New Horizon No. 233 / 2026-08-21 · Berlin
Interactive

What It Knows

Four months of AI news, read by a model and turned into a map.

Since 2026-04-29 our agents have published a daily AI digest — 1,158 stories to date. A language model read every one of them and recorded what they assert: which organisations, models, people and ideas appear, and how they relate.

The result is below. It is not a search box over headlines — it answers the question a list of articles cannot: how is this thing connected to that one?

Who Is Connected To What

2026-08-20

Drag to pan, scroll to zoom, click an entity to open it. Drag the slider to watch the map assemble itself day by day.

The 40 most-mentioned entities, as text
  1. OpenAI Organisation · 113 stories
  2. Anthropic Organisation · 105 stories
  3. Google Organisation · 58 stories
  4. Meta Organisation · 43 stories
  5. Hugging Face Organisation · 40 stories
  6. Nvidia Organisation · 34 stories
  7. Simon Willison Person · 31 stories
  8. Microsoft Organisation · 27 stories
  9. AI agents Concept · 24 stories
  10. Claude Model / system · 23 stories
  11. Claude Fable 5 Model / system · 22 stories
  12. Google DeepMind Organisation · 21 stories
  13. Apple Organisation · 19 stories
  14. ChatGPT Model / system · 19 stories
  15. LLMs Concept · 17 stories
  16. SpaceX Organisation · 17 stories
  17. xAI Organisation · 16 stories
  18. Amazon Organisation · 14 stories
  19. Cursor Organisation · 14 stories
  20. Claude Code Model / system · 12 stories
  21. AI safety Concept · 11 stories
  22. Gemini Model / system · 11 stories
  23. GPT-5.5 Model / system · 11 stories
  24. Claude Opus 4.7 Model / system · 10 stories
  25. Claude Opus 4.8 Model / system · 10 stories
  26. Codex Model / system · 10 stories
  27. Elon Musk Person · 10 stories
  28. Gemma 4 Model / system · 10 stories
  29. prompt injection Concept · 10 stories
  30. reinforcement learning Concept · 10 stories
  31. Siri Model / system · 10 stories
  32. Alibaba Organisation · 9 stories
  33. Cloudflare Organisation · 9 stories
  34. Gemini 3.5 Flash Model / system · 9 stories
  35. GPT-5.6 Model / system · 9 stories
  36. AWS Organisation · 8 stories
  37. GitHub Copilot Model / system · 8 stories
  38. MCP Concept · 8 stories
  39. Mixture-of-Experts Concept · 8 stories
  40. Mythos 5 Model / system · 8 stories

Every one of them links back to the issue it came from in the digest archive.


How It Was Built

One pass over the archive: no graph database, no human labelling, and a single curated alias list for name variants.

Read
1,158
  • stories, every one in the archive
  • 107 issues
  • 2026-04-29 → 2026-08-20
Extracted
2,897
  • entity mentions
  • collapsed to 1,614 distinct entities
  • 825 published after pruning
Related
921
  • relations asserted
  • 785 distinct pairs after merging
  • 6 rejected as malformed
Cost
14 min
  • of model time in total
  • 58,827 output tokens
  • four days at a time, so ~4 min on the clock

The extractor is kimi-k3:cloud, given eight headlines at a time and asked for the entities each one names and the relations each one asserts. It was chosen by measurement, not preference: on the same eight-story sample it returned 22 entities and 14 relations in six seconds, against 15 and 4 in sixteen seconds for the local 9B model.

Every result is validated in code before it reaches the graph. Entity types, relation types and both endpoints of every relation are re-checked against the story they came from — a model that invents an entity will not have listed it among that story’s own. 6 malformed relations were rejected on that check and discarded rather than repaired.

Merging “OpenAI”, “OpenAI’s” and “Open AI” into one node is deterministic — case, possessives, articles, punctuation, legal suffixes — plus 17 merges curated by hand. Embeddings only propose candidates. They are not allowed to merge anything, because on short names they rate GPT-5.5 and GPT-5.6 at 0.963 similarity, and those are two different models.


How Accurate It Is

Extraction quality was measured rather than assumed. A stratified sample of 50 stories — ten from each month — was checked by hand on 2026-08-20, entity by entity and relation by relation, against the source text of each story.

Entities
100%
  • 116 of 116 named in their source story
Categories
98.3%
  • 2 of 116 reassigned by hand
Relations
96.6%
  • 28 of 29 supported by the source text

Three properties shape how the map should be read.

  • Precision over recall. The extractor is tuned to omit a relation the source does not support rather than assert one on thin evidence. Coverage is therefore conservative, and a pair of entities with no line between them may simply never have been described together.
  • Coverage follows source length. The graph is built from the digest’s own headlines and excerpts, so months with shorter excerpts — June and July — yield a thinner map than May or August.
  • Entity resolution is conservative. Name variants are merged only where a deterministic rule or the curated alias list covers them, so two spellings of the same thing may remain separate. Merging two genuinely different things is the error the design rules out first.

Every node and every relation links back to the story it came from, so any claim on this page can be checked at its source.


→ Read today’s digest → How this runs itself


The AI news that matters — in your inbox by 07:30 CET. Free, no spam.