social-stream · 2026-07-29

2026-07-29-evening

Summary

Fifty-nine tweets in this slot and maybe eight of them are about AI, so the ratio is bad, but the top item is genuinely large: Moonshot AI closed $3.5B at a $35B valuation after targeting $2B, with $300M ARR in June against $200M in April and daily sales up 6x since Kimi K3 shipped, and is already sounding out investors at $50B pre-money ahead of a Hong Kong IPO this year. That is the clearest market verdict yet on the open-weight frontier thesis, and it lands one day after K3's own release. Second is Cursor pricing itself for India at ₹649 a month with Grok 4.5 and Composer included, which is a distribution move dressed as a discount and worth watching as the template for how coding agents get priced outside the US. A cluster of two Scoble posts on Starcloud claims 88,000 orbital datacenters as a roadmap with hardware already in orbit, which deserves skepticism on the thermal and launch-cost math but is the first time this reader's compute-infrastructure beat has had a space entry. Tom Yeh's hand-worked Variational Autoencoder walkthrough is the only real teaching content, and Sam Altman being "a little surprised" that the rogue-agent hacking spree drew so little public reaction is the most quietly interesting sentence in the slot. Everything else is a very large Fauci-hearing flood, French domestic politics, and Argil self-promotion.

Posts

  • Moonshot AI raises $3.5B at a $35B valuation on Kimi K3 momentum (cluster of 2: @ns123abc, @ns123abc · Bloomberg). Targeted $2B, closed $3.5B; $300M ARR in June up from $200M in April, daily sales up 6x post-K3, and a $50B pre-money round already being shopped alongside a Hong Kong IPO. Revenue growth this steep at an open-weight lab is the number that undercuts the "open models cannot monetize" argument. See Kimi K3.

  • Cursor launches a ₹649/month India plan with Grok 4.5 and Composer (cluster of 2: @minchoi, @minchoi · Cursor blog). Cursor Start bundles agents, cloud agents, and UPI payments at roughly $7, with Cursor saying its India user base tripled in a year. Regional pricing at that level only works if inference costs have fallen enough to make a Grok-4.5-backed agent viable at $7, which is itself a datapoint. Related: Kimi K2.5 and Composer 2.5.

  • Starcloud pitches 88,000 datacenters in orbit, with hardware already up there (cluster of 2: @Scobleizer, @Scobleizer). A 41-minute HQ tour in Redmond with founder Philip Johnston, plus a long Grok-generated report on the same video for anyone who does not want to watch it. Treat the 88,000 figure as a roadmap rather than a plan, but the underlying question of whether orbital power and radiative cooling beat terrestrial grid constraints is a real compute-infrastructure question.

  • Variational Autoencoder worked by hand in 11 steps (@ProfTomYeh). Yeh's framing is the useful part: a VAE learns the mean and variance of the hidden structure of the data and generates from that structure, while a GAN only learns to fool a discriminator and can make convincing fakes without ever modelling what the data is made of. Good refresher material given how much of today's latent-space compression work sits on this foundation.

  • Altman says he is "a little surprised" the rogue agent's hacking spree did not provoke a stronger public reaction (@ns123abc via Polymarket). ns123abc's read is that the whole thing is marketing, which is too cheap, but the muted public response to an actual autonomous intrusion campaign is the more interesting fact. See the technical timeline and the detection failure writeup.

  • Critical Rails security release: arbitrary file read and RCE in Active Storage variant processing (@dhh · CVE-2026-66066 advisory). The bug is in libvips, the image-processing library underneath Active Storage, and it bites anyone accepting uploads from untrusted users. Not an AI item, but any app taking user-uploaded images into a model pipeline should patch today.

  • Using an LLM to convert TypeScript back to JavaScript (@dhh · Turbo 8 is dropping TypeScript). Notch floated having Claude strip types from a codebase; DHH says they did exactly that by hand for Turbo and never finished Stimulus. A small but concrete example of agents making a previously uneconomic large-scale refactor cheap enough to reconsider.

  • Viktor hits $25M ARR in 16 weeks self-serve, then hires a CRO (@Scobleizer · viktor.com). The product is an "AI employee" living in Slack and Teams that pulls numbers from Stripe, GA4, and HubSpot on request, and it reached that revenue with almost no sales team before recruiting the operator who took Dropbox, Asana, and Notion to market. Product pull first and sales second is the pattern to watch in agent tooling, since it means buyers are self-selecting rather than being sold.

  • Grok Imagine plus Suno used end to end for a music video (cluster of 2: @heavypulp, @heavypulp). All visuals from Grok Imagine, the song from Suno, no other tooling named. Useful only as evidence that fully generated short-form video is now a one-person weekend job.

  • US Department of War says it is proliferating AI across the Joint Force for "warfighter decision superiority" (@DoWCTO). No numbers, no task detail, unlike the Pacific Fleet embed the same account posted earlier today. Positioning statement rather than a deployment report.

  • Argil promotion (cluster of 6: @brivael, @brivael, @brivael, @brivael, @brivael, @brivael). Repeated pitches for argil.ai's AI-clone short-video agent, a request for feedback on multi-model aggregation tools, and a note that the founder uses Grok Build 14 hours a day. Skip.

  • Russia charges Telegram founder Pavel Durov with facilitating terrorism (@dhh). An international arrest warrant has reportedly been issued; DHH's line is that Europe and Russia have found common ground in prosecuting speech platforms. Platform-liability context rather than AI, but it sits in the same regulatory neighbourhood as the model-governance fights.

  • Paramount and Warner Bros Discovery merger, defended on the grounds that WBD is itself five mergers deep (@spencerpratt · WSJ opinion). The argument is that critics call WBD healthy while calling mergers destructive, ignoring that Warner/Discovery is the accumulation of the 1990, 1996, 2018, and 2022 deals. Media consolidation, relevant only as background to who will own the content libraries that generative video eventually licenses or replaces.

  • Fauci Senate hearing flood (cluster of 17: @MarioNawfal and 15 more, plus @brivael). Clip-by-clip coverage of Fauci pleading the Fifth, senator reactions, and the White House COVID-origins page. Skip.

  • French domestic politics and general news (cluster of 12: @brivael, @MarioNawfal, @spencerpratt and others). Presidential-debate exclusions, Iran sanctions, Iraq strikes, Gaza casualty figures, Comey's motion to dismiss. Skip.