AI social digest — Sep 28, 2026
Top line: The X layer's day: NVIDIA's safety-platform launch with Perplexity pledging open source and publishing its SPACE red-team, Flight 14's orbital insertion crossing the feed in real time, Sonnet 5.5's announcement, and OpenClaw's business-model question getting a hand-wavy answer from its own co-founder.
Perplexity × NVIDIA: the safety layer, with receipts promised
- Arav: "Safety is an engineering problem… building safe and secure agent sandboxes with the right guardrails. And we intend to open source all of it." The detail layer: Perplexity's security team "spent the last month trying to break SPACE, the sandbox platform that runs all of Perplexity Computer" — 9 models (Opus 5, GPT-5.6 Sol, Kimi K3, Gemini 3.1 Pro) given root access inside the VM; the writeup of what broke is the receipt to watch. Jensen's RT names the platform: OpenShell and Sentry, 100+ partners.
- Also from Perplexity: the Agent API grows Profiles, Skills, and managed connectors — configurable reusable agents; and a research call (continual learning, multi-agent parallel workers, synthetic-data evals) that got LeCun's RT. The perplexity-index thesis keeps compounding.
- Links: x.com/AravSrinivas · x.com/AravSrinivas · x.com/AravSrinivas
Flight 14 crossed the feed in real time (cross-beat)
- Musk's own account carried the launch: "launches in ~30 mins" (8:14 AM ET), then the SpaceX RTs — "Starship performs its orbital insertion burn and enters orbit of Earth for the first time" — then Sawyer Merritt's congratulations framing it as "Starship's first revenue-generating flight." Outcome per wire copy: orbit achieved, 26 V3 Starlinks deployed, mission truncated after an engine shutdown; ship lost on return. The space beat owns the full read.
- Links: x.com/elonmusk · x.com/elonmusk
Sonnet 5.5 announced
- Anthropic's account: "Claude Sonnet 5.5 is now available… a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% [less]." The mid-band's price war is on — full read in the ai digest.
- Link: x.com/AnthropicAI
OpenClaw's business model, still unanswered
- nichochar on co-founder Noah's economics answer: "Apple takes a 30% cut, Amazon takes 10%, we don't yet know where we are going to be along this curve" — his verdict: "a little hand-wavy… ignoring costs is a little tricky." The open harness's monetization question stays open days after Microsoft's Autopilot launch made it urgent — the free-to-use commons now has a $2T-adjacent company building on it.
- Link: x.com/nichochar
Smaller moves
- Agent-driven bank-run scenario (Apollo's chief economist, Musk RT): agents sweeping household cash into 3-5% accounts could drain bank deposits — an agent-economics thought experiment; sentiment, not modeling.
- Cursor drops server-side RAG for grep-like agent search ("no longer computing embeddings of your code") — the code-search tooling trend: agent tool calls beating indexes.
- LeCun lab's "Temporal Straightening for Latent Planning" — brain-inspired latent planning doubling goal-finding; the AMI Labs research line's first entry here.
- Joe's ex-OpenAI security retrospective (steipete RT) — an insider's article-length read on security/safety culture; worth a read alongside the containment-industry items.
- Kyle Polley's safety-engineering hiring note — the containment industry is hiring.