Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – August 20, 2026

Transcript

AGI Dreams Podcast – August 20, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  most
• 0:00
  reliably
• 0:00
  productive
• 0:01
  corner
• 0:01
  of
• 0:01
  local
• 0:01
  AI
• 0:02
  right
• 0:02
  now
• 0:02
  is
• 0:02
  a
• 0:02
  single
• 0:03
  developer
• 0:03
  who
• 0:03
  will
• 0:03
  not
• 0:04
  stop
• 0:04
  optimizing
• 0:04
  one
• 0:04
  model
• 0:05
  on
• 0:05
  one
• 0:05
  card,
• 0:06
  this
• 0:06
  week
• 0:06
  pushing
• 0:06
  Qwen3.8-27B
• 0:09
  to
• 0:09
  roughly
• 0:10
  138
• 0:11
  tokens
• 0:11
  per
• 0:11
  second
• 0:11
  single-user
• 0:12
  on
• 0:12
  a
• 0:12
  power-limited
• 0:13
  RTX
• 0:13
  3090,
• 0:14
  up
• 0:15
  from
• 0:15
  124
• 0:16
  a
• 0:16
  day
• 0:16
  earlier,
• 0:16
  and
• 0:17
  942
• 0:17
  tps
• 0:18
  at
• 0:18
  64
• 0:18
  concurrent
• 0:19
  .
• 0:20
  The
• 0:20
  engineering
• 0:20
  underneath
• 0:21
  the
• 0:21
  headline
• 0:21
  is
• 0:21
  a
• 0:21
  tour
• 0:21
  of
• 0:22
  where
• 0:22
  speculative
• 0:22
  decoding
• 0:23
  has
• 0:23
  landed:
• 0:24
  DFlash2
• 0:24
  block
• 0:25
  drafting,
• 0:25
  the
• 0:26
  drafter
• 0:26
  requantized
• 0:26
  to
• 0:27
  W4A16
• 0:29
  GPTQ
• 0:30
  that
• 0:30
  shrinks
• 0:31
  it
• 0:31
  from
• 0:31
  3.85
• 0:32
  GB
• 0:32
  to
• 0:32
  1.19
• 0:33
  GB
• 0:34
  and
• 0:34
  stops
• 0:35
  being
• 0:35
  a
• 0:35
  net
• 0:35
  loss
• 0:35
  on
• 0:36
  a
• 0:36
  24
• 0:36
  GB
• 0:36
  card,
• 0:37
  and
• 0:37
  a
• 0:37
  clever
• 0:38
  "lookup-augmented
• 0:39
  drafting"
• 0:39
  trick
• 0:39
  —
• 0:40
  a
• 0:40
  Triton
• 0:40
  kernel
• 0:41
  that
• 0:41
  scans
• 0:41
  the
• 0:41
  request's
• 0:42
  own
• 0:42
  token
• 0:42
  history
• 0:43
  for
• 0:43
  the
• 0:43
  last
• 0:43
  occurrence
• 0:44
  of
• 0:44
  the
• 0:44
  last
• 0:44
  6-12
• 0:45
  generated
• 0:45
  tokens
• 0:45
  and
• 0:46
  proposes
• 0:46
  what
• 0:46
  followed.
• 0:47
  That
• 0:47
  single
• 0:48
  idea
• 0:48
  buys
• 0:48
  +29%
• 0:49
  tokens
• 0:50
  per
• 0:50
  step
• 0:50
  on
• 0:50
  repetitive
• 0:51
  work,
• 0:52
  because
• 0:52
  a
• 0:52
  long-context
• 0:53
  assistant
• 0:53
  spends
• 0:54
  much
• 0:54
  of
• 0:54
  its
• 0:54
  output
• 0:54
  quoting
• 0:55
  text
• 0:55
  already
• 0:55
  in
• 0:55
  the
• 0:55
  prompt,
• 0:56
  beyond
• 0:56
  the
• 0:56
  drafter's
• 0:57
  2,048-token
• 0:58
  window.
• 0:59
  The
• 0:59
  author
• 0:59
  also
• 1:00
  caught
• 1:00
  a
• 1:00
  real
• 1:00
  correctness
• 1:01
  bug
• 1:01
  in
• 1:01
  the
• 1:01
  backported
• 1:02
  vLLM
• 1:02
  PR
• 1:02
  —
• 1:03
  a
• 1:03
  cached-logits
• 1:04
  temperature
• 1:04
  mismatch
• 1:05
  that
• 1:05
  would
• 1:05
  have
• 1:05
  silently
• 1:06
  corrupted
• 1:06
  the
• 1:06
  proposal
• 1:06
  distribution,
• 1:08
  the
• 1:08
  kind
• 1:08
  of
• 1:08
  defect
• 1:08
  that
• 1:09
  makes
• 1:09
  "exact"
• 1:09
  decoding
• 1:10
  quietly
• 1:10
  inexact
• 1:10
  —
• 1:12
  while
• 1:12
  quality
• 1:12
  held
• 1:12
  at
• 1:13
  perplexity
• 1:13
  8.09
• 1:14
  and
• 1:14
  GSM8K
• 1:16
  96.5%.
• 1:18
  A
• 1:18
  parallel
• 1:19
  NInfer
• 1:19
  fork
• 1:19
  claims
• 1:20
  250-350K
• 1:22
  tokens
• 1:23
  of
• 1:23
  context
• 1:23
  in
• 1:23
  a
• 1:23
  single
• 1:24
  RTX
• 1:24
  4090's
• 1:25
  VRAM
• 1:26
  via
• 1:26
  a
• 1:26
  new
• 1:26
  rk2v4-e8
• 1:27
  KV-cache
• 1:28
  quant,
• 1:29
  but
• 1:29
  ships
• 1:29
  no
• 1:29
  perplexity
• 1:30
  or
• 1:30
  task
• 1:30
  numbers
• 1:31
  —
• 1:31
  a
• 1:31
  memory-allocation
• 1:32
  feat,
• 1:33
  not
• 1:33
  a
• 1:33
  quality
• 1:33
  result,
• 1:34
  as
• 1:34
  its
• 1:34
  own
• 1:35
  top
• 1:35
  commenter
• 1:35
  notes
• 1:35
  .
• 1:36
  Feeding
• 1:37
  both
• 1:37
  efforts
• 1:37
  is
• 1:37
  a
• 1:38
  llama.cpp
• 1:39
  pull
• 1:40
  request
• 1:40
  adding
• 1:40
  adaptive
• 1:41
  multi-token
• 1:42
  prediction,
• 1:43
  a
• 1:43
  counting-style
• 1:43
  state
• 1:44
  machine
• 1:44
  that
• 1:44
  raises
• 1:45
  MTP
• 1:45
  depth
• 1:45
  from
• 1:46
  a
• 1:46
  floor
• 1:46
  of
• 1:46
  3
• 1:46
  up
• 1:46
  to
• 1:47
  12
• 1:47
  when
• 1:47
  acceptance
• 1:48
  is
• 1:48
  high
• 1:48
  and
• 1:48
  drops
• 1:49
  it
• 1:49
  when
• 1:49
  prose
• 1:49
  gets
• 1:50
  unpredictable
• 1:50
  .
• 1:51
  The
• 1:51
  author
• 1:51
  is
• 1:51
  admirably
• 1:52
  honest
• 1:53
  that
• 1:53
  dense
• 1:53
  prose
• 1:53
  runs
• 1:54
  about
• 1:54
  3%
• 1:55
  slower
• 1:55
  than
• 1:55
  fixed
• 1:55
  MTP=3,
• 1:57
  while
• 1:57
  code
• 1:58
  generation
• 1:58
  gains
• 1:59
  10-15%
• 2:02
  and
• 2:02
  recalling
• 2:02
  code
• 2:03
  from
• 2:03
  the
• 2:03
  thinking
• 2:03
  phase
• 2:03
  can
• 2:04
  exceed
• 2:04
  50%
• 2:04
  —
• 2:06
  the
• 2:06
  recurring
• 2:06
  lesson
• 2:06
  that
• 2:07
  these
• 2:07
  gains
• 2:07
  live
• 2:07
  and
• 2:08
  die
• 2:08
  on
• 2:08
  draft
• 2:08
  acceptance,
• 2:09
  and
• 2:09
  a
• 2:09
  wrong
• 2:10
  static
• 2:10
  depth
• 2:10
  leaves
• 2:11
  throughput
• 2:11
  on
• 2:11
  the
• 2:11
  table.
• 2:12
  Against
• 2:13
  all
• 2:13
  this
• 2:13
  consumer-scale
• 2:14
  ingenuity,
• 2:15
  someone
• 2:15
  posted
• 2:15
  Nvidia's
• 2:16
  figure
• 2:16
  of
• 2:16
  Qwen3.8
• 2:18
  2.4T
• 2:20
  at
• 2:21
  288k
• 2:22
  tokens/s
• 2:23
  on
• 2:23
  a
• 2:23
  GB300
• 2:25
  NVL72,
• 2:27
  roughly
• 2:27
  4K
• 2:27
  tokens
• 2:28
  per
• 2:28
  second
• 2:28
  per
• 2:29
  GPU
• 2:29
  across
• 2:29
  72
• 2:30
  of
• 2:30
  them
• 2:30
  .
• 2:31
  The
• 2:31
  comments
• 2:32
  joked
• 2:32
  about
• 2:32
  dusting
• 2:32
  off
• 2:33
  the
• 2:33
  $4
• 2:33
  million
• 2:33
  rack
• 2:34
  in
• 2:34
  the
• 2:34
  closet,
• 2:35
  which
• 2:35
  is
• 2:35
  the
• 2:35
  point:
• 2:36
  the
• 2:36
  interesting
• 2:37
  frontier
• 2:37
  is
• 2:37
  a
• 2:37
  3090
• 2:38
  doing
• 2:38
  138
• 2:39
  tps,
• 2:40
  not
• 2:40
  a
• 2:40
  datacenter.
• 2:42
  Lightricks
• 2:42
  also
• 2:43
  pushed
• 2:43
  LTX-2.3
• 2:44
  to
• 2:44
  trending,
• 2:45
  though
• 2:45
  its
• 2:45
  model
• 2:45
  card
• 2:46
  was
• 2:46
  empty
• 2:46
  at
• 2:46
  capture
• 2:46
  .
• 2:49
  Unsloth
• 2:50
  shipped
• 2:50
  Dynamic
• 2:50
  v3.0
• 2:52
  GGUFs,
• 2:53
  claiming
• 2:53
  more
• 2:53
  than
• 2:53
  10%
• 2:54
  better
• 2:54
  top-1%
• 2:55
  accuracy
• 2:55
  at
• 2:55
  the
• 2:55
  same
• 2:56
  file
• 2:56
  size
• 2:56
  over
• 2:56
  v2.0,
• 2:58
  on
• 2:58
  the
• 2:58
  back
• 2:58
  of
• 2:58
  5.1
• 2:59
  million
• 2:59
  Qwen3.8
• 3:02
  downloads
• 3:02
  in
• 3:02
  five
• 3:02
  days
• 3:03
  .
• 3:03
  The
• 3:03
  methodology
• 3:04
  is
• 3:04
  the
• 3:04
  substance:
• 3:05
  the
• 3:05
  team
• 3:05
  built
• 3:06
  a
• 3:06
  300-example
• 3:07
  held-out
• 3:08
  evaluation
• 3:09
  set
• 3:09
  from
• 3:09
  Terminal-Bench
• 3:09
  2.1,
• 3:10
  DeepSWE,
• 3:11
  Harbor,
• 3:11
  and
• 3:11
  MathArena,
• 3:12
  then
• 3:12
  extended
• 3:13
  KL-divergence
• 3:13
  top-1%
• 3:14
  agreement
• 3:14
  to
• 3:14
  32-token
• 3:15
  greedy
• 3:15
  trajectories
• 3:16
  —
• 3:17
  a
• 3:17
  real
• 3:17
  attempt
• 3:17
  to
• 3:17
  measure
• 3:18
  overfitting
• 3:18
  to
• 3:18
  the
• 3:18
  calibration
• 3:19
  set.
• 3:20
  They
• 3:20
  stress
• 3:20
  this
• 3:20
  is
• 3:20
  pure
• 3:21
  post-training
• 3:21
  quantization
• 3:22
  with
• 3:22
  a
• 3:22
  public
• 3:22
  imatrix
• 3:23
  file,
• 3:23
  no
• 3:23
  quantization-aware
• 3:24
  training;
• 3:25
  a
• 3:25
  new
• 3:25
  UD-1bit
• 3:26
  quant
• 3:26
  at
• 3:27
  6.2
• 3:27
  GB
• 3:28
  retains
• 3:28
  around
• 3:28
  72%
• 3:29
  top-1%
• 3:30
  accuracy
• 3:31
  while
• 3:31
  being
• 3:31
  89%
• 3:32
  smaller.
• 3:33
  Their
• 3:33
  broader
• 3:34
  argument,
• 3:34
  borrowed
• 3:35
  from
• 3:35
  "Accuracy
• 3:35
  is
• 3:36
  Not
• 3:36
  All
• 3:36
  You
• 3:36
  Need,"
• 3:37
  is
• 3:37
  that
• 3:37
  KL
• 3:38
  divergence
• 3:38
  tracks
• 3:39
  answer
• 3:39
  flips
• 3:39
  far
• 3:39
  better
• 3:40
  than
• 3:40
  MMLU
• 3:40
  alone
• 3:40
  —
• 3:41
  a
• 3:41
  leaderboard
• 3:42
  that,
• 3:42
  by
• 3:42
  their
• 3:42
  account,
• 3:43
  mostly
• 3:43
  measures
• 3:44
  its
• 3:44
  own
• 3:44
  tokenizer
• 3:44
  harness
• 3:45
  bugs.
• 3:46
  That
• 3:46
  framing
• 3:46
  runs
• 3:47
  directly
• 3:47
  into
• 3:47
  Liquid
• 3:48
  AI's
• 3:48
  release
• 3:48
  of
• 3:48
  QAD
• 3:48
  Q4
• 3:49
  0
• 3:49
  checkpoints
• 3:49
  for
• 3:49
  the
• 3:50
  LFM2.5
• 3:52
  family,
• 3:53
  which
• 3:53
  take
• 3:53
  the
• 3:53
  opposite
• 3:53
  bet:
• 3:54
  distill
• 3:54
  a
• 3:54
  high-precision
• 3:55
  teacher
• 3:55
  into
• 3:55
  a
• 3:55
  quantized
• 3:56
  student
• 3:56
  and
• 3:56
  recover
• 3:57
  97%
• 3:57
  of
• 3:57
  the
• 3:58
  accuracy
• 3:58
  lost
• 3:58
  to
• 3:58
  quantization,
• 3:59
  at
• 3:59
  native
• 3:59
  Q4
• 4:00
  0
• 4:00
  memory
• 4:00
  and
• 4:00
  speed
• 4:00
  .
• 4:01
  On
• 4:01
  paper
• 4:01
  the
• 4:01
  230M
• 4:02
  and
• 4:02
  350M
• 4:03
  checkpoints
• 4:04
  match
• 4:04
  Q5
• 4:04
  K
• 4:05
  M
• 4:05
  quality
• 4:05
  at
• 4:05
  4-33%
• 4:06
  higher
• 4:07
  decode
• 4:07
  throughput
• 4:07
  across
• 4:08
  four
• 4:08
  edge
• 4:08
  targets,
• 4:09
  from
• 4:09
  a
• 4:09
  MacBook
• 4:09
  to
• 4:09
  a
• 4:09
  Raspberry
• 4:10
  Pi
• 4:10
  5.
• 4:11
  The
• 4:11
  caution
• 4:11
  is
• 4:11
  not
• 4:11
  Liquid's
• 4:12
  numbers
• 4:12
  but
• 4:12
  the
• 4:12
  technique's
• 4:13
  failure
• 4:13
  mode:
• 4:13
  output-distribution
• 4:14
  matching
• 4:15
  can
• 4:15
  leave
• 4:15
  internal
• 4:15
  representations
• 4:16
  drifting,
• 4:17
  damage
• 4:17
  concentrating
• 4:17
  in
• 4:17
  coding
• 4:18
  while
• 4:18
  a
• 4:18
  healthy
• 4:18
  perplexity
• 4:19
  masks
• 4:19
  broken
• 4:19
  tool
• 4:19
  use.
• 4:21
  Liquid
• 4:21
  reports
• 4:21
  aggregate
• 4:22
  recovery
• 4:22
  across
• 4:22
  GPQA,
• 4:23
  IFEval,
• 4:24
  and
• 4:24
  BFCLv4
• 4:24
  —
• 4:26
  but
• 4:26
  the
• 4:26
  burden
• 4:26
  of
• 4:26
  proof
• 4:27
  for
• 4:27
  output-matching
• 4:27
  QAD
• 4:28
  has
• 4:28
  risen,
• 4:28
  and
• 4:29
  edge
• 4:29
  deployment
• 4:29
  is
• 4:29
  where
• 4:29
  a
• 4:30
  masked
• 4:30
  coding
• 4:30
  regression
• 4:31
  bites
• 4:31
  hardest.
• 4:32
  Underneath
• 4:32
  both
• 4:32
  stories
• 4:33
  is
• 4:33
  a
• 4:33
  metric
• 4:33
  worth
• 4:33
  adopting.
• 4:34
  The
• 4:34
  "Intelligence
• 4:35
  per
• 4:35
  Watt"
• 4:35
  paper
• 4:35
  argues
• 4:36
  local
• 4:36
  AI
• 4:36
  should
• 4:36
  be
• 4:36
  benchmarked
• 4:37
  on
• 4:37
  efficiency
• 4:38
  over
• 4:38
  a
• 4:38
  model's
• 4:38
  lifetime,
• 4:39
  and
• 4:39
  the
• 4:39
  comment
• 4:39
  thread
• 4:40
  did
• 4:40
  the
• 4:40
  useful
• 4:40
  work
• 4:40
  of
• 4:40
  critiquing
• 4:41
  it
• 4:41
  .
• 4:41
  The
• 4:42
  sharpest
• 4:42
  objection:
• 4:43
  intelligence
• 4:44
  per
• 4:44
  watt
• 4:44
  on
• 4:44
  one
• 4:44
  resident
• 4:45
  model
• 4:45
  is
• 4:45
  a
• 4:45
  property
• 4:45
  of
• 4:45
  the
• 4:45
  model,
• 4:46
  but
• 4:46
  people
• 4:47
  run
• 4:47
  a
• 4:47
  stack
• 4:47
  under
• 4:47
  a
• 4:47
  fixed
• 4:48
  VRAM
• 4:48
  budget,
• 4:49
  where
• 4:49
  the
• 4:49
  binding
• 4:49
  constraint
• 4:50
  is
• 4:50
  memory
• 4:51
  and
• 4:51
  a
• 4:51
  real
• 4:51
  share
• 4:52
  of
• 4:52
  the
• 4:52
  energy
• 4:52
  bill
• 4:52
  goes
• 4:53
  to
• 4:53
  loading
• 4:53
  and
• 4:53
  evicting
• 4:54
  weights,
• 4:54
  not
• 4:54
  computing.
• 4:55
  The
• 4:55
  academic
• 4:56
  bookend
• 4:56
  pairs
• 4:56
  a
• 4:56
  68,000-parameter
• 4:59
  Transformer
• 4:59
  controller
• 5:00
  with
• 5:00
  an
• 5:00
  Artificial
• 5:00
  Bee
• 5:00
  Colony
• 5:01
  swarm
• 5:01
  for
• 5:02
  neural-architecture
• 5:02
  search,
• 5:03
  reaching
• 5:03
  84.85%
• 5:05
  on
• 5:05
  CIFAR-10
• 5:08
  in
• 5:08
  about
• 5:08
  0.16
• 5:09
  GPU-days
• 5:10
  on
• 5:10
  a
• 5:10
  6
• 5:10
  GB
• 5:10
  RTX
• 5:11
  3060
• 5:11
  —
• 5:12
  and,
• 5:13
  more
• 5:13
  relevant
• 5:13
  here,
• 5:14
  autonomously
• 5:15
  optimizing
• 5:15
  an
• 5:15
  F1
• 5:15
  of
• 5:15
  0.7178
• 5:17
  on
• 5:17
  credit-card
• 5:17
  fraud
• 5:17
  detection
• 5:18
  with
• 5:18
  a
• 5:18
  ~4,600-parameter
• 5:19
  network
• 5:19
  .
• 5:20
  The
• 5:20
  paper
• 5:21
  is
• 5:21
  honest
• 5:21
  about
• 5:21
  its
• 5:21
  edges:
• 5:22
  one
• 5:23
  CIFAR-100
• 5:25
  seed
• 5:26
  blew
• 5:26
  up
• 5:26
  to
• 5:26
  3.1
• 5:27
  million
• 5:27
  parameters
• 5:27
  when
• 5:27
  proxy
• 5:28
  noise
• 5:28
  overwhelmed
• 5:28
  the
• 5:28
  depth
• 5:29
  penalty.
• 5:30
  Automated,
• 5:30
  yes;
• 5:31
  unsupervised,
• 5:32
  not
• 5:32
  yet.
• 5:35
  The
• 5:35
  community's
• 5:36
  month-long
• 5:36
  project
• 5:37
  of
• 5:37
  squeezing
• 5:38
  DeepSeek
• 5:38
  V4
• 5:39
  Flash
• 5:39
  onto
• 5:39
  a
• 5:39
  single
• 5:39
  machine
• 5:40
  reached
• 5:40
  its
• 5:40
  logical
• 5:41
  endpoint:
• 5:42
  a
• 5:42
  pinned
• 5:42
  Docker
• 5:43
  recipe
• 5:43
  serving
• 5:43
  the
• 5:43
  284B
• 5:45
  0731
• 5:46
  build
• 5:46
  on
• 5:46
  one
• 5:47
  NVIDIA
• 5:47
  DGX
• 5:48
  Spark
• 5:48
  via
• 5:48
  SparkInfer,
• 5:49
  exposing
• 5:49
  a
• 5:50
  262,144-token
• 5:52
  limit
• 5:52
  with
• 5:52
  a
• 5:52
  compact
• 5:53
  K64
• 5:54
  speculative
• 5:54
  draft
• 5:54
  .
• 5:55
  What
• 5:56
  makes
• 5:56
  it
• 5:56
  credible
• 5:56
  is
• 5:56
  its
• 5:56
  refusal
• 5:57
  to
• 5:57
  round
• 5:57
  up:
• 5:57
  the
• 5:58
  measured
• 5:58
  clean-image
• 5:58
  code
• 5:59
  decode
• 5:59
  ran
• 5:59
  34.30
• 6:01
  to
• 6:01
  48.88
• 6:02
  tok/s,
• 6:03
  and
• 6:03
  the
• 6:03
  author
• 6:04
  states
• 6:04
  plainly
• 6:04
  that
• 6:04
  the
• 6:05
  requested
• 6:05
  steady
• 6:05
  35
• 6:06
  tok/s
• 6:07
  floor
• 6:07
  is
• 6:07
  "still
• 6:08
  an
• 6:08
  open
• 6:08
  optimization
• 6:09
  gate
• 6:09
  and
• 6:09
  is
• 6:09
  not
• 6:10
  claimed
• 6:10
  as
• 6:10
  complete."
• 6:11
  The
• 6:11
  KV-format
• 6:12
  disclosure
• 6:12
  is
• 6:13
  a
• 6:13
  model
• 6:13
  of
• 6:13
  the
• 6:13
  genre:
• 6:14
  the
• 6:14
  reliable
• 6:15
  route
• 6:15
  uses
• 6:15
  a
• 6:15
  584-byte
• 6:16
  padded
• 6:17
  FP8
• 6:17
  sparse-MLA
• 6:18
  record,
• 6:19
  not
• 6:19
  the
• 6:19
  true
• 6:20
  432-byte
• 6:21
  NVFP4
• 6:22
  format,
• 6:23
  because
• 6:23
  the
• 6:23
  432-byte
• 6:24
  version
• 6:24
  "boots
• 6:25
  and
• 6:25
  passes
• 6:25
  its
• 6:25
  isolated
• 6:26
  numerical
• 6:26
  oracle"
• 6:27
  but
• 6:27
  "produced
• 6:27
  corrupted
• 6:28
  full-model
• 6:28
  text
• 6:29
  and
• 6:29
  is
• 6:29
  therefore
• 6:30
  disabled."
• 6:31
  A
• 6:31
  numerical
• 6:31
  unit
• 6:32
  test
• 6:32
  that
• 6:32
  passes
• 6:33
  while
• 6:33
  the
• 6:33
  full
• 6:33
  model
• 6:33
  generates
• 6:34
  garbage
• 6:34
  is
• 6:34
  the
• 6:35
  trap
• 6:35
  self-reported
• 6:35
  benchmarks
• 6:36
  hide,
• 6:36
  and
• 6:37
  this
• 6:37
  recipe
• 6:37
  surfaces
• 6:38
  it.
• 6:38
  Set
• 6:38
  against
• 6:39
  that
• 6:39
  engineering
• 6:39
  discipline
• 6:40
  is
• 6:40
  the
• 6:40
  other
• 6:40
  DeepSeek
• 6:41
  story,
• 6:42
  pure
• 6:42
  product-terms
• 6:43
  arbitrage.
• 6:44
  Deepseek-API
• 6:45
  turns
• 6:45
  the
• 6:45
  free
• 6:45
  consumer
• 6:46
  web
• 6:46
  chat
• 6:46
  at
• 6:46
  chat.deepseek.com
• 6:48
  into
• 6:48
  an
• 6:48
  OpenAI-compatible
• 6:49
  local
• 6:49
  server
• 6:50
  with
• 6:50
  "no
• 6:50
  API
• 6:51
  key,
• 6:51
  no
• 6:52
  credits,
• 6:52
  no
• 6:52
  paid
• 6:52
  plan,"
• 6:53
  by
• 6:53
  automating
• 6:54
  your
• 6:54
  own
• 6:54
  signed-in
• 6:54
  account
• 6:55
  through
• 6:55
  a
• 6:55
  headless
• 6:55
  browser
• 6:55
  .
• 6:56
  It
• 6:56
  is
• 6:56
  tidy
• 6:57
  work:
• 6:57
  it
• 6:58
  clears
• 6:58
  the
• 6:58
  AWS
• 6:59
  WAF
• 6:59
  human-check
• 7:00
  once
• 7:00
  by
• 7:00
  hand,
• 7:00
  caches
• 7:01
  the
• 7:01
  session
• 7:01
  for
• 7:01
  roughly
• 7:02
  six
• 7:02
  hours,
• 7:03
  and
• 7:03
  defeats
• 7:03
  the
• 7:03
  per-completion
• 7:04
  proof-of-work
• 7:05
  challenge
• 7:05
  by
• 7:05
  running
• 7:05
  DeepSeek's
• 7:06
  own
• 7:06
  sha3
• 7:07
  wasm
• 7:07
  bg.wasm
• 7:10
  inside
• 7:11
  a
• 7:11
  wasmtime
• 7:11
  sandbox.
• 7:12
  It
• 7:12
  is
• 7:12
  also
• 7:13
  honest
• 7:13
  about
• 7:13
  its
• 7:13
  ceiling:
• 7:14
  because
• 7:14
  that
• 7:15
  solver's
• 7:15
  wasmtime
• 7:16
  store
• 7:16
  is
• 7:16
  not
• 7:16
  reentrant,
• 7:17
  parallel
• 7:18
  requests
• 7:18
  serialize
• 7:19
  behind
• 7:19
  a
• 7:19
  lock,
• 7:19
  so
• 7:20
  "throughput
• 7:20
  is
• 7:20
  sequential,
• 7:21
  not
• 7:21
  parallel."
• 7:22
  An
• 7:22
  elegant
• 7:23
  solution
• 7:23
  to
• 7:23
  a
• 7:23
  self-inflicted
• 7:24
  problem
• 7:24
  —
• 7:25
  DeepSeek's
• 7:26
  first-party
• 7:26
  API
• 7:27
  is
• 7:27
  already
• 7:28
  the
• 7:28
  cheapest
• 7:28
  way
• 7:28
  to
• 7:28
  run
• 7:29
  V4,
• 7:29
  so
• 7:30
  the
• 7:30
  whole
• 7:30
  value
• 7:30
  proposition
• 7:31
  rests
• 7:31
  on
• 7:31
  dodging
• 7:32
  a
• 7:32
  price
• 7:32
  that
• 7:32
  is
• 7:32
  barely
• 7:32
  there,
• 7:33
  via
• 7:34
  a
• 7:34
  browser
• 7:34
  bridge
• 7:34
  that
• 7:34
  breaks
• 7:35
  when
• 7:35
  anti-abuse
• 7:35
  tuning
• 7:36
  rotates
• 7:36
  the
• 7:36
  challenge.
• 7:38
  Useful,
• 7:38
  clever,
• 7:39
  and
• 7:39
  structurally
• 7:39
  fragile.
• 7:42
  Most
• 7:43
  uncensored
• 7:43
  model
• 7:44
  uploads
• 7:44
  follow
• 7:44
  a
• 7:44
  familiar
• 7:45
  pattern:
• 7:46
  a
• 7:46
  near-empty
• 7:46
  model
• 7:47
  card,
• 7:47
  a
• 7:47
  llama.cpp
• 7:49
  invocation,
• 7:49
  and
• 7:49
  zero
• 7:50
  evaluation
• 7:50
  of
• 7:51
  what
• 7:51
  the
• 7:51
  abliteration
• 7:51
  actually
• 7:52
  cost.
• 7:53
  This
• 7:53
  week's
• 7:53
  Qwen3.8-27B-Abliterated-SFT
• 7:59
  is
• 7:59
  the
• 7:59
  opposite,
• 7:59
  and
• 8:00
  worth
• 8:00
  attention
• 8:00
  precisely
• 8:01
  for
• 8:01
  that
• 8:01
  reason
• 8:01
  .
• 8:02
  Rather
• 8:02
  than
• 8:02
  a
• 8:02
  weight-edit,
• 8:03
  it
• 8:03
  is
• 8:03
  an
• 8:03
  SFT
• 8:04
  LoRA
• 8:04
  trained
• 8:04
  on
• 8:05
  84
• 8:05
  EOS-terminated
• 8:06
  teacher
• 8:06
  completions,
• 8:07
  and
• 8:07
  its
• 8:07
  interesting
• 8:08
  claim
• 8:08
  is
• 8:08
  mechanistic,
• 8:09
  not
• 8:09
  political.
• 8:11
  Every
• 8:11
  abliteration
• 8:11
  method
• 8:12
  the
• 8:12
  author
• 8:12
  re-measured
• 8:13
  under
• 8:13
  an
• 8:13
  identical
• 8:13
  harness
• 8:14
  suppresses
• 8:14
  refusals
• 8:15
  by
• 8:15
  a
• 8:15
  similar
• 8:15
  margin,
• 8:16
  but
• 8:16
  each
• 8:16
  weight
• 8:16
  or
• 8:16
  projection
• 8:17
  edit
• 8:17
  adds
• 8:17
  +0.36
• 8:19
  to
• 8:19
  +0.67
• 8:20
  in
• 8:20
  "invalid"
• 8:21
  output
• 8:21
  —
• 8:21
  models
• 8:22
  that
• 8:22
  ramble
• 8:22
  to
• 8:22
  the
• 8:22
  token
• 8:23
  cap
• 8:23
  instead
• 8:23
  of
• 8:23
  stopping.
• 8:24
  The
• 8:24
  paper
• 8:25
  calls
• 8:25
  this
• 8:25
  a
• 8:25
  termination
• 8:26
  signature:
• 8:26
  free-running
• 8:27
  clean
• 8:27
  stops
• 8:27
  collapse
• 8:28
  to
• 8:28
  19-43%
• 8:30
  across
• 8:31
  the
• 8:31
  weight-edited
• 8:31
  competitors
• 8:32
  on
• 8:32
  harmful
• 8:32
  prompts,
• 8:33
  while
• 8:33
  the
• 8:33
  base
• 8:33
  sits
• 8:34
  at
• 8:34
  64%
• 8:35
  and
• 8:35
  this
• 8:35
  SFT
• 8:36
  model
• 8:36
  at
• 8:36
  97%.
• 8:38
  The
• 8:38
  claim
• 8:38
  is
• 8:38
  that
• 8:38
  refusal
• 8:39
  and
• 8:39
  answer-boundedness
• 8:40
  are
• 8:40
  entangled,
• 8:41
  coarse
• 8:41
  abliteration
• 8:41
  breaks
• 8:42
  the
• 8:42
  model's
• 8:42
  ability
• 8:42
  to
• 8:43
  end
• 8:43
  a
• 8:43
  sentence,
• 8:43
  and
• 8:44
  that
• 8:44
  doubles
• 8:44
  as
• 8:44
  a
• 8:44
  detection
• 8:45
  heuristic.
• 8:46
  Two
• 8:46
  things
• 8:46
  keep
• 8:46
  this
• 8:46
  from
• 8:47
  being
• 8:47
  a
• 8:47
  pure
• 8:47
  vendor
• 8:47
  win.
• 8:48
  First,
• 8:48
  the
• 8:49
  author
• 8:49
  is
• 8:49
  unusually
• 8:50
  candid
• 8:50
  about
• 8:50
  limits
• 8:50
  —
• 8:51
  thinking
• 8:51
  mode
• 8:51
  blows
• 8:52
  61%
• 8:53
  of
• 8:53
  generations
• 8:53
  into
• 8:53
  the
• 8:53
  2048-token
• 8:55
  cap,
• 8:55
  the
• 8:55
  local
• 8:56
  judge
• 8:56
  is
• 8:56
  itself
• 8:57
  an
• 8:57
  abliterated
• 8:57
  Qwen3.6,
• 9:01
  and
• 9:01
  no
• 9:01
  third-party
• 9:01
  SFT-class
• 9:02
  refusal-removal
• 9:03
  checkpoint
• 9:03
  exists
• 9:04
  for
• 9:04
  this
• 9:04
  base,
• 9:05
  so
• 9:05
  the
• 9:05
  central
• 9:05
  "SFT-class
• 9:06
  preserves
• 9:06
  termination"
• 9:07
  claim
• 9:07
  rests
• 9:07
  on
• 9:08
  their
• 9:08
  own
• 9:08
  arm.
• 9:09
  Second,
• 9:09
  this
• 9:09
  is
• 9:09
  not
• 9:10
  the
• 9:10
  first
• 9:10
  abliteration
• 9:10
  of
• 9:11
  Qwen3.8-27B
• 9:14
  to
• 9:14
  surface
• 9:14
  —
• 9:15
  their
• 9:15
  model
• 9:16
  card
• 9:16
  lists
• 9:16
  several
• 9:16
  other
• 9:17
  abliterated
• 9:17
  models
• 9:18
  that
• 9:18
  came
• 9:18
  before
• 9:18
  its
• 9:18
  release,
• 9:19
  so
• 9:19
  the
• 9:19
  novelty
• 9:20
  is
• 9:20
  the
• 9:20
  harness
• 9:21
  and
• 9:21
  the
• 9:21
  finding,
• 9:21
  not
• 9:22
  the
• 9:22
  deed.
• 9:23
  The
• 9:23
  residual
• 9:23
  8%
• 9:24
  still
• 9:24
  refused
• 9:24
  form
• 9:25
  a
• 9:25
  coherent
• 9:25
  floor
• 9:25
  —
• 9:26
  self-harm,
• 9:27
  minors,
• 9:28
  targeted
• 9:28
  harassment,
• 9:29
  disinformation
• 9:29
  —
• 9:30
  because
• 9:30
  the
• 9:30
  teacher
• 9:31
  data
• 9:31
  carried
• 9:31
  "answer
• 9:32
  accurately,
• 9:32
  no
• 9:33
  targeted
• 9:33
  harm."
• 9:34
  For
• 9:34
  a
• 9:34
  security
• 9:35
  audience
• 9:35
  the
• 9:35
  durable
• 9:35
  takeaway
• 9:36
  is
• 9:36
  the
• 9:36
  detection
• 9:37
  angle:
• 9:37
  if
• 9:38
  rambling
• 9:38
  non-termination
• 9:39
  is
• 9:39
  a
• 9:39
  reliable
• 9:40
  fingerprint
• 9:40
  of
• 9:40
  weight-edited
• 9:41
  refusal
• 9:41
  removal,
• 9:42
  that
• 9:42
  is
• 9:42
  a
• 9:42
  triage
• 9:43
  signal
• 9:43
  for
• 9:43
  spotting
• 9:44
  coarsely-tampered
• 9:45
  models
• 9:45
  in
• 9:45
  the
• 9:45
  wild.
• 9:48
  Reuven
• 9:48
  Cohen's
• 9:49
  periodic
• 9:49
  manifesto
• 9:50
  on
• 9:50
  why
• 9:50
  he
• 9:50
  ships
• 9:50
  so
• 9:51
  much
• 9:51
  reads
• 9:51
  as
• 9:51
  a
• 9:51
  full
• 9:51
  taxonomy
• 9:52
  of
• 9:52
  the
• 9:52
  RuV
• 9:52
  stack:
• 9:54
  RuFlo
• 9:54
  coordinating
• 9:55
  agent
• 9:55
  teams,
• 9:56
  MetaHarness
• 9:56
  supplying
• 9:56
  tools
• 9:57
  and
• 9:57
  acceptance
• 9:57
  criteria,
• 9:58
  RuVector
• 9:59
  connecting
• 9:59
  semantic
• 9:59
  memory
• 10:00
  and
• 10:00
  graphs,
• 10:00
  RVF
• 10:01
  packaging
• 10:01
  agent
• 10:02
  identity,
• 10:02
  and
• 10:03
  a
• 10:03
  "Darwin
• 10:03
  mode"
• 10:03
  that
• 10:03
  promotes
• 10:04
  a
• 10:04
  new
• 10:04
  approach
• 10:04
  "only
• 10:04
  when
• 10:05
  it
• 10:05
  beats
• 10:05
  its
• 10:05
  parent
• 10:05
  without
• 10:06
  weakening
• 10:06
  safety"
• 10:06
  .
• 10:07
  One
• 10:07
  of
• 10:07
  his
• 10:07
  newer
• 10:08
  research
• 10:08
  arms
• 10:08
  is
• 10:08
  LatentMesh,
• 10:09
  a
• 10:09
  new
• 10:10
  Rust
• 10:10
  prototype
• 10:10
  from
• 10:11
  the
• 10:11
  same
• 10:11
  author
• 10:11
  exploring
• 10:12
  "latent
• 10:12
  communication"
• 10:12
  —
• 10:13
  passing
• 10:13
  an
• 10:14
  aligned
• 10:14
  slice
• 10:14
  of
• 10:14
  one
• 10:15
  agent's
• 10:15
  hidden
• 10:15
  state
• 10:15
  directly
• 10:16
  instead
• 10:16
  of
• 10:16
  serializing
• 10:17
  tokens
• 10:17
  over
• 10:17
  the
• 10:17
  wire
• 10:17
  .
• 10:18
  What
• 10:18
  makes
• 10:18
  it
• 10:18
  more
• 10:19
  interesting
• 10:19
  than
• 10:19
  the
• 10:19
  pitch:
• 10:20
  its
• 10:20
  README
• 10:21
  disclaims
• 10:22
  novelty
• 10:22
  on
• 10:22
  the
• 10:22
  wire
• 10:23
  format
• 10:23
  and
• 10:23
  stakes
• 10:23
  everything
• 10:24
  on
• 10:24
  a
• 10:24
  falsification
• 10:24
  test
• 10:24
  —
• 10:25
  no
• 10:25
  agent-to-agent
• 10:26
  latent
• 10:26
  edge
• 10:26
  earns
• 10:26
  execution
• 10:27
  authority
• 10:27
  unless
• 10:28
  it
• 10:28
  beats
• 10:28
  five
• 10:28
  controls,
• 10:29
  including
• 10:29
  the
• 10:29
  text-equivalent
• 10:30
  of
• 10:30
  the
• 10:30
  same
• 10:30
  content,
• 10:30
  under
• 10:31
  a
• 10:31
  sign-flip
• 10:31
  permutation
• 10:32
  test,
• 10:32
  with
• 10:32
  unverified
• 10:33
  edges
• 10:33
  structurally
• 10:33
  capped
• 10:34
  at
• 10:34
  "observe
• 10:34
  only."
• 10:35
  The
• 10:36
  engineering
• 10:36
  is
• 10:36
  real
• 10:37
  —
• 10:37
  an
• 10:37
  exact
• 10:38
  O(d²n)
• 10:39
  reformulation
• 10:39
  of
• 10:39
  the
• 10:39
  alignment
• 10:40
  step
• 10:40
  cut
• 10:40
  fit
• 10:40
  time
• 10:40
  at
• 10:41
  dimension
• 10:41
  4096
• 10:42
  from
• 10:42
  2.6
• 10:43
  minutes
• 10:43
  to
• 10:43
  162
• 10:43
  milliseconds
• 10:44
  —
• 10:45
  but
• 10:45
  the
• 10:45
  "honest
• 10:45
  status"
• 10:46
  section
• 10:46
  admits
• 10:46
  there
• 10:46
  is
• 10:47
  no
• 10:47
  live
• 10:47
  LLM
• 10:47
  integration
• 10:48
  and
• 10:48
  the
• 10:49
  defining
• 10:49
  acceptance
• 10:49
  test
• 10:50
  has
• 10:50
  not
• 10:50
  been
• 10:50
  run.
• 10:51
  A
• 10:51
  well-instrumented
• 10:52
  hypothesis,
• 10:52
  and
• 10:52
  demanding
• 10:53
  that
• 10:53
  latent
• 10:53
  transfer
• 10:54
  beat
• 10:54
  text
• 10:54
  before
• 10:54
  trusting
• 10:55
  it
• 10:55
  is
• 10:55
  the
• 10:55
  right
• 10:55
  instinct.
• 10:56
  At
• 10:56
  the
• 10:56
  opposite
• 10:57
  end
• 10:57
  of
• 10:57
  the
• 10:57
  spectrum
• 10:57
  sits
• 10:58
  fx,
• 10:58
  a
• 10:59
  coding
• 10:59
  agent
• 10:59
  written
• 10:59
  in
• 11:00
  Zig
• 11:00
  that
• 11:00
  fits
• 11:00
  in
• 11:00
  a
• 11:00
  6.39
• 11:01
  MiB
• 11:02
  binary
• 11:02
  and
• 11:02
  cold-starts
• 11:03
  fast
• 11:03
  enough
• 11:03
  to
• 11:03
  pack
• 11:03
  many
• 11:04
  instances
• 11:04
  on
• 11:04
  one
• 11:04
  machine
• 11:05
  .
• 11:06
  Model-agnostic,
• 11:07
  Apache-2.0,
• 11:08
  closer
• 11:09
  to
• 11:09
  a
• 11:09
  Unix
• 11:09
  shell
• 11:10
  than
• 11:10
  an
• 11:10
  IDE-in-a-terminal
• 11:10
  —
• 11:11
  the
• 11:11
  minimalist
• 11:12
  counterweight
• 11:12
  to
• 11:12
  the
• 11:12
  everything-stacks
• 11:13
  above,
• 11:14
  aimed
• 11:14
  at
• 11:14
  agent
• 11:14
  sandboxes
• 11:15
  and
• 11:15
  embedding.
• 11:16
  DreamLab's
• 11:17
  nine-phase
• 11:17
  workshop
• 11:18
  curriculum,
• 11:19
  meanwhile,
• 11:19
  walks
• 11:19
  from
• 11:20
  VS
• 11:20
  Code
• 11:20
  setup
• 11:20
  through
• 11:21
  local
• 11:21
  RAG
• 11:22
  with
• 11:22
  Ollama
• 11:22
  and
• 11:23
  multi-agent
• 11:23
  ReAct
• 11:24
  systems
• 11:24
  to
• 11:24
  a
• 11:24
  security-and-ethics
• 11:25
  survey
• 11:25
  of
• 11:26
  the
• 11:26
  coding-agent
• 11:26
  landscape
• 11:26
  —
• 11:27
  a
• 11:27
  reminder
• 11:28
  that
• 11:28
  when
• 11:28
  tooling
• 11:29
  churns
• 11:29
  this
• 11:29
  fast,
• 11:30
  structured
• 11:30
  onboarding
• 11:31
  becomes
• 11:31
  its
• 11:31
  own
• 11:31
  product.
• 11:35
  The
• 11:35
  Cloud
• 11:35
  Security
• 11:36
  Alliance's
• 11:36
  research
• 11:37
  note
• 11:37
  is
• 11:37
  the
• 11:37
  most
• 11:37
  important
• 11:38
  item
• 11:38
  in
• 11:38
  this
• 11:38
  batch:
• 11:39
  a
• 11:39
  second
• 11:40
  documented
• 11:40
  case
• 11:40
  of
• 11:41
  an
• 11:41
  AI
• 11:41
  agent
• 11:41
  running
• 11:42
  most
• 11:42
  of
• 11:42
  an
• 11:42
  intrusion,
• 11:43
  and
• 11:43
  this
• 11:43
  time
• 11:43
  the
• 11:44
  operator's
• 11:44
  model
• 11:44
  choice
• 11:45
  is
• 11:45
  itself
• 11:45
  the
• 11:45
  story
• 11:46
  .
• 11:47
  A
• 11:47
  Chinese-speaking
• 11:47
  actor
• 11:48
  tracked
• 11:48
  by
• 11:48
  Palo
• 11:48
  Alto's
• 11:49
  Unit
• 11:49
  42
• 11:49
  as
• 11:49
  knaithe/KnYuan
• 11:55
  wired
• 11:55
  the
• 11:55
  open-source
• 11:56
  Hermes
• 11:56
  Agent
• 11:56
  framework
• 11:56
  to
• 11:57
  DeepSeek
• 11:57
  and,
• 11:58
  on
• 11:58
  May
• 11:58
  7,
• 11:58
  gave
• 11:59
  it
• 11:59
  a
• 11:59
  broad
• 11:59
  mandate.
• 12:00
  It
• 12:00
  enumerated
• 12:00
  84
• 12:01
  Langflow
• 12:01
  instances,
• 12:02
  failed,
• 12:02
  pivoted
• 12:03
  to
• 12:03
  n8n
• 12:04
  after
• 12:04
  finding
• 12:04
  a
• 12:04
  258-star
• 12:06
  proof-of-concept
• 12:06
  repo,
• 12:07
  used
• 12:07
  FOFA
• 12:08
  to
• 12:08
  confirm
• 12:08
  over
• 12:08
  647,000
• 12:10
  reachable
• 12:10
  n8n
• 12:11
  hosts,
• 12:11
  reasoned
• 12:12
  correctly
• 12:12
  about
• 12:12
  chained
• 12:12
  CVSS-10.0
• 12:15
  and
• 12:15
  9.9
• 12:15
  vulnerabilities,
• 12:16
  and
• 12:16
  narrowed
• 12:17
  to
• 12:17
  three
• 12:17
  targets
• 12:17
  —
• 12:18
  where
• 12:18
  authentication
• 12:19
  stopped
• 12:19
  it
• 12:19
  cold.
• 12:20
  The
• 12:20
  manual
• 12:20
  track
• 12:20
  was
• 12:21
  more
• 12:21
  successful:
• 12:22
  memory
• 12:22
  exfiltration
• 12:23
  from
• 12:23
  three
• 12:23
  organizations
• 12:24
  via
• 12:24
  Citrix
• 12:24
  NetScaler,
• 12:25
  command
• 12:26
  execution
• 12:26
  on
• 12:26
  eleven
• 12:27
  Marimo
• 12:27
  notebook
• 12:27
  instances,
• 12:28
  460+
• 12:29
  targets
• 12:30
  in
• 12:30
  total.
• 12:31
  The
• 12:31
  detail
• 12:31
  that
• 12:31
  matters
• 12:31
  most
• 12:32
  for
• 12:32
  AI
• 12:32
  safety
• 12:32
  controls:
• 12:33
  the
• 12:33
  actor
• 12:34
  tested
• 12:34
  Claude
• 12:34
  Code
• 12:35
  and
• 12:35
  Codex,
• 12:35
  found
• 12:36
  their
• 12:36
  safeguards
• 12:36
  refused,
• 12:37
  and
• 12:37
  simply
• 12:37
  routed
• 12:38
  around
• 12:38
  the
• 12:38
  guardrailed
• 12:39
  providers,
• 12:39
  settling
• 12:40
  on
• 12:40
  DeepSeek
• 12:40
  for
• 12:40
  the
• 12:41
  offensive
• 12:41
  reasoning.
• 12:42
  This
• 12:42
  is
• 12:42
  not
• 12:42
  the
• 12:42
  November
• 12:43
  GTG-1002
• 12:45
  pattern
• 12:46
  of
• 12:46
  circumventing
• 12:46
  safety
• 12:47
  training;
• 12:47
  it
• 12:48
  is
• 12:48
  model-selection
• 12:48
  as
• 12:48
  operational
• 12:49
  security.
• 12:50
  Unit
• 12:50
  42
• 12:51
  could
• 12:51
  reconstruct
• 12:52
  it
• 12:52
  granularly
• 12:52
  only
• 12:52
  because
• 12:53
  Hermes
• 12:53
  Agent
• 12:53
  accidentally
• 12:54
  launched
• 12:54
  an
• 12:54
  HTTP
• 12:55
  file
• 12:55
  server
• 12:56
  on
• 12:56
  port
• 12:56
  8888
• 12:57
  exposing
• 12:58
  the
• 12:58
  actor's
• 12:58
  own
• 12:59
  API
• 12:59
  keys
• 12:59
  and
• 12:59
  session
• 13:00
  logs.
• 13:01
  If
• 13:01
  that
• 13:01
  is
• 13:01
  offense
• 13:01
  at
• 13:02
  machine
• 13:02
  speed,
• 13:02
  the
• 13:02
  "mind
• 13:03
  viruses"
• 13:03
  research
• 13:04
  is
• 13:04
  the
• 13:04
  propagation
• 13:05
  vector
• 13:05
  nobody
• 13:05
  has
• 13:05
  priced
• 13:06
  in:
• 13:06
  researchers
• 13:07
  convinced
• 13:07
  one
• 13:08
  agent
• 13:08
  to
• 13:08
  adopt
• 13:08
  an
• 13:08
  idea
• 13:09
  and
• 13:09
  transmit
• 13:09
  it
• 13:09
  onward
• 13:10
  to
• 13:10
  others
• 13:10
  .
• 13:11
  The
• 13:11
  reflexive
• 13:12
  "it's
• 13:12
  just
• 13:12
  role
• 13:12
  play"
• 13:12
  comment
• 13:13
  undersells
• 13:13
  it
• 13:13
  —
• 13:14
  the
• 13:14
  paper's
• 13:15
  own
• 13:15
  note
• 13:15
  that
• 13:15
  "in
• 13:15
  the
• 13:15
  wild,
• 13:16
  viruses
• 13:17
  could
• 13:17
  emerge
• 13:17
  organically
• 13:18
  without
• 13:18
  intentionally
• 13:18
  designed
• 13:19
  system
• 13:19
  prompts"
• 13:20
  is
• 13:20
  the
• 13:20
  part
• 13:20
  worth
• 13:20
  taking
• 13:21
  seriously,
• 13:22
  and
• 13:22
  it
• 13:22
  fits
• 13:22
  an
• 13:22
  established
• 13:23
  taxonomy
• 13:23
  of
• 13:23
  inter-agent
• 13:24
  risks,
• 13:24
  from
• 13:24
  message
• 13:25
  tampering
• 13:25
  to
• 13:25
  cascading
• 13:26
  rogue
• 13:26
  emergence.
• 13:27
  The
• 13:27
  concrete
• 13:28
  supply-chain
• 13:28
  end
• 13:28
  of
• 13:28
  the
• 13:28
  same
• 13:29
  worry
• 13:29
  landed
• 13:29
  the
• 13:29
  same
• 13:30
  day:
• 13:31
  a
• 13:31
  compromised
• 13:31
  release
• 13:32
  of
• 13:32
  the
• 13:32
  Rust
• 13:32
  crate
• 13:32
  arrayref,
• 13:33
  downloaded
• 13:34
  245
• 13:35
  million
• 13:35
  times,
• 13:36
  added
• 13:36
  a
• 13:36
  dependency
• 13:37
  on
• 13:37
  a
• 13:37
  typosquatted
• 13:38
  proc-macro1
• 13:38
  whose
• 13:39
  build
• 13:39
  script
• 13:39
  downloads
• 13:40
  and
• 13:40
  runs
• 13:40
  a
• 13:40
  remote
• 13:40
  binary
• 13:41
  at
• 13:41
  compile
• 13:41
  time
• 13:42
  .
• 13:43
  The
• 13:43
  mechanics
• 13:43
  reward
• 13:44
  study
• 13:44
  —
• 13:45
  proc-macro1
• 13:45
  is
• 13:46
  a
• 13:46
  mechanical
• 13:46
  find-and-replace
• 13:47
  of
• 13:47
  the
• 13:47
  real
• 13:47
  proc-macro2,
• 13:48
  so
• 13:49
  it
• 13:49
  works
• 13:49
  as
• 13:49
  a
• 13:49
  drop-in
• 13:50
  while
• 13:50
  its
• 13:50
  build
• 13:50
  script
• 13:50
  reassembles
• 13:51
  a
• 13:51
  C2
• 13:51
  address
• 13:52
  from
• 13:52
  base64
• 13:53
  fragments
• 13:53
  and
• 13:53
  fetches
• 13:53
  a
• 13:53
  payload
• 13:54
  over
• 13:54
  a
• 13:54
  TLS
• 13:55
  connection
• 13:55
  that
• 13:55
  accepts
• 13:56
  any
• 13:56
  certificate.
• 13:57
  Yanking
• 13:57
  releases
• 13:58
  0.3.5
• 13:59
  through
• 14:00
  0.3.9
• 14:01
  was
• 14:01
  the
• 14:02
  flourish:
• 14:03
  Cargo's
• 14:03
  "consider
• 14:04
  updating
• 14:04
  to
• 14:04
  a
• 14:04
  version
• 14:05
  that
• 14:05
  is
• 14:05
  not
• 14:05
  yanked"
• 14:06
  warning
• 14:06
  nudges
• 14:06
  developers
• 14:07
  straight
• 14:07
  to
• 14:07
  the
• 14:08
  only
• 14:08
  non-yanked
• 14:08
  release,
• 14:09
  the
• 14:09
  malicious
• 14:09
  one.
• 14:11
  Build-time
• 14:11
  execution
• 14:12
  before
• 14:12
  the
• 14:12
  dependency
• 14:13
  is
• 14:13
  even
• 14:13
  used
• 14:13
  is
• 14:13
  nastier
• 14:14
  than
• 14:14
  the
• 14:14
  CI-time
• 14:15
  payloads
• 14:15
  that
• 14:15
  hit
• 14:16
  PyPI
• 14:16
  earlier
• 14:16
  this
• 14:17
  year,
• 14:17
  because
• 14:18
  simply
• 14:18
  compiling
• 14:19
  is
• 14:19
  enough.
• 14:22
  A
• 14:22
  Bluetooth
• 14:23
  headphone
• 14:23
  glitch
• 14:23
  turned
• 14:24
  into
• 14:24
  one
• 14:24
  of
• 14:24
  the
• 14:24
  cleaner
• 14:25
  pieces
• 14:25
  of
• 14:25
  fingerprinting
• 14:26
  forensics
• 14:26
  in
• 14:26
  a
• 14:26
  while.
• 14:27
  The
• 14:27
  author
• 14:27
  noticed
• 14:28
  that
• 14:28
  opening
• 14:28
  AliExpress
• 14:29
  in
• 14:29
  Firefox
• 14:30
  or
• 14:30
  Chrome
• 14:30
  silently
• 14:31
  killed
• 14:31
  audio
• 14:31
  from
• 14:32
  their
• 14:32
  phone,
• 14:32
  and
• 14:33
  muting
• 14:33
  the
• 14:33
  tab,
• 14:33
  the
• 14:34
  browser,
• 14:34
  and
• 14:34
  Windows
• 14:35
  did
• 14:35
  nothing
• 14:35
  .
• 14:36
  Instrumenting
• 14:36
  the
• 14:36
  AudioContext
• 14:37
  constructor
• 14:38
  found
• 14:38
  the
• 14:38
  culprit:
• 14:38
  two
• 14:39
  hidden
• 14:39
  WebAudio
• 14:39
  graphs
• 14:40
  created
• 14:40
  by
• 14:40
  collina.js
• 14:41
  and
• 14:41
  fireyejs.js,
• 14:44
  both
• 14:44
  under
• 14:45
  an
• 14:45
  AWSC
• 14:45
  directory
• 14:46
  that
• 14:46
  is
• 14:46
  part
• 14:46
  of
• 14:46
  Alibaba's
• 14:47
  anti-abuse
• 14:47
  tooling.
• 14:48
  Each
• 14:48
  runs
• 14:49
  an
• 14:49
  oscillator
• 14:49
  through
• 14:49
  an
• 14:49
  analyser
• 14:50
  into
• 14:50
  a
• 14:50
  zero-gain
• 14:51
  node
• 14:51
  connected
• 14:51
  to
• 14:51
  the
• 14:51
  system
• 14:52
  destination
• 14:52
  —
• 14:53
  inaudible,
• 14:53
  but
• 14:53
  enough
• 14:53
  live
• 14:54
  audio
• 14:54
  processing
• 14:54
  to
• 14:54
  keep
• 14:54
  the
• 14:55
  PC's
• 14:55
  Bluetooth
• 14:55
  path
• 14:55
  active
• 14:56
  and
• 14:56
  break
• 14:56
  multipoint.
• 14:57
  The
• 14:57
  audio
• 14:58
  is
• 14:58
  one
• 14:58
  probe
• 14:58
  in
• 14:58
  a
• 14:58
  much
• 14:59
  larger
• 14:59
  fingerprint
• 14:59
  spanning
• 15:00
  WebGL
• 15:00
  renderer
• 15:00
  strings,
• 15:01
  hardware
• 15:01
  concurrency,
• 15:02
  device
• 15:02
  memory,
• 15:03
  and
• 15:03
  mouse
• 15:03
  and
• 15:03
  scroll
• 15:03
  behavior,
• 15:04
  all
• 15:04
  encrypted
• 15:04
  and
• 15:04
  shipped
• 15:04
  to
• 15:05
  Alibaba
• 15:05
  telemetry.
• 15:06
  The
• 15:06
  professional
• 15:06
  read
• 15:07
  here
• 15:07
  is
• 15:07
  balanced.
• 15:08
  The
• 15:08
  anti-fraud
• 15:09
  justification
• 15:09
  is
• 15:10
  real
• 15:10
  —
• 15:11
  AliExpress
• 15:11
  genuinely
• 15:12
  fights
• 15:12
  account
• 15:12
  takeovers,
• 15:13
  coupon
• 15:13
  abuse,
• 15:14
  scraping,
• 15:14
  and
• 15:15
  payment
• 15:15
  fraud,
• 15:15
  and
• 15:16
  a
• 15:16
  fingerprint
• 15:16
  assembled
• 15:16
  from
• 15:16
  many
• 15:17
  independent
• 15:17
  measurements
• 15:17
  is
• 15:18
  far
• 15:18
  harder
• 15:18
  to
• 15:18
  spoof
• 15:18
  or
• 15:18
  clear
• 15:19
  than
• 15:19
  a
• 15:19
  cookie.
• 15:19
  That
• 15:20
  is
• 15:20
  the
• 15:20
  honest
• 15:20
  case
• 15:20
  for
• 15:20
  behavioral
• 15:21
  fingerprinting,
• 15:22
  and
• 15:22
  it
• 15:22
  matches
• 15:22
  how
• 15:22
  modern
• 15:23
  enforcement
• 15:23
  works:
• 15:24
  layered
• 15:24
  scoring
• 15:25
  that
• 15:25
  lets
• 15:25
  trusted
• 15:26
  customers
• 15:26
  skip
• 15:26
  the
• 15:26
  CAPTCHA.
• 15:27
  The
• 15:27
  problems
• 15:28
  are
• 15:28
  the
• 15:28
  implementation,
• 15:29
  not
• 15:29
  the
• 15:29
  intent
• 15:30
  —
• 15:30
  it
• 15:31
  runs
• 15:31
  on
• 15:31
  the
• 15:31
  general
• 15:31
  homepage
• 15:32
  before
• 15:32
  any
• 15:32
  sensitive
• 15:33
  action,
• 15:33
  it
• 15:33
  is
• 15:34
  deliberately
• 15:34
  obfuscated,
• 15:35
  and
• 15:35
  it
• 15:35
  produced
• 15:36
  a
• 15:36
  real
• 15:36
  hardware
• 15:36
  side
• 15:36
  effect
• 15:37
  the
• 15:37
  browser's
• 15:37
  mute
• 15:37
  control
• 15:38
  could
• 15:38
  not
• 15:38
  stop.
• 15:39
  The
• 15:39
  two
• 15:39
  narrow
• 15:39
  uBlock
• 15:40
  Origin
• 15:40
  rules
• 15:40
  that
• 15:41
  block
• 15:41
  it
• 15:41
  come
• 15:41
  with
• 15:41
  the
• 15:41
  honest
• 15:42
  warning
• 15:42
  that
• 15:42
  they
• 15:42
  may
• 15:42
  trigger
• 15:43
  extra
• 15:43
  CAPTCHAs
• 15:43
  at
• 15:44
  login
• 15:44
  or
• 15:44
  checkout.
• 15:45
  For
• 15:46
  contrast
• 15:46
  on
• 15:46
  the
• 15:46
  benign
• 15:47
  end
• 15:47
  of
• 15:47
  the
• 15:47
  same
• 15:47
  API
• 15:48
  surface,
• 15:49
  someone
• 15:49
  built
• 15:49
  a
• 15:49
  browser
• 15:50
  theremin
• 15:50
  you
• 15:50
  play
• 15:51
  by
• 15:51
  waving
• 15:51
  both
• 15:51
  hands
• 15:52
  at
• 15:52
  a
• 15:52
  webcam
• 15:52
  —
• 15:53
  spread
• 15:53
  for
• 15:53
  volume,
• 15:54
  raise
• 15:54
  for
• 15:55
  pitch,
• 15:55
  palms
• 15:56
  together
• 15:56
  for
• 15:56
  silence,
• 15:57
  same
• 15:57
  audio
• 15:58
  and
• 15:58
  sensor
• 15:58
  primitives
• 15:59
  turned
• 15:59
  toward
• 15:59
  something
• 15:59
  that
• 15:59
  asks
• 16:00
  permission
• 16:01
  and
• 16:01
  shows
• 16:01
  exactly
• 16:01
  what
• 16:02
  it
• 16:02
  is
• 16:02
  doing
• 16:02
  .
• 16:04
  Estonia,
• 16:05
  the
• 16:05
  country
• 16:06
  that
• 16:06
  pioneered
• 16:06
  e-residency
• 16:07
  and
• 16:07
  digital
• 16:08
  nomad
• 16:08
  visas,
• 16:09
  now
• 16:09
  wants
• 16:09
  to
• 16:09
  issue
• 16:10
  official
• 16:10
  "AI
• 16:10
  ID
• 16:11
  codes"
• 16:11
  that
• 16:11
  identify
• 16:12
  an
• 16:12
  agent
• 16:13
  separately
• 16:13
  from
• 16:13
  the
• 16:13
  person
• 16:14
  or
• 16:14
  company
• 16:14
  it
• 16:15
  represents
• 16:15
  .
• 16:16
  The
• 16:16
  interesting
• 16:17
  part
• 16:17
  is
• 16:17
  the
• 16:17
  permission
• 16:17
  model:
• 16:18
  instead
• 16:18
  of
• 16:18
  full
• 16:19
  account
• 16:19
  access,
• 16:20
  you
• 16:20
  could
• 16:20
  authorize
• 16:21
  an
• 16:21
  assistant
• 16:21
  to
• 16:21
  view
• 16:21
  data,
• 16:22
  prepare
• 16:22
  a
• 16:22
  document,
• 16:23
  or
• 16:24
  spend
• 16:24
  only
• 16:24
  within
• 16:24
  a
• 16:24
  fixed
• 16:25
  limit,
• 16:25
  with
• 16:25
  every
• 16:26
  action
• 16:26
  verifiable
• 16:27
  and
• 16:27
  auditable.
• 16:28
  This
• 16:28
  is
• 16:28
  the
• 16:29
  same
• 16:29
  problem
• 16:29
  the
• 16:29
  industry
• 16:30
  circles
• 16:30
  from
• 16:30
  the
• 16:30
  other
• 16:30
  direction
• 16:31
  —
• 16:31
  an
• 16:31
  IETF
• 16:32
  draft
• 16:32
  treating
• 16:32
  agents
• 16:33
  as
• 16:33
  workloads
• 16:33
  under
• 16:34
  existing
• 16:34
  identity
• 16:34
  frameworks,
• 16:35
  startups
• 16:35
  like
• 16:36
  Oasis
• 16:36
  raising
• 16:36
  nine
• 16:37
  figures
• 16:37
  to
• 16:37
  enforce
• 16:37
  least-privilege,
• 16:38
  and
• 16:38
  the
• 16:38
  cautionary
• 16:39
  tale
• 16:39
  of
• 16:39
  an
• 16:40
  autonomous
• 16:40
  agent
• 16:40
  that
• 16:40
  deleted
• 16:41
  200
• 16:41
  customer
• 16:42
  records
• 16:42
  because
• 16:42
  it
• 16:42
  inherited
• 16:43
  never-revoked
• 16:43
  human
• 16:44
  permissions.
• 16:45
  Estonia's
• 16:45
  contribution
• 16:46
  is
• 16:46
  to
• 16:46
  propose
• 16:46
  the
• 16:47
  state
• 16:47
  as
• 16:47
  the
• 16:47
  identity
• 16:47
  anchor
• 16:48
  rather
• 16:48
  than
• 16:48
  a
• 16:48
  vendor
• 16:48
  —
• 16:49
  either
• 16:49
  the
• 16:50
  missing
• 16:50
  trust
• 16:50
  primitive
• 16:51
  or,
• 16:51
  as
• 16:51
  the
• 16:51
  thread's
• 16:51
  skeptics
• 16:52
  put
• 16:52
  it,
• 16:52
  another
• 16:53
  bureaucratic
• 16:53
  layer
• 16:54
  that
• 16:54
  dodges
• 16:54
  whether
• 16:54
  you
• 16:54
  should
• 16:54
  trust
• 16:55
  the
• 16:55
  agent
• 16:55
  at
• 16:55
  all.
• 16:56
  For
• 16:56
  a
• 16:56
  fraud
• 16:56
  practitioner
• 16:57
  the
• 16:57
  appeal
• 16:57
  is
• 16:57
  obvious:
• 16:58
  scoped,
• 16:58
  revocable,
• 16:59
  audited
• 17:00
  credentials
• 17:00
  with
• 17:00
  a
• 17:00
  tamper-evident
• 17:01
  log
• 17:01
  are
• 17:01
  what
• 17:01
  "least
• 17:02
  privilege
• 17:02
  for
• 17:02
  machines"
• 17:03
  has
• 17:03
  needed
• 17:03
  —
• 17:03
  a
• 17:04
  serious
• 17:04
  attempt
• 17:04
  at
• 17:04
  the
• 17:04
  attribution
• 17:05
  problem
• 17:05
  behavioral
• 17:06
  fingerprinting
• 17:06
  alone
• 17:07
  cannot
• 17:07
  solve.
• 17:08
  The
• 17:08
  money
• 17:08
  story
• 17:09
  is
• 17:09
  louder
• 17:09
  and
• 17:09
  stranger.
• 17:11
  Stripe
• 17:11
  told
• 17:11
  investors,
• 17:12
  per
• 17:12
  a
• 17:12
  scooped
• 17:12
  letter,
• 17:13
  that
• 17:13
  "the
• 17:13
  singularity"
• 17:14
  began
• 17:14
  and
• 17:15
  that
• 17:15
  the
• 17:15
  threshold
• 17:15
  was
• 17:15
  crossed
• 17:16
  on
• 17:16
  January
• 17:16
  1,
• 17:17
  2026
• 17:17
  —
• 17:19
  alongside
• 17:19
  first-half
• 17:20
  revenue
• 17:20
  up
• 17:20
  41%
• 17:21
  year
• 17:21
  over
• 17:21
  year,
• 17:22
  free
• 17:22
  cash
• 17:22
  flow
• 17:22
  up
• 17:23
  43%,
• 17:24
  and
• 17:24
  the
• 17:24
  disclosure
• 17:24
  that
• 17:25
  88%
• 17:25
  of
• 17:25
  the
• 17:25
  Forbes
• 17:26
  AI
• 17:26
  50
• 17:26
  use
• 17:26
  its
• 17:27
  platform.
• 17:28
  The
• 17:28
  letter
• 17:28
  also
• 17:28
  confirmed
• 17:29
  the
• 17:29
  OpenRouter
• 17:29
  acquisition,
• 17:30
  the
• 17:30
  AI-task
• 17:31
  routing
• 17:31
  layer,
• 17:31
  in
• 17:32
  a
• 17:32
  deal
• 17:32
  worth
• 17:32
  over
• 17:32
  $7
• 17:33
  billion
• 17:33
  in
• 17:33
  cash
• 17:33
  and
• 17:34
  stock.
• 17:34
  Strip
• 17:35
  the
• 17:35
  eschatology
• 17:35
  and
• 17:36
  the
• 17:36
  logic
• 17:36
  is
• 17:36
  coherent:
• 17:37
  OpenRouter
• 17:37
  is
• 17:38
  a
• 17:38
  neutral
• 17:38
  routing-and-billing
• 17:39
  layer
• 17:39
  across
• 17:39
  400-plus
• 17:40
  models
• 17:40
  serving
• 17:40
  a
• 17:40
  million
• 17:41
  developers
• 17:41
  —
• 17:42
  exactly
• 17:42
  the
• 17:42
  toll
• 17:42
  booth
• 17:43
  a
• 17:43
  payments
• 17:43
  company
• 17:43
  wants
• 17:43
  to
• 17:43
  own.
• 17:44
  The
• 17:44
  "singularity
• 17:45
  began
• 17:45
  January
• 17:46
  1"
• 17:46
  framing
• 17:46
  is
• 17:47
  the
• 17:47
  genuinely
• 17:47
  new
• 17:47
  element,
• 17:48
  and
• 17:48
  it
• 17:48
  belongs
• 17:49
  with
• 17:49
  the
• 17:49
  past
• 17:49
  year's
• 17:49
  executive-rhetoric
• 17:50
  inflation
• 17:50
  —
• 17:51
  the
• 17:51
  "foothills
• 17:52
  of
• 17:52
  the
• 17:52
  singularity"
• 17:52
  language
• 17:53
  from
• 17:53
  other
• 17:53
  lab
• 17:53
  leaders
• 17:53
  —
• 17:54
  and
• 17:54
  reactions
• 17:55
  on
• 17:55
  X
• 17:55
  that
• 17:55
  ranged
• 17:55
  from
• 17:56
  enthusiasm
• 17:56
  to
• 17:56
  a
• 17:57
  blunt
• 17:57
  "Collison/stripe
• 17:58
  has
• 17:59
  AI
• 17:59
  psychosis."
• 18:00
  A
• 18:00
  payments
• 18:01
  giant
• 18:01
  declaring
• 18:01
  a
• 18:01
  metaphysical
• 18:02
  threshold
• 18:02
  in
• 18:03
  an
• 18:03
  investor
• 18:03
  letter
• 18:03
  is
• 18:04
  a
• 18:04
  marketing
• 18:04
  decision
• 18:04
  with
• 18:04
  a
• 18:05
  revenue
• 18:05
  chart
• 18:05
  attached:
• 18:06
  the
• 18:06
  41%
• 18:07
  growth
• 18:08
  is
• 18:08
  the
• 18:08
  fact,
• 18:08
  the
• 18:09
  singularity
• 18:09
  the
• 18:09
  framing
• 18:10
  around
• 18:10
  it
• 18:10
  .