Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 02, 2026

Transcript

AGI Dreams Podcast – April 02, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Anthropic's
• 0:00
  source
• 0:00
  map
• 0:01
  fiasco
• 0:01
  dropped
• 0:01
  the
• 0:02
  full
• 0:02
  Claude
• 0:02
  Code
• 0:02
  TypeScript
• 0:03
  source
• 0:03
  onto
• 0:03
  npm
• 0:04
  for
• 0:04
  the
• 0:04
  world
• 0:04
  to
• 0:04
  mirror,
• 0:05
  and
• 0:05
  the
• 0:05
  learnings
• 0:06
  are
• 0:06
  still
• 0:06
  expanding.
• 0:07
  Two
• 0:07
  detailed
• 0:08
  teardowns
• 0:08
  —
• 0:09
  one
• 0:09
  by
• 0:09
  Alex
• 0:10
  Kim,
• 0:10
  one
• 0:10
  by
• 0:10
  Kuber
• 0:11
  Mehta
• 0:11
  —
• 0:12
  have
• 0:12
  cataloged
• 0:12
  interesting
• 0:13
  artifacts
• 0:13
  from
• 0:13
  the
• 0:14
  900-file,
• 0:15
  512,000-line
• 0:17
  bundle.
• 0:18
  We
• 0:18
  now
• 0:18
  have
• 0:18
  sketches
• 0:19
  of
• 0:19
  their
• 0:19
  unpublished
• 0:19
  product
• 0:20
  roadmap:
• 0:20
  KAIROS,
• 0:23
  an
• 0:23
  unreleased
• 0:23
  always-on
• 0:24
  autonomous
• 0:24
  agent
• 0:25
  mode
• 0:25
  with
• 0:25
  5-minute
• 0:25
  refresh
• 0:26
  cycles
• 0:26
  and
• 0:26
  push
• 0:26
  notifications;
• 0:28
  an
• 0:28
  anti-distillation
• 0:29
  system
• 0:29
  that
• 0:29
  injects
• 0:30
  fake
• 0:30
  tool
• 0:30
  definitions
• 0:31
  into
• 0:31
  API
• 0:31
  requests
• 0:32
  to
• 0:32
  poison
• 0:32
  competitor
• 0:32
  training
• 0:33
  data;
• 0:33
  "Undercover
• 0:34
  Mode,"
• 0:35
  which
• 0:35
  tells
• 0:35
  Claude
• 0:35
  to
• 0:36
  hide
• 0:36
  the
• 0:36
  fact
• 0:36
  it's
• 0:36
  an
• 0:36
  AI
• 0:37
  when
• 0:37
  Anthropic
• 0:37
  employees
• 0:38
  contribute
• 0:38
  to
• 0:38
  open-source
• 0:39
  repos;
• 0:39
  and
• 0:39
  native
• 0:40
  client
• 0:40
  attestation
• 0:41
  that
• 0:41
  uses
• 0:41
  Zig-level
• 0:42
  hash
• 0:42
  replacement
• 0:42
  below
• 0:43
  the
• 0:43
  JavaScript
• 0:43
  runtime
• 0:43
  —
• 0:44
  essentially
• 0:45
  DRM
• 0:45
  for
• 0:45
  API
• 0:45
  calls,
• 0:46
  and
• 0:46
  the
• 0:46
  technical
• 0:47
  enforcement
• 0:47
  behind
• 0:47
  Anthropic's
• 0:48
  legal
• 0:48
  fight
• 0:48
  with
• 0:48
  OpenCode.
• 0:49
  The
• 0:50
  anti-distillation
• 0:50
  mechanism
• 0:51
  sounds
• 0:51
  impressive
• 0:51
  until
• 0:52
  you
• 0:52
  read
• 0:52
  the
• 0:52
  activation
• 0:52
  logic:
• 0:53
  a
• 0:53
  MITM
• 0:54
  proxy
• 0:54
  stripping
• 0:55
  one
• 0:55
  field
• 0:55
  from
• 0:55
  request
• 0:55
  bodies
• 0:56
  would
• 0:56
  bypass
• 0:57
  it
• 0:57
  entirely,
• 0:58
  and
• 0:58
  setting
• 0:58
  a
• 0:58
  single
• 0:58
  environment
• 0:59
  variable
• 0:59
  disables
• 1:00
  the
• 1:00
  whole
• 1:00
  thing.
• 1:01
  The
• 1:01
  real
• 1:01
  protection,
• 1:01
  as
• 1:02
  Kim
• 1:02
  notes,
• 1:02
  was
• 1:03
  always
• 1:03
  legal
• 1:03
  rather
• 1:03
  than
• 1:04
  technical.
• 1:05
  The
• 1:05
  client
• 1:05
  attestation
• 1:06
  is
• 1:06
  similarly
• 1:07
  fragile
• 1:07
  —
• 1:08
  it
• 1:08
  only
• 1:08
  works
• 1:08
  inside
• 1:08
  the
• 1:09
  official
• 1:09
  Bun
• 1:09
  binary,
• 1:10
  and
• 1:10
  a
• 1:10
  server-side
• 1:11
  function
• 1:11
  that
• 1:11
  "tolerates
• 1:12
  unknown
• 1:12
  extra
• 1:12
  fields"
• 1:13
  suggests
• 1:13
  the
• 1:13
  validation
• 1:14
  may
• 1:14
  be
• 1:14
  more
• 1:14
  forgiving
• 1:15
  than
• 1:15
  a
• 1:15
  DRM
• 1:16
  system
• 1:16
  should
• 1:16
  be.
• 1:17
  The
• 1:17
  250,000
• 1:18
  wasted
• 1:19
  API
• 1:19
  calls
• 1:19
  per
• 1:19
  day
• 1:20
  from
• 1:20
  a
• 1:20
  compaction
• 1:20
  bug,
• 1:21
  fixed
• 1:21
  with
• 1:21
  three
• 1:21
  lines
• 1:22
  of
• 1:22
  code,
• 1:23
  is
• 1:23
  the
• 1:23
  kind
• 1:23
  of
• 1:23
  detail
• 1:24
  that
• 1:24
  makes
• 1:24
  the
• 1:24
  Tamagotchi
• 1:25
  companion
• 1:25
  pet
• 1:25
  system
• 1:26
  ("Buddy,"
• 1:27
  with
• 1:27
  18
• 1:27
  species,
• 1:28
  gacha
• 1:28
  rarity
• 1:29
  tiers,
• 1:30
  and
• 1:30
  RPG
• 1:30
  stats
• 1:31
  like
• 1:31
  DEBUGGING
• 1:32
  and
• 1:32
  SNARK)
• 1:32
  feel
• 1:32
  less
• 1:32
  whimsical
• 1:33
  and
• 1:33
  more
• 1:34
  indicative
• 1:34
  of
• 1:34
  a
• 1:34
  codebase
• 1:35
  growing
• 1:35
  faster
• 1:35
  than
• 1:35
  its
• 1:36
  review
• 1:36
  processes.
• 1:37
  The
• 1:37
  leak's
• 1:38
  root
• 1:38
  cause
• 1:38
  —
• 1:39
  a
• 1:39
  known
• 1:39
  Bun
• 1:40
  bug
• 1:40
  shipping
• 1:40
  source
• 1:40
  maps
• 1:41
  in
• 1:41
  production
• 1:41
  mode,
• 1:42
  filed
• 1:42
  on
• 1:42
  March
• 1:43
  11
• 1:43
  and
• 1:43
  still
• 1:43
  open
• 1:43
  —
• 1:44
  is
• 1:45
  the
• 1:45
  kind
• 1:45
  of
• 1:45
  irony
• 1:45
  that
• 1:46
  writes
• 1:46
  itself
• 1:46
  when
• 1:46
  your
• 1:46
  toolchain
• 1:47
  was
• 1:47
  partly
• 1:47
  built
• 1:48
  by
• 1:48
  your
• 1:48
  own
• 1:48
  AI.
• 1:49
  What
• 1:50
  happened
• 1:50
  next
• 1:50
  is
• 1:50
  arguably
• 1:51
  more
• 1:51
  consequential
• 1:52
  than
• 1:52
  the
• 1:52
  leak
• 1:52
  itself.
• 1:53
  Within
• 1:53
  hours,
• 1:54
  a
• 1:54
  clean-room
• 1:54
  Rust
• 1:55
  reimplementation
• 1:55
  appeared
• 1:56
  —
• 1:56
  spec
• 1:57
  first,
• 1:57
  implementation
• 1:58
  second,
• 1:59
  explicitly
• 1:59
  modeled
• 2:00
  on
• 2:00
  the
• 2:00
  Phoenix
• 2:00
  v.
• 2:01
  IBM
• 2:02
  BIOS
• 2:02
  precedent.
• 2:03
  Meanwhile,
• 2:04
  Dragos
• 2:04
  Ruiu
• 2:05
  posted
• 2:05
  a
• 2:05
  detailed
• 2:06
  walkthrough
• 2:06
  of
• 2:06
  Claude
• 2:07
  building
• 2:07
  a
• 2:07
  complete
• 2:08
  FreeBSD
• 2:08
  remote
• 2:09
  kernel
• 2:09
  RCE
• 2:10
  (CVE-2026-4747),
• 2:15
  from
• 2:15
  advisory
• 2:16
  to
• 2:16
  root
• 2:16
  shell
• 2:16
  in
• 2:16
  about
• 2:17
  four
• 2:17
  hours.
• 2:18
  The
• 2:18
  target
• 2:18
  was
• 2:19
  relatively
• 2:19
  soft
• 2:19
  —
• 2:20
  FreeBSD
• 2:21
  14.x
• 2:22
  ships
• 2:22
  without
• 2:23
  KASLR,
• 2:23
  the
• 2:23
  overflowed
• 2:24
  buffer
• 2:24
  had
• 2:24
  no
• 2:25
  stack
• 2:25
  canaries
• 2:25
  —
• 2:26
  but
• 2:26
  the
• 2:26
  model
• 2:27
  solved
• 2:27
  ROP
• 2:28
  chain
• 2:28
  construction,
• 2:29
  multi-packet
• 2:30
  shellcode
• 2:30
  delivery,
• 2:31
  and
• 2:31
  kernel-to-userland
• 2:32
  transition
• 2:33
  with
• 2:33
  roughly
• 2:33
  44
• 2:34
  human
• 2:34
  redirections
• 2:35
  across
• 2:35
  the
• 2:35
  session.
• 2:36
  Strip
• 2:37
  away
• 2:37
  those
• 2:37
  steering
• 2:37
  prompts
• 2:38
  and
• 2:38
  the
• 2:38
  model
• 2:38
  doesn't
• 2:39
  get
• 2:39
  there
• 2:39
  on
• 2:39
  its
• 2:39
  own.
• 2:40
  Not
• 2:40
  yet.
• 2:41
  But
• 2:42
  the
• 2:42
  gap
• 2:42
  between
• 2:42
  "controlled
• 2:43
  demo
• 2:43
  on
• 2:43
  a
• 2:43
  soft
• 2:44
  target"
• 2:44
  and
• 2:45
  "reliable
• 2:45
  autonomous
• 2:46
  exploitation"
• 2:47
  is
• 2:47
  closing
• 2:48
  faster
• 2:48
  than
• 2:48
  most
• 2:49
  defenders
• 2:49
  are
• 2:49
  calibrating
• 2:50
  for.
• 2:52
  The
• 2:53
  most
• 2:53
  thought-provoking
• 2:54
  piece
• 2:54
  this
• 2:54
  week
• 2:54
  argues
• 2:55
  that
• 2:55
  AI
• 2:55
  coding
• 2:56
  agents
• 2:56
  are
• 2:56
  about
• 2:56
  to
• 2:56
  make
• 2:56
  free
• 2:57
  software
• 2:57
  matter
• 2:57
  again
• 2:58
  —
• 2:58
  not
• 2:59
  open
• 2:59
  source
• 2:59
  in
• 2:59
  the
• 2:59
  corporate-friendly
• 3:00
  sense,
• 3:00
  but
• 3:01
  Stallman's
• 3:01
  original
• 3:02
  four
• 3:02
  freedoms.
• 3:02
  The
• 3:03
  reasoning
• 3:03
  is
• 3:03
  straightforward:
• 3:04
  when
• 3:04
  an
• 3:04
  agent
• 3:04
  can
• 3:05
  read,
• 3:05
  understand,
• 3:06
  and
• 3:06
  modify
• 3:07
  source
• 3:07
  code
• 3:07
  on
• 3:08
  your
• 3:08
  behalf,
• 3:09
  access
• 3:09
  to
• 3:09
  that
• 3:09
  source
• 3:10
  stops
• 3:10
  being
• 3:10
  a
• 3:10
  symbolic
• 3:11
  right
• 3:11
  for
• 3:11
  programmers
• 3:12
  and
• 3:12
  becomes
• 3:13
  a
• 3:13
  practical
• 3:13
  capability
• 3:14
  for
• 3:14
  everyone.
• 3:15
  The
• 3:15
  author
• 3:15
  tested
• 3:16
  this
• 3:16
  by
• 3:16
  trying
• 3:16
  to
• 3:16
  build
• 3:17
  a
• 3:17
  simple
• 3:17
  tweet-to-task
• 3:18
  workflow
• 3:18
  using
• 3:18
  Sunsama,
• 3:19
  a
• 3:19
  closed
• 3:20
  SaaS
• 3:20
  product.
• 3:21
  The
• 3:21
  result:
• 3:21
  six
• 3:22
  layers
• 3:22
  of
• 3:22
  workarounds,
• 3:23
  three
• 3:23
  authentication
• 3:24
  mechanisms,
• 3:25
  a
• 3:25
  dependency
• 3:26
  on
• 3:26
  a
• 3:26
  stranger's
• 3:26
  reverse-engineered
• 3:27
  API,
• 3:28
  and
• 3:28
  an
• 3:28
  iOS
• 3:29
  Shortcut
• 3:29
  he
• 3:29
  had
• 3:29
  to
• 3:29
  build
• 3:29
  by
• 3:29
  hand
• 3:30
  because
• 3:30
  Apple
• 3:30
  provides
• 3:31
  no
• 3:31
  programmatic
• 3:31
  way
• 3:32
  to
• 3:32
  create
• 3:32
  one.
• 3:33
  With
• 3:33
  free
• 3:33
  software,
• 3:34
  the
• 3:34
  agent
• 3:35
  reads
• 3:35
  the
• 3:35
  source,
• 3:35
  understands
• 3:36
  the
• 3:36
  data
• 3:36
  model,
• 3:37
  makes
• 3:37
  the
• 3:37
  change.
• 3:38
  Ten
• 3:38
  minutes.
• 3:39
  The
• 3:39
  AGPL's
• 3:40
  limited
• 3:40
  adoption,
• 3:41
  Google's
• 3:42
  public
• 3:42
  ban
• 3:42
  on
• 3:42
  AGPL
• 3:43
  code,
• 3:44
  and
• 3:44
  the
• 3:44
  licensing
• 3:44
  chaos
• 3:45
  from
• 3:45
  MongoDB
• 3:45
  to
• 3:45
  HashiCorp
• 3:46
  to
• 3:46
  Redis
• 3:46
  all
• 3:47
  get
• 3:47
  a
• 3:47
  clear-eyed
• 3:47
  review.
• 3:48
  The
• 3:48
  counterpoint
• 3:49
  is
• 3:49
  equally
• 3:49
  honest:
• 3:50
  a
• 3:51
  CEU
• 3:51
  working
• 3:51
  paper
• 3:52
  argues
• 3:52
  vibe-coding
• 3:53
  kills
• 3:53
  open
• 3:53
  source
• 3:53
  by
• 3:54
  severing
• 3:54
  the
• 3:54
  user-maintainer
• 3:55
  feedback
• 3:55
  loop,
• 3:56
  Adam
• 3:56
  Wathan
• 3:56
  reports
• 3:57
  Tailwind
• 3:57
  documentation
• 3:58
  traffic
• 3:58
  down
• 3:58
  40%
• 3:59
  and
• 3:59
  revenue
• 3:59
  down
• 4:00
  80%,
• 4:01
  and
• 4:01
  Mitchell
• 4:01
  Hashimoto
• 4:01
  moved
• 4:02
  Ghostty
• 4:02
  to
• 4:02
  a
• 4:02
  vouch-based
• 4:03
  contribution
• 4:03
  model
• 4:03
  to
• 4:04
  stem
• 4:04
  the
• 4:04
  AI-generated
• 4:05
  PR
• 4:05
  flood.
• 4:06
  GitHub's
• 4:06
  own
• 4:07
  relationship
• 4:07
  with
• 4:07
  developers
• 4:08
  took
• 4:08
  another
• 4:08
  hit
• 4:08
  when
• 4:09
  Copilot
• 4:09
  started
• 4:09
  injecting
• 4:10
  ads
• 4:10
  —
• 4:11
  GitHub
• 4:11
  calls
• 4:11
  them
• 4:11
  "tips"
• 4:12
  —
• 4:12
  into
• 4:13
  pull
• 4:13
  requests.
• 4:14
  A
• 4:14
  developer
• 4:14
  discovered
• 4:15
  that
• 4:15
  after
• 4:15
  a
• 4:15
  coworker
• 4:16
  asked
• 4:16
  Copilot
• 4:16
  to
• 4:16
  fix
• 4:17
  a
• 4:17
  typo,
• 4:17
  the
• 4:17
  bot
• 4:18
  inserted
• 4:18
  a
• 4:18
  promotion
• 4:19
  for
• 4:19
  Raycast
• 4:19
  into
• 4:19
  his
• 4:20
  PR
• 4:20
  description.
• 4:21
  Over
• 4:21
  11,400
• 4:23
  PRs
• 4:23
  received
• 4:23
  similar
• 4:24
  injections.
• 4:25
  GitHub's
• 4:25
  VP
• 4:26
  of
• 4:26
  developer
• 4:26
  relations
• 4:27
  acknowledged
• 4:27
  that
• 4:28
  letting
• 4:28
  Copilot
• 4:28
  touch
• 4:29
  PRs
• 4:29
  it
• 4:29
  didn't
• 4:30
  create
• 4:30
  "became
• 4:30
  icky,"
• 4:31
  and
• 4:31
  the
• 4:31
  feature
• 4:32
  was
• 4:32
  killed
• 4:32
  within
• 4:32
  hours.
• 4:33
  The
• 4:33
  official
• 4:34
  line
• 4:34
  —
• 4:35
  "GitHub
• 4:35
  does
• 4:35
  not
• 4:35
  and
• 4:35
  does
• 4:36
  not
• 4:36
  plan
• 4:36
  to
• 4:36
  include
• 4:37
  advertisements"
• 4:37
  —
• 4:38
  is
• 4:38
  technically
• 4:39
  true
• 4:39
  only
• 4:39
  if
• 4:39
  you
• 4:39
  accept
• 4:40
  that
• 4:40
  a
• 4:40
  promotional
• 4:40
  message
• 4:41
  with
• 4:41
  a
• 4:41
  download
• 4:41
  link
• 4:42
  inserted
• 4:42
  into
• 4:42
  someone
• 4:43
  else's
• 4:43
  code
• 4:43
  review
• 4:44
  is
• 4:44
  a
• 4:44
  "tip"
• 4:44
  rather
• 4:45
  than
• 4:45
  an
• 4:45
  ad.
• 4:46
  The
• 4:46
  trust
• 4:46
  problem
• 4:47
  extends
• 4:47
  deeper
• 4:47
  than
• 4:47
  ads.
• 4:48
  A
• 4:49
  developer
• 4:49
  building
• 4:49
  Claude
• 4:50
  Code
• 4:50
  skills
• 4:50
  for
• 4:50
  iOS
• 4:51
  design
• 4:51
  system
• 4:51
  auditing
• 4:52
  discovered
• 4:52
  his
• 4:52
  tools
• 4:53
  missed
• 4:53
  more
• 4:53
  than
• 4:53
  half
• 4:54
  the
• 4:54
  violations
• 4:54
  in
• 4:54
  his
• 4:55
  codebase.
• 4:56
  The
• 4:56
  root
• 4:56
  cause:
• 4:57
  many
• 4:57
  icons
• 4:58
  inherited
• 4:58
  their
• 4:58
  color
• 4:59
  from
• 4:59
  the
• 4:59
  system
• 4:59
  accent
• 4:59
  —
• 5:00
  there
• 5:00
  was
• 5:00
  nothing
• 5:01
  in
• 5:01
  the
• 5:01
  code
• 5:01
  to
• 5:01
  grep
• 5:01
  for.
• 5:02
  The
• 5:02
  tool
• 5:02
  was
• 5:02
  searching
• 5:03
  for
• 5:03
  things
• 5:03
  that
• 5:03
  looked
• 5:03
  wrong
• 5:04
  but
• 5:04
  couldn't
• 5:04
  find
• 5:05
  things
• 5:05
  that
• 5:05
  looked
• 5:05
  like
• 5:05
  nothing.
• 5:06
  His
• 5:06
  fix
• 5:06
  —
• 5:06
  enumerate
• 5:07
  every
• 5:07
  candidate
• 5:08
  file,
• 5:08
  then
• 5:08
  verify
• 5:09
  each
• 5:09
  one
• 5:09
  for
• 5:09
  the
• 5:09
  correct
• 5:10
  pattern
• 5:10
  —
• 5:10
  found
• 5:11
  71
• 5:11
  violations
• 5:12
  where
• 5:12
  grep-based
• 5:12
  scanning
• 5:13
  found
• 5:13
  31.
• 5:14
  The
• 5:14
  insight
• 5:15
  generalizes:
• 5:16
  any
• 5:16
  AI
• 5:17
  audit
• 5:17
  tool
• 5:17
  that
• 5:17
  searches
• 5:18
  for
• 5:18
  anti-patterns
• 5:18
  will
• 5:18
  miss
• 5:19
  violations
• 5:19
  defined
• 5:20
  by
• 5:20
  the
• 5:20
  absence
• 5:20
  of
• 5:20
  correct
• 5:21
  patterns.
• 5:22
  Greptile's
• 5:23
  analysis
• 5:23
  pushes
• 5:24
  back
• 5:24
  on
• 5:24
  the
• 5:24
  doom
• 5:24
  narrative
• 5:24
  from
• 5:25
  the
• 5:25
  other
• 5:25
  direction:
• 5:26
  AI
• 5:26
  models
• 5:27
  will
• 5:27
  write
• 5:27
  good
• 5:27
  code
• 5:27
  because
• 5:28
  economic
• 5:28
  incentives
• 5:28
  demand
• 5:29
  it.
• 5:30
  Complex
• 5:30
  code
• 5:30
  requires
• 5:31
  more
• 5:31
  tokens
• 5:31
  and
• 5:32
  more
• 5:32
  compute;
• 5:32
  simple,
• 5:33
  maintainable
• 5:34
  code
• 5:34
  is
• 5:34
  cheaper
• 5:34
  to
• 5:34
  generate
• 5:35
  and
• 5:35
  cheaper
• 5:35
  to
• 5:35
  modify.
• 5:37
  Competition
• 5:37
  between
• 5:38
  models
• 5:38
  will
• 5:38
  select
• 5:38
  for
• 5:38
  quality.
• 5:40
  The
• 5:40
  argument
• 5:40
  is
• 5:40
  sound
• 5:41
  in
• 5:41
  theory,
• 5:41
  though
• 5:42
  the
• 5:42
  current
• 5:42
  evidence
• 5:42
  —
• 5:44
  median
• 5:44
  PR
• 5:45
  size
• 5:45
  up
• 5:45
  33%,
• 5:46
  outages
• 5:47
  steadily
• 5:47
  increasing
• 5:48
  since
• 5:48
  2022
• 5:49
  —
• 5:50
  suggests
• 5:50
  the
• 5:50
  selection
• 5:51
  pressure
• 5:51
  hasn't
• 5:51
  kicked
• 5:52
  in
• 5:52
  yet.
• 5:56
  NVIDIA's
• 5:56
  OpenShell
• 5:57
  landed
• 5:57
  this
• 5:57
  week
• 5:57
  as
• 5:58
  an
• 5:58
  open-source
• 5:58
  runtime
• 5:59
  that
• 5:59
  sits
• 5:59
  between
• 6:00
  coding
• 6:00
  agents
• 6:00
  and
• 6:00
  infrastructure,
• 6:02
  governing
• 6:02
  execution,
• 6:03
  filesystem
• 6:03
  visibility,
• 6:04
  GPU
• 6:05
  pass-through,
• 6:05
  and
• 6:06
  inference
• 6:06
  routing.
• 6:07
  One
• 6:07
  policy
• 6:08
  update
• 6:08
  to
• 6:08
  wire
• 6:08
  agents
• 6:09
  to
• 6:09
  local
• 6:09
  inference,
• 6:10
  zero
• 6:10
  code
• 6:10
  changes
• 6:10
  —
• 6:11
  Claude
• 6:12
  Code,
• 6:12
  OpenCode,
• 6:13
  or
• 6:13
  any
• 6:13
  agent
• 6:14
  runs
• 6:14
  unmodified
• 6:15
  inside
• 6:15
  the
• 6:15
  sandbox.
• 6:16
  The
• 6:16
  pitch
• 6:16
  is
• 6:16
  "toddlers
• 6:17
  with
• 6:17
  sudo
• 6:17
  privileges
• 6:18
  need
• 6:18
  a
• 6:18
  chaperone
• 6:18
  with
• 6:19
  security
• 6:19
  clearance,"
• 6:20
  which
• 6:20
  is
• 6:20
  more
• 6:20
  honest
• 6:21
  than
• 6:21
  most
• 6:21
  agent
• 6:21
  infrastructure
• 6:22
  marketing.
• 6:23
  On
• 6:23
  the
• 6:23
  desktop
• 6:23
  agent
• 6:24
  front,
• 6:24
  H
• 6:24
  Company's
• 6:25
  Holo3
• 6:25
  posted
• 6:26
  78.85%
• 6:27
  on
• 6:27
  OSWorld-Verified,
• 6:29
  establishing
• 6:29
  a
• 6:29
  new
• 6:29
  state
• 6:30
  of
• 6:30
  the
• 6:30
  art
• 6:30
  for
• 6:30
  computer
• 6:30
  use
• 6:30
  —
• 6:31
  and
• 6:31
  it
• 6:31
  does
• 6:31
  so
• 6:31
  with
• 6:31
  only
• 6:32
  10B
• 6:32
  active
• 6:32
  parameters
• 6:33
  (122B
• 6:34
  total
• 6:34
  MoE),
• 6:35
  at
• 6:35
  a
• 6:35
  fraction
• 6:35
  of
• 6:35
  the
• 6:36
  cost
• 6:36
  of
• 6:36
  GPT
• 6:36
  5.4
• 6:37
  or
• 6:37
  Opus
• 6:37
  4.6.
• 6:39
  The
• 6:39
  training
• 6:39
  pipeline
• 6:40
  is
• 6:40
  built
• 6:40
  around
• 6:40
  a
• 6:40
  synthetic
• 6:41
  environment
• 6:41
  factory
• 6:42
  that
• 6:42
  reproduces
• 6:42
  enterprise
• 6:43
  systems
• 6:43
  and
• 6:43
  generates
• 6:44
  verifiable
• 6:44
  tasks,
• 6:45
  with
• 6:45
  a
• 6:45
  486-task
• 6:47
  corporate
• 6:47
  benchmark
• 6:47
  spanning
• 6:48
  e-commerce,
• 6:49
  collaboration,
• 6:49
  and
• 6:50
  multi-app
• 6:50
  workflows.
• 6:51
  The
• 6:51
  harder
• 6:51
  tasks
• 6:52
  require
• 6:52
  coordinating
• 6:53
  across
• 6:53
  multiple
• 6:54
  applications
• 6:54
  —
• 6:55
  retrieving
• 6:55
  equipment
• 6:56
  prices
• 6:56
  from
• 6:56
  a
• 6:56
  PDF,
• 6:57
  cross-referencing
• 6:58
  budgets,
• 6:58
  and
• 6:59
  sending
• 6:59
  personalized
• 6:59
  approval
• 7:00
  emails
• 7:00
  —
• 7:01
  the
• 7:01
  kind
• 7:01
  of
• 7:01
  sustained
• 7:02
  multi-step
• 7:02
  reasoning
• 7:03
  where
• 7:03
  most
• 7:03
  agents
• 7:03
  lose
• 7:03
  state
• 7:04
  or
• 7:04
  intent.
• 7:05
  The
• 7:05
  35B-A3B
• 7:07
  variant
• 7:07
  ships
• 7:08
  under
• 7:08
  Apache
• 7:08
  2.0.
• 7:10
  OpenYak
• 7:10
  takes
• 7:10
  the
• 7:10
  desktop
• 7:11
  agent
• 7:11
  local:
• 7:12
  an
• 7:12
  open-source
• 7:13
  AGPL-licensed
• 7:14
  app
• 7:14
  built
• 7:15
  on
• 7:15
  Tauri
• 7:15
  v2
• 7:16
  (Rust)
• 7:16
  +
• 7:17
  Next.js
• 7:17
  +
• 7:18
  FastAPI
• 7:19
  that
• 7:19
  connects
• 7:19
  to
• 7:19
  100+
• 7:20
  models
• 7:20
  from
• 7:21
  20+
• 7:21
  providers
• 7:22
  and
• 7:22
  runs
• 7:22
  20+
• 7:23
  built-in
• 7:23
  tools
• 7:23
  for
• 7:23
  file
• 7:24
  management,
• 7:25
  data
• 7:25
  analysis,
• 7:25
  and
• 7:26
  document
• 7:26
  creation
• 7:26
  —
• 7:27
  all
• 7:28
  without
• 7:28
  data
• 7:28
  leaving
• 7:28
  your
• 7:29
  machine.
• 7:30
  It
• 7:30
  auto-detects
• 7:31
  local
• 7:31
  Ollama
• 7:31
  models
• 7:32
  with
• 7:32
  full
• 7:32
  tool-calling
• 7:33
  support.
• 7:34
  Meanwhile,
• 7:35
  Cole
• 7:35
  Medin's
• 7:35
  live
• 7:35
  demo
• 7:36
  of
• 7:36
  multi-agent
• 7:37
  coding
• 7:37
  workflows
• 7:38
  with
• 7:38
  persistent
• 7:38
  memory
• 7:39
  shows
• 7:39
  where
• 7:39
  the
• 7:39
  orchestration
• 7:40
  layer
• 7:40
  is
• 7:40
  heading:
• 7:41
  agents
• 7:42
  that
• 7:42
  know
• 7:42
  your
• 7:42
  codebase,
• 7:43
  remember
• 7:43
  decisions
• 7:44
  from
• 7:44
  weeks
• 7:44
  ago,
• 7:45
  and
• 7:45
  coordinate
• 7:46
  research,
• 7:46
  implementation,
• 7:47
  and
• 7:48
  verification
• 7:48
  phases
• 7:49
  in
• 7:49
  parallel.
• 7:54
  Apple's
• 7:55
  M5
• 7:55
  Max
• 7:55
  is
• 7:56
  delivering
• 7:56
  real
• 7:56
  gains
• 7:57
  for
• 7:57
  local
• 7:57
  inference.
• 7:58
  Benchmarks
• 7:58
  against
• 7:58
  M3
• 7:59
  Max
• 7:59
  (both
• 8:00
  128GB,
• 8:00
  40
• 8:01
  GPU
• 8:01
  cores)
• 8:02
  across
• 8:02
  three
• 8:02
  Qwen
• 8:02
  3.5
• 8:03
  models
• 8:03
  show
• 8:03
  1.4x
• 8:04
  to
• 8:04
  1.7x
• 8:05
  improvement
• 8:06
  in
• 8:06
  token
• 8:06
  generation
• 8:07
  speed,
• 8:07
  with
• 8:07
  the
• 8:07
  gap
• 8:08
  widening
• 8:08
  dramatically
• 8:09
  at
• 8:09
  longer
• 8:09
  contexts
• 8:09
  —
• 8:10
  the
• 8:10
  27B
• 8:11
  dense
• 8:11
  model
• 8:11
  hits
• 8:11
  19.6
• 8:12
  tok/s
• 8:13
  at
• 8:13
  65K
• 8:14
  context
• 8:14
  on
• 8:14
  M5
• 8:15
  Max
• 8:15
  versus
• 8:15
  6.8
• 8:16
  on
• 8:16
  M3
• 8:16
  Max,
• 8:17
  a
• 8:17
  2.9x
• 8:17
  advantage.
• 8:18
  Prefill
• 8:19
  improvements
• 8:19
  reach
• 8:19
  4x
• 8:20
  at
• 8:20
  long
• 8:20
  context,
• 8:21
  driven
• 8:21
  by
• 8:21
  GPU
• 8:22
  Neural
• 8:22
  Accelerators.
• 8:23
  For
• 8:23
  agentic
• 8:24
  workloads,
• 8:24
  batching
• 8:25
  matters
• 8:25
  most:
• 8:26
  M5
• 8:26
  Max
• 8:26
  scales
• 8:27
  to
• 8:27
  2.54x
• 8:28
  throughput
• 8:28
  at
• 8:28
  4x
• 8:29
  batch
• 8:29
  on
• 8:29
  the
• 8:29
  35B-A3B
• 8:30
  MoE,
• 8:31
  while
• 8:31
  M3
• 8:31
  Max
• 8:32
  batching
• 8:32
  on
• 8:32
  dense
• 8:32
  models
• 8:33
  actually
• 8:33
  degrades.
• 8:34
  The
• 8:34
  614
• 8:35
  GB/s
• 8:36
  versus
• 8:36
  400
• 8:36
  GB/s
• 8:37
  bandwidth
• 8:37
  gap
• 8:37
  is
• 8:38
  significant
• 8:38
  for
• 8:38
  multi-step
• 8:39
  agent
• 8:39
  loops.
• 8:40
  At
• 8:40
  the
• 8:40
  extreme
• 8:40
  end
• 8:41
  of
• 8:41
  efficiency,
• 8:41
  Prism's
• 8:42
  1-Bit
• 8:42
  Bonsai
• 8:43
  models
• 8:43
  claim
• 8:43
  the
• 8:43
  title
• 8:44
  of
• 8:44
  first
• 8:44
  commercially
• 8:44
  viable
• 8:45
  1-bit
• 8:45
  LLMs.
• 8:46
  The
• 8:46
  8B
• 8:46
  variant
• 8:47
  requires
• 8:47
  only
• 8:47
  1.15GB
• 8:50
  of
• 8:50
  memory
• 8:50
  —
• 8:51
  a
• 8:51
  14x
• 8:51
  reduction
• 8:52
  from
• 8:52
  full
• 8:52
  precision
• 8:52
  —
• 8:53
  runs
• 8:53
  8x
• 8:53
  faster,
• 8:54
  uses
• 8:54
  5x
• 8:55
  less
• 8:55
  energy,
• 8:56
  and
• 8:56
  reportedly
• 8:56
  matches
• 8:57
  leading
• 8:57
  8B
• 8:57
  models
• 8:57
  on
• 8:58
  benchmarks
• 8:58
  spanning
• 8:58
  IFEval,
• 8:59
  GSM8K,
• 9:01
  HumanEval+,
• 9:02
  BFCL,
• 9:02
  MuSR,
• 9:03
  and
• 9:03
  MMLU-Redux.
• 9:05
  The
• 9:05
  4B
• 9:06
  model
• 9:06
  hits
• 9:06
  132
• 9:07
  tok/s
• 9:08
  on
• 9:08
  an
• 9:08
  M4
• 9:08
  Pro;
• 9:09
  the
• 9:09
  1.7B
• 9:10
  reaches
• 9:10
  130
• 9:11
  tok/s
• 9:12
  on
• 9:12
  an
• 9:12
  iPhone
• 9:13
  17
• 9:13
  Pro
• 9:13
  Max.
• 9:14
  If
• 9:14
  the
• 9:15
  benchmark
• 9:15
  claims
• 9:15
  hold
• 9:15
  under
• 9:16
  independent
• 9:16
  scrutiny,
• 9:17
  this
• 9:17
  moves
• 9:17
  1-bit
• 9:18
  quantization
• 9:18
  from
• 9:19
  research
• 9:19
  curiosity
• 9:19
  to
• 9:20
  deployment
• 9:20
  option
• 9:20
  for
• 9:20
  robotics,
• 9:21
  real-time
• 9:22
  agents,
• 9:22
  and
• 9:22
  edge
• 9:23
  computing.
• 9:24
  AMD's
• 9:24
  Lemonade
• 9:25
  server
• 9:25
  positions
• 9:25
  itself
• 9:26
  as
• 9:26
  the
• 9:26
  open-source
• 9:26
  local
• 9:27
  inference
• 9:27
  layer
• 9:28
  that
• 9:28
  works
• 9:28
  across
• 9:28
  GPUs
• 9:29
  and
• 9:29
  NPUs
• 9:30
  —
• 9:31
  a
• 9:31
  2MB
• 9:31
  service
• 9:32
  with
• 9:32
  automatic
• 9:32
  dependency
• 9:33
  configuration,
• 9:34
  multi-model
• 9:34
  concurrency,
• 9:35
  and
• 9:35
  OpenAI-compatible
• 9:36
  APIs
• 9:37
  across
• 9:37
  Windows,
• 9:38
  Linux,
• 9:38
  and
• 9:38
  macOS.
• 9:39
  It
• 9:40
  integrates
• 9:40
  llama.cpp,
• 9:42
  Ryzen
• 9:42
  AI
• 9:42
  SW,
• 9:43
  and
• 9:43
  FastFlowLM,
• 9:44
  with
• 9:44
  Qwen3.5-4B
• 9:48
  now
• 9:48
  available
• 9:48
  on
• 9:48
  NPU.
• 9:49
  On
• 9:50
  the
• 9:50
  quantization
• 9:50
  craft
• 9:51
  side,
• 9:51
  one
• 9:52
  developer
• 9:52
  spent
• 9:52
  48
• 9:53
  hours
• 9:53
  saturating
• 9:54
  Qwen
• 9:54
  3.5
• 9:55
  with
• 9:55
  2
• 9:55
  million
• 9:56
  tokens
• 9:56
  of
• 9:56
  high-density
• 9:57
  calibration
• 9:57
  data
• 9:58
  to
• 9:58
  produce
• 9:58
  what
• 9:58
  he
• 9:58
  calls
• 9:59
  the
• 9:59
  "Sovereign
• 9:59
  Series"
• 10:00
  —
• 10:01
  75
• 10:01
  GGUF
• 10:02
  models
• 10:02
  from
• 10:02
  0.8B
• 10:03
  to
• 10:03
  27B
• 10:04
  targeting
• 10:05
  what
• 10:05
  he
• 10:05
  terms
• 10:05
  "quantization-slop,"
• 10:07
  the
• 10:07
  subtle
• 10:08
  blurring
• 10:08
  of
• 10:08
  logic
• 10:08
  and
• 10:09
  linguistic
• 10:09
  nuance
• 10:09
  that
• 10:10
  standard
• 10:10
  40K-line
• 10:11
  calibration
• 10:11
  gists
• 10:12
  introduce.
• 10:13
  The
• 10:13
  27B
• 10:14
  "Colossus"
• 10:15
  hit
• 10:15
  8.71
• 10:16
  PPL
• 10:16
  after
• 10:17
  12
• 10:17
  hours
• 10:17
  of
• 10:17
  dedicated
• 10:18
  compute.
• 10:19
  Community
• 10:19
  testing
• 10:20
  was
• 10:20
  mixed:
• 10:21
  one
• 10:21
  user
• 10:21
  found
• 10:21
  the
• 10:21
  9B
• 10:22
  model
• 10:22
  couldn't
• 10:22
  follow
• 10:23
  a
• 10:23
  "respond
• 10:23
  in
• 10:23
  30
• 10:24
  words"
• 10:24
  constraint
• 10:25
  that
• 10:25
  vanilla
• 10:25
  Qwen3.5
• 10:28
  handled
• 10:28
  cleanly.
• 10:29
  The
• 10:29
  most
• 10:30
  ambitious
• 10:30
  application
• 10:30
  of
• 10:31
  local
• 10:31
  inference
• 10:31
  this
• 10:31
  week
• 10:32
  is
• 10:32
  a
• 10:32
  solo
• 10:32
  developer's
• 10:33
  medieval
• 10:33
  RPG
• 10:34
  where
• 10:34
  every
• 10:34
  significant
• 10:35
  NPC
• 10:35
  runs
• 10:36
  on
• 10:36
  a
• 10:36
  local
• 10:36
  uncensored
• 10:37
  LLM.
• 10:38
  The
• 10:38
  stack:
• 10:39
  Unreal
• 10:40
  Engine
• 10:40
  5,
• 10:40
  Ollama
• 10:41
  as
• 10:41
  a
• 10:41
  child
• 10:41
  process,
• 10:42
  Dolphin-Mistral
• 10:43
  7B
• 10:43
  Q4,
• 10:44
  Whisper
• 10:44
  for
• 10:45
  voice
• 10:45
  input,
• 10:46
  and
• 10:46
  Piper
• 10:46
  TTS
• 10:47
  for
• 10:47
  per-NPC
• 10:47
  voices
• 10:48
  with
• 10:48
  lip
• 10:48
  sync.
• 10:49
  No
• 10:49
  dialogue
• 10:50
  wheel
• 10:50
  —
• 10:51
  you
• 10:51
  say
• 10:51
  whatever
• 10:51
  you
• 10:52
  want,
• 10:52
  and
• 10:52
  the
• 10:52
  NPC
• 10:53
  responds
• 10:53
  and
• 10:54
  remembers.
• 10:55
  The
• 10:55
  community
• 10:55
  response
• 10:56
  was
• 10:56
  instructive:
• 10:57
  experienced
• 10:58
  builders
• 10:58
  immediately
• 10:59
  flagged
• 10:59
  context
• 10:59
  window
• 11:00
  degradation
• 11:00
  over
• 11:01
  long
• 11:01
  sessions,
• 11:02
  suggested
• 11:02
  fine-tuning
• 11:03
  over
• 11:03
  system
• 11:04
  prompts
• 11:04
  for
• 11:04
  character
• 11:04
  consistency,
• 11:06
  recommended
• 11:06
  switching
• 11:07
  from
• 11:07
  Ollama
• 11:07
  to
• 11:08
  raw
• 11:08
  llama.cpp
• 11:09
  for
• 11:10
  performance,
• 11:11
  and
• 11:11
  from
• 11:11
  Piper
• 11:11
  to
• 11:11
  Kokoro
• 11:12
  TTS
• 11:13
  for
• 11:13
  quality.
• 11:14
  The
• 11:14
  consensus:
• 11:15
  the
• 11:15
  concept
• 11:16
  is
• 11:16
  compelling,
• 11:17
  but
• 11:17
  "production
• 11:18
  hell
• 11:18
  awaits."
• 11:19
  Memory
• 11:19
  management
• 11:20
  as
• 11:20
  context
• 11:20
  grows
• 11:21
  is
• 11:21
  the
• 11:21
  decisive
• 11:21
  bottleneck,
• 11:22
  and
• 11:23
  no
• 11:23
  one
• 11:23
  has
• 11:23
  solved
• 11:23
  it
• 11:23
  at
• 11:24
  scale
• 11:24
  yet.
• 11:28
  TII's
• 11:29
  Falcon
• 11:29
  Perception
• 11:30
  is
• 11:30
  a
• 11:30
  0.6B-parameter
• 11:31
  early-fusion
• 11:32
  Transformer
• 11:32
  that
• 11:32
  handles
• 11:33
  open-vocabulary
• 11:34
  grounding
• 11:34
  and
• 11:34
  segmentation
• 11:35
  from
• 11:35
  natural
• 11:35
  language
• 11:35
  prompts
• 11:36
  —
• 11:36
  no
• 11:36
  separate
• 11:37
  vision
• 11:37
  backbone,
• 11:37
  no
• 11:37
  decoder
• 11:38
  pipeline,
• 11:39
  just
• 11:39
  one
• 11:39
  model
• 11:39
  processing
• 11:40
  image
• 11:40
  patches
• 11:40
  and
• 11:40
  text
• 11:41
  tokens
• 11:41
  in
• 11:41
  a
• 11:41
  shared
• 11:41
  parameter
• 11:42
  space
• 11:42
  with
• 11:42
  a
• 11:42
  hybrid
• 11:42
  attention
• 11:43
  mask.
• 11:44
  On
• 11:44
  the
• 11:44
  SA-Co
• 11:44
  benchmark
• 11:45
  it
• 11:45
  reaches
• 11:45
  72.7
• 11:46
  mAP
• 11:46
  versus
• 11:47
  SAM
• 11:47
  3's
• 11:48
  68.3,
• 11:49
  with
• 11:49
  outsized
• 11:49
  gains
• 11:50
  on
• 11:50
  attribute-heavy
• 11:50
  (+8.2)
• 11:51
  and
• 11:52
  food/drink
• 11:52
  (+12.2)
• 11:53
  categories.
• 11:54
  The
• 11:54
  Chain-of-Perception
• 11:55
  output
• 11:55
  interface
• 11:56
  decomposes
• 11:56
  each
• 11:56
  instance
• 11:57
  into
• 11:57
  coordinate,
• 11:58
  size,
• 11:58
  and
• 11:58
  segmentation
• 11:59
  steps
• 11:59
  using
• 11:59
  Fourier
• 11:59
  features
• 12:00
  and
• 12:00
  dot-product
• 12:01
  masks
• 12:01
  rather
• 12:01
  than
• 12:01
  Hungarian
• 12:02
  matching.
• 12:03
  Alongside
• 12:03
  it,
• 12:03
  Falcon
• 12:04
  OCR
• 12:04
  —
• 12:05
  a
• 12:05
  0.3B
• 12:06
  variant
• 12:06
  trained
• 12:06
  from
• 12:06
  scratch
• 12:07
  —
• 12:07
  hits
• 12:08
  80.3%
• 12:09
  on
• 12:09
  olmOCR
• 12:10
  and
• 12:10
  88.6
• 12:11
  on
• 12:11
  OmniDocBench
• 12:12
  while
• 12:12
  running
• 12:12
  3x
• 12:13
  faster
• 12:13
  than
• 12:13
  0.9B-class
• 12:14
  competitors.
• 12:15
  Trained
• 12:16
  on
• 12:16
  54M
• 12:16
  images
• 12:17
  with
• 12:17
  195M
• 12:18
  positive
• 12:18
  expressions
• 12:19
  and
• 12:19
  488M
• 12:20
  hard
• 12:20
  negatives
• 12:21
  across
• 12:21
  700
• 12:21
  GPU-hours,
• 12:23
  the
• 12:23
  model
• 12:23
  treats
• 12:23
  presence
• 12:24
  calibration
• 12:24
  as
• 12:24
  a
• 12:24
  first-class
• 12:25
  target
• 12:25
  with
• 12:25
  strict
• 12:26
  1:1
• 12:26
  positive-to-negative
• 12:27
  sampling.
• 12:28
  Hugging
• 12:28
  Face's
• 12:29
  TRL
• 12:29
  reaches
• 12:29
  v1.0,
• 12:31
  marking
• 12:32
  the
• 12:32
  post-training
• 12:32
  library's
• 12:33
  transition
• 12:33
  from
• 12:33
  research
• 12:34
  codebase
• 12:34
  to
• 12:34
  stable
• 12:34
  infrastructure.
• 12:36
  Now
• 12:36
  implementing
• 12:37
  75+
• 12:38
  methods
• 12:38
  across
• 12:38
  SFT,
• 12:39
  DPO,
• 12:40
  GRPO,
• 12:40
  RLOO,
• 12:41
  and
• 12:41
  reward
• 12:41
  modeling,
• 12:42
  TRL
• 12:43
  is
• 12:43
  downloaded
• 12:43
  3
• 12:44
  million
• 12:44
  times
• 12:44
  a
• 12:44
  month.
• 12:45
  The
• 12:45
  design
• 12:46
  philosophy
• 12:46
  is
• 12:46
  deliberately
• 12:47
  anti-abstraction:
• 12:48
  independent
• 12:49
  implementations
• 12:50
  over
• 12:50
  shared
• 12:50
  base
• 12:51
  classes,
• 12:51
  accepted
• 12:52
  code
• 12:52
  duplication
• 12:53
  over
• 12:53
  premature
• 12:54
  generalization.
• 12:55
  The
• 12:55
  roadmap
• 12:56
  targets
• 12:56
  asynchronous
• 12:57
  RL
• 12:57
  (decoupling
• 12:58
  generation
• 12:58
  from
• 12:58
  training),
• 12:59
  making
• 12:59
  training
• 13:00
  legible
• 13:00
  to
• 13:00
  agents
• 13:01
  via
• 13:01
  structured
• 13:01
  actionable
• 13:02
  warnings,
• 13:02
  and
• 13:03
  deeper
• 13:03
  MoE
• 13:03
  support.
• 13:04
  LeWorldModel
• 13:05
  (LeWM)
• 13:06
  is
• 13:06
  the
• 13:06
  first
• 13:06
  stable
• 13:07
  end-to-end
• 13:07
  JEPA
• 13:08
  that
• 13:08
  trains
• 13:08
  from
• 13:09
  raw
• 13:09
  pixels
• 13:09
  using
• 13:10
  only
• 13:10
  two
• 13:10
  loss
• 13:10
  terms
• 13:11
  —
• 13:12
  next-embedding
• 13:12
  prediction
• 13:13
  and
• 13:13
  a
• 13:13
  Gaussian
• 13:14
  regularizer
• 13:14
  —
• 13:15
  reducing
• 13:15
  tunable
• 13:16
  hyperparameters
• 13:16
  from
• 13:17
  six
• 13:17
  to
• 13:17
  one
• 13:17
  compared
• 13:18
  to
• 13:18
  alternatives.
• 13:19
  With
• 13:19
  ~15M
• 13:20
  parameters
• 13:21
  trainable
• 13:21
  on
• 13:21
  a
• 13:21
  single
• 13:22
  GPU
• 13:22
  in
• 13:22
  hours,
• 13:23
  it
• 13:23
  plans
• 13:24
  up
• 13:24
  to
• 13:24
  48x
• 13:25
  faster
• 13:25
  than
• 13:25
  foundation-model-based
• 13:26
  world
• 13:26
  models
• 13:27
  while
• 13:27
  remaining
• 13:28
  competitive
• 13:28
  across
• 13:28
  2D
• 13:29
  and
• 13:29
  3D
• 13:29
  control
• 13:30
  tasks.
• 13:31
  The
• 13:31
  latent
• 13:31
  space
• 13:32
  encodes
• 13:32
  meaningful
• 13:32
  physical
• 13:33
  structure,
• 13:34
  confirmed
• 13:34
  through
• 13:34
  probing
• 13:35
  and
• 13:35
  surprise
• 13:35
  evaluation
• 13:36
  that
• 13:36
  detects
• 13:37
  physically
• 13:37
  implausible
• 13:38
  events.
• 13:39
  This
• 13:39
  is
• 13:39
  the
• 13:39
  first
• 13:40
  concrete
• 13:40
  implementation
• 13:41
  in
• 13:41
  the
• 13:41
  lineage
• 13:41
  of
• 13:42
  LeCun's
• 13:42
  argument
• 13:43
  that
• 13:43
  autoregressive
• 13:43
  models
• 13:44
  are
• 13:44
  fundamentally
• 13:44
  limited
• 13:45
  by
• 13:45
  error
• 13:45
  compounding.
• 13:47
  On
• 13:47
  the
• 13:47
  abliteration
• 13:48
  front,
• 13:48
  Jim
• 13:49
  Lai's
• 13:49
  ORBA
• 13:49
  paper
• 13:50
  provides
• 13:50
  geometric
• 13:51
  grounding
• 13:51
  for
• 13:51
  why
• 13:51
  difference-of-means
• 13:52
  contrast
• 13:53
  vectors
• 13:53
  work:
• 13:54
  unit-normalized
• 13:55
  difference-of-means
• 13:56
  is
• 13:56
  exactly
• 13:56
  the
• 13:56
  normal
• 13:57
  of
• 13:57
  the
• 13:57
  Householder
• 13:57
  reflector
• 13:58
  mapping
• 13:58
  harmless
• 13:59
  to
• 13:59
  forbidden
• 13:59
  directions.
• 14:00
  The
• 14:01
  practical
• 14:01
  finding
• 14:01
  is
• 14:02
  negative
• 14:02
  —
• 14:03
  Householder
• 14:03
  reflection
• 14:04
  amplifies
• 14:04
  angular
• 14:05
  error
• 14:05
  into
• 14:05
  misdirected
• 14:06
  sign-flips
• 14:06
  that
• 14:07
  are
• 14:07
  more
• 14:07
  disruptive
• 14:07
  than
• 14:08
  projection-based
• 14:08
  zeroing,
• 14:09
  because
• 14:09
  models
• 14:10
  have
• 14:10
  effectively
• 14:10
  encountered
• 14:11
  magnitude-reduced
• 14:12
  activations
• 14:12
  via
• 14:13
  dropout
• 14:13
  but
• 14:14
  never
• 14:14
  encountered
• 14:14
  negated
• 14:15
  projections
• 14:15
  during
• 14:15
  training.
• 14:16
  Directional
• 14:17
  ablation
• 14:17
  remains
• 14:18
  the
• 14:18
  recommended
• 14:18
  primitive.
• 14:21
  A
• 14:21
  federal
• 14:22
  judge
• 14:22
  has
• 14:22
  blocked
• 14:23
  the
• 14:23
  Pentagon
• 14:23
  from
• 14:23
  designating
• 14:24
  Anthropic
• 14:25
  a
• 14:25
  national
• 14:25
  security
• 14:26
  threat,
• 14:26
  with
• 14:26
  Judge
• 14:27
  Rita
• 14:27
  Lin
• 14:27
  writing
• 14:28
  that
• 14:28
  the
• 14:28
  "financial
• 14:29
  and
• 14:29
  reputational
• 14:29
  harm"
• 14:30
  from
• 14:30
  the
• 14:30
  "likely
• 14:31
  unlawful"
• 14:31
  designation
• 14:32
  "risks
• 14:33
  crippling
• 14:33
  the
• 14:33
  company."
• 14:34
  The
• 14:34
  designation
• 14:34
  —
• 14:35
  historically
• 14:35
  reserved
• 14:36
  for
• 14:36
  foreign
• 14:36
  adversaries
• 14:37
  like
• 14:37
  Huawei
• 14:37
  —
• 14:38
  was
• 14:38
  apparently
• 14:39
  triggered
• 14:39
  by
• 14:39
  Anthropic's
• 14:40
  refusal
• 14:40
  to
• 14:40
  remove
• 14:41
  safety
• 14:41
  guardrails
• 14:41
  on
• 14:42
  autonomous
• 14:42
  weapons
• 14:43
  and
• 14:43
  mass
• 14:43
  surveillance
• 14:43
  applications.
• 14:45
  Lin
• 14:45
  noted
• 14:45
  the
• 14:45
  Pentagon
• 14:46
  retains
• 14:46
  discretion
• 14:46
  over
• 14:47
  which
• 14:47
  AI
• 14:47
  products
• 14:47
  it
• 14:48
  uses
• 14:48
  but
• 14:48
  argued
• 14:48
  the
• 14:48
  supply-chain-risk
• 14:49
  label
• 14:49
  doesn't
• 14:50
  appear
• 14:50
  directed
• 14:50
  at
• 14:50
  the
• 14:50
  government's
• 14:51
  stated
• 14:51
  national
• 14:51
  security
• 14:52
  interests.
• 14:53
  The
• 14:53
  administration
• 14:54
  has
• 14:54
  seven
• 14:54
  days
• 14:54
  to
• 14:55
  appeal.
• 14:56
  On
• 14:56
  the
• 14:56
  other
• 14:56
  side
• 14:56
  of
• 14:56
  the
• 14:56
  Atlantic,
• 14:57
  governance
• 14:57
  is
• 14:57
  being
• 14:57
  formalized
• 14:58
  rather
• 14:58
  than
• 14:58
  weaponized.
• 14:59
  A
• 14:59
  pilot
• 15:00
  study
• 15:00
  from
• 15:00
  TriStiX
• 15:01
  evaluates
• 15:01
  three
• 15:01
  t-norm
• 15:02
  operators
• 15:02
  —
• 15:03
  Lukasiewicz,
• 15:03
  Product,
• 15:04
  and
• 15:04
  Gödel
• 15:04
  —
• 15:05
  as
• 15:05
  logical
• 15:06
  conjunction
• 15:06
  mechanisms
• 15:06
  in
• 15:07
  a
• 15:07
  neuro-symbolic
• 15:07
  system
• 15:08
  for
• 15:08
  EU
• 15:08
  AI
• 15:08
  Act
• 15:09
  compliance
• 15:09
  classification.
• 15:10
  Using
• 15:11
  the
• 15:11
  LGGT+
• 15:12
  engine
• 15:12
  (Logic-Guided
• 15:13
  Graph
• 15:13
  Transformers
• 15:13
  Plus)
• 15:14
  across
• 15:14
  1,035
• 15:15
  annotated
• 15:16
  AI
• 15:16
  system
• 15:16
  descriptions
• 15:17
  spanning
• 15:17
  four
• 15:18
  risk
• 15:18
  categories,
• 15:19
  the
• 15:19
  study
• 15:19
  finds
• 15:19
  that
• 15:20
  Gödel's
• 15:20
  min-semantics
• 15:21
  achieves
• 15:21
  the
• 15:21
  highest
• 15:21
  accuracy
• 15:22
  (84.5%)
• 15:23
  and
• 15:23
  best
• 15:24
  borderline
• 15:24
  recall
• 15:24
  (85%)
• 15:25
  but
• 15:26
  introduces
• 15:26
  a
• 15:26
  0.8%
• 15:27
  false
• 15:28
  positive
• 15:28
  rate
• 15:28
  via
• 15:28
  over-classification.
• 15:30
  Lukasiewicz
• 15:30
  and
• 15:30
  Product
• 15:31
  maintain
• 15:31
  zero
• 15:31
  false
• 15:32
  positives
• 15:32
  but
• 15:33
  miss
• 15:33
  more
• 15:33
  borderline
• 15:33
  cases
• 15:34
  (81.2%
• 15:35
  and
• 15:35
  78.5%
• 15:36
  respectively).
• 15:37
  The
• 15:38
  deeper
• 15:38
  finding:
• 15:39
  operator
• 15:39
  choice
• 15:39
  is
• 15:39
  secondary
• 15:40
  to
• 15:40
  rule
• 15:40
  base
• 15:40
  completeness.
• 15:41
  Most
• 15:42
  false
• 15:42
  negatives
• 15:43
  stem
• 15:43
  from
• 15:43
  systems
• 15:44
  where
• 15:44
  AI
• 15:44
  plays
• 15:44
  an
• 15:44
  advisory
• 15:45
  but
• 15:45
  not
• 15:45
  decisive
• 15:46
  role
• 15:46
  —
• 15:47
  the
• 15:47
  classifier
• 15:47
  correctly
• 15:48
  identifies
• 15:48
  weak
• 15:49
  conditions,
• 15:50
  but
• 15:50
  the
• 15:50
  expert
• 15:50
  applies
• 15:51
  proportionality
• 15:51
  reasoning
• 15:52
  beyond
• 15:52
  what
• 15:52
  the
• 15:52
  rules
• 15:53
  encode.
• 15:54
  The
• 15:54
  paper
• 15:54
  maps
• 15:55
  a
• 15:55
  genuine
• 15:55
  legal
• 15:55
  interpretive
• 15:56
  question
• 15:56
  —
• 15:57
  whether
• 15:57
  rule
• 15:57
  conditions
• 15:58
  require
• 15:58
  strong
• 15:59
  joint
• 15:59
  confirmation
• 16:00
  or
• 16:00
  merely
• 16:00
  individual
• 16:01
  threshold-crossing
• 16:01
  —
• 16:02
  onto
• 16:03
  formal
• 16:03
  aggregation
• 16:04
  operators,
• 16:04
  with
• 16:04
  a
• 16:04
  mixed-semantics
• 16:05
  classifier
• 16:06
  (rule-specific
• 16:06
  t-norms
• 16:07
  annotated
• 16:07
  by
• 16:08
  legal
• 16:08
  experts)
• 16:08
  as
• 16:08
  the
• 16:09
  proposed
• 16:09
  next
• 16:09
  step.
• 16:10
  The
• 16:10
  full
• 16:11
  engine
• 16:11
  (201/201
• 16:12
  tests
• 16:13
  passing)
• 16:13
  ships
• 16:13
  under
• 16:14
  Apache
• 16:14
  2.0.