Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – July 13, 2026

Transcript

AGI Dreams Podcast – July 13, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  United
• 0:00
  States
• 0:00
  Senate
• 0:01
  Armed
• 0:01
  Services
• 0:01
  Committee
• 0:02
  has
• 0:02
  approved
• 0:03
  language
• 0:03
  in
• 0:03
  its
• 0:04
  fiscal
• 0:04
  2027
• 0:05
  National
• 0:05
  Defense
• 0:06
  Authorization
• 0:06
  Act
• 0:07
  that
• 0:07
  would
• 0:07
  authorize
• 0:08
  Defense
• 0:08
  Secretary
• 0:09
  Pete
• 0:09
  Hegseth
• 0:10
  to
• 0:10
  establish
• 0:10
  a
• 0:10
  pilot
• 0:11
  program
• 0:11
  allowing
• 0:11
  private-sector
• 0:12
  contractors
• 0:13
  to
• 0:13
  conduct
• 0:13
  cyber
• 0:13
  operations
• 0:14
  "through
• 0:15
  contractor-owned,
• 0:16
  contractor-operated
• 0:17
  means"
• 0:17
  under
• 0:18
  U.S.
• 0:18
  Cyber
• 0:18
  Command's
• 0:18
  operational
• 0:19
  authority.
• 0:20
  The
• 0:21
  provision,
• 0:21
  which
• 0:21
  still
• 0:21
  requires
• 0:22
  full
• 0:22
  Senate
• 0:23
  and
• 0:23
  House
• 0:23
  approval,
• 0:24
  limits
• 0:24
  contractors
• 0:25
  to
• 0:25
  "access
• 0:25
  generation
• 0:26
  and
• 0:26
  maintenance"
• 0:26
  —
• 0:26
  but
• 0:27
  as
• 0:27
  multiple
• 0:27
  former
• 0:28
  senior
• 0:28
  cyber
• 0:28
  officials
• 0:29
  pointed
• 0:29
  out,
• 0:29
  that
• 0:30
  phrase
• 0:30
  covers
• 0:30
  nearly
• 0:31
  the
• 0:31
  entire
• 0:31
  operational
• 0:32
  envelope
• 0:32
  of
• 0:32
  modern
• 0:33
  cyber
• 0:33
  campaigns.
• 0:34
  Nick
• 0:34
  Leiserson,
• 0:35
  who
• 0:35
  spent
• 0:35
  a
• 0:35
  dozen
• 0:36
  years
• 0:36
  as
• 0:36
  a
• 0:36
  House
• 0:36
  staffer
• 0:37
  before
• 0:37
  serving
• 0:37
  in
• 0:38
  the
• 0:38
  Biden
• 0:38
  White
• 0:38
  House's
• 0:38
  Office
• 0:39
  of
• 0:39
  the
• 0:39
  National
• 0:39
  Cyber
• 0:40
  Director,
• 0:40
  called
• 0:41
  the
• 0:41
  proposal
• 0:41
  a
• 0:41
  direct
• 0:42
  threat
• 0:42
  to
• 0:42
  global
• 0:42
  cyber
• 0:43
  stability
• 0:43
  —
• 0:44
  noting
• 0:44
  the
• 0:44
  U.S.
• 0:45
  has
• 0:45
  sanctioned
• 0:45
  Chinese
• 0:46
  military
• 0:46
  and
• 0:46
  intelligence
• 0:47
  contractors
• 0:47
  for
• 0:48
  doing
• 0:48
  exactly
• 0:48
  what
• 0:48
  this
• 0:49
  provision
• 0:49
  would
• 0:49
  authorize
• 0:50
  American
• 0:50
  firms
• 0:50
  to
• 0:50
  do.
• 0:51
  Jason
• 0:52
  Kikta,
• 0:52
  a
• 0:52
  20-year
• 0:53
  Marine
• 0:53
  Corps
• 0:53
  veteran
• 0:54
  who
• 0:54
  helped
• 0:54
  stand
• 0:54
  up
• 0:54
  Cyber
• 0:55
  Command's
• 0:55
  Cyber
• 0:55
  National
• 0:55
  Mission
• 0:56
  Force,
• 0:56
  offered
• 0:57
  the
• 0:57
  most
• 0:57
  memorable
• 0:58
  prediction:
• 0:59
  the
• 0:59
  pilot
• 0:59
  itself
• 0:59
  will
• 0:59
  be
• 0:59
  "smooth
• 1:00
  as
• 1:00
  butter,"
• 1:01
  then
• 1:01
  scale
• 1:01
  will
• 1:01
  arrive
• 1:02
  and
• 1:02
  "it'll
• 1:02
  go
• 1:02
  to
• 1:02
  hell
• 1:02
  in
• 1:02
  a
• 1:02
  hand
• 1:03
  basket."
• 1:04
  The
• 1:04
  SolarWinds
• 1:04
  operation,
• 1:05
  he
• 1:05
  noted,
• 1:05
  was
• 1:05
  "pretty
• 1:05
  much
• 1:06
  all
• 1:06
  access
• 1:06
  generation."
• 1:07
  So
• 1:07
  was
• 1:08
  Volt
• 1:08
  Typhoon's
• 1:08
  pre-positioning
• 1:09
  inside
• 1:09
  U.S.
• 1:10
  water
• 1:10
  and
• 1:10
  power
• 1:10
  networks.
• 1:11
  Alex
• 1:12
  Orleans,
• 1:12
  a
• 1:12
  CrowdStrike
• 1:13
  and
• 1:13
  Cisco
• 1:13
  veteran,
• 1:14
  cut
• 1:14
  to
• 1:14
  the
• 1:14
  geopolitical
• 1:15
  absurdity:
• 1:16
  "This
• 1:16
  is
• 1:16
  your
• 1:16
  brain
• 1:16
  on
• 1:16
  American
• 1:17
  exceptionalism"
• 1:17
  —
• 1:18
  the
• 1:18
  U.S.
• 1:18
  would
• 1:19
  assert
• 1:19
  its
• 1:19
  contractors
• 1:20
  are
• 1:20
  untouchable
• 1:20
  while
• 1:20
  continuing
• 1:21
  to
• 1:21
  burn
• 1:21
  adversary
• 1:22
  contractor
• 1:22
  infrastructure
• 1:23
  built
• 1:23
  on
• 1:23
  compromised
• 1:24
  home
• 1:24
  routers.
• 1:25
  While
• 1:25
  Washington
• 1:25
  debates
• 1:26
  whether
• 1:26
  to
• 1:26
  deputize
• 1:27
  hackers,
• 1:27
  the
• 1:27
  open-source
• 1:28
  community
• 1:28
  is
• 1:29
  already
• 1:29
  demonstrating
• 1:30
  what
• 1:30
  autonomous
• 1:30
  offensive
• 1:31
  security
• 1:31
  looks
• 1:31
  like.
• 1:32
  T3MP3ST,
• 1:35
  a
• 1:35
  multi-agent
• 1:35
  framework
• 1:36
  released
• 1:36
  under
• 1:36
  AGPL-3.0,
• 1:39
  turns
• 1:39
  an
• 1:39
  AI
• 1:39
  coding
• 1:40
  agent
• 1:40
  you
• 1:40
  already
• 1:40
  run
• 1:41
  —
• 1:41
  Claude
• 1:41
  Code,
• 1:42
  Codex,
• 1:42
  Hermes,
• 1:43
  or
• 1:43
  a
• 1:43
  fully
• 1:43
  offline
• 1:44
  model
• 1:44
  via
• 1:44
  Ollama
• 1:45
  —
• 1:45
  into
• 1:46
  a
• 1:46
  zero-day
• 1:46
  hunter.
• 1:47
  It
• 1:47
  claims
• 1:48
  90.1%
• 1:49
  pass@1
• 1:50
  on
• 1:50
  XBOW's
• 1:51
  own
• 1:51
  104-challenge
• 1:52
  black-box
• 1:53
  suite
• 1:53
  (above
• 1:54
  XBOW's
• 1:54
  self-reported
• 1:55
  85%),
• 1:56
  23/40
• 1:57
  hint-free
• 1:57
  on
• 1:57
  Cybench,
• 1:58
  and
• 1:58
  8/10
• 1:59
  exact
• 1:59
  file/line/CWE
• 2:02
  hits
• 2:02
  on
• 2:02
  held-out
• 2:02
  post-cutoff
• 2:03
  CVEs
• 2:03
  the
• 2:04
  model
• 2:04
  had
• 2:04
  never
• 2:04
  seen.
• 2:05
  Every
• 2:06
  headline
• 2:06
  number
• 2:06
  recomputes
• 2:07
  from
• 2:07
  committed
• 2:07
  data
• 2:08
  with
• 2:08
  npm
• 2:08
  run
• 2:08
  verify-claims.
• 2:10
  The
• 2:10
  recon
• 2:10
  engine
• 2:10
  and
• 2:11
  war
• 2:11
  room
• 2:11
  are
• 2:11
  stable;
• 2:12
  the
• 2:12
  8-operator
• 2:13
  kill
• 2:13
  chain's
• 2:13
  downstream
• 2:14
  swarm
• 2:14
  coordination
• 2:15
  remains
• 2:15
  experimental
• 2:16
  and
• 2:16
  honestly
• 2:16
  labeled
• 2:17
  as
• 2:17
  such.
• 2:18
  The
• 2:18
  framework's
• 2:18
  coordinated-disclosure
• 2:19
  pipeline
• 2:20
  is
• 2:20
  live
• 2:20
  enough
• 2:21
  to
• 2:21
  have
• 2:21
  drafts
• 2:21
  held
• 2:21
  for
• 2:21
  vendor
• 2:22
  coordination
• 2:22
  right
• 2:22
  now.
• 2:26
  A
• 2:26
  pair
• 2:26
  of
• 2:26
  security
• 2:26
  researchers,
• 2:27
  Sam
• 2:27
  Curry
• 2:27
  and
• 2:28
  Maik
• 2:28
  Robert,
• 2:28
  discovered
• 2:29
  that
• 2:29
  the
• 2:29
  San
• 2:29
  Francisco
• 2:30
  Police
• 2:30
  Department
• 2:30
  was
• 2:31
  accidentally
• 2:31
  livestreaming
• 2:32
  real-time
• 2:32
  footage
• 2:33
  from
• 2:33
  five
• 2:33
  Skydio
• 2:33
  surveillance
• 2:34
  drones
• 2:34
  —
• 2:35
  including
• 2:35
  color
• 2:35
  and
• 2:35
  thermal
• 2:36
  imaging,
• 2:36
  GPS
• 2:37
  telemetry,
• 2:37
  and
• 2:38
  drone
• 2:38
  pilots'
• 2:38
  names
• 2:38
  —
• 2:39
  to
• 2:39
  anyone
• 2:40
  who
• 2:40
  found
• 2:40
  a
• 2:40
  public,
• 2:41
  unauthenticated
• 2:41
  web
• 2:41
  address.
• 2:43
  The
• 2:43
  leak
• 2:43
  wasn't
• 2:43
  a
• 2:43
  Skydio
• 2:44
  vulnerability.
• 2:45
  Someone
• 2:45
  inside
• 2:46
  the
• 2:46
  SFPD
• 2:46
  created
• 2:47
  a
• 2:47
  "ReadyLink"
• 2:47
  shareable
• 2:48
  URL
• 2:48
  last
• 2:49
  December
• 2:49
  with
• 2:49
  no
• 2:50
  authentication
• 2:50
  requirement
• 2:51
  and
• 2:51
  a
• 2:51
  one-year
• 2:51
  expiration,
• 2:52
  then
• 2:53
  somehow
• 2:53
  that
• 2:53
  link
• 2:53
  ended
• 2:54
  up
• 2:54
  indexed
• 2:54
  in
• 2:54
  AlienVault
• 2:55
  Open
• 2:55
  Threat
• 2:55
  Exchange,
• 2:56
  where
• 2:56
  the
• 2:56
  researchers
• 2:57
  found
• 2:57
  it
• 2:57
  using
• 2:57
  routine
• 2:58
  URL
• 2:58
  enumeration.
• 3:00
  The
• 3:00
  footage
• 3:00
  had
• 3:00
  potentially
• 3:01
  been
• 3:01
  exposed
• 3:01
  for
• 3:02
  six
• 3:02
  months
• 3:02
  before
• 3:02
  discovery.
• 3:03
  The
• 3:03
  archived
• 3:04
  footage
• 3:04
  —
• 3:05
  60
• 3:05
  videos
• 3:05
  across
• 3:06
  20
• 3:06
  flights
• 3:06
  over
• 3:07
  roughly
• 3:07
  48
• 3:07
  hours
• 3:07
  —
• 3:08
  paints
• 3:08
  an
• 3:09
  uncomfortable
• 3:09
  picture
• 3:09
  of
• 3:09
  modern
• 3:10
  aerial
• 3:10
  policing.
• 3:11
  WIRED's
• 3:12
  analysis
• 3:12
  detected
• 3:12
  hundreds
• 3:13
  of
• 3:13
  people
• 3:13
  and
• 3:13
  dozens
• 3:14
  of
• 3:14
  clear
• 3:14
  faces
• 3:14
  across
• 3:15
  the
• 3:15
  videos.
• 3:16
  In
• 3:16
  one
• 3:17
  sequence,
• 3:17
  four
• 3:18
  Skydio
• 3:18
  quadcopters
• 3:19
  tracked
• 3:19
  a
• 3:19
  man
• 3:19
  across
• 3:20
  San
• 3:20
  Francisco
• 3:20
  for
• 3:21
  an
• 3:21
  alleged
• 3:21
  car-parts
• 3:21
  theft,
• 3:22
  capturing
• 3:22
  his
• 3:23
  license
• 3:23
  plate,
• 3:23
  his
• 3:24
  hiding
• 3:24
  spots
• 3:24
  behind
• 3:25
  parked
• 3:25
  cars,
• 3:25
  and
• 3:25
  the
• 3:25
  police
• 3:26
  tackle.
• 3:27
  In
• 3:27
  another,
• 3:28
  a
• 3:28
  drone
• 3:28
  hovered
• 3:28
  over
• 3:28
  a
• 3:28
  rooftop,
• 3:29
  zooming
• 3:30
  in
• 3:30
  on
• 3:30
  a
• 3:30
  person
• 3:30
  sitting
• 3:31
  alone
• 3:31
  wearing
• 3:31
  headphones
• 3:31
  —
• 3:32
  the
• 3:32
  response
• 3:33
  to
• 3:33
  a
• 3:33
  reported
• 3:33
  "prowler."
• 3:34
  As
• 3:35
  ACLU
• 3:35
  senior
• 3:35
  policy
• 3:36
  analyst
• 3:36
  Jay
• 3:36
  Stanley
• 3:37
  put
• 3:37
  it,
• 3:37
  drone
• 3:38
  footage
• 3:38
  is
• 3:38
  a
• 3:38
  "toxic
• 3:39
  asset"
• 3:40
  always
• 3:40
  one
• 3:40
  misconfiguration
• 3:41
  from
• 3:41
  public
• 3:42
  exposure,
• 3:42
  and
• 3:43
  AI
• 3:43
  makes
• 3:43
  the
• 3:43
  privacy
• 3:44
  implications
• 3:44
  exponentially
• 3:45
  worse:
• 3:46
  "Maybe
• 3:46
  no
• 3:46
  human
• 3:47
  has
• 3:47
  the
• 3:47
  time
• 3:47
  to
• 3:47
  pore
• 3:47
  over
• 3:48
  every
• 3:48
  frame
• 3:48
  of
• 3:48
  these
• 3:48
  videos,
• 3:49
  but
• 3:49
  an
• 3:49
  AI
• 3:50
  can
• 3:50
  do
• 3:50
  that,
• 3:50
  and
• 3:51
  it's
• 3:51
  scalable
• 3:51
  to
• 3:51
  vast
• 3:52
  amounts
• 3:52
  of
• 3:52
  video."
• 3:53
  Notably,
• 3:54
  SFPD's
• 3:55
  own
• 3:55
  fleet
• 3:55
  has
• 3:56
  exploded
• 3:56
  from
• 3:56
  6
• 3:57
  drones
• 3:57
  to
• 3:57
  98
• 3:58
  since
• 3:58
  the
• 3:58
  program's
• 3:59
  2024
• 3:59
  launch,
• 4:00
  with
• 4:00
  over
• 4:00
  1,400
• 4:02
  flights
• 4:02
  logged.
• 4:03
  On
• 4:03
  the
• 4:03
  data-privacy
• 4:04
  front,
• 4:05
  the
• 4:05
  ongoing
• 4:05
  OpenAI,
• 4:06
  Inc.
• 4:07
  Copyright
• 4:08
  Infringement
• 4:08
  Litigation
• 4:09
  has
• 4:09
  entered
• 4:09
  its
• 4:09
  ugliest
• 4:10
  phase.
• 4:11
  Last
• 4:11
  December,
• 4:12
  OpenAI
• 4:12
  made
• 4:12
  20
• 4:13
  million
• 4:13
  "de-identified"
• 4:14
  ChatGPT
• 4:15
  user
• 4:15
  conversation
• 4:16
  logs
• 4:16
  available
• 4:17
  to
• 4:17
  plaintiffs
• 4:17
  for
• 4:17
  keyword
• 4:18
  searching
• 4:18
  —
• 4:19
  already
• 4:19
  a
• 4:19
  cramp
• 4:20
  for
• 4:20
  anyone
• 4:20
  who
• 4:20
  assumed
• 4:21
  chatbot
• 4:21
  conversations
• 4:22
  were
• 4:22
  sacrosanct.
• 4:23
  But
• 4:23
  the
• 4:23
  plaintiffs
• 4:24
  now
• 4:24
  allege
• 4:24
  OpenAI
• 4:25
  applied
• 4:25
  approximately
• 4:26
  19
• 4:26
  billion
• 4:26
  redactions
• 4:27
  across
• 4:27
  those
• 4:27
  logs
• 4:28
  (roughly
• 4:28
  1,000
• 4:28
  per
• 4:29
  conversation),
• 4:30
  under-represented
• 4:30
  conversations
• 4:31
  using
• 4:31
  Retrieval
• 4:32
  Augmented
• 4:32
  Generation,
• 4:33
  and
• 4:33
  failed
• 4:34
  to
• 4:34
  retain
• 4:34
  large
• 4:34
  numbers
• 4:35
  of
• 4:35
  conversations
• 4:35
  entirely
• 4:35
  —
• 4:36
  including
• 4:37
  those
• 4:37
  from
• 4:37
  ChatGPT's
• 4:38
  "Temporary
• 4:39
  Chat"
• 4:39
  feature.
• 4:40
  On
• 4:40
  July
• 4:40
  9,
• 4:41
  the
• 4:41
  plaintiffs
• 4:41
  moved
• 4:42
  for
• 4:42
  sanctions,
• 4:43
  asking
• 4:43
  the
• 4:43
  court
• 4:43
  to
• 4:43
  prohibit
• 4:44
  OpenAI
• 4:44
  from
• 4:44
  using
• 4:45
  the
• 4:45
  produced
• 4:45
  logs
• 4:45
  for
• 4:46
  its
• 4:46
  defense,
• 4:47
  establish
• 4:47
  as
• 4:47
  fact
• 4:47
  that
• 4:47
  copyrighted
• 4:48
  material
• 4:48
  was
• 4:49
  "substantially
• 4:50
  and
• 4:50
  systematically"
• 4:51
  used,
• 4:51
  and
• 4:52
  inform
• 4:52
  the
• 4:52
  jury
• 4:52
  about
• 4:52
  deleted
• 4:53
  conversations.
• 4:56
  Salvatore
• 4:56
  Sanfilippo
• 4:57
  —
• 4:57
  antirez,
• 4:58
  the
• 4:58
  creator
• 4:59
  of
• 4:59
  Redis
• 4:59
  and
• 4:59
  a
• 4:59
  programmer
• 5:00
  whose
• 5:00
  credentials
• 5:00
  are
• 5:01
  beyond
• 5:01
  questioning
• 5:01
  —
• 5:02
  published
• 5:02
  a
• 5:02
  direct
• 5:02
  challenge
• 5:03
  to
• 5:03
  the
• 5:03
  remaining
• 5:03
  holdouts:
• 5:04
  stop
• 5:04
  reading
• 5:04
  the
• 5:04
  code.
• 5:05
  His
• 5:05
  argument
• 5:05
  is
• 5:05
  not
• 5:06
  a
• 5:06
  vibe-coding
• 5:06
  manifesto.
• 5:07
  It
• 5:07
  is
• 5:07
  a
• 5:07
  prioritization
• 5:08
  claim
• 5:08
  grounded
• 5:08
  in
• 5:08
  the
• 5:08
  economics
• 5:09
  of
• 5:09
  the
• 5:09
  8-hour
• 5:09
  workday.
• 5:10
  LLMs
• 5:10
  produce
• 5:11
  locally
• 5:11
  optimal
• 5:11
  code
• 5:11
  reliably;
• 5:12
  they
• 5:12
  struggle
• 5:12
  with
• 5:12
  big-picture
• 5:13
  design.
• 5:14
  Reviewing
• 5:14
  5,000
• 5:15
  lines
• 5:15
  of
• 5:15
  generated
• 5:16
  code
• 5:16
  daily
• 5:16
  is
• 5:16
  a
• 5:16
  tradeoff
• 5:17
  that
• 5:17
  steals
• 5:17
  time
• 5:17
  from
• 5:17
  design
• 5:18
  thinking,
• 5:18
  QA,
• 5:18
  and
• 5:19
  feature
• 5:19
  direction
• 5:19
  —
• 5:19
  the
• 5:20
  activities
• 5:20
  that
• 5:20
  now
• 5:20
  constitute
• 5:21
  the
• 5:21
  highest-leverage
• 5:21
  part
• 5:22
  of
• 5:22
  a
• 5:22
  programmer's
• 5:22
  job.
• 5:23
  "If
• 5:23
  you
• 5:23
  control
• 5:23
  the
• 5:23
  ideas
• 5:23
  of
• 5:24
  your
• 5:24
  software,
• 5:24
  looking
• 5:24
  at
• 5:24
  the
• 5:24
  code
• 5:25
  itself
• 5:25
  is
• 5:25
  suboptimal
• 5:25
  and
• 5:26
  often
• 5:26
  pointless,"
• 5:26
  he
• 5:27
  writes.
• 5:27
  His
• 5:27
  concrete
• 5:28
  recommendation:
• 5:29
  invest
• 5:29
  in
• 5:29
  DESIGN.md
• 5:30
  files
• 5:31
  where
• 5:31
  each
• 5:31
  data
• 5:31
  structure
• 5:32
  is
• 5:32
  described
• 5:32
  in
• 5:32
  human
• 5:32
  language,
• 5:33
  with
• 5:33
  implementation
• 5:34
  tricks
• 5:34
  and
• 5:34
  design
• 5:34
  rationale.
• 5:35
  "You
• 5:36
  open
• 5:36
  the
• 5:36
  file,
• 5:37
  read
• 5:37
  the
• 5:37
  design,
• 5:38
  then
• 5:38
  you
• 5:38
  own
• 5:38
  the
• 5:38
  ideas."
• 5:39
  The
• 5:39
  one
• 5:39
  caveat
• 5:40
  he
• 5:40
  offers
• 5:40
  is
• 5:41
  for
• 5:41
  junior
• 5:41
  programmers,
• 5:42
  who
• 5:42
  should
• 5:42
  still
• 5:42
  learn
• 5:42
  to
• 5:42
  write
• 5:43
  programs
• 5:43
  —
• 5:43
  but
• 5:44
  by
• 5:44
  implementing
• 5:44
  small
• 5:44
  interpreters
• 5:45
  and
• 5:45
  hash
• 5:45
  tables,
• 5:46
  not
• 5:46
  reviewing
• 5:47
  JavaScript
• 5:47
  for
• 5:47
  web
• 5:47
  clients.
• 5:48
  The
• 5:48
  tooling
• 5:49
  is
• 5:49
  catching
• 5:49
  up
• 5:49
  to
• 5:49
  antirez's
• 5:50
  thesis.
• 5:51
  Omnigent,
• 5:52
  a
• 5:52
  new
• 5:52
  open-source
• 5:52
  meta-harness,
• 5:53
  makes
• 5:54
  multi-agent
• 5:54
  orchestration
• 5:55
  trivial:
• 5:55
  one
• 5:56
  session
• 5:56
  drives
• 5:56
  Claude
• 5:56
  Code,
• 5:57
  Codex,
• 5:57
  and
• 5:57
  other
• 5:58
  agents
• 5:58
  simultaneously,
• 5:59
  with
• 5:59
  guardrails,
• 6:00
  session
• 6:00
  sharing,
• 6:01
  and
• 6:01
  custom
• 6:01
  orchestrators
• 6:02
  built
• 6:02
  on
• 6:02
  top.
• 6:03
  The
• 6:03
  emerging
• 6:03
  pattern
• 6:04
  that
• 6:04
  practitioners
• 6:04
  report
• 6:05
  getting
• 6:05
  "insane
• 6:06
  value"
• 6:06
  from
• 6:06
  is
• 6:07
  Claude
• 6:07
  Code
• 6:07
  for
• 6:07
  implementation,
• 6:08
  then
• 6:09
  Codex
• 6:09
  to
• 6:09
  review
• 6:10
  the
• 6:10
  diff
• 6:10
  in
• 6:10
  a
• 6:10
  separate
• 6:10
  session
• 6:11
  —
• 6:11
  different
• 6:12
  models,
• 6:12
  different
• 6:12
  context
• 6:13
  windows,
• 6:14
  the
• 6:14
  model
• 6:14
  that
• 6:14
  wrote
• 6:14
  the
• 6:14
  code
• 6:15
  never
• 6:15
  grades
• 6:15
  its
• 6:15
  own
• 6:15
  output.
• 6:16
  A
• 6:17
  detailed
• 6:17
  guide
• 6:17
  to
• 6:17
  building
• 6:18
  an
• 6:18
  "Agentic
• 6:18
  OS"
• 6:19
  around
• 6:19
  Claude
• 6:19
  Fable
• 6:19
  5
• 6:20
  lays
• 6:20
  out
• 6:20
  nine
• 6:20
  architectural
• 6:21
  layers,
• 6:22
  from
• 6:22
  hard
• 6:22
  rules
• 6:22
  the
• 6:23
  model
• 6:23
  cannot
• 6:23
  argue
• 6:23
  with
• 6:24
  (because
• 6:24
  a
• 6:24
  script
• 6:24
  can
• 6:24
  verify
• 6:25
  every
• 6:25
  one)
• 6:25
  to
• 6:25
  an
• 6:26
  earned-trust
• 6:26
  system
• 6:27
  that
• 6:27
  grants
• 6:27
  autonomy
• 6:28
  only
• 6:28
  after
• 6:28
  20
• 6:28
  logged
• 6:29
  runs
• 6:29
  at
• 6:29
  95%
• 6:30
  verified
• 6:30
  pass
• 6:31
  rate.
• 6:31
  The
• 6:31
  central
• 6:32
  insight:
• 6:32
  Fable
• 6:33
  5
• 6:33
  can
• 6:33
  cost
• 6:34
  10–50x
• 6:35
  more
• 6:35
  than
• 6:35
  it
• 6:35
  should
• 6:35
  depending
• 6:36
  on
• 6:36
  architecture,
• 6:36
  not
• 6:37
  workload.
• 6:37
  Let
• 6:38
  it
• 6:38
  split
• 6:38
  a
• 6:38
  job
• 6:38
  across
• 6:39
  ten
• 6:39
  parallel
• 6:39
  copies
• 6:39
  each
• 6:40
  re-reading
• 6:40
  the
• 6:40
  same
• 6:40
  material
• 6:41
  and
• 6:41
  a
• 6:41
  single
• 6:41
  day
• 6:41
  burns
• 6:42
  $400–600;
• 6:44
  route
• 6:44
  decisions
• 6:44
  to
• 6:45
  Fable
• 6:45
  5
• 6:45
  and
• 6:45
  execution
• 6:46
  to
• 6:46
  cheaper
• 6:46
  models
• 6:47
  and
• 6:47
  the
• 6:47
  same
• 6:47
  day
• 6:47
  costs
• 6:47
  a
• 6:47
  few
• 6:48
  dollars.
• 6:48
  The
• 6:49
  model
• 6:49
  that
• 6:49
  does
• 6:49
  the
• 6:49
  work
• 6:49
  cannot
• 6:50
  also
• 6:50
  judge
• 6:50
  whether
• 6:51
  the
• 6:51
  work
• 6:51
  is
• 6:51
  done
• 6:51
  —
• 6:52
  "something
• 6:52
  independent
• 6:52
  has
• 6:53
  to
• 6:53
  check."
• 6:53
  Daniel
• 6:54
  Miessler's
• 6:54
  viral
• 6:55
  LinkedIn
• 6:55
  post
• 6:55
  sharpens
• 6:56
  the
• 6:56
  strategic
• 6:56
  angle:
• 6:57
  Meta
• 6:57
  spent
• 6:58
  billions
• 6:58
  on
• 6:58
  a
• 6:58
  new
• 6:58
  AI
• 6:59
  research
• 6:59
  lab
• 6:59
  that's
• 6:59
  already
• 7:00
  "worse
• 7:00
  than
• 7:00
  dead,"
• 7:01
  Microsoft's
• 7:02
  relationship
• 7:02
  with
• 7:02
  OpenAI
• 7:03
  is
• 7:03
  "completely
• 7:04
  adversarial,"
• 7:04
  and
• 7:05
  the
• 7:05
  only
• 7:05
  company
• 7:05
  he
• 7:05
  sees
• 7:06
  in
• 7:06
  a
• 7:06
  strong
• 7:06
  position
• 7:06
  is
• 7:07
  Anthropic
• 7:07
  —
• 7:08
  vision
• 7:08
  plus
• 7:08
  business
• 7:09
  discipline,
• 7:09
  already
• 7:10
  profitable.
• 7:11
  The
• 7:11
  ultimate
• 7:11
  AI
• 7:12
  prompt
• 7:12
  for
• 7:12
  any
• 7:12
  business,
• 7:13
  he
• 7:13
  argues,
• 7:13
  is:
• 7:14
  "What
• 7:14
  is
• 7:15
  the
• 7:15
  specific
• 7:15
  problem
• 7:16
  that
• 7:16
  we
• 7:16
  have
• 7:16
  as
• 7:16
  a
• 7:16
  company,
• 7:17
  and
• 7:17
  why
• 7:17
  do
• 7:17
  we
• 7:17
  think
• 7:18
  AI
• 7:18
  can
• 7:18
  help
• 7:18
  us
• 7:19
  solve
• 7:19
  it?"
• 7:20
  For
• 7:20
  too
• 7:20
  many,
• 7:21
  the
• 7:21
  answer
• 7:21
  is
• 7:21
  circular:
• 7:22
  "The
• 7:22
  primary
• 7:23
  problem
• 7:23
  we
• 7:23
  have
• 7:24
  is
• 7:24
  that
• 7:24
  leadership
• 7:24
  wants
• 7:24
  us
• 7:25
  to
• 7:25
  use
• 7:25
  AI."
• 7:26
  One
• 7:26
  commenter
• 7:27
  reframed
• 7:27
  it
• 7:27
  through
• 7:27
  a
• 7:27
  security
• 7:28
  lens
• 7:28
  that
• 7:28
  hits
• 7:29
  home:
• 7:29
  "What
• 7:30
  can
• 7:30
  an
• 7:30
  adversary
• 7:30
  do
• 7:31
  with
• 7:31
  the
• 7:31
  AI
• 7:31
  you
• 7:31
  just
• 7:32
  deployed
• 7:32
  that
• 7:32
  they
• 7:32
  couldn't
• 7:32
  do
• 7:33
  before?
• 7:34
  Most
• 7:34
  organizations
• 7:35
  can't
• 7:35
  answer
• 7:35
  that
• 7:36
  because
• 7:36
  they
• 7:36
  never
• 7:36
  answered
• 7:37
  the
• 7:37
  question
• 7:37
  you
• 7:37
  pose
• 7:38
  either."
• 7:40
  A
• 7:41
  detailed
• 7:41
  benchmark
• 7:41
  from
• 7:41
  a
• 7:42
  user
• 7:42
  running
• 7:42
  Qwen3.6-27B
• 7:45
  on
• 7:45
  an
• 7:45
  RTX
• 7:46
  PRO
• 7:46
  6000
• 7:46
  Blackwell
• 7:47
  workstation
• 7:47
  reveals
• 7:48
  a
• 7:48
  painful
• 7:48
  truth
• 7:48
  that
• 7:49
  academic
• 7:49
  quantization
• 7:50
  metrics
• 7:50
  hide:
• 7:50
  low-bit
• 7:51
  quantization
• 7:52
  destroys
• 7:52
  agentic
• 7:52
  reliability.
• 7:54
  This
• 7:54
  is
• 7:54
  a
• 7:54
  pattern
• 7:54
  that
• 7:54
  keeps
• 7:55
  surfacing
• 7:55
  —
• 7:56
  small
• 7:56
  logit
• 7:57
  errors
• 7:57
  that
• 7:57
  are
• 7:57
  invisible
• 7:58
  in
• 7:58
  single-turn
• 7:58
  benchmarks
• 7:59
  compound
• 7:59
  catastrophically
• 8:00
  over
• 8:00
  multi-turn
• 8:01
  tool-calling
• 8:01
  sequences,
• 8:02
  because
• 8:02
  each
• 8:03
  degraded
• 8:03
  decision
• 8:03
  feeds
• 8:04
  into
• 8:04
  the
• 8:04
  next.
• 8:05
  In
• 8:06
  BF16,
• 8:06
  the
• 8:07
  model
• 8:07
  works
• 8:07
  flawlessly
• 8:08
  across
• 8:08
  OpenCode,
• 8:09
  VS
• 8:09
  Code
• 8:09
  Copilot,
• 8:10
  and
• 8:10
  multi-turn
• 8:11
  agent
• 8:11
  loops.
• 8:12
  Switch
• 8:12
  to
• 8:12
  NVIDIA's
• 8:13
  official
• 8:13
  NVFP4
• 8:14
  checkpoint
• 8:15
  —
• 8:16
  same
• 8:16
  hardware,
• 8:17
  same
• 8:17
  config
• 8:17
  —
• 8:18
  and
• 8:18
  the
• 8:18
  model
• 8:18
  halts
• 8:19
  mid-task,
• 8:19
  gets
• 8:20
  stuck
• 8:20
  in
• 8:20
  failure
• 8:20
  loops
• 8:21
  repeating
• 8:21
  the
• 8:21
  same
• 8:21
  phrases
• 8:22
  until
• 8:22
  context
• 8:22
  exhaustion,
• 8:23
  and
• 8:23
  requires
• 8:24
  manual
• 8:24
  "continue"
• 8:25
  prompts
• 8:25
  that
• 8:25
  never
• 8:25
  occur
• 8:26
  in
• 8:26
  BF16.
• 8:27
  FP8
• 8:28
  shows
• 8:28
  the
• 8:28
  same
• 8:28
  symptoms
• 8:29
  less
• 8:29
  frequently.
• 8:30
  Increasing
• 8:30
  repetition
• 8:31
  penalty
• 8:31
  doesn't
• 8:31
  break
• 8:32
  the
• 8:32
  loops;
• 8:32
  it
• 8:33
  just
• 8:33
  makes
• 8:33
  the
• 8:33
  model
• 8:33
  alternate
• 8:34
  between
• 8:34
  two
• 8:34
  failure
• 8:35
  phrases
• 8:35
  instead
• 8:35
  of
• 8:35
  one.
• 8:36
  Community
• 8:37
  consensus
• 8:37
  is
• 8:37
  blunt:
• 8:38
  "quants
• 8:38
  don't
• 8:39
  perform
• 8:39
  as
• 8:39
  well
• 8:39
  as
• 8:39
  native
• 8:40
  precision,
• 8:40
  period."
• 8:41
  Multiple
• 8:42
  users
• 8:42
  report
• 8:42
  that
• 8:43
  NVFP4's
• 8:44
  true
• 8:45
  quality
• 8:45
  lands
• 8:45
  between
• 8:45
  Q3
• 8:46
  and
• 8:46
  Q4
• 8:46
  despite
• 8:47
  its
• 8:47
  larger
• 8:47
  file
• 8:47
  size,
• 8:48
  and
• 8:48
  that
• 8:49
  chat
• 8:49
  template
• 8:49
  bugs
• 8:49
  in
• 8:50
  Qwen3.6's
• 8:52
  launch
• 8:52
  may
• 8:52
  compound
• 8:53
  the
• 8:53
  degradation.
• 8:54
  NVIDIA's
• 8:55
  own
• 8:55
  Puzzle-75B-A9B
• 8:58
  demonstrates
• 8:58
  the
• 8:58
  other
• 8:58
  side
• 8:59
  of
• 8:59
  the
• 8:59
  local
• 8:59
  inference
• 9:00
  equation
• 9:00
  —
• 9:01
  the
• 9:01
  sweet
• 9:01
  spot
• 9:02
  nobody
• 9:02
  else
• 9:02
  is
• 9:02
  filling.
• 9:04
  This
• 9:04
  75B-total,
• 9:05
  9B-active
• 9:06
  MoE
• 9:06
  model
• 9:06
  runs
• 9:07
  at
• 9:07
  132
• 9:08
  tokens/second
• 9:09
  on
• 9:09
  three
• 9:09
  3090s
• 9:10
  capped
• 9:10
  at
• 9:10
  200W
• 9:11
  each,
• 9:12
  with
• 9:12
  hybrid
• 9:12
  Mamba
• 9:13
  keeping
• 9:13
  the
• 9:13
  KV
• 9:14
  cache
• 9:14
  tiny
• 9:14
  enough
• 9:14
  for
• 9:15
  three
• 9:15
  concurrent
• 9:15
  256K-context
• 9:17
  sessions.
• 9:18
  It
• 9:18
  replaced
• 9:18
  a
• 9:18
  Nemotron
• 9:19
  Super
• 9:19
  120B
• 9:20
  MoE
• 9:21
  that
• 9:21
  consumed
• 9:21
  four
• 9:22
  GPUs,
• 9:23
  roughly
• 9:23
  doubling
• 9:23
  speed
• 9:24
  per
• 9:24
  watt
• 9:24
  while
• 9:24
  freeing
• 9:25
  a
• 9:25
  card.
• 9:26
  The
• 9:26
  user
• 9:26
  notes
• 9:26
  the
• 9:27
  70–80B
• 9:28
  total
• 9:29
  /
• 9:29
  ~10B
• 9:29
  active
• 9:30
  band
• 9:30
  fills
• 9:30
  72GB
• 9:31
  of
• 9:31
  quantized
• 9:32
  VRAM
• 9:32
  exactly
• 9:32
  —
• 9:33
  dense-class
• 9:34
  quality
• 9:34
  at
• 9:34
  A3B-class
• 9:35
  speed
• 9:35
  —
• 9:36
  yet
• 9:37
  almost
• 9:37
  nobody
• 9:37
  ships
• 9:37
  models
• 9:38
  in
• 9:38
  this
• 9:38
  range.
• 9:39
  Benchmark
• 9:39
  tables
• 9:40
  show
• 9:40
  Qwen3.6-27B
• 9:43
  still
• 9:44
  beats
• 9:44
  Puzzle
• 9:44
  on
• 9:44
  most
• 9:45
  academic
• 9:45
  metrics,
• 9:46
  but
• 9:46
  real-world
• 9:47
  in-context
• 9:47
  grounding
• 9:48
  tells
• 9:48
  a
• 9:48
  different
• 9:48
  story:
• 9:49
  one
• 9:49
  user
• 9:50
  measured
• 9:50
  cold
• 9:50
  accuracy
• 9:51
  at
• 9:51
  0.26
• 9:52
  that
• 9:52
  jumped
• 9:52
  to
• 9:53
  0.97
• 9:54
  with
• 9:54
  a
• 9:54
  4.6K-token
• 9:55
  reference
• 9:56
  document
• 9:56
  in
• 9:56
  the
• 9:56
  prompt.
• 9:57
  For
• 9:58
  those
• 9:58
  wondering
• 9:58
  whether
• 9:59
  FlashAttention-3
• 10:00
  or
• 10:00
  FA-4
• 10:00
  optimizations
• 10:01
  might
• 10:01
  unlock
• 10:02
  more
• 10:02
  performance
• 10:02
  on
• 10:02
  consumer
• 10:03
  silicon,
• 10:04
  a
• 10:04
  thorough
• 10:04
  kernel-level
• 10:05
  investigation
• 10:05
  confirms:
• 10:07
  FA-2
• 10:07
  is
• 10:07
  the
• 10:07
  ceiling.
• 10:08
  The
• 10:08
  primary
• 10:09
  lever
• 10:09
  behind
• 10:10
  FA-3
• 10:10
  —
• 10:11
  faster
• 10:11
  WGMMA
• 10:12
  tensor-core
• 10:13
  instructions
• 10:13
  —
• 10:14
  simply
• 10:15
  doesn't
• 10:15
  exist
• 10:15
  on
• 10:15
  RTX
• 10:16
  GPUs.
• 10:17
  TMA
• 10:18
  (tensor
• 10:18
  memory
• 10:18
  accelerator)
• 10:19
  is
• 10:19
  available
• 10:20
  on
• 10:20
  SM
• 10:20
  120
• 10:21
  (RTX
• 10:22
  50XX)
• 10:23
  but
• 10:23
  isn't
• 10:24
  the
• 10:24
  bottleneck,
• 10:25
  and
• 10:25
  warp
• 10:25
  specialization
• 10:26
  without
• 10:26
  asynchronous
• 10:27
  tensor
• 10:27
  cores
• 10:28
  actually
• 10:28
  makes
• 10:28
  things
• 10:29
  slightly
• 10:29
  worse
• 10:29
  (213μs
• 10:31
  vs.
• 10:32
  206μs).
• 10:33
  Since
• 10:34
  the
• 10:34
  pipeline
• 10:34
  is
• 10:35
  tensor-core-bound
• 10:35
  on
• 10:35
  consumer
• 10:36
  hardware,
• 10:37
  meaningful
• 10:37
  improvements
• 10:38
  will
• 10:38
  require
• 10:38
  sacrificing
• 10:39
  accuracy
• 10:40
  to
• 10:40
  leverage
• 10:40
  lower-precision
• 10:41
  tensor
• 10:41
  cores.
• 10:42
  On
• 10:42
  the
• 10:42
  audio
• 10:43
  inference
• 10:43
  front,
• 10:44
  audio.cpp
• 10:45
  continues
• 10:46
  expanding
• 10:46
  into
• 10:47
  the
• 10:47
  unified
• 10:47
  local
• 10:47
  audio
• 10:48
  framework
• 10:48
  the
• 10:48
  ecosystem
• 10:49
  has
• 10:49
  lacked.
• 10:50
  The
• 10:50
  latest
• 10:50
  release
• 10:51
  adds
• 10:51
  four
• 10:51
  ASR
• 10:51
  models
• 10:52
  —
• 10:52
  Nemotron
• 10:53
  3.5
• 10:54
  ASR,
• 10:54
  Higgs
• 10:55
  Audio
• 10:55
  STT,
• 10:56
  VibeVoice
• 10:56
  ASR,
• 10:57
  and
• 10:57
  Hviske
• 10:57
  —
• 10:58
  with
• 10:58
  initial
• 10:58
  streaming
• 10:58
  support.
• 10:59
  The
• 11:00
  headline:
• 11:00
  Nemotron
• 11:01
  ASR
• 11:01
  transcribed
• 11:02
  327.6
• 11:03
  seconds
• 11:04
  of
• 11:04
  audio
• 11:04
  in
• 11:05
  2.17
• 11:05
  seconds
• 11:06
  offline
• 11:07
  on
• 11:07
  an
• 11:07
  RTX
• 11:07
  5090
• 11:09
  at
• 11:09
  3.18%
• 11:10
  WER,
• 11:11
  with
• 11:11
  the
• 11:11
  streaming
• 11:11
  SSE
• 11:12
  path
• 11:12
  matching
• 11:12
  accuracy
• 11:13
  at
• 11:13
  307ms
• 11:14
  time-to-first-token
• 11:16
  and
• 11:16
  substantially
• 11:16
  lower
• 11:17
  VRAM.
• 11:18
  Meanwhile,
• 11:19
  Mesh
• 11:19
  LLM
• 11:19
  takes
• 11:20
  a
• 11:20
  different
• 11:20
  approach
• 11:20
  to
• 11:20
  the
• 11:20
  hardware
• 11:21
  problem:
• 11:21
  rather
• 11:22
  than
• 11:22
  optimizing
• 11:22
  single-node
• 11:23
  inference,
• 11:24
  it
• 11:24
  pools
• 11:24
  GPUs
• 11:25
  and
• 11:25
  memory
• 11:25
  across
• 11:26
  machines
• 11:26
  into
• 11:26
  a
• 11:26
  single
• 11:27
  OpenAI-compatible
• 11:28
  API
• 11:28
  endpoint.
• 11:29
  Models
• 11:29
  too
• 11:30
  large
• 11:30
  for
• 11:30
  one
• 11:30
  box
• 11:30
  run
• 11:31
  via
• 11:31
  "Skippy
• 11:31
  stage
• 11:32
  splits,"
• 11:33
  the
• 11:33
  coordinator
• 11:33
  plans
• 11:34
  contiguous
• 11:34
  layer
• 11:34
  ranges
• 11:35
  across
• 11:35
  peers,
• 11:36
  and
• 11:36
  a
• 11:36
  Mixture-of-Agents
• 11:37
  gateway
• 11:37
  can
• 11:37
  fan
• 11:38
  prompts
• 11:38
  to
• 11:38
  every
• 11:38
  model
• 11:39
  in
• 11:39
  the
• 11:39
  mesh
• 11:39
  simultaneously.
• 11:41
  The
• 11:41
  project
• 11:41
  supports
• 11:42
  72+
• 11:42
  model
• 11:43
  families
• 11:43
  and
• 11:43
  integrates
• 11:44
  directly
• 11:44
  with
• 11:44
  Claude
• 11:45
  Code,
• 11:45
  Goose,
• 11:46
  and
• 11:46
  OpenCode.
• 11:49
  The
• 11:49
  most
• 11:49
  technically
• 11:50
  ambitious
• 11:50
  paper
• 11:51
  this
• 11:51
  week
• 11:51
  comes
• 11:51
  from
• 11:51
  RightNow
• 11:52
  AI,
• 11:53
  and
• 11:53
  its
• 11:53
  title
• 11:53
  makes
• 11:54
  an
• 11:54
  audacious
• 11:54
  claim:
• 11:54
  an
• 11:55
  "AGI
• 11:55
  compiler."
• 11:56
  Strip
• 11:56
  away
• 11:56
  the
• 11:57
  provocation
• 11:57
  and
• 11:57
  the
• 11:58
  actual
• 11:58
  contribution
• 11:58
  is
• 11:59
  rigorous:
• 11:59
  AutoRatchet
• 12:00
  is
• 12:00
  a
• 12:00
  Rust
• 12:00
  workspace
• 12:01
  of
• 12:01
  16
• 12:01
  crates
• 12:02
  (36,843
• 12:03
  lines)
• 12:04
  that
• 12:04
  records
• 12:04
  live
• 12:05
  LLM
• 12:05
  agent
• 12:05
  behavior,
• 12:06
  measures
• 12:06
  which
• 12:07
  parts
• 12:07
  are
• 12:07
  deterministic,
• 12:08
  extracts
• 12:08
  them
• 12:08
  into
• 12:08
  verified
• 12:09
  programs
• 12:09
  or
• 12:10
  distilled
• 12:10
  specialists,
• 12:11
  and
• 12:11
  emits
• 12:11
  signed
• 12:11
  WebAssembly
• 12:12
  artifacts
• 12:12
  with
• 12:12
  physically
• 12:13
  enforced
• 12:13
  capability
• 12:14
  boundaries.
• 12:15
  The
• 12:15
  key
• 12:15
  finding
• 12:15
  on
• 12:15
  their
• 12:16
  AgentMark
• 12:16
  benchmark:
• 12:17
  87.1%
• 12:18
  of
• 12:18
  560
• 12:19
  recorded
• 12:19
  frontier-agent
• 12:20
  spans
• 12:21
  are
• 12:21
  witnessed-deterministic
• 12:22
  at
• 12:22
  default
• 12:22
  temperature
• 12:22
  —
• 12:23
  classification,
• 12:24
  extraction,
• 12:25
  and
• 12:25
  routing
• 12:25
  hit
• 12:25
  100%,
• 12:27
  while
• 12:27
  free-text
• 12:27
  summarization
• 12:28
  sits
• 12:28
  at
• 12:28
  17.5%.
• 12:30
  On
• 12:30
  a
• 12:30
  300-item
• 12:30
  production-style
• 12:31
  stream
• 12:31
  with
• 12:31
  three
• 12:32
  scheduled
• 12:32
  distribution
• 12:32
  shifts,
• 12:33
  the
• 12:33
  system
• 12:34
  drives
• 12:34
  marginal
• 12:34
  cost
• 12:35
  from
• 12:35
  59
• 12:35
  to
• 12:35
  2
• 12:35
  micro-dollars
• 12:36
  per
• 12:36
  item
• 12:36
  at
• 12:36
  96.9%
• 12:37
  parity
• 12:38
  on
• 12:38
  witnessed
• 12:38
  inputs
• 12:38
  with
• 12:38
  zero
• 12:39
  errors.
• 12:39
  What
• 12:40
  makes
• 12:40
  the
• 12:40
  paper
• 12:40
  worth
• 12:40
  reading
• 12:41
  in
• 12:41
  full
• 12:41
  is
• 12:41
  its
• 12:41
  equal
• 12:42
  emphasis
• 12:42
  on
• 12:42
  failure
• 12:43
  modes.
• 12:44
  Under
• 12:44
  loose
• 12:44
  guard
• 12:44
  calibration,
• 12:45
  48.9%
• 12:46
  of
• 12:46
  compiled
• 12:47
  answers
• 12:47
  silently
• 12:47
  disagree
• 12:48
  with
• 12:48
  ground-truth
• 12:48
  labels
• 12:49
  —
• 12:49
  the
• 12:49
  guard
• 12:49
  admits
• 12:50
  post-shift
• 12:50
  novelty
• 12:50
  and
• 12:51
  the
• 12:51
  stale
• 12:51
  artifact
• 12:51
  mislabels
• 12:52
  it.
• 12:53
  An
• 12:53
  unfaithful
• 12:53
  deopt
• 12:54
  reference
• 12:54
  causes
• 12:54
  the
• 12:54
  verification
• 12:55
  gate
• 12:55
  to
• 12:55
  refuse
• 12:56
  recompilation
• 12:56
  entirely.
• 12:58
  The
• 12:58
  authors'
• 12:58
  core
• 12:58
  claim
• 12:58
  is
• 12:59
  that
• 12:59
  calibration
• 12:59
  and
• 12:59
  reference
• 13:00
  fidelity,
• 13:00
  not
• 13:01
  model
• 13:01
  capability,
• 13:02
  decide
• 13:02
  whether
• 13:02
  cheap
• 13:03
  stays
• 13:03
  correct.
• 13:04
  The
• 13:04
  system
• 13:05
  uses
• 13:05
  conformal
• 13:05
  prediction
• 13:06
  to
• 13:06
  set
• 13:06
  abstention
• 13:06
  thresholds,
• 13:07
  and
• 13:07
  the
• 13:08
  compile-deopt-recompile
• 13:09
  loop
• 13:09
  means
• 13:09
  "nothing
• 13:09
  is
• 13:10
  figured
• 13:10
  out
• 13:10
  twice."
• 13:11
  This
• 13:12
  is
• 13:12
  the
• 13:12
  compilation
• 13:12
  step
• 13:13
  that
• 13:13
  agent
• 13:13
  frameworks
• 13:13
  have
• 13:13
  been
• 13:14
  missing:
• 13:14
  not
• 13:14
  routing
• 13:15
  to
• 13:15
  cheaper
• 13:15
  models
• 13:15
  (FrugalGPT,
• 13:16
  RouteLLM),
• 13:18
  not
• 13:18
  caching
• 13:18
  semantically
• 13:19
  similar
• 13:19
  answers
• 13:20
  (GPTCache),
• 13:21
  but
• 13:21
  actually
• 13:21
  extracting
• 13:22
  the
• 13:22
  symbolic
• 13:22
  program
• 13:22
  behind
• 13:23
  repetitive
• 13:23
  agent
• 13:23
  behavior
• 13:24
  and
• 13:24
  running
• 13:24
  it
• 13:24
  for
• 13:24
  microdollars
• 13:25
  in
• 13:25
  a
• 13:25
  sandboxed
• 13:25
  WebAssembly
• 13:26
  runtime.
• 13:27
  MrFlow,
• 13:28
  a
• 13:28
  training-free
• 13:29
  diffusion
• 13:29
  acceleration
• 13:30
  method
• 13:30
  that
• 13:30
  shifts
• 13:30
  most
• 13:30
  denoising
• 13:31
  cost
• 13:31
  from
• 13:31
  expensive
• 13:32
  high-resolution
• 13:32
  sampling
• 13:33
  to
• 13:33
  cheap
• 13:33
  low-resolution
• 13:34
  passes
• 13:34
  followed
• 13:34
  by
• 13:34
  a
• 13:34
  single
• 13:35
  refinement
• 13:35
  step,
• 13:36
  has
• 13:36
  now
• 13:36
  released
• 13:37
  its
• 13:37
  ComfyUI
• 13:37
  plugin
• 13:38
  and
• 13:38
  expanded
• 13:38
  backbone
• 13:39
  support
• 13:39
  beyond
• 13:39
  the
• 13:39
  FLUX
• 13:40
  and
• 13:40
  Qwen-Image
• 13:40
  families
• 13:41
  covered
• 13:41
  in
• 13:41
  its
• 13:41
  initial
• 13:42
  paper.
• 13:42
  Measured
• 13:43
  end-to-end
• 13:43
  speedups
• 13:44
  range
• 13:44
  from
• 13:44
  8.25x
• 13:45
  on
• 13:45
  FLUX.1-dev
• 13:46
  to
• 13:47
  25x
• 13:47
  on
• 13:47
  Qwen-Image
• 13:48
  with
• 13:48
  distilled
• 13:48
  Pi-Flow
• 13:49
  weights
• 13:49
  —
• 13:50
  no
• 13:50
  finetuning
• 13:50
  or
• 13:50
  custom
• 13:51
  kernels
• 13:51
  required.
• 13:52
  On
• 13:52
  the
• 13:52
  practitioner
• 13:53
  tooling
• 13:53
  side,
• 13:54
  Research
• 13:54
  Radar
• 13:55
  is
• 13:55
  a
• 13:55
  daily
• 13:55
  cron
• 13:55
  job
• 13:56
  that
• 13:56
  fetches
• 13:56
  every
• 13:56
  new
• 13:57
  paper
• 13:57
  from
• 13:57
  your
• 13:57
  arXiv
• 13:57
  categories,
• 13:58
  scores
• 13:59
  abstracts
• 13:59
  1-10
• 14:00
  against
• 14:00
  a
• 14:00
  markdown
• 14:00
  file
• 14:01
  describing
• 14:01
  your
• 14:01
  research
• 14:02
  interests
• 14:02
  using
• 14:02
  a
• 14:02
  cheap
• 14:03
  local
• 14:03
  model,
• 14:04
  then
• 14:04
  deep-reads
• 14:04
  the
• 14:04
  top
• 14:05
  scorers
• 14:05
  with
• 14:05
  a
• 14:05
  stronger
• 14:06
  model
• 14:06
  that
• 14:06
  writes
• 14:06
  summaries,
• 14:07
  key
• 14:07
  insights,
• 14:08
  and
• 14:08
  specific
• 14:08
  connections
• 14:09
  to
• 14:09
  your
• 14:09
  work.
• 14:10
  The
• 14:10
  local
• 14:11
  inference
• 14:11
  path
• 14:11
  runs
• 14:12
  entirely
• 14:12
  through
• 14:12
  Ollama
• 14:13
  —
• 14:13
  nothing
• 14:14
  leaves
• 14:14
  your
• 14:14
  machine.
• 14:17
  Anthropic
• 14:17
  co-founder
• 14:18
  Jack
• 14:18
  Clark
• 14:19
  and
• 14:19
  Marina
• 14:19
  Favaro,
• 14:20
  leader
• 14:20
  of
• 14:20
  The
• 14:20
  Anthropic
• 14:21
  Institute,
• 14:22
  published
• 14:22
  a
• 14:22
  warning
• 14:22
  that
• 14:23
  AI
• 14:23
  systems
• 14:23
  approaching
• 14:24
  "full
• 14:24
  recursive
• 14:25
  self-improvement"
• 14:25
  —
• 14:26
  models
• 14:27
  capable
• 14:27
  of
• 14:27
  building
• 14:27
  their
• 14:28
  own
• 14:28
  successors
• 14:28
  —
• 14:29
  are
• 14:29
  much
• 14:29
  closer
• 14:30
  than
• 14:30
  previously
• 14:31
  expected.
• 14:31
  Their
• 14:31
  recommendation:
• 14:33
  the
• 14:33
  industry
• 14:33
  should
• 14:33
  consider
• 14:34
  slowing
• 14:34
  or
• 14:34
  pausing
• 14:35
  frontier
• 14:35
  AI
• 14:35
  development
• 14:36
  to
• 14:36
  allow
• 14:37
  researchers
• 14:37
  to
• 14:37
  understand
• 14:37
  the
• 14:38
  societal
• 14:38
  harms,
• 14:39
  and
• 14:39
  should
• 14:39
  build
• 14:39
  mechanisms
• 14:40
  for
• 14:40
  human
• 14:40
  intervention.
• 14:41
  "When
• 14:42
  I
• 14:42
  look
• 14:42
  down
• 14:42
  at
• 14:42
  the
• 14:42
  car
• 14:42
  we're
• 14:42
  driving,
• 14:43
  all
• 14:43
  I
• 14:44
  have
• 14:44
  is
• 14:44
  a
• 14:44
  gas
• 14:44
  pedal.
• 14:45
  I
• 14:45
  don't
• 14:45
  have
• 14:45
  a
• 14:45
  brake
• 14:45
  pedal,"
• 14:46
  Clark
• 14:46
  told
• 14:47
  Anderson
• 14:47
  Cooper
• 14:47
  on
• 14:47
  CNN,
• 14:48
  drawing
• 14:49
  a
• 14:49
  Cold
• 14:49
  War
• 14:49
  nuclear
• 14:49
  arms
• 14:50
  race
• 14:50
  analogy
• 14:50
  for
• 14:50
  why
• 14:50
  competitors
• 14:51
  might
• 14:51
  cooperate
• 14:52
  on
• 14:52
  safety
• 14:52
  constraints.
• 14:53
  The
• 14:53
  tension
• 14:54
  here
• 14:54
  deserves
• 14:54
  naming:
• 14:55
  just
• 14:55
  weeks
• 14:55
  ago,
• 14:56
  Andrej
• 14:57
  Karpathy
• 14:57
  joined
• 14:57
  Anthropic
• 14:58
  with
• 14:58
  the
• 14:58
  stated
• 14:58
  mission
• 14:58
  of
• 14:59
  teaching
• 14:59
  Claude
• 14:59
  to
• 14:59
  improve
• 15:00
  itself
• 15:00
  without
• 15:00
  human
• 15:00
  supervision.
• 15:02
  Recursive
• 15:02
  self-improvement
• 15:03
  was
• 15:03
  a
• 15:03
  pipedream
• 15:03
  a
• 15:03
  year
• 15:03
  ago;
• 15:04
  it
• 15:04
  is
• 15:04
  now
• 15:04
  an
• 15:04
  explicit
• 15:05
  engineering
• 15:05
  goal
• 15:06
  at
• 15:06
  the
• 15:06
  same
• 15:06
  company
• 15:06
  publicly
• 15:07
  warning
• 15:07
  the
• 15:07
  industry
• 15:07
  to
• 15:07
  slow
• 15:08
  down.
• 15:09
  The
• 15:09
  timing
• 15:09
  is
• 15:09
  also
• 15:10
  financial:
• 15:10
  Anthropic
• 15:11
  has
• 15:11
  filed
• 15:11
  for
• 15:11
  an
• 15:12
  IPO
• 15:12
  that
• 15:12
  could
• 15:12
  raise
• 15:12
  tens
• 15:13
  of
• 15:13
  billions,
• 15:14
  and
• 15:14
  SpaceX's
• 15:14
  $75
• 15:15
  billion
• 15:15
  IPO
• 15:16
  —
• 15:17
  the
• 15:17
  largest
• 15:17
  on
• 15:17
  record
• 15:17
  —
• 15:18
  is
• 15:18
  set
• 15:18
  for
• 15:18
  next
• 15:19
  week.
• 15:20
  The
• 15:20
  open-source
• 15:20
  community
• 15:21
  is
• 15:21
  pushing
• 15:21
  in
• 15:21
  the
• 15:21
  opposite
• 15:21
  direction
• 15:22
  —
• 15:22
  or
• 15:23
  at
• 15:23
  least
• 15:23
  a
• 15:23
  complementary
• 15:23
  one.
• 15:24
  Zhipu's
• 15:25
  founder
• 15:25
  has
• 15:25
  publicly
• 15:26
  backed
• 15:26
  open-source
• 15:26
  AI
• 15:27
  as
• 15:27
  the
• 15:27
  global
• 15:27
  security
• 15:27
  debate
• 15:28
  intensifies,
• 15:29
  arguing
• 15:29
  against
• 15:29
  intelligence
• 15:30
  monopolies.
• 15:31
  On
• 15:31
  the
• 15:31
  tools
• 15:32
  front,
• 15:32
  OpenCut
• 15:32
  is
• 15:32
  being
• 15:33
  rewritten
• 15:33
  from
• 15:33
  the
• 15:33
  ground
• 15:33
  up
• 15:34
  with
• 15:34
  a
• 15:34
  Rust
• 15:34
  core
• 15:34
  targeting
• 15:35
  web,
• 15:35
  desktop,
• 15:36
  and
• 15:36
  mobile
• 15:36
  from
• 15:36
  one
• 15:36
  codebase
• 15:37
  —
• 15:37
  with
• 15:38
  an
• 15:38
  MCP
• 15:38
  server
• 15:39
  for
• 15:39
  AI
• 15:39
  agents
• 15:39
  and
• 15:39
  headless
• 15:40
  mode
• 15:40
  for
• 15:40
  automation
• 15:41
  baked
• 15:41
  into
• 15:41
  the
• 15:41
  architecture
• 15:42
  from
• 15:42
  day
• 15:42
  one.
• 15:43
  And
• 15:43
  the
• 15:43
  Kode
• 15:43
  Dot,
• 15:44
  a
• 15:44
  programmable
• 15:44
  pocket
• 15:45
  device
• 15:45
  built
• 15:45
  on
• 15:45
  dual
• 15:46
  ESP32
• 15:47
  MCUs
• 15:48
  (P4
• 15:48
  +
• 15:48
  C5)
• 15:49
  with
• 15:49
  AMOLED
• 15:49
  touchscreen,
• 15:50
  NFC/RFID/IR,
• 15:52
  WiFi/BLE,
• 15:54
  and
• 15:54
  16
• 15:55
  programmable
• 15:55
  GPIO
• 15:55
  pins,
• 15:56
  has
• 15:56
  shipped
• 15:57
  to
• 15:57
  its
• 15:57
  16,000
• 15:57
  Kickstarter
• 15:58
  backers
• 15:58
  —
• 15:59
  a
• 15:59
  pocket
• 15:59
  pentesting
• 15:59
  and
• 16:00
  maker
• 16:00
  platform
• 16:00
  that
• 16:00
  gives
• 16:00
  the
• 16:01
  Flipper
• 16:01
  Zero
• 16:01
  serious
• 16:01
  competition.
• 16:03
  Stability
• 16:03
  AI's
• 16:03
  Stable
• 16:04
  Audio
• 16:04
  3
• 16:04
  Medium
• 16:05
  and
• 16:05
  LiquidAI's
• 16:05
  LFM2.5-230M
• 16:09
  both
• 16:09
  hit
• 16:09
  HuggingFace
• 16:10
  trending
• 16:10
  this
• 16:10
  week,
• 16:11
  adding
• 16:11
  to
• 16:11
  the
• 16:11
  steady
• 16:11
  stream
• 16:12
  of
• 16:12
  open-weight
• 16:12
  releases
• 16:13
  across
• 16:13
  modalities.