Jin MiaoSignals
← July 11, 2026

Start of day · analyzed 2026-07-11 06:38:13 PT

Morning brief

Saturday, July 11, 2026

Overnight developments and what deserves attention today.

41sources scanned
33new signals
16edge cases kept
5confirmed
ListenEnglish edition

📡 Jin Miao Signals — Morning Brief · 2026-07-11

1. Top 5 — what actually matters today

  • Prismata: containing cross-site prompt injection in web agents — first serious "firewall" primitive for the exact failure mode that's blocking real agent deployment; if you ship browsing/computer-use agents, read this before your next release arxiv.
  • Xiaomi's MiMo v2.5 pushes hybrid SWA inference efficiency — the Asia-overnight story for engineers: sliding-window attention tuning that squeezes more tokens/sec out of the same silicon, the unglamorous work that actually moves unit economics mimo.xiaomi.com.
  • Meta pulls its Instagram AI feature after backlash — the average-user signal of the day: "use your public content to generate stuff" hit the consent wall and got yanked; the digital-identity/consent line is now a product constraint, not a philosophy debate techcrunch.
  • EdVisorly raises $13.3M Series A to automate college-transfer back-office [PRIORITY] — founder read: AI-native workflow plays into a specific institutional-pain niche keep clearing Series A; boring vertical + real buyer beats another horizontal copilot (⚠️ Rumor — Crunchbase exclusive, amount not yet in a primary filing) crunchbase.
  • Microsoft reports a 25% emissions jump, driven by AI datacenters — the markets/energy carryover: the capex-and-power bill of the buildout is showing up on the balance sheet; watch the datacenter-power complex as context, not a trade windowscentral.

2. New-direction sparks

  • A font humans can read but AI cannot — non-obvious because it inverts the whole adversarial-OCR arms race into a content-owner tool: type designed to be legible to eyes and opaque to scrapers/VLMs. Early, gimmicky, but points at a real "cognitive-sovereignty" primitive mixfont.
  • Reconstructing a closed-source LLM tokenizer from two chat-API oracles — if it holds, it's a quiet reminder that "closed" model surfaces leak structure through their APIs; a genuinely new angle on model IP and probing [reddit/r/deeplearning].

3. Threads worth watching

  • The human-vs-AI web-access boundary is hardening — Prismata (agents attacked by the open web), the residential-proxy/scraper escalation, and the ghost-font all pushed the same "who is allowed to read/act on this content" thread in one day lwn.

4. Contrarian watch

  • **"Latent reasoning without decoding" — an instrumented negative result** — against the consensus that skipping the token bottleneck buys you reasoning for free; a clean null is more useful than the tenth hype thread, and worth tracking as the reasoning-architecture debate matures [reddit/r/deeplearning].
  • Data, not the model, as the agent moat — NVIDIA's open "Data for Agents" push (imp5, ONGOING) cuts against model-centric roadmaps; the edge may be in agent-trajectory data, not the next checkpoint huggingface.

5. Verification flags

  • ⚠️ EdVisorly $13.3M Series A — do not act on yet — needs primary source (Crunchbase exclusive, no filing) crunchbase.
  • ⚠️ Paradigm's $1.2B "technical frontier" fund — do not act on yet; also ~3 days old, not fresh techcrunch.
  • ⚠️ Microsoft's 25% emissions figure — do not act on yet — needs the primary sustainability report, not secondary coverage windowscentral.
  • ⚠️ "GPT-2 fully decoded / black box fully open" — do not act on yet — social claim, no verified writeup [reddit/r/deeplearning].

Markets context only — not financial advice.

Private founder layer

Co-founder confidential

Strategic synthesis and adversarial review, encrypted in the page source.

Source ledgerEvery scored item, including outliers
  1. ConfirmedNEWOutlier
    i4 / e5
  2. RumorNEWOutlier
    GPT-2 Fully Decoded Internally Black Box Fully Open With Demoreddit/r/deeplearning
    i4 / e5
  3. RumorNEWOutlier
    I made a live visualizer for Anthropic's new "Jacobian lens" paper!reddit/r/deeplearning
    i4 / e5
  4. RumorNEWOutlier
    Can we reconstruct a closed-source LLM tokenizer using only two oracles from the chat API?reddit/r/deeplearning
    i4 / e5
  5. RumorNEWOutlier
    Latent reasoning without decoding: an instrumented negative result [R]reddit/r/deeplearning
    i4 / e5
  6. RumorNEWOutlier
    Dropped a 201M Masked Diffusion LM checkpoint on HF (Open code + weights). Seeking feedback on parallel text generation!reddit/r/deeplearning
    i4 / e5
  7. ConfirmedONGOINGOutlier
    i5 / e4
  8. ReportedNEWOutlier
    i4 / e4
  9. RumorONGOINGOutlier
    i4 / e4
  10. RumorONGOINGOutlier
    i4 / e4
  11. ReportedNEWOutlier
    i3 / e4
  12. ReportedNEWOutlier
    i3 / e4
  13. RumorNEWOutlier
    Predicting human preference for generated image pairs using HPSv3 [P]reddit/r/MachineLearning
    i3 / e4
  14. ReportedNEWOutlier
    i4 / e3
  15. ReportedNEWOutlier
    i2 / e4
  16. RumorONGOINGOutlier
    i3 / e3
  17. ConfirmedNEW
    i3 / e3
  18. ReportedNEW
    i3 / e3
  19. ReportedNEW
    i3 / e3
  20. ReportedNEW
    i4 / e2
  21. ReportedNEW
    i2 / e3
  22. RumorNEW
    Withdraw from ACL ARR and resubmit to a workshop? [D]reddit/r/MachineLearning
    i2 / e3
  23. ReportedONGOING
    i2 / e3
  24. ReportedNEW
    i2 / e3
  25. ReportedNEW
    i2 / e3
  26. RumorNEW
    i3 / e2
  27. ReportedNEW
    i3 / e2
  28. ReportedNEW
    i3 / e2
  29. RumorNEW
    i3 / e2
  30. ConfirmedONGOING
    i3 / e2
  31. ReportedNEW
    i3 / e2
  32. RumorNEW
    I built a variational AE with pytorch/PIL! Here is the model framework.reddit/r/deeplearning
    i2 / e2
  33. RumorNEW
    About Autonomous Model Trainingreddit/r/deeplearning
    i2 / e2
  34. ReportedONGOING
    i1 / e2
  35. ReportedNEW
    i1 / e2
  36. RumorNEW
    How does *ACL conferences acceptance work [D]reddit/r/MachineLearning
    i1 / e2
  37. RumorNEW
    Transformer Decoder from Scratchreddit/r/deeplearning
    i2 / e1
  38. ReportedONGOING
    i2 / e1
  39. ConfirmedNEW
    i1 / e1
  40. RumorNEW
    Self taught, how to advance?reddit/r/deeplearning
    i1 / e1
  41. RumorNEW
    I'm an undergraduate studying ai, should I commit sewer slide?reddit/r/deeplearning
    i1 / e1