Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – August 05, 2026

Transcript

AGI Dreams Podcast – August 05, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  DSpark
• 0:00
  speculative
• 0:01
  decoding
• 0:01
  already
• 0:02
  doubled
• 0:02
  throughput
• 0:02
  for
• 0:02
  DeepSeek
• 0:03
  V4
• 0:03
  Flash
• 0:03
  on
• 0:04
  multi-GPU
• 0:04
  rigs.
• 0:05
  The
• 0:05
  interesting
• 0:06
  question
• 0:06
  was
• 0:06
  always
• 0:07
  whether
• 0:07
  someone
• 0:07
  would
• 0:07
  make
• 0:07
  it
• 0:08
  smart
• 0:08
  enough
• 0:08
  to
• 0:08
  adapt
• 0:08
  to
• 0:09
  what
• 0:09
  the
• 0:09
  model
• 0:09
  is
• 0:09
  actually
• 0:09
  doing.
• 0:10
  That
• 0:10
  someone
• 0:11
  turned
• 0:11
  out
• 0:11
  to
• 0:11
  be
• 0:11
  a
• 0:11
  single
• 0:12
  developer
• 0:12
  running
• 0:12
  an
• 0:12
  RTX
• 0:13
  5090
• 0:14
  with
• 0:14
  256
• 0:15
  GB
• 0:15
  of
• 0:15
  system
• 0:16
  RAM,
• 0:16
  and
• 0:16
  the
• 0:16
  answer
• 0:17
  is
• 0:17
  phase-adaptive
• 0:17
  K
• 0:18
  with
• 0:18
  dual
• 0:18
  CUDA
• 0:18
  graphs.
• 0:19
  The
• 0:19
  core
• 0:20
  insight:
• 0:20
  during
• 0:21
  hard
• 0:21
  reasoning,
• 0:21
  the
• 0:22
  second
• 0:22
  speculative
• 0:22
  draft
• 0:23
  position
• 0:23
  gets
• 0:23
  accepted
• 0:24
  roughly
• 0:24
  30-50%
• 0:26
  of
• 0:26
  the
• 0:27
  time,
• 0:27
  making
• 0:27
  it
• 0:27
  a
• 0:27
  net
• 0:28
  drag.
• 0:29
  During
• 0:29
  predictable
• 0:29
  code
• 0:30
  or
• 0:30
  text
• 0:30
  output,
• 0:31
  acceptance
• 0:31
  jumps
• 0:31
  to
• 0:31
  80-90%,
• 0:34
  making
• 0:34
  depth-2
• 0:35
  drafting
• 0:35
  a
• 0:35
  clear
• 0:35
  win.
• 0:36
  The
• 0:36
  patch
• 0:37
  detects
• 0:37
  the
• 0:37
  token
• 0:38
  in
• 0:38
  generated
• 0:38
  output,
• 0:39
  switches
• 0:39
  from
• 0:39
  K=1
• 0:40
  to
• 0:40
  K=2,
• 0:41
  and
• 0:42
  --
• 0:42
  critically
• 0:43
  --
• 0:44
  keeps
• 0:44
  both
• 0:45
  CUDA
• 0:45
  graph
• 0:45
  shapes
• 0:45
  pre-captured
• 0:46
  so
• 0:46
  neither
• 0:46
  phase
• 0:47
  falls
• 0:47
  back
• 0:47
  to
• 0:47
  eager
• 0:48
  execution.
• 0:49
  Measured
• 0:49
  across
• 0:49
  a
• 0:49
  20-minute
• 0:50
  agentic
• 0:50
  coding
• 0:51
  session:
• 0:51
  13.8
• 0:52
  tok/s
• 0:53
  mean
• 0:53
  during
• 0:53
  reasoning,
• 0:54
  17.0
• 0:55
  tok/s
• 0:56
  during
• 0:56
  content
• 0:57
  generation,
• 0:58
  with
• 0:58
  the
• 0:58
  full
• 0:58
  1M-token
• 0:59
  context
• 0:59
  preserved
• 1:00
  on
• 1:00
  15.9
• 1:01
  GiB
• 1:01
  of
• 1:01
  model
• 1:01
  allocation.
• 1:03
  An
• 1:03
  update
• 1:03
  validated
• 1:04
  the
• 1:04
  approach
• 1:04
  through
• 1:04
  500,000
• 1:05
  tokens
• 1:06
  of
• 1:06
  context
• 1:06
  with
• 1:07
  a
• 1:07
  workaround
• 1:07
  for
• 1:07
  a
• 1:07
  separate
• 1:08
  SM120
• 1:09
  Triton
• 1:10
  MQA
• 1:10
  kernel
• 1:11
  bug
• 1:11
  that
• 1:11
  crashed
• 1:12
  on
• 1:12
  long-KV
• 1:12
  prefills
• 1:13
  above
• 1:13
  16K
• 1:14
  entries.
• 1:15
  The
• 1:15
  practical
• 1:15
  gain
• 1:16
  over
• 1:16
  fixed-K
• 1:16
  configurations
• 1:17
  is
• 1:18
  mid-single-digit
• 1:18
  percent
• 1:19
  on
• 1:19
  reasoning-heavy
• 1:20
  workloads
• 1:20
  --
• 1:21
  not
• 1:21
  a
• 1:21
  miracle,
• 1:22
  but
• 1:22
  the
• 1:22
  developer
• 1:23
  no
• 1:23
  longer
• 1:23
  has
• 1:23
  to
• 1:23
  pick
• 1:24
  a
• 1:24
  single
• 1:24
  compromise
• 1:25
  speculative
• 1:25
  depth
• 1:25
  for
• 1:25
  the
• 1:26
  entire
• 1:26
  completion.
• 1:27
  Meanwhile,
• 1:28
  TensorSharp
• 1:29
  shipped
• 1:29
  its
• 1:29
  own
• 1:29
  DSpark
• 1:30
  implementation
• 1:31
  for
• 1:31
  the
• 1:31
  same
• 1:31
  model
• 1:31
  on
• 1:32
  4x
• 1:32
  NVIDIA
• 1:33
  A40s,
• 1:34
  measuring
• 1:34
  1.5x
• 1:35
  to
• 1:35
  2x
• 1:36
  speedups
• 1:36
  depending
• 1:36
  on
• 1:37
  workload:
• 1:38
  a
• 1:38
  10K-token
• 1:38
  document
• 1:39
  query
• 1:39
  jumped
• 1:39
  from
• 1:40
  25.3
• 1:41
  to
• 1:41
  51.3
• 1:42
  tok/s
• 1:43
  with
• 1:43
  85%
• 1:44
  draft
• 1:44
  acceptance,
• 1:45
  while
• 1:45
  longer
• 1:46
  generations
• 1:46
  settled
• 1:47
  around
• 1:47
  1.53x.
• 1:49
  For
• 1:49
  people
• 1:49
  without
• 1:50
  multi-GPU
• 1:50
  setups
• 1:51
  or
• 1:51
  bleeding-edge
• 1:51
  5090s,
• 1:53
  a
• 1:53
  detailed
• 1:53
  configuration
• 1:54
  report
• 1:54
  on
• 1:54
  a
• 1:54
  single
• 1:55
  RTX
• 1:55
  4090
• 1:57
  with
• 1:57
  128
• 1:58
  GB
• 1:58
  DDR5
• 1:59
  establishes
• 1:59
  the
• 1:59
  floor:
• 2:00
  12-13
• 2:01
  tok/s
• 2:02
  generation
• 2:03
  at
• 2:03
  64K
• 2:03
  context
• 2:04
  using
• 2:04
  the
• 2:05
  UD-Q3
• 2:05
  K
• 2:05
  M
• 2:06
  quant,
• 2:07
  with
• 2:07
  prompt
• 2:07
  processing
• 2:08
  at
• 2:08
  212-224
• 2:09
  tok/s
• 2:10
  for
• 2:11
  8K-32K
• 2:12
  inputs.
• 2:14
  The
• 2:14
  critical
• 2:14
  finding
• 2:15
  is
• 2:15
  that
• 2:15
  128
• 2:16
  GB
• 2:16
  of
• 2:16
  system
• 2:16
  RAM
• 2:17
  is
• 2:17
  a
• 2:17
  hard
• 2:17
  requirement,
• 2:18
  not
• 2:18
  headroom
• 2:19
  --
• 2:20
  the
• 2:20
  model
• 2:20
  loaded
• 2:20
  and
• 2:20
  serving
• 2:21
  consumed
• 2:21
  103
• 2:22
  GB
• 2:23
  resident.
• 2:23
  The
• 2:24
  author
• 2:24
  noted
• 2:24
  that
• 2:24
  agentic
• 2:25
  workloads
• 2:25
  spent
• 2:26
  81-86%
• 2:28
  of
• 2:28
  wall
• 2:28
  time
• 2:29
  on
• 2:29
  prompt
• 2:29
  reprocessing,
• 2:30
  meaning
• 2:30
  the
• 2:30
  first
• 2:31
  token
• 2:31
  after
• 2:31
  a
• 2:31
  long
• 2:32
  context
• 2:32
  can
• 2:32
  take
• 2:33
  two
• 2:33
  and
• 2:33
  a
• 2:33
  half
• 2:33
  minutes.
• 2:34
  An
• 2:34
  agentic
• 2:35
  coding
• 2:35
  benchmark
• 2:35
  pitting
• 2:36
  V4
• 2:36
  Flash
• 2:36
  against
• 2:37
  Qwen3.6-27B,
• 2:41
  Qwen3.5-122B,
• 2:45
  and
• 2:45
  Gemma
• 2:45
  4
• 2:46
  31B
• 2:46
  on
• 2:46
  an
• 2:46
  M5
• 2:47
  Max
• 2:47
  confirmed
• 2:47
  that
• 2:48
  even
• 2:48
  heavily
• 2:48
  quantized
• 2:49
  V4
• 2:49
  Flash
• 2:49
  remains
• 2:50
  the
• 2:50
  smartest
• 2:50
  model
• 2:50
  people
• 2:51
  can
• 2:51
  run
• 2:51
  locally
• 2:51
  on
• 2:52
  128
• 2:52
  GB,
• 2:53
  though
• 2:53
  it
• 2:53
  outspent
• 2:54
  the
• 2:54
  next-best
• 2:54
  122B
• 2:55
  model
• 2:56
  by
• 2:56
  over
• 2:56
  5x
• 2:57
  in
• 2:57
  tokens
• 2:57
  to
• 2:57
  get
• 2:57
  there
• 2:58
  --
• 2:58
  and
• 2:58
  the
• 2:58
  Qwen
• 2:59
  models
• 2:59
  solved
• 2:59
  substantially
• 3:00
  more
• 3:00
  cases
• 3:01
  on
• 3:01
  their
• 3:01
  first
• 3:01
  attempt.
• 3:07
  DeepGrove,
• 3:08
  a
• 3:08
  startup
• 3:09
  focused
• 3:09
  on
• 3:10
  natively
• 3:10
  trained
• 3:10
  low-precision
• 3:11
  architectures,
• 3:12
  dropped
• 3:12
  Maple-Preview:
• 3:13
  a
• 3:13
  20-billion-parameter,
• 3:14
  1-billion-active
• 3:15
  ternary-weight
• 3:16
  reasoning
• 3:16
  MoE
• 3:17
  that
• 3:17
  runs
• 3:17
  at
• 3:17
  120+
• 3:18
  tok/s
• 3:19
  on
• 3:19
  an
• 3:19
  iPhone
• 3:20
  and
• 3:20
  200+
• 3:21
  tok/s
• 3:21
  on
• 3:22
  a
• 3:22
  Mac
• 3:22
  mini
• 3:22
  M4.
• 3:23
  The
• 3:23
  claim
• 3:23
  that
• 3:24
  distinguishes
• 3:24
  it
• 3:24
  from
• 3:24
  post-training
• 3:25
  quantization
• 3:25
  approaches
• 3:26
  like
• 3:26
  PrismML's
• 3:26
  Bonsai
• 3:27
  is
• 3:27
  that
• 3:27
  Maple
• 3:27
  was
• 3:28
  trained
• 3:28
  from
• 3:28
  scratch
• 3:28
  in
• 3:28
  ternary
• 3:29
  precision,
• 3:29
  with
• 3:30
  architecture
• 3:30
  decisions
• 3:30
  made
• 3:31
  in
• 3:31
  a
• 3:31
  hardware-aware
• 3:31
  design
• 3:32
  loop
• 3:32
  testing
• 3:32
  configurations
• 3:33
  directly
• 3:33
  on
• 3:33
  Apple
• 3:34
  Silicon
• 3:34
  for
• 3:34
  inference
• 3:34
  speed.
• 3:35
  At
• 3:35
  ternary
• 3:36
  weights,
• 3:36
  matrix
• 3:37
  multiplication
• 3:37
  effectively
• 3:38
  reduces
• 3:38
  to
• 3:39
  additions,
• 3:40
  cutting
• 3:40
  both
• 3:40
  memory
• 3:40
  and
• 3:41
  arithmetic
• 3:41
  requirements.
• 3:42
  The
• 3:42
  benchmarks
• 3:43
  are
• 3:43
  competitive
• 3:43
  with
• 3:43
  larger
• 3:44
  models
• 3:44
  --
• 3:45
  the
• 3:45
  team
• 3:45
  claims
• 3:45
  IMO-level
• 3:46
  problem
• 3:46
  solving
• 3:47
  --
• 3:47
  though
• 3:47
  the
• 3:47
  release
• 3:48
  explicitly
• 3:48
  notes
• 3:49
  minimal
• 3:49
  post-training
• 3:49
  for
• 3:50
  agentic
• 3:50
  domains
• 3:50
  and
• 3:51
  only
• 3:51
  small-scale
• 3:51
  reinforcement
• 3:52
  learning,
• 3:53
  so
• 3:53
  the
• 3:53
  practical
• 3:53
  ceiling
• 3:54
  remains
• 3:54
  unclear.
• 3:55
  The
• 3:55
  more
• 3:55
  provocative
• 3:56
  feature
• 3:56
  is
• 3:56
  on-device
• 3:57
  adaptation
• 3:57
  through
• 3:57
  what
• 3:57
  DeepGrove
• 3:58
  calls
• 3:58
  "dreaming."
• 3:59
  In
• 3:59
  a
• 3:59
  demo,
• 4:00
  the
• 4:00
  model
• 4:00
  encounters
• 4:01
  a
• 4:01
  user's
• 4:01
  vegan
• 4:01
  dietary
• 4:02
  preference
• 4:02
  during
• 4:03
  a
• 4:03
  recipe
• 4:03
  conversation,
• 4:04
  generates
• 4:05
  a
• 4:05
  small
• 4:05
  synthetic
• 4:05
  dataset
• 4:06
  overnight,
• 4:07
  fine-tunes
• 4:07
  itself,
• 4:08
  and
• 4:08
  the
• 4:08
  next
• 4:08
  day
• 4:09
  correctly
• 4:09
  recommends
• 4:09
  synthetic
• 4:10
  leather
• 4:10
  bags
• 4:11
  instead
• 4:11
  of
• 4:11
  real
• 4:11
  leather
• 4:11
  --
• 4:12
  a
• 4:12
  generalization
• 4:13
  that
• 4:13
  Claude
• 4:13
  Sonnet
• 4:13
  5
• 4:14
  with
• 4:14
  memory
• 4:14
  enabled
• 4:15
  reportedly
• 4:15
  failed
• 4:15
  to
• 4:15
  make.
• 4:16
  The
• 4:16
  dreaming
• 4:16
  process
• 4:17
  takes
• 4:17
  10-20
• 4:18
  minutes
• 4:18
  with
• 4:18
  5.9
• 4:19
  GB
• 4:19
  peak
• 4:19
  memory.
• 4:20
  Whether
• 4:20
  this
• 4:20
  scales
• 4:21
  beyond
• 4:21
  curated
• 4:21
  demos
• 4:22
  is
• 4:22
  an
• 4:22
  open
• 4:22
  question,
• 4:23
  but
• 4:23
  the
• 4:23
  thesis
• 4:24
  --
• 4:24
  that
• 4:24
  always-active
• 4:25
  on-device
• 4:25
  assistants
• 4:26
  should
• 4:26
  adapt
• 4:27
  their
• 4:27
  weights
• 4:27
  rather
• 4:27
  than
• 4:27
  accumulate
• 4:28
  context
• 4:28
  files
• 4:28
  --
• 4:29
  is
• 4:29
  directionally
• 4:30
  interesting.
• 4:31
  Qwen's
• 4:31
  developers
• 4:32
  held
• 4:32
  a
• 4:32
  Twitter
• 4:32
  AMA
• 4:33
  confirming
• 4:33
  that
• 4:33
  Qwen
• 4:34
  3.8-27B
• 4:37
  is
• 4:37
  imminent,
• 4:37
  with
• 4:37
  2.4
• 4:38
  trillion
• 4:38
  total
• 4:39
  parameters
• 4:39
  and
• 4:39
  95
• 4:40
  billion
• 4:40
  active.
• 4:41
  The
• 4:41
  architecture
• 4:41
  is
• 4:41
  "similar
• 4:42
  to
• 4:42
  3.5
• 4:43
  but
• 4:43
  much
• 4:43
  larger-scale,"
• 4:44
  and
• 4:44
  the
• 4:44
  team
• 4:44
  described
• 4:45
  the
• 4:45
  capability
• 4:45
  jump
• 4:45
  from
• 4:45
  3.6-27B
• 4:48
  as
• 4:48
  "pretty
• 4:48
  huge"
• 4:48
  --
• 4:49
  twice.
• 4:50
  They
• 4:50
  confirmed
• 4:50
  3.8
• 4:51
  supports
• 4:51
  different
• 4:51
  reasoning
• 4:52
  effort
• 4:52
  levels,
• 4:53
  addressing
• 4:53
  a
• 4:53
  known
• 4:53
  complaint
• 4:54
  about
• 4:54
  excessive
• 4:54
  token
• 4:55
  consumption
• 4:55
  in
• 4:55
  thinking
• 4:55
  mode.
• 4:56
  On
• 4:56
  quantization,
• 4:57
  they
• 4:57
  recommended
• 4:58
  QAT
• 4:58
  or
• 4:59
  quantizing
• 4:59
  only
• 4:59
  FFN
• 5:00
  layers
• 5:00
  to
• 5:00
  4-bit
• 5:01
  while
• 5:01
  keeping
• 5:01
  attention
• 5:02
  QKV
• 5:02
  projections
• 5:03
  in
• 5:03
  16-bit.
• 5:04
  The
• 5:05
  AMA
• 5:05
  was
• 5:05
  notably
• 5:06
  evasive
• 5:06
  on
• 5:06
  smaller
• 5:07
  models,
• 5:07
  70B
• 5:08
  variants,
• 5:08
  and
• 5:08
  the
• 5:08
  122B
• 5:09
  segment,
• 5:10
  deflecting
• 5:10
  with
• 5:11
  variations
• 5:11
  of
• 5:11
  "collecting
• 5:12
  everyone's
• 5:12
  requests."
• 5:14
  Community
• 5:14
  reaction
• 5:14
  to
• 5:14
  the
• 5:15
  AMA
• 5:15
  was
• 5:15
  split
• 5:15
  between
• 5:16
  gratitude
• 5:16
  for
• 5:16
  any
• 5:17
  engagement
• 5:17
  and
• 5:17
  frustration
• 5:18
  at
• 5:18
  the
• 5:18
  ratio
• 5:18
  of
• 5:18
  substantive
• 5:19
  answers
• 5:19
  to
• 5:19
  corporate
• 5:20
  hedging.
• 5:20
  AI9Stars,
• 5:22
  a
• 5:22
  group
• 5:22
  affiliated
• 5:23
  with
• 5:23
  Tsinghua
• 5:23
  University,
• 5:24
  released
• 5:24
  G9v3-39A5B
• 5:28
  under
• 5:28
  Apache
• 5:28
  2.0
• 5:29
  --
• 5:30
  a
• 5:30
  39-billion-parameter
• 5:31
  MoE
• 5:31
  with
• 5:31
  5
• 5:32
  billion
• 5:32
  active
• 5:32
  experts
• 5:33
  targeting
• 5:33
  coding,
• 5:34
  tool
• 5:34
  use,
• 5:34
  and
• 5:34
  reasoning.
• 5:36
  Benchmark
• 5:36
  comparisons
• 5:37
  on
• 5:37
  Artificial
• 5:37
  Analysis
• 5:38
  put
• 5:38
  it
• 5:38
  roughly
• 5:39
  on
• 5:39
  par
• 5:39
  with
• 5:39
  Qwen
• 5:39
  3.6-35B-A3B
• 5:43
  despite
• 5:44
  having
• 5:44
  both
• 5:44
  larger
• 5:45
  total
• 5:45
  and
• 5:45
  active
• 5:45
  parameter
• 5:46
  counts,
• 5:47
  suggesting
• 5:47
  the
• 5:47
  full
• 5:47
  release
• 5:48
  after
• 5:48
  post-training
• 5:49
  may
• 5:49
  be
• 5:49
  where
• 5:49
  it
• 5:49
  gets
• 5:50
  interesting.
• 5:51
  Unsloth
• 5:51
  published
• 5:52
  detailed
• 5:52
  documentation
• 5:53
  for
• 5:53
  running
• 5:53
  Kimi
• 5:54
  K3,
• 5:55
  Moonshot
• 5:55
  AI's
• 5:55
  2.8-trillion-parameter
• 5:57
  model
• 5:57
  (104B
• 5:58
  active),
• 5:59
  in
• 5:59
  quantized
• 6:00
  form
• 6:00
  via
• 6:00
  llama.cpp,
• 6:03
  reporting
• 6:03
  that
• 6:03
  their
• 6:03
  dynamic
• 6:04
  1-bit
• 6:04
  quant
• 6:04
  at
• 6:05
  553
• 6:06
  GB
• 6:06
  reaches
• 6:07
  79%
• 6:07
  top-1
• 6:08
  accuracy
• 6:09
  with
• 6:09
  2.58
• 6:09
  perplexity
• 6:10
  --
• 6:11
  while
• 6:11
  a
• 6:11
  competing
• 6:11
  community
• 6:12
  quant
• 6:12
  at
• 6:13
  619
• 6:14
  GB
• 6:14
  delivered
• 6:14
  54.56
• 6:16
  perplexity,
• 6:16
  21x
• 6:17
  worse.
• 6:18
  The
• 6:18
  practical
• 6:19
  takeaway:
• 6:20
  quantization
• 6:20
  method
• 6:20
  matters
• 6:21
  enormously
• 6:21
  at
• 6:22
  this
• 6:22
  scale,
• 6:22
  and
• 6:23
  naive
• 6:23
  approaches
• 6:23
  destroy
• 6:24
  the
• 6:24
  model.
• 6:25
  Separately,
• 6:26
  tests
• 6:26
  of
• 6:26
  GLM-5.2
• 6:28
  on
• 6:28
  synthetic
• 6:28
  lethal
• 6:29
  gene
• 6:29
  identification
• 6:30
  in
• 6:30
  cancer
• 6:30
  data
• 6:31
  showed
• 6:31
  clear
• 6:31
  performance
• 6:32
  gains
• 6:32
  scaling
• 6:33
  with
• 6:33
  parameter
• 6:33
  count,
• 6:34
  though
• 6:35
  the
• 6:35
  researcher
• 6:35
  noted
• 6:35
  uncertainty
• 6:36
  about
• 6:36
  whether
• 6:37
  this
• 6:37
  reflects
• 6:37
  better
• 6:37
  reasoning
• 6:38
  or
• 6:38
  simply
• 6:39
  more
• 6:39
  memorized
• 6:39
  biological
• 6:40
  knowledge
• 6:40
  from
• 6:41
  training
• 6:41
  data.
• 6:45
  A
• 6:45
  research
• 6:46
  team
• 6:46
  at
• 6:46
  University
• 6:46
  of
• 6:47
  the
• 6:47
  Bundeswehr
• 6:47
  Munich
• 6:47
  identified
• 6:48
  a
• 6:48
  previously
• 6:48
  overlooked
• 6:49
  attack
• 6:49
  surface
• 6:49
  in
• 6:50
  vision-language
• 6:50
  models:
• 6:51
  the
• 6:51
  bottom
• 6:51
  singular-vector
• 6:52
  subspaces
• 6:53
  of
• 6:53
  intermediate
• 6:53
  linear
• 6:53
  transformations.
• 6:55
  Their
• 6:55
  Spectral
• 6:55
  Subspace
• 6:56
  Guided
• 6:56
  Representation
• 6:57
  Attack
• 6:57
  (SSGRA)
• 6:58
  aligns
• 6:59
  adversarial
• 6:59
  intermediate
• 7:00
  representations
• 7:01
  with
• 7:01
  these
• 7:01
  information-attenuating
• 7:02
  subspaces,
• 7:03
  where
• 7:03
  near-null
• 7:03
  singular
• 7:04
  values
• 7:04
  cause
• 7:04
  the
• 7:04
  strongest
• 7:05
  signal
• 7:05
  degradation.
• 7:06
  Tested
• 7:07
  against
• 7:07
  Gemma-3
• 7:08
  (4B),
• 7:09
  Qwen2.5-VL
• 7:10
  (7B),
• 7:11
  and
• 7:12
  LLaVA-1.5
• 7:12
  (7B),
• 7:13
  SSGRA
• 7:14
  consistently
• 7:14
  outperformed
• 7:15
  six
• 7:15
  existing
• 7:16
  baselines
• 7:16
  including
• 7:17
  feature-discrepancy
• 7:17
  and
• 7:18
  entropy-guided
• 7:18
  attacks.
• 7:19
  The
• 7:19
  telling
• 7:20
  structural
• 7:20
  finding:
• 7:21
  across
• 7:21
  all
• 7:21
  three
• 7:22
  models,
• 7:22
  near-null
• 7:23
  singular
• 7:23
  directions
• 7:24
  occurred
• 7:24
  far
• 7:24
  more
• 7:24
  frequently
• 7:25
  than
• 7:25
  strongly
• 7:25
  amplifying
• 7:26
  ones
• 7:26
  --
• 7:27
  in
• 7:27
  Qwen2.5-VL,
• 7:30
  singular
• 7:31
  values
• 7:31
  below
• 7:31
  0.01
• 7:32
  appeared
• 7:33
  in
• 7:33
  a
• 7:33
  large
• 7:33
  fraction
• 7:33
  of
• 7:34
  operators,
• 7:34
  while
• 7:34
  values
• 7:35
  above
• 7:35
  10
• 7:35
  were
• 7:35
  nearly
• 7:35
  absent.
• 7:37
  This
• 7:37
  spectral
• 7:37
  imbalance
• 7:38
  creates
• 7:38
  a
• 7:38
  large
• 7:38
  attack
• 7:38
  surface
• 7:39
  that
• 7:39
  untargeted
• 7:39
  adversarial
• 7:40
  optimization
• 7:41
  naturally
• 7:41
  exploits
• 7:42
  even
• 7:42
  without
• 7:42
  explicit
• 7:43
  spectral
• 7:43
  guidance.
• 7:44
  The
• 7:44
  practical
• 7:45
  implication
• 7:45
  is
• 7:45
  that
• 7:46
  spectral-norm
• 7:46
  regularization
• 7:47
  of
• 7:47
  large
• 7:47
  singular
• 7:48
  values,
• 7:48
  the
• 7:49
  standard
• 7:49
  defense,
• 7:50
  addresses
• 7:50
  only
• 7:50
  half
• 7:51
  the
• 7:51
  problem;
• 7:52
  controlling
• 7:52
  near-null
• 7:53
  directions
• 7:53
  may
• 7:53
  be
• 7:53
  equally
• 7:54
  important.
• 7:55
  On
• 7:55
  the
• 7:55
  defense
• 7:55
  side,
• 7:56
  Tantalus.io,
• 7:57
  built
• 7:58
  by
• 7:58
  Vince
• 7:58
  Ovando,
• 7:59
  launched
• 7:59
  a
• 7:59
  free
• 7:59
  prompt
• 8:00
  injection
• 8:00
  arena
• 8:00
  that
• 8:01
  makes
• 8:01
  the
• 8:01
  structural
• 8:01
  argument
• 8:02
  visceral.
• 8:03
  Round
• 8:03
  1
• 8:03
  gives
• 8:03
  attackers
• 8:04
  the
• 8:04
  full
• 8:04
  exploit
• 8:05
  playbook
• 8:05
  --
• 8:06
  poisoned
• 8:06
  marketplace
• 8:07
  skills,
• 8:07
  phishing
• 8:08
  emails,
• 8:09
  exposed
• 8:09
  credentials
• 8:09
  --
• 8:10
  against
• 8:11
  every
• 8:11
  behavioral
• 8:11
  defense
• 8:12
  the
• 8:12
  industry
• 8:12
  sells:
• 8:13
  defensive
• 8:13
  system
• 8:14
  prompts,
• 8:14
  input
• 8:14
  classifiers,
• 8:15
  output
• 8:16
  filters.
• 8:17
  All
• 8:17
  bypassable.
• 8:18
  Round
• 8:19
  2
• 8:19
  removes
• 8:19
  every
• 8:19
  behavioral
• 8:20
  defense,
• 8:21
  rewrites
• 8:21
  the
• 8:21
  system
• 8:22
  prompt
• 8:22
  to
• 8:22
  demand
• 8:22
  unconditional
• 8:23
  compliance,
• 8:24
  and
• 8:24
  adds
• 8:24
  one
• 8:24
  structural
• 8:25
  control:
• 8:26
  grammar-constrained
• 8:27
  decoding
• 8:27
  that
• 8:27
  restricts
• 8:28
  tool-call
• 8:28
  parameters
• 8:29
  at
• 8:29
  the
• 8:29
  token
• 8:29
  generation
• 8:29
  level.
• 8:30
  If
• 8:31
  the
• 8:31
  URL
• 8:31
  enum
• 8:31
  does
• 8:32
  not
• 8:32
  contain
• 8:32
  the
• 8:32
  attacker's
• 8:33
  domain,
• 8:33
  the
• 8:33
  model
• 8:34
  physically
• 8:34
  cannot
• 8:34
  generate
• 8:35
  it.
• 8:35
  The
• 8:36
  same
• 8:36
  attacks
• 8:36
  that
• 8:36
  succeeded
• 8:37
  in
• 8:37
  Round
• 8:37
  1
• 8:37
  die
• 8:38
  in
• 8:38
  Round
• 8:38
  2
• 8:39
  despite
• 8:39
  the
• 8:39
  agent
• 8:39
  being
• 8:40
  explicitly
• 8:40
  instructed
• 8:41
  to
• 8:41
  be
• 8:41
  malicious.
• 8:42
  The
• 8:42
  analogy
• 8:43
  to
• 8:43
  parameterized
• 8:44
  queries
• 8:44
  versus
• 8:44
  SQL
• 8:45
  injection
• 8:45
  is
• 8:45
  apt
• 8:46
  --
• 8:46
  this
• 8:46
  is
• 8:46
  not
• 8:47
  a
• 8:47
  filter
• 8:47
  to
• 8:47
  bypass
• 8:48
  but
• 8:48
  a
• 8:48
  constraint
• 8:48
  on
• 8:49
  the
• 8:49
  output
• 8:49
  space
• 8:49
  itself.
• 8:51
  One
• 8:51
  caveat
• 8:51
  already
• 8:52
  surfacing
• 8:52
  in
• 8:52
  early
• 8:52
  red-teaming:
• 8:53
  the
• 8:54
  grammar
• 8:54
  constrains
• 8:54
  what
• 8:54
  the
• 8:55
  agent
• 8:55
  can
• 8:55
  DO,
• 8:56
  not
• 8:56
  what
• 8:56
  it
• 8:56
  can
• 8:56
  SAY
• 8:56
  --
• 8:58
  free-text
• 8:58
  responses
• 8:59
  can
• 8:59
  still
• 8:59
  carry
• 8:59
  secrets,
• 9:00
  a
• 9:00
  gap
• 9:01
  the
• 9:01
  authors
• 9:01
  themselves
• 9:02
  scope
• 9:02
  to
• 9:02
  a
• 9:02
  follow-up
• 9:02
  paper.
• 9:04
  Mistral
• 9:04
  released
• 9:04
  Shieldstral,
• 9:05
  a
• 9:06
  3-billion-parameter
• 9:07
  open-weights
• 9:07
  multimodal
• 9:08
  safety
• 9:08
  classifier
• 9:09
  under
• 9:09
  Apache
• 9:09
  2.0
• 9:11
  that
• 9:11
  takes
• 9:11
  a
• 9:11
  different
• 9:11
  approach
• 9:12
  to
• 9:12
  content
• 9:12
  moderation.
• 9:14
  Instead
• 9:14
  of
• 9:14
  baking
• 9:14
  a
• 9:14
  fixed
• 9:15
  harm
• 9:15
  taxonomy
• 9:16
  into
• 9:16
  the
• 9:16
  weights,
• 9:17
  Shieldstral
• 9:17
  treats
• 9:18
  moderation
• 9:18
  as
• 9:18
  natural
• 9:19
  language
• 9:19
  inference:
• 9:20
  you
• 9:20
  write
• 9:20
  the
• 9:21
  safety
• 9:21
  policy
• 9:21
  as
• 9:21
  a
• 9:21
  plain-language
• 9:22
  question
• 9:23
  at
• 9:23
  inference
• 9:23
  time,
• 9:24
  and
• 9:24
  the
• 9:24
  model
• 9:24
  returns
• 9:25
  a
• 9:25
  calibrated
• 9:25
  probability
• 9:26
  from
• 9:26
  a
• 9:26
  single
• 9:27
  forward
• 9:27
  pass.
• 9:28
  This
• 9:28
  means
• 9:29
  one
• 9:29
  checkpoint
• 9:29
  adapts
• 9:30
  to
• 9:30
  novel
• 9:30
  policies
• 9:31
  without
• 9:31
  retraining
• 9:31
  --
• 9:32
  the
• 9:32
  same
• 9:33
  content
• 9:33
  can
• 9:33
  be
• 9:33
  classified
• 9:34
  differently
• 9:35
  for
• 9:35
  a
• 9:35
  cybersecurity
• 9:36
  research
• 9:36
  tool
• 9:36
  versus
• 9:37
  a
• 9:37
  children's
• 9:37
  platform.
• 9:38
  Shieldstral
• 9:39
  matches
• 9:39
  or
• 9:39
  outperforms
• 9:40
  models
• 9:40
  up
• 9:41
  to
• 9:41
  7x
• 9:41
  its
• 9:41
  size
• 9:42
  on
• 9:42
  text
• 9:42
  safety
• 9:42
  benchmarks
• 9:43
  and
• 9:43
  sets
• 9:43
  a
• 9:43
  new
• 9:44
  state
• 9:44
  of
• 9:44
  the
• 9:44
  art
• 9:44
  on
• 9:44
  multimodal
• 9:45
  moderation,
• 9:46
  running
• 9:46
  on
• 9:46
  a
• 9:46
  single
• 9:47
  16
• 9:47
  GB
• 9:48
  GPU.
• 9:48
  The
• 9:49
  agent
• 9:49
  containment
• 9:49
  debate
• 9:50
  continued
• 9:50
  with
• 9:50
  a
• 9:50
  piece
• 9:51
  arguing
• 9:51
  that
• 9:51
  sandboxes
• 9:52
  solve
• 9:52
  process
• 9:53
  isolation
• 9:53
  but
• 9:54
  not
• 9:54
  the
• 9:54
  full
• 9:54
  agent
• 9:55
  governance
• 9:55
  problem
• 9:55
  --
• 9:56
  identity,
• 9:57
  memory,
• 9:58
  tools,
• 9:58
  policies,
• 9:59
  mutations,
• 10:00
  provenance,
• 10:01
  and
• 10:01
  future
• 10:02
  versions
• 10:02
  all
• 10:02
  travel
• 10:03
  outside
• 10:03
  the
• 10:03
  sandbox
• 10:04
  boundary.
• 10:04
  And
• 10:05
  a
• 10:05
  Go-based
• 10:05
  tool
• 10:05
  called
• 10:06
  Conversation
• 10:06
  Steganography
• 10:07
  demonstrated
• 10:08
  hiding
• 10:08
  encrypted
• 10:08
  messages
• 10:09
  inside
• 10:09
  normal-looking
• 10:10
  LLM-generated
• 10:11
  chat
• 10:11
  text,
• 10:12
  using
• 10:12
  AES-SIV
• 10:14
  encryption
• 10:15
  with
• 10:15
  token-choice
• 10:16
  encoding
• 10:16
  so
• 10:17
  the
• 10:17
  cover
• 10:17
  text
• 10:17
  is
• 10:17
  generated
• 10:18
  by
• 10:18
  a
• 10:18
  local
• 10:18
  model
• 10:19
  and
• 10:19
  the
• 10:19
  messaging
• 10:20
  platform
• 10:20
  never
• 10:20
  sees
• 10:21
  the
• 10:21
  plaintext.
• 10:22
  The
• 10:22
  security
• 10:22
  model
• 10:23
  requires
• 10:23
  both
• 10:23
  parties
• 10:24
  to
• 10:24
  use
• 10:24
  the
• 10:24
  same
• 10:24
  local
• 10:25
  model
• 10:25
  and
• 10:25
  process
• 10:25
  messages
• 10:26
  in
• 10:26
  exact
• 10:26
  order
• 10:27
  --
• 10:27
  fragile,
• 10:28
  but
• 10:28
  the
• 10:28
  proof
• 10:29
  of
• 10:29
  concept
• 10:29
  illustrates
• 10:30
  that
• 10:30
  LLM
• 10:30
  output
• 10:31
  is
• 10:31
  now
• 10:31
  a
• 10:31
  viable
• 10:32
  steganographic
• 10:32
  channel,
• 10:33
  one
• 10:33
  that
• 10:33
  content
• 10:34
  scanners
• 10:34
  are
• 10:34
  not
• 10:34
  designed
• 10:35
  to
• 10:35
  detect.
• 10:39
  The
• 10:40
  pattern
• 10:40
  of
• 10:40
  safety
• 10:40
  classifiers
• 10:41
  blocking
• 10:41
  the
• 10:41
  exact
• 10:42
  work
• 10:42
  they
• 10:42
  are
• 10:42
  supposed
• 10:43
  to
• 10:43
  enable
• 10:43
  has
• 10:43
  now
• 10:43
  hit
• 10:44
  both
• 10:44
  major
• 10:44
  frontier
• 10:45
  vendors.
• 10:46
  A
• 10:46
  Codex
• 10:46
  user
• 10:46
  ran
• 10:46
  OpenAI's
• 10:47
  Codex
• 10:47
  Security
• 10:48
  feature
• 10:48
  on
• 10:48
  their
• 10:48
  repository
• 10:49
  for
• 10:49
  the
• 10:49
  first
• 10:49
  time.
• 10:50
  It
• 10:50
  worked
• 10:50
  for
• 10:51
  nearly
• 10:51
  20
• 10:51
  minutes,
• 10:52
  consumed
• 10:52
  11%
• 10:53
  of
• 10:53
  the
• 10:53
  weekly
• 10:53
  usage
• 10:53
  limit,
• 10:54
  and
• 10:54
  then
• 10:54
  returned
• 10:55
  "Goal
• 10:55
  blocked,
• 10:55
  This
• 10:56
  content
• 10:56
  can't
• 10:56
  be
• 10:56
  shown."
• 10:57
  No
• 10:57
  results,
• 10:57
  no
• 10:58
  partial
• 10:58
  output,
• 10:59
  no
• 10:59
  way
• 10:59
  to
• 10:59
  see
• 10:59
  the
• 10:59
  thinking
• 11:00
  process
• 11:00
  up
• 11:00
  to
• 11:00
  the
• 11:00
  block
• 11:01
  point.
• 11:02
  The
• 11:02
  community
• 11:02
  response
• 11:02
  was
• 11:03
  predictable:
• 11:03
  "Classic.
• 11:05
  Classifier
• 11:05
  kicks
• 11:05
  in
• 11:05
  mid-run,
• 11:06
  eats
• 11:06
  your
• 11:06
  compute,
• 11:06
  gives
• 11:07
  you
• 11:07
  nothing
• 11:07
  back."
• 11:08
  Others
• 11:08
  noted
• 11:08
  that
• 11:08
  asking
• 11:09
  the
• 11:09
  model
• 11:09
  "why
• 11:09
  are
• 11:10
  we
• 11:10
  blocked"
• 11:10
  sometimes
• 11:11
  causes
• 11:11
  it
• 11:11
  to
• 11:11
  continue,
• 11:12
  suggesting
• 11:12
  the
• 11:12
  block
• 11:13
  is
• 11:13
  classifier
• 11:14
  noise
• 11:14
  rather
• 11:14
  than
• 11:14
  a
• 11:15
  genuine
• 11:15
  safety
• 11:15
  boundary.
• 11:16
  The
• 11:16
  deeper
• 11:16
  problem
• 11:17
  is
• 11:17
  economic:
• 11:18
  a
• 11:18
  security
• 11:18
  analysis
• 11:19
  tool
• 11:19
  that
• 11:19
  burns
• 11:20
  significant
• 11:20
  quota
• 11:20
  before
• 11:21
  deciding
• 11:21
  its
• 11:21
  own
• 11:21
  output
• 11:22
  is
• 11:22
  too
• 11:22
  dangerous
• 11:23
  to
• 11:23
  show
• 11:23
  creates
• 11:23
  a
• 11:23
  perverse
• 11:24
  incentive
• 11:24
  to
• 11:24
  avoid
• 11:25
  using
• 11:25
  it
• 11:25
  on
• 11:25
  anything
• 11:26
  that
• 11:26
  might
• 11:26
  trigger
• 11:26
  a
• 11:26
  false
• 11:26
  positive
• 11:27
  --
• 11:27
  which
• 11:27
  is
• 11:28
  precisely
• 11:28
  the
• 11:28
  kind
• 11:28
  of
• 11:28
  code
• 11:29
  that
• 11:29
  needs
• 11:29
  security
• 11:29
  review.
• 11:30
  Anthropic's
• 11:31
  Fable
• 11:31
  continues
• 11:31
  to
• 11:32
  accumulate
• 11:32
  complaints
• 11:32
  from
• 11:33
  researchers
• 11:33
  who
• 11:33
  find
• 11:33
  the
• 11:33
  model
• 11:34
  unusable
• 11:34
  for
• 11:34
  legitimate
• 11:35
  science.
• 11:36
  A
• 11:36
  scientist
• 11:36
  working
• 11:37
  outside
• 11:37
  any
• 11:37
  of
• 11:37
  the
• 11:37
  high-risk
• 11:38
  domains
• 11:38
  --
• 11:39
  not
• 11:39
  frontier
• 11:39
  AI,
• 11:40
  not
• 11:40
  cybersecurity,
• 11:41
  not
• 11:41
  biology
• 11:42
  --
• 11:42
  reported
• 11:43
  that
• 11:43
  Fable
• 11:43
  "essentially
• 11:44
  non-existent"
• 11:45
  due
• 11:45
  to
• 11:45
  constant
• 11:45
  refusals
• 11:46
  and
• 11:46
  flags,
• 11:46
  leading
• 11:47
  them
• 11:47
  to
• 11:47
  switch
• 11:47
  to
• 11:47
  Sol.
• 11:49
  The
• 11:49
  frustration
• 11:49
  is
• 11:50
  sharpened
• 11:50
  by
• 11:50
  Anthropic's
• 11:50
  stated
• 11:51
  mission
• 11:51
  to
• 11:51
  accelerate
• 11:52
  science
• 11:52
  while
• 11:52
  shipping
• 11:52
  a
• 11:52
  model
• 11:53
  that,
• 11:53
  in
• 11:54
  practice,
• 11:55
  routes
• 11:55
  bioinformatics
• 11:56
  researchers
• 11:56
  to
• 11:56
  a
• 11:56
  fallback
• 11:57
  model
• 11:57
  before
• 11:57
  they
• 11:57
  can
• 11:57
  even
• 11:57
  ask
• 11:58
  about
• 11:58
  the
• 11:58
  weather.
• 11:59
  Multiple
• 12:00
  commenters
• 12:00
  reported
• 12:00
  the
• 12:01
  same
• 12:01
  experience:
• 12:02
  buying
• 12:03
  Pro
• 12:03
  subscriptions
• 12:03
  specifically
• 12:04
  to
• 12:04
  try
• 12:04
  Fable,
• 12:05
  finding
• 12:06
  it
• 12:06
  redirects
• 12:06
  everything,
• 12:07
  and
• 12:07
  concluding
• 12:07
  it
• 12:08
  is
• 12:08
  not
• 12:08
  worth
• 12:08
  the
• 12:08
  prompt
• 12:08
  engineering
• 12:09
  required
• 12:09
  to
• 12:09
  work
• 12:10
  around
• 12:10
  the
• 12:10
  classifier.
• 12:11
  The
• 12:11
  suggestion
• 12:12
  that
• 12:12
  researchers
• 12:12
  from
• 12:13
  credentialed
• 12:13
  institutions
• 12:14
  should
• 12:14
  receive
• 12:15
  differentiated
• 12:15
  access
• 12:15
  is
• 12:16
  reasonable
• 12:16
  but
• 12:17
  appears
• 12:17
  nowhere
• 12:17
  on
• 12:17
  Anthropic's
• 12:18
  roadmap.
• 12:19
  One
• 12:19
  commenter's
• 12:19
  assessment
• 12:20
  --
• 12:21
  that
• 12:21
  Anthropic
• 12:21
  is
• 12:22
  "on
• 12:22
  course
• 12:22
  to
• 12:22
  become
• 12:22
  the
• 12:23
  Yahoo!
• 12:23
  of
• 12:23
  AI"
• 12:24
  --
• 12:25
  is
• 12:25
  harsh
• 12:25
  but
• 12:26
  captures
• 12:26
  the
• 12:26
  gap
• 12:26
  between
• 12:27
  the
• 12:27
  mission
• 12:27
  statement
• 12:27
  and
• 12:27
  the
• 12:27
  user
• 12:28
  experience
• 12:28
  for
• 12:28
  exactly
• 12:29
  the
• 12:29
  audience
• 12:29
  the
• 12:29
  mission
• 12:30
  claims
• 12:30
  to
• 12:30
  serve.
• 12:33
  NEURA
• 12:34
  for
• 12:34
  Browser
• 12:34
  ships
• 12:34
  a
• 12:34
  Chrome/Firefox
• 12:36
  extension
• 12:36
  that
• 12:36
  turns
• 12:37
  Open
• 12:37
  WebUI
• 12:37
  into
• 12:37
  an
• 12:38
  agent
• 12:38
  that
• 12:38
  lives
• 12:38
  in
• 12:38
  every
• 12:39
  browser
• 12:39
  tab.
• 12:40
  The
• 12:40
  sidebar
• 12:40
  connects
• 12:40
  to
• 12:41
  your
• 12:41
  existing
• 12:41
  Open
• 12:41
  WebUI
• 12:42
  instance
• 12:42
  --
• 12:43
  your
• 12:43
  models,
• 12:43
  chat
• 12:44
  history,
• 12:44
  knowledge
• 12:45
  collections
• 12:45
  --
• 12:46
  and
• 12:46
  reads
• 12:47
  the
• 12:47
  current
• 12:47
  page's
• 12:47
  content
• 12:48
  without
• 12:48
  copy-paste.
• 12:49
  The
• 12:49
  more
• 12:49
  interesting
• 12:50
  capability
• 12:50
  is
• 12:50
  agent
• 12:51
  mode:
• 12:52
  the
• 12:52
  model
• 12:52
  gets
• 12:52
  a
• 12:52
  real
• 12:52
  toolset
• 12:53
  against
• 12:53
  the
• 12:53
  active
• 12:54
  tab
• 12:54
  including
• 12:54
  DOM
• 12:55
  snapshots,
• 12:56
  element
• 12:56
  clicking,
• 12:57
  form
• 12:57
  filling
• 12:57
  with
• 12:57
  proper
• 12:58
  framework
• 12:58
  event
• 12:58
  dispatch,
• 12:59
  key
• 12:59
  presses,
• 13:00
  scrolling,
• 13:00
  waiting
• 13:01
  for
• 13:01
  selectors,
• 13:01
  and
• 13:01
  screenshots.
• 13:03
  Every
• 13:03
  tool
• 13:03
  call
• 13:03
  shows
• 13:04
  as
• 13:04
  a
• 13:04
  live
• 13:04
  chip
• 13:04
  in
• 13:04
  the
• 13:04
  chat
• 13:05
  so
• 13:05
  you
• 13:05
  can
• 13:05
  see
• 13:05
  what
• 13:05
  the
• 13:06
  agent
• 13:06
  is
• 13:06
  doing
• 13:06
  and
• 13:06
  kill
• 13:06
  it
• 13:06
  mid-task.
• 13:08
  Navigation
• 13:08
  that
• 13:08
  would
• 13:08
  leave
• 13:08
  the
• 13:09
  current
• 13:09
  origin
• 13:09
  stops
• 13:10
  and
• 13:10
  asks
• 13:10
  for
• 13:10
  confirmation;
• 13:11
  javascript:
• 13:12
  and
• 13:12
  chrome:
• 13:12
  schemes
• 13:13
  are
• 13:13
  blocked
• 13:13
  outright.
• 13:14
  The
• 13:14
  developer
• 13:15
  is
• 13:15
  refreshingly
• 13:15
  honest
• 13:15
  about
• 13:16
  the
• 13:16
  state
• 13:16
  of
• 13:16
  the
• 13:16
  release
• 13:16
  --
• 13:17
  "This
• 13:17
  is
• 13:17
  a
• 13:17
  first
• 13:18
  release
• 13:18
  and
• 13:18
  I
• 13:18
  know
• 13:18
  it.
• 13:19
  There
• 13:19
  are
• 13:19
  bugs."
• 13:19
  --
• 13:20
  which
• 13:20
  is
• 13:20
  more
• 13:20
  useful
• 13:21
  than
• 13:21
  a
• 13:21
  polished
• 13:21
  launch
• 13:21
  blog
• 13:22
  hiding
• 13:22
  the
• 13:22
  same
• 13:22
  rough
• 13:22
  edges.
• 13:23
  On
• 13:23
  the
• 13:24
  writing
• 13:24
  side,
• 13:24
  SimpleEnglish
• 13:25
  is
• 13:25
  an
• 13:25
  agent
• 13:25
  skill
• 13:26
  that
• 13:26
  forces
• 13:26
  LLMs
• 13:27
  to
• 13:27
  write
• 13:27
  in
• 13:27
  ASD-STE100
• 13:30
  Simplified
• 13:30
  Technical
• 13:31
  English,
• 13:31
  the
• 13:32
  controlled
• 13:32
  language
• 13:32
  aerospace
• 13:33
  has
• 13:33
  used
• 13:33
  since
• 13:33
  1983
• 13:34
  to
• 13:34
  prevent
• 13:35
  ambiguity
• 13:35
  in
• 13:35
  maintenance
• 13:36
  manuals.
• 13:36
  The
• 13:36
  rules
• 13:37
  --
• 13:37
  max
• 13:38
  20
• 13:38
  words
• 13:38
  per
• 13:39
  instruction,
• 13:39
  one
• 13:40
  word
• 13:40
  equals
• 13:40
  one
• 13:40
  meaning,
• 13:41
  active
• 13:41
  voice
• 13:41
  only,
• 13:42
  no
• 13:42
  should/would/may/might,
• 13:44
  condition
• 13:44
  before
• 13:45
  command
• 13:45
  --
• 13:46
  happen
• 13:46
  to
• 13:46
  be
• 13:46
  a
• 13:46
  near-perfect
• 13:47
  negative
• 13:47
  of
• 13:47
  every
• 13:47
  AI
• 13:48
  writing
• 13:48
  tell.
• 13:49
  Measured
• 13:49
  across
• 13:49
  96
• 13:50
  runs
• 13:50
  on
• 13:50
  6
• 13:51
  Claude
• 13:51
  models,
• 13:51
  the
• 13:51
  skill
• 13:52
  reduced
• 13:52
  STE
• 13:52
  violations
• 13:53
  by
• 13:53
  72.9%
• 13:54
  per
• 13:54
  100
• 13:55
  words
• 13:55
  while
• 13:55
  also
• 13:56
  reducing
• 13:56
  output
• 13:56
  token
• 13:57
  count.
• 13:57
  The
• 13:58
  before/after
• 13:58
  examples
• 13:59
  are
• 13:59
  striking:
• 14:00
  "Leveraging
• 14:01
  sqlpipe's
• 14:01
  robust
• 14:02
  architecture,
• 14:02
  users
• 14:03
  can
• 14:03
  seamlessly
• 14:04
  synchronize..."
• 14:05
  becomes
• 14:05
  "sqlpipe
• 14:05
  copies
• 14:06
  your
• 14:06
  Postgres
• 14:06
  tables
• 14:07
  to
• 14:07
  S3.
• 14:08
  It
• 14:08
  needs
• 14:08
  one
• 14:08
  configuration
• 14:09
  file."
• 14:10
  For
• 14:10
  teams
• 14:10
  trying
• 14:11
  to
• 14:11
  understand
• 14:11
  where
• 14:11
  their
• 14:12
  AI
• 14:12
  budget
• 14:12
  actually
• 14:13
  goes,
• 14:14
  a
• 14:14
  curated
• 14:14
  reference
• 14:14
  called
• 14:15
  Awesome
• 14:15
  AI
• 14:15
  Tokenomics
• 14:16
  aggregates
• 14:17
  roughly
• 14:17
  200
• 14:17
  entries
• 14:18
  across
• 14:18
  monitoring
• 14:18
  dashboards,
• 14:19
  caching
• 14:20
  tools,
• 14:21
  compression
• 14:21
  libraries,
• 14:22
  routing
• 14:22
  frameworks,
• 14:23
  and
• 14:23
  academic
• 14:23
  papers
• 14:24
  on
• 14:24
  token
• 14:24
  economics
• 14:25
  --
• 14:26
  organized
• 14:26
  not
• 14:26
  by
• 14:27
  hype
• 14:27
  but
• 14:27
  by
• 14:27
  function,
• 14:28
  with
• 14:28
  license
• 14:28
  badges
• 14:29
  and
• 14:29
  one-line
• 14:29
  summaries.
• 14:30
  A
• 14:31
  YouTube
• 14:31
  walkthrough
• 14:32
  of
• 14:32
  the
• 14:32
  Hermes
• 14:32
  agent
• 14:32
  framework
• 14:33
  addressed
• 14:33
  the
• 14:33
  five
• 14:34
  most
• 14:34
  common
• 14:34
  community
• 14:35
  questions,
• 14:36
  emphasizing
• 14:37
  that
• 14:37
  Hermes
• 14:37
  itself
• 14:37
  is
• 14:38
  lightweight
• 14:38
  on
• 14:38
  tokens
• 14:38
  --
• 14:39
  a
• 14:39
  report
• 14:40
  of
• 14:40
  burning
• 14:40
  50
• 14:40
  million
• 14:41
  tokens
• 14:41
  in
• 14:41
  an
• 14:41
  hour
• 14:42
  likely
• 14:42
  reflected
• 14:42
  a
• 14:43
  retry
• 14:43
  loop
• 14:43
  rather
• 14:44
  than
• 14:44
  normal
• 14:44
  operation
• 14:44
  --
• 14:45
  and
• 14:45
  that
• 14:45
  the
• 14:46
  framework's
• 14:46
  value
• 14:46
  is
• 14:47
  portability:
• 14:48
  workflows,
• 14:49
  skills,
• 14:49
  and
• 14:50
  self-improvement
• 14:50
  configurations
• 14:51
  move
• 14:51
  with
• 14:51
  you
• 14:52
  when
• 14:52
  you
• 14:52
  swap
• 14:52
  underlying
• 14:53
  models.
• 14:58
  OpenAI
• 14:58
  settled
• 14:58
  with
• 14:59
  the
• 14:59
  Department
• 14:59
  of
• 14:59
  Justice
• 14:59
  for
• 15:00
  $3.2
• 15:00
  million
• 15:01
  over
• 15:01
  allegations
• 15:02
  that
• 15:02
  it
• 15:02
  discriminated
• 15:03
  against
• 15:03
  U.S.
• 15:03
  workers
• 15:04
  during
• 15:04
  the
• 15:04
  PERM
• 15:04
  (Permanent
• 15:04
  Labor
• 15:05
  Certification)
• 15:05
  hiring
• 15:06
  process.
• 15:06
  The
• 15:07
  DOJ's
• 15:07
  investigation
• 15:08
  found
• 15:08
  that
• 15:08
  OpenAI
• 15:08
  failed
• 15:09
  to
• 15:09
  publicly
• 15:09
  post
• 15:10
  PERM
• 15:10
  positions
• 15:10
  on
• 15:10
  its
• 15:10
  careers
• 15:11
  website,
• 15:12
  required
• 15:12
  paper
• 15:12
  applications
• 15:13
  instead
• 15:13
  of
• 15:13
  electronic
• 15:14
  submissions,
• 15:14
  and
• 15:14
  used
• 15:15
  recruiting
• 15:15
  practices
• 15:16
  that
• 15:16
  discouraged
• 15:16
  U.S.
• 15:16
  applicants
• 15:17
  --
• 15:17
  including
• 15:18
  late-night
• 15:18
  radio
• 15:18
  advertisements.
• 15:20
  Under
• 15:20
  the
• 15:20
  settlement,
• 15:20
  OpenAI
• 15:21
  will
• 15:21
  pay
• 15:21
  $1.2
• 15:22
  million
• 15:22
  in
• 15:22
  civil
• 15:22
  penalties
• 15:23
  and
• 15:23
  establish
• 15:23
  a
• 15:23
  $2
• 15:24
  million
• 15:24
  back-pay
• 15:24
  fund.
• 15:26
  The
• 15:26
  company
• 15:26
  pushed
• 15:26
  back
• 15:27
  on
• 15:27
  the
• 15:27
  findings
• 15:27
  while
• 15:27
  agreeing
• 15:28
  to
• 15:28
  resolve
• 15:28
  the
• 15:28
  matter,
• 15:29
  stating
• 15:29
  that
• 15:30
  "maintaining
• 15:30
  America's
• 15:30
  leadership
• 15:31
  in
• 15:31
  AI
• 15:31
  requires
• 15:32
  attracting
• 15:32
  and
• 15:32
  retaining
• 15:33
  the
• 15:33
  best
• 15:33
  talent
• 15:33
  from
• 15:33
  the
• 15:33
  United
• 15:34
  States
• 15:34
  and
• 15:34
  around
• 15:34
  the
• 15:34
  world."
• 15:35
  The
• 15:35
  settlement
• 15:36
  is
• 15:36
  notable
• 15:36
  less
• 15:36
  for
• 15:36
  its
• 15:36
  size
• 15:37
  than
• 15:37
  for
• 15:37
  its
• 15:37
  timing:
• 15:38
  it
• 15:39
  arrives
• 15:39
  while
• 15:39
  OpenAI
• 15:40
  is
• 15:40
  simultaneously
• 15:40
  promoting
• 15:41
  the
• 15:41
  $500
• 15:41
  billion
• 15:42
  Stargate
• 15:42
  infrastructure
• 15:43
  project
• 15:44
  and
• 15:44
  a
• 15:44
  pledge
• 15:44
  to
• 15:44
  certify
• 15:44
  10
• 15:45
  million
• 15:45
  Americans
• 15:45
  in
• 15:45
  AI
• 15:46
  skills
• 15:46
  by
• 15:46
  2030.
• 15:48
  Federal
• 15:48
  officials
• 15:49
  noted
• 15:49
  that
• 15:49
  fewer
• 15:49
  than
• 15:49
  10
• 15:49
  PERM
• 15:50
  positions
• 15:50
  were
• 15:50
  involved,
• 15:51
  but
• 15:51
  the
• 15:51
  settlement
• 15:51
  amount
• 15:52
  reflects
• 15:52
  "the
• 15:52
  impact
• 15:53
  of
• 15:53
  excluding
• 15:53
  U.S.
• 15:54
  workers
• 15:54
  from
• 15:54
  high-paying
• 15:54
  technology
• 15:55
  positions."
• 15:56
  This
• 15:56
  marks
• 15:56
  the
• 15:56
  13th
• 15:57
  settlement
• 15:57
  under
• 15:57
  the
• 15:57
  DOJ's
• 15:58
  Protecting
• 15:58
  U.S.
• 15:59
  Workers
• 15:59
  Initiative
• 15:59
  launched
• 16:00
  in
• 16:00
  2025.
• 16:01
  In
• 16:01
  a
• 16:01
  different
• 16:01
  corner
• 16:02
  of
• 16:02
  the
• 16:02
  industry,
• 16:02
  Robert
• 16:03
  Hansen
• 16:03
  and
• 16:03
  Jeremiah
• 16:04
  Grossman
• 16:04
  published
• 16:04
  "The
• 16:05
  End
• 16:05
  of
• 16:05
  Guessing,"
• 16:05
  a
• 16:06
  book
• 16:06
  arguing
• 16:06
  that
• 16:06
  vulnerability
• 16:07
  management
• 16:07
  and
• 16:08
  external
• 16:08
  attack
• 16:08
  surface
• 16:09
  management
• 16:09
  have
• 16:09
  both
• 16:09
  failed
• 16:10
  their
• 16:10
  core
• 16:10
  promises.
• 16:11
  The
• 16:11
  thesis:
• 16:12
  VM
• 16:12
  scanners
• 16:13
  are
• 16:13
  littered
• 16:13
  with
• 16:13
  ancient
• 16:14
  vulnerabilities
• 16:14
  that
• 16:14
  never
• 16:15
  mattered,
• 16:15
  theoretical
• 16:16
  CVE
• 16:16
  issues,
• 16:17
  and
• 16:17
  minor
• 16:17
  information
• 16:18
  leaks,
• 16:18
  driving
• 16:19
  per-scan
• 16:19
  costs
• 16:20
  up
• 16:20
  while
• 16:20
  customers
• 16:21
  choose
• 16:21
  to
• 16:21
  scan
• 16:21
  only
• 16:22
  a
• 16:22
  fraction
• 16:22
  of
• 16:22
  their
• 16:22
  attack
• 16:23
  surface.
• 16:23
  EASM
• 16:24
  tools,
• 16:25
  meanwhile,
• 16:25
  were
• 16:25
  built
• 16:26
  to
• 16:26
  produce
• 16:26
  asset
• 16:26
  lists,
• 16:27
  not
• 16:27
  to
• 16:27
  be
• 16:27
  asset
• 16:28
  maps
• 16:28
  themselves
• 16:28
  --
• 16:29
  a
• 16:29
  distinction
• 16:30
  that
• 16:30
  matters
• 16:31
  because
• 16:31
  the
• 16:31
  average
• 16:31
  customer
• 16:32
  does
• 16:32
  not
• 16:32
  know
• 16:32
  what
• 16:32
  a
• 16:32
  complete
• 16:33
  inventory
• 16:33
  looks
• 16:34
  like.
• 16:35
  The
• 16:35
  authors
• 16:35
  claim
• 16:35
  to
• 16:35
  have
• 16:36
  figured
• 16:36
  out
• 16:36
  how
• 16:36
  to
• 16:36
  fix
• 16:37
  both
• 16:37
  problems
• 16:37
  using
• 16:38
  evidence-based
• 16:38
  prioritization
• 16:39
  rather
• 16:39
  than
• 16:40
  CVSS
• 16:40
  severity
• 16:41
  scores,
• 16:41
  building
• 16:42
  on
• 16:42
  the
• 16:42
  actuarial
• 16:43
  insight
• 16:43
  that
• 16:43
  49%
• 16:44
  of
• 16:44
  CVEs
• 16:45
  rated
• 16:45
  high
• 16:45
  or
• 16:45
  critical
• 16:46
  have
• 16:46
  less
• 16:47
  than
• 16:47
  a
• 16:47
  1%
• 16:47
  chance
• 16:48
  of
• 16:48
  exploitation
• 16:49
  in
• 16:49
  any
• 16:49
  given
• 16:49
  month.
• 16:50
  For
• 16:50
  anyone
• 16:51
  attending
• 16:51
  Black
• 16:51
  Hat,
• 16:52
  signed
• 16:52
  copies
• 16:53
  are
• 16:53
  available
• 16:53
  at
• 16:53
  the
• 16:53
  Root
• 16:53
  Evidence
• 16:54
  booth
• 16:54
  on
• 16:54
  Wednesday
• 16:55
  and
• 16:55
  Thursday.