Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 18, 2026

Transcript

AGI Dreams Podcast – June 18, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Anthropic
• 0:00
  CEO
• 0:01
  Dario
• 0:01
  Amodei
• 0:02
  and
• 0:02
  Google
• 0:02
  DeepMind's
• 0:02
  Demis
• 0:03
  Hassabis
• 0:03
  showed
• 0:03
  up
• 0:04
  at
• 0:04
  the
• 0:04
  G7
• 0:04
  this
• 0:04
  week
• 0:05
  to
• 0:05
  pitch
• 0:05
  a
• 0:05
  U.S.-led
• 0:06
  AI
• 0:06
  coalition.
• 0:07
  The
• 0:07
  community
• 0:07
  response
• 0:08
  was
• 0:08
  immediate
• 0:08
  and
• 0:09
  uniformly
• 0:09
  skeptical.
• 0:10
  The
• 0:10
  top-voted
• 0:11
  comment
• 0:11
  captured
• 0:11
  the
• 0:11
  mood:
• 0:12
  "If
• 0:12
  my
• 0:12
  trillion-dollar
• 0:13
  industry
• 0:13
  is
• 0:14
  only
• 0:14
  six
• 0:14
  months
• 0:14
  ahead
• 0:14
  of
• 0:15
  a
• 0:15
  Chinese
• 0:15
  PhD
• 0:15
  with
• 0:16
  a
• 0:16
  bunch
• 0:16
  of
• 0:16
  GPUs,
• 0:17
  I
• 0:17
  would
• 0:17
  also
• 0:17
  seek
• 0:18
  a
• 0:18
  coalition."
• 0:19
  Others
• 0:19
  pointed
• 0:20
  out
• 0:20
  the
• 0:20
  obvious
• 0:20
  credibility
• 0:21
  problem
• 0:21
  —
• 0:22
  the
• 0:22
  same
• 0:22
  U.S.
• 0:22
  government
• 0:23
  that
• 0:23
  started
• 0:23
  a
• 0:23
  tariff
• 0:24
  war
• 0:24
  with
• 0:24
  Europe,
• 0:25
  threatened
• 0:25
  Greenland,
• 0:26
  and
• 0:26
  backed
• 0:26
  pro-Russia
• 0:27
  politicians
• 0:27
  in
• 0:28
  EU
• 0:28
  elections
• 0:29
  is
• 0:29
  now
• 0:29
  asking
• 0:29
  allies
• 0:30
  to
• 0:30
  co-fund
• 0:30
  American
• 0:31
  AI
• 0:31
  infrastructure.
• 0:32
  The
• 0:32
  F-35
• 0:33
  analogy
• 0:33
  kept
• 0:33
  surfacing:
• 0:34
  the
• 0:34
  U.S.
• 0:35
  retains
• 0:35
  total
• 0:35
  control,
• 0:36
  partners
• 0:36
  fund
• 0:37
  and
• 0:37
  get
• 0:37
  minimal
• 0:37
  contribution
• 0:38
  rights.
• 0:39
  None
• 0:39
  of
• 0:39
  this
• 0:39
  is
• 0:39
  new.
• 0:40
  Back
• 0:41
  in
• 0:41
  OpenAI's
• 0:41
  early
• 0:41
  days,
• 0:42
  Altman's
• 0:43
  team
• 0:43
  outlined
• 0:43
  how
• 0:43
  a
• 0:44
  "coalition
• 0:44
  of
• 0:44
  labs"
• 0:45
  could
• 0:45
  coordinate
• 0:45
  through
• 0:46
  an
• 0:46
  IAEA-like
• 0:46
  body.
• 0:47
  That
• 0:47
  vision
• 0:48
  never
• 0:48
  materialized.
• 0:49
  What
• 0:50
  changed
• 0:50
  is
• 0:50
  that
• 0:50
  DeepSeek
• 0:51
  proved
• 0:51
  the
• 0:51
  cost-moat
• 0:51
  argument
• 0:52
  is
• 0:52
  thin,
• 0:52
  and
• 0:53
  now
• 0:53
  the
• 0:53
  labs
• 0:53
  want
• 0:53
  regulatory
• 0:54
  barriers
• 0:54
  instead.
• 0:55
  As
• 0:55
  one
• 0:55
  commenter
• 0:56
  put
• 0:56
  it:
• 0:56
  "Without
• 0:57
  regulation,
• 0:57
  no
• 0:58
  moat."
• 0:58
  Meanwhile,
• 0:59
  a
• 0:59
  user
• 1:00
  discovered
• 1:00
  that
• 1:00
  Anthropic
• 1:01
  hard-blocks
• 1:01
  Claude
• 1:02
  from
• 1:02
  outputting
• 1:02
  the
• 1:02
  verbatim
• 1:03
  text
• 1:03
  of
• 1:03
  copyleft
• 1:04
  licenses
• 1:05
  —
• 1:06
  AGPL-3,
• 1:07
  GPL-3,
• 1:08
  MPL-2.0,
• 1:10
  LGPL-2.1,
• 1:11
  EPL-2.0,
• 1:12
  Apache
• 1:13
  2.0,
• 1:14
  SSPL,
• 1:15
  and
• 1:15
  BSL
• 1:16
  all
• 1:16
  refused.
• 1:17
  Permissive
• 1:17
  licenses
• 1:18
  like
• 1:18
  MIT,
• 1:19
  BSD-2,
• 1:20
  BSD-3,
• 1:21
  and
• 1:21
  Unlicense
• 1:21
  went
• 1:22
  through
• 1:22
  fine.
• 1:23
  The
• 1:23
  blocked
• 1:23
  list
• 1:24
  maps
• 1:24
  almost
• 1:24
  perfectly
• 1:25
  to
• 1:25
  licenses
• 1:25
  longer
• 1:26
  than
• 1:26
  300
• 1:26
  words,
• 1:27
  which
• 1:27
  makes
• 1:28
  the
• 1:28
  most
• 1:28
  plausible
• 1:28
  explanation
• 1:29
  a
• 1:29
  verbatim-reproduction
• 1:30
  safeguard
• 1:31
  rather
• 1:31
  than
• 1:31
  an
• 1:31
  ideological
• 1:32
  stance.
• 1:33
  Multiple
• 1:34
  commenters
• 1:34
  noted
• 1:35
  the
• 1:35
  irony
• 1:35
  of
• 1:35
  spending
• 1:35
  tokens
• 1:36
  generating
• 1:36
  a
• 1:36
  LICENSE
• 1:37
  file
• 1:37
  instead
• 1:38
  of
• 1:38
  copy-pasting
• 1:38
  it.
• 1:39
  But
• 1:39
  there
• 1:40
  is
• 1:40
  a
• 1:40
  sharper
• 1:40
  read:
• 1:40
  if
• 1:41
  agentic
• 1:41
  coding
• 1:41
  makes
• 1:42
  copyleft
• 1:42
  matter
• 1:42
  again
• 1:42
  —
• 1:43
  because
• 1:43
  agents
• 1:44
  need
• 1:44
  source
• 1:44
  access
• 1:45
  to
• 1:45
  modify
• 1:45
  software
• 1:45
  —
• 1:46
  then
• 1:46
  the
• 1:47
  largest
• 1:47
  agent
• 1:47
  provider
• 1:48
  selectively
• 1:48
  blocking
• 1:48
  copyleft
• 1:49
  text
• 1:49
  is
• 1:50
  worth
• 1:50
  watching,
• 1:51
  even
• 1:51
  if
• 1:51
  the
• 1:51
  mechanism
• 1:52
  is
• 1:52
  mundane.
• 1:53
  Daniel
• 1:53
  Miessler
• 1:54
  offered
• 1:54
  what
• 1:54
  he
• 1:55
  calls
• 1:55
  a
• 1:55
  "unified
• 1:55
  theory"
• 1:56
  of
• 1:56
  AI
• 1:56
  and
• 1:56
  jobs.
• 1:57
  His
• 1:57
  thesis:
• 1:58
  both
• 1:59
  camps
• 1:59
  are
• 1:59
  right.
• 2:00
  Lots
• 2:00
  of
• 2:00
  jobs
• 2:01
  will
• 2:01
  vanish,
• 2:01
  lots
• 2:02
  will
• 2:02
  appear,
• 2:03
  but
• 2:03
  the
• 2:03
  new
• 2:03
  employability
• 2:04
  bar
• 2:04
  will
• 2:04
  be
• 2:04
  dramatically
• 2:05
  higher,
• 2:05
  producing
• 2:06
  a
• 2:06
  K-shaped
• 2:06
  split
• 2:07
  between
• 2:07
  those
• 2:07
  who
• 2:07
  adapt
• 2:08
  and
• 2:08
  those
• 2:08
  who
• 2:08
  do
• 2:08
  not.
• 2:10
  He
• 2:10
  frames
• 2:10
  this
• 2:10
  as
• 2:10
  "escape
• 2:11
  velocity"
• 2:11
  —
• 2:12
  a
• 2:12
  binary
• 2:13
  threshold,
• 2:13
  not
• 2:13
  a
• 2:13
  gradient.
• 2:15
  The
• 2:15
  list
• 2:15
  of
• 2:15
  thriving-vs-struggling
• 2:16
  traits
• 2:17
  reads
• 2:17
  more
• 2:17
  like
• 2:18
  a
• 2:18
  self-help
• 2:18
  manifesto
• 2:19
  than
• 2:19
  labor
• 2:19
  economics
• 2:20
  (curiosity
• 2:20
  good,
• 2:21
  Netflix
• 2:21
  bad),
• 2:22
  and
• 2:22
  it
• 2:22
  conveniently
• 2:23
  sidesteps
• 2:23
  structural
• 2:24
  questions
• 2:24
  about
• 2:25
  who
• 2:25
  gets
• 2:25
  access
• 2:25
  to
• 2:25
  retraining.
• 2:27
  But
• 2:27
  the
• 2:27
  K-shape
• 2:27
  framing
• 2:28
  itself
• 2:28
  has
• 2:28
  legs.
• 2:29
  We
• 2:29
  have
• 2:29
  tracked
• 2:29
  this
• 2:30
  since
• 2:30
  Edition
• 2:30
  234,
• 2:32
  when
• 2:32
  Microsoft/ACM
• 2:33
  research
• 2:34
  showed
• 2:34
  AI
• 2:34
  creating
• 2:35
  "seniority-biased
• 2:36
  technological
• 2:36
  change"
• 2:37
  —
• 2:37
  amplifying
• 2:38
  experienced
• 2:39
  engineers
• 2:39
  while
• 2:39
  dragging
• 2:40
  down
• 2:40
  juniors.
• 2:41
  Miessler's
• 2:42
  contribution
• 2:42
  is
• 2:42
  packaging
• 2:43
  the
• 2:43
  split
• 2:43
  as
• 2:43
  a
• 2:43
  personal
• 2:44
  optimization
• 2:44
  problem.
• 2:45
  Whether
• 2:46
  that
• 2:46
  framing
• 2:46
  helps
• 2:47
  or
• 2:47
  merely
• 2:47
  flatters
• 2:47
  the
• 2:47
  already-thriving
• 2:48
  is
• 2:49
  the
• 2:49
  question
• 2:49
  he
• 2:49
  does
• 2:49
  not
• 2:49
  answer.
• 2:53
  The
• 2:53
  4x
• 2:54
  RTX
• 2:54
  5060
• 2:55
  Ti
• 2:56
  build
• 2:56
  saga
• 2:56
  continues.
• 2:57
  One
• 2:57
  builder
• 2:57
  assembled
• 2:58
  a
• 2:58
  quad-card
• 2:58
  rig
• 2:59
  on
• 2:59
  an
• 2:59
  MSI
• 2:59
  MEG
• 2:59
  Z890
• 3:00
  Unify-X
• 3:01
  board,
• 3:01
  using
• 3:02
  M.2-to-PCIe
• 3:03
  adapters
• 3:03
  to
• 3:03
  slot
• 3:04
  two
• 3:04
  extra
• 3:04
  GPUs
• 3:05
  via
• 3:05
  PCIe
• 3:05
  5.0
• 3:06
  x4
• 3:06
  CPU
• 3:07
  lanes.
• 3:08
  Memory
• 3:08
  overclocking
• 3:08
  hit
• 3:09
  +6000
• 3:09
  MT/s
• 3:10
  on
• 3:10
  most
• 3:10
  cards.
• 3:11
  The
• 3:11
  goal:
• 3:12
  Qwen
• 3:12
  3.6
• 3:12
  27B
• 3:13
  at
• 3:13
  Q8
• 3:13
  with
• 3:14
  64
• 3:14
  GB
• 3:14
  of
• 3:14
  combined
• 3:15
  VRAM.
• 3:15
  The
• 3:15
  update
• 3:16
  was
• 3:16
  less
• 3:16
  triumphant
• 3:16
  —
• 3:17
  both
• 3:17
  M.2-connected
• 3:18
  GPUs
• 3:18
  drop
• 3:18
  off
• 3:19
  the
• 3:19
  PCIe
• 3:19
  bus
• 3:19
  under
• 3:20
  vLLM
• 3:20
  load.
• 3:21
  Another
• 3:21
  user
• 3:21
  reported
• 3:21
  their
• 3:22
  4x
• 3:22
  5060
• 3:23
  Ti
• 3:23
  rig
• 3:23
  hitting
• 3:24
  120
• 3:24
  tokens/second
• 3:25
  on
• 3:26
  Qwen
• 3:26
  3.6
• 3:26
  35B
• 3:27
  via
• 3:27
  llama.cpp,
• 3:29
  which
• 3:29
  is
• 3:29
  the
• 3:29
  real
• 3:29
  benchmark
• 3:30
  people
• 3:30
  care
• 3:30
  about.
• 3:31
  The
• 3:31
  5060
• 3:32
  Ti
• 3:32
  at
• 3:32
  $400-ish
• 3:33
  remains
• 3:33
  the
• 3:33
  price-performance
• 3:34
  sweet
• 3:34
  spot
• 3:35
  for
• 3:35
  multi-GPU
• 3:35
  local
• 3:36
  inference,
• 3:36
  but
• 3:37
  PCIe
• 3:37
  bandwidth
• 3:37
  constraints
• 3:38
  on
• 3:38
  consumer
• 3:38
  boards
• 3:39
  keep
• 3:39
  biting.
• 3:40
  On
• 3:40
  the
• 3:40
  Mac
• 3:40
  side,
• 3:41
  Antirez's
• 3:42
  ds4
• 3:42
  engine
• 3:42
  now
• 3:43
  runs
• 3:43
  DeepSeek
• 3:43
  4
• 3:44
  Flash
• 3:44
  on
• 3:44
  an
• 3:44
  M3
• 3:44
  Max
• 3:45
  with
• 3:45
  96
• 3:46
  GB
• 3:46
  unified
• 3:46
  memory
• 3:47
  via
• 3:47
  SSD
• 3:47
  streaming.
• 3:49
  Performance:
• 3:50
  11-13
• 3:50
  tokens/second
• 3:51
  generation,
• 3:52
  3-5
• 3:53
  second
• 3:53
  time-to-first-token
• 3:54
  after
• 3:54
  warm
• 3:54
  cache.
• 3:55
  The
• 3:55
  catch
• 3:55
  is
• 3:56
  prefill
• 3:56
  —
• 3:57
  36K
• 3:57
  tokens
• 3:58
  takes
• 3:58
  two
• 3:58
  and
• 3:58
  a
• 3:58
  half
• 3:59
  minutes,
• 3:59
  which
• 4:00
  kills
• 4:00
  any
• 4:00
  agentic
• 4:00
  coding
• 4:01
  loop
• 4:01
  where
• 4:01
  context
• 4:01
  rebuilds
• 4:02
  between
• 4:02
  tool
• 4:02
  calls.
• 4:03
  One
• 4:04
  user's
• 4:04
  workaround:
• 4:05
  pair
• 4:05
  DS4
• 4:05
  Flash
• 4:06
  with
• 4:06
  a
• 4:06
  smaller
• 4:06
  model
• 4:07
  for
• 4:07
  the
• 4:07
  file-reading
• 4:07
  step,
• 4:08
  reserve
• 4:08
  the
• 4:09
  big
• 4:09
  model
• 4:09
  for
• 4:09
  final
• 4:09
  generation
• 4:10
  only.
• 4:11
  Another
• 4:11
  reported
• 4:12
  that
• 4:12
  on
• 4:12
  an
• 4:12
  M3
• 4:12
  Ultra,
• 4:13
  DS4
• 4:13
  Flash
• 4:14
  at
• 4:14
  IQ2XXS
• 4:16
  matched
• 4:17
  Gemma
• 4:17
  4
• 4:17
  31B
• 4:18
  Q4
• 4:18
  speeds
• 4:19
  at
• 4:19
  32
• 4:20
  tokens/second.
• 4:21
  The
• 4:21
  SSD
• 4:22
  wear
• 4:22
  question
• 4:22
  is
• 4:22
  legitimate
• 4:23
  for
• 4:23
  daily
• 4:23
  use,
• 4:24
  but
• 4:24
  the
• 4:24
  capability
• 4:25
  floor
• 4:25
  keeps
• 4:25
  dropping:
• 4:26
  a
• 4:26
  model
• 4:27
  that
• 4:27
  needed
• 4:27
  datacenter
• 4:28
  hardware
• 4:28
  a
• 4:28
  year
• 4:28
  ago
• 4:29
  now
• 4:29
  runs
• 4:29
  on
• 4:29
  a
• 4:29
  laptop,
• 4:30
  slowly.
• 4:31
  The
• 4:31
  Anthropic
• 4:32
  service
• 4:32
  interruption
• 4:32
  this
• 4:32
  week
• 4:33
  catalyzed
• 4:34
  a
• 4:34
  builder
• 4:34
  who
• 4:34
  had
• 4:34
  been
• 4:35
  expecting
• 4:35
  exactly
• 4:35
  this
• 4:36
  moment.
• 4:37
  Bantz
• 4:37
  is
• 4:37
  a
• 4:37
  fully
• 4:37
  local
• 4:38
  AI
• 4:38
  personal
• 4:39
  assistant
• 4:39
  running
• 4:39
  on
• 4:39
  Gemma
• 4:40
  4B:
• 4:41
  Gmail
• 4:41
  summarization
• 4:42
  by
• 4:42
  category,
• 4:43
  Google
• 4:43
  Calendar
• 4:43
  integration,
• 4:44
  web
• 4:45
  search,
• 4:45
  real-time
• 4:46
  system
• 4:46
  monitoring,
• 4:47
  scheduled
• 4:47
  tasks,
• 4:48
  and
• 4:48
  early
• 4:48
  Wayland
• 4:48
  desktop
• 4:49
  control
• 4:49
  —
• 4:50
  all
• 4:50
  on
• 4:50
  CPU,
• 4:51
  no
• 4:51
  GPU
• 4:52
  required.
• 4:53
  The
• 4:53
  comments
• 4:54
  were
• 4:54
  more
• 4:54
  interesting
• 4:54
  than
• 4:54
  the
• 4:54
  project
• 4:55
  itself.
• 4:56
  The
• 4:56
  highest-rated
• 4:57
  response
• 4:57
  nailed
• 4:57
  the
• 4:57
  real
• 4:58
  difficulty:
• 4:59
  "The
• 4:59
  model
• 4:59
  was
• 4:59
  not
• 4:59
  the
• 4:59
  hard
• 5:00
  bit.
• 5:00
  The
• 5:00
  hard
• 5:01
  bit
• 5:01
  was
• 5:01
  everything
• 5:01
  around
• 5:02
  it:
• 5:02
  memory
• 5:02
  that
• 5:03
  doesn't
• 5:03
  turn
• 5:03
  into
• 5:03
  junk,
• 5:04
  permissions
• 5:05
  narrow
• 5:05
  enough
• 5:05
  to
• 5:05
  trust,
• 5:06
  retries
• 5:06
  when
• 5:06
  a
• 5:07
  task
• 5:07
  half-fails."
• 5:08
  The
• 5:09
  boundary
• 5:09
  between
• 5:09
  what
• 5:09
  the
• 5:09
  assistant
• 5:10
  can
• 5:10
  see,
• 5:11
  what
• 5:11
  it
• 5:11
  can
• 5:11
  act
• 5:11
  on,
• 5:12
  and
• 5:12
  what
• 5:12
  still
• 5:12
  requires
• 5:13
  explicit
• 5:13
  approval
• 5:14
  —
• 5:14
  that
• 5:15
  turned
• 5:15
  out
• 5:15
  to
• 5:15
  be
• 5:15
  the
• 5:15
  design
• 5:16
  center.
• 5:16
  Not
• 5:17
  the
• 5:17
  chat
• 5:17
  interface.
• 5:18
  Not
• 5:18
  the
• 5:18
  persona.
• 5:19
  A
• 5:19
  streaming
• 5:20
  copilot
• 5:20
  project
• 5:21
  pushed
• 5:21
  local
• 5:21
  inference
• 5:22
  to
• 5:22
  the
• 5:22
  edge
• 5:22
  —
• 5:23
  literally.
• 5:24
  Running
• 5:24
  on
• 5:24
  a
• 5:25
  Jetson
• 5:25
  AGX
• 5:25
  Orin
• 5:25
  with
• 5:26
  61
• 5:26
  GB
• 5:27
  RAM,
• 5:27
  it
• 5:27
  does
• 5:27
  live
• 5:28
  transcription
• 5:28
  via
• 5:28
  Faster
• 5:29
  Whisper
• 5:29
  large-v3-turbo
• 5:30
  on
• 5:30
  CUDA
• 5:31
  (~0.06
• 5:32
  RTF,
• 5:33
  basically
• 5:34
  real
• 5:34
  time),
• 5:35
  feeds
• 5:35
  suggestions
• 5:36
  through
• 5:36
  a
• 5:36
  local
• 5:36
  Ollama
• 5:37
  model
• 5:37
  (Gemma
• 5:37
  4),
• 5:38
  and
• 5:38
  surfaces
• 5:38
  them
• 5:39
  in
• 5:39
  a
• 5:39
  web
• 5:39
  dashboard.
• 5:40
  Multi-platform
• 5:41
  chat
• 5:41
  integration
• 5:42
  (Twitch,
• 5:42
  YouTube,
• 5:43
  TikTok),
• 5:44
  post-stream
• 5:44
  report
• 5:44
  cards,
• 5:45
  session
• 5:45
  replay.
• 5:46
  No
• 5:46
  audio
• 5:47
  or
• 5:47
  transcript
• 5:47
  leaves
• 5:48
  the
• 5:48
  device.
• 5:49
  The
• 5:49
  Jetson
• 5:49
  angle
• 5:49
  is
• 5:50
  new
• 5:50
  territory
• 5:50
  for
• 5:50
  the
• 5:51
  local-AI
• 5:51
  movement
• 5:51
  —
• 5:52
  the
• 5:52
  jump
• 5:52
  from
• 5:53
  "can
• 5:53
  I
• 5:53
  run
• 5:53
  a
• 5:53
  model
• 5:53
  on
• 5:53
  my
• 5:54
  desktop"
• 5:54
  to
• 5:54
  "can
• 5:54
  I
• 5:55
  run
• 5:55
  a
• 5:55
  real-time
• 5:55
  streaming
• 5:56
  copilot
• 5:56
  on
• 5:56
  edge
• 5:56
  hardware."
• 6:01
  Local
• 6:02
  coding
• 6:02
  agents
• 6:02
  are
• 6:03
  good
• 6:03
  now,
• 6:03
  but
• 6:03
  the
• 6:04
  babysitting
• 6:04
  problem
• 6:05
  has
• 6:05
  not
• 6:05
  gone
• 6:05
  away.
• 6:06
  One
• 6:06
  practitioner's
• 6:07
  workflow
• 6:08
  is
• 6:08
  the
• 6:08
  same
• 6:08
  loop
• 6:08
  everyone
• 6:09
  converges
• 6:09
  on:
• 6:10
  small
• 6:10
  task,
• 6:11
  run
• 6:11
  tests,
• 6:12
  check
• 6:12
  diff,
• 6:13
  fix
• 6:13
  the
• 6:13
  weird
• 6:13
  part,
• 6:14
  repeat.
• 6:15
  Local
• 6:15
  models
• 6:16
  drift
• 6:16
  when
• 6:16
  task
• 6:17
  boundaries
• 6:17
  are
• 6:17
  vague
• 6:17
  —
• 6:18
  "refactor
• 6:19
  this
• 6:19
  module"
• 6:19
  fails,
• 6:20
  "change
• 6:20
  this
• 6:20
  method
• 6:21
  signature
• 6:21
  and
• 6:21
  update
• 6:22
  all
• 6:22
  callers
• 6:22
  in
• 6:22
  this
• 6:22
  file"
• 6:23
  mostly
• 6:23
  works.
• 6:24
  The
• 6:24
  most
• 6:24
  sophisticated
• 6:25
  workaround
• 6:26
  reported:
• 6:27
  use
• 6:27
  Opus
• 6:27
  to
• 6:27
  babysit
• 6:28
  a
• 6:28
  local
• 6:28
  Qwen,
• 6:29
  having
• 6:29
  the
• 6:29
  frontier
• 6:30
  model
• 6:30
  write
• 6:30
  notes
• 6:30
  so
• 6:31
  the
• 6:31
  cheap
• 6:31
  model
• 6:31
  can
• 6:31
  handle
• 6:32
  the
• 6:32
  same
• 6:32
  task
• 6:32
  in
• 6:32
  future
• 6:33
  runs.
• 6:34
  Another
• 6:34
  user
• 6:34
  runs
• 6:35
  Qwen
• 6:35
  3.6:27B
• 6:38
  inside
• 6:38
  an
• 6:39
  ephemeral
• 6:39
  Docker
• 6:39
  container
• 6:40
  with
• 6:40
  Opus
• 6:40
  overseeing
• 6:40
  —
• 6:41
  an
• 6:42
  80/20
• 6:42
  token
• 6:42
  ratio
• 6:43
  of
• 6:43
  local
• 6:43
  to
• 6:43
  frontier,
• 6:44
  molecular
• 6:45
  task
• 6:45
  slices,
• 6:46
  checkpoint-and-destroy
• 6:47
  after
• 6:47
  each
• 6:47
  feature
• 6:47
  slice.
• 6:48
  The
• 6:48
  pattern
• 6:49
  is
• 6:49
  clear:
• 6:50
  local
• 6:50
  models
• 6:50
  need
• 6:51
  a
• 6:51
  supervisor,
• 6:52
  and
• 6:52
  the
• 6:52
  supervisor
• 6:53
  is
• 6:53
  currently
• 6:53
  another
• 6:53
  model.
• 6:54
  A
• 6:54
  YouTube
• 6:55
  walkthrough
• 6:55
  of
• 6:55
  a
• 6:55
  "Fable
• 6:56
  5
• 6:56
  Agentic
• 6:56
  OS"
• 6:57
  —
• 6:58
  really
• 6:58
  a
• 6:58
  Claude
• 6:58
  Code
• 6:59
  wrapper
• 6:59
  with
• 6:59
  a
• 6:59
  visual
• 6:59
  HUD
• 7:00
  layer
• 7:00
  —
• 7:01
  demonstrated
• 7:01
  the
• 7:01
  skill
• 7:01
  architecture
• 7:02
  approach.
• 7:03
  Faster
• 7:03
  Whisper
• 7:04
  handles
• 7:04
  speech-to-text,
• 7:05
  a
• 7:06
  three-tier
• 7:06
  routing
• 7:06
  system
• 7:07
  (regex,
• 7:07
  Haiku,
• 7:08
  local
• 7:09
  model)
• 7:09
  dispatches
• 7:09
  commands,
• 7:10
  headless
• 7:11
  Claude
• 7:11
  Code
• 7:11
  (-p
• 7:12
  flag)
• 7:12
  executes
• 7:12
  them,
• 7:13
  and
• 7:13
  Kokoro
• 7:14
  TTS
• 7:14
  voices
• 7:15
  the
• 7:15
  responses.
• 7:16
  The
• 7:16
  key
• 7:17
  insight
• 7:17
  is
• 7:17
  not
• 7:17
  the
• 7:17
  Fable
• 7:18
  5
• 7:18
  branding
• 7:18
  (the
• 7:18
  system
• 7:19
  does
• 7:19
  not
• 7:19
  require
• 7:20
  Fable
• 7:20
  5
• 7:20
  and
• 7:20
  mostly
• 7:21
  runs
• 7:21
  on
• 7:21
  Opus/Sonnet)
• 7:21
  —
• 7:23
  it
• 7:23
  is
• 7:23
  the
• 7:23
  skill
• 7:23
  backbone.
• 7:24
  Every
• 7:25
  manual
• 7:25
  workflow
• 7:26
  becomes
• 7:26
  a
• 7:26
  reusable
• 7:27
  skill
• 7:27
  with
• 7:27
  a
• 7:27
  button.
• 7:28
  Obsidian
• 7:29
  integration
• 7:29
  provides
• 7:30
  persistent
• 7:30
  knowledge.
• 7:31
  The
• 7:31
  $200/month
• 7:32
  API
• 7:33
  credits
• 7:33
  coming
• 7:34
  for
• 7:34
  headless
• 7:34
  mode
• 7:34
  will
• 7:34
  lower
• 7:35
  the
• 7:35
  barrier,
• 7:35
  but
• 7:36
  the
• 7:36
  architecture
• 7:36
  pattern
• 7:37
  matters
• 7:37
  more
• 7:37
  than
• 7:37
  the
• 7:37
  specific
• 7:38
  implementation.
• 7:39
  The
• 7:39
  editorial
• 7:40
  framing
• 7:40
  that
• 7:40
  pulls
• 7:41
  all
• 7:41
  of
• 7:41
  this
• 7:41
  together
• 7:42
  came
• 7:42
  from
• 7:42
  a
• 7:42
  video
• 7:43
  essay
• 7:43
  arguing
• 7:43
  that
• 7:44
  agent
• 7:44
  maintenance
• 7:44
  is
• 7:45
  the
• 7:45
  real
• 7:45
  story
• 7:45
  of
• 7:46
  2026.
• 7:47
  The
• 7:47
  central
• 7:48
  example:
• 7:48
  Vercel
• 7:49
  improved
• 7:49
  its
• 7:49
  sales
• 7:50
  agent
• 7:50
  by
• 7:50
  removing
• 7:51
  80%
• 7:51
  of
• 7:51
  its
• 7:52
  tools.
• 7:53
  Four
• 7:53
  principles
• 7:53
  emerged.
• 7:54
  First,
• 7:55
  agents
• 7:55
  break
• 7:56
  when
• 7:56
  models
• 7:56
  get
• 7:56
  better
• 7:57
  —
• 7:57
  a
• 7:57
  harness
• 7:58
  tuned
• 7:58
  for
• 7:58
  a
• 7:58
  weak
• 7:58
  model
• 7:59
  becomes
• 7:59
  a
• 7:59
  drag
• 8:00
  on
• 8:00
  a
• 8:00
  stronger
• 8:00
  one.
• 8:01
  Second,
• 8:02
  agents
• 8:02
  inherit
• 8:03
  organizational
• 8:03
  crud:
• 8:04
  stale
• 8:04
  wikis,
• 8:05
  outdated
• 8:06
  processes,
• 8:07
  drifted
• 8:07
  definitions.
• 8:09
  Third,
• 8:09
  the
• 8:10
  big
• 8:10
  labs
• 8:10
  bet
• 8:10
  on
• 8:11
  the
• 8:11
  harness
• 8:11
  flywheel
• 8:11
  —
• 8:12
  better
• 8:12
  models
• 8:13
  help
• 8:13
  build
• 8:13
  better
• 8:14
  harnesses,
• 8:14
  which
• 8:14
  make
• 8:15
  agents
• 8:15
  more
• 8:15
  useful,
• 8:16
  which
• 8:16
  justify
• 8:17
  better
• 8:17
  models.
• 8:18
  Fourth,
• 8:19
  the
• 8:19
  question
• 8:20
  everyone
• 8:20
  needs
• 8:20
  to
• 8:20
  answer
• 8:21
  is
• 8:21
  "what
• 8:21
  is
• 8:21
  my
• 8:21
  harness?"
• 8:23
  Stewart
• 8:23
  Brand's
• 8:23
  Maintenance
• 8:24
  of
• 8:24
  Everything
• 8:24
  provided
• 8:25
  the
• 8:25
  frame:
• 8:26
  agents
• 8:26
  are
• 8:26
  more
• 8:27
  like
• 8:27
  sailboats
• 8:27
  than
• 8:28
  apps.
• 8:29
  You
• 8:29
  do
• 8:29
  not
• 8:29
  launch
• 8:29
  them
• 8:30
  once.
• 8:31
  The
• 8:31
  weather
• 8:31
  changes,
• 8:32
  the
• 8:32
  lines
• 8:32
  loosen,
• 8:33
  salt
• 8:33
  gets
• 8:33
  into
• 8:33
  everything.
• 8:35
  Five
• 8:35
  checks:
• 8:36
  sources
• 8:36
  (are
• 8:36
  they
• 8:36
  current?),
• 8:37
  reach
• 8:37
  (permissions
• 8:38
  still
• 8:38
  right?),
• 8:39
  job
• 8:39
  definition
• 8:40
  (has
• 8:40
  it
• 8:40
  silently
• 8:40
  changed?),
• 8:41
  proof
• 8:41
  trails
• 8:42
  (can
• 8:42
  a
• 8:42
  human
• 8:42
  verify?),
• 8:43
  and
• 8:43
  value
• 8:43
  (does
• 8:43
  anyone
• 8:44
  read
• 8:44
  the
• 8:44
  output?).
• 8:45
  On
• 8:45
  the
• 8:45
  tooling
• 8:46
  side,
• 8:46
  OpenClaw
• 8:47
  published
• 8:47
  a
• 8:47
  shared
• 8:48
  agent-skills
• 8:48
  repository
• 8:48
  —
• 8:50
  canonical
• 8:50
  SKILL.md
• 8:51
  workflows
• 8:52
  for
• 8:52
  review
• 8:52
  closeout,
• 8:53
  remote
• 8:53
  validation,
• 8:54
  session
• 8:55
  transcripts,
• 8:55
  and
• 8:55
  task
• 8:56
  handoff,
• 8:56
  installable
• 8:57
  via
• 8:57
  symlinks
• 8:58
  into
• 8:58
  Claude
• 8:58
  Code
• 8:59
  or
• 8:59
  Codex
• 8:59
  skill
• 9:00
  directories.
• 9:01
  The
• 9:01
  goal:
• 9:02
  write
• 9:02
  a
• 9:02
  workflow
• 9:02
  once,
• 9:03
  reuse
• 9:03
  everywhere,
• 9:04
  stop
• 9:04
  hand-copying
• 9:05
  SKILL.md
• 9:06
  files
• 9:06
  across
• 9:07
  repos.
• 9:08
  Trail
• 9:08
  of
• 9:08
  Bits,
• 9:09
  by
• 9:09
  comparison,
• 9:10
  has
• 9:10
  accumulated
• 9:11
  94
• 9:11
  plugins
• 9:11
  with
• 9:12
  201+
• 9:13
  skills
• 9:13
  and
• 9:13
  84
• 9:14
  specialized
• 9:15
  agents.
• 9:16
  The
• 9:16
  gap
• 9:16
  between
• 9:17
  "five
• 9:17
  shared
• 9:17
  skills
• 9:18
  in
• 9:18
  a
• 9:18
  GitHub
• 9:18
  repo"
• 9:19
  and
• 9:19
  "400+
• 9:20
  reference
• 9:20
  files
• 9:21
  encoding
• 9:21
  domain
• 9:21
  expertise"
• 9:22
  is
• 9:23
  exactly
• 9:23
  the
• 9:23
  maintenance
• 9:24
  problem
• 9:24
  the
• 9:24
  editorial
• 9:25
  described.
• 9:26
  The
• 9:26
  Agentic
• 9:26
  Resource
• 9:27
  Discovery
• 9:27
  (ARD)
• 9:28
  specification,
• 9:29
  developed
• 9:29
  by
• 9:29
  contributors
• 9:30
  from
• 9:30
  Microsoft,
• 9:31
  Google,
• 9:31
  GoDaddy,
• 9:32
  and
• 9:32
  Hugging
• 9:33
  Face,
• 9:34
  addresses
• 9:34
  the
• 9:34
  layer
• 9:34
  beneath
• 9:35
  skills:
• 9:36
  how
• 9:36
  do
• 9:36
  agents
• 9:36
  find
• 9:37
  tools
• 9:37
  and
• 9:37
  capabilities
• 9:38
  at
• 9:38
  runtime
• 9:38
  instead
• 9:39
  of
• 9:39
  needing
• 9:39
  them
• 9:39
  pre-installed?
• 9:41
  ARD
• 9:41
  defines
• 9:41
  a
• 9:41
  static
• 9:42
  manifest
• 9:42
  format
• 9:43
  (ai-catalog.json
• 9:45
  at
• 9:45
  a
• 9:45
  well-known
• 9:46
  URL)
• 9:46
  and
• 9:46
  a
• 9:46
  dynamic
• 9:47
  registry
• 9:47
  REST
• 9:48
  API.
• 9:49
  Hugging
• 9:49
  Face
• 9:50
  shipped
• 9:50
  a
• 9:50
  reference
• 9:50
  implementation
• 9:51
  over
• 9:51
  their
• 9:51
  Hub,
• 9:52
  wrapping
• 9:52
  existing
• 9:53
  Spaces
• 9:53
  search
• 9:54
  into
• 9:54
  ARD
• 9:54
  catalog
• 9:55
  entries
• 9:55
  for
• 9:55
  skills,
• 9:56
  MCP
• 9:56
  servers,
• 9:57
  and
• 9:57
  raw
• 9:57
  Space
• 9:58
  metadata.
• 9:59
  The
• 9:59
  verify
• 10:00
  step
• 10:00
  matters
• 10:00
  most
• 10:00
  —
• 10:01
  discovery
• 10:02
  without
• 10:02
  publisher
• 10:03
  verification
• 10:03
  just
• 10:03
  industrializes
• 10:04
  trusting
• 10:05
  strangers.
• 10:06
  The
• 10:06
  av-curator
• 10:06
  repo
• 10:07
  from
• 10:07
  henliveira
• 10:07
  represents
• 10:08
  the
• 10:08
  other
• 10:08
  end
• 10:08
  of
• 10:08
  the
• 10:09
  spectrum
• 10:09
  —
• 10:10
  a
• 10:10
  single-purpose
• 10:10
  agent
• 10:11
  tool
• 10:11
  for
• 10:11
  audiovisual
• 10:12
  content
• 10:12
  curation,
• 10:13
  the
• 10:13
  kind
• 10:13
  of
• 10:13
  narrow
• 10:14
  utility
• 10:14
  that
• 10:14
  ARD
• 10:15
  registries
• 10:15
  are
• 10:15
  designed
• 10:16
  to
• 10:16
  surface.
• 10:19
  VibeThinker-3B
• 10:20
  is
• 10:20
  reportedly
• 10:21
  crushing
• 10:21
  MathQA
• 10:22
  benchmarks
• 10:22
  at
• 10:22
  levels
• 10:23
  typically
• 10:23
  associated
• 10:24
  with
• 10:24
  30B-parameter
• 10:25
  models.
• 10:25
  The
• 10:25
  community
• 10:26
  reaction
• 10:26
  was
• 10:26
  appropriately
• 10:27
  split.
• 10:28
  The
• 10:28
  impressed
• 10:28
  camp
• 10:28
  noted
• 10:29
  the
• 10:29
  chain-of-thought
• 10:29
  quality:
• 10:30
  "What's
• 10:30
  impressive
• 10:30
  isn't
• 10:31
  that
• 10:31
  it
• 10:31
  solved
• 10:31
  them,
• 10:31
  it's
• 10:31
  that
• 10:32
  it
• 10:32
  had
• 10:32
  a
• 10:32
  coherent
• 10:32
  chain
• 10:33
  in
• 10:33
  solving
• 10:33
  them
• 10:33
  and
• 10:33
  didn't
• 10:34
  just
• 10:34
  throw
• 10:34
  stuff
• 10:34
  at
• 10:34
  the
• 10:34
  wall."
• 10:35
  The
• 10:35
  skeptics
• 10:36
  countered:
• 10:36
  "A
• 10:37
  3B
• 10:37
  crushing
• 10:37
  MathQA
• 10:38
  is
• 10:38
  narrow
• 10:38
  RL
• 10:39
  on
• 10:39
  that
• 10:39
  task
• 10:39
  shape,
• 10:40
  not
• 10:40
  general
• 10:40
  smarts."
• 10:41
  Both
• 10:42
  are
• 10:42
  right.
• 10:43
  The
• 10:43
  floor
• 10:43
  for
• 10:43
  specialized
• 10:44
  math
• 10:44
  performance
• 10:44
  keeps
• 10:44
  dropping
• 10:45
  —
• 10:45
  from
• 10:46
  PaCoRe's
• 10:46
  8B
• 10:46
  beating
• 10:47
  GPT-5
• 10:47
  on
• 10:47
  HMMT25
• 10:49
  in
• 10:49
  December
• 10:49
  to
• 10:49
  3B
• 10:50
  now.
• 10:51
  But
• 10:51
  no
• 10:51
  tool
• 10:51
  calling
• 10:51
  means
• 10:52
  no
• 10:52
  agentic
• 10:52
  coding,
• 10:53
  which
• 10:53
  limits
• 10:53
  the
• 10:53
  practical
• 10:54
  surface
• 10:54
  area.
• 10:55
  Reasoning
• 10:55
  ability
• 10:55
  compresses
• 10:56
  better
• 10:56
  than
• 10:56
  world
• 10:56
  knowledge.
• 10:58
  That
• 10:58
  is
• 10:58
  a
• 10:58
  real
• 10:58
  finding,
• 10:59
  even
• 10:59
  if
• 11:00
  the
• 11:00
  model
• 11:00
  is
• 11:00
  a
• 11:00
  sharp
• 11:00
  comb
• 11:00
  with
• 11:01
  few
• 11:01
  teeth.
• 11:02
  A
• 11:02
  LocalLLaMA
• 11:02
  thread
• 11:03
  proposed
• 11:03
  building
• 11:03
  a
• 11:03
  community
• 11:04
  model
• 11:04
  via
• 11:04
  Branch-Train-Stitch:
• 11:06
  distribute
• 11:06
  a
• 11:06
  prototype
• 11:06
  dense
• 11:07
  model
• 11:07
  to
• 11:07
  volunteers,
• 11:08
  each
• 11:08
  trains
• 11:09
  on
• 11:09
  their
• 11:09
  own
• 11:09
  hardware
• 11:09
  for
• 11:10
  a
• 11:10
  narrow
• 11:10
  domain,
• 11:11
  then
• 11:11
  organizers
• 11:11
  stitch
• 11:12
  the
• 11:12
  submodels
• 11:12
  into
• 11:12
  a
• 11:12
  large
• 11:13
  MoE.
• 11:14
  The
• 11:14
  logistics
• 11:14
  are
• 11:14
  nontrivial
• 11:15
  —
• 11:16
  sabotage
• 11:16
  detection
• 11:17
  requires
• 11:17
  non-uniform
• 11:18
  utilization
• 11:18
  in
• 11:18
  the
• 11:18
  router,
• 11:19
  final
• 11:19
  healing
• 11:20
  and
• 11:20
  RL
• 11:20
  need
• 11:20
  full-model
• 11:21
  VRAM
• 11:21
  (H200-class
• 11:22
  rigs),
• 11:23
  and
• 11:23
  the
• 11:23
  data
• 11:23
  recipe
• 11:24
  is
• 11:24
  the
• 11:24
  hardest
• 11:24
  unsolved
• 11:25
  piece.
• 11:26
  But
• 11:26
  the
• 11:26
  architecture
• 11:26
  is
• 11:26
  sound:
• 11:27
  independently
• 11:28
  trained
• 11:28
  experts,
• 11:28
  stitched
• 11:29
  at
• 11:29
  merge
• 11:29
  time,
• 11:30
  lowering
• 11:30
  coordination
• 11:31
  cost
• 11:31
  below
• 11:31
  what
• 11:31
  distributed
• 11:32
  training
• 11:32
  normally
• 11:33
  demands.
• 11:33
  A
• 11:33
  2B
• 11:34
  prototype
• 11:34
  would
• 11:35
  maximize
• 11:35
  participation;
• 11:36
  a
• 11:36
  7B
• 11:37
  prototype
• 11:37
  would
• 11:37
  produce
• 11:38
  a
• 11:38
  better
• 11:38
  final
• 11:38
  model
• 11:39
  but
• 11:39
  limit
• 11:39
  contributors
• 11:40
  to
• 11:40
  32GB
• 11:41
  VRAM
• 11:41
  and
• 11:41
  above.
• 11:42
  Whether
• 11:42
  the
• 11:42
  community
• 11:43
  actually
• 11:43
  executes
• 11:44
  is
• 11:44
  TBD,
• 11:45
  but
• 11:45
  the
• 11:45
  technical
• 11:45
  path
• 11:46
  is
• 11:46
  real.
• 11:47
  Bartowski
• 11:47
  dropped
• 11:48
  GGUF
• 11:48
  quantizations
• 11:49
  of
• 11:49
  Cohere's
• 11:49
  Command-A-Plus,
• 11:50
  making
• 11:51
  the
• 11:51
  enterprise-RAG-optimized
• 11:52
  model
• 11:52
  available
• 11:52
  for
• 11:53
  local
• 11:53
  inference.
• 11:54
  Early
• 11:54
  reports
• 11:55
  were
• 11:55
  mixed
• 11:55
  —
• 11:56
  some
• 11:56
  users
• 11:56
  hit
• 11:57
  incoherent
• 11:57
  outputs
• 11:58
  and
• 11:58
  infinite
• 11:58
  thinking
• 11:59
  loops,
• 11:59
  suggesting
• 12:00
  llama.cpp
• 12:01
  template
• 12:02
  support
• 12:02
  may
• 12:02
  not
• 12:02
  be
• 12:03
  fully
• 12:03
  baked.
• 12:04
  At
• 12:04
  25B
• 12:05
  active
• 12:05
  parameters
• 12:05
  in
• 12:06
  an
• 12:06
  MoE
• 12:06
  configuration,
• 12:07
  the
• 12:07
  speed-quality
• 12:08
  tradeoff
• 12:08
  is
• 12:08
  aggressive.
• 12:09
  The
• 12:09
  128
• 12:10
  GB
• 12:11
  unified
• 12:11
  RAM
• 12:11
  crowd
• 12:12
  still
• 12:12
  cannot
• 12:12
  fit
• 12:12
  it
• 12:13
  comfortably.
• 12:13
  The
• 12:14
  quantization
• 12:14
  pipeline
• 12:15
  grinds
• 12:15
  on
• 12:15
  —
• 12:16
  everything
• 12:16
  gets
• 12:16
  GGUF'd
• 12:17
  eventually
• 12:17
  —
• 12:18
  but
• 12:18
  surviving
• 12:19
  quantization
• 12:19
  well
• 12:19
  remains
• 12:20
  model-dependent.
• 12:21
  NVIDIA
• 12:22
  published
• 12:22
  Nemotron-Labs-Diffusion-14B
• 12:24
  on
• 12:24
  Hugging
• 12:24
  Face,
• 12:25
  their
• 12:25
  entry
• 12:26
  into
• 12:26
  diffusion-based
• 12:27
  language
• 12:27
  modeling.
• 12:28
  At
• 12:28
  14B
• 12:29
  parameters
• 12:29
  this
• 12:30
  is
• 12:30
  substantially
• 12:30
  larger
• 12:31
  than
• 12:31
  prior
• 12:31
  attempts
• 12:31
  like
• 12:32
  Tencent's
• 12:32
  WeDLM
• 12:32
  8B,
• 12:33
  which
• 12:34
  we
• 12:34
  last
• 12:34
  covered
• 12:34
  with
• 12:35
  skepticism
• 12:35
  in
• 12:35
  December
• 12:36
  2025.
• 12:37
  Whether
• 12:38
  NVIDIA's
• 12:38
  engineering
• 12:39
  muscle
• 12:39
  changes
• 12:40
  the
• 12:40
  viability
• 12:40
  picture
• 12:41
  for
• 12:41
  diffusion-based
• 12:42
  text
• 12:42
  generation
• 12:42
  remains
• 12:43
  to
• 12:43
  be
• 12:43
  seen
• 12:43
  from
• 12:43
  independent
• 12:44
  benchmarks.
• 12:47
  npm
• 12:47
  v12
• 12:48
  is
• 12:48
  shipping
• 12:49
  with
• 12:49
  post-install
• 12:49
  scripts
• 12:50
  disabled
• 12:50
  by
• 12:50
  default.
• 12:51
  This
• 12:52
  is
• 12:52
  genuinely
• 12:52
  significant.
• 12:53
  The
• 12:54
  Shy
• 12:54
  Hallude
• 12:54
  worm
• 12:54
  has
• 12:55
  been
• 12:55
  rampaging
• 12:55
  the
• 12:55
  npm
• 12:56
  ecosystem
• 12:57
  since
• 12:57
  late
• 12:57
  2025,
• 12:58
  propagating
• 12:59
  via
• 12:59
  transitive
• 13:00
  post-install
• 13:00
  scripts
• 13:01
  that
• 13:01
  steal
• 13:02
  AWS,
• 13:02
  GCP,
• 13:03
  and
• 13:03
  Azure
• 13:03
  tokens,
• 13:04
  then
• 13:04
  use
• 13:04
  compromised
• 13:05
  npm
• 13:05
  credentials
• 13:06
  to
• 13:06
  publish
• 13:06
  more
• 13:07
  malicious
• 13:07
  packages.
• 13:08
  The
• 13:08
  attack
• 13:09
  vector
• 13:09
  is
• 13:09
  brutally
• 13:09
  simple:
• 13:10
  you
• 13:10
  run
• 13:10
  npm
• 13:11
  install
• 13:11
  axios,
• 13:12
  and
• 13:12
  if
• 13:13
  any
• 13:13
  package
• 13:13
  in
• 13:13
  the
• 13:13
  entire
• 13:14
  dependency
• 13:14
  tree
• 13:15
  has
• 13:15
  a
• 13:15
  malicious
• 13:15
  post-install
• 13:16
  script,
• 13:16
  it
• 13:17
  executes
• 13:17
  on
• 13:17
  your
• 13:17
  machine.
• 13:18
  Three
• 13:19
  new
• 13:19
  defaults
• 13:19
  address
• 13:19
  this:
• 13:20
  allow-scripts
• 13:21
  off
• 13:21
  (no
• 13:21
  pre/post-install
• 13:22
  execution),
• 13:23
  allow-git
• 13:23
  none
• 13:24
  (no
• 13:24
  malicious
• 13:24
  .npmrc
• 13:25
  overrides),
• 13:26
  and
• 13:26
  allow-remote
• 13:26
  restricted
• 13:27
  (no
• 13:27
  dependency
• 13:27
  resolution
• 13:28
  from
• 13:28
  remote
• 13:28
  URLs).
• 13:29
  A
• 13:29
  new
• 13:30
  minimum-release-age
• 13:30
  setting
• 13:31
  lets
• 13:31
  you
• 13:31
  refuse
• 13:32
  packages
• 13:32
  younger
• 13:32
  than
• 13:32
  N
• 13:33
  days,
• 13:33
  which
• 13:33
  would
• 13:33
  have
• 13:34
  caught
• 13:34
  most
• 13:34
  Shy
• 13:34
  Hallude
• 13:35
  variants
• 13:35
  that
• 13:35
  were
• 13:35
  detected
• 13:35
  within
• 13:36
  hours
• 13:36
  to
• 13:36
  days.
• 13:37
  The
• 13:37
  CI/CD
• 13:38
  angle
• 13:38
  is
• 13:38
  the
• 13:38
  critical
• 13:39
  one
• 13:39
  —
• 13:39
  automated
• 13:40
  pipelines
• 13:40
  that
• 13:40
  run
• 13:41
  npm
• 13:41
  install
• 13:41
  with
• 13:41
  post-install
• 13:42
  scripts
• 13:42
  enabled
• 13:43
  are
• 13:43
  handing
• 13:43
  attacker
• 13:44
  code
• 13:44
  execution
• 13:44
  in
• 13:45
  environments
• 13:45
  with
• 13:45
  GitHub
• 13:45
  PATs
• 13:46
  and
• 13:46
  deployment
• 13:47
  credentials.
• 13:48
  This
• 13:48
  fix
• 13:48
  removes
• 13:49
  the
• 13:49
  highway.
• 13:50
  Credit
• 13:50
  to
• 13:50
  @Koba,
• 13:51
  who
• 13:51
  opened
• 13:51
  an
• 13:52
  issue
• 13:52
  about
• 13:52
  this
• 13:52
  in
• 13:52
  2019
• 13:53
  and
• 13:53
  was
• 13:54
  ignored
• 13:54
  for
• 13:54
  years.
• 13:55
  On
• 13:55
  the
• 13:55
  offensive
• 13:56
  side,
• 13:56
  OpenAnt
• 13:57
  (Knostic,
• 13:57
  Apache
• 13:58
  2.0)
• 13:58
  published
• 13:59
  a
• 13:59
  full
• 13:59
  paper
• 13:59
  on
• 14:00
  their
• 14:00
  six-stage
• 14:00
  LLM-powered
• 14:01
  vulnerability
• 14:02
  discovery
• 14:02
  pipeline.
• 14:03
  The
• 14:03
  architecture:
• 14:04
  static
• 14:05
  filtering
• 14:05
  reduces
• 14:06
  analysis
• 14:06
  surface
• 14:06
  by
• 14:07
  97%
• 14:07
  via
• 14:08
  reachability
• 14:08
  from
• 14:08
  external
• 14:09
  entry
• 14:09
  points;
• 14:10
  Claude
• 14:10
  Sonnet
• 14:11
  classifies
• 14:11
  exposure;
• 14:12
  Claude
• 14:13
  Opus
• 14:13
  detects
• 14:13
  vulnerabilities
• 14:14
  with
• 14:14
  a
• 14:14
  language-agnostic
• 14:15
  prompt;
• 14:16
  adversarial
• 14:17
  verification
• 14:17
  simulates
• 14:18
  a
• 14:18
  constrained
• 14:18
  attacker;
• 14:19
  dynamic
• 14:20
  verification
• 14:20
  generates
• 14:21
  exploit
• 14:21
  environments
• 14:22
  in
• 14:22
  sandboxed
• 14:22
  Docker
• 14:23
  containers
• 14:23
  that
• 14:23
  are
• 14:24
  discarded
• 14:24
  after
• 14:24
  use.
• 14:25
  Results
• 14:25
  across
• 14:26
  OpenSSL,
• 14:27
  WordPress,
• 14:27
  and
• 14:27
  Flowise:
• 14:28
  64,132
• 14:30
  functions
• 14:31
  filtered
• 14:31
  to
• 14:31
  190
• 14:32
  confirmed
• 14:33
  vulnerabilities,
• 14:34
  144
• 14:35
  dynamically
• 14:36
  verified
• 14:36
  (75.8%
• 14:38
  confirmation
• 14:38
  rate).
• 14:39
  Cost:
• 14:40
  $1,461
• 14:42
  for
• 14:42
  eight
• 14:42
  repositories.
• 14:43
  Over
• 14:44
  30
• 14:44
  vulnerability
• 14:45
  classes
• 14:45
  identified.
• 14:46
  The
• 14:46
  paper
• 14:46
  advances
• 14:47
  the
• 14:47
  field
• 14:47
  past
• 14:48
  the
• 14:48
  "Claude
• 14:48
  found
• 14:48
  some
• 14:49
  bugs"
• 14:49
  stage
• 14:49
  into
• 14:50
  a
• 14:50
  reproducible
• 14:50
  methodology
• 14:51
  with
• 14:51
  cost
• 14:51
  accounting
• 14:52
  and
• 14:52
  false-positive
• 14:53
  reduction.
• 14:54
  The
• 14:54
  closed-loop
• 14:55
  design
• 14:55
  —
• 14:56
  semantic
• 14:56
  reasoning
• 14:57
  followed
• 14:57
  by
• 14:57
  actual
• 14:58
  exploit
• 14:58
  execution
• 14:58
  —
• 14:59
  is
• 15:00
  the
• 15:00
  right
• 15:00
  architecture
• 15:00
  for
• 15:01
  scaling
• 15:01
  this.
• 15:05
  A
• 15:05
  new
• 15:05
  paper
• 15:06
  introduces
• 15:06
  "concept
• 15:07
  frustration"
• 15:07
  —
• 15:08
  a
• 15:08
  geometric
• 15:08
  framework
• 15:09
  for
• 15:09
  diagnosing
• 15:10
  when
• 15:10
  a
• 15:10
  model's
• 15:10
  concept
• 15:11
  ontology
• 15:11
  is
• 15:11
  incomplete.
• 15:13
  The
• 15:13
  core
• 15:13
  idea:
• 15:14
  when
• 15:14
  human-interpretable
• 15:15
  bottleneck
• 15:15
  concepts
• 15:16
  do
• 15:16
  not
• 15:16
  cover
• 15:16
  all
• 15:16
  the
• 15:17
  information
• 15:17
  the
• 15:17
  data
• 15:17
  actually
• 15:18
  carries,
• 15:19
  the
• 15:19
  model's
• 15:19
  internal
• 15:20
  representations
• 15:20
  become
• 15:21
  frustrated,
• 15:22
  carrying
• 15:22
  information
• 15:22
  the
• 15:22
  concept
• 15:23
  labels
• 15:23
  cannot
• 15:24
  express.
• 15:25
  The
• 15:25
  key
• 15:25
  technical
• 15:26
  contribution
• 15:26
  is
• 15:27
  using
• 15:27
  the
• 15:27
  Fisher
• 15:27
  information
• 15:28
  metric
• 15:28
  to
• 15:28
  detect
• 15:28
  this
• 15:29
  frustration
• 15:29
  in
• 15:29
  representation
• 15:30
  space
• 15:30
  —
• 15:31
  something
• 15:31
  invisible
• 15:32
  in
• 15:32
  Euclidean
• 15:32
  distance.
• 15:33
  The
• 15:33
  closed-form
• 15:34
  Bayes-optimal
• 15:34
  accuracy
• 15:35
  decomposition
• 15:35
  tells
• 15:36
  you
• 15:36
  exactly
• 15:36
  how
• 15:36
  much
• 15:36
  performance
• 15:37
  you
• 15:37
  are
• 15:37
  losing
• 15:37
  to
• 15:37
  missing
• 15:38
  concepts.
• 15:39
  Real
• 15:39
  experiments
• 15:40
  on
• 15:40
  sarcasm
• 15:40
  detection
• 15:41
  (DeBERTa)
• 15:41
  and
• 15:41
  bird
• 15:42
  classification
• 15:42
  (CLIP)
• 15:43
  showed
• 15:43
  that
• 15:43
  adding
• 15:43
  a
• 15:43
  missing
• 15:44
  concept
• 15:44
  reorganizes
• 15:45
  all
• 15:45
  existing
• 15:45
  concept
• 15:46
  representations,
• 15:47
  not
• 15:47
  just
• 15:47
  the
• 15:47
  new
• 15:47
  one.
• 15:48
  This
• 15:48
  has
• 15:49
  direct
• 15:49
  implications
• 15:49
  for
• 15:49
  interpretable
• 15:50
  AI:
• 15:51
  if
• 15:51
  your
• 15:51
  concept
• 15:51
  bottleneck
• 15:52
  model
• 15:52
  is
• 15:52
  underperforming,
• 15:53
  the
• 15:53
  bottleneck
• 15:54
  itself
• 15:54
  may
• 15:54
  be
• 15:54
  the
• 15:54
  problem,
• 15:55
  and
• 15:55
  the
• 15:55
  failure
• 15:56
  mode
• 15:56
  is
• 15:56
  geometrically
• 15:57
  detectable
• 15:57
  before
• 15:57
  you
• 15:58
  know
• 15:58
  what
• 15:58
  concept
• 15:58
  is
• 15:58
  missing.
• 15:59
  The
• 15:59
  Artificial
• 16:00
  Tripartite
• 16:00
  Intelligence
• 16:01
  paper
• 16:01
  proposes
• 16:02
  a
• 16:02
  bio-inspired,
• 16:03
  sensor-first
• 16:03
  architecture
• 16:04
  for
• 16:04
  physical
• 16:04
  AI
• 16:04
  that
• 16:05
  inverts
• 16:05
  the
• 16:05
  usual
• 16:05
  hierarchy.
• 16:06
  Three
• 16:07
  tiers:
• 16:07
  a
• 16:08
  brainstem
• 16:08
  L1
• 16:08
  layer
• 16:09
  handling
• 16:09
  reflexive
• 16:09
  responses
• 16:10
  in
• 16:10
  under
• 16:10
  10
• 16:11
  milliseconds,
• 16:12
  a
• 16:12
  cerebellum
• 16:12
  L2
• 16:13
  layer
• 16:13
  doing
• 16:13
  calibration
• 16:14
  via
• 16:14
  contextual
• 16:15
  bandits,
• 16:15
  and
• 16:16
  cerebral
• 16:16
  L3/L4
• 16:19
  layers
• 16:19
  for
• 16:19
  local
• 16:19
  and
• 16:20
  cloud
• 16:20
  inference.
• 16:21
  A
• 16:21
  Galaxy
• 16:22
  S25
• 16:22
  prototype
• 16:23
  went
• 16:23
  from
• 16:23
  50%
• 16:24
  to
• 16:24
  88%
• 16:24
  accuracy
• 16:25
  with
• 16:25
  43.3%
• 16:26
  fewer
• 16:27
  cloud
• 16:27
  calls.
• 16:28
  The
• 16:28
  architecture
• 16:29
  is
• 16:29
  a
• 16:29
  direct
• 16:29
  response
• 16:29
  to
• 16:29
  a
• 16:30
  problem
• 16:30
  we
• 16:30
  have
• 16:30
  tracked:
• 16:31
  language
• 16:31
  abstracts
• 16:32
  away
• 16:32
  forces,
• 16:33
  friction,
• 16:33
  and
• 16:33
  timing,
• 16:34
  making
• 16:34
  LLM-centric
• 16:35
  physical
• 16:35
  AI
• 16:35
  brittle.
• 16:36
  By
• 16:36
  prioritizing
• 16:37
  raw
• 16:37
  sensory
• 16:38
  processing
• 16:39
  over
• 16:39
  language-mediated
• 16:40
  reasoning,
• 16:40
  the
• 16:41
  tripartite
• 16:41
  design
• 16:41
  sidesteps
• 16:42
  the
• 16:42
  embodiment-induced
• 16:43
  mismatches
• 16:44
  that
• 16:44
  make
• 16:44
  current
• 16:44
  robotics
• 16:45
  systems
• 16:45
  fail
• 16:46
  in
• 16:46
  ways
• 16:46
  their
• 16:46
  language
• 16:46
  models
• 16:47
  cannot
• 16:47
  predict.
• 16:48
  Accepted
• 16:49
  at
• 16:49
  MobiSys
• 16:49
  2026.
• 16:51
  Classifier-free
• 16:52
  guidance
• 16:52
  (CFG)
• 16:52
  has
• 16:53
  been
• 16:53
  the
• 16:53
  workhorse
• 16:53
  trick
• 16:53
  for
• 16:54
  improving
• 16:54
  diffusion
• 16:54
  model
• 16:55
  outputs,
• 16:56
  but
• 16:56
  a
• 16:56
  new
• 16:56
  paper
• 16:56
  proves
• 16:57
  it
• 16:57
  breaks
• 16:57
  probability
• 16:57
  conservation
• 16:58
  in
• 16:58
  flow-matching
• 16:59
  models.
• 17:00
  The
• 17:00
  analysis
• 17:01
  decomposes
• 17:01
  the
• 17:01
  damage
• 17:02
  into
• 17:02
  two
• 17:02
  terms:
• 17:03
  divergence
• 17:03
  that
• 17:03
  blows
• 17:04
  up
• 17:04
  structurally
• 17:04
  near
• 17:04
  the
• 17:05
  data
• 17:05
  manifold,
• 17:06
  and
• 17:06
  a
• 17:06
  score-parallel
• 17:06
  flux
• 17:07
  that
• 17:07
  drives
• 17:07
  the
• 17:07
  saturation
• 17:08
  artifacts
• 17:08
  everyone
• 17:09
  has
• 17:09
  learned
• 17:09
  to
• 17:09
  live
• 17:09
  with.
• 17:10
  AdaMaG
• 17:11
  attenuates
• 17:11
  the
• 17:11
  parallel
• 17:12
  component
• 17:12
  with
• 17:12
  a
• 17:12
  time-dependent
• 17:13
  schedule,
• 17:14
  conserving
• 17:14
  probability
• 17:15
  without
• 17:15
  adding
• 17:16
  any
• 17:16
  inference
• 17:16
  cost.
• 17:17
  On
• 17:17
  SD3,
• 17:19
  SD3.5,
• 17:20
  and
• 17:20
  Flux,
• 17:21
  it
• 17:21
  beats
• 17:22
  CFG,
• 17:22
  APG,
• 17:23
  and
• 17:23
  TAG
• 17:24
  across
• 17:24
  metrics.
• 17:26
  Human
• 17:26
  evaluation:
• 17:27
  60%
• 17:28
  preferred
• 17:28
  over
• 17:28
  standard
• 17:29
  CFG.
• 17:30
  The
• 17:30
  practical
• 17:30
  implication
• 17:31
  is
• 17:31
  immediate
• 17:31
  —
• 17:32
  if
• 17:32
  you
• 17:33
  are
• 17:33
  generating
• 17:33
  images
• 17:33
  with
• 17:34
  flow-matching
• 17:34
  models
• 17:35
  and
• 17:35
  tolerating
• 17:35
  saturation
• 17:36
  artifacts
• 17:36
  as
• 17:36
  "the
• 17:36
  cost
• 17:37
  of
• 17:37
  guidance,"
• 17:38
  there
• 17:38
  is
• 17:38
  now
• 17:38
  a
• 17:38
  zero-cost
• 17:39
  fix
• 17:39
  with
• 17:39
  a
• 17:39
  theoretical
• 17:40
  justification
• 17:41
  for
• 17:41
  why
• 17:41
  it
• 17:41
  works.