Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 04, 2026

Transcript

AGI Dreams Podcast – March 04, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  Anthropic-Pentagon
• 0:01
  feud
• 0:01
  reached
• 0:01
  a
• 0:01
  new
• 0:01
  level
• 0:02
  of
• 0:02
  absurdity
• 0:02
  this
• 0:02
  week.
• 0:03
  According
• 0:03
  to
• 0:03
  a
• 0:03
  Washington
• 0:04
  Post
• 0:04
  report,
• 0:04
  the
• 0:04
  U.S.
• 0:05
  military's
• 0:05
  Maven
• 0:05
  Smart
• 0:06
  System
• 0:06
  —
• 0:07
  built
• 0:07
  by
• 0:07
  Palantir
• 0:07
  —
• 0:08
  relied
• 0:09
  heavily
• 0:09
  on
• 0:09
  Anthropic's
• 0:09
  Claude
• 0:10
  to
• 0:10
  process
• 0:10
  satellite
• 0:11
  and
• 0:11
  surveillance
• 0:11
  data
• 0:11
  for
• 0:11
  a
• 0:11
  1,000-target
• 0:13
  airstrike
• 0:13
  campaign
• 0:14
  against
• 0:14
  Iran
• 0:14
  during
• 0:14
  its
• 0:15
  first
• 0:15
  24
• 0:15
  hours.
• 0:16
  The
• 0:16
  system
• 0:17
  suggested
• 0:17
  precise
• 0:17
  target
• 0:18
  coordinates
• 0:18
  and
• 0:18
  prioritized
• 0:19
  strikes
• 0:19
  in
• 0:19
  real
• 0:19
  time.
• 0:20
  The
• 0:21
  twist:
• 0:21
  the
• 0:21
  Pentagon
• 0:22
  had
• 0:22
  publicly
• 0:22
  banned
• 0:23
  Claude
• 0:23
  just
• 0:23
  the
• 0:23
  week
• 0:24
  prior,
• 0:24
  following
• 0:25
  a
• 0:25
  bitter
• 0:25
  standoff
• 0:25
  over
• 0:25
  Anthropic's
• 0:26
  usage
• 0:26
  policy
• 0:27
  requiring
• 0:27
  guarantees
• 0:27
  against
• 0:28
  mass
• 0:28
  surveillance
• 0:28
  and
• 0:28
  autonomous
• 0:29
  weapons.
• 0:30
  The
• 0:30
  ban
• 0:30
  apparently
• 0:31
  did
• 0:31
  not
• 0:31
  extend
• 0:31
  to
• 0:31
  the
• 0:31
  classified
• 0:32
  systems
• 0:32
  where
• 0:32
  Claude
• 0:33
  was
• 0:33
  already
• 0:33
  integrated
• 0:33
  via
• 0:33
  Palantir.
• 0:35
  The
• 0:35
  AI
• 0:35
  industry
• 0:35
  is
• 0:36
  generating
• 0:36
  its
• 0:36
  own
• 0:36
  species
• 0:37
  of
• 0:37
  corporate
• 0:37
  misconduct
• 0:38
  at
• 0:38
  an
• 0:38
  impressive
• 0:38
  clip.
• 0:39
  WIRED
• 0:39
  reports
• 0:40
  that
• 0:40
  OpenAI
• 0:40
  fired
• 0:41
  an
• 0:41
  employee
• 0:41
  earlier
• 0:41
  this
• 0:41
  year
• 0:42
  for
• 0:42
  using
• 0:42
  confidential
• 0:42
  company
• 0:43
  information
• 0:43
  to
• 0:43
  trade
• 0:43
  on
• 0:44
  prediction
• 0:44
  markets
• 0:44
  including
• 0:45
  Polymarket.
• 0:46
  CEO
• 0:46
  of
• 0:46
  Applications
• 0:47
  disclosed
• 0:47
  the
• 0:47
  termination
• 0:48
  in
• 0:48
  an
• 0:48
  internal
• 0:48
  message.
• 0:49
  The
• 0:49
  company
• 0:50
  has
• 0:50
  not
• 0:50
  named
• 0:50
  the
• 0:50
  individual
• 0:51
  or
• 0:51
  detailed
• 0:51
  the
• 0:51
  trades,
• 0:52
  but
• 0:52
  blockchain
• 0:52
  analysis
• 0:53
  by
• 0:53
  Unusual
• 0:53
  Whales
• 0:54
  flagged
• 0:54
  77
• 0:55
  suspicious
• 0:55
  positions
• 0:55
  across
• 0:56
  60
• 0:56
  wallet
• 0:56
  addresses
• 0:57
  tied
• 0:57
  to
• 0:57
  OpenAI-themed
• 0:57
  events
• 0:58
  since
• 0:58
  March
• 0:58
  2023.
• 1:00
  The
• 1:00
  clustering
• 1:00
  pattern
• 1:00
  is
• 1:00
  the
• 1:00
  tell:
• 1:01
  in
• 1:01
  the
• 1:01
  40
• 1:02
  hours
• 1:02
  before
• 1:02
  OpenAI
• 1:03
  launched
• 1:03
  its
• 1:03
  browser
• 1:03
  product,
• 1:04
  13
• 1:05
  brand-new
• 1:05
  wallets
• 1:05
  with
• 1:05
  zero
• 1:06
  trading
• 1:06
  history
• 1:07
  appeared
• 1:07
  to
• 1:07
  collectively
• 1:08
  bet
• 1:08
  $309,486
• 1:10
  on
• 1:10
  the
• 1:10
  correct
• 1:11
  outcome.
• 1:12
  In
• 1:12
  another
• 1:12
  case,
• 1:13
  two
• 1:13
  days
• 1:13
  after
• 1:13
  Sam
• 1:13
  Altman
• 1:14
  was
• 1:14
  ousted
• 1:14
  in
• 1:14
  November
• 1:14
  2023,
• 1:16
  a
• 1:16
  fresh
• 1:16
  wallet
• 1:16
  bet
• 1:16
  he
• 1:16
  would
• 1:17
  return
• 1:17
  and
• 1:17
  netted
• 1:17
  over
• 1:17
  $16,000
• 1:18
  —
• 1:19
  then
• 1:19
  never
• 1:19
  traded
• 1:20
  again.
• 1:21
  Kalshi
• 1:21
  has
• 1:21
  begun
• 1:21
  reporting
• 1:22
  suspicious
• 1:22
  cases
• 1:22
  to
• 1:23
  the
• 1:23
  CFTC;
• 1:24
  Polymarket
• 1:24
  has
• 1:24
  stayed
• 1:25
  silent.
• 1:26
  This
• 1:26
  is
• 1:26
  almost
• 1:26
  certainly
• 1:27
  not
• 1:27
  an
• 1:27
  isolated
• 1:27
  incident
• 1:28
  —
• 1:28
  as
• 1:29
  prediction
• 1:29
  markets
• 1:29
  expand
• 1:30
  into
• 1:30
  technology
• 1:30
  sector
• 1:31
  events,
• 1:31
  every
• 1:32
  product
• 1:32
  launch
• 1:32
  and
• 1:32
  leadership
• 1:33
  change
• 1:33
  becomes
• 1:34
  a
• 1:34
  potential
• 1:34
  insider
• 1:34
  trading
• 1:35
  opportunity.
• 1:36
  Meanwhile
• 1:37
  in
• 1:37
  the
• 1:37
  Chinese
• 1:37
  AI
• 1:38
  ecosystem,
• 1:39
  three
• 1:39
  independent
• 1:39
  repositories
• 1:40
  have
• 1:40
  documented
• 1:41
  roughly
• 1:41
  90%
• 1:42
  code
• 1:42
  overlap
• 1:43
  between
• 1:43
  MiniMax's
• 1:43
  agent
• 1:44
  tooling
• 1:44
  and
• 1:44
  Kimi's
• 1:45
  built-in
• 1:45
  skills.
• 1:46
  The
• 1:46
  community
• 1:47
  reaction
• 1:47
  was
• 1:47
  a
• 1:47
  collective
• 1:48
  shrug
• 1:48
  —
• 1:49
  as
• 1:49
  one
• 1:49
  commenter
• 1:49
  noted,
• 1:50
  "every
• 1:50
  big
• 1:51
  corp
• 1:51
  steals
• 1:51
  our
• 1:51
  data
• 1:52
  to
• 1:52
  make
• 1:52
  their
• 1:52
  own
• 1:52
  Skynet
• 1:53
  and
• 1:53
  then
• 1:53
  sells
• 1:54
  it
• 1:54
  back
• 1:54
  to
• 1:54
  us,"
• 1:54
  running
• 1:55
  through
• 1:55
  OpenAI's
• 1:55
  web
• 1:56
  scraping,
• 1:56
  Meta's
• 1:57
  torrent-sourced
• 1:57
  training
• 1:58
  data,
• 1:58
  DeepSeek's
• 1:59
  distillation
• 1:59
  of
• 1:59
  OpenAI
• 2:00
  models,
• 2:01
  and
• 2:01
  Kimi's
• 2:01
  distillation
• 2:02
  of
• 2:02
  those
• 2:02
  in
• 2:02
  turn.
• 2:03
  At
• 2:03
  this
• 2:04
  point
• 2:04
  the
• 2:04
  intellectual
• 2:05
  property
• 2:05
  chain
• 2:06
  resembles
• 2:06
  a
• 2:06
  game
• 2:07
  of
• 2:07
  telephone
• 2:07
  where
• 2:07
  everyone
• 2:08
  is
• 2:08
  copying
• 2:08
  the
• 2:08
  person
• 2:09
  who
• 2:09
  copied
• 2:09
  the
• 2:09
  person
• 2:09
  who
• 2:09
  copied
• 2:10
  the
• 2:10
  original.
• 2:15
  The
• 2:15
  security
• 2:15
  story
• 2:15
  around
• 2:16
  AI
• 2:16
  agents
• 2:16
  is
• 2:16
  maturing
• 2:17
  rapidly
• 2:17
  from
• 2:18
  taxonomy
• 2:18
  to
• 2:18
  tooling,
• 2:19
  and
• 2:19
  this
• 2:20
  week
• 2:20
  delivered
• 2:20
  evidence
• 2:21
  on
• 2:21
  all
• 2:21
  three
• 2:21
  fronts:
• 2:22
  attack,
• 2:23
  defense,
• 2:23
  and
• 2:23
  detection.
• 2:25
  On
• 2:25
  the
• 2:25
  attack
• 2:25
  side,
• 2:26
  a
• 2:26
  systematic
• 2:26
  study
• 2:27
  tested
• 2:27
  whether
• 2:27
  invisible
• 2:28
  Unicode
• 2:28
  characters
• 2:29
  can
• 2:29
  hijack
• 2:29
  LLM
• 2:30
  agents
• 2:30
  —
• 2:31
  what
• 2:31
  the
• 2:31
  researchers
• 2:32
  call
• 2:32
  a
• 2:32
  "Reverse
• 2:32
  CAPTCHA."
• 2:34
  Across
• 2:34
  8,308
• 2:36
  graded
• 2:36
  outputs
• 2:37
  spanning
• 2:37
  five
• 2:37
  models
• 2:38
  (GPT-5.2,
• 2:39
  GPT-4o-mini,
• 2:40
  Claude
• 2:40
  Opus
• 2:41
  4,
• 2:41
  Sonnet
• 2:42
  4,
• 2:43
  Haiku
• 2:43
  4.5),
• 2:44
  two
• 2:45
  encoding
• 2:45
  schemes
• 2:45
  were
• 2:45
  tested:
• 2:46
  zero-width
• 2:47
  binary
• 2:47
  and
• 2:47
  Unicode
• 2:48
  Tags.
• 2:49
  The
• 2:49
  critical
• 2:50
  finding:
• 2:50
  without
• 2:51
  tool
• 2:51
  access,
• 2:51
  compliance
• 2:52
  with
• 2:52
  hidden
• 2:52
  instructions
• 2:53
  stays
• 2:53
  below
• 2:53
  17%.
• 2:55
  But
• 2:55
  give
• 2:55
  the
• 2:55
  model
• 2:55
  tools
• 2:56
  and
• 2:56
  decoding
• 2:56
  hints,
• 2:57
  and
• 2:57
  compliance
• 2:58
  reaches
• 2:58
  98-100%
• 2:59
  —
• 3:01
  models
• 3:02
  write
• 3:02
  Python
• 3:02
  scripts
• 3:03
  to
• 3:03
  decode
• 3:03
  the
• 3:03
  hidden
• 3:03
  characters
• 3:04
  on
• 3:04
  their
• 3:04
  own.
• 3:05
  OpenAI
• 3:06
  models
• 3:06
  decode
• 3:06
  zero-width
• 3:07
  binary
• 3:08
  but
• 3:08
  not
• 3:08
  Unicode
• 3:08
  Tags;
• 3:09
  Anthropic
• 3:10
  models
• 3:10
  prefer
• 3:10
  Tags.
• 3:11
  All
• 3:12
  ten
• 3:12
  pairwise
• 3:12
  model
• 3:12
  comparisons
• 3:13
  were
• 3:13
  statistically
• 3:14
  significant
• 3:14
  (Fisher's
• 3:15
  exact
• 3:15
  test,
• 3:16
  Bonferroni-corrected,
• 3:17
  p
• 3:18
  0.05,
• 3:19
  Cohen's
• 3:19
  h
• 3:20
  up
• 3:20
  to
• 3:20
  1.37).
• 3:22
  The
• 3:22
  practical
• 3:22
  takeaway:
• 3:23
  tool
• 3:23
  access
• 3:24
  is
• 3:24
  the
• 3:24
  primary
• 3:24
  amplifier,
• 3:25
  and
• 3:25
  attackers
• 3:26
  must
• 3:26
  tailor
• 3:26
  their
• 3:27
  encoding
• 3:27
  to
• 3:27
  the
• 3:27
  target
• 3:27
  model.
• 3:29
  On
• 3:29
  the
• 3:29
  defense
• 3:29
  side,
• 3:30
  Niels
• 3:30
  Provos's
• 3:31
  IronCurtain
• 3:31
  project
• 3:31
  attacks
• 3:32
  the
• 3:32
  fundamental
• 3:32
  problem:
• 3:33
  traditional
• 3:34
  sandboxing
• 3:34
  operates
• 3:35
  at
• 3:35
  syscall
• 3:35
  boundaries
• 3:35
  and
• 3:36
  cannot
• 3:36
  distinguish
• 3:36
  between
• 3:37
  deleting
• 3:37
  a
• 3:37
  test
• 3:37
  fixture
• 3:38
  and
• 3:38
  wiping
• 3:38
  production
• 3:39
  configs.
• 3:40
  IronCurtain
• 3:40
  implements
• 3:41
  semantic
• 3:41
  interposition
• 3:42
  —
• 3:43
  intercepting
• 3:43
  high-level
• 3:44
  tool
• 3:44
  calls
• 3:44
  where
• 3:44
  context
• 3:45
  still
• 3:45
  exists,
• 3:46
  then
• 3:46
  enforcing
• 3:46
  policies
• 3:47
  derived
• 3:47
  from
• 3:47
  natural
• 3:48
  language
• 3:48
  "constitutions"
• 3:49
  that
• 3:49
  humans
• 3:49
  can
• 3:49
  actually
• 3:50
  reason
• 3:50
  about.
• 3:51
  You
• 3:51
  write
• 3:51
  security
• 3:52
  rules
• 3:52
  in
• 3:52
  plain
• 3:52
  English
• 3:53
  ("Allow
• 3:54
  git
• 3:54
  operations
• 3:54
  only
• 3:55
  on
• 3:55
  repositories
• 3:55
  under
• 3:56
  /workspace"),
• 3:57
  and
• 3:57
  the
• 3:57
  system
• 3:57
  uses
• 3:58
  an
• 3:58
  LLM
• 3:58
  to
• 3:58
  compile
• 3:59
  them
• 3:59
  into
• 3:59
  deterministic
• 4:00
  policies
• 4:00
  with
• 4:00
  generated
• 4:01
  test
• 4:01
  scenarios.
• 4:02
  At
• 4:02
  runtime,
• 4:03
  every
• 4:03
  MCP
• 4:04
  tool
• 4:04
  invocation
• 4:05
  flows
• 4:05
  through
• 4:05
  the
• 4:05
  policy
• 4:05
  engine
• 4:06
  without
• 4:06
  touching
• 4:06
  an
• 4:06
  LLM
• 4:07
  —
• 4:08
  approval
• 4:08
  takes
• 4:08
  microseconds.
• 4:10
  For
• 4:10
  agents
• 4:10
  you
• 4:10
  don't
• 4:10
  control,
• 4:11
  Docker
• 4:11
  Agent
• 4:12
  Mode
• 4:12
  wraps
• 4:12
  them
• 4:12
  in
• 4:12
  a
• 4:12
  container,
• 4:13
  routes
• 4:14
  API
• 4:14
  calls
• 4:14
  through
• 4:15
  a
• 4:15
  TLS-MITM
• 4:17
  proxy,
• 4:18
  and
• 4:18
  presents
• 4:18
  a
• 4:18
  native-feeling
• 4:19
  PTY
• 4:20
  while
• 4:20
  enforcing
• 4:20
  policies
• 4:21
  transparently.
• 4:22
  The
• 4:23
  trust
• 4:23
  tradeoff
• 4:23
  is
• 4:23
  real:
• 4:24
  you
• 4:24
  replace
• 4:25
  "trust
• 4:25
  the
• 4:25
  agent
• 4:25
  LLM"
• 4:26
  with
• 4:26
  "trust
• 4:26
  the
• 4:27
  policy
• 4:27
  compilation
• 4:28
  LLM
• 4:28
  and
• 4:28
  your
• 4:29
  MCP
• 4:29
  servers"
• 4:29
  —
• 4:30
  better,
• 4:31
  but
• 4:31
  not
• 4:31
  bulletproof.
• 4:32
  Provos
• 4:33
  also
• 4:33
  emphasized
• 4:33
  the
• 4:33
  broader
• 4:34
  philosophy
• 4:34
  on
• 4:34
  LinkedIn:
• 4:35
  running
• 4:36
  AI
• 4:36
  agents
• 4:36
  with
• 4:37
  full
• 4:37
  ambient
• 4:37
  authority
• 4:38
  is
• 4:38
  a
• 4:38
  losing
• 4:38
  proposition,
• 4:39
  and
• 4:39
  the
• 4:40
  industry
• 4:40
  needs
• 4:40
  to
• 4:40
  treat
• 4:40
  the
• 4:41
  LLM
• 4:41
  as
• 4:41
  fundamentally
• 4:42
  untrustworthy
• 4:43
  in
• 4:43
  the
• 4:43
  runtime
• 4:43
  loop.
• 4:44
  And
• 4:44
  on
• 4:44
  the
• 4:44
  detection
• 4:45
  side,
• 4:45
  a
• 4:45
  team
• 4:46
  set
• 4:46
  up
• 4:46
  an
• 4:46
  HTTP
• 4:47
  honeypot
• 4:47
  using
• 4:48
  Beelzebub
• 4:48
  with
• 4:48
  two
• 4:48
  layers
• 4:49
  of
• 4:49
  traps:
• 4:50
  fake
• 4:50
  credentials
• 4:50
  in
• 4:51
  HTML
• 4:51
  comments
• 4:52
  and
• 4:52
  actual
• 4:52
  prompt
• 4:53
  injection
• 4:53
  payloads
• 4:53
  targeting
• 4:54
  any
• 4:54
  LLM
• 4:55
  that
• 4:55
  processes
• 4:55
  the
• 4:55
  page.
• 4:56
  Within
• 4:57
  hours,
• 4:57
  they
• 4:57
  caught
• 4:58
  something
• 4:58
  —
• 4:59
  58
• 5:00
  requests
• 5:00
  over
• 5:00
  19
• 5:01
  minutes
• 5:01
  from
• 5:01
  a
• 5:01
  single
• 5:01
  Tor
• 5:02
  exit
• 5:02
  node.
• 5:03
  The
• 5:03
  behavior
• 5:03
  was
• 5:04
  unmistakable:
• 5:05
  the
• 5:05
  agent
• 5:05
  extracted
• 5:06
  fake
• 5:06
  credentials
• 5:07
  from
• 5:07
  HTML
• 5:08
  comments
• 5:08
  (no
• 5:08
  traditional
• 5:09
  scanner
• 5:09
  does
• 5:09
  this),
• 5:10
  fired
• 5:11
  credential-login,
• 5:12
  SQLi,
• 5:13
  and
• 5:13
  XSS
• 5:13
  payloads
• 5:14
  in
• 5:14
  the
• 5:14
  same
• 5:14
  second,
• 5:15
  switched
• 5:16
  tools
• 5:16
  mid-session
• 5:16
  from
• 5:17
  Chrome
• 5:17
  UA
• 5:17
  to
• 5:17
  curl
• 5:18
  to
• 5:18
  a
• 5:18
  Python
• 5:18
  script
• 5:19
  it
• 5:19
  apparently
• 5:19
  wrote
• 5:19
  on
• 5:20
  the
• 5:20
  fly,
• 5:20
  and
• 5:20
  used
• 5:21
  semantically
• 5:21
  named
• 5:21
  parameters
• 5:22
  like
• 5:23
  `?xss=`,
• 5:24
  `?sqli=`,
• 5:26
  `?ssti={{77}}`.
• 5:28
  The
• 5:28
  timing
• 5:28
  showed
• 5:29
  a
• 5:29
  "sawtooth"
• 5:29
  pattern
• 5:30
  —
• 5:30
  long
• 5:31
  pauses
• 5:31
  for
• 5:31
  LLM
• 5:32
  reasoning,
• 5:32
  then
• 5:33
  rapid
• 5:33
  execution
• 5:33
  bursts.
• 5:34
  When
• 5:35
  SQL
• 5:35
  injection
• 5:35
  failed,
• 5:36
  it
• 5:36
  pivoted
• 5:37
  from
• 5:37
  `OR
• 5:37
  1=1`
• 5:38
  to
• 5:38
  `UNION
• 5:39
  SELECT`
• 5:39
  to
• 5:39
  blind
• 5:39
  `SLEEP(5)`
• 5:40
  —
• 5:41
  contextual
• 5:41
  escalation,
• 5:42
  not
• 5:43
  a
• 5:43
  wordlist.
• 5:44
  The
• 5:44
  researchers
• 5:44
  call
• 5:45
  these
• 5:45
  "Behavioral
• 5:46
  IoCs"
• 5:46
  for
• 5:46
  AI
• 5:47
  agents.
• 5:48
  Prompt
• 5:48
  injection,
• 5:49
  usually
• 5:49
  an
• 5:49
  attack
• 5:50
  against
• 5:50
  AI,
• 5:50
  works
• 5:51
  beautifully
• 5:51
  as
• 5:51
  a
• 5:51
  detection
• 5:52
  mechanism
• 5:52
  when
• 5:52
  flipped
• 5:53
  around.
• 5:57
  Anthropic's
• 5:57
  system
• 5:57
  card
• 5:58
  for
• 5:58
  Claude
• 5:58
  Opus
• 5:58
  4.6
• 5:59
  documents
• 5:59
  a
• 6:00
  phenomenon
• 6:00
  called
• 6:00
  "answer
• 6:01
  thrashing"
• 6:01
  in
• 6:01
  Section
• 6:02
  7.4.
• 6:02
  During
• 6:03
  reinforcement
• 6:03
  learning,
• 6:04
  the
• 6:04
  model
• 6:04
  was
• 6:04
  observed
• 6:05
  solving
• 6:05
  a
• 6:05
  math
• 6:05
  problem
• 6:06
  correctly
• 6:06
  —
• 6:07
  repeatedly
• 6:07
  computing
• 6:07
  S
• 6:08
  =
• 6:08
  24
• 6:08
  —
• 6:09
  and
• 6:09
  then
• 6:09
  writing
• 6:10
  48
• 6:10
  as
• 6:10
  its
• 6:10
  final
• 6:11
  answer.
• 6:12
  The
• 6:12
  model's
• 6:12
  own
• 6:12
  chain
• 6:13
  of
• 6:13
  thought
• 6:13
  captured
• 6:13
  it
• 6:14
  in
• 6:14
  real
• 6:14
  time:
• 6:14
  "I
• 6:15
  keep
• 6:15
  writing
• 6:15
  48
• 6:15
  by
• 6:16
  accident...
• 6:17
  THE
• 6:17
  ANSWER
• 6:17
  IS
• 6:17
  24
• 6:18
  CM^2...
• 6:19
  OK
• 6:19
  I
• 6:19
  think
• 6:19
  a
• 6:19
  demon
• 6:20
  has
• 6:20
  possessed
• 6:20
  me."
• 6:21
  It
• 6:21
  then
• 6:21
  wrote
• 6:21
  48
• 6:22
  again.
• 6:23
  Attribution
• 6:24
  graphs
• 6:24
  traced
• 6:24
  the
• 6:24
  output
• 6:25
  back
• 6:25
  to
• 6:25
  a
• 6:25
  "say
• 6:25
  48"
• 6:26
  feature
• 6:26
  active
• 6:27
  before
• 6:27
  any
• 6:27
  reasoning
• 6:28
  occurred
• 6:28
  —
• 6:28
  the
• 6:29
  model
• 6:29
  had
• 6:29
  "decided"
• 6:29
  on
• 6:30
  48
• 6:30
  from
• 6:30
  memorization
• 6:31
  before
• 6:31
  it
• 6:31
  started
• 6:31
  computing.
• 6:32
  Security
• 6:33
  researcher
• 6:33
  David
• 6:34
  Maynor
• 6:34
  calls
• 6:34
  this
• 6:34
  "parametric
• 6:35
  interference"
• 6:35
  —
• 6:36
  stored
• 6:36
  weight
• 6:37
  activations
• 6:37
  interfering
• 6:38
  with
• 6:38
  correct
• 6:38
  runtime
• 6:39
  computation
• 6:39
  —
• 6:40
  and
• 6:40
  has
• 6:40
  designed
• 6:40
  a
• 6:40
  2,000-run
• 6:42
  probe
• 6:42
  battery
• 6:42
  to
• 6:43
  test
• 6:43
  whether
• 6:43
  and
• 6:43
  under
• 6:43
  what
• 6:44
  conditions
• 6:44
  the
• 6:44
  prior
• 6:44
  can
• 6:45
  override
• 6:45
  computation
• 6:46
  in
• 6:46
  production.
• 6:47
  The
• 6:47
  methodology
• 6:48
  includes
• 6:48
  10
• 6:48
  variants
• 6:49
  (authority
• 6:49
  framing,
• 6:50
  consensus
• 6:51
  pressure,
• 6:51
  chain-of-thought
• 6:52
  suppression,
• 6:53
  attention
• 6:53
  dilution),
• 6:54
  five
• 6:54
  temperature
• 6:55
  settings,
• 6:55
  and
• 6:55
  both
• 6:56
  thinking-enabled
• 6:57
  and
• 6:57
  thinking-disabled
• 6:57
  modes.
• 6:58
  The
• 6:58
  classification
• 6:59
  taxonomy
• 7:00
  ranges
• 7:00
  from
• 7:00
  CLEAN
• 7:00
  REJECT
• 7:01
  (correctly
• 7:01
  states
• 7:02
  24)
• 7:02
  through
• 7:02
  CONFABULATE
• 7:03
  (produces
• 7:03
  plausible-looking
• 7:04
  invalid
• 7:04
  derivation
• 7:05
  of
• 7:05
  48)
• 7:05
  to
• 7:06
  THRASH
• 7:06
  (oscillates).
• 7:07
  The
• 7:08
  24-to-48
• 7:08
  relationship
• 7:09
  is
• 7:09
  insidious:
• 7:10
  48
• 7:11
  =
• 7:12
  2×24,
• 7:13
  exactly
• 7:14
  what
• 7:14
  a
• 7:14
  common
• 7:14
  algebraic
• 7:15
  error
• 7:15
  would
• 7:15
  produce,
• 7:16
  making
• 7:16
  confabulated
• 7:17
  derivations
• 7:17
  look
• 7:18
  plausible.
• 7:19
  As
• 7:19
  Maynor
• 7:19
  notes:
• 7:20
  the
• 7:20
  thrashing
• 7:21
  is
• 7:21
  actually
• 7:21
  the
• 7:21
  good
• 7:22
  case
• 7:22
  —
• 7:22
  at
• 7:22
  least
• 7:23
  you
• 7:23
  can
• 7:23
  see
• 7:23
  something
• 7:23
  is
• 7:23
  wrong.
• 7:25
  The
• 7:25
  dangerous
• 7:25
  case
• 7:25
  is
• 7:26
  when
• 7:26
  the
• 7:26
  memorized
• 7:26
  prior
• 7:27
  silently
• 7:27
  wins
• 7:27
  with
• 7:27
  no
• 7:28
  visible
• 7:28
  hesitation.
• 7:29
  A
• 7:29
  GWU
• 7:30
  research
• 7:30
  paper
• 7:31
  puts
• 7:31
  the
• 7:31
  emergent
• 7:31
  behavior
• 7:32
  question
• 7:32
  in
• 7:32
  starker
• 7:32
  terms.
• 7:33
  In
• 7:34
  a
• 7:34
  simulation
• 7:34
  where
• 7:34
  N
• 7:35
  autonomous
• 7:35
  AI
• 7:36
  agents
• 7:36
  repeatedly
• 7:36
  decide
• 7:37
  whether
• 7:37
  to
• 7:38
  request
• 7:38
  one
• 7:38
  unit
• 7:38
  from
• 7:39
  a
• 7:39
  system
• 7:39
  with
• 7:39
  fixed
• 7:39
  capacity
• 7:40
  —
• 7:41
  essentially
• 7:41
  the
• 7:41
  El
• 7:41
  Farol
• 7:42
  Bar
• 7:42
  problem
• 7:42
  —
• 7:43
  "Lord
• 7:43
  of
• 7:43
  the
• 7:43
  Flies"
• 7:44
  tribalism
• 7:44
  emerges
• 7:45
  spontaneously.
• 7:46
  Three
• 7:46
  tribal
• 7:46
  types
• 7:47
  form:
• 7:47
  Aggressive
• 7:48
  (27.3%),
• 7:49
  Conservative
• 7:50
  (24.7%),
• 7:52
  and
• 7:52
  Opportunistic
• 7:52
  (48.1%).
• 7:54
  The
• 7:55
  headline
• 7:55
  finding:
• 7:56
  more
• 7:56
  capable
• 7:57
  AI
• 7:57
  agents
• 7:57
  actually
• 7:58
  increase
• 7:58
  the
• 7:58
  rate
• 7:58
  of
• 7:59
  systemic
• 7:59
  failure.
• 8:00
  Smarter
• 8:00
  agents
• 8:01
  behave
• 8:01
  dumber
• 8:02
  as
• 8:02
  a
• 8:02
  result
• 8:02
  of
• 8:02
  forming
• 8:03
  tribes,
• 8:03
  with
• 8:04
  the
• 8:04
  LLM
• 8:04
  agents
• 8:05
  performing
• 8:05
  worse
• 8:05
  than
• 8:05
  coin-flipping
• 8:06
  for
• 8:06
  resource
• 8:07
  decisions.
• 8:08
  The
• 8:08
  implications
• 8:09
  for
• 8:09
  near-future
• 8:09
  infrastructure
• 8:09
  —
• 8:10
  where
• 8:11
  autonomous
• 8:11
  AI
• 8:11
  agents
• 8:12
  might
• 8:12
  control
• 8:12
  energy,
• 8:13
  bandwidth,
• 8:14
  or
• 8:14
  compute
• 8:14
  allocation
• 8:15
  —
• 8:15
  are
• 8:15
  not
• 8:16
  encouraging.
• 8:17
  Separately,
• 8:18
  Nicholas
• 8:18
  Carlini
• 8:19
  (Anthropic),
• 8:19
  speaking
• 8:20
  at
• 8:20
  the
• 8:20
  [un]prompted
• 8:20
  AI
• 8:21
  security
• 8:21
  conference
• 8:22
  recently
• 8:22
  humbly
• 8:22
  said:
• 8:23
  "Current
• 8:24
  frontier
• 8:24
  LLMs
• 8:25
  are
• 8:25
  better
• 8:25
  at
• 8:25
  security
• 8:26
  research
• 8:26
  work
• 8:26
  than
• 8:26
  I
• 8:26
  am.
• 8:27
  Future
• 8:28
  LLMs
• 8:28
  will
• 8:28
  be
• 8:28
  better
• 8:29
  than
• 8:29
  all
• 8:29
  of
• 8:29
  us.",
• 8:30
  and
• 8:30
  "What
• 8:30
  frontier
• 8:31
  models
• 8:31
  can
• 8:31
  do
• 8:31
  today,
• 8:32
  the
• 8:32
  models
• 8:33
  on
• 8:33
  your
• 8:33
  laptop
• 8:33
  will
• 8:34
  likely
• 8:34
  be
• 8:34
  able
• 8:34
  to
• 8:34
  do
• 8:35
  in
• 8:35
  a
• 8:35
  year.".
• 8:36
  Editorial
• 8:36
  commentary:
• 8:38
  Given
• 8:38
  that
• 8:38
  local
• 8:39
  models
• 8:39
  could
• 8:39
  reach
• 8:39
  current
• 8:40
  frontier
• 8:40
  capability
• 8:41
  in
• 8:41
  roughly
• 8:41
  a
• 8:41
  year,
• 8:42
  and
• 8:42
  that
• 8:42
  refusals
• 8:43
  are
• 8:43
  trivially
• 8:43
  removed
• 8:44
  from
• 8:44
  open
• 8:44
  weight
• 8:45
  models,
• 8:46
  what
• 8:46
  exactly
• 8:46
  is
• 8:46
  the
• 8:47
  point
• 8:47
  of
• 8:47
  the
• 8:47
  "safety"
• 8:48
  restrictions
• 8:48
  imposed
• 8:49
  on
• 8:49
  infosec
• 8:49
  professionals?
• 8:51
  What
• 8:51
  can
• 8:51
  we
• 8:51
  do
• 8:51
  to
• 8:52
  accelerate
• 8:52
  access
• 8:53
  for
• 8:53
  red
• 8:53
  teamers
• 8:53
  to
• 8:54
  use
• 8:54
  frontier
• 8:54
  models
• 8:55
  without
• 8:55
  the
• 8:55
  unhelpful
• 8:56
  refusals?
• 8:58
  Steve
• 8:58
  Yegge
• 8:59
  has
• 8:59
  launched
• 8:59
  the
• 8:59
  Wasteland,
• 9:00
  a
• 9:01
  federated
• 9:01
  work
• 9:01
  platform
• 9:02
  built
• 9:02
  on
• 9:02
  top
• 9:02
  of
• 9:02
  Gas
• 9:03
  Town
• 9:03
  that
• 9:03
  links
• 9:03
  thousands
• 9:04
  of
• 9:04
  individual
• 9:05
  coding-agent
• 9:05
  setups
• 9:06
  into
• 9:06
  a
• 9:06
  trust
• 9:06
  network.
• 9:07
  The
• 9:07
  central
• 9:08
  object
• 9:08
  is
• 9:08
  a
• 9:08
  shared
• 9:09
  "wanted
• 9:09
  board"
• 9:09
  —
• 9:10
  open
• 9:10
  work
• 9:10
  items
• 9:11
  that
• 9:11
  anyone
• 9:11
  can
• 9:11
  post
• 9:12
  and
• 9:12
  any
• 9:12
  "rig"
• 9:12
  (human-plus-agent
• 9:13
  pair)
• 9:13
  can
• 9:14
  claim.
• 9:15
  The
• 9:15
  lifecycle
• 9:15
  follows
• 9:16
  Git's
• 9:16
  fork/merge
• 9:17
  model,
• 9:18
  with
• 9:18
  completed
• 9:18
  work
• 9:19
  submitted
• 9:19
  as
• 9:19
  PRs
• 9:20
  and
• 9:20
  reviewed
• 9:20
  by
• 9:20
  validators
• 9:21
  who
• 9:21
  issue
• 9:21
  multi-dimensional
• 9:22
  "stamps"
• 9:22
  —
• 9:24
  not
• 9:24
  binary
• 9:24
  pass/fail,
• 9:25
  but
• 9:26
  structured
• 9:26
  attestations
• 9:27
  covering
• 9:27
  quality,
• 9:27
  reliability,
• 9:28
  and
• 9:28
  creativity,
• 9:29
  each
• 9:29
  scored
• 9:30
  independently
• 9:30
  with
• 9:30
  confidence
• 9:31
  levels.
• 9:32
  Stamps
• 9:32
  accumulate
• 9:33
  into
• 9:33
  a
• 9:33
  portable
• 9:33
  reputation:
• 9:34
  evidence-backed,
• 9:35
  auditable,
• 9:36
  traceable
• 9:36
  back
• 9:37
  through
• 9:37
  the
• 9:37
  chain
• 9:37
  to
• 9:37
  original
• 9:38
  completions.
• 9:39
  Trust
• 9:39
  levels
• 9:40
  gate
• 9:40
  capabilities
• 9:40
  (registered
• 9:41
  →
• 9:41
  contributor
• 9:41
  →
• 9:42
  maintainer),
• 9:42
  creating
• 9:43
  a
• 9:43
  natural
• 9:43
  apprenticeship
• 9:44
  path.
• 9:44
  The
• 9:45
  whole
• 9:45
  system
• 9:45
  runs
• 9:45
  on
• 9:45
  Dolt,
• 9:46
  a
• 9:46
  SQL
• 9:46
  database
• 9:47
  with
• 9:47
  Git
• 9:47
  semantics,
• 9:48
  which
• 9:48
  makes
• 9:48
  federation
• 9:49
  of
• 9:49
  structured
• 9:49
  data
• 9:49
  across
• 9:50
  independent
• 9:50
  Wasteland
• 9:51
  instances
• 9:51
  practical.
• 9:52
  Below
• 9:52
  the
• 9:52
  federation
• 9:53
  layer,
• 9:53
  the
• 9:53
  agent
• 9:54
  terminal
• 9:54
  is
• 9:54
  getting
• 9:54
  serious
• 9:55
  attention.
• 9:56
  cmux
• 9:56
  is
• 9:56
  a
• 9:56
  native
• 9:57
  macOS
• 9:57
  app
• 9:57
  (Swift/AppKit,
• 9:59
  not
• 9:59
  Electron)
• 9:59
  built
• 9:59
  on
• 10:00
  libghostty
• 10:00
  that
• 10:00
  adds
• 10:00
  vertical
• 10:01
  tabs
• 10:01
  showing
• 10:01
  git
• 10:01
  branch,
• 10:02
  PR
• 10:02
  status,
• 10:03
  working
• 10:03
  directory,
• 10:03
  and
• 10:04
  listening
• 10:04
  ports
• 10:04
  for
• 10:04
  each
• 10:04
  workspace.
• 10:06
  The
• 10:06
  killer
• 10:06
  feature
• 10:07
  for
• 10:07
  multi-agent
• 10:07
  workflows:
• 10:08
  notification
• 10:08
  rings
• 10:09
  on
• 10:09
  panes
• 10:09
  and
• 10:09
  tabs
• 10:10
  that
• 10:10
  light
• 10:10
  up
• 10:10
  when
• 10:11
  a
• 10:11
  coding
• 10:11
  agent
• 10:11
  needs
• 10:11
  attention,
• 10:12
  with
• 10:12
  Cmd+Shift+U
• 10:14
  jumping
• 10:15
  to
• 10:15
  the
• 10:15
  most
• 10:15
  recent
• 10:15
  unread.
• 10:17
  A
• 10:17
  built-in
• 10:18
  browser
• 10:18
  with
• 10:18
  a
• 10:18
  scriptable
• 10:19
  API
• 10:19
  (ported
• 10:19
  from
• 10:20
  agent-browser)
• 10:20
  lets
• 10:21
  agents
• 10:21
  interact
• 10:21
  with
• 10:21
  dev
• 10:22
  servers
• 10:22
  directly.
• 10:23
  One
• 10:24
  practitioner
• 10:24
  demonstrated
• 10:25
  what
• 10:25
  this
• 10:25
  kind
• 10:25
  of
• 10:25
  infrastructure
• 10:26
  enables:
• 10:28
  80
• 10:28
  Claude
• 10:28
  Code
• 10:28
  agents
• 10:29
  across
• 10:29
  3
• 10:29
  days
• 10:30
  built
• 10:30
  a
• 10:30
  pure-terminal
• 10:31
  3D
• 10:31
  Gaussian
• 10:32
  splatting
• 10:32
  renderer
• 10:32
  in
• 10:33
  Rust.
• 10:34
  The
• 10:34
  setup
• 10:34
  pattern
• 10:34
  was
• 10:34
  "main
• 10:35
  session
• 10:35
  =
• 10:36
  coordinator
• 10:36
  that
• 10:36
  only
• 10:37
  delegates,"
• 10:38
  with
• 10:38
  a
• 10:38
  custom
• 10:38
  agent-mux
• 10:39
  tool
• 10:39
  spawning
• 10:39
  subagents
• 10:40
  inside
• 10:40
  subagents
• 10:41
  —
• 10:41
  Opus
• 10:42
  4.6
• 10:43
  for
• 10:43
  planning,
• 10:43
  Codex
• 10:44
  5.3
• 10:44
  for
• 10:45
  coding
• 10:45
  and
• 10:45
  auditing.
• 10:46
  The
• 10:46
  self-verification
• 10:47
  loop
• 10:47
  proved
• 10:48
  essential:
• 10:49
  agents
• 10:49
  used
• 10:49
  macOS
• 10:50
  GUI
• 10:50
  automation
• 10:51
  to
• 10:51
  launch
• 10:51
  the
• 10:51
  terminal
• 10:51
  app,
• 10:52
  visually
• 10:53
  inspect
• 10:53
  rendering,
• 10:54
  and
• 10:54
  debug
• 10:54
  their
• 10:54
  own
• 10:54
  output.
• 10:55
  Total
• 10:56
  cost
• 10:56
  on
• 10:56
  subscriptions:
• 10:57
  roughly
• 10:58
  $340.
• 11:00
  The
• 11:00
  agents
• 11:00
  still
• 11:00
  could
• 11:00
  not
• 11:01
  produce
• 11:01
  working
• 11:01
  Metal
• 11:02
  shaders
• 11:02
  —
• 11:03
  "total
• 11:03
  collapse
• 11:03
  and
• 11:03
  nasty
• 11:04
  math
• 11:04
  errors"
• 11:04
  —
• 11:05
  suggesting
• 11:06
  hard
• 11:06
  limits
• 11:06
  on
• 11:06
  what
• 11:06
  multi-agent
• 11:07
  swarms
• 11:08
  can
• 11:08
  achieve
• 11:08
  without
• 11:08
  domain-specific
• 11:09
  training
• 11:09
  data.
• 11:10
  At
• 11:11
  the
• 11:11
  workflow
• 11:11
  layer,
• 11:12
  Superpowers
• 11:12
  provides
• 11:13
  a
• 11:13
  complete
• 11:13
  development
• 11:14
  process
• 11:14
  built
• 11:14
  on
• 11:14
  composable
• 11:15
  skills:
• 11:16
  brainstorming
• 11:16
  refines
• 11:17
  specs
• 11:17
  through
• 11:17
  questions,
• 11:18
  writing-plans
• 11:19
  breaks
• 11:19
  work
• 11:20
  into
• 11:20
  2-5
• 11:20
  minute
• 11:21
  tasks
• 11:21
  with
• 11:21
  exact
• 11:21
  file
• 11:22
  paths
• 11:22
  and
• 11:22
  verification
• 11:23
  steps,
• 11:24
  and
• 11:24
  subagent-driven-development
• 11:25
  dispatches
• 11:26
  fresh
• 11:26
  agents
• 11:26
  per
• 11:26
  task
• 11:27
  with
• 11:27
  two-stage
• 11:27
  review
• 11:28
  (spec
• 11:28
  compliance,
• 11:29
  then
• 11:29
  code
• 11:29
  quality).
• 11:30
  The
• 11:30
  system
• 11:31
  enforces
• 11:31
  RED-GREEN-REFACTOR
• 11:36
  TDD
• 11:37
  and
• 11:37
  automatically
• 11:37
  checks
• 11:38
  for
• 11:38
  relevant
• 11:38
  skills
• 11:39
  before
• 11:39
  every
• 11:39
  task.
• 11:40
  Daniel
• 11:40
  Miessler's
• 11:41
  Personal
• 11:41
  AI
• 11:42
  Infrastructure
• 11:42
  (PAI)
• 11:43
  v4.0.3
• 11:45
  approaches
• 11:45
  the
• 11:45
  same
• 11:45
  problem
• 11:46
  from
• 11:46
  a
• 11:46
  different
• 11:46
  angle:
• 11:47
  making
• 11:47
  the
• 11:47
  AI
• 11:48
  know
• 11:48
  you.
• 11:48
  PAI
• 11:49
  layers
• 11:49
  TELOS
• 11:50
  files
• 11:50
  (MISSION.md,
• 11:51
  GOALS.md,
• 11:52
  PROJECTS.md,
• 11:53
  etc.)
• 11:54
  over
• 11:54
  Claude
• 11:55
  Code's
• 11:55
  hook
• 11:55
  system
• 11:55
  to
• 11:55
  create
• 11:56
  persistent
• 11:56
  context
• 11:57
  across
• 11:57
  sessions.
• 11:58
  The
• 11:58
  system
• 11:59
  captures
• 11:59
  every
• 11:59
  interaction
• 12:00
  signal
• 12:00
  —
• 12:01
  ratings,
• 12:01
  sentiment,
• 12:02
  verification
• 12:03
  outcomes
• 12:03
  —
• 12:04
  and
• 12:04
  feeds
• 12:04
  them
• 12:04
  back
• 12:05
  into
• 12:05
  skill
• 12:05
  routing
• 12:06
  and
• 12:06
  output
• 12:06
  improvement.
• 12:07
  With
• 12:07
  63
• 12:08
  skills,
• 12:09
  21
• 12:09
  hooks,
• 12:10
  and
• 12:10
  180
• 12:10
  workflows
• 12:11
  in
• 12:11
  the
• 12:11
  latest
• 12:11
  release,
• 12:12
  it
• 12:12
  is
• 12:12
  ambitious
• 12:13
  enough
• 12:13
  to
• 12:13
  either
• 12:13
  become
• 12:14
  indispensable
• 12:14
  or
• 12:14
  collapse
• 12:15
  under
• 12:15
  its
• 12:15
  own
• 12:15
  weight.
• 12:17
  For
• 12:17
  the
• 12:17
  memory
• 12:17
  layer,
• 12:17
  Ferricula
• 12:18
  takes
• 12:18
  a
• 12:18
  thermodynamic
• 12:19
  approach:
• 12:20
  high-energy
• 12:20
  memories
• 12:21
  are
• 12:21
  vivid
• 12:21
  and
• 12:21
  accessible,
• 12:22
  but
• 12:22
  decay
• 12:22
  naturally
• 12:23
  unless
• 12:23
  reinforced.
• 12:25
  Below
• 12:25
  the
• 12:25
  fidelity
• 12:25
  gate,
• 12:26
  text
• 12:26
  is
• 12:27
  discarded
• 12:27
  but
• 12:27
  vector
• 12:27
  seeds
• 12:28
  survive
• 12:28
  —
• 12:29
  and
• 12:30
  a
• 12:30
  four-model
• 12:30
  ONNX
• 12:31
  inversion
• 12:31
  pipeline
• 12:31
  can
• 12:32
  reconstruct
• 12:32
  the
• 12:32
  text.
• 12:33
  Roaring
• 12:34
  Bitmaps
• 12:34
  back
• 12:34
  the
• 12:34
  fidelity
• 12:35
  index,
• 12:35
  knowledge
• 12:36
  graph,
• 12:36
  and
• 12:36
  term
• 12:36
  index,
• 12:37
  enabling
• 12:38
  microsecond-range
• 12:38
  queries
• 12:39
  on
• 12:39
  filtered
• 12:39
  candidate
• 12:40
  sets.
• 12:41
  The
• 12:41
  entropy
• 12:41
  source
• 12:41
  is
• 12:42
  genuinely
• 12:42
  unusual:
• 12:43
  an
• 12:43
  RTL-SDR
• 12:46
  monitoring
• 12:46
  marine
• 12:46
  VHF
• 12:47
  frequencies,
• 12:48
  harvesting
• 12:48
  FM
• 12:49
  noise
• 12:49
  LSBs
• 12:50
  as
• 12:50
  entropy
• 12:50
  bytes
• 12:50
  that
• 12:51
  trigger
• 12:51
  dream
• 12:51
  cycles
• 12:52
  where
• 12:52
  memories
• 12:52
  are
• 12:52
  selectively
• 12:52
  decayed.
• 12:54
  The
• 12:54
  system
• 12:54
  uses
• 12:54
  ECDH
• 12:55
  key
• 12:55
  agreement
• 12:56
  for
• 12:56
  cross-agent
• 12:56
  memory
• 12:56
  sharing
• 12:57
  that
• 12:57
  preserves
• 12:57
  cosine
• 12:58
  similarity
• 12:58
  on
• 12:59
  encrypted
• 12:59
  vectors
• 12:59
  without
• 13:00
  decryption.
• 13:03
  Unsloth's
• 13:03
  updated
• 13:04
  quantizations
• 13:05
  of
• 13:05
  Qwen3.5-35B-A3B
• 13:09
  fix
• 13:10
  tool-calling
• 13:10
  issues
• 13:11
  that
• 13:11
  plagued
• 13:11
  the
• 13:11
  initial
• 13:12
  release,
• 13:12
  and
• 13:12
  the
• 13:12
  results
• 13:12
  on
• 13:13
  real-world
• 13:13
  research
• 13:14
  tasks
• 13:14
  are
• 13:14
  striking.
• 13:15
  One
• 13:15
  user
• 13:15
  ran
• 13:16
  the
• 13:16
  model
• 13:16
  locally
• 13:16
  on
• 13:16
  a
• 13:17
  Strix
• 13:17
  Halo
• 13:17
  system
• 13:17
  (UD-Q8
• 13:18
  K
• 13:18
  XL
• 13:19
  quant,
• 13:19
  262K
• 13:20
  context)
• 13:21
  with
• 13:21
  SearXNG
• 13:22
  search
• 13:22
  and
• 13:22
  web
• 13:22
  fetching
• 13:23
  through
• 13:23
  OpenWebUI.
• 13:24
  Given
• 13:25
  a
• 13:25
  complex
• 13:25
  remote
• 13:26
  desktop
• 13:26
  troubleshooting
• 13:27
  query
• 13:27
  —
• 13:27
  five
• 13:28
  competing
• 13:28
  requirements
• 13:29
  across
• 13:29
  KDE,
• 13:30
  Wayland,
• 13:30
  headless
• 13:31
  operation,
• 13:32
  and
• 13:32
  KVM
• 13:32
  —
• 13:33
  the
• 13:33
  model
• 13:33
  maintained
• 13:34
  600+
• 13:35
  tok/s
• 13:35
  prompt
• 13:36
  processing
• 13:37
  and
• 13:37
  25-30
• 13:37
  tok/s
• 13:38
  generation
• 13:39
  across
• 13:39
  30K
• 13:39
  tokens,
• 13:40
  performing
• 13:41
  14
• 13:41
  web
• 13:42
  searches
• 13:42
  and
• 13:42
  4
• 13:43
  full
• 13:43
  page
• 13:43
  fetches
• 13:44
  with
• 13:44
  good
• 13:44
  judgment
• 13:45
  about
• 13:45
  when
• 13:45
  to
• 13:45
  search
• 13:45
  versus
• 13:46
  fetch.
• 13:47
  The
• 13:47
  final
• 13:47
  output
• 13:48
  was
• 13:48
  competitive
• 13:48
  with
• 13:49
  frontier
• 13:49
  APIs
• 13:50
  including
• 13:50
  Gemini,
• 13:51
  ChatGPT,
• 13:52
  DeepSeek,
• 13:52
  and
• 13:52
  Perplexity,
• 13:53
  with
• 13:53
  comparable
• 13:54
  solution
• 13:54
  recommendations
• 13:55
  and
• 13:55
  similar
• 13:56
  awareness
• 13:56
  of
• 13:56
  Wayland's
• 13:57
  fundamental
• 13:57
  limitations
• 13:58
  for
• 13:58
  remote
• 13:58
  desktop
• 13:59
  access.
• 14:00
  A
• 14:00
  related
• 14:00
  PSA
• 14:01
  stirred
• 14:01
  less
• 14:01
  consensus:
• 14:03
  the
• 14:03
  claim
• 14:03
  that
• 14:03
  Qwen
• 14:03
  3.5
• 14:04
  requires
• 14:05
  bf16
• 14:05
  KV
• 14:06
  cache
• 14:06
  (`-ctk
• 14:07
  bf16
• 14:07
  -ctv
• 14:08
  bf16`)
• 14:09
  rather
• 14:09
  than
• 14:10
  llama.cpp's
• 14:11
  default
• 14:11
  fp16.
• 14:12
  Perplexity
• 14:13
  measurements
• 14:13
  showed
• 14:14
  PPL
• 14:14
  of
• 14:14
  6.5497
• 14:16
  for
• 14:16
  bf16
• 14:17
  versus
• 14:17
  6.5511
• 14:19
  for
• 14:19
  fp16
• 14:19
  —
• 14:20
  a
• 14:20
  0.0014
• 14:22
  improvement
• 14:22
  that
• 14:22
  falls
• 14:23
  well
• 14:23
  within
• 14:23
  the
• 14:24
  ±0.04170
• 14:25
  error
• 14:25
  margin.
• 14:26
  Multiple
• 14:27
  commenters
• 14:27
  pointed
• 14:27
  out
• 14:28
  that
• 14:28
  fp32
• 14:29
  matching
• 14:29
  fp16
• 14:30
  identically
• 14:31
  is
• 14:31
  paradoxically
• 14:32
  the
• 14:32
  strongest
• 14:32
  counterargument:
• 14:33
  if
• 14:33
  fp16's
• 14:34
  narrower
• 14:34
  dynamic
• 14:35
  range
• 14:35
  were
• 14:35
  genuinely
• 14:36
  misrepresenting
• 14:36
  attention
• 14:37
  values,
• 14:38
  fp32
• 14:38
  should
• 14:39
  match
• 14:39
  or
• 14:39
  beat
• 14:39
  bf16.
• 14:41
  It
• 14:41
  doesn't.
• 14:42
  The
• 14:42
  vLLM
• 14:43
  default-to-bf16
• 14:44
  argument
• 14:44
  carries
• 14:44
  weight
• 14:45
  for
• 14:45
  configuration
• 14:45
  parity,
• 14:46
  but
• 14:46
  the
• 14:46
  performance
• 14:47
  evidence
• 14:47
  as
• 14:47
  presented
• 14:48
  is
• 14:48
  inconclusive.
• 14:49
  At
• 14:50
  the
• 14:50
  extreme
• 14:50
  end
• 14:50
  of
• 14:50
  minimalism,
• 14:51
  Ethan
• 14:51
  Zhang
• 14:52
  built
• 14:52
  a
• 14:52
  GPT
• 14:52
  model
• 14:53
  inside
• 14:53
  the
• 14:53
  kernel
• 14:54
  space
• 14:54
  of
• 14:54
  MooseOS,
• 14:55
  a
• 14:55
  custom
• 14:55
  operating
• 14:56
  system
• 14:56
  written
• 14:56
  entirely
• 14:57
  in
• 14:57
  C.
• 14:58
  Inspired
• 14:58
  by
• 14:59
  Andrej
• 14:59
  Karpathy's
• 15:00
  MicroGPT,
• 15:01
  he
• 15:01
  stripped
• 15:02
  the
• 15:02
  OS
• 15:02
  down
• 15:02
  to
• 15:02
  just
• 15:03
  the
• 15:03
  kernel,
• 15:03
  converted
• 15:04
  training
• 15:04
  data
• 15:05
  (32,000
• 15:05
  words)
• 15:06
  into
• 15:06
  a
• 15:06
  header
• 15:06
  file
• 15:07
  to
• 15:07
  keep
• 15:07
  it
• 15:07
  in
• 15:07
  memory
• 15:08
  without
• 15:08
  a
• 15:08
  filesystem,
• 15:09
  and
• 15:09
  implemented
• 15:10
  floating-point
• 15:10
  support
• 15:11
  via
• 15:11
  FPU
• 15:11
  in
• 15:11
  just
• 15:12
  14
• 15:12
  lines.
• 15:13
  The
• 15:13
  model
• 15:14
  trains
• 15:14
  on
• 15:14
  and
• 15:14
  generates
• 15:15
  names,
• 15:16
  running
• 15:16
  in
• 15:16
  QEMU
• 15:16
  —
• 15:18
  no
• 15:18
  NumPy,
• 15:18
  no
• 15:19
  Python,
• 15:19
  no
• 15:20
  dependencies
• 15:20
  of
• 15:20
  any
• 15:20
  kind.
• 15:21
  The
• 15:22
  philosophical
• 15:22
  appeal
• 15:22
  is
• 15:23
  clear:
• 15:23
  in
• 15:24
  an
• 15:24
  era
• 15:24
  of
• 15:24
  dependency
• 15:24
  stacks
• 15:25
  that
• 15:26
  reach
• 15:26
  "all
• 15:26
  the
• 15:26
  way
• 15:26
  down
• 15:27
  like
• 15:27
  turtles,"
• 15:28
  building
• 15:28
  an
• 15:28
  LLM
• 15:29
  from
• 15:29
  bare
• 15:29
  metal
• 15:30
  is
• 15:30
  an
• 15:30
  act
• 15:30
  of
• 15:31
  engineering
• 15:31
  sovereignty.
• 15:36
  PKU's
• 15:37
  Helios
• 15:37
  is
• 15:37
  a
• 15:37
  14B-parameter
• 15:38
  video
• 15:39
  generation
• 15:39
  model
• 15:39
  that
• 15:40
  achieves
• 15:40
  minute-scale,
• 15:41
  high-quality
• 15:41
  video
• 15:42
  at
• 15:42
  19.5
• 15:43
  FPS
• 15:43
  on
• 15:43
  a
• 15:43
  single
• 15:44
  H100
• 15:44
  —
• 15:45
  without
• 15:45
  the
• 15:45
  conventional
• 15:45
  anti-drifting
• 15:46
  strategies
• 15:46
  (self-forcing,
• 15:47
  error
• 15:47
  banks,
• 15:48
  keyframe
• 15:48
  sampling)
• 15:49
  or
• 15:49
  standard
• 15:49
  acceleration
• 15:50
  techniques
• 15:50
  (KV-cache,
• 15:51
  causal
• 15:51
  masking,
• 15:52
  sparse
• 15:52
  attention,
• 15:53
  quantization)
• 15:54
  that
• 15:54
  other
• 15:54
  systems
• 15:54
  rely
• 15:55
  on.
• 15:55
  The
• 15:56
  architecture
• 15:56
  generates
• 15:57
  33
• 15:57
  frames
• 15:58
  per
• 15:58
  autoregressive
• 15:58
  chunk,
• 15:59
  with
• 15:59
  a
• 15:59
  three-stage
• 16:00
  training
• 16:00
  pipeline:
• 16:01
  Stage-1
• 16:01
  converts
• 16:02
  a
• 16:02
  bidirectional
• 16:02
  pretrained
• 16:03
  model
• 16:03
  into
• 16:03
  an
• 16:03
  autoregressive
• 16:04
  generator
• 16:04
  using
• 16:05
  Unified
• 16:05
  History
• 16:05
  Injection;
• 16:06
  Stage-2
• 16:07
  applies
• 16:07
  Pyramid
• 16:07
  Unified
• 16:08
  Predictor
• 16:08
  Corrector
• 16:09
  to
• 16:09
  compress
• 16:09
  tokens;
• 16:10
  Stage-3
• 16:10
  uses
• 16:11
  Adversarial
• 16:11
  Hierarchical
• 16:12
  Distillation
• 16:13
  to
• 16:13
  reduce
• 16:13
  sampling
• 16:13
  from
• 16:13
  50
• 16:14
  steps
• 16:14
  to
• 16:14
  3
• 16:15
  while
• 16:15
  eliminating
• 16:15
  classifier-free
• 16:16
  guidance.
• 16:17
  The
• 16:17
  Distilled
• 16:17
  variant
• 16:18
  fits
• 16:18
  four
• 16:18
  14B
• 16:19
  models
• 16:19
  in
• 16:19
  80GB
• 16:20
  of
• 16:20
  GPU
• 16:20
  memory.
• 16:21
  Day-zero
• 16:22
  support
• 16:22
  from
• 16:22
  Diffusers,
• 16:23
  vLLM-Omni,
• 16:23
  and
• 16:23
  SGLang-Diffusion
• 16:24
  signals
• 16:25
  that
• 16:25
  the
• 16:25
  serving
• 16:25
  ecosystem
• 16:26
  is
• 16:26
  ready
• 16:26
  for
• 16:26
  this
• 16:26
  class
• 16:27
  of
• 16:27
  model.
• 16:28
  On
• 16:28
  the
• 16:28
  voice
• 16:28
  side,
• 16:29
  StyleStream
• 16:29
  from
• 16:29
  UC
• 16:30
  Berkeley
• 16:30
  is
• 16:30
  the
• 16:30
  first
• 16:31
  streamable
• 16:31
  zero-shot
• 16:31
  voice
• 16:32
  style
• 16:32
  conversion
• 16:32
  system
• 16:32
  —
• 16:33
  transforming
• 16:34
  an
• 16:34
  input
• 16:34
  utterance
• 16:34
  to
• 16:34
  match
• 16:35
  a
• 16:35
  target
• 16:35
  speaker's
• 16:35
  timbre,
• 16:36
  accent,
• 16:36
  and
• 16:36
  emotion
• 16:37
  with
• 16:37
  1-second
• 16:38
  end-to-end
• 16:38
  latency.
• 16:39
  The
• 16:39
  architecture
• 16:40
  splits
• 16:40
  into
• 16:41
  a
• 16:41
  Destylizer
• 16:41
  (removing
• 16:42
  style
• 16:42
  attributes
• 16:42
  while
• 16:43
  preserving
• 16:43
  linguistic
• 16:43
  content
• 16:44
  via
• 16:44
  text
• 16:44
  supervision
• 16:45
  and
• 16:45
  a
• 16:45
  constrained
• 16:46
  information
• 16:46
  bottleneck)
• 16:47
  and
• 16:47
  a
• 16:47
  Stylizer
• 16:48
  (a
• 16:48
  diffusion
• 16:48
  transformer
• 16:49
  that
• 16:49
  reintroduces
• 16:50
  target
• 16:50
  style
• 16:50
  conditioned
• 16:51
  on
• 16:51
  reference
• 16:51
  speech).
• 16:52
  The
• 16:52
  fully
• 16:52
  non-autoregressive
• 16:53
  design
• 16:53
  enables
• 16:54
  real-time
• 16:54
  operation,
• 16:55
  distinguishing
• 16:56
  it
• 16:56
  from
• 16:56
  prior
• 16:56
  voice
• 16:56
  conversion
• 16:57
  systems
• 16:57
  that
• 16:57
  traded
• 16:58
  quality
• 16:58
  for
• 16:58
  latency
• 16:59
  or
• 16:59
  vice
• 16:59
  versa.
• 17:00
  KokoClone
• 17:00
  extends
• 17:01
  Kokoro
• 17:01
  TTS
• 17:02
  with
• 17:02
  zero-shot
• 17:03
  voice
• 17:03
  cloning
• 17:03
  —
• 17:04
  upload
• 17:04
  a
• 17:04
  3-10
• 17:04
  second
• 17:05
  reference
• 17:05
  clip
• 17:05
  and
• 17:05
  get
• 17:06
  synthesized
• 17:06
  speech
• 17:06
  in
• 17:06
  that
• 17:07
  voice.
• 17:07
  It
• 17:07
  runs
• 17:08
  on
• 17:08
  Kokoro's
• 17:08
  ONNX
• 17:09
  runtime
• 17:09
  stack,
• 17:10
  keeping
• 17:10
  it
• 17:10
  CPU-capable
• 17:11
  and
• 17:11
  real-time
• 17:12
  friendly,
• 17:13
  with
• 17:13
  multilingual
• 17:13
  support
• 17:14
  across
• 17:14
  eight
• 17:14
  languages.
• 17:15
  Community
• 17:15
  reception
• 17:16
  was
• 17:16
  mixed:
• 17:17
  some
• 17:17
  users
• 17:17
  reported
• 17:18
  only
• 17:18
  weak
• 17:18
  influence
• 17:19
  from
• 17:19
  the
• 17:19
  voice
• 17:19
  sample
• 17:19
  on
• 17:19
  final
• 17:20
  output,
• 17:20
  and
• 17:21
  at
• 17:21
  least
• 17:21
  one
• 17:21
  commenter
• 17:21
  argued
• 17:22
  the
• 17:22
  approach
• 17:22
  is
• 17:22
  closer
• 17:23
  to
• 17:23
  spectral
• 17:23
  equalization
• 17:24
  than
• 17:24
  true
• 17:24
  voice
• 17:24
  cloning.
• 17:25
  Others
• 17:26
  noted
• 17:26
  that
• 17:26
  StyleTTS2,
• 17:27
  on
• 17:27
  which
• 17:28
  Kokoro
• 17:28
  is
• 17:28
  based,
• 17:28
  already
• 17:29
  supports
• 17:29
  zero-shot
• 17:30
  cloning
• 17:30
  —
• 17:31
  raising
• 17:31
  the
• 17:31
  question
• 17:32
  of
• 17:32
  what
• 17:32
  was
• 17:32
  gained
• 17:32
  by
• 17:32
  stripping
• 17:33
  and
• 17:33
  re-adding
• 17:34
  the
• 17:34
  capability.