06-reference

not boring wdoo 202

2026-07-17·reference·source: Not Boring·by Packy McCormick

"Weekly Dose of Optimism #202" — @packyM

Why this is in the vault

WDoO #202 is a "pairs" issue — Packy explicitly frames it as two-of-everything: two Saronic stories, two Austin stories, two robotics stories, two open-weight model stories. The AI-models pair (item 5) is the primary RDCO signal: both Kimi K3 and Inkling dropped this week, and together they represent the first open-weight models competing in the frontier band with Fable 5 and GPT-5.6 Sol. The rest of the issue (defense drones, energy storage, construction robots, humanoids, pharma) is high-quality optimism but tangential to RDCO's core stack.

⚠️ Sponsor / bias notes

Self-sponsor (not a paid external sponsor): "Today's Weekly Dose is brought to you by… Not Boring." This is a paid-subscriber upsell block, not a third-party sponsor placement. No external company paid for this issue's top slot.

Undisclosed promotional mention — Ramp (item 5): The section closes with a redirected link to "Ramp's new AI Token Spend Management tool." Ramp has served as a Not Boring paid sponsor in past issues; this appears to be an editorial/affiliate placement without a sponsor disclosure. Treat Ramp coverage in this WDoO as potentially non-neutral.

Portfolio disclosure — Unnatural Products (item 4): Packy explicitly flags "not boring capital portfolio company Unnatural Products" in the Lipfendra section. The $220M Merck/Unnatural Products deal is the structural reason Lipfendra gets covered. Disclosure is present but embedded in editorial, not headlined.

Unverified — Saronic, Base Power, TerraFirma, Walden Robotics, Sunday Robotics, Moonshot AI, Thinking Machines: No Not Boring Capital disclosures present for these companies. WDoO standard caveat applies: Packy does not always surface angel positions in curation. Enthusiasm ≠ neutral, but no relationship confirmed.

Issue contents

(1) Saronic's Very Big Week [defense-tech / autonomous systems / unverified]

Two Saronic events in one week: (a) three 24-foot Corsair autonomous boats used in combat for the first time — striking an Iranian submarine and ship-maintenance facility at Bandar Abbas; and (b) Saronic announced "Port Alpha," a $3B+ greenfield shipyard in Brownsville, TX (835 acres, expandable to 4,400). The shipyard will build crewed and uncrewed military vessels, container ships, tankers, and icebreakers using the same autonomy and advanced-manufacturing stack as the drone boats. Operations targeted for 2028; up to 10,000 direct jobs. The site was previously destined for California Forever's plan — California lost it to TX after failing to pass permitting legislation.

Key signal: a four-year-old startup went from prototype to live combat deployment in months; then announced a capital program usually reserved for legacy industrial incumbents.

(2) Austin Week: Base Power + TerraFirma [energy storage / construction robotics / unverified]

Base Power Company begins operating inside Austin Energy's service territory under a 40-megawatt residential battery agreement: Base installs and maintains batteries; Austin Energy can dispatch the fleet during demand/price spikes; homeowners get whole-home backup. Win-Win-Win structure.

TerraFirma (SpaceX vets, "Building giant robots to build a brighter future") raised $100M Kleiner Perkins-led Series A for robotic construction tech — targeting order-of-magnitude improvements in construction efficiency. Cap table includes Anduril's Matt Grimm, Hadrian's Chris Power, and Base's Justin Lopas.

(3) Walden Robotics + Sunday Robotics [humanoids / factory + home robotics / unverified]

Walden Robotics emerged from stealth with $300M in seed funding at a $1.1B valuation (co-led by Deviation Capital + Toyota Motor Corp, Toyota Invention Partners, Toyota Ventures). Spun out of Toyota Research Institute in January by CEO Russ Tedrake (MIT professor, former TRI robotics + ML lead); robots were in production on a Toyota factory floor within two months of founding. Full-stack approach: hardware, software, models, and deployment application layer. Learn via demonstrations, improve through practice.

Sunday Robotics introduced ACT-2 Preview for its home robot Memo: learns a new behavior from a single fine-tuning example, generalizes zero-shot to unseen homes, claims 99% success rate. Introduced "Solves" — a framing to contextualize robotics demos: "A Solve declares two boundaries: scope and adaptation cost. Without both, 99% has no context." The general recipe: scale pretraining, then hill-climb with minimal in-house data.

(4) Lipfendra FDA Approval [pharma / cardiovascular / NBC-portfolio disclosed]

FDA approved Merck's Lipfendra — first oral PCSK9 inhibitor. Once-daily pill achieves injectable PCSK9-level efficacy: 56% placebo-adjusted LDL reduction (broad trial), 59% in familial hypercholesterolemia. Mechanism: macrocyclic peptide folded to survive the gut, enabling antibody-like potency in oral form. Price: ~$10.50/day.

NBC portfolio disclosure: Not Boring Capital portfolio company Unnatural Products signed a $220M "biobucks" deal with Merck in 2024 around macrocyclic peptide drug discovery. Lipfendra is a macrocyclic peptide. The deal is the structural reason this item appears in the Dose; disclosure is present but embedded.

Context: joins Eli Lilly's PCSK9-targeting gene therapy (covered WDoO #195) as part of a broader attack on LDL cholesterol's dominance in cardiovascular disease.

(5) Kimi K3 + Inkling: Open-Weight Models Hit the Frontier [AI models / open-weight / agentic / most RDCO-relevant]

Kimi K3 (Moonshot AI): First open-weight model in the 3-trillion-parameter class. 2.8T total params, 1M-token context, native vision. Sparse: activates 16 of 896 experts at inference. New architecture: Kimi Delta Attention (6.3× faster decoding at long context), Attention Residuals (25% higher training efficiency at <2% cost increase). Claims 2.5× scaling efficiency improvement over K2. Moonshot says K3 still trails Fable 5 and GPT-5.6 Sol overall, but wins specific agentic/coding benchmarks: 91.2 on BrowseComp vs. Fable's 88.0; 88.3 on Terminal Bench 2.1 vs. Fable/Opus 4.8's 84.6. Weights release July 27.

Inkling (Thinking Machines Lab / Mira Murati): 975B-parameter MoE (41B active), Apache 2.0 license. Accepts text, images, audio; 1M-token context; fine-tunable via Thinking Machines' Tinker platform. Not the best model overall — positioned as a strong, broad base for domain-specific fine-tuning. Demo: the team had Inkling write and run its own fine-tuning job, evaluate the weights, and switch itself to a version that never uses the letter "e." Architecture draws on DeepSeek-V3; post-training used data generated by Moonshot's Kimi K2.5. Cross-lab compounding in action.

Ramp plug (undisclosed): Packy closes item 5 with a link to Ramp's AI Token Spend Management tool — no sponsor disclosure. Ramp has served as a Not Boring paid sponsor previously; treat this as a potentially non-neutral editorial placement.

Extra Doses [paywalled — science breakthroughs, space tissues, Senra]

Science Breakthroughs from Ulkar Aghayeva and additional items (Space Tissues, Senra) are behind the paid-subscriber wall. Only teaser line available: "EXTRA DOSES: Science Breakthroughs, Space Tissues, Senra..."

Mapping against Ray Data Co

Lead connection: The Kimi K3 + Inkling releases directly shift RDCO's model-selection calculus for the Claude Code agent toolchain and phData client work. Inkling's Apache 2.0 license + Tinker fine-tuning means RDCO can now access a frontier-competitive base model without closed-API lock-in — and customize it for specific agent behaviors (the self-fine-tuning demo is exactly the kind of agentic loop that the always-on agent harness runs). Kimi K3's 91.2 BrowseComp score besting Fable 5 at 88.0 is the first time an open model has demonstrably outperformed a closed frontier model on an agentic task benchmark. This matters for the "value accrues to companies building on AI, not to the labs" thesis from [[2026-07-12-innermost-loop-singularity-yeshiva-agent-layer-war]].

Secondary: Walden Robotics' "full-stack + learn-from-demos" architecture mirrors the agent-layer pattern Ray is building for data workflows — the insight that you need hardware + software + models + deployment layer to deliver reliable autonomous output applies to AI data agents as much as factory robots.

Personal (not RDCO): Lipfendra's oral PCSK9 inhibitor is worth tracking given Ray's MASLD profile and cardiovascular risk factors — but this is a personal health signal, not a business one.

What this doesn't change: Defense tech (Saronic), energy storage (Base Power), and construction robots (TerraFirma) are exciting but don't intersect RDCO's current bets or phData pipeline.

Related