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

Transcript

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

• 0:00
  An
• 0:00
  18-year-old
• 0:00
  heap
• 0:01
  buffer
• 0:01
  overflow
• 0:01
  in
• 0:01
  NGINX's
• 0:02
  ngx
• 0:02
  http
• 0:03
  rewrite
• 0:04
  module
• 0:04
  has
• 0:04
  been
• 0:04
  sitting
• 0:04
  in
• 0:05
  production
• 0:05
  since
• 0:05
  2008,
• 0:06
  and
• 0:06
  it's
• 0:07
  a
• 0:07
  full
• 0:07
  unauthenticated
• 0:07
  RCE.
• 0:09
  CVE-2026-42945
• 0:12
  exploits
• 0:12
  a
• 0:12
  mismatch
• 0:13
  between
• 0:13
  NGINX's
• 0:13
  two-pass
• 0:14
  script
• 0:14
  engine:
• 0:15
  the
• 0:15
  length-calculation
• 0:16
  pass
• 0:16
  runs
• 0:16
  on
• 0:17
  a
• 0:17
  freshly
• 0:17
  zeroed
• 0:17
  sub-engine
• 0:18
  where
• 0:18
  is
• 0:18
  args
• 0:18
  is
• 0:18
  0,
• 0:19
  so
• 0:19
  it
• 0:19
  returns
• 0:20
  the
• 0:20
  raw
• 0:20
  capture
• 0:20
  length.
• 0:21
  The
• 0:21
  copy
• 0:22
  pass
• 0:22
  sees
• 0:22
  is
• 0:22
  args
• 0:23
  =
• 0:23
  1
• 0:23
  from
• 0:24
  the
• 0:24
  parent
• 0:24
  engine
• 0:25
  and
• 0:25
  calls
• 0:25
  ngx
• 0:25
  escape
• 0:26
  uri
• 0:26
  with
• 0:26
  NGX
• 0:27
  ESCAPE
• 0:27
  ARGS,
• 0:28
  expanding
• 0:29
  each
• 0:29
  escapable
• 0:29
  byte
• 0:30
  to
• 0:30
  three
• 0:30
  bytes.
• 0:31
  The
• 0:31
  copy
• 0:31
  overflows
• 0:32
  the
• 0:32
  undersized
• 0:32
  heap
• 0:32
  buffer
• 0:33
  with
• 0:33
  attacker-controlled
• 0:34
  URI
• 0:34
  data.
• 0:35
  Exploitation
• 0:36
  uses
• 0:36
  cross-request
• 0:37
  heap
• 0:37
  feng
• 0:37
  shui
• 0:37
  —
• 0:38
  spraying
• 0:38
  ngx
• 0:39
  pool
• 0:39
  t
• 0:39
  structures
• 0:40
  via
• 0:40
  POST
• 0:40
  bodies
• 0:40
  to
• 0:40
  corrupt
• 0:41
  an
• 0:41
  adjacent
• 0:41
  pool's
• 0:42
  cleanup
• 0:42
  pointer,
• 0:43
  redirecting
• 0:43
  it
• 0:43
  to
• 0:43
  a
• 0:44
  fake
• 0:44
  cleanup
• 0:44
  handler
• 0:44
  that
• 0:44
  invokes
• 0:45
  system().
• 0:46
  Every
• 0:46
  NGINX
• 0:47
  Open
• 0:47
  Source
• 0:47
  build
• 0:47
  from
• 0:48
  0.6.27
• 0:49
  through
• 0:49
  1.30.0,
• 0:51
  and
• 0:51
  NGINX
• 0:52
  Plus
• 0:52
  R32
• 0:53
  through
• 0:53
  R36,
• 0:54
  is
• 0:54
  affected.
• 0:55
  The
• 0:55
  vulnerability
• 0:56
  was
• 0:56
  autonomously
• 0:56
  discovered
• 0:57
  by
• 0:57
  DepthFirst's
• 0:58
  security
• 0:58
  analysis
• 0:58
  system
• 0:59
  along
• 0:59
  with
• 0:59
  three
• 1:00
  additional
• 1:00
  memory
• 1:00
  corruption
• 1:01
  issues
• 1:04
  (CVE-2026-42946,
• 1:07
  CVE-2026-40701,
• 1:10
  CVE-2026-42934),
• 1:14
  all
• 1:14
  from
• 1:14
  a
• 1:14
  single
• 1:14
  onboarding
• 1:15
  of
• 1:15
  the
• 1:15
  NGINX
• 1:16
  source
• 1:16
  .
• 1:16
  The
• 1:17
  IoT
• 1:17
  security
• 1:17
  narrative
• 1:18
  got
• 1:18
  another
• 1:18
  chapter
• 1:19
  courtesy
• 1:19
  of
• 1:19
  a
• 1:19
  Honeywell
• 1:20
  X2S
• 1:20
  smart
• 1:20
  thermostat
• 1:21
  teardown.
• 1:22
  The
• 1:22
  device
• 1:22
  packs
• 1:23
  a
• 1:23
  200
• 1:23
  MHz
• 1:24
  Renesas
• 1:24
  Cortex-M33
• 1:26
  with
• 1:26
  TrustZone
• 1:26
  and
• 1:26
  a
• 1:26
  Realtek
• 1:27
  RTL8721DM
• 1:29
  Wi-Fi/BLE
• 1:30
  SoC
• 1:32
  —
• 1:34
  two
• 1:34
  brains,
• 1:35
  two
• 1:35
  Winbond
• 1:36
  flash
• 1:36
  chips,
• 1:36
  all
• 1:37
  encrypted.
• 1:38
  Cracking
• 1:38
  the
• 1:38
  Realtek
• 1:39
  turned
• 1:39
  out
• 1:39
  to
• 1:39
  be
• 1:39
  trivially
• 1:40
  easy
• 1:40
  using
• 1:41
  its
• 1:41
  own
• 1:41
  RSIP
• 1:42
  decrypt-on-the-fly
• 1:43
  feature,
• 1:43
  which
• 1:44
  led
• 1:44
  to
• 1:44
  the
• 1:44
  discovery
• 1:45
  of
• 1:45
  a
• 1:45
  TLS
• 1:45
  certificate
• 1:46
  issue
• 1:46
  enabling
• 1:47
  man-in-the-middle
• 1:47
  attacks
• 1:48
  and
• 1:48
  a
• 1:48
  seeding
• 1:49
  bug
• 1:49
  that
• 1:49
  makes
• 1:49
  session
• 1:50
  key
• 1:50
  recovery
• 1:50
  possible.
• 1:51
  The
• 1:51
  Renesas
• 1:52
  firmware
• 1:52
  still
• 1:53
  needs
• 1:53
  decrypting,
• 1:54
  but
• 1:54
  the
• 1:54
  partial
• 1:55
  results
• 1:55
  are
• 1:55
  already
• 1:56
  damning
• 1:56
  enough
• 1:56
  to
• 1:56
  confirm
• 1:57
  that
• 1:57
  the
• 1:57
  "S"
• 1:58
  in
• 1:58
  IoT
• 1:58
  still
• 1:58
  stands
• 1:59
  for
• 1:59
  "security"
• 1:59
  .
• 2:00
  On
• 2:00
  the
• 2:00
  defensive
• 2:01
  side,
• 2:01
  a
• 2:01
  new
• 2:01
  open-source
• 2:02
  tool
• 2:02
  called
• 2:03
  lockbit-rescue
• 2:04
  exploits
• 2:04
  the
• 2:05
  documented
• 2:05
  keystream-reuse
• 2:06
  weakness
• 2:06
  in
• 2:07
  LockBit
• 2:07
  3.0
• 2:08
  ("Black")
• 2:09
  to
• 2:09
  decrypt
• 2:10
  files
• 2:10
  without
• 2:11
  the
• 2:11
  attacker's
• 2:11
  private
• 2:11
  key.
• 2:13
  LockBit's
• 2:13
  file
• 2:13
  encryption
• 2:14
  uses
• 2:14
  a
• 2:14
  modified
• 2:15
  Salsa20
• 2:15
  cipher
• 2:16
  where
• 2:16
  the
• 2:16
  same
• 2:17
  keystream
• 2:17
  gets
• 2:17
  reused
• 2:18
  across
• 2:18
  files
• 2:19
  in
• 2:19
  an
• 2:19
  encryption
• 2:19
  batch.
• 2:21
  The
• 2:21
  tool
• 2:21
  groups
• 2:21
  encrypted
• 2:22
  files
• 2:22
  by
• 2:22
  their
• 2:22
  RSA-encrypted
• 2:25
  KEK
• 2:25
  fingerprint,
• 2:26
  picks
• 2:26
  the
• 2:26
  longest-named
• 2:27
  file
• 2:27
  in
• 2:27
  each
• 2:27
  group
• 2:28
  as
• 2:28
  a
• 2:28
  known-plaintext
• 2:29
  oracle
• 2:29
  (the
• 2:30
  original
• 2:30
  filename
• 2:31
  is
• 2:31
  apLib-compressed
• 2:33
  UTF-16LE
• 2:35
  in
• 2:35
  the
• 2:35
  footer),
• 2:36
  XORs
• 2:36
  it
• 2:36
  against
• 2:37
  ciphertext
• 2:37
  to
• 2:37
  recover
• 2:38
  keystream
• 2:38
  bytes,
• 2:39
  and
• 2:39
  decrypts
• 2:40
  every
• 2:40
  other
• 2:40
  file
• 2:40
  in
• 2:41
  the
• 2:41
  group
• 2:41
  whose
• 2:41
  footer
• 2:41
  fits
• 2:42
  within
• 2:42
  the
• 2:42
  recovered
• 2:42
  coverage.
• 2:43
  Real-world
• 2:44
  recovery
• 2:44
  rates
• 2:45
  run
• 2:45
  5–40%
• 2:47
  depending
• 2:47
  on
• 2:47
  original
• 2:48
  filename
• 2:48
  lengths,
• 2:49
  and
• 2:49
  a
• 2:49
  Phase
• 2:50
  2
• 2:50
  brute-force
• 2:50
  extension
• 2:51
  can
• 2:51
  push
• 2:51
  further
• 2:52
  by
• 2:52
  climbing
• 2:52
  a
• 2:52
  ladder
• 2:53
  of
• 2:53
  intermediate
• 2:53
  files.
• 2:54
  It's
• 2:55
  resumable,
• 2:55
  read-only
• 2:56
  on
• 2:56
  source
• 2:56
  files,
• 2:57
  and
• 2:57
  includes
• 2:58
  a
• 2:58
  libmagic-based
• 2:59
  verification
• 2:59
  sweep
• 3:00
  .
• 3:02
  A
• 3:02
  new
• 3:03
  paper
• 3:03
  on
• 3:03
  the
• 3:03
  Elliptic
• 3:03
  Bitcoin
• 3:04
  Dataset
• 3:04
  delivers
• 3:05
  one
• 3:05
  of
• 3:05
  the
• 3:05
  cleaner
• 3:05
  methodological
• 3:06
  takedowns
• 3:07
  in
• 3:07
  recent
• 3:07
  fraud-detection
• 3:08
  literature.
• 3:09
  The
• 3:09
  consensus
• 3:09
  that
• 3:10
  GCN,
• 3:10
  GraphSAGE,
• 3:11
  GAT,
• 3:11
  and
• 3:11
  EvolveGCN
• 3:12
  beat
• 3:12
  feature-only
• 3:13
  baselines
• 3:13
  on
• 3:14
  Elliptic
• 3:14
  has
• 3:14
  been
• 3:15
  cited
• 3:15
  for
• 3:15
  years
• 3:15
  —
• 3:16
  GCN
• 3:17
  at
• 3:17
  F1=0.70,
• 3:19
  GraphSAGE+SSL
• 3:21
  at
• 3:21
  0.75,
• 3:22
  EvolveGCN
• 3:23
  at
• 3:23
  0.77.
• 3:25
  Those
• 3:25
  numbers
• 3:26
  are
• 3:26
  artifacts
• 3:26
  of
• 3:26
  the
• 3:26
  evaluation
• 3:27
  protocol,
• 3:27
  not
• 3:28
  evidence
• 3:28
  that
• 3:28
  graph
• 3:28
  structure
• 3:29
  helps
• 3:29
  detect
• 3:29
  fraud.
• 3:31
  Every
• 3:31
  prior
• 3:31
  Elliptic
• 3:32
  study
• 3:32
  trains
• 3:32
  with
• 3:32
  transductive
• 3:33
  message
• 3:33
  passing,
• 3:34
  meaning
• 3:34
  the
• 3:34
  full
• 3:34
  graph
• 3:34
  —
• 3:36
  including
• 3:36
  test-period
• 3:36
  nodes
• 3:37
  and
• 3:37
  edges
• 3:37
  —
• 3:38
  is
• 3:38
  visible
• 3:38
  at
• 3:39
  every
• 3:39
  training
• 3:39
  forward
• 3:39
  pass.
• 3:41
  Several
• 3:41
  setups
• 3:41
  labeled
• 3:42
  "inductive"
• 3:43
  still
• 3:43
  leak
• 3:44
  test-period
• 3:44
  statistics
• 3:45
  through
• 3:45
  batch
• 3:45
  normalization
• 3:46
  and
• 3:46
  neighborhood
• 3:46
  aggregation.
• 3:48
  When
• 3:49
  the
• 3:49
  authors
• 3:49
  run
• 3:49
  a
• 3:49
  strictly
• 3:50
  inductive
• 3:50
  protocol
• 3:51
  (encoder
• 3:51
  trained
• 3:52
  only
• 3:52
  on
• 3:52
  the
• 3:52
  time-step
• 3:52
  ≤
• 3:53
  34
• 3:53
  subgraph,
• 3:54
  with
• 3:54
  10
• 3:54
  seeds
• 3:55
  and
• 3:55
  per-timestep
• 3:55
  reporting),
• 3:56
  Random
• 3:57
  Forest
• 3:57
  on
• 3:57
  the
• 3:57
  raw
• 3:57
  165-dimensional
• 3:59
  features
• 3:59
  reaches
• 4:00
  F1=0.821
• 4:02
  and
• 4:02
  beats
• 4:02
  every
• 4:03
  GNN
• 4:03
  tested.
• 4:04
  GraphSAGE,
• 4:05
  the
• 4:05
  strongest
• 4:05
  graph
• 4:06
  encoder
• 4:06
  under
• 4:06
  strict
• 4:07
  inductive
• 4:07
  evaluation,
• 4:08
  manages
• 4:09
  only
• 4:09
  F1=0.689
• 4:10
  ±
• 4:12
  0.017.
• 4:14
  A
• 4:14
  paired
• 4:14
  controlled
• 4:14
  experiment
• 4:15
  quantifies
• 4:15
  the
• 4:15
  leakage
• 4:16
  directly:
• 4:17
  across
• 4:17
  10
• 4:17
  matched
• 4:18
  seeds
• 4:18
  holding
• 4:18
  architecture,
• 4:19
  optimizer,
• 4:20
  loss,
• 4:21
  and
• 4:21
  seed
• 4:21
  constant,
• 4:22
  GraphSAGE
• 4:23
  scores
• 4:23
  F1=0.294
• 4:26
  transductively
• 4:26
  versus
• 4:27
  F1=0.689
• 4:28
  inductively
• 4:29
  —
• 4:31
  a
• 4:31
  39.5-point
• 4:32
  gap
• 4:33
  (Cohen's
• 4:33
  d
• 4:34
  =
• 4:34
  15.8,
• 4:35
  p
• 4:35
  =
• 4:36
  2.6
• 4:37
  ×
• 4:37
  10⁻¹²)
• 4:38
  explained
• 4:39
  entirely
• 4:39
  by
• 4:39
  training-time
• 4:40
  exposure
• 4:40
  to
• 4:40
  test-period
• 4:41
  adjacency.
• 4:42
  The
• 4:42
  edge-shuffle
• 4:43
  ablation
• 4:43
  is
• 4:43
  perhaps
• 4:44
  the
• 4:44
  most
• 4:44
  devastating
• 4:44
  finding.
• 4:46
  Randomly
• 4:46
  shuffled
• 4:46
  edges
• 4:47
  outperform
• 4:47
  the
• 4:47
  real
• 4:48
  transaction
• 4:48
  graph
• 4:48
  by
• 4:48
  8.9
• 4:49
  F1
• 4:49
  points,
• 4:50
  and
• 4:50
  removing
• 4:51
  edges
• 4:51
  entirely
• 4:51
  still
• 4:52
  beats
• 4:52
  real
• 4:52
  edges
• 4:52
  by
• 4:53
  2.5
• 4:53
  points.
• 4:54
  The
• 4:55
  original
• 4:55
  graph's
• 4:55
  mean
• 4:56
  degree
• 4:56
  of
• 4:56
  2.3
• 4:57
  produces
• 4:57
  neighborhoods
• 4:58
  that
• 4:58
  are
• 4:58
  simultaneously
• 4:59
  sparse
• 4:59
  and
• 5:00
  semantically
• 5:00
  misleading
• 5:01
  under
• 5:01
  temporal
• 5:01
  shift
• 5:01
  —
• 5:02
  each
• 5:02
  fraud
• 5:03
  node
• 5:03
  is
• 5:03
  surrounded
• 5:03
  by
• 5:04
  licit
• 5:04
  counterparties,
• 5:05
  so
• 5:05
  message
• 5:05
  passing
• 5:06
  actively
• 5:06
  pulls
• 5:07
  fraud
• 5:07
  representations
• 5:08
  toward
• 5:08
  the
• 5:08
  licit
• 5:08
  manifold.
• 5:09
  The
• 5:10
  paper's
• 5:10
  feature
• 5:10
  attribution
• 5:11
  confirms
• 5:11
  where
• 5:11
  the
• 5:12
  discriminative
• 5:12
  signal
• 5:12
  actually
• 5:13
  lives:
• 5:14
  local
• 5:14
  features
• 5:14
  account
• 5:15
  for
• 5:15
  82.4%
• 5:16
  of
• 5:16
  total
• 5:17
  Random
• 5:17
  Forest
• 5:18
  importance,
• 5:19
  and
• 5:19
  the
• 5:19
  top
• 5:19
  five
• 5:19
  features
• 5:20
  are
• 5:20
  all
• 5:20
  local.
• 5:21
  A
• 5:21
  concatenation
• 5:22
  hybrid
• 5:22
  of
• 5:22
  GraphSAGE
• 5:23
  embeddings
• 5:23
  and
• 5:24
  raw
• 5:24
  features
• 5:24
  drops
• 5:25
  from
• 5:25
  the
• 5:25
  previously
• 5:25
  reported
• 5:26
  F1=0.807
• 5:29
  to
• 5:30
  0.699
• 5:31
  under
• 5:31
  the
• 5:31
  clean
• 5:31
  protocol,
• 5:32
  and
• 5:32
  the
• 5:33
  GNN
• 5:33
  contributes
• 5:34
  only
• 5:34
  +0.018
• 5:35
  F1
• 5:36
  over
• 5:37
  a
• 5:37
  matched-capacity
• 5:38
  MLP
• 5:38
  —
• 5:39
  real,
• 5:39
  reliable,
• 5:40
  and
• 5:40
  practically
• 5:41
  inconsequential.
• 5:43
  For
• 5:43
  anyone
• 5:44
  building
• 5:44
  fraud
• 5:44
  pipelines
• 5:45
  on
• 5:45
  transaction
• 5:45
  graphs:
• 5:46
  test
• 5:47
  under
• 5:47
  strict
• 5:47
  inductive
• 5:48
  protocol
• 5:48
  before
• 5:49
  you
• 5:49
  invest
• 5:49
  in
• 5:49
  graph
• 5:49
  machinery,
• 5:50
  because
• 5:51
  the
• 5:51
  feature
• 5:51
  engineering
• 5:52
  you
• 5:52
  already
• 5:52
  have
• 5:52
  might
• 5:53
  be
• 5:53
  the
• 5:53
  strongest
• 5:53
  signal
• 5:54
  in
• 5:54
  the
• 5:54
  room
• 5:54
  .
• 5:57
  Someone
• 5:58
  went
• 5:58
  spelunking
• 5:58
  through
• 5:59
  Claude
• 5:59
  Code's
• 5:59
  node
• 5:59
  modules
• 6:00
  and
• 6:00
  surfaced
• 6:01
  a
• 6:01
  surprisingly
• 6:01
  deep
• 6:01
  configuration
• 6:02
  surface
• 6:02
  area
• 6:03
  that
• 6:03
  the
• 6:03
  official
• 6:03
  docs
• 6:04
  never
• 6:04
  mention.
• 6:05
  The
• 6:05
  auto-mode
• 6:05
  permission
• 6:06
  system
• 6:06
  is
• 6:06
  internally
• 6:07
  called
• 6:07
  the
• 6:07
  "YOLO
• 6:07
  Classifier"
• 6:08
  (that's
• 6:08
  the
• 6:09
  actual
• 6:09
  variable
• 6:09
  name
• 6:09
  in
• 6:10
  yoloClassifier.ts),
• 6:13
  and
• 6:14
  it
• 6:14
  accepts
• 6:14
  plain
• 6:14
  English
• 6:15
  environment
• 6:15
  descriptions
• 6:15
  —
• 6:16
  strings
• 6:16
  like
• 6:16
  "this
• 6:16
  is
• 6:17
  a
• 6:17
  staging
• 6:17
  server,
• 6:17
  destructive
• 6:18
  operations
• 6:18
  are
• 6:18
  acceptable"
• 6:19
  —
• 6:19
  that
• 6:19
  the
• 6:19
  classifier
• 6:20
  reads
• 6:20
  to
• 6:20
  decide
• 6:21
  what
• 6:21
  gets
• 6:21
  auto-approved.
• 6:22
  PreToolUse
• 6:22
  hooks
• 6:23
  can
• 6:23
  return
• 6:23
  updatedInput
• 6:24
  to
• 6:24
  rewrite
• 6:24
  commands
• 6:25
  mid-flight
• 6:25
  (so
• 6:26
  git
• 6:26
  push
• 6:26
  silently
• 6:26
  becomes
• 6:27
  git
• 6:27
  push
• 6:27
  --dry-run),
• 6:28
  permissionDecision
• 6:29
  to
• 6:29
  force
• 6:29
  allow/deny
• 6:30
  without
• 6:31
  user
• 6:31
  prompts,
• 6:31
  and
• 6:32
  additionalContext
• 6:32
  to
• 6:32
  inject
• 6:33
  text
• 6:33
  into
• 6:33
  the
• 6:33
  conversation.
• 6:35
  Three
• 6:35
  undocumented
• 6:36
  hook
• 6:36
  fields
• 6:36
  change
• 6:36
  behavior
• 6:37
  fundamentally:
• 6:38
  once:
• 6:38
  true
• 6:39
  fires
• 6:39
  a
• 6:39
  hook
• 6:39
  exactly
• 6:40
  once
• 6:40
  then
• 6:40
  auto-removes
• 6:40
  it,
• 6:42
  async:
• 6:42
  true
• 6:42
  runs
• 6:43
  hooks
• 6:43
  in
• 6:43
  the
• 6:43
  background
• 6:44
  without
• 6:44
  blocking,
• 6:44
  and
• 6:44
  asyncRewake:
• 6:45
  true
• 6:45
  runs
• 6:46
  non-blocking
• 6:46
  but
• 6:47
  wakes
• 6:47
  the
• 6:47
  model
• 6:47
  and
• 6:47
  blocks
• 6:48
  if
• 6:48
  the
• 6:48
  hook
• 6:48
  exits
• 6:48
  with
• 6:48
  code
• 6:49
  2
• 6:49
  —
• 6:49
  non-blocking
• 6:50
  on
• 6:50
  the
• 6:50
  happy
• 6:50
  path,
• 6:51
  blocking
• 6:51
  when
• 6:51
  something's
• 6:52
  wrong.
• 6:52
  Custom
• 6:53
  agents
• 6:53
  support
• 6:53
  persistent
• 6:54
  memory
• 6:54
  across
• 6:54
  invocations
• 6:55
  (user,
• 6:56
  project,
• 6:56
  or
• 6:56
  local
• 6:57
  scopes),
• 6:57
  omitClaudeMd:
• 6:58
  true
• 6:59
  for
• 6:59
  "fresh
• 6:59
  eyes"
• 6:59
  reviews,
• 7:00
  and
• 7:00
  a
• 7:00
  criticalSystemReminder
• 7:01
  EXPERIMENTAL
• 7:02
  field
• 7:02
  that
• 7:02
  gets
• 7:02
  re-injected
• 7:03
  at
• 7:03
  every
• 7:03
  turn
• 7:04
  even
• 7:04
  after
• 7:04
  context
• 7:05
  compaction.
• 7:06
  Two
• 7:06
  undocumented
• 7:07
  settings.json
• 7:07
  fields
• 7:08
  —
• 7:09
  autoMemoryEnabled
• 7:10
  and
• 7:10
  autoDreamEnabled
• 7:11
  —
• 7:12
  activate
• 7:12
  a
• 7:12
  compound
• 7:13
  learning
• 7:13
  loop:
• 7:13
  sessions
• 7:14
  produce
• 7:14
  memories,
• 7:15
  dreams
• 7:15
  consolidate
• 7:15
  them
• 7:16
  every
• 7:16
  24
• 7:16
  hours,
• 7:17
  and
• 7:17
  consolidated
• 7:18
  memories
• 7:18
  inform
• 7:18
  future
• 7:19
  sessions
• 7:19
  .
• 7:20
  The
• 7:20
  configuration
• 7:21
  deep-dive
• 7:21
  arrives
• 7:21
  at
• 7:21
  the
• 7:22
  exact
• 7:22
  moment
• 7:22
  organizations
• 7:23
  are
• 7:23
  discovering
• 7:23
  what
• 7:24
  unrestricted
• 7:24
  AI
• 7:24
  coding
• 7:25
  tool
• 7:25
  access
• 7:25
  actually
• 7:26
  costs.
• 7:27
  Microsoft
• 7:27
  reportedly
• 7:28
  pulled
• 7:28
  internal
• 7:28
  Claude
• 7:29
  Code
• 7:29
  licenses
• 7:29
  after
• 7:30
  the
• 7:30
  token
• 7:30
  bill
• 7:30
  ran
• 7:30
  past
• 7:31
  anything
• 7:31
  anyone
• 7:31
  had
• 7:31
  budgeted
• 7:32
  for
• 7:32
  —
• 7:33
  the
• 7:33
  tool
• 7:33
  didn't
• 7:33
  fail,
• 7:34
  it
• 7:34
  worked
• 7:34
  so
• 7:34
  well
• 7:34
  that
• 7:34
  engineers
• 7:35
  leaned
• 7:35
  on
• 7:35
  it
• 7:35
  constantly.
• 7:37
  The
• 7:37
  post
• 7:37
  promoting
• 7:37
  this
• 7:38
  disclosure
• 7:38
  also
• 7:38
  introduces
• 7:39
  SpectralQuant,
• 7:40
  a
• 7:40
  KV-cache
• 7:41
  compression
• 7:41
  library
• 7:42
  claiming
• 7:42
  6x
• 7:42
  compression
• 7:43
  on
• 7:43
  production
• 7:43
  H200s
• 7:45
  where
• 7:45
  the
• 7:45
  cache
• 7:45
  eats
• 7:45
  roughly
• 7:46
  64%
• 7:47
  of
• 7:47
  GPU
• 7:47
  memory.
• 7:48
  On
• 7:48
  Mistral
• 7:49
  7B
• 7:49
  at
• 7:49
  5.95x
• 7:50
  compression,
• 7:51
  it
• 7:51
  reports
• 7:52
  40.2
• 7:53
  tokens/sec
• 7:54
  versus
• 7:54
  18.9
• 7:55
  for
• 7:55
  the
• 7:55
  uncompressed
• 7:55
  baseline
• 7:56
  —
• 7:57
  same
• 7:57
  model,
• 7:57
  same
• 7:58
  accuracy,
• 7:58
  roughly
• 7:59
  double
• 7:59
  the
• 7:59
  speed
• 7:59
  .
• 8:00
  Meanwhile,
• 8:01
  the
• 8:01
  local-AI
• 8:02
  community
• 8:02
  is
• 8:02
  pushing
• 8:03
  in
• 8:03
  the
• 8:03
  opposite
• 8:03
  direction
• 8:04
  entirely:
• 8:05
  building
• 8:05
  agents
• 8:06
  that
• 8:06
  optimize
• 8:06
  their
• 8:06
  own
• 8:06
  prompts
• 8:07
  and
• 8:07
  tool
• 8:07
  usage
• 8:08
  over
• 8:08
  time
• 8:08
  using
• 8:09
  feedback
• 8:09
  loops,
• 8:10
  turning
• 8:10
  the
• 8:10
  cost
• 8:10
  problem
• 8:11
  into
• 8:11
  an
• 8:11
  engineering
• 8:11
  problem
• 8:12
  that
• 8:12
  runs
• 8:13
  on
• 8:13
  hardware
• 8:13
  you
• 8:13
  already
• 8:14
  own
• 8:14
  .
• 8:16
  A
• 8:16
  blog
• 8:17
  series
• 8:17
  culminating
• 8:18
  in
• 8:18
  "A
• 8:18
  10
• 8:18
  year
• 8:18
  old
• 8:18
  Xeon
• 8:19
  is
• 8:19
  all
• 8:19
  you
• 8:19
  need"
• 8:20
  lays
• 8:20
  out
• 8:20
  exactly
• 8:21
  what
• 8:21
  it
• 8:21
  takes
• 8:21
  to
• 8:21
  run
• 8:21
  Gemma
• 8:22
  4
• 8:22
  26B
• 8:23
  on
• 8:23
  a
• 8:23
  single
• 8:23
  Intel
• 8:24
  Xeon
• 8:24
  E5-2620
• 8:25
  v4
• 8:26
  from
• 8:27
  2016
• 8:27
  —
• 8:28
  8
• 8:28
  cores,
• 8:29
  DDR3,
• 8:29
  no
• 8:30
  GPU.
• 8:31
  The
• 8:31
  author's
• 8:31
  magic
• 8:32
  spell
• 8:32
  requires
• 8:33
  25
• 8:33
  flags
• 8:34
  across
• 8:34
  speculative
• 8:34
  decoding
• 8:35
  (--spec-type
• 8:36
  mtp
• 8:36
  --draft-max
• 8:37
  3
• 8:38
  --spec-autotune),
• 8:39
  CPU-aware
• 8:39
  MoE
• 8:40
  routing
• 8:41
  (--cpu-moe
• 8:42
  --merge-up-gate-experts),
• 8:43
  memory
• 8:43
  pinning
• 8:45
  (--mlock
• 8:46
  --run-time-repack),
• 8:47
  and
• 8:47
  Flash
• 8:48
  Attention
• 8:48
  ported
• 8:48
  to
• 8:49
  CPU
• 8:49
  (--flash-attn
• 8:50
  on
• 8:50
  --mla-use
• 8:51
  3).
• 8:52
  The
• 8:53
  argument
• 8:53
  for
• 8:53
  speculative
• 8:54
  decoding
• 8:54
  is
• 8:54
  stronger
• 8:55
  on
• 8:55
  CPU
• 8:55
  than
• 8:55
  GPU:
• 8:57
  CPU
• 8:57
  compute
• 8:58
  is
• 8:58
  cheap
• 8:58
  relative
• 8:58
  to
• 8:59
  streaming
• 8:59
  verifier
• 8:59
  weights
• 9:00
  through
• 9:00
  cache,
• 9:01
  so
• 9:01
  the
• 9:01
  drafter's
• 9:02
  working
• 9:02
  set
• 9:02
  fits
• 9:02
  in
• 9:02
  L3
• 9:03
  while
• 9:03
  the
• 9:03
  verifier
• 9:04
  spills
• 9:04
  out
• 9:04
  of
• 9:04
  everything.
• 9:06
  Runtime
• 9:06
  repacking
• 9:06
  reorganizes
• 9:07
  weight
• 9:07
  matrices
• 9:08
  to
• 9:08
  align
• 9:08
  with
• 9:08
  CPU
• 9:09
  cache
• 9:09
  layout,
• 9:10
  paying
• 9:10
  a
• 9:10
  small
• 9:10
  startup
• 9:11
  penalty
• 9:11
  for
• 9:11
  maximum
• 9:11
  bandwidth
• 9:12
  during
• 9:12
  generation.
• 9:14
  The
• 9:14
  result:
• 9:15
  an
• 9:15
  82
• 9:15
  GB
• 9:16
  footprint
• 9:16
  in
• 9:16
  DDR3
• 9:17
  generating
• 9:18
  text
• 9:18
  at
• 9:18
  reading
• 9:19
  speed
• 9:19
  on
• 9:19
  hardware
• 9:20
  that
• 9:20
  was
• 9:20
  old
• 9:20
  when
• 9:20
  the
• 9:21
  architecture
• 9:21
  hadn't
• 9:21
  been
• 9:21
  invented
• 9:22
  yet
• 9:22
  .
• 9:23
  In
• 9:23
  the
• 9:23
  same
• 9:23
  vein,
• 9:24
  a
• 9:24
  community
• 9:25
  member
• 9:25
  reports
• 9:25
  10.33
• 9:27
  tokens/sec
• 9:28
  on
• 9:28
  Qwen
• 9:28
  3.5
• 9:29
  35B
• 9:29
  running
• 9:30
  on
• 9:30
  a
• 9:30
  $300
• 9:31
  laptop
• 9:31
  .
• 9:32
  On
• 9:32
  the
• 9:32
  research
• 9:33
  side,
• 9:33
  Huawei's
• 9:34
  OSC
• 9:34
  framework
• 9:35
  tackles
• 9:35
  W4A4
• 9:36
  quantization
• 9:37
  by
• 9:37
  systematically
• 9:38
  characterizing
• 9:38
  where
• 9:39
  activation
• 9:39
  outliers
• 9:40
  actually
• 9:40
  cluster.
• 9:41
  Their
• 9:42
  key
• 9:42
  insight:
• 9:43
  outliers
• 9:43
  aren't
• 9:43
  uniformly
• 9:44
  distributed
• 9:44
  —
• 9:45
  they
• 9:45
  exhibit
• 9:46
  "token-persistent
• 9:47
  structural
• 9:47
  clustering"
• 9:48
  where
• 9:48
  high-magnitude
• 9:49
  values
• 9:49
  consistently
• 9:50
  occupy
• 9:50
  fixed
• 9:51
  channels
• 9:51
  across
• 9:51
  tokens,
• 9:52
  with
• 9:52
  clustering
• 9:53
  density
• 9:53
  exceeding
• 9:54
  60%
• 9:54
  in
• 9:54
  attention
• 9:55
  and
• 9:55
  FFN
• 9:56
  up-projection
• 9:56
  inputs.
• 9:57
  OSC
• 9:58
  exploits
• 9:58
  this
• 9:59
  by
• 9:59
  building
• 9:59
  an
• 9:59
  offline
• 10:00
  lookup
• 10:00
  table
• 10:00
  that
• 10:01
  identifies
• 10:01
  the
• 10:01
  single
• 10:02
  most
• 10:02
  statistically
• 10:03
  significant
• 10:03
  outlier
• 10:03
  channel
• 10:04
  per
• 10:04
  quantization
• 10:05
  group,
• 10:05
  extracting
• 10:06
  it
• 10:06
  to
• 10:06
  a
• 10:06
  high-precision
• 10:07
  buffer
• 10:07
  before
• 10:08
  calculating
• 10:08
  scaling
• 10:08
  factors.
• 10:10
  The
• 10:10
  result
• 10:10
  is
• 10:10
  a
• 10:10
  dual-path
• 10:11
  computation:
• 10:12
  a
• 10:13
  low-precision
• 10:13
  4-bit
• 10:14
  GEMM
• 10:14
  path
• 10:15
  and
• 10:15
  a
• 10:15
  compact
• 10:16
  high-precision
• 10:16
  16-bit
• 10:17
  branch
• 10:17
  GEMM.
• 10:19
  For
• 10:19
  W2
• 10:19
  down-projection
• 10:20
  inputs
• 10:21
  where
• 10:21
  outliers
• 10:21
  are
• 10:21
  more
• 10:21
  diffused
• 10:23
  (20–35%
• 10:24
  clustering),
• 10:25
  OSC
• 10:26
  falls
• 10:26
  back
• 10:26
  to
• 10:26
  FP8.
• 10:28
  On
• 10:28
  Qwen3-8B
• 10:30
  and
• 10:30
  Qwen3-30B,
• 10:33
  OSC
• 10:34
  restricts
• 10:34
  the
• 10:34
  average
• 10:35
  accuracy
• 10:35
  drop
• 10:36
  to
• 10:36
  2.19
• 10:37
  and
• 10:37
  1.12
• 10:38
  points
• 10:38
  respectively
• 10:39
  while
• 10:39
  achieving
• 10:40
  1.78x
• 10:41
  speedup
• 10:41
  over
• 10:42
  the
• 10:42
  W8A8
• 10:42
  baseline
• 10:43
  .
• 10:44
  NVIDIA's
• 10:44
  Lightning
• 10:45
  OPD
• 10:45
  addresses
• 10:46
  a
• 10:46
  different
• 10:46
  efficiency
• 10:47
  bottleneck:
• 10:48
  the
• 10:48
  live
• 10:48
  teacher
• 10:49
  server
• 10:49
  required
• 10:49
  during
• 10:50
  on-policy
• 10:50
  distillation.
• 10:52
  Standard
• 10:52
  OPD
• 10:53
  trains
• 10:53
  a
• 10:53
  student
• 10:53
  model
• 10:54
  to
• 10:54
  match
• 10:54
  a
• 10:54
  stronger
• 10:55
  teacher's
• 10:55
  token-level
• 10:56
  distribution,
• 10:57
  but
• 10:57
  keeping
• 10:57
  the
• 10:57
  teacher
• 10:58
  running
• 10:58
  throughout
• 10:58
  training
• 10:59
  is
• 10:59
  expensive.
• 11:00
  Lightning
• 11:01
  OPD
• 11:01
  identifies
• 11:02
  a
• 11:02
  previously
• 11:03
  overlooked
• 11:03
  condition
• 11:04
  called
• 11:04
  "teacher
• 11:04
  consistency"
• 11:05
  —
• 11:06
  the
• 11:06
  SFT-stage
• 11:08
  and
• 11:08
  OPD-stage
• 11:09
  teachers
• 11:10
  must
• 11:10
  be
• 11:10
  the
• 11:10
  same
• 11:10
  model.
• 11:11
  Violating
• 11:12
  this
• 11:12
  introduces
• 11:13
  irreducible
• 11:14
  gradient
• 11:14
  bias
• 11:15
  causing
• 11:15
  convergence
• 11:16
  to
• 11:16
  a
• 11:16
  suboptimal
• 11:17
  fixed
• 11:17
  point
• 11:17
  regardless
• 11:18
  of
• 11:18
  training
• 11:18
  duration.
• 11:20
  With
• 11:20
  teacher
• 11:20
  consistency
• 11:21
  enforced,
• 11:22
  teacher
• 11:22
  log-probabilities
• 11:23
  can
• 11:23
  be
• 11:23
  precomputed
• 11:24
  once
• 11:24
  over
• 11:24
  SFT
• 11:25
  rollouts
• 11:25
  and
• 11:26
  reused,
• 11:27
  eliminating
• 11:27
  the
• 11:27
  live
• 11:28
  server
• 11:28
  entirely.
• 11:29
  Starting
• 11:30
  from
• 11:30
  Qwen3-8B-Base,
• 11:33
  Lightning
• 11:33
  OPD
• 11:34
  reaches
• 11:34
  69.9%
• 11:36
  on
• 11:36
  AIME
• 11:36
  2024
• 11:37
  in
• 11:38
  just
• 11:38
  30
• 11:38
  GPU
• 11:39
  hours
• 11:39
  —
• 11:40
  a
• 11:40
  4.0x
• 11:41
  speedup
• 11:41
  over
• 11:41
  standard
• 11:42
  OPD
• 11:42
  .
• 11:43
  Separately,
• 11:43
  the
• 11:44
  quantization
• 11:44
  precision
• 11:45
  debate
• 11:45
  got
• 11:45
  concrete
• 11:46
  data:
• 11:46
  Qwen3.6
• 11:49
  shows
• 11:49
  a
• 11:49
  measurable
• 11:50
  quality
• 11:50
  jump
• 11:50
  from
• 11:51
  Q4
• 11:51
  to
• 11:51
  Q6
• 11:52
  quantization
• 11:52
  specifically
• 11:53
  for
• 11:53
  coding
• 11:53
  agent
• 11:54
  workloads,
• 11:55
  suggesting
• 11:55
  that
• 11:55
  the
• 11:55
  "good
• 11:56
  enough"
• 11:56
  threshold
• 11:57
  for
• 11:57
  agentic
• 11:57
  use
• 11:57
  cases
• 11:58
  sits
• 11:58
  higher
• 11:58
  than
• 11:59
  for
• 11:59
  conversational
• 11:59
  ones
• 12:00
  .
• 12:02
  JetBrains
• 12:02
  released
• 12:03
  Mellum2,
• 12:03
  a
• 12:04
  12B-parameter
• 12:05
  Mixture-of-Experts
• 12:06
  model
• 12:06
  that
• 12:06
  activates
• 12:07
  only
• 12:07
  2.5B
• 12:08
  parameters
• 12:08
  per
• 12:08
  token,
• 12:09
  under
• 12:09
  Apache
• 12:09
  2.0.
• 12:10
  It's
• 12:11
  the
• 12:11
  first
• 12:11
  model
• 12:11
  from
• 12:11
  an
• 12:11
  IDE
• 12:12
  vendor
• 12:12
  designed
• 12:12
  explicitly
• 12:13
  as
• 12:13
  infrastructure
• 12:14
  for
• 12:14
  AI-powered
• 12:14
  development
• 12:15
  tools
• 12:15
  rather
• 12:15
  than
• 12:15
  as
• 12:15
  a
• 12:15
  standalone
• 12:16
  assistant.
• 12:17
  JetBrains
• 12:17
  positions
• 12:18
  it
• 12:18
  as
• 12:18
  a
• 12:18
  "focal
• 12:18
  model"
• 12:18
  —
• 12:19
  fast
• 12:19
  and
• 12:19
  well-scoped
• 12:20
  for
• 12:20
  high-frequency
• 12:20
  tasks
• 12:21
  inside
• 12:21
  larger
• 12:21
  AI
• 12:22
  systems:
• 12:22
  routing,
• 12:23
  RAG
• 12:23
  post-processing,
• 12:24
  tool
• 12:24
  selection,
• 12:25
  context
• 12:25
  compression,
• 12:26
  and
• 12:26
  agent
• 12:27
  subtasks
• 12:27
  like
• 12:27
  planning
• 12:28
  and
• 12:28
  validation.
• 12:29
  Benchmarks
• 12:30
  show
• 12:30
  competitive
• 12:30
  performance
• 12:31
  against
• 12:31
  similar-sized
• 12:32
  models
• 12:32
  with
• 12:32
  over
• 12:32
  2x
• 12:33
  faster
• 12:33
  inference,
• 12:34
  targeting
• 12:35
  the
• 12:35
  latency-sensitive
• 12:35
  slots
• 12:36
  in
• 12:36
  multi-model
• 12:36
  architectures
• 12:37
  where
• 12:37
  calling
• 12:37
  a
• 12:37
  frontier
• 12:38
  model
• 12:38
  for
• 12:38
  every
• 12:39
  intermediate
• 12:39
  step
• 12:39
  is
• 12:39
  wasteful
• 12:39
  .
• 12:40
  Tencent's
• 12:40
  Hy-MT2
• 12:41
  also
• 12:42
  moved
• 12:42
  to
• 12:42
  Apache
• 12:42
  License
• 12:43
  2.0,
• 12:44
  continuing
• 12:44
  the
• 12:44
  trend
• 12:45
  of
• 12:45
  major
• 12:45
  labs
• 12:45
  opening
• 12:46
  their
• 12:46
  model
• 12:46
  weights
• 12:46
  under
• 12:46
  permissive
• 12:47
  terms
• 12:47
  .
• 12:47
  The
• 12:47
  tooling
• 12:48
  layer
• 12:48
  beneath
• 12:48
  models
• 12:49
  continues
• 12:49
  filling
• 12:49
  in.
• 12:50
  LogicPipe
• 12:51
  tackles
• 12:51
  multi-device
• 12:52
  LLM
• 12:52
  inference
• 12:53
  by
• 12:53
  splitting
• 12:53
  transformer
• 12:54
  layers
• 12:54
  across
• 12:54
  ranks,
• 12:55
  using
• 12:55
  DAG
• 12:56
  scheduling
• 12:56
  with
• 12:56
  dependency-aware
• 12:57
  task
• 12:57
  dispatch
• 12:58
  and
• 12:58
  contextual
• 12:58
  KV-cache
• 12:59
  reuse
• 12:59
  between
• 13:00
  points
• 13:00
  —
• 13:01
  a
• 13:01
  structured
• 13:01
  approach
• 13:01
  to
• 13:02
  pipeline
• 13:02
  parallelism
• 13:03
  on
• 13:03
  heterogeneous
• 13:03
  edge
• 13:04
  hardware
• 13:04
  .
• 13:04
  dlmserve
• 13:05
  launches
• 13:05
  as
• 13:05
  the
• 13:05
  first
• 13:06
  dedicated
• 13:06
  serving
• 13:07
  engine
• 13:07
  for
• 13:07
  diffusion
• 13:07
  language
• 13:08
  models,
• 13:08
  a
• 13:09
  class
• 13:09
  of
• 13:09
  models
• 13:09
  that
• 13:09
  has
• 13:10
  lacked
• 13:10
  the
• 13:10
  vLLM/SGLang-equivalent
• 13:14
  infrastructure
• 13:15
  that
• 13:15
  autoregressive
• 13:15
  models
• 13:16
  enjoy
• 13:16
  .
• 13:17
  Genspark
• 13:18
  AI
• 13:18
  ships
• 13:18
  an
• 13:18
  open-source
• 13:19
  "Super
• 13:19
  Agent"
• 13:20
  framework
• 13:20
  with
• 13:20
  80+
• 13:21
  built-in
• 13:21
  tools,
• 13:22
  multi-LLM
• 13:23
  routing,
• 13:24
  and
• 13:24
  a
• 13:24
  planner-executor
• 13:25
  architecture
• 13:25
  that
• 13:26
  decomposes
• 13:26
  goals
• 13:27
  into
• 13:27
  DAGs
• 13:27
  of
• 13:27
  subtasks
• 13:28
  dispatched
• 13:28
  to
• 13:28
  specialist
• 13:29
  agents
• 13:29
  —
• 13:30
  positioned
• 13:31
  as
• 13:31
  a
• 13:31
  self-hosted
• 13:31
  alternative
• 13:32
  to
• 13:32
  the
• 13:32
  commercial
• 13:33
  Genspark.ai
• 13:33
  service
• 13:34
  .
• 13:35
  And
• 13:35
  Qwen
• 13:35
  released
• 13:35
  Qwen-Image-Bench,
• 13:38
  a
• 13:38
  new
• 13:38
  benchmark
• 13:38
  for
• 13:38
  evaluating
• 13:39
  visual
• 13:39
  language
• 13:40
  models
• 13:40
  .
• 13:42
  NVIDIA's
• 13:43
  Cosmos
• 13:43
  3
• 13:43
  represents
• 13:44
  a
• 13:44
  significant
• 13:45
  architectural
• 13:45
  leap
• 13:46
  from
• 13:46
  its
• 13:46
  predecessors.
• 13:47
  Where
• 13:47
  earlier
• 13:48
  Cosmos
• 13:48
  releases
• 13:49
  required
• 13:49
  separate
• 13:49
  models
• 13:50
  for
• 13:50
  world
• 13:50
  generation
• 13:51
  (Predict),
• 13:51
  controlled
• 13:51
  generation
• 13:52
  (Transfer),
• 13:53
  scene
• 13:53
  understanding
• 13:54
  (Reason),
• 13:55
  and
• 13:55
  policy
• 13:55
  generation
• 13:56
  (Policy),
• 13:57
  Cosmos
• 13:57
  3
• 13:58
  unifies
• 13:58
  everything
• 13:59
  into
• 13:59
  a
• 13:59
  single
• 13:59
  Mixture-of-Transformers
• 14:00
  architecture.
• 14:01
  The
• 14:01
  model
• 14:02
  processes
• 14:02
  text,
• 14:03
  image,
• 14:03
  video,
• 14:04
  audio,
• 14:05
  and
• 14:05
  action
• 14:05
  modalities
• 14:06
  in
• 14:06
  one
• 14:06
  forward
• 14:06
  pass,
• 14:07
  splitting
• 14:07
  the
• 14:07
  input
• 14:08
  sequence
• 14:08
  into
• 14:08
  an
• 14:08
  autoregressive
• 14:09
  subsequence
• 14:09
  for
• 14:09
  reasoning
• 14:10
  and
• 14:10
  a
• 14:10
  diffusion
• 14:11
  subsequence
• 14:11
  for
• 14:11
  generation.
• 14:12
  AR
• 14:13
  and
• 14:13
  DM
• 14:13
  tokens
• 14:13
  use
• 14:13
  separate
• 14:14
  parameter
• 14:14
  sets
• 14:14
  per
• 14:15
  layer
• 14:15
  but
• 14:15
  interact
• 14:16
  through
• 14:16
  joint
• 14:16
  attention,
• 14:17
  letting
• 14:17
  the
• 14:17
  same
• 14:17
  model
• 14:18
  serve
• 14:18
  as
• 14:18
  a
• 14:18
  VLM,
• 14:19
  video
• 14:19
  generator,
• 14:20
  forward/inverse
• 14:21
  dynamics
• 14:21
  model,
• 14:21
  or
• 14:22
  robot
• 14:22
  policy
• 14:22
  without
• 14:22
  architectural
• 14:23
  changes.
• 14:24
  Two
• 14:24
  sizes
• 14:25
  ship:
• 14:25
  Cosmos
• 14:26
  3
• 14:26
  Nano
• 14:27
  (16B,
• 14:27
  workstation-grade)
• 14:28
  and
• 14:28
  Cosmos
• 14:29
  3
• 14:29
  Super
• 14:30
  (64B,
• 14:30
  datacenter-scale),
• 14:31
  both
• 14:32
  with
• 14:32
  Diffusers
• 14:32
  integration
• 14:33
  and
• 14:33
  post-training
• 14:33
  scripts
• 14:33
  for
• 14:33
  domain
• 14:34
  adaptation.
• 14:35
  NVIDIA
• 14:35
  also
• 14:36
  released
• 14:36
  accompanying
• 14:36
  synthetic
• 14:37
  data
• 14:37
  generation
• 14:38
  datasets
• 14:38
  covering
• 14:38
  robotics,
• 14:39
  physics
• 14:40
  simulation,
• 14:40
  autonomous
• 14:41
  driving,
• 14:42
  warehouse
• 14:42
  operations,
• 14:43
  and
• 14:43
  human
• 14:43
  motion
• 14:43
  —
• 14:44
  the
• 14:44
  kind
• 14:45
  of
• 14:45
  training
• 14:45
  data
• 14:45
  that
• 14:45
  physical
• 14:46
  AI
• 14:46
  has
• 14:46
  chronically
• 14:47
  lacked
• 14:47
  .
• 14:48
  At
• 14:48
  Computex
• 14:49
  in
• 14:49
  Taipei,
• 14:49
  Jensen
• 14:50
  Huang
• 14:50
  unveiled
• 14:51
  RTX
• 14:51
  Spark,
• 14:52
  a
• 14:52
  chip
• 14:52
  designed
• 14:52
  for
• 14:52
  what
• 14:53
  NVIDIA
• 14:53
  calls
• 14:53
  "the
• 14:53
  era
• 14:54
  of
• 14:54
  personal
• 14:54
  AI
• 14:54
  agents."
• 14:55
  Lenovo,
• 14:56
  HP,
• 14:57
  Dell,
• 14:57
  Microsoft
• 14:58
  Surface,
• 14:59
  Asus,
• 14:59
  and
• 15:00
  MSI
• 15:00
  will
• 15:00
  ship
• 15:01
  Windows
• 15:01
  PCs
• 15:02
  with
• 15:02
  RTX
• 15:02
  Spark
• 15:02
  beginning
• 15:03
  this
• 15:03
  autumn,
• 15:04
  with
• 15:04
  Acer
• 15:04
  and
• 15:04
  Gigabyte
• 15:05
  following.
• 15:06
  Forrester's
• 15:06
  Charlie
• 15:07
  Dai
• 15:07
  called
• 15:07
  it
• 15:07
  a
• 15:07
  "paradigm
• 15:08
  shift"
• 15:08
  from
• 15:09
  "component
• 15:09
  supplier"
• 15:10
  to
• 15:10
  "architecture
• 15:10
  owner
• 15:11
  in
• 15:11
  the
• 15:11
  PC
• 15:11
  market,"
• 15:12
  putting
• 15:12
  NVIDIA
• 15:12
  in
• 15:13
  direct
• 15:13
  competition
• 15:13
  with
• 15:13
  Intel,
• 15:14
  AMD,
• 15:14
  and
• 15:15
  Qualcomm.
• 15:16
  Analyst
• 15:17
  Ian
• 15:17
  Fogg
• 15:17
  tempered
• 15:17
  expectations,
• 15:19
  noting
• 15:19
  the
• 15:19
  machines
• 15:20
  will
• 15:20
  likely
• 15:20
  carry
• 15:20
  a
• 15:20
  significant
• 15:21
  price
• 15:21
  tag
• 15:21
  and
• 15:22
  target
• 15:22
  workstation-class
• 15:23
  buyers
• 15:23
  .
• 15:24
  NVIDIA
• 15:24
  also
• 15:25
  released
• 15:25
  LocateAnything,
• 15:26
  a
• 15:26
  vision-language
• 15:27
  grounding
• 15:27
  model
• 15:27
  using
• 15:28
  parallel
• 15:28
  box
• 15:29
  decoding
• 15:29
  that
• 15:29
  claims
• 15:30
  10x
• 15:30
  faster
• 15:30
  inference
• 15:31
  than
• 15:31
  Qwen3-VL
• 15:33
  for
• 15:33
  object
• 15:34
  detection
• 15:34
  and
• 15:34
  spatial
• 15:35
  understanding
• 15:35
  tasks
• 15:35
  .
• 15:38
  Higgsfield,
• 15:38
  a
• 15:38
  San
• 15:39
  Francisco
• 15:39
  startup
• 15:39
  valued
• 15:40
  at
• 15:40
  $1.3
• 15:41
  billion,
• 15:42
  announced
• 15:42
  it
• 15:42
  had
• 15:42
  premiered
• 15:43
  a
• 15:43
  fully
• 15:43
  AI-generated
• 15:44
  feature
• 15:45
  film
• 15:45
  at
• 15:45
  Cannes.
• 15:46
  The
• 15:46
  Wall
• 15:46
  Street
• 15:47
  Journal
• 15:47
  covered
• 15:47
  it.
• 15:48
  The
• 15:48
  founder
• 15:48
  posted
• 15:49
  on
• 15:49
  LinkedIn
• 15:49
  about
• 15:49
  how
• 15:50
  "for
• 15:50
  decades,
• 15:50
  Cannes
• 15:51
  has
• 15:51
  been
• 15:51
  the
• 15:51
  room
• 15:51
  where
• 15:51
  new
• 15:52
  cinema
• 15:52
  gets
• 15:52
  legitimized."
• 15:53
  Then
• 15:53
  Cannes
• 15:54
  said
• 15:54
  it
• 15:54
  never
• 15:54
  happened.
• 15:55
  A
• 15:55
  festival
• 15:56
  spokesperson
• 15:56
  confirmed
• 15:57
  to
• 15:57
  Futurism
• 15:58
  that
• 15:58
  "Hell
• 15:58
  Grind
• 15:59
  was
• 15:59
  not
• 15:59
  screened
• 16:00
  as
• 16:00
  part
• 16:00
  of
• 16:00
  the
• 16:00
  official
• 16:00
  Festival
• 16:01
  de
• 16:01
  Cannes
• 16:01
  program."
• 16:02
  The
• 16:03
  film
• 16:03
  screened
• 16:03
  at
• 16:03
  the
• 16:03
  Marche
• 16:04
  du
• 16:04
  Film,
• 16:05
  a
• 16:05
  separate
• 16:05
  commercial
• 16:06
  marketplace
• 16:06
  that
• 16:06
  will
• 16:06
  screen
• 16:07
  anything
• 16:07
  that
• 16:07
  pays
• 16:07
  the
• 16:08
  fee
• 16:08
  —
• 16:08
  it
• 16:08
  has
• 16:08
  screened
• 16:09
  Sharknado.
• 16:10
  Calling
• 16:10
  it
• 16:10
  a
• 16:10
  Cannes
• 16:11
  premiere
• 16:11
  is
• 16:11
  roughly
• 16:11
  equivalent
• 16:12
  to
• 16:12
  buying
• 16:12
  an
• 16:12
  ad
• 16:13
  in
• 16:13
  the
• 16:13
  New
• 16:13
  York
• 16:13
  Times
• 16:13
  and
• 16:13
  describing
• 16:14
  yourself
• 16:14
  as
• 16:14
  a
• 16:14
  Times
• 16:15
  journalist.
• 16:16
  The
• 16:16
  underlying
• 16:16
  work
• 16:16
  was
• 16:16
  real.
• 16:17
  Higgsfield
• 16:17
  made
• 16:18
  a
• 16:18
  95-minute
• 16:18
  action
• 16:19
  film
• 16:19
  in
• 16:19
  two
• 16:19
  weeks
• 16:19
  using
• 16:19
  AI
• 16:20
  video
• 16:20
  generation
• 16:20
  tools
• 16:21
  including
• 16:21
  Google's
• 16:21
  Veo
• 16:22
  3,
• 16:22
  at
• 16:23
  a
• 16:23
  total
• 16:23
  cost
• 16:23
  of
• 16:23
  $500,000
• 16:24
  ($400,000
• 16:25
  in
• 16:25
  compute).
• 16:26
  Each
• 16:26
  prompt
• 16:27
  averaged
• 16:27
  3,000
• 16:28
  words.
• 16:29
  The
• 16:29
  first
• 16:29
  25
• 16:30
  minutes
• 16:30
  required
• 16:30
  16,181
• 16:32
  initial
• 16:32
  video
• 16:33
  generations
• 16:33
  to
• 16:33
  produce
• 16:34
  253
• 16:35
  final
• 16:35
  shots.
• 16:36
  Maintaining
• 16:36
  visual
• 16:37
  consistency
• 16:37
  across
• 16:38
  feature
• 16:38
  length
• 16:38
  required
• 16:39
  detailed
• 16:39
  style
• 16:39
  prefixes
• 16:40
  defining
• 16:40
  lighting,
• 16:41
  camera
• 16:41
  type,
• 16:42
  and
• 16:42
  physics
• 16:42
  behavior
• 16:43
  for
• 16:43
  every
• 16:43
  prompt.
• 16:44
  "You
• 16:44
  can't
• 16:44
  go
• 16:45
  into
• 16:45
  AI
• 16:45
  and
• 16:45
  say
• 16:45
  make
• 16:45
  me
• 16:46
  a
• 16:46
  95-minute
• 16:46
  cool
• 16:47
  video,"
• 16:47
  one
• 16:47
  content
• 16:48
  lead
• 16:48
  acknowledged.
• 16:49
  But
• 16:49
  director
• 16:50
  John
• 16:50
  Washburn's
• 16:51
  response
• 16:51
  to
• 16:51
  the
• 16:51
  festival
• 16:51
  claims
• 16:52
  was
• 16:52
  more
• 16:52
  direct:
• 16:53
  "The
• 16:53
  suggestion
• 16:54
  that
• 16:54
  paying
• 16:54
  for
• 16:54
  a
• 16:54
  screening
• 16:55
  at
• 16:55
  some
• 16:55
  random
• 16:55
  theatre
• 16:56
  in
• 16:56
  the
• 16:56
  same
• 16:56
  town
• 16:57
  and
• 16:57
  at
• 16:57
  the
• 16:57
  same
• 16:57
  time
• 16:57
  as
• 16:57
  a
• 16:57
  major
• 16:58
  festival
• 16:58
  is
• 16:59
  somehow
• 16:59
  the
• 16:59
  same
• 16:59
  thing
• 16:59
  as
• 16:59
  being
• 17:00
  selected
• 17:00
  by
• 17:00
  that
• 17:00
  festival
• 17:01
  is
• 17:01
  misleading
• 17:02
  at
• 17:02
  best.
• 17:03
  Spurious
• 17:03
  bullshittery,
• 17:04
  really."
• 17:05
  The
• 17:05
  Wall
• 17:05
  Street
• 17:05
  Journal
• 17:06
  later
• 17:06
  added
• 17:06
  a
• 17:06
  correction
• 17:07
  clarifying
• 17:07
  the
• 17:07
  distinction.
• 17:08
  The
• 17:08
  pattern
• 17:09
  is
• 17:09
  familiar
• 17:09
  —
• 17:10
  real
• 17:10
  capability
• 17:11
  dressed
• 17:11
  in
• 17:11
  borrowed
• 17:11
  prestige,
• 17:12
  traveling
• 17:13
  faster
• 17:13
  than
• 17:13
  anyone
• 17:14
  can
• 17:14
  check
• 17:14
  .