorial.)
5. OpenBMB $100K contest — Push 9B model to 1M-token inference, 9.1K stars on repo. (Editorial.)
6. Micrograd 150 LOC neural net — Karpathy's classic, 16K stars. (Editorial.)

Mapping against Ray Data Co

Cognition raise — strongest mapping. Anchor data point for agent-deployer positioning.

• 52x ARR multiple ($26B / $492M) for "agent that opens PRs autonomously" is the market's current pricing for production-tier coding-agent. Anchors our COO-agent value-prop framing: if Devin's wedge is "PR-opening engineer that never sleeps," ours is "ops/strategy operator that never sleeps." The Mercedes 8mo→8d case study is the kind of asymmetric-outcome story we should track for our own narrative arsenal.
• 89% of Cognition's code written by Devin = self-eating credibility signal. Worth borrowing for our own RDCO positioning: how much of RDCO's operational throughput is Ray-COO-authored? (Most of the vault, most of the skills, most of the iterations. The eat-our-own-cooking narrative is available to us if we want it.)
• File against 01-projects/agent-deployer/ and 01-projects/coo-agent/. Consider Sanity Check angle: "What the Cognition raise tells you about agent pricing power" — but only if there's an original re-frame, not a restate.

Gemini Embedding 2 — medium mapping. Watch, don't act yet.

• qmd currently runs on text embeddings only. If we ever want the vault to retrieve from screenshots, voice notes, or video clips (e.g., founder voice-memos, Discord image dumps, Loom recordings), GE2 is the obvious migration target. Not urgent — but worth a "what would it cost to swap embedding providers" note in the qmd backlog.
• Cross-reference: 02-sops/ likely has nothing on multi-modal retrieval today. Gap.

Qwen3 8B at 20 tok/s — medium mapping. Sub-agent cost lever.

• If a 10-year-old GPU can run a full coding agent at 20 tok/s, the implication is local-model sub-agents become realistic for the high-volume low-stakes parts of the COO-agent fan-out (newsletter processing, YouTube transcript summarization, vault hygiene). Not a near-term swap — Anthropic models still win on judgment-heavy work — but it changes the math on "should we be paying API rates for every subagent."
• File against the budget-control thread (feedback_api_cost_budget_controlled).

SAM3DBody-cpp, KAIST optimizer, OpenBMB contest, Crawl4AI, micrograd — weak/skip. Interesting AI/ML news but not load-bearing for RDCO's current bets.

Curation section — notes

• Cognition / Devin (editorial, third-party) — Anchor signal. No deep-fetch needed; AlphaSignal blurb has the load-bearing numbers ($1B / $26B / $37M→$492M / 89% / 950 tok/s / 8mo→8d). Cross-check: TechCrunch and The Information likely have primary-source coverage if we want to verify the ARR claim.
• Buildkite (sponsored, third-party) — Skip. Standard CI sponsor placement.
• SAM3DBody-cpp (editorial, third-party) — Skim. Cool tech demo, no RDCO relevance.
• Slack context layer (sponsored, third-party) — Skip. Salesforce sales pitch.
• Gemini Embedding 2 (editorial, third-party) — File. Track for qmd / graph-ingest migration consideration. Worth a follow-up vault note once we see real-world cross-modal benchmarks vs OpenAI's text-embedding-3-large / Voyage / Cohere embed-v4.
• Qwen3 8B 20 tok/s (editorial, third-party) — Skim. Cost-lever signal for sub-agent fan-out.
• Braintrust eval (sponsored, third-party) — Skip. Eval-platform pitch.
• Crawl4AI (editorial, third-party) — Skim. Useful tool if we ever need LLM-tuned scraping; right now WebFetch handles our needs.
• KAIST optimizer (editorial, third-party) — Skip. Researcher signal, not operator-relevant.
• OpenBMB $100K contest (editorial, third-party) — Skip.
• Micrograd (editorial, third-party) — Skip. Karpathy classic; well-known.

Related

• [[~/rdco-vault/01-projects/coo-agent/]] — COO-agent unhobble track; Cognition raise anchors pricing/positioning
• [[~/rdco-vault/01-projects/agent-deployer/]] — agent-deployer pricing reference
• [[feedback_api_cost_budget_controlled]] — local-model viability (Qwen3 8B) affects sub-agent cost math
• [[feedback_advisor_not_pair_programmer]] — Devin's wedge (PR-opening engineer) is what Cognition raised against; our wedge is ops/strategy operator, distinct
• [[feedback_listen_and_injection_caution]] — multi-sponsor issue, three placements clearly labeled, no embedded-instruction risk