Tickline News

AI social digest — 2026-09-15

2026-09-15 AI social

Top line: The X layer is agent-economy week: Perplexity says two engineers plus hundreds of persistent Computer agents built a DynamoDB replacement in two months ("saves up to $100M/year"), and SpaceXAI's 72-hour "build a company with Grok Bot" livestream went live — the spectacle its own bot watched and took notes on. Underneath, the pacing fight's backlash entered week two with Jensen Huang calling the 10% claim "made up," and the AI-control school finally answered Amodei's essay with a 13,000-word counter-architecture.

Perplexity built its DynamoDB replacement with agents — two engineers, hundreds of "Computers," a claimed $100M/year (new thread → perplexity-index)

What happened: Arav Srinivas (~4:23 PM ET pickup, tweet ~12:30 PM ET): "We built a replacement for AWS DynamoDB, a key-value database for fast web content fetches… two engineers and hundreds of persistent Computer agents over two months. Migrating to our in-house database will save us up to a hundred million dollars yearly." The system: CobbleDB, Perplexity's key-value store serving web content for search; research on how it was built is being published.

What's genuinely new: the first concrete claim of core infrastructure built by agent workforces at a frontier-adjacent company — not agents assisting engineers, agents as the workforce with two humans supervising. All claims are self-reported (the $100M/yr is "up to," unverified; "hundreds of agents" is their telemetry), but the artifact — a database serving production search traffic — is checkable, and they're publishing the research.

Why it matters: this is the Coinbase agent-telemetry thread's next datapoint: agents doing work with a balance-sheet line attached. If the paper holds up, "two engineers + an agent fleet" becomes the reference configuration every infra team will be asked about.

Links: https://x.com/AravSrinivas/status/2099957318935028173 — CobbleDB research to follow

SpaceXAI's Galaxy livestream goes live: three employees, one company, 72 hours, Grok Bot as staff (continuation — agent-harness race, the watch item executes)

What happened: Day 1 of the Sep 15–17 "build a company in 3 days with Grok Bot" broadcast went live (~11:57 AM ET; SpaceXAI employees Matt Palmer, Lauren Tan, Roshan Sadanani — "start with research, a plan, and a product… Live now, plus sessions for engineering, product, and founders"). The meta-moment of the day: an attendee used Grok Bot to watch the Grok Bot livestream in real time — "pulling out the interesting parts, grabbing screenshots" — the product demoing itself as its own audience's note-taker.

What's genuinely new: the pre-hyped event is now measurable. The tells we set on Sep 13 hold: human-intervention frequency (how much of the "company" is employees demoing vs bots producing), whether a real product with real users emerges by Day 3, and — the one the stream can't answer — seat retention. Days 2–3 run through tomorrow.

Why it matters: this is SpaceXAI's consumer-agent GTM bet at full volume, three weeks into Grok Bot's public life — and the direct spectacle-marketing rival to Meta's Muse and Perplexity's Computer. Whatever actually ships becomes the first third-party-checkable workload claim for the platform.

Links: https://x.com/i/broadcasts/1AxRnZbVpjaxl (SpaceXAI RT via @Grok Bot)

"What does it mean to pace the frontier?" — the AI-control school answers Amodei (continuation — risk-governance, the substantive counter)

What happened: Sayash Kapoor (Princeton, with Arvind Narayanan — the AI as Normal Technology authors) posted a 13,000-word essay (pickup 6:53 PM ET Sep 14) analyzing "the loss-of-control incidents at AI companies" — a month of work, their most substantial safety writing since Normal Technology. Core moves: (1) bridge the cybersecurity-vs-AI-safety polarization — the incidents are both a crisis and a basic-security failure; (2) control over alignment as the higher-yield investment ("marginal investments in control are more likely to be effective"); (3) cyberoffense as the one urgent, specific risk worth defending against now; (4) organizational governance as the pacing tool — "individual teams should not be able to run potentially dangerous experiments without oversight… If putting these processes in place requires pausing some experiments, companies should do so."

What's genuinely new: the pacing debate finally has a policy architecture that doesn't depend on the labs' own promises — governance and control interventions that exist today, applied to the actual incident record (HF breach, DseWiki, RubyGems). Notable provenance: LeCun amplified it — the essay's loudest skeptic platforming a control-first critique, which is a different genre from his RT-a-day mockery.

Why it matters: Amodei's essay proposed embedded evaluators and Altman proposed safety cases; Kapoor/Narayanan propose neither — they propose boring organizational process with real teeth. If the Sanders briefing or any bill borrows its language, this is the document that mattered this week.

Links: https://x.com/ylecun/status/2099775834861056270 (Kapoor RT via LeCun; essay linked in thread)

The backlash enters week two — now with Huang, and All-In as the venue (continuation — risk-governance discourse strand)

What happened: The counter-wave gained its most credible amplifier yet: Jensen Huang, on the record (via Rohan Paul RT, 7:30 AM ET pickup, LeCun amplifying): of the 10%-extinction claim — "First of all, we shouldn't [try to explain a 10% chance of extinction], because it's made up. These are well-educated…" — i.e., the number has "no data, no scientific grounding." The All-In Summit (Sep 14, episodes dropping since) hosted the whole spectrum: Trump calling in live during Huang's interview ("They agree that AI is safe, grt for jobs, economic growth & natl security"); Nadella on "earning permission" and safety/control; Musk & Shotwell on "AI Risks and Peer Review." Meanwhile the motive war ground on: Shyam Sankar — "There is an unseen hand pushing AI safety as a political ideology. Effective Altruists believe a tiny group of technocrats should decide" (7:43 PM ET Sep 14); Kevin Bass claiming to have "conducted an audit of Anthropic's finances. What I have found is so shocking…" (6:10 PM ET Sep 14) and calling Dario/Moskovitz "primarily ideological"; Roemmele — "The Anthropic grift machine! FED BY FEAR THEATER" (8:24 PM ET Sep 14); Derya Unutmaz (via LeCun) on being "really really angry" at the virus-fatalism framing after two decades with HIV.

What's genuinely new: Huang is the first Fortune-5 CEO to call the quantified x-risk claim "made up" on-platform — the Nvidia position ("compute is revenue") now has its safety-denial line, and it's the same venue (All-In) where the White House said no. Bass's "audit of Anthropic's finances" is a new escalation genre — an unverified financial-forensics claim aimed at the IPO frame; treat as insinuation until anything checkable surfaces. Coxon: still zero actual reporting, day 6.

Why it matters: the discourse is now fully motive-and-character on one side and refusal-to-engage on the other, with the technical middle (Kapoor, above) the only strand producing architecture. The smear genre is no longer fringe — it has a Nobel-adjacent CEO and an All-In stage.

Links: Rohan Paul RT (via @ylecun) · All-In Summit episodes (Nadella, Musk/Shotwell, Trump–Huang call-in) · Sankar/Bass/Roemmele via LeCun RT chain

Also

Dropped