Manas Bihani
About

the questions

  1. What is a moat in an AI world?
  2. Why do AI products converge?
  3. What becomes scarce when intelligence becomes cheap?
  4. Does distribution matter more than technology?
  5. Why might human-made things become more valuable?
  6. What happens to expertise when everyone has the same models?
  7. Which parts of an AI startup are actually defensible?
  8. Where does value move when intelligence becomes commoditized?

everything on the desk

  1. The periodic table of the AI stackVisualization
  2. What is a moat when the model isn't yours?Note
  3. The problem-selection premiumNote
  4. Same model, different wiringNote
  5. Selection is the new bottleneckNote
  6. Get friendly with the AI raceEssay
  7. The convergence taxNote
  8. The luxury of realityNote
  9. The non-technical technical advantageNote
  10. The verification economyNote
  11. The bets against the wallVisualization
  12. You can't buy your way outVisualization
  13. How a chatbot writes one wordVisualization
  14. The grid is the last wallVisualization
  15. Who got paidVisualization
  16. Why this paper mattersExplainer
  17. Transformer: Why did transformers replace RNNs?Vaswani et al., NeurIPS 2017
  18. KV cache: Why does a long conversation get slower and cost more than a short one?Shazeer, 2019
  19. Mixture of experts: Why do some AI models have experts?Fedus, Zoph and Shazeer, 2021
  20. FlashAttention: Why is attention slow when the GPU is barely doing any arithmetic?Dao et al., NeurIPS 2022
  21. Mamba: Why does a model reread the whole conversation instead of just remembering it?Gu & Dao, 2023
  22. PagedAttention: Why does a GPU with free memory still refuse new requests?Kwon et al., SOSP 2023
  23. DeepSeek: How did DeepSeek train a frontier model so cheaply?DeepSeek-AI, 2024
  24. Jamba: Why does Jamba matter?Lieber et al., AI21 Labs, 2024
  25. BitNet: Why does BitNet matter?Ma et al., Microsoft Research, 2025
  26. DeepSeek-R1: Can a small AI model learn to reason like a huge one?DeepSeek-AI, 2025
  27. Kimi K2: Why does Kimi K2 matter?Kimi Team, Moonshot AI, 2025
  28. Sliding-window attention: How do models handle huge context windows without the memory bill exploding?Gemma Team, Google DeepMind, 2025
  29. How electricity becomes intelligenceVisualization
  30. This desk, as a datasetDataset
  31. The first version of this roomNote
  32. The aura dividendNote
  33. Distribution is rented attentionNote
  34. The Convergence TestNote
  35. A shelf for thinking about cheap intelligenceCollection
  36. Anatomy of an AI startupNote
  37. Six shocks to expertiseNote
  38. Nineteen Public KeysEssay
  39. The value migration machineModel
  40. AAA-Rated GPUsEssay
  41. Moats, before and afterVisualization
  42. The rhinoceros problemNote
  43. What becomes scarce when intelligence becomes cheap?Essay
  44. AI Has Passed Every Exam. It Has Never Had an Idea.Essay
  45. What Becomes Scarce After Intelligence?Essay
  46. India’s Carbon Markets : A New Test for Global Climate PolicyEssay
  47. Google Wants AI to Become BoringEssay
  48. The Wall That Wasn’t YoursEssay
  49. The Rate-Limiting StepEssay
  50. The Speed of Being WrongEssay
  51. Uber Burned a Year of AI Budget in Four Months. A Rat Catcher in 1902 Knew WhyEssay
  52. Finding a Flat in India Is Broken. We Have the Technology to Fix It. Nobody With Power Wants To.Essay
  53. Why We Can Never Have Good Social MediaEssay
  54. Gen Z Is Going OfflineEssay

rooms

  1. Home
  2. Writing
  3. Projects
  4. Reading & Watching
  5. All the questions
  6. Everything, as a contact sheet
  7. About

What I'm trying to figure out

Each question, with everything I've made while trying to answer it laid out next to it.

What is a moat in an AI world?

If a competitor can rebuild your product on the same model in a month, what exactly are you defending?

Investigating

What is a moat when the model isn't yours?
What is a moat when the model isn't yours?
The convergence tax
The convergence tax
memoryinterconnectgridGroqlicensedCerebraspublicCelestial AIacquiredAyar LabsprivateOklopublicFervopublic
The bets against the wall
packagingmemorytransformergrid queuecount the models
You can't buy your way out
Distribution is rented attention
Distribution is rented attention
ABCDEFGHB · A: 0.86C · A: 0.81C · B: 0.84D · A: 0.78D · B: 0.80D · C: 0.83E · A: 0.74E · B: 0.77E · C: 0.79E · D: 0.82F · A: 0.58F · B: 0.61F · C: 0.55F · D: 0.60F · E: 0.63G · A: 0.52G · B: 0.57G · C: 0.54G · D: 0.59G · E: 0.61G · F: 0.68H · A: 0.24H · B: 0.21H · C: 0.29H · D: 0.26H · E: 0.31H · F: 0.34H · G: 0.38
The Convergence Test
The Work of Art in the Age of Mechanical ReproductionThe Tacit DimensionPerforming Arts: The Economic DilemmaThe Innovator’s DilemmaInformation RulesTechnological Revolutions and Financial CapitalStrategy Letter VAggregation Theory7 PowersAttention Is All You NeedThe Bitter Lesson
A shelf for thinking about cheap intelligence
Anatomy of an AI startup
BeforeAfter cheap intelligenceFeatures & functionality: 6 → 1Technical know-how: 8 → 3Switching costs: 7 → 4Workflow lock-in: 7 → 5Scale economies: 6 → 6Regulatory licences: 5 → 6Network effects: 8 → 8Proprietary data: 5 → 7Distribution: 6 → 8Brand & trust: 5 → 9
Moats, before and after

Why do AI products converge?

Same models, same interface patterns, same launch videos. Is the sameness a phase, or where things settle?

Investigating

Same model, different wiring
Same model, different wiring
The convergence tax
The convergence tax
ABCDEFGHB · A: 0.86C · A: 0.81C · B: 0.84D · A: 0.78D · B: 0.80D · C: 0.83E · A: 0.74E · B: 0.77E · C: 0.79E · D: 0.82F · A: 0.58F · B: 0.61F · C: 0.55F · D: 0.60F · E: 0.63G · A: 0.52G · B: 0.57G · C: 0.54G · D: 0.59G · E: 0.61G · F: 0.68H · A: 0.24H · B: 0.21H · C: 0.29H · D: 0.26H · E: 0.31H · F: 0.34H · G: 0.38
The Convergence Test

What becomes scarce when intelligence becomes cheap?

When the marginal cost of competent thinking approaches zero, what sits next to it and gets more valuable?

Provisional answer

HHeLiBeBCNOFNeNaMgAlSiPSClArKCaScTiVCrMnFeCoNiCuZnGaGeAsSeBrKrRbSrYZrNbMoTcRuRhPdAgCdInSnSbTeIXeCsBaLaCePrNdPmSmEuGdTbDyHoErTmYbLuHfTaWReOsIrPtAuHgTlPbBiPoAtRnFrRaAcThPaUNpPuAmCmBkCfEsFmMdNoLrRfDbSgBhHsMtDsRgCnNhFlMcLvTsOg
The periodic table of the AI stack
Selection is the new bottleneck
Selection is the new bottleneck
Get friendly with the AI race
Get friendly with the AI race
The luxury of reality
The luxury of reality
The verification economy
The verification economy
memoryinterconnectgridGroqlicensedCerebraspublicCelestial AIacquiredAyar LabsprivateOklopublicFervopublic
The bets against the wall
packagingmemorytransformergrid queuecount the models
You can't buy your way out
Whydoesachatbotwriteonewordatwordone token, followed all the way through
How a chatbot writes one word
the last wall
The grid is the last wall
acceleratorpackagingrackmargin, while it binds
Who got paid
FlashAttention: Fast andMemory-Efficient ExactAttention withDao et al. · 2022AbstractSwitch Transformers:Scaling to TrillionParameter Models withFedus, Zoph and Shazeer · 2021AbstractFast TransformerDecoding: One Write-Headis All You NeedShazeer · 2019AbstractAttention Is All YouNeedVaswani et al. · 2017Abstract12 papers, and why each mattered
Why this paper matters
OAM-CLASS ACCELERATOR · PLAN≈1200 WCOMPUTE DIE ×2the constraintwent here next
How electricity becomes intelligence
The Work of Art in the Age of Mechanical ReproductionThe Tacit DimensionPerforming Arts: The Economic DilemmaThe Innovator’s DilemmaInformation RulesTechnological Revolutions and Financial CapitalStrategy Letter VAggregation Theory7 PowersAttention Is All You NeedThe Bitter Lesson
A shelf for thinking about cheap intelligence
The value migration machine
Dürer's engraving Melencolia I: a winged figure sits brooding among unused tools, a polyhedron, a sphere, an hourglass and a magic square.
What becomes scarce when intelligence becomes cheap?

Does distribution matter more than technology?

When anyone can build it, maybe the only question left is who can get it in front of people.

Open

Distribution is rented attention
Distribution is rented attention

Why might human-made things become more valuable?

Perfect copies once made originals more valuable. Could generated abundance do the same for things made by people?

Open

The luxury of reality
The luxury of reality
The verification economy
The verification economy
The aura dividend
The aura dividend
The Work of Art in the Age of Mechanical ReproductionThe Tacit DimensionPerforming Arts: The Economic DilemmaThe Innovator’s DilemmaInformation RulesTechnological Revolutions and Financial CapitalStrategy Letter VAggregation Theory7 PowersAttention Is All You NeedThe Bitter Lesson
A shelf for thinking about cheap intelligence
  1. 1455Gutenberg’s Bible
  2. +296 yrs1751The Encyclopédie, volume I
  3. +88 yrs1839The daguerreotype is announced
  4. +133 yrs1972The HP-35 pocket calculator
  5. +26 yrs1998Google is founded
  6. +24 yrs2022ChatGPT is released
Six shocks to expertise

What happens to expertise when everyone has the same models?

If the median expert is available to everyone for free, where does the edge go?

Investigating

The problem-selection premium
The problem-selection premium
Selection is the new bottleneck
Selection is the new bottleneck
The non-technical technical advantage
The non-technical technical advantage
  1. 1455Gutenberg’s Bible
  2. +296 yrs1751The Encyclopédie, volume I
  3. +88 yrs1839The daguerreotype is announced
  4. +133 yrs1972The HP-35 pocket calculator
  5. +26 yrs1998Google is founded
  6. +24 yrs2022ChatGPT is released
Six shocks to expertise
Dürer's 1515 woodcut of a rhinoceros covered in armour-like plates, with a small extra horn on its back.
The rhinoceros problem

Which parts of an AI startup are actually defensible?

Take a typical AI company apart, layer by layer. Which layers could a well-funded rival not copy?

Investigating

What is a moat when the model isn't yours?
What is a moat when the model isn't yours?
Same model, different wiring
Same model, different wiring
Anatomy of an AI startup
BeforeAfter cheap intelligenceFeatures & functionality: 6 → 1Technical know-how: 8 → 3Switching costs: 7 → 4Workflow lock-in: 7 → 5Scale economies: 6 → 6Regulatory licences: 5 → 6Network effects: 8 → 8Proprietary data: 5 → 7Distribution: 6 → 8Brand & trust: 5 → 9
Moats, before and after

Where does value move when intelligence becomes commoditized?

Follow the money as the price of thinking falls. Who ends up holding it?

Investigating

HHeLiBeBCNOFNeNaMgAlSiPSClArKCaScTiVCrMnFeCoNiCuZnGaGeAsSeBrKrRbSrYZrNbMoTcRuRhPdAgCdInSnSbTeIXeCsBaLaCePrNdPmSmEuGdTbDyHoErTmYbLuHfTaWReOsIrPtAuHgTlPbBiPoAtRnFrRaAcThPaUNpPuAmCmBkCfEsFmMdNoLrRfDbSgBhHsMtDsRgCnNhFlMcLvTsOg
The periodic table of the AI stack
The problem-selection premium
The problem-selection premium
Get friendly with the AI race
Get friendly with the AI race
the last wall
The grid is the last wall
acceleratorpackagingrackmargin, while it binds
Who got paid
OAM-CLASS ACCELERATOR · PLAN≈1200 WCOMPUTE DIE ×2the constraintwent here next
How electricity becomes intelligence
The Work of Art in the Age of Mechanical ReproductionThe Tacit DimensionPerforming Arts: The Economic DilemmaThe Innovator’s DilemmaInformation RulesTechnological Revolutions and Financial CapitalStrategy Letter VAggregation Theory7 PowersAttention Is All You NeedThe Bitter Lesson
A shelf for thinking about cheap intelligence
The value migration machine
Dürer's engraving Melencolia I: a winged figure sits brooding among unused tools, a polyhedron, a sphere, an hourglass and a magic square.
What becomes scarce when intelligence becomes cheap?