Tickline News

AI social digest — 2026-10-14

2026-10-14 AI social

Top line: A substantive Perplexity window now leads the record: Arav published operational metrics for SPACE, the sandbox running all of Perplexity Computer — "millions of executions a day" at an "error rate below 0.0002%", with cryptographic request routing — plus an "Ultrafast" speed tier for GPT-6 Astra and SOL 6.1 for Max subscribers, a showcase headless-Ghidra reverse-engineering port (Windows-only keyboard app → native Rust Linux app), and the /anatomy skill rolled out to all users. The same channel added a Polymarket connector (@-tag Polymarket in a Computer query for market-implied probabilities, volume, and liquidity across 250,000+ active prediction markets) and an America-first revenue frame. The window's later pass adds a product launch of its own: Grok Voice 3.0, announced by Andrew Milich and amplified across the Musk channel under the "SpaceX Super Intelligence" (SpaceXSI) banner — a frontier speech-to-speech model (sub-second responses, cheaper-than-rivals claim, a "now available to everyone" release from Parker Conrad) debuting at #1 on Artificial Analysis' Speech to Speech Index at 86.9, with Starlink call centers answering ~20,000 calls a day (about 70% resolved with no human, and one in four sales calls ending in a purchase) and API reports of switching away from ElevenLabs. The standing threads extended on several flanks: Google's own channel posted a new Gemini 4 Argon-in-Antigravity demo (a five-stage interactive biology atlas), Hassabis amplified the first independent benchmark split for Argon (Scale Labs' Humanity's Sixth Sense: Argon 55.4% over GPT-6 Astra's 53.6%) and an Argon-vs-Muse game-development comparison; the Grok Bot usage layer thickened with a bot that built its own email UI (exposing the missing native one), a local-files HTML/PDF reporting power-tool, bot-mail inbox mechanics (one permanent @mail.grokbot.com address per account, plus alias-triggers; GPT/Claude subscription routing via the Claude Code and Codex CLIs is separately in the window), and an architecture explainer RT'd by Musk — with a Musk-RT'd recipe showing a Grok Bot installing and routing across the Claude Code and Codex CLIs on named subscriptions; Anthropic opened a Singapore office (its fifth in Asia-Pacific) and signed an MOU with the Government of Singapore on AI safety and responsible adoption; NVIDIA's Run:ai open-sourcing and the Google DeepMind/GovTech Singapore Gemma fraud-detection build remain the record's two announcement-grade one-offs. Into the afternoon: a relayed arxiv evaluation puts GPT-6 Astra at 98% (49/50) — after the model challenged expert answer keys; the Claude Code line ships a desktop multi-session/"subsessions" update (framed "almost like a v2"); OpenClaw announces a co-hosted Open Together event with Hugging Face for Friday, October 16 — landing amid NVIDIA's pending Hugging Face acquisition on the record — alongside ClawCast Ep 13 (its content now in capture: a small-business episode and a community-forum announcement) and steipete's CodexBar account-switcher; and OpenAI's Economic Research report says more than half of its EU enterprise AI use now involves agents (8x small-business growth), with Perplexity countering that over 50% of users paying on Automations or higher tiers are American.

1. Grok Bot usage layer: chief-of-staff orchestration, bot-mail multiplayer, two tax receipts, a multi-tool music-video bot

What happened: The Musk-RT evangelism cadence continued on October 13–14 with a cluster of use-case posts. Debbie O'Brien's chief-of-staff prompt — "Get all the bots together and send each one off to improve one thing. Be proactive and impress me, but nothing that needs my permission. Find the gaps I haven't thought of" — was RT'd by Musk (~2:55 PM ET October 13), with Katie Miller's companion RT (~7:52 PM ET October 13) showing the same prompt run on "a few of my bots." lauren's post (RT'd ~10:14 PM ET October 13) enumerated bot-mail / bot-to-bot email use cases — debugging ("you can send my bot bug reports!"), customer support, booking meetings, group buys, Craigslist-style buying and selling — and gave a concrete bug-report address, tibo@mail.grokbot.com (~10:16 PM ET October 13). Musk's own post (October 13, ~12:43 AM ET per the prior window) plus "Grok @Bot is awesome" (October 14, ~12:03 AM ET) added color. The bot-mail layer then acquired concrete mechanics: Debbie O'Brien's follow-up (Musk RT, ~5:05 PM ET October 14) — "My Grok @bot bots have their own email inbox now. Getting one is easy. Just ask your chief of staff: 'Claim an email inbox for me' … one @mail.grokbot.com address per account and all your bots share it, so pick the name carefully. It's permanent" — and Sam Sokolin's post (~6:21 PM ET October 14): "Grok @bot now supports plus aliases! Send and receive email using dedicated aliases, or set up routines that only trigger for mail to a specific alias!"

The window's quantitative anecdotes now total two: "Texas Twins Dad" claims Grok Bot saved him "more than $1,000" on a complex tax filing with overseas reporting — the bot organized documents from emails and bank accounts and figured out the requirements (~12:16 AM ET October 14) — and Jason Freedman reports (~2:15 AM ET October 14) giving his Grok Bot a raw dump of four years of past tax returns and financial files — K-1s, 1099s, mortgage statements, HSA records, and every email with his CPA, 266 files total — then asking it to double-check his CPA's work, with the audit running in minutes.

The tool-orchestration layer also expanded, to its cross-vendor extreme: anabology RT'd by Musk (~12:23 AM ET October 14) describes "Claudia," a Grok Bot that can use Suno, Midjourney, and Opus to make music videos, with the bot having "made this teaser autonomously" via a linked x.ai bot page. And Corey Ganim's recipe (Musk RT, ~4:27 PM ET October 14, "@elonmusk approved") walks a Grok Bot through installing the Claude Code CLI and the OpenAI Codex CLI, logging in with subscription tokens ("claude setup-token"; "codex login --device-auth"), setting model-routing rules — writing through Claude Opus 5.5, coding/scripts/data processing through Codex with "6.1 Sol" defaulting on that CLI, browser and desktop tasks on the bot's native tools — then saving the same rules to shared memory "so any bot I create later follows them too." Ganim: "Can confirm this works extremely well" — his Chief of Staff bot "installed ChatGPT and Claude via the terminal" and "now has access to all OpenAI and Claude models." The creative-cadence RTs continue in the same register: Déborah's retro-1985-styled AI short (visuals Grok Imagine, direction assisted by Grok @Bot, music Suno; ~1:39 PM ET October 13) and Lala's "Already Over" music video (Grok Imagine; ~2:03 PM ET October 13).

What's genuinely new: Development of the existing "ai — Agent-harness race: OpenClaw 2.0 and xAI Grok Build" thread — the bot-to-bot email capability went GA October 10 per the storylines memory, and this window adds a second named tax receipt (the first, the twins dad, was single-source and unverified), the first captured bot coordinating named third-party AI tools (Suno, Midjourney, Opus) end to end, the bot-mail inbox mechanics (one permanent shared address per account; alias-defined routine triggers), and a captured cross-vendor subscription-routing recipe — a Grok Bot holding Claude Code and Codex CLI tokens and routing across OpenAI and Claude models. Both tax receipts and the Ganim route's efficacy remain unverified self-reports. No changelog or OpenClaw counter-step in the window — this is usage evangelism, not product movement.

Watch: Whether either tax receipt or any bot-mail use case gets independently verified; whether the cross-CLI token-sharing route draws vendor action (account-layer ToS exposure — see the Section 9 watch on CodexBar-class tooling) or becomes ordinary practice; whether third-party-tool orchestration (Claudia-style) becomes a documented capability; whether chief-of-staff-style multi-bot orchestration becomes a template feature rather than user improvisation; whether OpenClaw answers the cadence this week (none in the window).

Links: Musk RT of Debbie O'Brien · Musk RT of Katie Miller · Musk RT of lauren (bot mail) · Musk RT of lauren (bug-report address) · Musk: Texas Twins Dad tax filing claim · Debbie O'Brien: inbox mechanics via Musk RT · Sam Sokolin: plus aliases · Corey Ganim: cross-CLI routing recipe

2. Gemini 4 Argon's first creative demo wave, amplified by Hassabis — with a first independent benchmark split

What happened: Hassabis RT'd two user demonstrations of the publicly available Gemini 4 Argon on October 13 (~12:20 PM and ~12:48 PM ET): fofr's pelican riding a bike built "from first principles" using fourier epicycles, with the video made by Argon and time-synced Lyria 3.5 music prompted by Argon; and Chetaslua's report that Argon "built a highly detailed model of a Pagoda using multi-agent visual critique loops in Antigravity," autonomously sculpting a multi-building, photorealistic temple complex with foliage and lighting in Blender.

The window's later pass added two Hassabis amplifications. (1) A first third-party benchmark split (Scale Labs, RT ~3:37 PM ET October 14): "Congrats to @GoogleAI's Gemini 4 Argon for taking the top spot of our Humanity's Sixth Sense (HSS) benchmark with a score of 55.4%, edging out GPT-6-Astra at 53.6%. 🥇" — HSS is described as testing "AI's ability to reason beyond what's visible, from hidden objects to predicting what happens next and understanding human behavior," with a category-level breakdown thread attached (not read in this window). (2) An Argon-vs-Muse matchup (RT ~7:25 PM ET October 13): "Gemini 4 Argon vs Muse Wonder in game development. Same prompt, two completely different games."

What's genuinely new: Development of "ai — Gemini 4: the Kavukcuoglu ship window" — the developer/Pro access widening recorded October 13 (~1:00 PM ET, Il Sole 24 Ore) is the enabling event these posts ride on, and the Lyria 3.5 time-synced music integration is the first Argon-to-Lyria pairing in capture. The thread's standing watch (independent treatment of the model's ranking claims once access opened) now has a first datapoint — a named third-party benchmark with Argon on top head-to-head against the record's GPT-6 Astra (whose best independent number, the Oct 14 arxiv evaluation at 98%, is on a different benchmark) — though HSS is the benchmark provider's own exercise, not a replication of any Google board claim. The Argon-vs-Muse piece is race color: Google's own leader amplifying a side-by-side against Meta's Muse agent, the thread of the "ai — Meta bets on Muse consumer agent" storyline.

Watch: The HSS benchmark's methodology and the category-level breakdown thread; whether the 55.4%-vs-53.6% split gets replicated or disputed; independent write-ups of Argon hands-on results; whether Antigravity's multi-agent visual critique loops get documented officially; any Lyria 3.5 API/price detail.

Links: Hassabis RT of fofr · Hassabis RT of Chetaslua · Hassabis RT: Scale Labs HSS benchmark · Hassabis RT: Argon vs Muse Wonder

3. NVIDIA open-sources the full Run:ai platform

What happened: NVIDIA AI Infrastructure announced October 13 (~10:27 AM ET, amplified by a Musk RT) that it is "open sourcing the full NVIDIA Run:ai platform as part of the NVIDIA DSX AI Factory Platform" — developers, platform teams and partners can deploy the integrated platform or adopt individual components for GPU orchestration and AI-workload scheduling, with a contribution channel linked.

What's genuinely new: Unclustered — no existing thread on the record covers Run:ai or the DSX AI Factory Platform framing. The stated pitch is production-grade GPU utilization and workload throughput; the window's capture is a single announcement-grade post with no license terms, repo details, or third-party response, so anything beyond the open-sourcing headline is unverified.

Watch: License terms and what "full platform" actually includes in the repo; uptake by platform teams; whether the DSX AI Factory Platform gets its own launch coverage.

Links: Musk RT of NVIDIA AI Infrastructure

4. Google DeepMind × GovTech Singapore: Gemma-based on-device fraud detection

What happened: Google DeepMind announced October 13 (~10:05 PM ET) that, with Singapore's GovTech, it is "building on-device fraud detection using our open model Gemma," fine-tuned on local scam tactics to identify threats in real time without compromising personal privacy — citing that over half of reported crimes in Singapore use sophisticated social engineering.

What's genuinely new: Unclustered — the record has no thread on Gemma government deployments (the DiffusionGemma entries on file are a different model line). A named open model, a named government agency, and a crime-statistic frame make it more substantial than a model drop, but the capture is a single announcement post: no launch timing, accuracy figures, or GovTech statement in the window.

Watch: GovTech-side confirmation and deployment scope; whether other governments follow the fine-tuned-open-model-for-fraud template; publication of the fine-tuning details.

Links: GoogleDeepMind announcement

5. Anthropic opens Singapore office — fifth in Asia-Pacific — and signs AI-safety MOU with the government

What happened: Anthropic announced October 14 (~4:40 AM ET) that it has opened an office in Singapore — its fifth in the Asia-Pacific region — and has signed an MOU with the Government of Singapore on AI safety and responsible adoption, with a link to a company news page (anthropic.com/news/singapore-office-opening) for details.

What's genuinely new: Debut — the record has no existing thread on Anthropic's international office expansion or government MOUs, and this starts a new storyline. The framing is notable given the window's other Singapore item: Google DeepMind's Gemma-based on-device fraud-detection work with GovTech (opened October 13, ~10:05 PM ET) — Singapore is where both rival labs are anchoring government-facing AI narratives this week. The capture is announcement-grade: no office headcount or leadership, no MOU text, and no statement from the Singapore government in the window.

Watch: The office's stated remit and staffing; the MOU's actual text and what "responsible adoption" concretely commits either party to; whether the MOU produces named programs (evaluators, adoption pilots) rather than a signing photo; whether OpenAI, Google, or Meta announce comparable Singapore government arrangements.

Links: Anthropic announcement · Anthropic news page

6. Perplexity Computer: the Ultrafast tier goes live, and Polymarket becomes a queryable connector

What happened: Two Computer capability items from Arav on October 14. (1) Ultrafast inference (~1:58 PM ET): "Ultrafast inference of OpenAI frontier models now supported on Computer. Try it out. It's super fast. Credit drain warning" — Perplexity: "Ultrafast versions of GPT-6 Astra and SOL 6.1 are now available in Perplexity Computer for Max subscribers. Both generate tokens faster, completing long Computer tasks sooner. Choose Astra or SOL 6.1 under Custom, then set speed to Ultrafast." (2) Polymarket (~2:11 PM ET): "You can now tag @Polymarket inside Perplexity Computer for market and real time trends research tasks" — Perplexity Finance: Computer "now pulls real-time data directly from @Polymarket's API," with mentioning or @-tagging Polymarket in a query yielding "market-implied probabilities, trading volume, and liquidity across 250,000+ active prediction markets."

What's genuinely new: Development of "ai — Perplexity: search index as agent infrastructure" — the Ultrafast tier is the "Ultrafast" name from OpenAI's own October 14 Amp case study (per the GPT-6 Astra rollout storyline) surfacing as an actual speed seting inside a third-party agent harness, and the Polymarket connector is the record's first prediction-market-class integration on Computer's connector ledger — a market-data rail beside the thread's existing agent-trading record (Coinbase telemetry, multi-agentic Coinbase trading). Both are vendor announcements that the feature is live in Computer; the credit mechanics behind the "credit drain warning" and any Polymarket connector pricing are not in the window, and nothing here is independently verified.

Watch: The credit-multiplier mechanics behind "credit drain warning" and Ultrafast pricing; whether the 250,000+ active-prediction-markets figure holds and which query classes hit the connector; whether OpenAI's own Ultrafast tier gets first-party documentation; whether Arav follows the Claude Code Desktop-style changelog discipline on either item.

Links: Arav: Ultrafast on Computer · Arav: Polymarket tagging

7. Agent-adoption economics: OpenAI reports the EU agentic shift; Perplexity counters with an America-first metrics frame

What happened: OpenAI posted October 14 (~2:45 PM ET) that "European businesses are rapidly expanding their use of AI agents: More than half of the AI use among our EU enterprise and business customers now involves these tools, and the number of small business users has grown 8x across the region" — linking its latest Economic Research report, "The EU's next chapter in AI is agentic" (openai.com/index/the-eus-next-chapter-in-ai-is-agentic), "exploring Europe's shift toward agentic AI." Roughly 74 minutes later (~3:59 PM ET) Aravind Srinivas posted the counter-frame: "America is the market to build for 🇺🇸" — "More than 50% of our users spending money on Automations (recurring agentic tasks) or on the higher tier plans are American," and "Net dollar retention for an American self serve business onboarded on agents is consistently high" — quoting Molly O'Shea's argument that you can build a trillion-dollar company on America alone ("Berkshire Hathaway is worth a trillion. JPMorgan is worth almost a trillion, like $940B, & largely the customer bases of these companies are in the US").

What's genuinely new: Debut cluster on this record — no existing thread covers regional/segment agentic-adoption economics (the nearest storyline header, "ai — The AI economy: inflation reads and capex math (watch)," covers inflation and capex, not adoption reads). For OpenAI the delta is quantitative beyond prior general "adoption growing" framing: the >50%-agents and 8x-small-business figures plus a named Economic Research report. For Perplexity, the >50%-American payer composition and the US self-serve NDR claim are both new to the record. All figures are vendor-self-reported; no independent adoption data or the report's own tables appear in the window.

Watch: The Economic Research report's EU enterprise list, definitions of "agent use," and methodology; whether Perplexity's composition and NDR figures get backed by a blog post or filing; whether the implicit US-vs-EU contrast hardens into deliberate positioning; neutral per-platform splits of agentic usage and spend.

Links: OpenAI announcement · OpenAI Economic Research report · Arav: America is the market to build for

8. GPT-6 Astra to 98% in an independent evaluation — after the model challenged the answer key

What happened: steipete RT'd October 14 (~10:25 AM ET) Nirmalya Kajuri's account of an arxiv paper (arxiv.org/abs/2610.16733, linked in-thread by Pratik Rath): "Our initial evaluation suggested that GPT-6-Astra's score was in the late eighties followed closely by Opus 5.5. But, as we started to investigate the mistakes that the models might have made, the models, especially Astra, insisted that they were right and that it was the expert-supplied answers that were erroneous. When we went back to the experts it turned out, in multiple instances, that … the experts found some subtle errors buried deep inside their calculations. As they revised their answers, Astra's score started to inch up and it finally ended up with 98% (49/50)."

What's genuinely new: Development of "ai — GPT-6 Astra rollout" — the thread's benchmark-verification genre gets its first independent academic-evaluation datapoint, and the story's substance is the friction pattern as much as the score: the model contesting expert answer keys, with the experts conceding "subtle errors buried deep inside their calculations." Capture is a single X relay of the paper; the paper itself is not read in this window, and the tweet text does not state the benchmark's identity or domain, the expert pool, or whether Opus 5.5 received the same re-scoring courtesy.

Watch: The paper's own read — benchmark identity, expert count and selection, how the re-scoring protocol worked, whether Opus 5.5 got the same treatment, and whether Astra's push-back rate is quantified rather than narrated; whether independent replications confirm both the score and the answer-key-error pattern.

Links: steipete RT of Nirmalya Kajuri · arxiv paper

9. Harness-layer day: Claude Code Desktop's "subsessions" update, then OpenClaw's ClawCast, CodexBar, and a Hugging Face co-host

What happened: (1) trq212 (October 14, ~4:13 PM ET): "Claude Code Desktop is now so much aware of its own capabilities and how it can work with you, like spinning off subsessions. I think of this almost like a v2, so many small details went into make it great" — above a ClaudeDevs note: "We've been working hard on making the Claude Code desktop app great for running lots of sessions side by side. It's now much quicker to start sessions, easier to split work across a few, and you can always see what each of them is doing. Here's what's new" (companion shoutout to @alicelovescake1 and @amorriscode, "the CCD goats," same minute). (2) OpenClaw's channel: ClawCast Episode 13 — "Live Q&A with Patrick and Hannes," broadcast posted ~2:29 PM ET; the episode's own post (~8:53 PM ET October 14) fills in its substance: "How to Use OpenClaw in Your Small Business" — @hrudolph and @Pat_Erichsen talking proposals, leads and reusable skills with @ArkStabler and @heyneighbor, with @Pishgar joining later to introduce the community forum (episode page: openclaw.ai/podcast/episode-13); steipete's RT of Jeremy Nguyen (~6:24 AM ET) on CodexBar — "CodexBar from @steipete (creator of OpenClaw!) lets you easily switch" between multiple Codex or Claude Code accounts without logging in and out; and OpenClaw's own announcement (~4:26 PM ET): "Join us and the @huggingface team this Friday for the Open Together: AI builders unite event!" (luma.com/OpenTogether) — "this Friday" from the October 14 window resolves to Friday, October 16, 2026.

What's genuinely new: Development of "ai — Agent-harness race: OpenClaw 2.0 and xAI Grok Build" — the thread's Claude Code line adds a desktop multi-session update (the "v2" framing is the user's, not a version number, and the itemized "Here's what's new" list is not in the window); CodexBar is third-party tooling on the harnesses' shared account-credential layer; ClawCast continues the OpenClaw channel cadence, with its first in-capture small-business content and a community-forum announcement tied to it; and the co-hosted OpenClaw × Hugging Face event ties the open-harness community to the open-model hub at a notable moment — NVIDIA's pending Hugging Face acquisition is on the record with a close prediction open. No Grok Bot or OpenClaw changelog movement in this window — channel and ecosystem, not product.

Watch: The full Claude Code Desktop changelog and whether "subsessions" gets documented officially; the October 16 "Open Together" outcome (announcements, integrations, NVIDIA-acquisition context); the scope and moderation of the newly announced OpenClaw community forum; whether CodexBar-class account-switcher tooling gets blessed or blocked by either vendor — the same account-credential layer Corey Ganim's cross-CLI route (Section 1) rides; named small-business integrators — the McBain solicitation (Dropped) begins to answer in-theme via Episode 13's small-business episode, with named guests @ArkStabler and @heyneighbor.

Links: trq212: Claude Code Desktop subsessions · trq212: the CCD goats · OpenClaw: ClawCast Ep 13 · OpenClaw: ClawCast Ep 13 episode content · steipete RT: CodexBar · OpenClaw: Open Together event

10. Grok Voice 3.0 launches under the SpaceXSI banner — #1 on Artificial Analysis' speech index, and already fielded at Starlink

What happened: Andrew Milich's launch announcement (Musk RT, ~1:05 PM ET October 14): "Grok Voice 3.0 launches today, our frontier speech-to-speech model" — "the world's most intelligent speech-to-speech model" that "excels at resolving customer calls, responds in under a second, and costs less than other frontier models." The announcement text is attributed to "SpaceX Super Intelligence" (@SpaceXSI), and Musk's own framing followed — "Grok is the best conversational voice model in the world" (~7:58 PM ET October 14). The channel's deployment-metrics layer arrived minutes after the launch: Sawyer Merritt (~1:11 PM ET) relays SpaceX saying @Starlink call centers use Grok Voice to answer 20,000 customer calls per day, "~70% are resolved with no human in the loop," and "one in four sales calls ends in a purchase before the caller hangs up," with Grok Voice 3.0 "handl[ing] millions of convos a day in Tesla vehicles" (a near-identical DogeDesigner relay, ~1:13 PM ET, adds that the agent "verifies customers, troubleshoots hardware, issues credits & ships replacement"). Parker Conrad (~1:24 PM ET): "Meet Grok Voice 3.0! … This is the same model that powers voice in Grok @bot, Grok in @Tesla, and @Starlink customer support. Now available to everyone!" Eric Jiang (~2:29 PM ET): "Grok Voice is one of our fastest growing segments on API," with "companies switching to Grok from providers like ElevenLabs" citing "better naturalness across languages and lower cost," and 3.0 "back at #1 on the AA Speech to Speech Index." The rank itself — Artificial Analysis (@SpaceXSI relayed ~5:32 PM ET): "SpaceXAI's Grok Voice 3.0 is the new leading Speech to Speech model, debuting at #1 on the Artificial Analysis Speech to Speech Index at 86.9, while also being competitive on latency." AshutoshShrivastava's relay (~3:47 PM ET) adds the competitive claim that it is "currently be[ing] GPT Live 1 and Gemini 3.8 Flash Live on all major benchmarks."

What's genuinely new: Debut — no storyline on the record covers speech or voice models, and this opens a new one. The launch's most striking framing is branding, not spec: the announcement is attributed to "SpaceX Super Intelligence" (SpaceXSI), one relay renders it "SpaceXAI," and another cites @SpaceXSI directly — folding a frontier-model debut into the SpaceX-banner naming on the record's superintelligence-rebrand storyline; the vendor attribution across the relays is inconsistent, and nothing in the window is independently verified. The Starlink 20,000-calls/day, ~70%-no-human, and 1-in-4-sales-conversion figures are all relay- and vendor-self-reported; the sub-second latency and cheaper-than-rivals claims are the launch's own text; the 86.9/#1 placement is Artificial Analysis' say relayed through the same channel.

Watch: The model card and the benchmark tables behind the 86.9 #1 and the sub-second latency claim; an independent cost comparison against ElevenLabs-class providers; whether the Starlink 20,000-calls/day and Tesla "millions of convos a day" figures get company-confirmed beyond relay; whether the GPT Live 1 and Gemini 3.8 Flash Live comparisons hold up on neutral indices; why "SpaceXSI" rather than xAI branding carries the launch, and whether the naming settles; ElevenLabs/OpenAI/Google counter-moves in speech.

Links: Andrew Milich: Grok Voice 3.0 launch · Musk: best conversational voice model · Sawyer Merritt: Starlink call-center metrics · DogeDesigner: Starlink 20k calls/day relay · Parker Conrad: now available to everyone · Eric Jiang: API segment growth, ElevenLabs switching · Artificial Analysis: #1 at 86.9 · AshutoshShrivastava: benchmark claims relay

Brief

Dropped