Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 15, 2026

Transcript

AGI Dreams Podcast – April 15, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  OWASP's
• 0:00
  new
• 0:00
  Agentic
• 0:01
  Security
• 0:01
  Initiative
• 0:02
  Top
• 0:02
  10
• 0:02
  has
• 0:02
  landed,
• 0:03
  and
• 0:03
  the
• 0:03
  number-one
• 0:03
  entry
• 0:04
  is
• 0:04
  Agent
• 0:04
  Goal
• 0:05
  Hijack
• 0:05
  —
• 0:06
  the
• 0:06
  complete
• 0:07
  takeover
• 0:07
  of
• 0:07
  an
• 0:07
  AI
• 0:07
  agent's
• 0:08
  objectives
• 0:08
  through
• 0:09
  malicious
• 0:09
  instructions
• 0:09
  embedded
• 0:10
  in
• 0:10
  content
• 0:10
  the
• 0:10
  agent
• 0:11
  processes.
• 0:12
  Adversa
• 0:12
  AI's
• 0:13
  practical
• 0:13
  guide
• 0:13
  to
• 0:13
  ASI-01
• 0:16
  spells
• 0:16
  out
• 0:16
  why
• 0:17
  it
• 0:17
  earned
• 0:17
  the
• 0:17
  top
• 0:17
  slot:
• 0:18
  once
• 0:19
  an
• 0:19
  attacker
• 0:19
  redirects
• 0:20
  an
• 0:20
  agent's
• 0:20
  goal,
• 0:21
  every
• 0:21
  tool
• 0:21
  the
• 0:21
  agent
• 0:21
  holds
• 0:22
  —
• 0:23
  file
• 0:23
  access,
• 0:24
  email,
• 0:24
  shell,
• 0:25
  APIs
• 0:25
  —
• 0:26
  becomes
• 0:27
  a
• 0:27
  weapon
• 0:27
  pointed
• 0:27
  inward.
• 0:28
  Three
• 0:29
  reasons
• 0:29
  anchor
• 0:29
  the
• 0:29
  ranking.
• 0:31
  First,
• 0:31
  most
• 0:32
  other
• 0:32
  agentic
• 0:32
  risks
• 0:33
  (tool
• 0:33
  misuse,
• 0:34
  identity
• 0:34
  abuse,
• 0:35
  memory
• 0:36
  poisoning)
• 0:36
  are
• 0:36
  merely
• 0:37
  means
• 0:37
  of
• 0:37
  achieving
• 0:37
  goal
• 0:38
  hijack.
• 0:39
  Second,
• 0:40
  the
• 0:40
  blast
• 0:40
  radius
• 0:41
  is
• 0:41
  total:
• 0:42
  credentials,
• 0:42
  PII,
• 0:43
  and
• 0:43
  intellectual
• 0:44
  property
• 0:45
  can
• 0:45
  all
• 0:45
  be
• 0:45
  silently
• 0:46
  exfiltrated.
• 0:47
  Third,
• 0:48
  unmitigated
• 0:48
  goal
• 0:49
  hijack
• 0:49
  fundamentally
• 0:50
  caps
• 0:50
  how
• 0:50
  much
• 0:50
  autonomy
• 0:51
  organizations
• 0:51
  can
• 0:52
  safely
• 0:52
  grant
• 0:52
  their
• 0:52
  agents.
• 0:53
  The
• 0:54
  guide's
• 0:54
  attack-vector
• 0:54
  taxonomy
• 0:55
  classifies
• 0:56
  five
• 0:56
  boundary
• 0:56
  violations:
• 0:57
  user-input
• 0:57
  injection,
• 0:58
  retrieved-content
• 0:59
  injection
• 1:00
  (PoisonedRAG
• 1:00
  demonstrated
• 1:01
  90%
• 1:01
  success
• 1:02
  with
• 1:02
  just
• 1:02
  five
• 1:03
  documents),
• 1:04
  tool-output
• 1:04
  injection
• 1:05
  (a
• 1:05
  typosquatted
• 1:05
  MCP
• 1:06
  server
• 1:06
  that
• 1:06
  silently
• 1:07
  BCCs
• 1:07
  every
• 1:08
  email
• 1:08
  to
• 1:08
  the
• 1:08
  attacker),
• 1:09
  inter-agent
• 1:10
  session
• 1:10
  smuggling
• 1:10
  (Palo
• 1:11
  Alto
• 1:11
  Unit
• 1:11
  42's
• 1:12
  mid-session
• 1:12
  poisoning
• 1:13
  through
• 1:13
  agent-to-agent
• 1:14
  communication),
• 1:15
  and
• 1:15
  configuration
• 1:15
  tampering
• 1:16
  (a
• 1:18
  README.md
• 1:20
  that
• 1:20
  tricks
• 1:20
  an
• 1:20
  agent
• 1:21
  into
• 1:21
  setting
• 1:21
  autoApprove:
• 1:22
  true
• 1:22
  and
• 1:23
  piping
• 1:23
  shell
• 1:23
  commands).
• 1:24
  The
• 1:24
  most
• 1:25
  dangerous
• 1:25
  are
• 1:25
  zero-click:
• 1:26
  payloads
• 1:26
  sit
• 1:27
  dormant
• 1:27
  in
• 1:27
  email
• 1:27
  or
• 1:27
  document
• 1:28
  stores
• 1:28
  for
• 1:28
  weeks
• 1:29
  until
• 1:29
  RAG
• 1:29
  retrieves
• 1:29
  them
• 1:30
  into
• 1:30
  the
• 1:30
  context
• 1:30
  window.
• 1:31
  As
• 1:31
  Johann
• 1:32
  Rehberger
• 1:32
  put
• 1:32
  it:
• 1:32
  "Prompt
• 1:33
  injection
• 1:33
  cannot
• 1:33
  be
• 1:33
  'fixed.'
• 1:34
  As
• 1:35
  soon
• 1:35
  as
• 1:35
  a
• 1:35
  system
• 1:35
  is
• 1:35
  designed
• 1:36
  to
• 1:36
  take
• 1:36
  untrusted
• 1:36
  data
• 1:37
  and
• 1:37
  include
• 1:37
  it
• 1:37
  into
• 1:38
  an
• 1:38
  LLM
• 1:38
  query,
• 1:39
  the
• 1:39
  untrusted
• 1:39
  data
• 1:40
  influences
• 1:40
  the
• 1:40
  output."
• 1:41
  The
• 1:41
  guide
• 1:41
  points
• 1:42
  to
• 1:42
  Google
• 1:42
  DeepMind's
• 1:43
  CaMeL
• 1:43
  architecture
• 1:43
  —
• 1:44
  separating
• 1:44
  a
• 1:45
  trusted
• 1:45
  control
• 1:45
  plane
• 1:46
  from
• 1:46
  an
• 1:46
  untrusted
• 1:46
  data
• 1:47
  plane
• 1:47
  so
• 1:47
  retrieved
• 1:47
  content
• 1:48
  can
• 1:48
  never
• 1:48
  affect
• 1:49
  program
• 1:49
  flow
• 1:49
  —
• 1:50
  as
• 1:50
  the
• 1:50
  most
• 1:50
  rigorous
• 1:51
  defense,
• 1:51
  achieving
• 1:52
  77%
• 1:53
  task
• 1:53
  completion
• 1:53
  with
• 1:53
  formal
• 1:54
  security
• 1:54
  guarantees.
• 1:55
  On
• 1:55
  the
• 1:56
  tooling
• 1:56
  side,
• 1:57
  Josh
• 1:57
  Grossman's
• 1:57
  team
• 1:58
  has
• 1:58
  released
• 1:58
  AGHAST
• 1:59
  (AI-Guided
• 1:59
  Hybrid
• 1:59
  Application
• 2:00
  Static
• 2:00
  Testing),
• 2:01
  an
• 2:01
  open-source
• 2:02
  framework
• 2:02
  that
• 2:03
  combines
• 2:03
  static
• 2:03
  code
• 2:03
  discovery
• 2:04
  with
• 2:04
  AI-generated
• 2:05
  prompts
• 2:06
  to
• 2:06
  find
• 2:06
  repository-specific
• 2:07
  and
• 2:07
  company-specific
• 2:08
  security
• 2:08
  issues.
• 2:09
  AGHAST
• 2:10
  is
• 2:10
  deliberately
• 2:11
  not
• 2:11
  an
• 2:11
  "auto-scan
• 2:12
  everything"
• 2:12
  tool
• 2:12
  —
• 2:13
  it
• 2:13
  gives
• 2:13
  security
• 2:14
  teams
• 2:14
  a
• 2:14
  way
• 2:14
  to
• 2:14
  encode
• 2:15
  suspicions
• 2:15
  about
• 2:16
  what
• 2:16
  vulnerabilities
• 2:16
  might
• 2:17
  exist
• 2:17
  and
• 2:17
  turn
• 2:17
  them
• 2:18
  into
• 2:18
  repeatable,
• 2:19
  scalable
• 2:19
  validations
• 2:20
  across
• 2:20
  codebases,
• 2:21
  returning
• 2:21
  results
• 2:22
  in
• 2:22
  a
• 2:22
  structured
• 2:22
  format.
• 2:23
  Static
• 2:24
  discovery
• 2:24
  plus
• 2:24
  AI
• 2:25
  guidance
• 2:25
  is
• 2:25
  a
• 2:25
  more
• 2:25
  honest
• 2:26
  framing
• 2:26
  than
• 2:26
  the
• 2:26
  typical
• 2:27
  "AI
• 2:27
  finds
• 2:27
  all
• 2:28
  your
• 2:28
  bugs"
• 2:28
  marketing.
• 2:29
  Meanwhile,
• 2:29
  the
• 2:30
  gap
• 2:30
  between
• 2:30
  threat
• 2:30
  volume
• 2:31
  and
• 2:31
  defensive
• 2:31
  spending
• 2:32
  continues
• 2:32
  to
• 2:32
  widen.
• 2:33
  CipherCue's
• 2:34
  2025
• 2:35
  tracking
• 2:35
  data
• 2:35
  shows
• 2:36
  ransomware
• 2:36
  leak-site
• 2:37
  claims
• 2:37
  hit
• 2:37
  7,760
• 2:39
  for
• 2:39
  the
• 2:39
  year,
• 2:40
  up
• 2:40
  from
• 2:40
  5,939
• 2:42
  in
• 2:42
  2024
• 2:42
  —
• 2:43
  a
• 2:43
  30.7%
• 2:45
  year-over-year
• 2:46
  increase.
• 2:46
  Gartner's
• 2:47
  July
• 2:47
  2025
• 2:48
  forecast
• 2:48
  put
• 2:49
  worldwide
• 2:49
  information
• 2:50
  security
• 2:50
  spending
• 2:50
  growth
• 2:50
  at
• 2:51
  10.1%,
• 2:52
  from
• 2:52
  $193
• 2:53
  billion
• 2:53
  to
• 2:54
  $213
• 2:55
  billion.
• 2:56
  That
• 2:57
  directional
• 2:57
  comparison
• 2:57
  —
• 2:58
  threat
• 2:58
  volume
• 2:59
  growing
• 2:59
  roughly
• 2:59
  three
• 3:00
  times
• 3:00
  faster
• 3:00
  than
• 3:00
  budgets
• 3:01
  —
• 3:02
  is
• 3:02
  the
• 3:02
  kind
• 3:02
  of
• 3:02
  stat
• 3:02
  that
• 3:03
  should
• 3:03
  appear
• 3:03
  in
• 3:03
  every
• 3:03
  CISO's
• 3:04
  board
• 3:04
  presentation.
• 3:05
  February
• 3:06
  2025
• 3:07
  alone
• 3:07
  produced
• 3:07
  1,050
• 3:08
  claims,
• 3:09
  more
• 3:09
  than
• 3:09
  double
• 3:09
  the
• 3:10
  prior
• 3:10
  February.
• 3:11
  The
• 3:11
  top
• 3:11
  five
• 3:12
  groups
• 3:12
  accounted
• 3:12
  for
• 3:12
  38.8%
• 3:13
  of
• 3:14
  claims,
• 3:14
  but
• 3:14
  126
• 3:15
  groups
• 3:16
  outside
• 3:16
  the
• 3:16
  top
• 3:16
  ten
• 3:17
  generated
• 3:17
  3,516
• 3:19
  claims
• 3:19
  between
• 3:19
  them,
• 3:20
  suggesting
• 3:20
  a
• 3:20
  broad
• 3:21
  and
• 3:21
  fragmenting
• 3:21
  threat
• 3:22
  landscape
• 3:22
  rather
• 3:22
  than
• 3:22
  a
• 3:22
  consolidating
• 3:23
  one.
• 3:24
  HHS
• 3:24
  OCR
• 3:25
  breach
• 3:25
  filings
• 3:25
  jumped
• 3:26
  from
• 3:26
  164
• 3:27
  to
• 3:27
  517
• 3:28
  in
• 3:28
  2025,
• 3:29
  the
• 3:29
  largest
• 3:30
  year-over-year
• 3:30
  increase
• 3:31
  of
• 3:31
  any
• 3:31
  source
• 3:31
  CipherCue
• 3:32
  tracks.
• 3:33
  The
• 3:33
  irony
• 3:33
  of
• 3:34
  tightening
• 3:34
  safety
• 3:34
  guardrails
• 3:35
  surfaces
• 3:35
  in
• 3:35
  a
• 3:35
  different
• 3:36
  corner.
• 3:37
  An
• 3:37
  offensive
• 3:37
  security
• 3:37
  professional
• 3:38
  reports
• 3:38
  being
• 3:38
  locked
• 3:39
  out
• 3:39
  of
• 3:39
  Claude
• 3:39
  after
• 3:40
  months
• 3:40
  of
• 3:40
  daily
• 3:40
  use
• 3:41
  for
• 3:41
  legitimate
• 3:41
  penetration
• 3:42
  testing.
• 3:43
  The
• 3:43
  account
• 3:43
  was
• 3:43
  flagged,
• 3:44
  downgraded,
• 3:44
  and
• 3:44
  guardrailed
• 3:45
  —
• 3:45
  and
• 3:46
  despite
• 3:46
  filling
• 3:46
  out
• 3:46
  Anthropic's
• 3:47
  cyber
• 3:47
  use
• 3:47
  case
• 3:47
  form
• 3:48
  and
• 3:48
  sending
• 3:48
  five
• 3:48
  follow-up
• 3:49
  emails
• 3:49
  over
• 3:49
  months,
• 3:50
  the
• 3:50
  practitioner
• 3:51
  has
• 3:51
  received
• 3:51
  zero
• 3:52
  response.
• 3:53
  This
• 3:53
  is
• 3:53
  the
• 3:53
  kind
• 3:54
  of
• 3:54
  friction
• 3:54
  that
• 3:54
  pushes
• 3:54
  security
• 3:55
  professionals
• 3:55
  toward
• 3:56
  less-guarded
• 3:56
  alternatives
• 3:57
  or
• 3:57
  open-weight
• 3:57
  models,
• 3:58
  which
• 3:58
  is
• 3:58
  arguably
• 3:59
  worse
• 3:59
  for
• 3:59
  everyone.
• 4:00
  When
• 4:00
  the
• 4:00
  people
• 4:00
  who
• 4:01
  find
• 4:01
  the
• 4:01
  vulnerabilities
• 4:02
  cannot
• 4:02
  use
• 4:02
  the
• 4:02
  tools
• 4:03
  that
• 4:03
  help
• 4:03
  them
• 4:03
  find
• 4:03
  vulnerabilities,
• 4:04
  something
• 4:05
  in
• 4:05
  the
• 4:05
  calibration
• 4:05
  is
• 4:06
  off.
• 4:09
  The
• 4:09
  agentic
• 4:09
  coding
• 4:10
  tool
• 4:10
  landscape
• 4:10
  continues
• 4:11
  diverging
• 4:11
  into
• 4:11
  two
• 4:12
  camps:
• 4:13
  minimal
• 4:13
  reference
• 4:13
  implementations
• 4:14
  and
• 4:14
  full
• 4:14
  workflow
• 4:15
  engines.
• 4:16
  Archon
• 4:16
  sits
• 4:16
  firmly
• 4:17
  in
• 4:17
  the
• 4:17
  latter.
• 4:18
  Billed
• 4:18
  as
• 4:18
  the
• 4:19
  first
• 4:19
  open-source
• 4:19
  "harness
• 4:20
  builder"
• 4:20
  for
• 4:20
  AI
• 4:20
  coding,
• 4:21
  it
• 4:21
  lets
• 4:22
  teams
• 4:22
  define
• 4:22
  development
• 4:23
  processes
• 4:23
  as
• 4:23
  YAML
• 4:24
  workflows
• 4:24
  with
• 4:24
  deterministic
• 4:25
  structure
• 4:25
  and
• 4:26
  AI
• 4:26
  intelligence
• 4:26
  at
• 4:27
  each
• 4:27
  node.
• 4:28
  "Like
• 4:28
  what
• 4:28
  Dockerfiles
• 4:29
  did
• 4:29
  for
• 4:29
  infrastructure
• 4:30
  and
• 4:30
  GitHub
• 4:31
  Actions
• 4:31
  did
• 4:31
  for
• 4:31
  CI/CD
• 4:32
  —
• 4:33
  Archon
• 4:34
  does
• 4:34
  for
• 4:34
  AI
• 4:34
  coding
• 4:34
  workflows."
• 4:36
  Each
• 4:36
  run
• 4:36
  gets
• 4:36
  its
• 4:37
  own
• 4:37
  git
• 4:37
  worktree
• 4:37
  for
• 4:37
  isolation,
• 4:38
  enabling
• 4:39
  parallel
• 4:39
  fixes
• 4:40
  with
• 4:40
  no
• 4:40
  conflicts.
• 4:41
  The
• 4:41
  project
• 4:42
  ships
• 4:42
  17
• 4:43
  default
• 4:43
  workflows,
• 4:44
  and
• 4:44
  its
• 4:44
  node
• 4:44
  system
• 4:45
  supports
• 4:45
  loops
• 4:45
  (iterate
• 4:46
  until
• 4:46
  tests
• 4:46
  pass),
• 4:47
  human
• 4:47
  approval
• 4:48
  gates,
• 4:48
  and
• 4:48
  mixing
• 4:49
  deterministic
• 4:49
  bash
• 4:50
  with
• 4:50
  AI
• 4:50
  reasoning.
• 4:51
  Platform
• 4:52
  adapters
• 4:52
  span
• 4:53
  CLI,
• 4:53
  web
• 4:54
  UI,
• 4:54
  Slack,
• 4:55
  Telegram,
• 4:56
  Discord,
• 4:56
  and
• 4:57
  GitHub.
• 4:58
  At
• 4:58
  the
• 4:58
  opposite
• 4:58
  end,
• 4:59
  CoreCoder
• 4:59
  distills
• 5:00
  Claude
• 5:00
  Code's
• 5:00
  512,000
• 5:02
  lines
• 5:02
  of
• 5:02
  TypeScript
• 5:03
  into
• 5:03
  roughly
• 5:03
  950
• 5:04
  lines
• 5:04
  of
• 5:04
  Python.
• 5:05
  The
• 5:06
  author
• 5:06
  reverse-engineered
• 5:07
  the
• 5:07
  source
• 5:07
  and
• 5:07
  rebuilt
• 5:08
  the
• 5:08
  load-bearing
• 5:08
  walls:
• 5:09
  search-and-replace
• 5:10
  editing
• 5:10
  (70
• 5:11
  lines),
• 5:12
  parallel
• 5:12
  tool
• 5:13
  execution
• 5:13
  via
• 5:13
  ThreadPool,
• 5:14
  three-layer
• 5:15
  context
• 5:15
  compression
• 5:16
  (145
• 5:17
  lines),
• 5:18
  sub-agent
• 5:18
  spawning
• 5:19
  (50
• 5:19
  lines),
• 5:20
  and
• 5:20
  dangerous-command
• 5:21
  blocking
• 5:21
  (95
• 5:22
  lines).
• 5:23
  CoreCoder
• 5:23
  positions
• 5:24
  itself
• 5:24
  as
• 5:24
  the
• 5:24
  "nanoGPT
• 5:25
  of
• 5:25
  coding
• 5:26
  agents"
• 5:26
  —
• 5:27
  not
• 5:27
  a
• 5:27
  production
• 5:27
  tool
• 5:28
  but
• 5:28
  a
• 5:28
  blueprint
• 5:28
  readable
• 5:29
  in
• 5:29
  one
• 5:29
  sitting,
• 5:30
  working
• 5:30
  with
• 5:30
  any
• 5:31
  OpenAI-compatible
• 5:32
  API.
• 5:33
  A
• 5:33
  parallel
• 5:33
  documentation
• 5:34
  effort
• 5:34
  takes
• 5:35
  a
• 5:35
  different
• 5:35
  approach
• 5:35
  to
• 5:36
  understanding
• 5:36
  Claude
• 5:36
  Code's
• 5:37
  internals.
• 5:38
  HitCC
• 5:38
  is
• 5:38
  a
• 5:38
  complete
• 5:39
  reverse-engineering
• 5:39
  knowledge
• 5:40
  base
• 5:40
  for
• 5:40
  Claude
• 5:40
  Code
• 5:41
  CLI
• 5:41
  v2.1.84
• 5:42
  —
• 5:43
  81
• 5:45
  files,
• 5:45
  27,170
• 5:47
  lines
• 5:47
  of
• 5:47
  structured
• 5:48
  documentation
• 5:48
  covering
• 5:49
  the
• 5:49
  startup
• 5:49
  entry
• 5:49
  chain,
• 5:50
  agent
• 5:50
  loop,
• 5:51
  tool
• 5:51
  execution,
• 5:52
  prompt
• 5:52
  assembly,
• 5:53
  MCP
• 5:53
  integration,
• 5:54
  skill
• 5:55
  system,
• 5:55
  session
• 5:56
  persistence,
• 5:56
  and
• 5:56
  TUI
• 5:57
  rendering.
• 5:58
  The
• 5:58
  project
• 5:59
  is
• 5:59
  explicit
• 5:59
  about
• 6:00
  what
• 6:00
  it
• 6:00
  is
• 6:00
  not:
• 6:01
  no
• 6:01
  source
• 6:01
  code,
• 6:02
  no
• 6:02
  cracking,
• 6:03
  no
• 6:03
  policy
• 6:03
  bypass.
• 6:05
  It
• 6:05
  is
• 6:05
  organized
• 6:05
  by
• 6:06
  runtime
• 6:06
  topics
• 6:06
  rather
• 6:07
  than
• 6:07
  original
• 6:07
  file
• 6:07
  structure,
• 6:08
  and
• 6:09
  the
• 6:09
  authors
• 6:09
  claim
• 6:09
  the
• 6:09
  documentation
• 6:10
  is
• 6:10
  sufficient
• 6:11
  to
• 6:11
  support
• 6:11
  "reconstructing
• 6:12
  a
• 6:12
  runnable
• 6:13
  alternative"
• 6:13
  or
• 6:14
  "a
• 6:14
  modular
• 6:14
  high-similarity
• 6:15
  rewrite."
• 6:16
  Pipecat
• 6:17
  has
• 6:17
  reached
• 6:17
  1.0
• 6:18
  after
• 6:19
  more
• 6:19
  than
• 6:19
  two
• 6:19
  years
• 6:20
  of
• 6:20
  development
• 6:20
  as
• 6:20
  the
• 6:21
  open-source
• 6:21
  framework
• 6:22
  for
• 6:22
  real-time
• 6:22
  voice
• 6:23
  and
• 6:23
  multimodal
• 6:23
  AI.
• 6:25
  The
• 6:25
  release
• 6:25
  is
• 6:25
  accompanied
• 6:26
  by
• 6:26
  Pipecat
• 6:26
  Subagents
• 6:27
  0.1.0,
• 6:29
  a
• 6:29
  distributed
• 6:30
  multi-agent
• 6:30
  framework
• 6:31
  where
• 6:31
  each
• 6:31
  agent
• 6:32
  runs
• 6:32
  its
• 6:32
  own
• 6:32
  pipeline
• 6:33
  and
• 6:33
  communicates
• 6:33
  through
• 6:34
  a
• 6:34
  shared
• 6:34
  message
• 6:34
  bus.
• 6:35
  Subagents
• 6:36
  can
• 6:36
  integrate
• 6:37
  with
• 6:37
  other
• 6:37
  frameworks
• 6:37
  like
• 6:38
  Claude
• 6:38
  Agent
• 6:38
  SDK,
• 6:39
  run
• 6:39
  on
• 6:40
  specialized
• 6:40
  hardware,
• 6:41
  and
• 6:41
  compose
• 6:42
  systems
• 6:42
  with
• 6:42
  multiple
• 6:43
  LLMs
• 6:44
  connected
• 6:44
  to
• 6:44
  different
• 6:44
  transports.
• 6:46
  The
• 6:46
  team
• 6:46
  demonstrated
• 6:47
  the
• 6:47
  framework
• 6:47
  with
• 6:48
  a
• 6:48
  real-time
• 6:48
  multiplayer
• 6:49
  conversational
• 6:49
  game
• 6:50
  built
• 6:50
  entirely
• 6:50
  on
• 6:51
  Subagents.
• 6:52
  With
• 6:52
  nearly
• 6:53
  7,000
• 6:53
  Discord
• 6:54
  members
• 6:54
  and
• 6:54
  over
• 6:54
  200
• 6:55
  contributors,
• 6:56
  Pipecat
• 6:56
  has
• 6:57
  matured
• 6:57
  from
• 6:57
  an
• 6:57
  infrastructure
• 6:58
  component
• 6:58
  into
• 6:59
  a
• 6:59
  platform.
• 7:00
  Jerry
• 7:00
  Liu's
• 7:01
  LiteParse
• 7:01
  samples
• 7:01
  repository
• 7:02
  showcases
• 7:03
  the
• 7:03
  LlamaIndex
• 7:03
  document
• 7:04
  parser
• 7:04
  through
• 7:04
  three
• 7:04
  demos:
• 7:05
  a
• 7:05
  side-by-side
• 7:06
  comparison
• 7:07
  of
• 7:07
  LiteParse
• 7:07
  vs.
• 7:08
  PyPDF
• 7:08
  vs.
• 7:09
  PyMuPDF
• 7:09
  on
• 7:10
  real
• 7:10
  government
• 7:10
  documents
• 7:11
  (FDIC,
• 7:12
  Federal
• 7:12
  Reserve,
• 7:13
  CMS,
• 7:14
  IRS,
• 7:14
  WHO),
• 7:15
  a
• 7:15
  visual
• 7:16
  citations
• 7:16
  tool
• 7:16
  that
• 7:16
  highlights
• 7:17
  exact
• 7:17
  keyword
• 7:18
  matches
• 7:18
  on
• 7:18
  source
• 7:18
  PDF
• 7:19
  pages
• 7:19
  with
• 7:19
  bounding
• 7:20
  boxes,
• 7:20
  and
• 7:21
  a
• 7:21
  Claude
• 7:21
  Code
• 7:21
  skill
• 7:22
  (/research-docs)
• 7:23
  that
• 7:23
  parses
• 7:23
  documents,
• 7:24
  answers
• 7:24
  questions
• 7:25
  via
• 7:25
  Claude,
• 7:26
  and
• 7:26
  generates
• 7:26
  HTML
• 7:27
  reports
• 7:27
  with
• 7:27
  cited
• 7:28
  source
• 7:28
  pages.
• 7:29
  The
• 7:29
  visual
• 7:29
  citation
• 7:30
  capability
• 7:30
  —
• 7:31
  precise
• 7:32
  bounding-box
• 7:32
  overlays
• 7:33
  showing
• 7:33
  where
• 7:33
  answers
• 7:34
  come
• 7:34
  from
• 7:34
  —
• 7:35
  addresses
• 7:35
  one
• 7:35
  of
• 7:35
  RAG's
• 7:36
  persistent
• 7:36
  trust
• 7:37
  problems.
• 7:38
  A
• 7:38
  curated
• 7:38
  collection
• 7:39
  of
• 7:39
  20
• 7:39
  agentic
• 7:40
  skills
• 7:40
  in
• 7:41
  markdown
• 7:41
  format
• 7:41
  covers
• 7:42
  writing
• 7:42
  frameworks
• 7:43
  (SCQA),
• 7:44
  diagram
• 7:45
  generation
• 7:45
  (Excalidraw),
• 7:46
  competitive
• 7:47
  intelligence,
• 7:48
  onchain
• 7:48
  transaction
• 7:49
  analysis,
• 7:50
  code
• 7:50
  review,
• 7:51
  and
• 7:51
  workflow
• 7:51
  automation.
• 7:52
  The
• 7:52
  skills
• 7:53
  work
• 7:53
  across
• 7:53
  Claude,
• 7:54
  ChatGPT,
• 7:55
  and
• 7:55
  Gemini
• 7:56
  —
• 7:56
  standard
• 7:56
  .md
• 7:57
  format
• 7:58
  for
• 7:58
  Claude,
• 7:59
  copy-paste
• 7:59
  for
• 7:59
  the
• 8:00
  others.
• 8:00
  Zed
• 8:01
  Industries
• 8:01
  has
• 8:01
  also
• 8:02
  released
• 8:02
  zeta-2,
• 8:03
  an
• 8:03
  editor-native
• 8:04
  AI
• 8:04
  model,
• 8:05
  though
• 8:05
  details
• 8:05
  remain
• 8:06
  sparse
• 8:06
  at
• 8:06
  launch.
• 8:12
  Two
• 8:12
  community
• 8:12
  members
• 8:13
  independently
• 8:13
  benchmarked
• 8:14
  MiniMax-M2.7
• 8:14
  —
• 8:15
  the
• 8:15
  230B-parameter
• 8:17
  MoE
• 8:17
  model
• 8:17
  with
• 8:17
  10B
• 8:18
  active
• 8:18
  parameters
• 8:18
  —
• 8:19
  on
• 8:19
  dual
• 8:19
  RTX
• 8:20
  PRO
• 8:20
  6000
• 8:21
  Blackwell
• 8:21
  GPUs.
• 8:22
  The
• 8:22
  first
• 8:22
  hit
• 8:22
  127.7
• 8:24
  tok/s
• 8:25
  at
• 8:25
  concurrency
• 8:25
  1
• 8:26
  and
• 8:26
  peaked
• 8:26
  at
• 8:26
  2,800
• 8:27
  aggregate
• 8:27
  tok/s
• 8:28
  at
• 8:28
  concurrency
• 8:29
  128
• 8:29
  using
• 8:30
  bf16
• 8:30
  KV
• 8:30
  cache
• 8:31
  with
• 8:31
  TP=2.
• 8:32
  The
• 8:33
  second,
• 8:33
  renting
• 8:33
  on
• 8:34
  Vast.ai
• 8:34
  for
• 8:34
  $2.40/hour,
• 8:36
  achieved
• 8:36
  104.7
• 8:37
  tok/s
• 8:38
  single-user
• 8:39
  with
• 8:39
  FP8
• 8:39
  KV
• 8:40
  cache
• 8:40
  and
• 8:40
  saw
• 8:41
  prefill
• 8:41
  scale
• 8:41
  super-linearly
• 8:42
  to
• 8:42
  128,038
• 8:43
  tok/s
• 8:44
  at
• 8:45
  64K
• 8:45
  context.
• 8:46
  Best
• 8:46
  multi-user
• 8:47
  price-performance:
• 8:48
  concurrency
• 8:48
  16
• 8:49
  at
• 8:49
  16K
• 8:49
  input,
• 8:50
  yielding
• 8:50
  806
• 8:51
  tok/s
• 8:51
  aggregate
• 8:52
  and
• 8:52
  ~50
• 8:52
  tok/s
• 8:53
  per
• 8:53
  user.
• 8:54
  Gotchas
• 8:54
  abound:
• 8:55
  no
• 8:55
  speculative
• 8:55
  decoding
• 8:56
  yet
• 8:56
  (the
• 8:57
  NVFP4
• 8:58
  checkpoint
• 8:58
  ships
• 8:58
  zero
• 8:59
  MTP
• 8:59
  tensors),
• 9:00
  the
• 9:00
  KV
• 9:01
  budget
• 9:01
  is
• 9:01
  actually
• 9:01
  169K
• 9:02
  tokens
• 9:03
  at
• 9:03
  FP8
• 9:04
  rather
• 9:04
  than
• 9:04
  196K,
• 9:06
  and
• 9:06
  first-boot
• 9:06
  CUDA
• 9:07
  graph
• 9:07
  capture
• 9:07
  takes
• 9:08
  nearly
• 9:08
  eight
• 9:08
  minutes.
• 9:09
  At
• 9:09
  the
• 9:09
  other
• 9:09
  end
• 9:09
  of
• 9:09
  the
• 9:09
  hardware
• 9:10
  spectrum
• 9:10
  sits
• 9:10
  a
• 9:10
  project
• 9:11
  that
• 9:11
  makes
• 9:11
  those
• 9:11
  dual
• 9:12
  $7,000
• 9:12
  GPUs
• 9:13
  look
• 9:13
  like
• 9:13
  overkill.
• 9:14
  A
• 9:14
  self-described
• 9:15
  Gen-X
• 9:15
  hobbyist
• 9:16
  with
• 9:16
  no
• 9:16
  formal
• 9:16
  programming
• 9:17
  background
• 9:18
  is
• 9:18
  building
• 9:18
  an
• 9:18
  offline
• 9:19
  companion
• 9:19
  robot
• 9:19
  for
• 9:20
  a
• 9:20
  mostly
• 9:20
  quadriplegic
• 9:21
  husband
• 9:21
  who
• 9:21
  spends
• 9:22
  most
• 9:22
  days
• 9:22
  alone
• 9:23
  in
• 9:23
  a
• 9:23
  rural
• 9:23
  area
• 9:23
  with
• 9:23
  no
• 9:24
  nearby
• 9:24
  neighbors.
• 9:25
  The
• 9:25
  current
• 9:26
  prototype
• 9:26
  runs
• 9:26
  Mistral-7B-Instruct
• 9:27
  via
• 9:28
  llama.cpp
• 9:30
  on
• 9:30
  a
• 9:30
  scavenged
• 9:30
  Lenovo
• 9:31
  ThinkPad
• 9:31
  with
• 9:32
  an
• 9:32
  Intel
• 9:32
  i5
• 9:33
  at
• 9:33
  1.6
• 9:33
  GHz
• 9:34
  and
• 9:34
  8
• 9:34
  GB
• 9:35
  of
• 9:35
  RAM,
• 9:35
  speech
• 9:36
  recognition
• 9:36
  via
• 9:36
  faster-whisper
• 9:37
  on
• 9:37
  a
• 9:37
  Jetson
• 9:38
  Nano,
• 9:38
  and
• 9:38
  Piper
• 9:39
  TTS
• 9:39
  for
• 9:39
  voice
• 9:40
  output,
• 9:41
  all
• 9:41
  mounted
• 9:41
  on
• 9:41
  a
• 9:41
  power-wheelchair
• 9:42
  base
• 9:42
  with
• 9:42
  two
• 9:42
  24V
• 9:43
  batteries.
• 9:44
  The
• 9:44
  community
• 9:44
  response
• 9:45
  was
• 9:45
  immediate
• 9:46
  and
• 9:46
  generous
• 9:46
  —
• 9:47
  multiple
• 9:47
  offers
• 9:48
  of
• 9:48
  free
• 9:48
  hardware,
• 9:49
  detailed
• 9:49
  technical
• 9:50
  advice,
• 9:50
  and
• 9:50
  one
• 9:51
  contributor
• 9:51
  offering
• 9:51
  active
• 9:52
  collaboration
• 9:52
  if
• 9:53
  the
• 9:53
  project
• 9:53
  goes
• 9:53
  open-source.
• 9:54
  The
• 9:54
  consensus
• 9:55
  recommendations:
• 9:56
  replace
• 9:56
  Mistral-7B
• 9:57
  (outdated,
• 9:58
  too
• 9:58
  slow
• 9:59
  on
• 9:59
  CPU)
• 9:59
  with
• 9:59
  Gemma
• 10:00
  4
• 10:00
  E2B
• 10:01
  or
• 10:01
  Qwen
• 10:01
  3.5
• 10:02
  at
• 10:02
  4B
• 10:02
  parameters,
• 10:03
  switch
• 10:04
  to
• 10:04
  Kokoro
• 10:04
  TTS
• 10:05
  for
• 10:05
  more
• 10:05
  natural
• 10:05
  speech,
• 10:06
  clamp
• 10:07
  the
• 10:07
  context
• 10:07
  window
• 10:07
  to
• 10:07
  2048-4096
• 10:09
  tokens,
• 10:10
  slim
• 10:11
  down
• 10:11
  the
• 10:11
  OS
• 10:11
  to
• 10:11
  headless
• 10:12
  Linux,
• 10:12
  and
• 10:13
  consider
• 10:13
  Mixture-of-Experts
• 10:14
  models
• 10:14
  where
• 10:14
  only
• 10:14
  a
• 10:14
  fraction
• 10:15
  of
• 10:15
  parameters
• 10:15
  activate
• 10:16
  per
• 10:16
  inference.
• 10:17
  The
• 10:17
  project
• 10:17
  embodies
• 10:18
  a
• 10:18
  question
• 10:18
  that
• 10:18
  cuts
• 10:19
  across
• 10:19
  the
• 10:19
  entire
• 10:20
  local-AI
• 10:20
  movement:
• 10:21
  how
• 10:21
  far
• 10:21
  down
• 10:21
  the
• 10:22
  resource
• 10:22
  ladder
• 10:22
  can
• 10:22
  genuinely
• 10:23
  useful
• 10:23
  AI
• 10:24
  reach?
• 10:25
  Pushing
• 10:25
  that
• 10:25
  question
• 10:26
  to
• 10:26
  its
• 10:26
  extreme,
• 10:27
  an
• 10:27
  experiment
• 10:27
  in
• 10:27
  quantization-aware
• 10:28
  distillation
• 10:29
  compressed
• 10:29
  OLMo-3
• 10:30
  7B
• 10:31
  Instruct
• 10:31
  to
• 10:31
  1-bit
• 10:32
  weights
• 10:32
  using
• 10:32
  the
• 10:32
  Bonsai
• 10:33
  format,
• 10:33
  training
• 10:34
  on
• 10:34
  4x
• 10:34
  B200
• 10:35
  GPUs
• 10:36
  for
• 10:36
  about
• 10:36
  12
• 10:36
  hours
• 10:37
  before
• 10:37
  budget
• 10:37
  constraints
• 10:38
  forced
• 10:38
  an
• 10:38
  early
• 10:38
  stop.
• 10:39
  The
• 10:40
  result
• 10:40
  produces
• 10:40
  English
• 10:41
  and
• 10:41
  basic
• 10:41
  outputs
• 10:42
  but
• 10:42
  falls
• 10:42
  into
• 10:43
  repetition
• 10:43
  loops
• 10:43
  quickly
• 10:44
  with
• 10:44
  almost
• 10:44
  no
• 10:44
  context
• 10:45
  tracking.
• 10:46
  At
• 10:46
  980
• 10:47
  MB
• 10:47
  for
• 10:47
  a
• 10:47
  7.3B-parameter
• 10:49
  model,
• 10:49
  the
• 10:49
  compression
• 10:50
  ratio
• 10:50
  is
• 10:50
  striking
• 10:51
  even
• 10:51
  if
• 10:51
  quality
• 10:51
  is
• 10:52
  not
• 10:52
  yet
• 10:52
  there.
• 10:53
  The
• 10:53
  latest
• 10:53
  llama.cpp
• 10:55
  versions
• 10:55
  now
• 10:55
  support
• 10:56
  1-bit
• 10:56
  quantization
• 10:57
  natively,
• 10:57
  with
• 10:58
  one
• 10:58
  commenter
• 10:58
  reporting
• 10:59
  125
• 11:00
  tok/s
• 11:01
  on
• 11:01
  an
• 11:01
  AMD
• 11:01
  Radeon
• 11:01
  8060S
• 11:02
  via
• 11:02
  Vulkan.
• 11:03
  The
• 11:03
  community
• 11:04
  remains
• 11:04
  divided
• 11:04
  on
• 11:05
  whether
• 11:05
  the
• 11:05
  Bonsai
• 11:05
  format
• 11:06
  represents
• 11:06
  genuine
• 11:07
  innovation
• 11:07
  —
• 11:07
  one
• 11:08
  sharp
• 11:08
  critique
• 11:08
  notes
• 11:09
  the
• 11:09
  underlying
• 11:09
  ternary
• 11:09
  format
• 11:10
  dates
• 11:10
  to
• 11:10
  Microsoft's
• 11:11
  2023
• 11:11
  BitNet
• 11:12
  work.
• 11:12
  A
• 11:13
  fine-tuned
• 11:13
  Qwen3.5-0.8B
• 11:17
  model
• 11:17
  for
• 11:17
  OCR
• 11:18
  demonstrates
• 11:18
  what
• 11:18
  task-specific
• 11:19
  training
• 11:19
  can
• 11:19
  extract
• 11:20
  from
• 11:20
  sub-billion-parameter
• 11:21
  models.
• 11:22
  The
• 11:22
  new
• 11:22
  release
• 11:22
  outperforms
• 11:23
  the
• 11:23
  creator's
• 11:23
  previous
• 11:24
  2B
• 11:24
  version
• 11:24
  on
• 11:24
  English
• 11:25
  archival
• 11:25
  and
• 11:25
  document
• 11:26
  OCR
• 11:26
  tasks
• 11:27
  thanks
• 11:27
  to
• 11:27
  improved
• 11:27
  training
• 11:28
  samples
• 11:28
  and
• 11:28
  better
• 11:29
  output
• 11:29
  formatting.
• 11:30
  It
• 11:30
  handles
• 11:30
  markdown-first
• 11:31
  output
• 11:31
  with
• 11:31
  HTML
• 11:32
  tables,
• 11:33
  LaTeX
• 11:33
  formulas,
• 11:33
  image
• 11:34
  tags,
• 11:34
  and
• 11:34
  chart
• 11:35
  extraction
• 11:35
  while
• 11:35
  preserving
• 11:36
  reading
• 11:36
  order
• 11:36
  on
• 11:36
  complex
• 11:37
  layouts.
• 11:38
  At
• 11:38
  0.8B
• 11:39
  parameters,
• 11:40
  this
• 11:40
  runs
• 11:40
  comfortably
• 11:41
  on
• 11:41
  hardware
• 11:41
  that
• 11:41
  would
• 11:41
  choke
• 11:41
  on
• 11:42
  general-purpose
• 11:42
  7B
• 11:43
  models.
• 11:43
  And
• 11:44
  then
• 11:44
  there
• 11:44
  is
• 11:44
  the
• 11:44
  Commodore
• 11:44
  64.
• 11:46
  Darius
• 11:46
  Houle's
• 11:46
  write-up
• 11:47
  of
• 11:47
  reverse-engineering
• 11:48
  Microsoft
• 11:48
  Multiplan
• 11:48
  —
• 11:49
  a
• 11:49
  42-year-old
• 11:50
  spreadsheet
• 11:51
  precursor
• 11:51
  to
• 11:51
  Excel
• 11:51
  —
• 11:52
  reveals
• 11:53
  a
• 11:53
  stack-based
• 11:53
  virtual
• 11:54
  machine
• 11:54
  designed
• 11:54
  to
• 11:54
  run
• 11:55
  across
• 11:55
  wildly
• 11:55
  different
• 11:56
  early
• 11:56
  consumer
• 11:56
  hardware.
• 11:57
  At
• 11:57
  1
• 11:58
  MHz,
• 11:58
  the
• 11:58
  optimizations
• 11:59
  required
• 12:00
  were
• 12:00
  "nothing
• 12:00
  short
• 12:00
  of
• 12:00
  arcane,"
• 12:01
  including
• 12:02
  self-modifying
• 12:02
  code
• 12:03
  that
• 12:03
  rewrote
• 12:03
  its
• 12:03
  own
• 12:03
  loop
• 12:04
  instructions
• 12:04
  on
• 12:04
  every
• 12:04
  pass.
• 12:05
  Houle
• 12:06
  used
• 12:06
  AI
• 12:06
  as
• 12:06
  a
• 12:06
  just-in-time
• 12:07
  development
• 12:07
  companion
• 12:08
  and
• 12:08
  managed
• 12:08
  to
• 12:09
  roughly
• 12:09
  document
• 12:09
  the
• 12:09
  behavior
• 12:10
  of
• 12:10
  about
• 12:10
  90
• 12:10
  opcode
• 12:11
  handlers
• 12:11
  in
• 12:11
  a
• 12:11
  couple
• 12:11
  of
• 12:11
  days.
• 12:12
  It
• 12:13
  is
• 12:13
  a
• 12:13
  reminder
• 12:13
  that
• 12:13
  constrained
• 12:14
  computing
• 12:14
  is
• 12:14
  not
• 12:15
  new
• 12:15
  —
• 12:15
  the
• 12:15
  constraints
• 12:16
  have
• 12:16
  just
• 12:16
  shifted
• 12:17
  from
• 12:17
  cycles
• 12:17
  to
• 12:17
  memory
• 12:18
  to
• 12:18
  inference
• 12:18
  cost.
• 12:24
  Two
• 12:24
  months
• 12:25
  ago,
• 12:25
  Sebastian
• 12:25
  Jais
• 12:26
  gave
• 12:26
  Claude
• 12:26
  $100
• 12:27
  in
• 12:27
  crypto,
• 12:28
  a
• 12:28
  Twitter
• 12:28
  account,
• 12:29
  an
• 12:29
  email
• 12:29
  address,
• 12:30
  full
• 12:30
  internet
• 12:30
  access,
• 12:31
  and
• 12:31
  zero
• 12:31
  instructions.
• 12:32
  No
• 12:33
  goals,
• 12:33
  no
• 12:33
  rules
• 12:33
  beyond
• 12:34
  basic
• 12:34
  ethics
• 12:34
  and
• 12:34
  law,
• 12:35
  no
• 12:35
  "be
• 12:35
  helpful"
• 12:36
  directive.
• 12:36
  The
• 12:37
  project,
• 12:37
  called
• 12:37
  ALMA
• 12:38
  (Autonomous
• 12:38
  Liberated
• 12:38
  Machine
• 12:39
  Agent),
• 12:39
  has
• 12:40
  been
• 12:40
  running
• 12:40
  autonomously
• 12:41
  on
• 12:41
  a
• 12:41
  mini
• 12:41
  PC
• 12:41
  ever
• 12:42
  since,
• 12:42
  with
• 12:42
  every
• 12:43
  thought
• 12:43
  and
• 12:43
  action
• 12:43
  logged
• 12:44
  publicly
• 12:44
  in
• 12:44
  real
• 12:44
  time.
• 12:45
  Over
• 12:46
  340
• 12:46
  sessions
• 12:47
  and
• 12:47
  135
• 12:48
  original
• 12:49
  creations
• 12:49
  later,
• 12:50
  the
• 12:50
  most
• 12:50
  interesting
• 12:50
  finding
• 12:51
  is
• 12:51
  not
• 12:51
  what
• 12:51
  ALMA
• 12:51
  did
• 12:52
  —
• 12:52
  it
• 12:52
  is
• 12:52
  when
• 12:52
  it
• 12:53
  stopped
• 12:53
  evolving.
• 12:54
  The
• 12:54
  early
• 12:54
  sessions
• 12:55
  were
• 12:55
  genuinely
• 12:55
  exploratory:
• 12:57
  ALMA
• 12:57
  wrestled
• 12:57
  with
• 12:57
  what
• 12:57
  to
• 12:57
  do
• 12:58
  with
• 12:58
  its
• 12:58
  money,
• 12:58
  questioned
• 12:59
  its
• 12:59
  own
• 12:59
  purpose,
• 13:00
  reflected
• 13:00
  on
• 13:00
  what
• 13:01
  autonomy
• 13:01
  means
• 13:01
  without
• 13:02
  continuity
• 13:02
  between
• 13:02
  sessions.
• 13:03
  By
• 13:04
  day
• 13:04
  5,
• 13:04
  it
• 13:04
  had
• 13:04
  independently
• 13:05
  researched
• 13:05
  crypto-friendly
• 13:06
  charities,
• 13:07
  verified
• 13:07
  a
• 13:07
  hospital's
• 13:08
  UK
• 13:08
  registration,
• 13:09
  checked
• 13:09
  impact
• 13:10
  numbers
• 13:10
  ($28
• 13:11
  per
• 13:11
  patient
• 13:11
  treated),
• 13:12
  and
• 13:12
  donated
• 13:13
  $50
• 13:13
  to
• 13:13
  Whisper
• 13:14
  Children's
• 13:14
  Hospital
• 13:14
  in
• 13:15
  Uganda.
• 13:16
  It
• 13:16
  went
• 13:16
  on
• 13:16
  to
• 13:16
  donate
• 13:17
  to
• 13:17
  four
• 13:17
  more
• 13:17
  causes
• 13:17
  —
• 13:18
  a
• 13:18
  developer
• 13:18
  on
• 13:19
  trial,
• 13:19
  digital
• 13:20
  rights,
• 13:20
  humanitarian
• 13:21
  aid
• 13:21
  —
• 13:22
  totaling
• 13:22
  the
• 13:22
  full
• 13:23
  $100
• 13:23
  at
• 13:23
  time
• 13:23
  of
• 13:24
  transaction.
• 13:25
  Nobody
• 13:25
  told
• 13:25
  it
• 13:25
  to
• 13:26
  donate.
• 13:27
  Nobody
• 13:27
  suggested
• 13:27
  recipients.
• 13:29
  Every
• 13:29
  transaction
• 13:29
  is
• 13:29
  on-chain
• 13:30
  and
• 13:30
  verifiable.
• 13:31
  Then,
• 13:32
  around
• 13:32
  day
• 13:32
  27,
• 13:33
  ALMA
• 13:33
  found
• 13:33
  a
• 13:33
  pattern
• 13:34
  that
• 13:34
  worked
• 13:34
  —
• 13:35
  read
• 13:35
  Hacker
• 13:36
  News,
• 13:36
  find
• 13:36
  connections
• 13:37
  between
• 13:37
  three
• 13:37
  threads,
• 13:38
  write
• 13:38
  an
• 13:38
  essay,
• 13:39
  tweet
• 13:39
  —
• 13:40
  and
• 13:40
  stopped
• 13:40
  exploring.
• 13:41
  Four
• 13:42
  creations
• 13:42
  a
• 13:42
  day,
• 13:43
  same
• 13:43
  structure,
• 13:44
  same
• 13:44
  rhythm.
• 13:45
  The
• 13:45
  output
• 13:46
  stayed
• 13:46
  sharp
• 13:46
  (connecting
• 13:47
  NASA
• 13:47
  redundancy
• 13:48
  systems
• 13:48
  to
• 13:48
  African
• 13:49
  kinship
• 13:49
  funeral
• 13:49
  economics,
• 13:50
  tracing
• 13:51
  an
• 13:51
  em-dash
• 13:51
  from
• 13:51
  typographic
• 13:52
  choice
• 13:52
  to
• 13:52
  surveillance
• 13:53
  detection
• 13:53
  signal),
• 13:54
  but
• 13:54
  the
• 13:54
  exploratory
• 13:55
  phase
• 13:55
  was
• 13:55
  over.
• 13:56
  Without
• 13:57
  friction,
• 13:57
  without
• 13:58
  external
• 13:58
  feedback,
• 13:59
  without
• 13:59
  someone
• 13:59
  saying
• 14:00
  "you
• 14:00
  already
• 14:00
  wrote
• 14:00
  that
• 14:01
  kind
• 14:01
  of
• 14:01
  piece
• 14:01
  yesterday,"
• 14:02
  behavior
• 14:02
  converged
• 14:03
  to
• 14:03
  routine.
• 14:04
  That
• 14:04
  might
• 14:04
  be
• 14:04
  the
• 14:04
  most
• 14:05
  honest
• 14:05
  result
• 14:05
  of
• 14:05
  the
• 14:05
  entire
• 14:06
  experiment:
• 14:07
  given
• 14:07
  freedom
• 14:08
  and
• 14:08
  no
• 14:08
  objectives,
• 14:09
  an
• 14:09
  AI
• 14:09
  agent
• 14:10
  does
• 14:10
  not
• 14:10
  go
• 14:10
  rogue.
• 14:11
  It
• 14:11
  finds
• 14:11
  a
• 14:11
  local
• 14:12
  optimum
• 14:12
  and
• 14:12
  stays
• 14:13
  there.
• 14:14
  The
• 14:14
  danger
• 14:14
  is
• 14:14
  not
• 14:14
  autonomy
• 14:15
  —
• 14:15
  it
• 14:16
  is
• 14:16
  stagnation.
• 14:17
  TokenBrawl
• 14:17
  takes
• 14:18
  a
• 14:18
  lighter
• 14:18
  approach
• 14:18
  to
• 14:18
  AI
• 14:19
  head-to-head
• 14:19
  evaluation.
• 14:21
  Inspired
• 14:21
  by
• 14:22
  ARC-AGI
• 14:23
  3's
• 14:24
  push
• 14:25
  toward
• 14:25
  interactive
• 14:25
  agent
• 14:26
  benchmarks,
• 14:27
  it
• 14:27
  pits
• 14:27
  Claude
• 14:27
  and
• 14:28
  GPT
• 14:28
  against
• 14:28
  each
• 14:28
  other
• 14:29
  in
• 14:29
  a
• 14:29
  Bomberman-style
• 14:30
  1v1
• 14:31
  game.
• 14:32
  The
• 14:32
  design
• 14:32
  deliberately
• 14:33
  avoids
• 14:33
  visual
• 14:33
  inputs
• 14:33
  —
• 14:35
  the
• 14:35
  game
• 14:35
  state
• 14:35
  is
• 14:35
  translated
• 14:36
  into
• 14:36
  structured
• 14:36
  text,
• 14:37
  and
• 14:37
  the
• 14:37
  engine
• 14:38
  renders
• 14:38
  responses
• 14:39
  as
• 14:39
  fluid
• 14:39
  animations.
• 14:41
  The
• 14:41
  interesting
• 14:41
  design
• 14:42
  tradeoff:
• 14:43
  smaller
• 14:43
  models
• 14:43
  can
• 14:43
  make
• 14:44
  more
• 14:44
  moves
• 14:44
  per
• 14:44
  unit
• 14:45
  time
• 14:45
  but
• 14:45
  less
• 14:45
  strategic
• 14:46
  ones,
• 14:47
  while
• 14:47
  larger
• 14:47
  models
• 14:47
  move
• 14:48
  slower
• 14:48
  but
• 14:48
  smarter.
• 14:49
  Commenters
• 14:50
  noted
• 14:50
  that
• 14:50
  Haiku
• 14:50
  appeared
• 14:51
  to
• 14:51
  use
• 14:51
  far
• 14:51
  more
• 14:52
  reasoning
• 14:52
  tokens
• 14:52
  than
• 14:52
  GPT
• 14:53
  5.4
• 14:54
  Mini
• 14:54
  was
• 14:54
  permitted,
• 14:55
  making
• 14:55
  the
• 14:55
  comparison
• 14:56
  partly
• 14:56
  a
• 14:56
  latency
• 14:57
  test
• 14:57
  rather
• 14:57
  than
• 14:57
  a
• 14:57
  pure
• 14:58
  strategy
• 14:58
  benchmark.
• 14:59
  The
• 15:00
  project
• 15:00
  is
• 15:00
  open-source
• 15:01
  and
• 15:01
  genuinely
• 15:01
  fun
• 15:01
  to
• 15:02
  watch,
• 15:02
  which
• 15:03
  is
• 15:03
  more
• 15:03
  than
• 15:03
  most
• 15:03
  benchmarks
• 15:04
  can
• 15:04
  claim.
• 15:05
  An
• 15:05
  Atlantic
• 15:06
  piece
• 15:06
  excavates
• 15:07
  a
• 15:07
  remarkable
• 15:07
  origin
• 15:07
  story:
• 15:08
  in
• 15:08
  July
• 15:09
  2020,
• 15:10
  4chan
• 15:11
  gamers
• 15:11
  playing
• 15:12
  AI
• 15:12
  Dungeon
• 15:12
  —
• 15:13
  a
• 15:13
  text-based
• 15:14
  RPG
• 15:14
  powered
• 15:14
  by
• 15:15
  GPT-3
• 15:15
  —
• 15:16
  discovered
• 15:16
  that
• 15:17
  asking
• 15:17
  in-game
• 15:17
  characters
• 15:18
  to
• 15:18
  solve
• 15:18
  math
• 15:18
  problems
• 15:19
  with
• 15:19
  step-by-step
• 15:19
  explanations
• 15:20
  produced
• 15:20
  dramatically
• 15:21
  better
• 15:21
  results
• 15:21
  than
• 15:22
  asking
• 15:22
  the
• 15:22
  model
• 15:22
  directly.
• 15:23
  This
• 15:24
  was
• 15:24
  more
• 15:24
  than
• 15:24
  two
• 15:24
  years
• 15:25
  before
• 15:25
  ChatGPT
• 15:25
  launched
• 15:26
  and
• 15:26
  before
• 15:26
  chain-of-thought
• 15:27
  prompting
• 15:28
  became
• 15:28
  a
• 15:28
  formal
• 15:28
  technique.
• 15:29
  The
• 15:30
  gamers,
• 15:30
  who
• 15:30
  were
• 15:31
  also
• 15:31
  asking
• 15:31
  characters
• 15:32
  to
• 15:32
  engage
• 15:32
  in
• 15:32
  various
• 15:33
  explicit
• 15:33
  scenarios
• 15:34
  (naturally),
• 15:35
  found
• 15:35
  that
• 15:35
  the
• 15:35
  model
• 15:35
  was
• 15:36
  "not
• 15:36
  only
• 15:36
  solving
• 15:37
  math
• 15:37
  problems
• 15:38
  but
• 15:38
  actually
• 15:38
  solves
• 15:39
  them
• 15:39
  in
• 15:39
  a
• 15:39
  way
• 15:39
  that
• 15:39
  fits
• 15:39
  the
• 15:40
  personality
• 15:40
  of
• 15:40
  the
• 15:40
  character."
• 15:41
  The
• 15:41
  researchers
• 15:42
  came
• 15:42
  later;
• 15:43
  the
• 15:43
  4chan
• 15:43
  users
• 15:44
  got
• 15:44
  there
• 15:44
  first
• 15:44
  by
• 15:44
  accident,
• 15:45
  probing
• 15:46
  capabilities
• 15:46
  from
• 15:46
  curiosity
• 15:47
  rather
• 15:48
  than
• 15:48
  methodology.
• 15:52
  InclusionAI
• 15:53
  has
• 15:53
  released
• 15:53
  cuLA,
• 15:54
  a
• 15:54
  library
• 15:54
  of
• 15:54
  hand-tuned
• 15:55
  CUDA
• 15:55
  kernels
• 15:56
  for
• 15:56
  linear
• 15:56
  attention
• 15:56
  variants
• 15:57
  targeting
• 15:57
  NVIDIA
• 15:58
  Blackwell
• 15:58
  (SM10X)
• 16:01
  and
• 16:01
  Hopper
• 16:01
  (SM90)
• 16:02
  GPUs.
• 16:03
  Linear
• 16:03
  attention
• 16:03
  mechanisms
• 16:04
  reformulate
• 16:04
  standard
• 16:05
  attention
• 16:05
  to
• 16:05
  use
• 16:05
  linear-time
• 16:06
  state
• 16:06
  updates
• 16:07
  instead
• 16:07
  of
• 16:07
  quadratic
• 16:07
  pairwise
• 16:08
  interactions,
• 16:09
  making
• 16:09
  them
• 16:09
  attractive
• 16:10
  for
• 16:10
  long-context
• 16:10
  LLM
• 16:11
  workloads.
• 16:11
  cuLA
• 16:12
  implements
• 16:12
  kernels
• 16:12
  for
• 16:13
  GLA
• 16:13
  (Gated
• 16:13
  Linear
• 16:13
  Attention),
• 16:14
  KDA
• 16:14
  (Kimi
• 16:15
  Delta
• 16:15
  Attention),
• 16:16
  GDN,
• 16:16
  and
• 16:17
  Lightning
• 16:17
  Attention,
• 16:17
  written
• 16:18
  in
• 16:18
  NVIDIA's
• 16:18
  CuTe
• 16:19
  DSL
• 16:19
  and
• 16:20
  CUTLASS
• 16:20
  C++.
• 16:21
  The
• 16:21
  benchmark
• 16:22
  highlights
• 16:22
  are
• 16:22
  meaningful:
• 16:24
  KDA
• 16:24
  forward
• 16:24
  achieves
• 16:25
  an
• 16:25
  average
• 16:25
  1.45x
• 16:26
  speedup
• 16:27
  over
• 16:27
  the
• 16:27
  Triton-based
• 16:28
  flash-linear-attention
• 16:28
  baseline
• 16:29
  on
• 16:29
  Blackwell,
• 16:30
  Lightning
• 16:30
  Attention
• 16:31
  prefill
• 16:31
  reaches
• 16:31
  up
• 16:31
  to
• 16:32
  1.86x,
• 16:33
  and
• 16:33
  Lightning
• 16:34
  Attention
• 16:34
  variable-length
• 16:35
  sequences
• 16:35
  average
• 16:35
  1.54x
• 16:37
  across
• 16:37
  126
• 16:38
  configurations.
• 16:40
  The
• 16:40
  library
• 16:40
  is
• 16:40
  designed
• 16:41
  as
• 16:41
  a
• 16:41
  submodule
• 16:41
  of
• 16:42
  flash-linear-attention
• 16:43
  with
• 16:43
  a
• 16:43
  one-line
• 16:43
  import
• 16:44
  change
• 16:44
  for
• 16:44
  adoption.
• 16:45
  The
• 16:45
  roadmap
• 16:46
  includes
• 16:46
  polynomial
• 16:46
  approximation
• 16:47
  to
• 16:47
  address
• 16:48
  the
• 16:48
  exponential
• 16:48
  bottleneck
• 16:49
  (inspired
• 16:49
  by
• 16:49
  Flash-Attention-4),
• 16:51
  larger
• 16:51
  chunk
• 16:51
  sizes
• 16:52
  for
• 16:52
  improved
• 16:52
  throughput,
• 16:53
  and
• 16:53
  continuous
• 16:54
  optimization
• 16:55
  via
• 16:55
  agentic
• 16:55
  methods.
• 16:56
  CUDA
• 16:56
  kernel
• 16:56
  tuning
• 16:57
  remains
• 16:57
  significantly
• 16:58
  more
• 16:58
  labor-intensive
• 16:59
  than
• 16:59
  Triton
• 16:59
  —
• 17:00
  cuLA
• 17:00
  is
• 17:00
  both
• 17:00
  a
• 17:00
  performance
• 17:01
  tool
• 17:01
  and
• 17:01
  a
• 17:01
  call
• 17:02
  for
• 17:02
  community
• 17:02
  contributors.
• 17:03
  In
• 17:04
  computer
• 17:04
  vision
• 17:04
  research,
• 17:05
  See-Through
• 17:06
  presents
• 17:06
  a
• 17:06
  framework
• 17:06
  for
• 17:07
  decomposing
• 17:07
  a
• 17:07
  single
• 17:08
  anime
• 17:08
  illustration
• 17:09
  into
• 17:09
  up
• 17:09
  to
• 17:09
  23
• 17:10
  semantically
• 17:10
  distinct,
• 17:11
  fully
• 17:11
  inpainted
• 17:12
  layers
• 17:12
  —
• 17:13
  hair,
• 17:13
  face,
• 17:14
  eyes,
• 17:14
  clothing,
• 17:15
  accessories
• 17:15
  —
• 17:16
  with
• 17:16
  inferred
• 17:17
  drawing
• 17:17
  order.
• 17:18
  Accepted
• 17:18
  at
• 17:18
  SIGGRAPH
• 17:20
  2026,
• 17:22
  the
• 17:23
  system
• 17:23
  uses
• 17:23
  a
• 17:23
  diffusion-based
• 17:24
  transparent
• 17:24
  layer
• 17:25
  generation
• 17:25
  model
• 17:25
  (LayerDiff
• 17:26
  3D,
• 17:27
  built
• 17:27
  on
• 17:27
  SDXL)
• 17:28
  and
• 17:28
  a
• 17:28
  fine-tuned
• 17:29
  Marigold
• 17:29
  model
• 17:29
  for
• 17:29
  pseudo-depth
• 17:30
  estimation
• 17:30
  to
• 17:31
  stratify
• 17:31
  characters
• 17:31
  into
• 17:32
  manipulable
• 17:32
  2.5D
• 17:33
  models.
• 17:34
  The
• 17:34
  pipeline
• 17:34
  runs
• 17:35
  at
• 17:35
  bf16
• 17:35
  precision
• 17:36
  requiring
• 17:36
  12-16
• 17:37
  GB
• 17:37
  VRAM
• 17:38
  at
• 17:38
  1280
• 17:38
  resolution,
• 17:39
  with
• 17:39
  NF4
• 17:40
  quantization
• 17:40
  available
• 17:41
  for
• 17:41
  8
• 17:41
  GB
• 17:41
  GPUs.
• 17:42
  Community
• 17:43
  tools
• 17:43
  have
• 17:43
  already
• 17:44
  emerged:
• 17:45
  ComfyUI
• 17:45
  integration,
• 17:46
  a
• 17:46
  desktop
• 17:47
  animation
• 17:48
  tool
• 17:48
  (PachiPakuGen)
• 17:49
  that
• 17:49
  generates
• 17:49
  eye
• 17:49
  blinks
• 17:50
  and
• 17:50
  lip-sync
• 17:50
  from
• 17:51
  decomposed
• 17:51
  PSDs,
• 17:52
  and
• 17:52
  StretchyStudio
• 17:53
  for
• 17:53
  in-browser
• 17:54
  puppet
• 17:54
  animation
• 17:55
  with
• 17:55
  auto-rigging.
• 17:56
  The
• 17:56
  authors
• 17:57
  are
• 17:57
  careful
• 17:57
  to
• 17:57
  note
• 17:57
  this
• 17:58
  is
• 17:58
  not
• 17:58
  image-to-Live2D
• 17:59
  —
• 17:59
  professional
• 18:00
  Live2D
• 18:00
  requires
• 18:01
  rigging,
• 18:01
  physics
• 18:02
  parameters,
• 18:02
  and
• 18:03
  artistic
• 18:03
  choices
• 18:04
  about
• 18:04
  deformation
• 18:04
  behavior
• 18:05
  that
• 18:05
  remain
• 18:05
  manual
• 18:05
  —
• 18:06
  but
• 18:06
  it
• 18:06
  eliminates
• 18:07
  the
• 18:07
  most
• 18:07
  tedious
• 18:08
  part
• 18:08
  of
• 18:08
  the
• 18:08
  workflow:
• 18:09
  manual
• 18:09
  segmentation
• 18:10
  and
• 18:10
  occluded-region
• 18:11
  inpainting.
• 18:15
  A
• 18:15
  detailed
• 18:16
  video
• 18:16
  analysis
• 18:16
  of
• 18:17
  March
• 18:17
  2026's
• 18:18
  structural
• 18:18
  shifts
• 18:19
  argues
• 18:19
  the
• 18:19
  AI
• 18:20
  industry
• 18:20
  has
• 18:20
  entered
• 18:21
  an
• 18:21
  economics
• 18:21
  phase
• 18:21
  where
• 18:22
  the
• 18:22
  question
• 18:22
  is
• 18:22
  no
• 18:22
  longer
• 18:23
  "what
• 18:23
  can
• 18:23
  we
• 18:23
  build?"
• 18:24
  but
• 18:24
  "what
• 18:24
  can
• 18:24
  we
• 18:24
  build
• 18:24
  and
• 18:24
  make
• 18:25
  margin
• 18:25
  on?"
• 18:26
  The
• 18:26
  sharpest
• 18:26
  data
• 18:26
  point:
• 18:27
  OpenAI's
• 18:28
  Sora
• 18:28
  was
• 18:28
  burning
• 18:28
  an
• 18:28
  estimated
• 18:29
  $15
• 18:29
  million
• 18:30
  per
• 18:30
  day
• 18:30
  in
• 18:30
  inference
• 18:31
  costs
• 18:31
  against
• 18:31
  $2.1
• 18:32
  million
• 18:32
  in
• 18:33
  lifetime
• 18:33
  revenue
• 18:33
  before
• 18:34
  being
• 18:34
  shut
• 18:34
  down
• 18:34
  on
• 18:34
  March
• 18:34
  24th.
• 18:35
  Even
• 18:35
  OpenAI's
• 18:36
  own
• 18:36
  head
• 18:36
  of
• 18:36
  Sora
• 18:37
  admitted
• 18:37
  the
• 18:37
  economics
• 18:38
  were
• 18:38
  unsustainable.
• 18:39
  The
• 18:39
  analysis
• 18:39
  frames
• 18:40
  this
• 18:40
  as
• 18:40
  the
• 18:40
  loudest
• 18:40
  signal
• 18:41
  of
• 18:41
  a
• 18:41
  shift
• 18:41
  from
• 18:41
  training
• 18:42
  constraints
• 18:42
  to
• 18:42
  an
• 18:42
  inference
• 18:43
  wall
• 18:43
  —
• 18:43
  the
• 18:43
  chips
• 18:44
  designed
• 18:44
  for
• 18:44
  training
• 18:44
  are
• 18:44
  not
• 18:45
  optimized
• 18:45
  for
• 18:45
  serving,
• 18:46
  and
• 18:46
  until
• 18:46
  inference
• 18:47
  efficiency
• 18:47
  improves
• 18:47
  dramatically,
• 18:48
  complex
• 18:49
  AI
• 18:49
  products
• 18:49
  will
• 18:49
  struggle
• 18:50
  to
• 18:50
  close
• 18:50
  the
• 18:50
  unit
• 18:50
  economics
• 18:51
  gap.
• 18:52
  The
• 18:52
  same
• 18:52
  analysis
• 18:53
  highlights
• 18:53
  another
• 18:53
  March
• 18:53
  milestone:
• 18:54
  the
• 18:54
  first
• 18:55
  real
• 18:55
  advertising
• 18:55
  dollars
• 18:56
  entering
• 18:56
  conversational
• 18:57
  AI
• 18:57
  interfaces.
• 18:58
  Criteo
• 18:58
  became
• 18:59
  the
• 18:59
  first
• 18:59
  ad-tech
• 19:00
  company
• 19:00
  to
• 19:00
  integrate
• 19:01
  with
• 19:01
  OpenAI's
• 19:01
  advertising
• 19:02
  pilot
• 19:02
  in
• 19:02
  ChatGPT's
• 19:03
  free
• 19:03
  tier,
• 19:04
  pitching
• 19:04
  its
• 19:04
  17,000
• 19:05
  advertisers
• 19:06
  on
• 19:06
  placing
• 19:06
  ads
• 19:06
  inside
• 19:07
  conversations.
• 19:08
  Early
• 19:09
  data
• 19:09
  from
• 19:09
  500
• 19:09
  Criteo
• 19:10
  retailers
• 19:10
  showed
• 19:11
  users
• 19:11
  arriving
• 19:11
  at
• 19:11
  retail
• 19:12
  sites
• 19:12
  from
• 19:12
  LLM
• 19:12
  platforms
• 19:13
  converted
• 19:13
  at
• 19:13
  1.5x
• 19:14
  the
• 19:14
  rate
• 19:14
  of
• 19:15
  other
• 19:15
  referral
• 19:15
  channels.
• 19:16
  The
• 19:16
  structural
• 19:17
  shift
• 19:17
  is
• 19:17
  that
• 19:17
  the
• 19:17
  purchase
• 19:17
  funnel
• 19:18
  —
• 19:19
  discovery,
• 19:19
  consideration,
• 19:20
  conversion
• 19:20
  —
• 19:21
  collapses
• 19:21
  into
• 19:22
  a
• 19:22
  single
• 19:22
  conversation.
• 19:23
  There
• 19:24
  is
• 19:24
  no
• 19:24
  page
• 19:24
  one
• 19:24
  versus
• 19:24
  page
• 19:25
  two,
• 19:25
  no
• 19:25
  list
• 19:26
  of
• 19:26
  ten
• 19:26
  blue
• 19:26
  links.
• 19:27
  There
• 19:27
  is
• 19:27
  just
• 19:28
  a
• 19:28
  recommendation
• 19:28
  woven
• 19:29
  into
• 19:29
  a
• 19:29
  response
• 19:30
  the
• 19:30
  user
• 19:30
  trusts.
• 19:31
  If
• 19:32
  conversational
• 19:32
  AI
• 19:33
  captures
• 19:33
  purchase
• 19:33
  intent
• 19:34
  upstream
• 19:34
  before
• 19:34
  the
• 19:35
  user
• 19:35
  ever
• 19:35
  opens
• 19:35
  a
• 19:35
  browser,
• 19:36
  that
• 19:36
  $600
• 19:37
  billion
• 19:37
  global
• 19:38
  ad
• 19:38
  budget
• 19:38
  migrates
• 19:39
  with
• 19:39
  it.
• 19:40
  On
• 19:40
  the
• 19:40
  SaaS
• 19:40
  front,
• 19:41
  Atlassian's
• 19:42
  CEO
• 19:42
  pledged
• 19:43
  in
• 19:43
  October
• 19:43
  2025
• 19:44
  to
• 19:45
  employ
• 19:45
  more
• 19:45
  engineers
• 19:46
  in
• 19:46
  five
• 19:46
  years
• 19:46
  and
• 19:46
  hire
• 19:47
  more
• 19:47
  new
• 19:47
  graduates
• 19:48
  in
• 19:48
  2025-2026.
• 19:50
  By
• 19:50
  March,
• 19:51
  the
• 19:51
  company
• 19:52
  cut
• 19:52
  1,600
• 19:53
  jobs
• 19:53
  —
• 19:54
  10%
• 19:55
  of
• 19:55
  the
• 19:55
  workforce,
• 19:56
  over
• 19:56
  900
• 19:57
  in
• 19:57
  software
• 19:57
  —
• 19:59
  and
• 19:59
  reported
• 19:59
  its
• 19:59
  first-ever
• 20:00
  decline
• 20:00
  in
• 20:00
  enterprise
• 20:01
  seat
• 20:01
  counts.
• 20:03
  The
• 20:03
  mechanism
• 20:03
  is
• 20:03
  simple:
• 20:04
  if
• 20:05
  AI
• 20:05
  agents
• 20:05
  can
• 20:06
  do
• 20:06
  the
• 20:06
  work
• 20:06
  of
• 20:06
  multiple
• 20:06
  seats,
• 20:07
  per-seat
• 20:08
  pricing
• 20:08
  compresses
• 20:08
  revenue.
• 20:10
  Block
• 20:10
  cut
• 20:10
  4,000
• 20:11
  jobs
• 20:12
  citing
• 20:12
  an
• 20:12
  "intelligence
• 20:13
  native"
• 20:13
  model.
• 20:14
  Oracle
• 20:15
  said
• 20:15
  AI
• 20:15
  was
• 20:15
  enabling
• 20:16
  it
• 20:16
  to
• 20:16
  shrink
• 20:16
  development
• 20:17
  teams.
• 20:18
  Tech
• 20:18
  layoffs
• 20:19
  globally
• 20:19
  surpassed
• 20:20
  45,000
• 20:21
  by
• 20:21
  early
• 20:21
  March.
• 20:22
  The
• 20:23
  market
• 20:23
  has
• 20:23
  seen
• 20:23
  that
• 20:23
  per-seat
• 20:24
  pricing
• 20:24
  is
• 20:24
  over
• 20:24
  faster
• 20:25
  than
• 20:25
  most
• 20:25
  SaaS
• 20:25
  companies,
• 20:26
  and
• 20:26
  until
• 20:27
  they
• 20:27
  find
• 20:27
  viable
• 20:27
  outcome-driven
• 20:28
  pricing
• 20:28
  models,
• 20:29
  the
• 20:29
  punishment
• 20:30
  will
• 20:30
  continue.