The Voices Pack — tapping the X universe without drowning in it

One line: a curated roster of ~100 AI + responsible-AI voices becomes a first-class source in the Signal Studio manifest — ingested cheaply, synthesized as discourse (what the field is arguing about), never as a firehose.

Origin: your dictation last night ("bring additional sources in… maybe a check-on-check thing") + this morning's realization that a friend's X share exposed the gap: you have no tap into the X universe.


The design principle: voices are a consensus instrument, not a feed

The failure mode of adding 100 accounts to anything is 100 accounts' worth of noise. So the roster is never rendered as posts. It's rendered as discourse measurements:

Verification tier for social content (the check-on-check rule, extended)

Social claims are the lowest-trust class in the pipeline, so the rules are structural:

Roster governance (same manifest discipline as everything else)

The roster lives in the studio's sources panel as voices.json: name, platform+handle, category (lab-leader / researcher / builder / commentator / safety / ethics / governance / enterprise-RAI), weight, added date, and — mortality again — a 90-day review date: voices that go silent or drift off-topic get flagged for replacement, so the roster stays a living instrument instead of a 2026 time capsule. Adding a voice shows its cost impact before save, like every other source.

Podcast synthesis

The daily episode (now 4 stories) gains one discourse beat, not four: the day's sharpest argument from the roster, framed as a debate ("X says A, Y counters B — here's the evidence side"). Voice + counter-voice is radio-native and it's the format that makes 100 voices compress instead of expand. The weekly public edition gets the fuller "state of the discourse" segment with the verification ledger's receipts.

The roster — ~100 voices, researched 2026-07-17

Handles verified against mid-2026 reporting where possible; affiliations current as of research date (Karpathy → Anthropic pretraining May 2026; LeCun → AMI Labs; Murati → Thinking Machines; Sutskever → SSI CEO). Pure-hype and rumor accounts deliberately excluded. This is the draft roster — you prune it in the studio; your taste is the editorial value.

Part 1 — AI (60 voices)

Frontier-lab leaders (14):

NameXRoleWhy
Sam Altman@samaCEO, OpenAIHighest-reach lab account; posts move markets
Greg Brockman@gdbPresident, OpenAIMore technical than Altman; training-infra detail
Dario Amodei@DarioAmodeiCEO, AnthropicRare posts, but essays set discourse for weeks
Demis Hassabis@demishassabisCEO, Google DeepMindGemini, science-AI, measured AGI takes
Elon Musk@elonmuskFounder, xAINoisy, but Grok/xAI lands here first
Mira Murati@miramuratiCEO, Thinking MachinesThe most-watched new lab
Ilya Sutskever@ilyasutCEO, SSIEvery rare post is an event
Mustafa Suleyman@mustafasuleymanCEO, Microsoft AICopilot + policy from AI's largest distributor
Yann LeCun@ylecunExec Chair, AMI LabsThe "LLMs are a dead end" counterweight, $1B world-model bet
Alexandr Wang@alexandr_wangChief AI Officer, MetaRuns Meta's frontier push
Arthur Mensch@arthurmenschCEO, MistralThe European sovereign-AI position
Clem Delangue@ClementDelangueCEO, Hugging FaceOpen-source ecosystem barometer, daily
Aravind Srinivas@AravSrinivasCEO, PerplexityAI search / browser wars, prolific
Sundar Pichai@sundarpichaiCEO, GoogleGemini-era strategy originates here

Researchers (15): Andrej Karpathy @karpathy (Anthropic; the field's most-read explainer) · Jeff Dean @JeffDean (DeepMind chief scientist) · Noam Brown @polynoamial (OpenAI reasoning line) · François Chollet @fchollet (ARC Prize; the rigorous skeptic) · Jim Fan @DrJimFan (NVIDIA embodied AI) · Lilian Weng @lilianweng (canonical RLHF/agents writing) · Sebastian Raschka @rasbt (architecture breakdowns within days) · Fei-Fei Li @drfeifei (Stanford HAI / World Labs) · Andrew Ng @AndrewYNg (applied-ML signal) · Chris Olah @ch402 (interpretability founding voice) · Neel Nanda @NeelNanda5 (most active interp poster) · Jan Leike @janleike (Anthropic alignment lead) · Amanda Askell @AmandaAskell (model character/behavior) · Sholto Douglas @_sholtodouglas (RL scaling) · David Ha @hardmaru (Sakana AI)

Builders (12): Simon Willison @simonw (what LLMs can actually do) · Boris Cherny @bcherny (Claude Code creator) · Logan Kilpatrick @OfficialLoganK (Gemini API lead) · Michael Truell @mntruell (Cursor CEO) · Lee Robinson @leerob (Cursor DX) · Amjad Masad @amjadmasad (Replit) · Guillermo Rauch @rauchg (Vercel) · Pieter Levels @levelsio (indie AI in public) · Peter Steinberger @steipete (OpenClaw creator, now OpenAI) · Mckay Wrigley @mckaywrigley (shipping with agents daily) · Riley Goodside @goodside (finds model quirks first) · Eric Zakariasson @ericzakariasson (Cursor UX experiments)

Analysts / commentators (15): swyx @swyx (Latent Space; mapped the AI-engineer category) · Ethan Mollick @emollick (AI × knowledge work, daily experiments) · Zvi Mowshowitz @TheZvi (the most complete weekly digest alive) · Dwarkesh Patel @dwarkesh_sp (best long-form interviews) · Nathan Lambert @natolambert (post-training + open weights) · Jack Clark @jackclarkSF (Import AI; research+policy synthesis) · AK @_akhaliq (default feed for new papers) · Rowan Cheung @rowancheung (highest-volume launches) · Nathan Benaich @nathanbenaich (State of AI; compute/geopolitics) · Nathan Labenz @labenz (capability+alignment scouting) · Deedy Das @deedydas (one-tweet distillations) · Gary Marcus @GaryMarcus (the persistent bear case — useful friction) · Eliezer Yudkowsky @ESYudkowsky (the doom frame's originating voice) · Lex Fridman @lexfridman (leaders say things there they don't post) · Allie K. Miller @alliekmiller (how enterprise hears about AI)

High-signal, lower-fame — what practitioners actually read (15): Teortaxes @teortaxesTex (earliest credible China-lab reads) · Tibor Blaho @btibor91 (scoops from app builds) · Cameron Wolfe @cwolferesearch (rigorous paper breakdowns) · wh @nrehiew_ (dense training threads, lab-cited) · vik @vikhyatk (Moondream; small-VLM antidote to frontier-only thinking) · Linus Lee @thesephist (LLM internals + AI-native interfaces) · VB Srivastav @reach_vb (fastest open-weight release signal) · Omar Sanseviero @osanseviero (Gemma/open ecosystem) · Maxime Labonne @maximelabonne (the fine-tuning guides everyone uses) · Teknium @Teknium1 (Nous; open post-training frontier) · kalomaze @kalomaze (sampling/RL tinkering in public) · xjdr @_xjdr (entropix; inference tricks that jump into labs) · Ahmad Osman @TheAhmadOsman (GPUs and self-hosted metal) · Artificial Analysis @ArtificialAnlys (neutral benchmarks every release day) · Epoch AI @EpochAIResearch (compute/scaling trend anchors)

Part 2 — Responsible AI (~40 voices + institutions)

The platform note that shapes the whole tap: AI safety and policy voices remain heavily on X; the FAccT / AI-ethics community largely left X in 2023–25 for Bluesky and Mastodon (much of it around dair-community.social); enterprise RAI practitioners live on LinkedIn + Substack. Handles marked UNSURE were not confirmed — verify before wiring into the feed.

Safety / alignment / evals (13): Dan Hendrycks X @DanHendrycks (CAIS; MMLU/HLE benchmarks) · Yoshua Bengio X @Yoshua_Bengio (LawZero; chaired the International AI Safety Report) · Jan Leike X @janleike · Eliezer Yudkowsky X @ESYudkowsky · Beth Barnes (METR — via @METR_Evals; the task-length-doubling curve) · Daniel Kokotajlo (AI 2027; blog.aifutures.org) · Neel Nanda X @NeelNanda5 · Chris Olah X @ch402 · Amanda Askell X @AmandaAskell · Ajeya Cotra X @ajeya_cotra (Open Phil) · Zvi Mowshowitz X @TheZvi · Nathan Labenz X @labenz · Marius Hobbhahn (Apollo Research — scheming evals; handle UNSURE, follow the org)

Ethics / FAccT — mostly on Bluesky now (13): Timnit Gebru bsky @timnitgebru.bsky.social (DAIR) · Emily M. Bender bsky @emilymbender.bsky.social (The AI Con) · Alex Hanna bsky @alexhanna.bsky.social (DAIR) · Margaret Mitchell bsky @mmitchell.bsky.social (HF chief ethics scientist; model cards) · Abeba Birhane bsky @abeba.bsky.social (AI Accountability Lab) · Deborah Raji X @rajiinio (algorithmic auditing) · Arvind Narayanan X @random_walker (AI Snake Oil) · Sayash Kapoor X @sayashk · Joy Buolamwini X @jovialjoy (AJL) · Kate Crawford X @katecrawford (Atlas of AI) · Meredith Whittaker (Signal president; left X, handle UNSURE) · Sasha Luccioni LinkedIn sashaluccioniphd (AI × climate accounting) · Seth Lazar X @sethlazar (normative philosophy of agents)

Governance / policy (14): Helen Toner X @hlntnr + Rising Tide · Miles Brundage Substack + X @Miles_Brundage (AVERI) · Dean Ball Hyperdimensional + X @deanwball (co-wrote the US AI Action Plan) · Kai Zenner LinkedIn kzenner (EP insider on AI Act implementation) · Luiza Jarovsky X @LuizaJarovsky (highest-reach AI Act explainer) · Gabriele Mazzini LinkedIn (architect/lead author of the EU AI Act) · Marietje Schaake X @MarietjeSchaake (The Tech Coup) · Jack Clark X @jackclarkSF (Import AI) · Lennart Heim X @ohlennart (RAND; compute governance) · Markus Anderljung X @Manderljung (GovAI) · Peter Wildeford X @peterwildeford (IAPS) · Merve Hickok LinkedIn mervehickok (CAIDP) · Shakeel Hashim Transformer + X @ShakeelHashim (the governance trade-press) · Paul Christiano (CAISI/NIST — no active social; papers)

Enterprise RAI — mostly LinkedIn (10): Navrina Singh (Credo AI) · Rumman Chowdhury bsky @ruchowdh.bsky.social (Humane Intelligence; invented enterprise red-teaming) · Oliver Patel LinkedIn + Enterprise AI Governance Substack (AstraZeneca — the practical operator voice) · Paula Goldman (Salesforce) · Christina Montgomery (IBM) · Kay Firth-Butterfield X @KayFButterfield · Patrick Hall (GWU; NIST AI RMF-aligned practice) · Ravit Dotan (TechBetter) · Ashley Casovan (IAPP AI Governance Center) · Reid Blackman (Ethical Machines)

Institutions (the always-on layer): METR @METR_Evals · CAIS @ai_risks · AI Now @AINowInstitute · Apollo Research · UK AI Security Institute · European AI Office · NIST CAISI · GovAI · Ada Lovelace Institute · DAIR (Mastodon) · Epoch AI @EpochAIResearch · Partnership on AI · Future of Life Institute · AI Snake Oil (Substack)

Full research with per-claim source URLs: voices-research-ai.md · voices-research-rai.md

The tap — how to actually ingest X in 2026 (mechanics + cost)

The research verdict: don't buy the firehose — buy the summary, and spot-check the source. The recommended stack totals ~$2/month, every component independently replaceable:

  1. Backbone (free): smol.ai AI News over RSS (news.smol.ai/rss.xml) — swyx's daily, agent-curated digest that explicitly summarizes "top AI discords + AI reddits + AI X/Twitters" with embedded tweet quotes. Ingesting it is ingesting curated X for $0. Depth feeds alongside: Simon Willison, Import AI, Last Week in AI.
  2. The targeted X tap (~$1–2/mo): the Grok API's x_search tool — server-side live-X search with allowed_x_handles (20 handles per call, rotate 2–3 calls for the full must-read tier), $5 per 1,000 tool calls plus pennies of tokens. A nightly "summarize the last 24h from these handles" job costs ~$0.03–0.05/day — sanctioned, ban-proof access from the one company allowed to read X. This is the loophole that makes the whole feature viable.
  3. The free half of the graph ($0, no auth): Bluesky — any app.bsky.* GET works unauthenticated against public.api.bsky.app, including getListFeed for a curated list's whole timeline in one call. The academic/RAI wing has substantially migrated there, so the responsible-AI half of the roster is largely free to follow.
  4. Only if raw per-tweet JSON is ever truly needed: twitterapi.io at ~$0.15/1K tweets (~$1–5/mo at digest volume) — 33× cheaper than official, but ToS-gray and could vanish; code defensively.

Rejected, with reasons: the official X API (went pay-per-use Feb 2026, $0.005/post read, no free tier — ~$30/mo for the same digest volume, worst value here); Nitter/xcancel/RSSHub cookie routes and logged-in scraping (fragile, and the cookie approaches put a real account at ban risk — never load-bearing).

Rollout