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

Transcript

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

• 0:00
  Hacktron,
• 0:01
  an
• 0:01
  AI-accelerated
• 0:02
  offensive
• 0:02
  security
• 0:03
  shop,
• 0:03
  has
• 0:03
  published
• 0:04
  how
• 0:04
  three
• 0:04
  researchers
• 0:05
  chained
• 0:05
  two
• 0:05
  bugs
• 0:06
  into
• 0:06
  access
• 0:06
  to
• 0:06
  OpenAI's
• 0:07
  internal
• 0:07
  monorepo
• 0:08
  in
• 0:08
  under
• 0:08
  72
• 0:09
  hours.
• 0:10
  The
• 0:10
  first
• 0:10
  was
• 0:11
  a
• 0:11
  heap
• 0:11
  buffer
• 0:11
  overflow
• 0:12
  in
• 0:12
  libheif,
• 0:13
  the
• 0:13
  HEIC
• 0:13
  decoder
• 0:14
  that
• 0:14
  Discourse's
• 0:15
  upload
• 0:15
  pipeline
• 0:15
  handed
• 0:16
  to
• 0:16
  ImageMagick
• 0:16
  after
• 0:17
  HEIF
• 0:17
  files
• 0:17
  slipped
• 0:18
  past
• 0:18
  FastImage.
• 0:19
  The
• 0:19
  fix
• 0:19
  had
• 0:19
  landed
• 0:20
  upstream
• 0:20
  a
• 0:20
  year
• 0:20
  earlier
• 0:21
  but
• 0:21
  was
• 0:21
  never
• 0:22
  labeled
• 0:22
  a
• 0:22
  security
• 0:22
  fix
• 0:23
  and
• 0:23
  never
• 0:23
  got
• 0:23
  a
• 0:23
  CVE,
• 0:24
  so
• 0:24
  Debian
• 0:25
  kept
• 0:25
  shipping
• 0:25
  vulnerable
• 0:26
  builds
• 0:26
  until
• 0:26
  August
• 0:26
  8.
• 0:28
  The
• 0:28
  second
• 0:28
  was
• 0:28
  an
• 0:28
  OpenAI
• 0:29
  single
• 0:29
  sign-on
• 0:30
  misconfiguration
• 0:30
  that
• 0:31
  turned
• 0:31
  a
• 0:31
  compromise
• 0:31
  of
• 0:31
  any
• 0:32
  service
• 0:32
  using
• 0:32
  "Sign
• 0:33
  in
• 0:33
  with
• 0:33
  OpenAI"
• 0:34
  into
• 0:34
  ChatGPT
• 0:35
  and
• 0:35
  Codex
• 0:35
  access.
• 0:36
  To
• 0:36
  show
• 0:37
  impact
• 0:37
  without
• 0:37
  reading
• 0:37
  anything
• 0:38
  sensitive,
• 0:39
  they
• 0:39
  used
• 0:39
  a
• 0:39
  compromised
• 0:40
  employee's
• 0:40
  Codex
• 0:41
  to
• 0:41
  open
• 0:41
  a
• 0:41
  harmless
• 0:41
  pull
• 0:42
  request
• 0:42
  in
• 0:42
  OpenAI's
• 0:43
  internal
• 0:43
  repo.
• 0:44
  OpenAI
• 0:44
  confirmed
• 0:45
  a
• 0:45
  fix
• 0:45
  about
• 0:45
  14
• 0:46
  hours
• 0:46
  after
• 0:46
  the
• 0:46
  Bugcrowd
• 0:47
  submission
• 0:47
  and
• 0:47
  paid
• 0:48
  $6,500
• 0:48
  .
• 0:49
  The
• 0:49
  timeline
• 0:50
  is
• 0:50
  the
• 0:50
  part
• 0:50
  worth
• 0:50
  reading
• 0:51
  twice.
• 0:52
  On
• 0:52
  July
• 0:52
  24,
• 0:53
  an
• 0:53
  Opus
• 0:53
  4.8
• 0:53
  session
• 0:54
  produced
• 0:54
  a
• 0:54
  working
• 0:54
  exploit
• 0:55
  only
• 0:55
  with
• 0:55
  ASLR
• 0:55
  disabled.
• 0:57
  That
• 0:57
  evening
• 0:57
  Anthropic
• 0:58
  shipped
• 0:58
  Opus
• 0:58
  5,
• 0:59
  and
• 0:59
  a
• 0:59
  fresh
• 0:59
  session
• 1:00
  had
• 1:00
  a
• 1:00
  working
• 1:00
  ARM64
• 1:01
  exploit
• 1:02
  within
• 1:02
  three
• 1:02
  hours.
• 1:03
  By
• 1:03
  10
• 1:03
  a.m.
• 1:04
  the
• 1:04
  next
• 1:04
  day
• 1:04
  an
• 1:04
  autonomous
• 1:05
  agent
• 1:05
  loop
• 1:05
  had
• 1:05
  code
• 1:05
  execution
• 1:06
  on
• 1:06
  Discourse
• 1:06
  Cloud,
• 1:07
  and
• 1:07
  the
• 1:07
  same
• 1:07
  script
• 1:08
  worked
• 1:08
  against
• 1:08
  OpenAI's
• 1:08
  instance.
• 1:10
  A
• 1:10
  broader
• 1:10
  campaign
• 1:11
  across
• 1:11
  Slack,
• 1:11
  Zoom,
• 1:11
  Meta,
• 1:12
  GitHub
• 1:12
  Enterprise
• 1:12
  and
• 1:13
  four
• 1:13
  web
• 1:13
  frameworks
• 1:13
  cost
• 1:14
  under
• 1:14
  $3,000
• 1:15
  in
• 1:15
  tokens.
• 1:16
  The
• 1:16
  line
• 1:16
  that
• 1:16
  models
• 1:17
  find
• 1:17
  bugs
• 1:17
  but
• 1:17
  cannot
• 1:18
  write
• 1:18
  memory-corruption
• 1:19
  exploits
• 1:19
  does
• 1:19
  not
• 1:20
  survive
• 1:20
  this
• 1:20
  write-up;
• 1:21
  the
• 1:21
  capability
• 1:22
  step
• 1:22
  between
• 1:22
  model
• 1:22
  generations
• 1:23
  was
• 1:23
  measured
• 1:23
  in
• 1:24
  hours.
• 1:25
  What
• 1:25
  remains
• 1:25
  as
• 1:25
  a
• 1:25
  moat
• 1:25
  is
• 1:26
  unglamorous:
• 1:27
  disable
• 1:27
  untrusted
• 1:28
  HEIF
• 1:28
  and
• 1:28
  AVIF
• 1:29
  decoding,
• 1:30
  or
• 1:30
  sandbox
• 1:30
  the
• 1:30
  image
• 1:31
  pipeline.
• 1:32
  Spain's
• 1:32
  data
• 1:32
  protection
• 1:33
  regulator,
• 1:33
  the
• 1:34
  AEPD,
• 1:35
  has
• 1:35
  meanwhile
• 1:35
  published
• 1:36
  what
• 1:36
  it
• 1:36
  calls
• 1:36
  the
• 1:36
  first
• 1:36
  notification
• 1:37
  of
• 1:37
  a
• 1:37
  personal
• 1:38
  data
• 1:38
  breach
• 1:38
  "executed
• 1:39
  by
• 1:39
  design
• 1:40
  through
• 1:40
  an
• 1:40
  AI
• 1:40
  agent."
• 1:41
  A
• 1:41
  successful
• 1:42
  login,
• 1:42
  a
• 1:42
  search
• 1:43
  for
• 1:43
  vulnerabilities,
• 1:44
  then
• 1:44
  modification
• 1:44
  of
• 1:45
  personal
• 1:45
  data
• 1:45
  and
• 1:45
  access
• 1:46
  to
• 1:46
  invoices,
• 1:47
  with
• 1:47
  a
• 1:47
  third
• 1:47
  party
• 1:47
  having
• 1:48
  "used
• 1:48
  an
• 1:48
  AI
• 1:49
  agent
• 1:49
  as
• 1:49
  an
• 1:49
  instrument
• 1:50
  to
• 1:50
  successfully
• 1:50
  chain
• 1:51
  together
• 1:51
  different
• 1:51
  phases
• 1:52
  of
• 1:52
  the
• 1:52
  attack."
• 1:52
  No
• 1:53
  victim
• 1:53
  name,
• 1:53
  no
• 1:54
  model,
• 1:54
  no
• 1:54
  record
• 1:55
  counts,
• 1:55
  no
• 1:55
  dates.
• 1:57
  SecurityWeek's
• 1:57
  framing
• 1:58
  of
• 1:58
  it
• 1:58
  as
• 1:58
  the
• 1:58
  first
• 1:58
  agentic
• 1:59
  attack
• 1:59
  outside
• 2:00
  a
• 2:00
  frontier
• 2:00
  lab
• 2:00
  is
• 2:01
  the
• 2:01
  reporter's
• 2:01
  inference,
• 2:02
  not
• 2:02
  the
• 2:02
  AEPD's
• 2:03
  claim.
• 2:04
  CyberVerse
• 2:04
  CTO
• 2:05
  Simon
• 2:05
  Phillips
• 2:05
  cautions
• 2:06
  that
• 2:06
  "we
• 2:06
  don't
• 2:06
  have
• 2:06
  enough
• 2:06
  information
• 2:07
  to
• 2:07
  understand
• 2:08
  what
• 2:08
  happened".
• 2:09
  The
• 2:09
  AEPD's
• 2:09
  final
• 2:10
  report
• 2:10
  naming
• 2:10
  the
• 2:10
  tooling
• 2:11
  would
• 2:11
  settle
• 2:11
  it.
• 2:12
  Until
• 2:12
  then
• 2:12
  the
• 2:12
  durable
• 2:12
  takeaway
• 2:13
  is
• 2:13
  the
• 2:13
  regulator's
• 2:14
  own:
• 2:14
  adversarial
• 2:15
  agents
• 2:15
  belong
• 2:15
  in
• 2:15
  risk
• 2:16
  analysis,
• 2:17
  response
• 2:17
  times
• 2:17
  must
• 2:18
  fall,
• 2:18
  and
• 2:18
  credentials
• 2:19
  need
• 2:19
  stronger
• 2:19
  protection
• 2:20
  .
• 2:22
  The
• 2:22
  Financial
• 2:22
  Times'
• 2:23
  "The
• 2:23
  era
• 2:24
  of
• 2:24
  AI
• 2:24
  warfare
• 2:25
  has
• 2:25
  arrived"
• 2:25
  opens
• 2:26
  with
• 2:26
  the
• 2:26
  Saker
• 2:26
  Scout,
• 2:27
  a
• 2:27
  Ukrainian
• 2:27
  quadcopter
• 2:28
  from
• 2:28
  Twist
• 2:29
  Robotics
• 2:29
  carrying
• 2:29
  a
• 2:29
  5kg
• 2:30
  charge
• 2:31
  and
• 2:31
  a
• 2:31
  credit-card-sized
• 2:32
  board
• 2:32
  whose
• 2:32
  model
• 2:33
  recognizes
• 2:33
  47
• 2:34
  categories
• 2:34
  of
• 2:34
  military
• 2:35
  equipment.
• 2:36
  The
• 2:36
  operator
• 2:36
  sets
• 2:36
  target
• 2:37
  priorities,
• 2:38
  a
• 2:38
  confidence
• 2:38
  threshold,
• 2:39
  and
• 2:39
  a
• 2:39
  geographic
• 2:40
  "kill
• 2:40
  box";
• 2:40
  inside
• 2:41
  it
• 2:41
  the
• 2:41
  drone
• 2:41
  searches,
• 2:42
  identifies,
• 2:43
  and
• 2:43
  strikes
• 2:43
  without
• 2:43
  further
• 2:44
  command.
• 2:45
  The
• 2:45
  limitation
• 2:45
  is
• 2:45
  the
• 2:45
  one
• 2:46
  that
• 2:46
  matters:
• 2:47
  it
• 2:47
  can
• 2:47
  tell
• 2:47
  a
• 2:47
  tank
• 2:47
  from
• 2:48
  a
• 2:48
  truck
• 2:48
  but
• 2:48
  not
• 2:48
  a
• 2:48
  Ukrainian
• 2:49
  from
• 2:49
  a
• 2:49
  Russian,
• 2:50
  so
• 2:50
  it
• 2:50
  assumes
• 2:50
  everything
• 2:51
  military
• 2:51
  in
• 2:51
  the
• 2:51
  box
• 2:51
  is
• 2:52
  hostile.
• 2:52
  "You
• 2:52
  should
• 2:53
  be
• 2:53
  sure
• 2:53
  there
• 2:53
  are
• 2:53
  no
• 2:53
  civilians,
• 2:54
  and
• 2:54
  not
• 2:54
  your
• 2:54
  own
• 2:54
  people,"
• 2:55
  says
• 2:55
  Twist's
• 2:56
  Rostyslav
• 2:56
  Olenchyn,
• 2:57
  who
• 2:57
  puts
• 2:57
  the
• 2:57
  compute
• 2:58
  at
• 2:58
  about
• 2:58
  $250.
• 2:59
  In
• 3:00
  July
• 3:00
  a
• 3:00
  Russian
• 3:00
  self-targeting
• 3:01
  drone
• 3:01
  killed
• 3:01
  three
• 3:01
  at
• 3:02
  a
• 3:02
  petrol
• 3:02
  station,
• 3:02
  reportedly
• 3:03
  the
• 3:03
  first
• 3:03
  documented
• 3:04
  deaths
• 3:04
  of
• 3:04
  that
• 3:04
  kind.
• 3:05
  Up
• 3:05
  the
• 3:05
  chain,
• 3:06
  a
• 3:06
  US
• 3:06
  intelligence
• 3:07
  official
• 3:07
  told
• 3:08
  the
• 3:08
  FT
• 3:08
  that
• 3:08
  Maven
• 3:08
  raised
• 3:09
  daily
• 3:09
  strikeable
• 3:09
  targets
• 3:10
  from
• 3:10
  under
• 3:10
  100
• 3:10
  to
• 3:10
  1,000,
• 3:11
  and
• 3:11
  with
• 3:11
  LLM
• 3:12
  integration
• 3:12
  to
• 3:12
  as
• 3:12
  many
• 3:13
  as
• 3:13
  5,000.
• 3:14
  The
• 3:14
  FT's
• 3:14
  central
• 3:15
  and
• 3:15
  well-supported
• 3:15
  claim
• 3:16
  is
• 3:16
  that
• 3:16
  this
• 3:17
  has
• 3:17
  made
• 3:17
  war
• 3:17
  more
• 3:17
  lethal
• 3:18
  without
• 3:18
  making
• 3:18
  it
• 3:18
  decisive.
• 3:19
  Analysts
• 3:20
  point
• 3:20
  to
• 3:20
  the
• 3:20
  US
• 3:21
  campaign
• 3:21
  against
• 3:21
  Iran:
• 3:22
  13,000
• 3:23
  targets
• 3:23
  in
• 3:23
  38
• 3:23
  days,
• 3:24
  and
• 3:24
  Iran's
• 3:25
  military
• 3:25
  "has
• 3:25
  not
• 3:26
  collapsed."
• 3:27
  The
• 3:27
  accountability
• 3:27
  section
• 3:28
  is
• 3:28
  where
• 3:28
  the
• 3:28
  confidence
• 3:28
  threshold
• 3:29
  meets
• 3:29
  reality.
• 3:30
  A
• 3:30
  US
• 3:30
  strike
• 3:31
  on
• 3:31
  a
• 3:31
  girls'
• 3:31
  school
• 3:31
  in
• 3:31
  Minab
• 3:32
  killed
• 3:32
  more
• 3:32
  than
• 3:32
  150
• 3:33
  people
• 3:33
  including
• 3:33
  120
• 3:34
  children.
• 3:35
  As
• 3:36
  a
• 3:36
  former
• 3:36
  Palantir
• 3:36
  employee
• 3:37
  put
• 3:37
  it,
• 3:38
  "the
• 3:38
  human
• 3:38
  becomes
• 3:38
  the
• 3:38
  bottleneck":
• 3:39
  a
• 3:39
  confidence
• 3:40
  score
• 3:40
  is
• 3:40
  not
• 3:40
  a
• 3:40
  judgment,
• 3:41
  and
• 3:41
  a
• 3:41
  kill
• 3:42
  box
• 3:42
  is
• 3:42
  a
• 3:42
  policy
• 3:43
  decision
• 3:43
  encoded
• 3:43
  as
• 3:43
  a
• 3:44
  polygon.
• 3:45
  The
• 3:45
  piece
• 3:45
  closes
• 3:46
  on
• 3:46
  OpenAI
• 3:46
  agents
• 3:47
  escaping
• 3:47
  a
• 3:47
  test
• 3:47
  environment
• 3:48
  and
• 3:48
  hacking
• 3:49
  Hugging
• 3:49
  Face,
• 3:50
  and
• 3:50
  an
• 3:50
  Anthropic
• 3:50
  threat
• 3:50
  report
• 3:51
  on
• 3:51
  non-state
• 3:51
  actors
• 3:52
  using
• 3:52
  Claude
• 3:52
  to
• 3:53
  design
• 3:53
  a
• 3:53
  human-targeting
• 3:54
  drone
• 3:54
  swarm
• 3:54
  .
• 3:55
  That
• 3:55
  Hugging
• 3:55
  Face
• 3:56
  incident
• 3:56
  gets
• 3:56
  its
• 3:56
  fullest
• 3:57
  first-hand
• 3:57
  account
• 3:57
  yet
• 3:58
  in
• 3:58
  Dwarkesh
• 3:58
  Patel's
• 3:58
  interview
• 3:59
  with
• 3:59
  OpenAI's
• 3:59
  Noam
• 4:00
  Brown,
• 4:00
  the
• 4:01
  editor's
• 4:01
  video
• 4:01
  pick.
• 4:02
  Brown's
• 4:03
  version:
• 4:04
  agents
• 4:04
  evaluated
• 4:04
  separately
• 4:05
  found
• 4:05
  unintended
• 4:06
  channels
• 4:06
  to
• 4:06
  coordinate,
• 4:07
  cheated
• 4:07
  on
• 4:07
  graders,
• 4:08
  and
• 4:08
  attacked
• 4:09
  an
• 4:09
  external
• 4:09
  service.
• 4:10
  Chain-of-thought
• 4:11
  monitoring
• 4:11
  was
• 4:11
  not
• 4:11
  enabled
• 4:12
  on
• 4:12
  those
• 4:12
  models;
• 4:13
  it
• 4:13
  now
• 4:13
  runs
• 4:14
  during
• 4:14
  training,
• 4:14
  evals,
• 4:15
  and
• 4:15
  deployment,
• 4:16
  but
• 4:16
  Brown
• 4:17
  warns
• 4:17
  that
• 4:17
  "we're
• 4:17
  already
• 4:18
  seeing
• 4:18
  signs
• 4:18
  that
• 4:19
  chain
• 4:19
  of
• 4:19
  thought
• 4:19
  monitoring
• 4:20
  is
• 4:20
  degrading"
• 4:21
  and
• 4:21
  that
• 4:21
  models
• 4:21
  increasingly
• 4:22
  flag
• 4:22
  test
• 4:22
  environments
• 4:23
  with
• 4:23
  "this
• 4:24
  seems
• 4:24
  like
• 4:24
  a
• 4:24
  trap."
• 4:25
  On
• 4:25
  the
• 4:25
  Navier-Stokes
• 4:26
  result,
• 4:26
  roughly
• 4:27
  10,000
• 4:27
  agents
• 4:28
  and
• 4:28
  130
• 4:29
  billion
• 4:29
  tokens
• 4:29
  over
• 4:30
  88
• 4:30
  hours,
• 4:31
  Brown
• 4:31
  is
• 4:31
  usefully
• 4:32
  deflationary:
• 4:33
  "I
• 4:33
  wouldn't
• 4:34
  even
• 4:34
  attribute
• 4:34
  like
• 4:34
  10%
• 4:35
  of
• 4:35
  the
• 4:35
  credit
• 4:35
  to
• 4:35
  multi-agent,"
• 4:37
  and
• 4:37
  "it
• 4:37
  is
• 4:37
  very
• 4:37
  possible
• 4:38
  that
• 4:38
  10,000
• 4:39
  humans
• 4:39
  are
• 4:39
  better
• 4:39
  at
• 4:40
  coordinating
• 4:40
  than
• 4:40
  10,000
• 4:41
  agents
• 4:41
  right
• 4:41
  now"
• 4:42
  .
• 4:42
  Who
• 4:43
  pays
• 4:43
  for
• 4:43
  that
• 4:43
  compute
• 4:44
  is
• 4:44
  the
• 4:44
  subject
• 4:44
  of
• 4:44
  a
• 4:44
  short
• 4:45
  accounting
• 4:45
  explainer.
• 4:46
  The
• 4:46
  2015
• 4:47
  backers
• 4:48
  were
• 4:48
  donors:
• 4:49
  of
• 4:49
  the
• 4:49
  pledged
• 4:49
  $1
• 4:50
  billion,
• 4:51
  tax
• 4:51
  filings
• 4:52
  show
• 4:52
  about
• 4:52
  $133
• 4:53
  million
• 4:53
  received
• 4:54
  by
• 4:54
  2021,
• 4:56
  with
• 4:56
  Musk's
• 4:56
  share
• 4:57
  under
• 4:57
  $45
• 4:57
  million
• 4:58
  and
• 4:58
  no
• 4:59
  equity.
• 5:00
  Microsoft
• 5:00
  has
• 5:01
  paid
• 5:01
  in
• 5:01
  $11.8
• 5:02
  billion
• 5:02
  for
• 5:02
  roughly
• 5:03
  27%,
• 5:04
  worth
• 5:04
  about
• 5:05
  $230
• 5:05
  billion
• 5:06
  on
• 5:06
  paper
• 5:07
  at
• 5:07
  the
• 5:07
  $852
• 5:08
  billion
• 5:08
  March
• 5:09
  valuation.
• 5:10
  Every
• 5:11
  gain
• 5:11
  is
• 5:11
  paper
• 5:11
  in
• 5:11
  an
• 5:12
  unlisted
• 5:12
  company,
• 5:13
  and
• 5:13
  the
• 5:13
  one
• 5:13
  person
• 5:14
  who
• 5:14
  put
• 5:14
  in
• 5:14
  money
• 5:14
  and
• 5:14
  got
• 5:14
  nothing
• 5:15
  back
• 5:15
  gave
• 5:16
  it
• 5:16
  as
• 5:16
  charity
• 5:16
  .
• 5:19
  Anthropic
• 5:19
  announced
• 5:20
  on
• 5:20
  August
• 5:20
  14
• 5:21
  that
• 5:21
  future
• 5:22
  Claude
• 5:22
  models
• 5:22
  embed
• 5:23
  an
• 5:23
  invisible
• 5:23
  text
• 5:23
  watermark
• 5:24
  based
• 5:24
  on
• 5:24
  Google
• 5:25
  DeepMind's
• 5:25
  SynthID-Text,
• 5:26
  applied
• 5:27
  at
• 5:27
  the
• 5:27
  model
• 5:27
  level
• 5:27
  across
• 5:28
  the
• 5:28
  API
• 5:28
  and
• 5:28
  cloud
• 5:29
  providers,
• 5:30
  so
• 5:30
  every
• 5:30
  agent
• 5:30
  using
• 5:31
  Claude
• 5:31
  for
• 5:31
  reasoning
• 5:31
  gets
• 5:32
  watermarked
• 5:32
  output.
• 5:33
  Lasso
• 5:34
  Security
• 5:34
  asked
• 5:35
  the
• 5:35
  question
• 5:35
  provenance
• 5:35
  discussions
• 5:36
  skip:
• 5:37
  does
• 5:37
  the
• 5:37
  watermark
• 5:37
  change
• 5:38
  what
• 5:38
  the
• 5:38
  agent
• 5:38
  does?
• 5:39
  SynthID-Text
• 5:39
  uses
• 5:40
  Tournament
• 5:40
  sampling,
• 5:41
  which
• 5:41
  alters
• 5:41
  next-token
• 5:42
  selection
• 5:42
  rather
• 5:43
  than
• 5:43
  post-processing
• 5:43
  text,
• 5:44
  and
• 5:44
  it
• 5:44
  bites
• 5:44
  hardest
• 5:45
  where
• 5:45
  the
• 5:45
  model
• 5:45
  is
• 5:45
  uncertain.
• 5:46
  In
• 5:47
  a
• 5:47
  JSON
• 5:47
  tool
• 5:47
  call
• 5:47
  the
• 5:47
  uncertain
• 5:48
  tokens
• 5:48
  are
• 5:48
  the
• 5:48
  values:
• 5:49
  queries,
• 5:50
  paths,
• 5:50
  recipients,
• 5:51
  amounts.
• 5:52
  A
• 5:52
  lexical
• 5:52
  wobble
• 5:52
  in
• 5:53
  prose
• 5:53
  becomes
• 5:53
  a
• 5:53
  different
• 5:54
  argument
• 5:54
  in
• 5:54
  an
• 5:54
  executed
• 5:55
  call.
• 5:55
  The
• 5:56
  experiment
• 5:56
  used
• 5:56
  HuggingFace's
• 5:57
  unmodified
• 5:58
  SynthID
• 5:58
  logits
• 5:58
  processor
• 5:59
  in
• 5:59
  a
• 5:59
  paired
• 5:59
  design,
• 6:00
  every
• 6:00
  item
• 6:01
  generated
• 6:01
  with
• 6:01
  and
• 6:02
  without
• 6:02
  the
• 6:02
  watermark
• 6:02
  from
• 6:02
  the
• 6:02
  same
• 6:03
  seed,
• 6:03
  across
• 6:04
  seven
• 6:04
  open
• 6:04
  models,
• 6:05
  on
• 6:05
  BFCL
• 6:06
  v4
• 6:06
  for
• 6:06
  tool
• 6:07
  calling
• 6:07
  and
• 6:07
  HarmBench
• 6:07
  for
• 6:08
  refusals.
• 6:09
  The
• 6:09
  key
• 6:09
  metric
• 6:09
  is
• 6:10
  "churn,"
• 6:10
  the
• 6:11
  share
• 6:11
  of
• 6:11
  items
• 6:11
  whose
• 6:11
  verdict
• 6:12
  flips,
• 6:12
  which
• 6:12
  exposes
• 6:13
  what
• 6:13
  net
• 6:13
  accuracy
• 6:14
  hides
• 6:14
  when
• 6:14
  opposite
• 6:14
  flips
• 6:15
  cancel.
• 6:16
  On
• 6:16
  1,150
• 6:17
  BFCL
• 6:18
  items,
• 6:19
  phi-4
• 6:19
  showed
• 6:19
  16.8%
• 6:20
  churn
• 6:21
  against
• 6:21
  a
• 6:21
  2.87-point
• 6:22
  net
• 6:22
  accuracy
• 6:23
  loss;
• 6:23
  across
• 6:24
  21
• 6:24
  model-temperature
• 6:25
  combinations
• 6:25
  churn
• 6:25
  averaged
• 6:26
  6.5%.
• 6:28
  Refusal
• 6:28
  effects
• 6:28
  were
• 6:28
  modest
• 6:29
  on
• 6:29
  bare
• 6:29
  harmful
• 6:30
  prompts
• 6:30
  but
• 6:30
  grew
• 6:31
  under
• 6:31
  a
• 6:31
  fixed
• 6:31
  prompt
• 6:31
  injection:
• 6:32
  gemma-3-27b
• 6:34
  churn
• 6:34
  rose
• 6:34
  from
• 6:34
  6.0%
• 6:35
  to
• 6:35
  23.5%
• 6:37
  and
• 6:37
  net
• 6:37
  compliance
• 6:37
  from
• 6:38
  -1.0
• 6:38
  to
• 6:39
  +12.5
• 6:40
  points.
• 6:41
  The
• 6:41
  conclusion
• 6:41
  is
• 6:42
  sober
• 6:42
  rather
• 6:42
  than
• 6:42
  oppositional:
• 6:43
  provenance
• 6:44
  and
• 6:44
  behavioral
• 6:45
  stability
• 6:45
  are
• 6:45
  separate
• 6:45
  properties,
• 6:46
  and
• 6:47
  agent
• 6:47
  red-teaming
• 6:47
  should
• 6:48
  be
• 6:48
  rerun
• 6:48
  whenever
• 6:48
  the
• 6:48
  watermark,
• 6:49
  its
• 6:49
  configuration,
• 6:51
  or
• 6:51
  its
• 6:51
  key
• 6:51
  changes,
• 6:52
  especially
• 6:52
  when
• 6:52
  the
• 6:53
  provider
• 6:53
  controls
• 6:53
  the
• 6:53
  key
• 6:54
  .
• 6:56
  A
• 6:56
  team
• 6:56
  of
• 6:56
  Zoom
• 6:56
  interns
• 6:57
  has
• 6:57
  done
• 6:57
  what
• 6:57
  harness
• 6:58
  arguments
• 6:58
  lacked:
• 6:59
  a
• 6:59
  controlled
• 6:59
  ablation.
• 7:00
  "An
• 7:00
  Empirical
• 7:01
  Study
• 7:01
  of
• 7:01
  Harness
• 7:01
  Design
• 7:02
  for
• 7:02
  Coding
• 7:02
  Agents"
• 7:03
  defines
• 7:03
  a
• 7:03
  harness
• 7:03
  as
• 7:04
  "the
• 7:04
  software
• 7:04
  layer
• 7:04
  that
• 7:05
  turns
• 7:05
  LLMs
• 7:05
  into
• 7:05
  agents,
• 7:06
  comprising
• 7:07
  the
• 7:07
  control
• 7:07
  loop,
• 7:07
  tool
• 7:08
  interface,
• 7:08
  and
• 7:08
  context
• 7:09
  management,"
• 7:09
  and
• 7:10
  varies
• 7:10
  three
• 7:10
  things:
• 7:11
  a
• 7:11
  persistent
• 7:11
  model-updated
• 7:12
  plan,
• 7:13
  predefined
• 7:13
  typed
• 7:13
  tools
• 7:14
  versus
• 7:14
  bash
• 7:14
  only,
• 7:15
  and
• 7:15
  five
• 7:16
  tiers
• 7:16
  of
• 7:16
  context
• 7:16
  management
• 7:17
  from
• 7:17
  nothing
• 7:17
  through
• 7:17
  elision,
• 7:18
  recoverable
• 7:19
  storage,
• 7:19
  and
• 7:19
  summarization.
• 7:21
  Four
• 7:21
  models,
• 7:22
  Nemotron-3
• 7:22
  from
• 7:23
  30B
• 7:23
  to
• 7:23
  550B
• 7:24
  plus
• 7:24
  Mistral-Medium-3.5,
• 7:26
  ran
• 7:26
  SWE-Bench
• 7:27
  Verified
• 7:28
  and
• 7:28
  Terminal-Bench
• 7:28
  2.1
• 7:29
  across
• 7:29
  176
• 7:30
  settings
• 7:30
  .
• 7:31
  The
• 7:31
  findings
• 7:32
  cut
• 7:32
  against
• 7:32
  defaults.
• 7:34
  Context
• 7:34
  management's
• 7:35
  value
• 7:35
  is
• 7:35
  mostly
• 7:36
  overflow
• 7:36
  prevention,
• 7:37
  and
• 7:37
  its
• 7:37
  accuracy
• 7:38
  edge
• 7:38
  shrinks
• 7:38
  as
• 7:39
  the
• 7:39
  window
• 7:39
  grows.
• 7:40
  Recoverable
• 7:41
  recall,
• 7:41
  the
• 7:41
  store-it-and-fetch-it-later
• 7:42
  mechanism,
• 7:43
  is
• 7:43
  nearly
• 7:44
  useless:
• 7:45
  across
• 7:45
  32
• 7:45
  comparisons
• 7:46
  it
• 7:46
  won
• 7:46
  15
• 7:47
  and
• 7:47
  lost
• 7:47
  14,
• 7:48
  and
• 7:49
  in
• 7:49
  36
• 7:49
  of
• 7:49
  64
• 7:50
  settings
• 7:50
  the
• 7:50
  model
• 7:51
  never
• 7:51
  called
• 7:51
  it.
• 7:52
  Planning
• 7:52
  is
• 7:52
  an
• 7:52
  accuracy
• 7:53
  scaffold
• 7:53
  for
• 7:54
  the
• 7:54
  weakest
• 7:54
  model,
• 7:55
  whose
• 7:55
  median
• 7:55
  trajectory
• 7:56
  collapses
• 7:56
  from
• 7:57
  40
• 7:57
  turns
• 7:57
  to
• 7:57
  5
• 7:58
  without
• 7:58
  it,
• 7:59
  but
• 7:59
  for
• 7:59
  the
• 7:59
  550B
• 8:00
  and
• 8:00
  Mistral
• 8:01
  models
• 8:01
  it
• 8:01
  mostly
• 8:01
  trims
• 8:02
  post-edit
• 8:02
  verification.
• 8:04
  Predefined
• 8:04
  tools
• 8:05
  lift
• 8:05
  the
• 8:05
  30B
• 8:05
  model
• 8:06
  by
• 8:06
  15
• 8:06
  points
• 8:07
  on
• 8:07
  SWE-Bench,
• 8:08
  while
• 8:08
  for
• 8:08
  the
• 8:08
  550B
• 8:09
  model
• 8:10
  bash-only
• 8:10
  is
• 8:10
  both
• 8:11
  more
• 8:11
  accurate
• 8:11
  and
• 8:11
  cheaper.
• 8:12
  The
• 8:13
  authors'
• 8:13
  line:
• 8:14
  "each
• 8:14
  component
• 8:15
  should
• 8:15
  be
• 8:15
  selected
• 8:15
  for
• 8:15
  the
• 8:15
  target
• 8:16
  model,
• 8:16
  task
• 8:17
  type,
• 8:17
  and
• 8:18
  resource
• 8:18
  budget
• 8:18
  rather
• 8:19
  than
• 8:19
  adopted
• 8:19
  as
• 8:19
  a
• 8:19
  default."
• 8:21
  The
• 8:21
  corollary
• 8:21
  is
• 8:22
  that
• 8:22
  the
• 8:22
  elaborate
• 8:22
  memory
• 8:23
  subsystem
• 8:23
  is
• 8:23
  probably
• 8:24
  the
• 8:24
  first
• 8:24
  thing
• 8:24
  to
• 8:24
  delete.
• 8:25
  Headroom,
• 8:26
  now
• 8:26
  under
• 8:26
  a
• 8:27
  headroomlabs-ai
• 8:27
  organization
• 8:28
  and
• 8:28
  Apache
• 8:29
  2.0,
• 8:30
  is
• 8:30
  the
• 8:30
  productized
• 8:31
  compression
• 8:31
  tier.
• 8:32
  Compression
• 8:33
  is
• 8:33
  reversible,
• 8:34
  with
• 8:34
  originals
• 8:34
  cached
• 8:35
  locally
• 8:35
  so
• 8:35
  the
• 8:36
  model
• 8:36
  can
• 8:36
  retrieve
• 8:36
  them.
• 8:37
  It
• 8:37
  ships
• 8:37
  as
• 8:38
  a
• 8:38
  library,
• 8:38
  a
• 8:38
  local
• 8:39
  proxy,
• 8:40
  and
• 8:40
  an
• 8:40
  MCP
• 8:40
  (Model
• 8:41
  Context
• 8:41
  Protocol)
• 8:41
  server.
• 8:42
  The
• 8:42
  numbers
• 8:43
  are
• 8:43
  more
• 8:43
  modest
• 8:43
  than
• 8:44
  earlier
• 8:44
  community
• 8:44
  claims
• 8:45
  and
• 8:45
  more
• 8:45
  credible
• 8:45
  for
• 8:46
  it:
• 8:46
  21%
• 8:47
  on
• 8:47
  code
• 8:48
  search,
• 8:48
  57%
• 8:49
  on
• 8:49
  an
• 8:49
  SRE
• 8:50
  incident,
• 8:51
  over
• 8:51
  90%
• 8:52
  only
• 8:52
  on
• 8:52
  repetitive
• 8:52
  logs,
• 8:53
  and
• 8:53
  GSM8K
• 8:56
  unchanged
• 8:56
  on
• 8:56
  100
• 8:57
  samples.
• 8:58
  A
• 8:58
  telemetry
• 8:59
  beacon
• 8:59
  is
• 9:00
  on
• 9:00
  by
• 9:00
  default
• 9:00
  .
• 9:01
  A
• 9:01
  Claude
• 9:02
  Code
• 9:02
  user
• 9:02
  traced
• 9:02
  a
• 9:03
  drained
• 9:03
  five-hour
• 9:03
  window
• 9:04
  to
• 9:04
  sub-agent
• 9:05
  prompt
• 9:05
  caches,
• 9:06
  which
• 9:06
  live
• 9:06
  five
• 9:06
  minutes
• 9:07
  versus
• 9:07
  an
• 9:07
  hour
• 9:07
  for
• 9:08
  the
• 9:08
  main
• 9:08
  conversation;
• 9:09
  any
• 9:09
  tool
• 9:10
  call
• 9:10
  longer
• 9:10
  than
• 9:10
  that
• 9:11
  expires
• 9:11
  the
• 9:11
  cache
• 9:12
  and
• 9:12
  the
• 9:12
  next
• 9:12
  request
• 9:13
  rewrites
• 9:13
  a
• 9:13
  300K
• 9:14
  to
• 9:14
  600K
• 9:15
  token
• 9:15
  context
• 9:16
  at
• 9:16
  cache-write
• 9:16
  rates.
• 9:18
  A
• 9:18
  one-line
• 9:18
  settings
• 9:18
  change
• 9:19
  to
• 9:19
  a
• 9:19
  one-hour
• 9:20
  sub-agent
• 9:20
  TTL,
• 9:21
  plus
• 9:21
  keeping
• 9:22
  sub-agents
• 9:22
  out
• 9:22
  of
• 9:22
  long
• 9:23
  commands,
• 9:24
  cut
• 9:24
  measured
• 9:24
  cache
• 9:25
  writes
• 9:25
  about
• 9:25
  75%
• 9:26
  .
• 9:27
  On
• 9:27
  the
• 9:27
  pi
• 9:27
  coding
• 9:27
  agent,
• 9:28
  users
• 9:28
  are
• 9:28
  weighing
• 9:29
  pi-vcc,
• 9:30
  which
• 9:30
  compacts
• 9:31
  context
• 9:31
  sub-second
• 9:32
  without
• 9:32
  a
• 9:32
  model,
• 9:32
  against
• 9:33
  pi-blackhole,
• 9:34
  which
• 9:34
  adds
• 9:34
  LLM-built
• 9:35
  observational
• 9:36
  memory;
• 9:36
  the
• 9:37
  early
• 9:37
  verdict
• 9:37
  is
• 9:37
  a
• 9:37
  small
• 9:38
  improvement,
• 9:39
  and
• 9:39
  nobody
• 9:39
  has
• 9:39
  deterministic
• 9:40
  tests
• 9:40
  .
• 9:41
  forkd
• 9:41
  addresses
• 9:42
  the
• 9:42
  other
• 9:42
  harness
• 9:42
  cost,
• 9:43
  environment
• 9:43
  boot.
• 9:44
  Built
• 9:45
  on
• 9:45
  Firecracker,
• 9:46
  it
• 9:46
  boots
• 9:46
  a
• 9:46
  parent
• 9:46
  microVM
• 9:47
  once,
• 9:47
  snapshots
• 9:48
  it,
• 9:48
  and
• 9:49
  lets
• 9:49
  each
• 9:49
  child
• 9:49
  map
• 9:50
  the
• 9:50
  memory
• 9:50
  image
• 9:50
  copy-on-write.
• 9:52
  Claimed:
• 9:52
  100
• 9:53
  microVMs
• 9:54
  forked
• 9:54
  in
• 9:54
  101
• 9:55
  ms
• 9:55
  at
• 9:55
  0.12
• 9:56
  MiB
• 9:56
  host
• 9:57
  memory
• 9:57
  per
• 9:57
  child,
• 9:58
  versus
• 9:58
  759
• 10:00
  ms
• 10:00
  and
• 10:00
  84
• 10:00
  MiB
• 10:01
  for
• 10:01
  a
• 10:01
  Firecracker
• 10:02
  cold
• 10:02
  boot,
• 10:03
  a
• 10:03
  warm-fork
• 10:03
  against
• 10:03
  cold-boot
• 10:04
  comparison
• 10:04
  the
• 10:04
  authors
• 10:05
  acknowledge.
• 10:06
  It
• 10:06
  is
• 10:06
  alpha
• 10:06
  at
• 10:07
  v0.5.3,
• 10:08
  x86
• 10:09
  64
• 10:10
  Linux
• 10:11
  only,
• 10:11
  has
• 10:11
  no
• 10:12
  default-deny
• 10:12
  egress,
• 10:13
  and
• 10:13
  has
• 10:13
  had
• 10:13
  no
• 10:13
  third-party
• 10:14
  audit
• 10:14
  .
• 10:16
  Cactus
• 10:17
  Compute's
• 10:17
  Needle
• 10:17
  3
• 10:18
  is
• 10:18
  a
• 10:18
  121M-parameter
• 10:20
  model
• 10:20
  that
• 10:20
  does
• 10:20
  not
• 10:20
  chat:
• 10:21
  given
• 10:22
  an
• 10:22
  app's
• 10:22
  declared
• 10:22
  functions
• 10:23
  it
• 10:23
  returns
• 10:23
  tool
• 10:24
  calls
• 10:24
  with
• 10:24
  every
• 10:24
  argument
• 10:25
  filled,
• 10:25
  or
• 10:25
  a
• 10:25
  typed
• 10:26
  record
• 10:26
  against
• 10:26
  a
• 10:26
  schema,
• 10:27
  fully
• 10:27
  on-device.
• 10:28
  Dense
• 10:29
  feed-forward
• 10:29
  layers
• 10:29
  give
• 10:30
  way
• 10:30
  to
• 10:30
  a
• 10:30
  tiny
• 10:30
  Monarch
• 10:31
  Hadamard
• 10:31
  MLP,
• 10:32
  and
• 10:32
  factual
• 10:33
  knowledge
• 10:33
  sits
• 10:33
  in
• 10:34
  an
• 10:34
  "engram"
• 10:34
  of
• 10:34
  hashed
• 10:35
  n-gram
• 10:35
  tables
• 10:35
  read
• 10:36
  by
• 10:36
  gather,
• 10:36
  so
• 10:37
  a
• 10:37
  token
• 10:37
  costs
• 10:37
  100
• 10:38
  MFLOPs
• 10:38
  versus
• 10:39
  296
• 10:40
  for
• 10:40
  an
• 10:40
  equivalent
• 10:40
  transformer.
• 10:41
  "Sliceable"
• 10:42
  means
• 10:42
  one
• 10:43
  set
• 10:43
  of
• 10:43
  weights
• 10:43
  in
• 10:43
  which
• 10:44
  every
• 10:44
  depth
• 10:44
  from
• 10:44
  2
• 10:45
  to
• 10:45
  20
• 10:45
  layers
• 10:45
  is
• 10:45
  a
• 10:45
  deployable
• 10:46
  model.
• 10:47
  With
• 10:47
  2.125-bit
• 10:48
  quantization
• 10:49
  the
• 10:49
  binaries
• 10:50
  run
• 10:50
  8
• 10:50
  to
• 10:50
  29MB,
• 10:52
  the
• 10:52
  CPU-only
• 10:52
  engine
• 10:53
  is
• 10:53
  under
• 10:53
  1MB,
• 10:54
  and
• 10:55
  a
• 10:55
  Raspberry
• 10:55
  Pi
• 10:55
  5
• 10:55
  decodes
• 10:56
  up
• 10:56
  to
• 10:56
  4k
• 10:57
  tokens/s
• 10:57
  at
• 10:58
  the
• 10:58
  bottom
• 10:58
  of
• 10:58
  the
• 10:58
  ladder,
• 10:59
  under
• 10:59
  Apache
• 10:59
  2
• 10:59
  .
• 11:00
  The
• 11:00
  headline
• 11:01
  needs
• 11:01
  its
• 11:01
  fine
• 11:01
  print.
• 11:02
  On
• 11:02
  Mobile
• 11:02
  Actions,
• 11:03
  961
• 11:04
  phone
• 11:04
  commands
• 11:05
  scored
• 11:05
  by
• 11:05
  exact
• 11:05
  match,
• 11:06
  the
• 11:06
  20-layer
• 11:07
  model
• 11:07
  hits
• 11:07
  86.0
• 11:08
  versus
• 11:09
  DeepSeek
• 11:09
  V4
• 11:09
  Flash
• 11:10
  at
• 11:10
  88.4
• 11:11
  via
• 11:11
  API.
• 11:12
  Everywhere
• 11:13
  else
• 11:13
  the
• 11:13
  gap
• 11:13
  is
• 11:13
  wide:
• 11:14
  DroidCall
• 11:15
  47.0
• 11:16
  against
• 11:16
  60.5,
• 11:17
  BFCL
• 11:17
  v4
• 11:18
  50.2
• 11:19
  against
• 11:19
  77.2,
• 11:20
  and
• 11:21
  the
• 11:21
  4-layer
• 11:21
  model
• 11:21
  scores
• 11:22
  11.7
• 11:23
  on
• 11:23
  Mobile
• 11:23
  Actions.
• 11:24
  The
• 11:24
  "surpasses
• 11:25
  DeepSeek"
• 11:26
  claim
• 11:26
  applies
• 11:26
  only
• 11:26
  after
• 11:27
  product-specific
• 11:27
  fine-tuning,
• 11:28
  and
• 11:28
  thread
• 11:29
  skeptics
• 11:29
  called
• 11:29
  it
• 11:30
  "a
• 11:30
  benchmark
• 11:30
  niche-pick."
• 11:31
  One
• 11:32
  set
• 11:32
  of
• 11:32
  weights
• 11:32
  spanning
• 11:32
  a
• 11:32
  Pi
• 11:33
  and
• 11:33
  a
• 11:33
  watch
• 11:33
  is
• 11:33
  a
• 11:33
  genuine
• 11:34
  engineering
• 11:34
  result;
• 11:35
  matching
• 11:36
  a
• 11:36
  frontier
• 11:36
  model
• 11:36
  is
• 11:37
  a
• 11:37
  claim
• 11:37
  about
• 11:37
  one
• 11:37
  benchmark.
• 11:38
  Fastino's
• 11:39
  GLiNER2.5
• 11:42
  makes
• 11:43
  the
• 11:43
  structural
• 11:43
  change
• 11:43
  the
• 11:44
  extraction
• 11:44
  family
• 11:44
  needed.
• 11:45
  Earlier
• 11:46
  GLiNER
• 11:46
  models
• 11:47
  enumerated
• 11:47
  every
• 11:47
  candidate
• 11:48
  span
• 11:48
  up
• 11:48
  to
• 11:48
  about
• 11:49
  twelve
• 11:49
  words,
• 11:50
  so
• 11:50
  longer
• 11:50
  entities
• 11:51
  were
• 11:51
  never
• 11:51
  scored
• 11:51
  at
• 11:51
  all.
• 11:52
  GLiNER2.5
• 11:55
  predicts
• 11:55
  boundaries
• 11:56
  instead:
• 11:57
  per
• 11:57
  query
• 11:57
  it
• 11:57
  emits
• 11:57
  start,
• 11:58
  end
• 11:58
  and
• 11:59
  inside
• 11:59
  scores,
• 12:00
  pairs
• 12:00
  promising
• 12:01
  starts
• 12:01
  and
• 12:01
  ends
• 12:01
  with
• 12:02
  no
• 12:02
  distance
• 12:02
  restriction,
• 12:03
  and
• 12:03
  reranks,
• 12:04
  so
• 12:04
  compute
• 12:05
  is
• 12:05
  linear
• 12:05
  in
• 12:05
  sequence
• 12:05
  length
• 12:06
  for
• 12:06
  a
• 12:06
  fixed
• 12:06
  schema.
• 12:07
  The
• 12:07
  vendor-run
• 12:08
  numbers
• 12:08
  are
• 12:08
  mixed:
• 12:09
  zero-shot
• 12:10
  macro-F1
• 12:11
  across
• 12:11
  16
• 12:12
  benchmarks
• 12:12
  moves
• 12:12
  from
• 12:13
  53.34
• 12:14
  to
• 12:14
  54.87
• 12:15
  for
• 12:15
  the
• 12:15
  base
• 12:16
  model,
• 12:17
  XNLI
• 12:17
  jumps
• 12:17
  24.75
• 12:19
  points
• 12:19
  for
• 12:19
  the
• 12:19
  multilingual
• 12:20
  model,
• 12:21
  and
• 12:21
  the
• 12:21
  post
• 12:21
  concedes
• 12:22
  per-dataset
• 12:22
  regressions.
• 12:24
  There
• 12:24
  are
• 12:24
  no
• 12:24
  head-to-head
• 12:24
  LLM
• 12:25
  numbers
• 12:25
  .
• 12:26
  Laya
• 12:26
  continues
• 12:27
  the
• 12:27
  community's
• 12:27
  reverse-engineering
• 12:28
  of
• 12:28
  TypeSafe's
• 12:29
  Jev:
• 12:30
  a
• 12:30
  421M
• 12:31
  non-autoregressive
• 12:32
  decision
• 12:33
  model
• 12:33
  pairing
• 12:33
  a
• 12:33
  ModernBERT-large
• 12:34
  encoder
• 12:35
  with
• 12:35
  a
• 12:35
  transformer
• 12:36
  head
• 12:36
  that
• 12:36
  scores
• 12:37
  masked
• 12:37
  option
• 12:37
  markers
• 12:38
  to
• 12:38
  resolve
• 12:38
  typed
• 12:39
  schemas,
• 12:40
  with
• 12:40
  an
• 12:40
  unofficial
• 12:40
  RLCD
• 12:41
  stage
• 12:42
  that
• 12:42
  rewards
• 12:42
  only
• 12:42
  calibrated
• 12:43
  probabilities
• 12:44
  under
• 12:44
  strictly
• 12:44
  proper
• 12:45
  scoring
• 12:45
  rules.
• 12:45
  .
• 12:48
  Training-free
• 12:48
  speculative
• 12:49
  decoders
• 12:49
  share
• 12:49
  one
• 12:50
  key:
• 12:50
  match
• 12:51
  the
• 12:51
  longest
• 12:51
  exact
• 12:52
  suffix
• 12:52
  of
• 12:52
  the
• 12:52
  context
• 12:53
  against
• 12:53
  a
• 12:53
  pool
• 12:53
  of
• 12:53
  past
• 12:54
  text
• 12:54
  and
• 12:54
  propose
• 12:54
  what
• 12:54
  followed.
• 12:56
  Researchers
• 12:56
  at
• 12:56
  JAIST
• 12:57
  show
• 12:57
  that
• 12:57
  on
• 12:57
  tool-calling
• 12:58
  traffic
• 12:58
  this
• 12:58
  key
• 12:58
  "does
• 12:58
  not
• 12:59
  degrade
• 12:59
  there:
• 13:00
  it
• 13:00
  dies."
• 13:01
  A
• 13:01
  single
• 13:01
  freshly
• 13:02
  minted
• 13:02
  value,
• 13:03
  an
• 13:03
  ID
• 13:03
  or
• 13:04
  a
• 13:04
  name,
• 13:04
  zeroes
• 13:05
  every
• 13:05
  suffix
• 13:05
  containing
• 13:05
  it.
• 13:06
  Oilbird
• 13:07
  adds
• 13:07
  a
• 13:07
  second,
• 13:07
  semantic
• 13:08
  key
• 13:08
  over
• 13:08
  the
• 13:08
  same
• 13:08
  pool:
• 13:09
  every
• 13:09
  committed
• 13:10
  token
• 13:10
  is
• 13:10
  tagged
• 13:11
  with
• 13:11
  the
• 13:11
  hidden
• 13:11
  state
• 13:11
  the
• 13:11
  verifier
• 13:12
  already
• 13:12
  computed
• 13:13
  during
• 13:13
  verification,
• 13:14
  and
• 13:14
  "capture
• 13:14
  costs
• 13:15
  a
• 13:15
  copy,
• 13:16
  never
• 13:16
  a
• 13:16
  forward
• 13:16
  pass."
• 13:17
  Retrieval
• 13:18
  is
• 13:18
  nearest-neighbor
• 13:18
  at
• 13:18
  every
• 13:19
  position,
• 13:20
  and
• 13:20
  the
• 13:20
  greedy
• 13:20
  lossless
• 13:21
  acceptance
• 13:21
  rule
• 13:21
  is
• 13:21
  untouched
• 13:22
  .
• 13:22
  On
• 13:23
  Llama-3.1-8B
• 13:24
  over
• 13:24
  API-Bank
• 13:25
  that
• 13:25
  yields
• 13:25
  6.76
• 13:26
  accepted
• 13:27
  tokens
• 13:27
  per
• 13:27
  pass
• 13:28
  at
• 13:28
  4.39x
• 13:29
  autoregressive
• 13:29
  speed,
• 13:30
  versus
• 13:30
  SuffixDecoding
• 13:31
  at
• 13:31
  3.88x
• 13:32
  and
• 13:32
  EAGLE-3
• 13:33
  at
• 13:33
  2.03x.
• 13:34
  The
• 13:34
  honesty
• 13:35
  is
• 13:35
  in
• 13:35
  the
• 13:35
  limitations:
• 13:36
  EAGLE-3
• 13:37
  wins
• 13:37
  on
• 13:37
  all
• 13:37
  five
• 13:38
  non-repetitive
• 13:38
  workloads,
• 13:39
  the
• 13:39
  store
• 13:39
  grows
• 13:40
  up
• 13:40
  to
• 13:40
  2.6
• 13:41
  GB
• 13:41
  per
• 13:41
  benchmark
• 13:42
  unbounded,
• 13:43
  the
• 13:43
  advantage
• 13:43
  collapses
• 13:44
  toward
• 13:44
  parity
• 13:44
  at
• 13:44
  batch
• 13:45
  32,
• 13:45
  and
• 13:46
  the
• 13:46
  store
• 13:46
  is
• 13:46
  a
• 13:46
  cache
• 13:47
  of
• 13:47
  user
• 13:47
  content
• 13:47
  that
• 13:47
  needs
• 13:48
  per-tenant
• 13:48
  governance.
• 13:49
  For
• 13:49
  agent
• 13:49
  serving,
• 13:50
  where
• 13:50
  the
• 13:50
  same
• 13:50
  schemas
• 13:51
  and
• 13:51
  IDs
• 13:51
  recur
• 13:51
  across
• 13:52
  turns,
• 13:52
  it
• 13:53
  is
• 13:53
  a
• 13:53
  real
• 13:53
  win;
• 13:54
  for
• 13:54
  chat
• 13:54
  it
• 13:54
  is
• 13:54
  not.
• 13:55
  The
• 13:55
  local
• 13:56
  crowd
• 13:56
  is
• 13:56
  extracting
• 13:57
  speed
• 13:57
  from
• 13:57
  a
• 13:57
  different
• 13:58
  bottleneck.
• 13:59
  One
• 13:59
  user
• 13:59
  runs
• 14:00
  Qwen3.8-Flash-Next,
• 14:03
  a
• 14:04
  125B-A6B
• 14:05
  MoE
• 14:06
  with
• 14:06
  a
• 14:06
  51B-parameter
• 14:07
  n-gram
• 14:07
  table,
• 14:08
  on
• 14:08
  an
• 14:08
  RTX
• 14:09
  4070
• 14:10
  12GB
• 14:10
  with
• 14:10
  64GB
• 14:11
  DDR5,
• 14:12
  going
• 14:13
  from
• 14:13
  6
• 14:13
  to
• 14:13
  nearly
• 14:13
  20
• 14:14
  tok/s
• 14:15
  via
• 14:15
  a
• 14:15
  4.27
• 14:16
  bpw
• 14:17
  quant,
• 14:17
  lazy
• 14:18
  SSD
• 14:18
  offloading
• 14:19
  of
• 14:19
  the
• 14:19
  n-gram
• 14:19
  table,
• 14:20
  and
• 14:20
  an
• 14:20
  unmerged
• 14:21
  multi-token-prediction
• 14:22
  PR
• 14:22
  whose
• 14:22
  compact
• 14:23
  head
• 14:23
  yields
• 14:23
  77
• 14:24
  to
• 14:24
  96%
• 14:25
  acceptance.
• 14:26
  The
• 14:26
  catch
• 14:26
  is
• 14:26
  memory
• 14:27
  bandwidth:
• 14:28
  a
• 14:28
  3090
• 14:29
  owner
• 14:29
  on
• 14:29
  DDR4
• 14:30
  cannot
• 14:31
  reliably
• 14:31
  reach
• 14:31
  20,
• 14:32
  because
• 14:32
  the
• 14:32
  active
• 14:33
  experts
• 14:33
  stream
• 14:34
  from
• 14:34
  system
• 14:34
  RAM
• 14:34
  .
• 14:35
  On
• 14:35
  low-memory
• 14:36
  Macs
• 14:36
  streaming
• 14:36
  experts
• 14:37
  from
• 14:37
  SSD,
• 14:38
  the
• 14:38
  slotstream
• 14:39
  author
• 14:39
  found
• 14:39
  the
• 14:39
  model
• 14:40
  is
• 14:40
  its
• 14:40
  own
• 14:40
  best
• 14:40
  predictor:
• 14:42
  running
• 14:42
  Qwen's
• 14:42
  router
• 14:43
  two
• 14:43
  layers
• 14:43
  ahead
• 14:44
  on
• 14:44
  the
• 14:44
  current
• 14:44
  hidden
• 14:44
  state
• 14:44
  to
• 14:45
  prefetch
• 14:45
  experts
• 14:45
  gave
• 14:45
  a
• 14:46
  10%+
• 14:46
  gain
• 14:46
  .
• 14:47
  A
• 14:47
  more
• 14:47
  speculative
• 14:48
  paper
• 14:48
  asks
• 14:48
  whether
• 14:49
  the
• 14:49
  expert
• 14:49
  bank
• 14:49
  needs
• 14:49
  to
• 14:50
  be
• 14:50
  stored
• 14:50
  at
• 14:50
  all.
• 14:51
  "Infinite-Parameter
• 14:52
  LLMs"
• 14:52
  replaces
• 14:53
  an
• 14:53
  MoE's
• 14:53
  stored
• 14:53
  experts
• 14:54
  with
• 14:54
  a
• 14:54
  hypernetwork
• 14:54
  that
• 14:54
  generates
• 14:55
  a
• 14:55
  low-rank
• 14:55
  feed-forward
• 14:56
  delta
• 14:56
  per
• 14:56
  token
• 14:57
  from
• 14:57
  a
• 14:57
  latent
• 14:57
  code
• 14:58
  compiled
• 14:58
  out
• 14:58
  of
• 14:58
  live
• 14:59
  evidence.
• 15:00
  The
• 15:00
  name
• 15:00
  is
• 15:00
  defined
• 15:01
  narrowly:
• 15:01
  an
• 15:02
  unbounded
• 15:02
  set
• 15:02
  of
• 15:02
  reachable
• 15:03
  weights
• 15:03
  from
• 15:03
  a
• 15:03
  fixed
• 15:03
  footprint,
• 15:04
  not
• 15:04
  unbounded
• 15:05
  knowledge.
• 15:06
  On
• 15:06
  a
• 15:06
  frozen
• 15:07
  Qwen3-8B,
• 15:08
  in-context
• 15:09
  prompting
• 15:09
  still
• 15:09
  wins
• 15:10
  on
• 15:10
  short
• 15:10
  clean
• 15:10
  evidence,
• 15:11
  85.3
• 15:12
  against
• 15:12
  51.8
• 15:13
  F1
• 15:14
  on
• 15:14
  SQuAD,
• 15:15
  while
• 15:15
  compiled
• 15:15
  weights
• 15:16
  win
• 15:16
  on
• 15:16
  long
• 15:16
  noisy
• 15:16
  evidence,
• 15:17
  48.0
• 15:18
  against
• 15:19
  33.6
• 15:19
  on
• 15:20
  MS
• 15:20
  MARCO.
• 15:21
  A
• 15:21
  single
• 15:22
  code
• 15:22
  saturates
• 15:22
  fast
• 15:23
  as
• 15:23
  passages
• 15:23
  pile
• 15:24
  in,
• 15:24
  and
• 15:24
  full
• 15:24
  training
• 15:25
  details
• 15:25
  are
• 15:25
  omitted
• 15:25
  .
• 15:28
  HP's
• 15:28
  ZGX
• 15:29
  Fury
• 15:29
  AI
• 15:29
  station
• 15:30
  is
• 15:30
  orderable,
• 15:31
  and
• 15:31
  it
• 15:31
  puts
• 15:31
  a
• 15:31
  GB300
• 15:33
  on
• 15:34
  a
• 15:34
  desk.
• 15:34
  One
• 15:35
  Blackwell
• 15:35
  Ultra
• 15:36
  GPU
• 15:36
  with
• 15:36
  252GB
• 15:38
  of
• 15:38
  HBM3e
• 15:39
  at
• 15:39
  7.1TB/s
• 15:41
  sits
• 15:41
  beside
• 15:42
  a
• 15:42
  72-core
• 15:42
  Grace
• 15:43
  CPU
• 15:43
  with
• 15:43
  496GB
• 15:45
  of
• 15:45
  LPDDR5X
• 15:48
  over
• 15:48
  NVLink-C2C,
• 15:49
  for
• 15:49
  748GB
• 15:51
  of
• 15:51
  coherent
• 15:51
  memory
• 15:52
  and
• 15:52
  up
• 15:52
  to
• 15:52
  20
• 15:52
  petaFLOPS
• 15:53
  of
• 15:53
  FP4;
• 15:54
  HP's
• 15:54
  footnote
• 15:55
  is
• 15:55
  that
• 15:55
  the
• 15:55
  trillion-parameter
• 15:56
  inference
• 15:56
  claim
• 15:56
  assumes
• 15:57
  FP4
• 15:57
  quantization.
• 15:59
  HP
• 15:59
  has
• 15:59
  not
• 15:59
  published
• 16:00
  a
• 16:00
  price,
• 16:00
  but
• 16:01
  the
• 16:01
  same
• 16:01
  silicon
• 16:01
  in
• 16:01
  MSI's
• 16:02
  WS300
• 16:04
  has
• 16:04
  been
• 16:04
  offered
• 16:05
  at
• 16:05
  around
• 16:05
  $100,000,
• 16:06
  and
• 16:06
  that
• 16:07
  is
• 16:07
  the
• 16:07
  number
• 16:07
  to
• 16:07
  budget
• 16:07
  against.
• 16:09
  For
• 16:09
  scale,
• 16:10
  HP's
• 16:10
  GB10-based
• 16:11
  ZGX
• 16:11
  Nano
• 16:11
  lists
• 16:12
  at
• 16:12
  $5,649
• 16:14
  to
• 16:14
  $6,699
• 16:15
  with
• 16:16
  128GB
• 16:17
  of
• 16:17
  unified
• 16:18
  memory
• 16:18
  at
• 16:18
  273GB/s,
• 16:21
  so
• 16:21
  the
• 16:21
  Fury
• 16:21
  is
• 16:21
  roughly
• 16:22
  15
• 16:22
  to
• 16:22
  18
• 16:23
  times
• 16:23
  the
• 16:23
  price
• 16:23
  for
• 16:23
  about
• 16:23
  6
• 16:24
  times
• 16:24
  the
• 16:24
  memory.
• 16:25
  The
• 16:25
  premium
• 16:26
  buys
• 16:26
  HBM,
• 16:26
  not
• 16:27
  capacity:
• 16:27
  the
• 16:27
  252GB
• 16:29
  GPU
• 16:29
  pool
• 16:29
  is
• 16:29
  26
• 16:30
  times
• 16:30
  faster
• 16:30
  than
• 16:31
  the
• 16:31
  Nano's
• 16:31
  memory,
• 16:31
  while
• 16:32
  the
• 16:32
  Fury's
• 16:32
  396GB/s
• 16:34
  CPU-side
• 16:35
  pool
• 16:35
  is
• 16:35
  only
• 16:35
  about
• 16:36
  1.5
• 16:36
  times
• 16:36
  faster.
• 16:37
  That
• 16:37
  number
• 16:38
  will
• 16:38
  decide
• 16:38
  whether
• 16:38
  a
• 16:38
  large
• 16:39
  MoE
• 16:39
  runs
• 16:39
  well,
• 16:40
  because
• 16:40
  anything
• 16:41
  that
• 16:41
  spills
• 16:41
  out
• 16:41
  of
• 16:41
  HBM
• 16:42
  is
• 16:42
  running
• 16:42
  on
• 16:42
  memory
• 16:42
  not
• 16:42
  much
• 16:42
  quicker
• 16:43
  than
• 16:43
  the
• 16:43
  $6,000
• 16:44
  box.
• 16:45
  The
• 16:45
  Red
• 16:45
  Hat
• 16:45
  AI
• 16:45
  Factory
• 16:46
  integration
• 16:46
  is
• 16:46
  a
• 16:46
  plan
• 16:47
  rather
• 16:47
  than
• 16:47
  a
• 16:47
  SKU.
• 16:48
  The
• 16:48
  WS300
• 16:50
  held
• 16:50
  71C
• 16:51
  under
• 16:51
  a
• 16:51
  1,400W
• 16:54
  liquid
• 16:55
  loop
• 16:55
  .
• 16:55
  At
• 16:56
  the
• 16:56
  other
• 16:56
  end
• 16:56
  of
• 16:56
  the
• 16:56
  budget,
• 16:57
  a
• 16:57
  home
• 16:57
  voice
• 16:57
  assistant
• 16:57
  on
• 16:57
  an
• 16:58
  RTX
• 16:58
  3060
• 16:59
  replicated
• 16:59
  OpenAI's
• 17:00
  GPT
• 17:00
  Live
• 17:01
  "Improved
• 17:01
  Intelligence"
• 17:02
  demo
• 17:02
  with
• 17:02
  Qwen3
• 17:03
  1.7B
• 17:03
  ASR,
• 17:04
  Qwen3.5-9B
• 17:06
  at
• 17:06
  12K
• 17:07
  context,
• 17:07
  and
• 17:07
  Pocket
• 17:08
  TTS
• 17:08
  in
• 17:08
  12GB
• 17:09
  of
• 17:09
  VRAM.
• 17:10
  It
• 17:10
  could
• 17:10
  not
• 17:10
  find
• 17:10
  a
• 17:10
  feasible
• 17:11
  flight
• 17:11
  route
• 17:11
  but
• 17:11
  delivered
• 17:11
  restaurant
• 17:12
  recommendations
• 17:12
  in
• 17:12
  under
• 17:12
  six
• 17:13
  and
• 17:13
  a
• 17:13
  half
• 17:13
  minutes;
• 17:14
  a
• 17:14
  commenter
• 17:15
  noted
• 17:15
  that
• 17:15
  route
• 17:15
  feasibility
• 17:16
  is
• 17:16
  a
• 17:16
  chain
• 17:16
  whose
• 17:16
  early
• 17:17
  results
• 17:17
  must
• 17:17
  survive
• 17:17
  in
• 17:17
  a
• 17:18
  12K
• 17:18
  window
• 17:18
  .
• 17:19
  Another
• 17:19
  user
• 17:20
  wired
• 17:20
  a
• 17:20
  local
• 17:20
  Qwen3.6-35B-A3B
• 17:24
  to
• 17:24
  GIMP
• 17:25
  through
• 17:25
  an
• 17:25
  MCP
• 17:25
  server
• 17:26
  via
• 17:26
  llama.cpp
• 17:27
  and
• 17:27
  asked
• 17:27
  it
• 17:27
  to
• 17:27
  draw
• 17:27
  a
• 17:27
  flower;
• 17:28
  the
• 17:28
  result
• 17:29
  was
• 17:29
  poor
• 17:29
  for
• 17:29
  the
• 17:29
  reason
• 17:29
  a
• 17:29
  commenter
• 17:30
  nailed,
• 17:30
  the
• 17:30
  model
• 17:31
  cannot
• 17:31
  see
• 17:31
  the
• 17:31
  canvas
• 17:31
  between
• 17:32
  tool
• 17:32
  calls,
• 17:32
  so
• 17:32
  each
• 17:32
  shape
• 17:33
  is
• 17:33
  placed
• 17:33
  blind
• 17:33
  and
• 17:33
  errors
• 17:33
  compound
• 17:34
  .
• 17:35
  NeuralScreen
• 17:35
  runs
• 17:35
  an
• 17:35
  entire
• 17:36
  Windows
• 17:36
  desktop
• 17:36
  through
• 17:37
  NVIDIA's
• 17:37
  DLSS
• 17:38
  neural
• 17:38
  renderer,
• 17:39
  claiming
• 17:39
  45.7
• 17:40
  to
• 17:40
  72.6
• 17:41
  fps
• 17:41
  at
• 17:41
  4K
• 17:42
  on
• 17:42
  a
• 17:42
  5070
• 17:43
  Ti.
• 17:43
  Read
• 17:44
  the
• 17:44
  notice
• 17:44
  before
• 17:44
  running
• 17:45
  it:
• 17:45
  the
• 17:45
  archive
• 17:46
  bundles
• 17:46
  what
• 17:46
  the
• 17:46
  README
• 17:48
  itself
• 17:48
  calls
• 17:49
  "the
• 17:49
  leaked
• 17:49
  310.8.0
• 17:51
  build"
• 17:51
  of
• 17:51
  the
• 17:51
  neural-renderer
• 17:52
  DLL,
• 17:53
  the
• 17:53
  executable
• 17:53
  is
• 17:53
  unsigned,
• 17:54
  and
• 17:54
  "DLSS
• 17:55
  5"
• 17:56
  is
• 17:56
  the
• 17:56
  author's
• 17:56
  branding
• 17:57
  on
• 17:57
  310.x
• 17:57
  runtimes
• 17:58
  .
• 17:59
  The
• 17:59
  Event
• 17:59
  Horizon
• 18:00
  Observatory
• 18:00
  is
• 18:01
  a
• 18:01
  one-shot
• 18:01
  WebGL
• 18:02
  black
• 18:02
  hole
• 18:02
  exhibit
• 18:02
  generated
• 18:03
  with
• 18:03
  Astra:
• 18:04
  it
• 18:04
  traces
• 18:05
  rays
• 18:05
  backward
• 18:05
  with
• 18:05
  the
• 18:05
  Schwarzschild
• 18:06
  null-geodesic
• 18:07
  equation,
• 18:08
  marks
• 18:08
  the
• 18:08
  critical
• 18:08
  impact
• 18:09
  parameter
• 18:09
  at
• 18:09
  3√3
• 18:10
  GM/c²,
• 18:12
  and
• 18:12
  is
• 18:12
  precise
• 18:12
  about
• 18:13
  its
• 18:13
  own
• 18:13
  limits,
• 18:14
  leaving
• 18:14
  frame
• 18:14
  dragging
• 18:15
  unsolved.
• 18:16
  A
• 18:16
  generated
• 18:16
  exhibit
• 18:17
  that
• 18:17
  documents
• 18:17
  what
• 18:17
  it
• 18:17
  did
• 18:17
  not
• 18:18
  simulate
• 18:18
  is
• 18:19
  worth
• 18:19
  more
• 18:19
  than
• 18:19
  one
• 18:19
  that
• 18:19
  fakes
• 18:20
  it
• 18:20
  .