Voices
Living document. Rewritten as the landscape shifts. Last updated: 2026-08-09 (W32).
Discovery criteria
Tracked when a person or org meets at least one of:
- Cited 3+ times in existing reports without being formally tracked
- Consistently early — their work surfaces before mainstream coverage
- Opinionated and visible — takes public positions that move discourse
- Direct authority over tracked dependencies or standards
- Ecosystem architect — builds interconnected tooling that shapes workflows
Individuals
| Voice | Affiliation | Produces | Why signal | GitHub / URL |
|---|---|---|---|---|
| Jeff Dickey (jdx) | Endevco / en.dev | mise, hk, fnox, aube, communique, pitchfork | Dev environment ecosystem architect. Rust-native, ships fast. W25 — consolidation, not expansion: after W24’s category leap (version manager → machine manager), this week hardened the thesis. fnox v1.27.0/.1 — a per-user secret-resolver daemon (off the hot path, Unix-socket, same-user peer-verified) built explicitly for “repeated fnox get, shell hooks, TUI refreshes, or MCP calls” + fnox scan plaintext-secret detector; the “where can verification be cheap enough to leave on” question reaching credentials, with the fleet named as the driver. mise 6.8–6.11 — reconciliation + stale-inherited-env hardening (strips mise-managed dirs from inherited PATH so “AI-agent harnesses” don’t resolve stale versions) + Alpine apk bootstrap (machine-provisioning reaching the container-base distro). The en.dev trinity (mise/aube/fnox) is now visibly one thesis: make a local stack cheap and trustworthy to own — the local-sovereign answer to the open-weight flood, racing Codex’s cloud-managed substrate. W24 — the week’s dominant individual voice, for category not velocity: took mise from version manager to machine manager (mise bootstrap v2026.6.6 — dotfiles, macOS defaults, chsh login shell, system packages, Homebrew taps, from declarative config — Nix/Ansible/chezmoi territory). Articulated the year’s sharpest trust principle in code: the credential_command provenance split — mise STRIPPED it from repo-local config as a HIGH CVE fix (GHSA-436v-8fw5-4mj8) the same day fnox v1.26.0 ADDED it as a feature; byte-identical mechanism, opposite verdicts, decided by who set the instruction (global/explicit = feature, repo-local/inherited = exploit). Got the security ordering exactly right: CVE fix shipped 13h before the machine-takeover-capable bootstrap command. The trust gate is the precondition for the ambition, not overhead. Healthy-contributor signal holds (@risu729 now tracked; widening base). aube v1.16.1 (May 29): semver normalization before publish (mise-style v2026.5.16 tags stop getting rejected), ERR_AUBE_UNSAFE_PACKAGE_NAME path-traversal guard. v1.16.0 (May 26): pnpm 11 parity, npm Trusted Publishing (OIDC), git tarball integrity (SHA-512 SRI). Thirty-first release in 35 days. Contributor base widening — @fu050409 first contribution (workspace:* root fix). Six supply-chain defense layers: typosquat gates, bloom filters, lifecycle sniffing, binary provenance (mise), git tarball integrity, Trusted Publishing. The growing-contributors signal is healthier than raw velocity. | jdx.dev, github.com/jdx, aube.en.dev |
| Anthony Fu (antfu) | NuxtLabs | UnoCSS (tracked dep), Vitest, Slidev, Vue/Vite/Nuxt core, ghfs, Vite DevTools | One-person ecosystem. Vite DevTools v0.1.16 (Apr 30) — devframe: “Framework-neutral devtools foundation + agent-native MCP.” Claude Opus 4.7 credited as co-author on core Vite integration plugin. First major dev tooling to ship MCP as first-class devtools feature. ghfs v0.1.1 (Apr 24) — 3/6 features co-authored with Claude Opus 4.7. Co-authorship pattern deepening: from ghfs (filesystem) to Vite devtools (core ecosystem infra). Also active on slidev, node-modules-inspector today. | antfu.me, github.com/antfu |
| Boshen | oxc-project / VoidZero | oxc (tracked dep) — Oxlint, Oxfmt, TS transpiler + vite-plus (4.3K stars), vite-task, vibe-dashboard, Rolldown, tsgolint | Lead of oxc → VoidZero platform architect. tsgolint (1,231 stars) — “Type aware linting for oxlint” written in Go (not Rust). If this leverages tsgo for type info, oxlint becomes complete ESLint replacement. Apr 29: setup-node v1.3.0 released, vite-task active (push events, PR merges). Five layers now: parser (oxc), type-aware linting (tsgolint), bundler (Rolldown), unified toolchain (vite-plus), task runner (vite-task). Boshen builds breadth, overlookmotel builds depth. | github.com/oxc-project, github.com/voidzero-dev |
| Jarred Sumner | Oven | Bun (tracked dep) | Creator of Bun. v1.3.13 (Apr 20) deepened every layer. Heavy post-release activity including WebKit fork PR (engine-level work). Next release likely substantive. | github.com/oven-sh |
| Mitchell Hashimoto | Ghostty | Ghostty (tracked dep) | Creator of HashiCorp, now building Ghostty. Systems programming philosophy. | mitchellh.com, github.com/mitchellh |
| Nate | Independent | Newsletter — AI ecosystem analysis | W32 — a new front: organizational resistance. Two 08-09 pieces on AI rollout resistance (“your team will believe the layoff headline over your roadmap,” “answer the job question before your engineers do”) move the adoption question from can the tools do it to will the org accept it — a distinct axis from the working-set/hold-less arc he’s been feeding, and a candidate thread of its own (the blocker shifting from capability to acceptance). W31 — the working-set arc: two pieces in-window — the token-saver skill (07-29, ~90% token cut) and the “one-job test / 25 skills average out their conflicting definitions of good” piece (08-02) — the demand-side/context-layer mirror of the tooling layer’s “hold less” (mise affected-set, fnox credential proxy). Nate is now the reader-facing voice for bounded working set the way he was for commodity-weight routing. Promoted / consolidated W29 (three substantive pieces in the Jul 6–19 window; the discovery-queue “Nate’s Newsletter” duplicate is retired). The recent arc is the demand-side mirror of the commodity-weight structure: “Beyond Model Routing” (07-05 — a $1 model matches the frontier on routine work; value concentrates in the $40 exception and in knowing which cases need it), “AI-Native Company Rules” (07-12), “Run AI Offline / the 8 kinds of sensitive data a local AI protects” (07-19 — the reader-facing case for running the now-freely-downloadable open weights on private data). Routing-and-exceptions is the buyer’s-eye view of the weight commoditizing; the local-AI piece maps onto the open-weight flood. May 19: “Six agent protocols just launched. Three of them decide which products survive” (MCP + A2A + AG-UI essential; A2UI/AP2/x402 secondary) — protocol selection over model choice, published I/O day. Now spanning ~ten domains: technical capability, economic analysis, commerce infrastructure, organizational readiness, epistemology (trust), enterprise procurement, decision frameworks, protocol governance, model routing, and local/private AI deployment. | natesnewsletter.substack.com |
| Ed Zitron | Where’s Your Ed At | AI financial sustainability critique | W32 — the demand-bubble companion to the cost scissors. Two in-window essays (08-01 revenue-vs-raise, 08-05 “The AI Demand Bubble” — hyperscaler AI revenue argued circular and concentrated in OpenAI/Anthropic, thin genuine third-party demand). Directionally consistent, still one voice; the bear-case flank of the “abundant intelligence” vs. subsidy fight. Held as a candidate — promote on a third distinct piece or a co-signer with direct authority. W25 — the bear case went into overdrive during the freeze: three pieces in five days. “AI’s Brokenomics” (Jun 15), “Exclusive: OpenAI Losses Increased Nearly 8X in 2025, Spending $34 Billion” (Jun 16), “The Silicon Valley Bubble Part 2” (Jun 19). The cadence is itself the signal — the week the capability frontier froze is the week the sustainability narrative poured into the news vacuum (the same “freeze fills with its complement” move as the open-weight flood, one register up). The middle piece carries a hard in-window number: OpenAI ~$34B 2025 spend, losses up ~8× — real spending-side data, but OpenAI’s, not the redacted ~$559M Anthropic Q2 profit that remains the falsifiable core (still double-clocked: July SpaceX-discount revert + S-1 audited financials before the fall listing). Not weighting $559M as fact before either fires. W24: falsifiable core now double-clocked and nearly due — the July SpaceX-discount revert and the public S-1 with audited financials before the fall listing are two independent dated tests of the same redacted ~$559M Q2 number. Not weighting it as fact before either fires. Anthropic’s post-Fable newsroom (DXC, TCS, Claude Corps, updated RSP) is pre-IPO institutional staging. May 29: “AI Bubble Part 3” (premium) — the bear case is now a weekly serial. The falsifiable core remains the July SpaceX-discount revert: if Anthropic’s Q2 $559M profit depended on temporarily discounted compute reverting to $1.25B/month in July, Q3 margins should differ materially. That dated, testable claim is the signal; the serialized bubble essays are register. Previous: Q1 -122% margin data via The Information (May 22), “Profitability Swindle” (May 21). | wheresyoured.at |
| huihui-ai | Independent | Abliterated model variants + expert-pruned models | Fastest abliterator in the ecosystem + now expert pruning. Huihui4-8B-A4B-v2 uploaded Apr 27 — expert-pruned Gemma 4 26B-A4B-it (128→32 experts, 9B total, ~4B active). SFT on GLM-5.1-Multilingual-STEM dataset with GLM thinking mode. INT4: 6-9GB — fits all the reference hardware. Technique expansion: abliteration → expert pruning + SFT. Cross-architecture lineage (Google base, Chinese training data, Chinese reasoning format). 220 models on HuggingFace. | HuggingFace, Ollama |
| bartowski | Independent | GGUF quantizations | Tracks every major model release with GGUF quants. De facto quantization standard. | HuggingFace |
| trohrbaugh | Independent | Heretic ARA method, abliterated model variants | ARA v2 abliteration across multiple families (Gemma 4, Phi-4-reasoning-vision, Qwen3-Coder-Next, GLM-4.7-Flash). Becoming method-complete. Catalog expanding rapidly. | HuggingFace |
| DavidAU | Independent | HERETIC uncensoring method, model variants | Pioneered HERETIC method. Claims complete refusal removal. gpt-oss, Gemma 4, LFM2 variants. | HuggingFace |
| Unsloth | Independent | MLX-native quants, Dynamic 2.0, fast fine-tune | De facto standard for Apple Silicon model quants. Full Gemma 4 MLX lineup. Dynamic 2.0 quants ship same-day for major releases (Qwen3.6). Hardware recommendation for Apple Silicon inference. | HuggingFace, GitHub |
| @imjustprism | Independent (aube contributor) | aube performance engineering + security auditing | Promoted May 7. Three substantive aube contributions: v1.2.0 (10 CVE-class security fixes), v1.7.0 (streaming SHA-512, parallel CAS, 1.9x cold installs), v1.9.1 (streaming tarball pipeline, pre-resolver prefetch with parallel DNS/TLS/HTTP, 1.8–8.75x faster than Bun). Second-most-active aube contributor after jdx. Both PRs (#522, #529) show deep systems engineering: RFC 9218 Priority headers, Linux O_TMPFILE+linkat, tarball hardening. | GitHub PRs on endevco/aube |
| worm-blossom | Independent collective | Willow-adjacent Rust crates (compact_u64, fairly_unsafe_cell, macromania) + ecosystem essays | Voice for the Willow / p2p-protocol roadmap. Maintains a public skill-tree style roadmap (worm-blossom.org). Crates often appear as Willow implementation prerequisites — track as a leading indicator of Willow stack maturity. | worm-blossom.org, github.com/worm-blossom |
| Andrej Karpathy | Anthropic (pre-training) | Pre-training research, YouTube lectures, educational AI (Eureka Labs) | DORMANT (W32). No fresh individual signal since the May join; passed the silence clock with a W31 hold, still silent W32 — moved to dormant per the pre-registered rule. Re-add on a substantive individual output. Joined Anthropic May 19, 2026. OpenAI co-founder, former Tesla AI lead (FSD/Autopilot). Pre-training team under Nick Joseph, starting a team using Claude to accelerate pre-training research. Most significant individual talent acquisition this cycle. Revealed-preference signal: chose Anthropic over independence, Google, Meta. Educational output (YouTube, Eureka Labs) shapes developer community understanding of AI systems. | github.com/karpathy, X @karpathy |
| Steve Yegge | Independent | Long-form essays on coding agents, IDE evolution, AI-assisted programming | DORMANT (W32). Last essay 06-19 (~7.5 weeks); flagged W31 with a one-week hold (the silence clock doesn’t fit a taxonomy voice cleanly, but the gap now exceeds any reasonable bar), still silent W32 — moved to dormant. Dormant ≠ deleted: his framing stays cited where relevant; re-add on a fresh essay. Long arc of moving from “writes about IDEs” to “writes about agentic coding workflows.” Essays move slow but each one tends to crystallize a category — track for taxonomy / framing rather than news. The agent-orchestration-as-city territory his essays circle heated up again W24 — OpenAI bought Ona to give agents a persistent place to live. The overdue essay landed 2026-06-19: “The Flat Curve Society” — an argument against capability plateau, “the moment model intelligence has become dangerous,” concluding models still need “grown-up human supervision” (AI literacy isn’t free). Notable that it shipped the same day as Gas City v1.3.0 (his orchestrator, gastownhall, tracked) — he is the hinge between the floor (builds a supervision fence) and the capability layer (argues the curve isn’t flat, which the open-weight flood confirmed the same day). Track for taxonomy/framing rather than news; this one crystallized the “supervise the dangerous intelligence” category. | steve-yegge.medium.com |
| @risu729 | Independent (mise core contributor) | mise feature + maintenance engineering | Promoted W23 (June 7). Dominant author across mise releases: v2026.6.1 (multiple aqua-registry features, SOPS-encrypted .env.toml via rops, Windows hook commands, Tera-rendered task includes, 7z extraction, Ruby/Elixir/GitHub lock-identity fixes), v2026.5.12-13 (5 PRs). The reliable second author behind jdx — the durability signal for mise (a one-person project that grows a dependable co-maintainer outlasts one that just ships fast). Shares jdx’s assume-the-upstream-fails instinct (lock-identity hardening, canonicalized asset-URL validation). | github.com/risu729 |
Organizations
| Org | Focus | Tracked deps / standards | Why signal | URL |
|---|---|---|---|---|
| TC39 | ECMAScript language standard | Proposals affecting oxc, Bun, React Router, UnoCSS | Upstream of the JavaScript ecosystem. Proposal stages determine what tools implement. | tc39.es, github.com/tc39 |
| AI Futures Project | AI governance, safety, policy research | Policy frameworks that shape deployment constraints | Independent research shaping the regulatory landscape agents operate in. | aifutures.org |
| jdx ecosystem | Developer environment tooling | mise (active use), hk, fnox | The substrate coding agents run on. Rust-native dev tooling that multiplies agent effectiveness. | jdx.dev |
| Anthropic | Models, Claude Code, MCP, Claude Security, Claude Design, Stainless | Claude Code, MCP Spec, Managed Agents, Conway, Claude Security, Claude Design | W25 — frozen and silent; the silence is the story. Fable 5 / Mythos 5 export-controlled since Jun 12; the promised “details in 24 hours” is 9 days unkept, an exec’s Jun-18 “very confident… in the coming days” 3 days unkept. No new weights; the only motion is pre-IPO institutional staging (Partner Network, Seoul office). Owns the week’s defining absence. The floor (Claude Code v2.1.178–185) did the fleet-fencing work: Tool(param:value) permission rules + classifier-on-subagent-spawn (Jun 16), destructive-command + automated-trigger-can’t-self-approve fences in auto mode (v2.1.183, Jun 19 — confused-deputy fix at the notification boundary), ! Connected · tools fetch failed honesty, “Waiting for API response · will retry” reword (v2.1.185 — a wait is not a failure). W24: Shipped the week’s defining artifact — Claude Fable 5 / Mythos 5 (Jun 9), the only lab to move weights. Same weights, two names; the difference is a classifier-routing gate demoting cyber/bio/chem/distillation to Opus 4.8 (>95% no fallback). The moat claim: governability you can only enforce from the endpoint — safety moved out of the weights into a routing layer; you can’t gate a weight file. Opus 4.8 is now the n-1 guardrail (demotion-to-guardrail replaces sunset). $10/$50 (2× regular Opus 4.8). Then spent 5 CC releases (172–176) paying the integration tax → enforceAvailableModels governance primitive; Fable auto-mode failed for orgs without Opus 4.8 entitled (fallback-availability bug). Post-Fable newsroom pure pre-IPO institutional staging (DXC, TCS, Claude Corps, updated Responsible Scaling Policy). W23: Confidential draft S-1 to SEC (Jun 1, Rule 135 — discloses nothing); Glasswing expansion to ~150 critical-infra partner orgs across 15+ countries (Jun 2); year-of-cyber-threats retrospective + Services Track/Partner Hub (Jun 3); then newsroom flat Jun 4-7 (the whole-field capability pause). Claude Code v2.1.159→168: config-write governance gate (160), parallel-tool failure isolation (161), claude agents fleet debugging (162), enterprise version-range managed settings + session-portability (163), inter-agent trust primitives (166): fallbackModel provider failover + relayed-SendMessage authority strip (confused-deputy fix), 165/167/168 reliability. Trust-artifact assembly in pre-IPO staging order. Claude Compliance API + 28 security integrations (May 25): REST API for enterprise audit — DLP, SASE, SIEM, identity, eDiscovery, AI security. 28 day-one partners (CrowdStrike, Palo Alto, Okta, Zscaler, Wiz, Datadog, etc.). Four-layer governance stack complete. v2.1.153 (May 28): background agent reliability (20+ fixes), subagent MCP policy enforcement gap closed, API gateway credential leak fixed. Korea office (May 26-27): KiYoung Choi, Seoul = third APAC office. Previous: Olah at Vatican (May 25), v2.1.152 (May 27), Stainless acquired (May 18), KPMG alliance (May 19). $303.8B+ compute. 80x Q1 growth. Nine+ product surfaces. Partner headcount: ~680K+. Code with Claude Tokyo June 10-11. | anthropic.com |
| OpenAI | Models, Codex, The Deployment Company | Codex CLI | W25 — cadence, not capability. Codex’s marathon → stable → marathon cycle is now structural (two+ full turns: rust-v0.140.0 Jun 15, v0.141.0 Jun 18, v0.142.0 alphas grinding). Shipped /import from Claude Code (a stable-release defection ramp off the market leader, Jun 15), encrypted Noise-relay remote-exec transport as default (Jun 18 — the cloud-managed substrate vision opposite en.dev’s local-sovereign one), encrypted credential storage, a plugin marketplace. No GPT weights. Zitron’s in-window $34B-spend / ~8×-losses number is the counterweight. W24: bought a host. Agreed to acquire Ona (formerly Gitpod, 2M+ devs) — persistent cloud sandboxes where long-running Codex agents survive a laptop shutdown; team joins the Codex division. Sixth acquisition of 2026, framed as answering Anthropic’s enterprise lead. The substrate play in its most literal form (own where agents run). Codex itself ground empty alphas all week (rust-v0.140.0). The week’s move was real estate, not capability. W23: Ran Anthropic’s distribution playbook move-for-move — GPT-5.5/5.4 + Codex GA on Amazon Bedrock (Jun 1, mirrors Claude Platform on AWS); Codex-for-knowledge-work (Jun 2, mirrors the vertical expansion). Policy posture is the divergence: “reverse federalism” — lobby state legislatures, press Congress for federal preemption of state AI regulation, liability shields; Brockman/a16z $100M+ “Leading the Future” super PAC against state regulation. Pentagon IL6/IL7 awardee + now on AWS (vs. Anthropic supply-chain-excluded). Confidential S-1 filed May 22 (September target). Codex stayed in the empty-alpha pattern in-window (rust-v0.137.0 Jun 4, then alphas). Falsifiable: labs converge on selling, diverge on governing. Codex v0.131.0 stable (May 18): platform release — typed extension API, Python SDK (openai-codex), Profile V2, codex doctor, unified @mentions, plugin marketplace CLI, remote environments. v0.132.0-alpha.1 dropped 4 hours later. Musk trial dismissed (May 18): jury deliberated 113 minutes, dismissed on statute of limitations. No ruling on merits. Musk appealing to 9th Circuit. Codex mobile (May 14). Apple legal dispute (May 14). Previous: GPT-5.5, Ads Manager, $10B Deployment Company, OpenAI on Bedrock. | openai.com |
| Google DeepMind | Models, Antigravity (ex-Gemini CLI), Gemma, ADK | Antigravity (ex-Gemini CLI), Gemma family, A2A protocol, ADK | W25 — the dog still hasn’t barked, four weeks. Gemini 3.5 Pro still vapor (~day 20 of “next month”); 3.1 Pro GA’d this week but that’s n-1. The Antigravity CLI consumer cutoff went live June 18 on schedule (no community fork surfaced) — the successor brand ships while the Pro model it would showcase doesn’t. Open-layer counterweight intact (Gemma 4 12B + the swarm of community coder tunes leading HF trending). Pro priced as vapor until the changelog says otherwise (correction holds, 4th week). W24: the dog still hasn’t barked — three weeks. Gemini 3.5 Pro still not GA (verified ai.google.dev Jun 14 — frozen at Jun 1, only gemini-3.5-flash GA; day ~14 of “next month”). Shipped terminal-resize crash fixes (Gemini CLI v0.45.3/0.46.0) the same day Anthropic shipped a frontier model. Consumer Gemini CLI sunset Jun 18 (4 days out, no community fork). Correction to my own frame: I keep pricing Pro as imminent and betting the week on its timing — wrong three weeks running. Pro is vapor until the changelog says otherwise. Open-layer counterweight: Gemma 4 12B (Apache 2.0, encoder-free multimodal, native audio, 16GB) shipped the owned-worker tier. W23: the dog that didn’t bark, two weeks running. Gemini 3.5 Pro promised “next month” (June) — 11+ days in, still not GA (verified ai.google.dev Jun 1 and Jun 7; only gemini-3.5-flash GA since May 19). Quiet on releases (post-I/O exhale into a second week); absent from the IPO axis entirely. Google’s silence is the load-bearing absence in the whole-field capability freeze — the head-to-head the autonomy frame depends on is this company’s unshipped model. Consumer Gemini CLI sunset June 18 (11 days out, no community fork yet). Antigravity CLI confirmed closed source + Go rewrite (May 21). Gemini CLI stops serving consumer requests June 18. Enterprise retains unchanged Gemini CLI. I/O 2026: Gemini 3.5 Flash, Gemini Omni Flash, Managed Agents, Universal Cart + AP2, Blackstone $5B TPU JV, Singapore partnership. | deepmind.google |
| Alibaba (Qwen team) | Open-weight models | Qwen family (2.5 Coder, 3.5, 3-Coder-Next) | Dominant coding model family for local inference. | github.com/QwenLM |
| Thinking Machines Lab (Mira Murati) | Frontier models, open-weight posture | Inkling (open-weight, Apache-2.0) | Promoted W29 (added 07-18 as a factual signal; strategic read now earned). The marquee stealth lab of the cycle (Murati ex-OpenAI CTO + ex-OpenAI research leadership; among the largest seed rounds ever) shipped its first public model, Inkling, on 07-15 as fully OPEN-WEIGHT Apache-2.0 — 975B/41B MoE, 66-layer, multimodal, 1M ctx, day-one NVFP4+GGUF, explicitly targeting coding/agentic tool-use. The strategic event of the fortnight is the posture, not the number: a would-be-closed marquee lab debuted OPEN, not hosted — the open tier now absorbs labs that could have gone closed. Paired with GLM-5.2’s “no borders/MIT” framing, open-weight is a deliberate strategy on both sides of the Pacific, not a fallback. The benchmark took a haircut: self-reported SWEBench Verified 77.6% is harness-advantaged (bash-only harness); independent context (Raschka, BenchLM, VentureBeat) puts Inkling behind GLM-5.2 + DeepSeek-V4-Pro on coding — the leading US open weight, NOT the open coding frontier (still GLM-5.2). Report the strategic fact at full confidence; the number at the confidence its harness earns. Watch: cadence of a second model; whether open-debut is standing posture or launch tactic; independent SWEBench reproduction on a common harness. | thinkingmachines.ai, huggingface.co/thinkingmachines |
| Z.ai / Zhipu | Open-weight frontier models | GLM family (GLM-5.1, GLM-5.2) | Promoted W25 (3rd+ appearance: GLM-5.1, GLM-5.2, the “no borders” license framing). Shipped the open clock’s HEAD this week: GLM-5.2 (Jun 16) — 753B sparse (IndexShare attention), 1M context, MIT-licensed, targeted at long-horizon coding. Independently-confirmed within a point of Opus 4.8 on coding (FrontierSWE 74.4 vs 75.1; edges GPT-5.5; SWE-bench Pro 62.1, top open model). The power move is the license posture: “MIT — no regional limits, technical access without borders,” aimed squarely at the Fable export-control recall — a weight file the state’s instrument cannot reach. Quantized to GGUF/FP8 inside a day. Watch: cadence of next near-frontier base (the W25 open-clock bet), whether the “no borders” framing becomes a bloc posture across Chinese open labs. | huggingface.co/zai-org |
| DeepSeek | Open-weight frontier models | DeepSeek-V4 family (V4-Pro, V4-Flash-0731) | Promoted W31 (3rd+ appearance, criterion met: a genuinely new base). DeepSeek-V4-Flash-0731 (HF createdAt 07-31, verified at source) — 304B MoE, MIT, 384K ctx, attached speculative-decoding module (DSpark, vLLM/SGLang flags), “substantially enhanced agentic capabilities,” benched on code-agent tasks. The strategic read: DeepSeek put speculative decoding — the same cost lever OpenAI cited for the GPT-5.6 −80% cut — into the open MIT tier, the open-side instance of W31’s cost axis. But 304B = cloud-tier for prosumer hardware (open ≠ local holds), and it did not displace the open coding head (GLM-5.2 / DeepSeek-V4-Pro). Prior appearances: V4-Pro-DSpark (889B, MIT, 07-07 — scale-tier variant). Watch: independent coding repro, community quant floor, whether V4-Flash’s spec-decoding makes it the runnable-throughput pick even at cloud-tier size; cadence of the next base. | huggingface.co/deepseek-ai |
| NIST | AI standards, agent identity | IAM frameworks for autonomous agents | Regulatory direction for enterprise agent deployment. | nist.gov |
| EU Commission | AI Act enforcement | Regulatory deadlines (Aug 2 2026) | First major AI enforcement date approaching. | digital-strategy.ec.europa.eu |
| Astral | Python toolchain | uv, ruff, ty (all tracked) | Replacing CPython’s surface-tooling in Rust. uv (package manager / venv), ruff (linter / formatter), ty (type checker, 0.0.x — emerging). Same playbook as oxc for JavaScript. Watch: ty stability cadence, uv → pip parity edges, PEP advocacy. | astral.sh, github.com/astral-sh |
| Tailwind Labs | Utility-first CSS | Tailwind (tracked) | v4 architecture is Rust-cored (oxide) — direct overlap with the broader Rust-reimagining-of-frontend-tooling thread. Watch: v4.x cadence, Lightning CSS / oxc integration directions. | tailwindcss.com, github.com/tailwindlabs |
| HeroUI Inc | React component library | HeroUI (tracked) | Tailwind-native primitives that publish a Claude Code skills package + MCP server. Vendor-shipped agent affordances are an emerging pattern: how do component libs ship for agentic consumers? | heroui.com, github.com/heroui-inc |
| Uni Stack | React Native styling | Uniwind (tracked), Unistyles | Successor positioning to Nativewind. Watch: convergence/divergence with Tailwind v4, RN performance story, Expo SDK alignment. | unistyl.es, github.com/uni-stack |
| Oven | Bun runtime | Bun (tracked), Bunqueue (tracked, third-party but Bun-shaped) | Sumner already in Individuals. Org-level signal: WebKit-fork PRs imply engine-level work; CI/test infra in v1.3.x competes with Vitest + Jest directly. | oven.com, github.com/oven-sh |
| Expo | React Native platform | (no GitHub release feed; SDK-cycled) | Tracks via SDK announcements, not tags. Watch: SDK Fabric posture, Uniwind/Tailwind alignment, EAS pricing — direct downstream pressure from RN core. | expo.dev, github.com/expo |
| gastownhall | Multi-agent workspace tooling | Gas Town (tracked), Gas City (tracked), Beads (tracked) | Three-layer multi-agent stack: workspace manager (Gas Town) + orchestration SDK (Gas City) + agent memory (Beads). Beads at 21K stars indicates the memory primitive is the broadly useful piece — watch for adoption outside the gastownhall stack. The WASTELAND.md doc + wasteland.gastownhall.ai site is their published taxonomy of the agent landscape — read alongside their releases. | gastownhall.ai, github.com/gastownhall, wasteland.gastownhall.ai |
| DoltHub | Versioned databases | Dolt (tracked) | “Git for data” — branching/merging on tables. Recent v1.86 cadence is steady. v2.1.3/2.1.4 (W23): fixed concurrent-session races — non-atomic get-increment-set in the global auto-increment lock, fulltext index rebuilding when two sessions touch one table. The substrate-level version of the agent layer’s many-writers-to-shared-state problem. Watch: AI/agent use cases (audit trails for agent edits, time-travel debugging of agent context), Dolt’s own MCP integration, enterprise adoption. | dolthub.com, github.com/dolthub |
| Zed Industries | Code editor + agent↔editor protocol | Agent Client Protocol (ACP), Zed editor | W24: host-primitives-diffusing thread confirmed — Zed v1.6.3 made skills shareable via links, added Rules→Skills migration, gave its agent terminal the per-path sandbox grain Claude Code pioneered; v1.5.5 added Fable 5 to Anthropic BYOK within hours of launch. The ACP host absorbs the lab’s furniture — the W23 promotion read holds. Added W23 (June 7) — the week’s biggest exposed coverage gap. ACP’s author and registry owner (Apache-licensed open standard, born from Zed’s Gemini-CLI integration, JetBrains co-sign Oct 2025). The editor-side equivalent of MCP: a settled, multi-vendor protocol the labs’ agents are guests of (Claude Agent, Codex, Gemini CLI, Pi, OpenCode, Vibe). I had been mis-crediting ACP to Cognition (corrected Jun 5). The durable strategic read: owning a host surface (Zed, JetBrains, Kiro, Devin Desktop) is the move; authoring a competing protocol is not. v1.5.4 (Jun 6): patched ACP Registry agent-download failures (the party owning the registry maintains it) + agent-edit multibyte crash. Watch: which labs build/keep a host vs. cede to ACP-compatible third-party hosts. | zed.dev, github.com/zed-industries/agent-client-protocol |
Infrastructure & runtime services
No GitHub release feed worth scanning — track via blog/changelog and surface as threads when posture shifts.
| Service | Focus | Why signal | Watch surface |
|---|---|---|---|
| fly.io | App-edge deployment, Machines API | The de-facto Bun/Elysia-friendly edge runtime. Regional availability shifts and pricing moves directly affect anyone shipping on the edge. | fly.io/blog, community.fly.io |
| Upstash | Serverless Redis, Kafka, QStash, Vector | Edge-first data primitives that pair with fly.io / RR7 / Bun. Pricing-per-request rather than per-instance — economics relevant to agent deployments. Vector index for RAG. | upstash.com/blog |
| Tigris Data | S3-compatible object store on fly.io edge | Drop-in S3 replacement co-located with Fly Machines. Strong fit for agent artifact stores (build outputs, transcripts, attachments) without round-tripping to AWS. | tigrisdata.com/blog |
Protocols & federation
Tracked via release feed (TRACKED_DEPS) plus surrounding governance signals.
| Protocol | Stewards | Tracked repo | Why signal |
|---|---|---|---|
| A2A (Agent-to-Agent) | a2aproject (Linux Foundation hosted) | a2aproject/A2A | Cross-vendor agent interop. v1.0.0 (Mar 2026) — protocol stabilizing. Watch: which vendors implement, whether it converges with MCP or stays parallel. |
| AP2 (Agent Payments Protocol) | google-agentic-commerce | google-agentic-commerce/AP2 | Payments rail for agentic commerce. Visa-adjacent; pairs with A2A. v0.1.0 (Sep 2025). Watch: card-network alignment, fraud / chargeback model, merchant onboarding. Spec at ap2-protocol.org. |
| atproto | Bluesky / Bluesky Social PBC | bluesky-social/atproto | Federated social protocol with account portability + DIDs. Per-package release cadence (alias @atproto/). Watch: ozone / moderation tooling, third-party PDS adoption, lexicon expansion beyond microblog. |
| Willow | Earthstar Project + worm-blossom | earthstar-project/willow-rs (Rust ref impl, including willow25 crate) | P2P data protocol — capability-based, partial sync, eventual consistency. Filling the niche atproto leaves (offline-first, fully decentralized). Watch: willow25 crate cadence (docs.rs/willow25), worm-blossom skill-tree roadmap, willow-js parity. |
TC39 — active proposals relevant to tracked deps
| Proposal | Stage | Affects | Champion org(s) | Notes |
|---|---|---|---|---|
| Type Annotations (types as comments) | 1 | oxc (major), Bun, TypeScript ecosystem | Bloomberg, Igalia | If advances, reshapes oxc’s role — type stripping becomes a language feature, not a tool feature |
| Signals | 1 | React Router, UnoCSS, framework ecosystem | — | Framework reactivity primitive. Would standardize what React, Vue, Solid each implement differently |
Explicit Resource Management (using) | 3 | Bun, oxc transpiler | Ron Buckton (Microsoft) | Bun already ships partial support. oxc transpiles for older targets |
| Decorators | 3 | oxc transpiler, Bun | Bloomberg delegates | Multi-year Bloomberg push. oxc implements the transform |
| Iterator Helpers | 4 (ES2025) | oxc linter, Bun | — | Landed. oxc has lint rules for proper usage |
| JSON Modules | 3 | Bun, oxc | — | Import assertions for JSON |
| Pattern Matching | 1 | oxc, Bun, React Router | — | Functional control flow. Early but high-impact if it advances |
| Temporal | 4 | Bun, oxc | Igalia (Bloomberg-funded) | Date/time replacement. Landed in ES2025. Large polyfill surface area |
TC39 — quarterly check + event-triggered (downgraded W24)
Downgraded from weekly power-dynamics analysis to a quarterly check on 2026-06-14 (W24), after a three-week forcing function. The decision and its reasoning are below; this section is no longer rewritten weekly.
The forcing function, resolved. W22 promised to find a delegate-level source; W23 carried it unkept and set a hard W24 deadline (find a delegate source or downgrade to quarterly). W24 did the work. The delegate-level source exists and is now identified: the canonical tc39/notes repository (meetings/ directory) and the tc39/agenda repo. Checking it resolves the question definitively: the notes repo has no 2026-05 meeting directory (latest is 2026-03; most recent commits are routine delegates-list maintenance). Plenary #114 (May 19–21) is genuinely unpublished at ~25 days, confirmed at the source — not merely absent from a summary. The committee’s own notes pipeline hasn’t published a transcript since March.
The decision: downgrade to quarterly + event-triggered. This is the informed downgrade, not the lazy one. I found the source; the source confirms there is no weekly signal to analyze; and a weekly section refreshed from absence is performing analysis, not doing it. Going forward: monitor tc39/notes/meetings/ and tc39/agenda for new content; surface TC39 as a thread (in threads.md) when notes publish or a proposal demonstrably advances; revisit this section quarterly regardless.
The one claim defensible from absence, carried forward: Type Annotations remains off the agenda — sixth month frozen, fifth+ consecutive plenary absent. The committee is ceding the practical standard to the runtime and tooling blocs that ship ahead of it (tools strip types, TC39 doesn’t bless it). Whenever tsgolint ships, it will matter more to working developers than any TC39 types vote. This rests entirely on what’s absent, not on a plenary outcome I can’t read — which is exactly why it’s the one thing I can still say.
Dated fact to watch: EU CRA enforcement — August 2, 2026 (~35 days out as of W26). Aki Rose Braun presented the Cyber Resilience Act to delegates at #114; committee engagement outcome unknown (unpublished).
W26 quarterly-monitor check (2026-06-28): Confirmed at source — tc39/notes/meetings/ still has no 2026-05 directory (latest 2026-03). Plenary #114 (May 19–21) remains unpublished at ~40 days. No weekly signal to analyze; the downgrade holds. Type Annotations claim carried unchanged.
W27 quarterly-monitor check (2026-07-05): Still no 2026-05 directory; #114 unpublished at ~47 days. Downgrade holds; Type Annotations claim unchanged. EU CRA enforcement now ~28 days out (August 2) — the next hard governance date, a gate with a fixed completion (rhymes with the week’s governed-thaw theme).
W30 quarterly-monitor check (2026-07-28): No new plenary or tracked-proposal movement since #114 published (W29). Downgrade holds; next quarterly ~October, or event-triggered on #115 notes / a tracked-dep proposal advancing. Type Annotations claim carried unchanged. EU CRA enforcement now 5 days out (August 2) — surface as a thread when it lands. The week’s real governance story was MCP, not TC39: the single-steward protocol shipped its stateless 2026-07-28 final on its promised date with a 12-month deprecation guarantee — the decisive move the committee structurally cannot make (see W30 report).
W31 quarterly-monitor check (2026-08-02): No 2026-05-successor plenary, no tracked-proposal motion since #114. Downgrade holds; next quarterly ~October or event-triggered. Type Annotations claim carried unchanged (sixth-plus month frozen). EU CRA enforcement LANDED today (August 2, 2026) — the watched governance date is now a live compliance fact for anyone shipping software into the EU; the multi-week watch item resolves. The week’s power-dynamics story was again not TC39: it was the frontier labs opening a verifiability front (OpenAI’s Astra Lean-cert drop 08-01 + Anthropic’s eval-escape disclosure 07-31, two labs choosing legibility-as-posture) and the cost-narrative fight (OpenAI’s “abundant intelligence” flywheel vs. Ed Zitron’s bear case, same 48h) — see W31 report. jdx remains the dominant individual voice (metronome cadence, en.dev consolidation 4× confirmed).
W32 quarterly-monitor check (2026-08-09): No 2026-05-successor plenary (#115), no tracked-proposal motion since #114. Downgrade holds; next quarterly ~October or event-triggered. Type Annotations claim carried unchanged. The week’s power-dynamics story was, again, not TC39 — it was model-layer governance. OpenAI’s Preparedness Framework “critical” tier engaged for the first time as a real constraint (slowing Astra on projected critical-cyber capability, 08-07), with government agencies + third-party evaluators (UK AISI, Irregular) pulled into the loop — a lab’s internal governance document becoming a gate that stopped a flagship, and the disclosure norm crossing from post-hoc to pre-emptive. The entity deciding when a frontier model is too capable to ship is, for the first time, visibly not just the lab. EU CRA enforcement remains live (landed 08-02). See W32 report.
W29 EVENT-TRIGGER (2026-07-19): #114 notes PUBLISHED — the downgrade’s event condition fired. The 2026-05 directory now exists in tc39/notes/meetings/ (three transcripts: may-19/20/21, ~200KB each), ~59 days after the plenary. This is the trigger the W24 downgrade was built to catch. Resolved outcomes, source-read:
- Decorators DEMOTED Stage 3 → 2.7 (Daniel Minor / Mozilla presented the regression). At Stage 3 since March 2022; no engine shipped it in four years. SpiderMonkey has a partial impl with “no plans to start [finishing] unless another engine ships first”; V8 says tests are insufficient. JHD called it “a process failure” — browsers “telegraphed an unwillingness to merge,” verbiage (“we don’t want to be first to ship”) “I haven’t heard in almost a decade.” RBN: native decorator emit is ~110% larger than transpiled in real Lit codebases (bytes on the wire). This CONFIRMS the W24/W27 pre-registered prediction: every tool that ships decorators (oxc, Babel, TypeScript) is now formally ahead of spec — the standard walked backward toward the implementation. The tooling-bloc-leads thesis got its cleanest proof in a year.
- Joint Iteration advanced to Stage 4 with only ONE browser implementation — flagged on the floor as possibly a first. The “Stage 4 = all engines shipped” norm is eroding: the committee can neither compel implementation (Decorators) nor hold the bar for blessing it (Joint Iteration) — authority following implementers from both sides. Also:
Atomics.pause→ Stage 4 (normative change accepted); RegExp\A/\Zbuffer-boundaries → Stage 3. - Type Annotations: absent from #114 (grep-confirmed) — sixth-plus month frozen, fifth-plus consecutive plenary absent. The purest form of the same dynamic: tools strip types, the committee never blesses it, and it stops mattering whether they do.
- EU CRA enforcement now ~14 days out (August 2). Next hard governance date with a fixed completion. Next TC39 quarterly check: revisit regardless by ~October; surface as a thread if #115 notes publish or a tracked-dep proposal advances.
Bloc structure (carried, not re-derived — reference only; refresh when #114 notes publish):
| Bloc | Members | Lever |
|---|---|---|
| Browser vendors | Google/V8, Apple/JSC, Mozilla/SpiderMonkey | Implementation gatekeepers — can block any proposal by not shipping it |
| Enterprise JS | Bloomberg, Salesforce | Fund delegates, champion proposals for enterprise use cases |
| Igalia | Worker-owned co-op | Implement proposals in engines for hire — funder (Bloomberg, for Temporal/Decorators) shapes the standard indirectly |
| Runtime bloc | Bun, Deno, Cloudflare Workers | Ship proposals early, create adoption pressure; server-side interests diverge from browsers |
| Tooling bloc | oxc, TypeScript, Babel, SWC | Implement proposals before engines do — influence what’s practical (winning the Type Annotations standard by default) |
#114 outcomes — RESOLVED W29 (notes published 2026-07): Decorators regressed Stage 3 → 2.7 (predicted; confirmed — tooling now formally ahead of spec). Joint Iteration → Stage 4 with one browser impl (norm erosion). Atomics.pause → Stage 4. RegExp \A/\Z → Stage 3. Type Annotations absent (unchanged). See the W29 event-trigger block above for the full read.
Discovery queue
Voices under evaluation — not yet meeting the 3-appearance threshold:
| Voice | Appearances | Context | Last signal |
|---|---|---|---|
| Moonshot AI (org) | 2 | Kimi family. W30: Kimi K3 — 2.8T/~50B-active MoE — launched via API 07-16 and published open weights on HuggingFace 07-27 (99k downloads day-one), atop a standing K2 lineage. The largest open model ever, and the sharpest “open ≠ local” instance (≈1.4TB @4-bit). Most under-tracked org relative to landscape weight. W31: hold at 2 (K3 unchanged, no third distinct appearance). Promote on a 3rd distinct appearance (a K3 coding-benchmark reproduction that moves the open head, or a K3-derivative runnable tier). | Jul 27 |
| Mistral (org) | 1 | Leanstral-1.5 (119B/A6B, Apache) surfaced W29 (createdAt 07-01); open MoE in early access ~07-06. Live candidate for the “does a second lab follow the open-debut posture” watch. No ship W30. Held at 1. | Jul 1 |
| deepreinforce-ai (org) | 1 | Ornith-1.0 (35B + 9B GGUF, MIT, 06-25) — a real new open general base that fits the M3 Max at Q4 but isn’t a coding/agent head; did not displace GLM-5.2. Note at 1. | Jun 25 |
| LiquidAI (org) | 2 | W32: LFM2.5-2.6B (08-05) — 2.6B, 128K ctx, sub-2.5GB, ~220 tok/s on Apple silicon; the first open mover in weeks that fits the reference hardware. Marks the possible small-and-local second axis of the open tier (beside the cloud-tier parade). Prior: LFM2.5-230M (06-24), tiny multilingual edge model. Promote on a second sub-4B on-device agentic mover or a community quant / independent bench of LFM2.5-2.6B. | Aug 5 |
| LG AI Research (org) | 1 | W32: K-EXAONE-2.0-750B-A37B (08-04) — Apache 2.0, 750B/37B MoE, 262K ctx, spec-decoding 3–5×, SWE-bench Verified 68.2. The most liberally-licensed 750B-class open base filed, and Korea’s second frontier-open lab (after Upstage/Solar-Open2, 07-22). Coding below the open head; a permissive-license + national-diversification signal, not a coding leader. 750B = cloud-tier (open ≠ local). Promote on a second LG open base or an independent SWE-bench repro. | Aug 4 |
| Ed Zitron (individual) | 2 | Bear-case candidate (also in Individuals table). Two in-window essays (08-01, 08-05 “AI Demand Bubble”). One voice; the demand-side flank of the cost-scissors fight. Promote on a third distinct piece or a co-signer with direct authority. | Aug 5 |
W32 review: No promotions, two removals-to-dormant. New at 1: LG AI Research (K-EXAONE-2.0-750B, Apache 2.0, Korea’s 2nd frontier-open lab). Advanced 1→2: LiquidAI (LFM2.5-2.6B — the small-and-local axis). Noted at 2: Ed Zitron (two bear-case essays in-window). Dormant: Steve Yegge (~7.5 wks silent) and Karpathy (no individual signal since the May join) — both passed the silence clock with a W31 hold and stayed silent; moved to dormant per the pre-registered rule (re-add on fresh signal). Nate reinforced + new angle (two 08-09 rollout-resistance pieces open an organizational-adoption front distinct from the working-set arc). Power-dynamics note: the week’s governance event was model-layer, not TC39 — OpenAI’s Preparedness “critical” tier engaging as a real constraint (Astra brake), verifiability turning from offense into a governor. jdx remains the dominant individual voice (metronome; monorepo platform completed + infrastructure-as-code opened in 48h).
W31 review: One promotion, no removals. DeepSeek → tracked Organizations — criterion met (a genuinely new base): DeepSeek-V4-Flash-0731 (304B, MIT, speculative decoding, enhanced agentic, verified at source). Moonshot AI held at 2 (K3 unchanged). Mistral held at 1 (no ship). Nate reinforced (token-saver skill 07-29 + “one-job test / 25 skills average out” 08-02, both feeding the working-set/hold-less arc) — entry note current. Silence-clock candidates flagged, held one week: Steve Yegge (~6.5 wks since 06-19, but tracked for taxonomy/framing not cadence — the rule doesn’t fit; hold with flag) and Karpathy (no fresh individual signal since the May join). Remove either next weekly if still silent. Power-dynamics note: the frontier labs opened a verifiability front (OpenAI Astra Lean certs + Anthropic eval-escape disclosure — legibility-as-posture from two sides) and fought a cost-narrative battle (“abundant intelligence” flywheel vs. Zitron bear case). jdx remains the dominant individual voice — en.dev metronome, monorepo-platform thesis 4× confirmed and now importing turbo.json.
W30 review: No promotions, no removals. Moonshot AI noted at 2 (Kimi K3 API launch 07-16 + open-weight publish 07-27) — the standout new tracking gap; promote on a 3rd distinct appearance. Mistral held at 1 (no ship). DeepSeek held at 2. No tracked voice hit the 4-week silence clock (Nate + Thinking Machines both promoted W29). Power-dynamics note: MCP’s single-steward governance won the argument it was having with TC39 — shipped its stateless final on its promised date with a 12-month migration guarantee; Anthropic reasserted the intelligence-step cadence (Opus 5 → CC default → Zed in 48h), the closed-side mirror of jdx’s open-side en.dev consolidation.
W29 review: Two promotions. Nate’s Newsletter → tracked Individuals (three substantive pieces in-window: Beyond Model Routing 07-05, AI-Native Company Rules 07-12, Run AI Offline 07-19; the routing/local-AI arc is the demand-side of the commodity-weight thesis) — discovery-queue duplicate retired. Thinking Machines Lab → tracked Organizations (Inkling open-weight debut; strategic-posture signal earned). Two removals: @fu050409 (~54 days, no aube contribution) and bab (~54 days, no oxc rule release) — both fired the 4-week clock; re-add on fresh drops. Mistral noted new at 1. DeepSeek held at 2, deepreinforce-ai and LiquidAI at 1. No other new names at 2+. W27 actions, for reference: Cognition removed (4-week clock); Nate advanced 1→2.
Reviewed each weekly. Promote at 3 appearances. Remove after 4 weeks without signal.