Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – July 30, 2026

Transcript

AGI Dreams Podcast – July 30, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Moonshot
• 0:00
  AI's
• 0:00
  Kimi
• 0:01
  K3
• 0:01
  is
• 0:01
  one
• 0:02
  of
• 0:02
  the
• 0:02
  largest
• 0:02
  open-weight
• 0:03
  models
• 0:03
  ever
• 0:03
  released:
• 0:04
  2.8
• 0:05
  trillion
• 0:05
  parameters
• 0:06
  in
• 0:06
  a
• 0:06
  mixture-of-experts
• 0:07
  architecture
• 0:07
  activating
• 0:08
  roughly
• 0:08
  104
• 0:08
  billion
• 0:09
  per
• 0:09
  token,
• 0:10
  built
• 0:10
  on
• 0:10
  Delta
• 0:10
  Attention
• 0:11
  (KDA)
• 0:11
  and
• 0:11
  a
• 0:11
  LatentMoE
• 0:12
  framework
• 0:12
  that
• 0:12
  selects
• 0:13
  16
• 0:13
  of
• 0:13
  896
• 0:14
  experts
• 0:14
  per
• 0:15
  forward
• 0:15
  pass.
• 0:16
  Moonshot's
• 0:16
  reported
• 0:17
  benchmarks
• 0:17
  are
• 0:17
  striking
• 0:17
  —
• 0:18
  though
• 0:18
  some
• 0:18
  models
• 0:19
  were
• 0:19
  evaluated
• 0:19
  with
• 0:20
  different
• 0:20
  agent
• 0:20
  harnesses,
• 0:21
  K3
• 0:21
  trades
• 0:21
  blows
• 0:22
  with
• 0:22
  Fable
• 0:22
  5
• 0:22
  and
• 0:22
  GPT-5.6
• 0:24
  Sol
• 0:24
  across
• 0:24
  coding,
• 0:25
  legal
• 0:25
  reasoning,
• 0:26
  and
• 0:26
  agentic
• 0:26
  tasks,
• 0:27
  and
• 0:28
  on
• 0:28
  Frontend
• 0:28
  Code
• 0:28
  Arena,
• 0:29
  it
• 0:29
  outright
• 0:29
  wins.
• 0:30
  On
• 0:30
  BrowseComp,
• 0:31
  Moonshot
• 0:31
  reports
• 0:31
  91.2%
• 0:33
  at
• 0:33
  a
• 0:33
  fraction
• 0:34
  of
• 0:34
  the
• 0:34
  per-task
• 0:34
  cost
• 0:35
  of
• 0:35
  closed
• 0:35
  competitors.
• 0:36
  Community-produced
• 0:37
  benchmarks
• 0:38
  and
• 0:38
  sponsored
• 0:38
  walkthroughs
• 0:39
  confirm
• 0:39
  the
• 0:40
  headline
• 0:40
  numbers,
• 0:41
  while
• 0:41
  also
• 0:41
  revealing
• 0:42
  the
• 0:42
  model's
• 0:42
  practical
• 0:43
  ceiling:
• 0:43
  self-hosting
• 0:44
  K3
• 0:45
  at
• 0:45
  usable
• 0:45
  speed
• 0:45
  (50
• 0:46
  tok/s)
• 0:47
  requires
• 0:47
  a
• 0:47
  full
• 0:47
  DGX
• 0:48
  B400
• 0:48
  server
• 0:48
  —
• 0:49
  eight
• 0:50
  GPUs,
• 0:50
  2.3
• 0:51
  TB
• 0:51
  of
• 0:52
  HBM,
• 0:52
  $300-400K
• 0:54
  in
• 0:54
  hardware.
• 0:55
  A
• 0:55
  more
• 0:55
  realistic
• 0:56
  hobbyist
• 0:56
  path,
• 0:57
  at
• 0:57
  5-10
• 0:57
  tok/s,
• 0:58
  still
• 0:59
  demands
• 0:59
  $16-22K
• 1:00
  in
• 1:00
  used
• 1:01
  GPUs.
• 1:02
  Cost
• 1:02
  per
• 1:03
  task,
• 1:03
  not
• 1:03
  cost
• 1:04
  per
• 1:04
  token,
• 1:04
  is
• 1:04
  the
• 1:05
  metric
• 1:05
  that
• 1:05
  matters
• 1:05
  —
• 1:06
  and
• 1:06
  on
• 1:06
  that
• 1:06
  measure,
• 1:07
  K3
• 1:07
  consistently
• 1:08
  undercuts
• 1:08
  Fable
• 1:09
  for
• 1:09
  comparable
• 1:09
  quality.
• 1:11
  The
• 1:11
  open-weight
• 1:11
  community
• 1:12
  has
• 1:12
  already
• 1:12
  started
• 1:12
  compressing.
• 1:14
  A
• 1:14
  1-bit
• 1:14
  quantization
• 1:15
  shrinks
• 1:15
  K3
• 1:15
  from
• 1:15
  1.56
• 1:16
  TB
• 1:17
  to
• 1:17
  594
• 1:18
  GB,
• 1:19
  retaining
• 1:19
  approximately
• 1:20
  78.9%
• 1:21
  top-1
• 1:21
  accuracy;
• 1:22
  a
• 1:23
  2-bit
• 1:23
  variant
• 1:23
  exceeds
• 1:24
  90%.
• 1:25
  Running
• 1:25
  the
• 1:26
  1-bit
• 1:26
  model
• 1:26
  still
• 1:27
  requires
• 1:27
  a
• 1:27
  Mac
• 1:27
  Studio
• 1:28
  with
• 1:28
  512
• 1:29
  GB
• 1:29
  RAM
• 1:29
  plus
• 1:30
  a
• 1:30
  secondary
• 1:30
  128
• 1:31
  GB
• 1:31
  machine
• 1:31
  —
• 1:33
  hardware
• 1:33
  that
• 1:33
  Apple
• 1:33
  has
• 1:33
  discontinued.
• 1:35
  At
• 1:35
  the
• 1:35
  other
• 1:35
  end,
• 1:36
  loading
• 1:36
  594
• 1:37
  GB
• 1:37
  on
• 1:37
  a
• 1:37
  128
• 1:38
  GB
• 1:38
  box
• 1:39
  forces
• 1:39
  over
• 1:39
  400
• 1:40
  GB
• 1:40
  onto
• 1:40
  NVMe
• 1:41
  swap,
• 1:41
  and
• 1:41
  throughput
• 1:42
  drops
• 1:42
  to
• 1:42
  a
• 1:42
  crawl.
• 1:43
  The
• 1:43
  honest
• 1:44
  takeaway:
• 1:45
  K3
• 1:45
  is
• 1:45
  an
• 1:45
  F1
• 1:46
  car
• 1:46
  in
• 1:46
  a
• 1:46
  showroom.
• 1:47
  You
• 1:47
  can
• 1:47
  look
• 1:47
  at
• 1:47
  it.
• 1:48
  Running
• 1:48
  it
• 1:48
  is
• 1:49
  another
• 1:49
  question.
• 1:50
  The
• 1:50
  gap
• 1:50
  between
• 1:51
  "open
• 1:51
  weights"
• 1:51
  and
• 1:52
  "open
• 1:52
  access"
• 1:53
  surfaced
• 1:53
  immediately
• 1:54
  when
• 1:54
  Ollama
• 1:54
  listed
• 1:54
  K3
• 1:55
  behind
• 1:55
  a
• 1:55
  pay-per-use
• 1:56
  gate,
• 1:56
  even
• 1:57
  for
• 1:57
  Max-tier
• 1:57
  subscribers.
• 1:59
  The
• 1:59
  community
• 1:59
  reaction
• 2:00
  was
• 2:00
  sharp:
• 2:01
  Ollama
• 2:01
  had
• 2:01
  always
• 2:01
  been
• 2:02
  positioned
• 2:02
  as
• 2:02
  a
• 2:02
  local-first
• 2:03
  inference
• 2:03
  server,
• 2:04
  and
• 2:04
  charging
• 2:04
  separate
• 2:05
  usage
• 2:05
  credits
• 2:05
  for
• 2:05
  an
• 2:05
  open
• 2:06
  model
• 2:06
  felt
• 2:06
  like
• 2:06
  a
• 2:07
  betrayal
• 2:07
  of
• 2:07
  that
• 2:07
  identity.
• 2:09
  Multiple
• 2:09
  users
• 2:09
  reported
• 2:10
  canceling
• 2:10
  subscriptions.
• 2:11
  Ollama's
• 2:12
  stated
• 2:12
  reason
• 2:12
  —
• 2:13
  capacity
• 2:14
  constraints
• 2:14
  on
• 2:14
  a
• 2:14
  model
• 2:15
  requiring
• 2:15
  8-10
• 2:16
  B200
• 2:16
  GPUs
• 2:17
  per
• 2:17
  concurrent
• 2:17
  user
• 2:18
  —
• 2:18
  is
• 2:19
  plausible,
• 2:20
  but
• 2:20
  it
• 2:20
  exposed
• 2:20
  the
• 2:21
  uncomfortable
• 2:21
  truth
• 2:21
  that
• 2:22
  hosting
• 2:22
  open
• 2:22
  models
• 2:23
  at
• 2:23
  frontier
• 2:23
  scale
• 2:24
  is
• 2:24
  not
• 2:24
  free.
• 2:25
  Meanwhile,
• 2:26
  the
• 2:26
  models
• 2:26
  people
• 2:27
  actually
• 2:27
  run
• 2:27
  every
• 2:27
  day
• 2:28
  tell
• 2:28
  a
• 2:28
  quieter
• 2:28
  story.
• 2:29
  A
• 2:29
  month
• 2:30
  after
• 2:30
  the
• 2:30
  initial
• 2:30
  excitement
• 2:31
  fades,
• 2:32
  the
• 2:32
  stack
• 2:32
  that
• 2:32
  survives
• 2:33
  is
• 2:33
  GLM
• 2:33
  5.2
• 2:34
  for
• 2:34
  high-end
• 2:35
  local
• 2:35
  hardware,
• 2:36
  Qwen3.6
• 2:38
  27B
• 2:40
  for
• 2:40
  knowledge-dense
• 2:41
  tasks,
• 2:41
  Gemma
• 2:42
  4
• 2:43
  26B-A4B
• 2:44
  for
• 2:44
  speed-critical
• 2:45
  agent
• 2:45
  slots,
• 2:46
  and
• 2:46
  DeepSeek
• 2:46
  V4
• 2:47
  Flash
• 2:47
  for
• 2:47
  its
• 2:47
  384K
• 2:49
  context
• 2:49
  window
• 2:50
  and
• 2:50
  sheer
• 2:50
  consistency.
• 2:51
  The
• 2:52
  highest
• 2:52
  praise
• 2:52
  in
• 2:52
  the
• 2:52
  local-model
• 2:53
  community
• 2:53
  remains:
• 2:54
  "I
• 2:54
  stopped
• 2:55
  thinking
• 2:55
  about
• 2:55
  it."
• 2:56
  One
• 2:56
  user
• 2:57
  runs
• 2:57
  Laguna
• 2:57
  S2.1
• 2:59
  as
• 2:59
  a
• 2:59
  red-team
• 3:00
  pen
• 3:00
  tester
• 3:00
  and
• 3:00
  gap
• 3:01
  finder.
• 3:02
  Another
• 3:02
  keeps
• 3:02
  GLM-4.5-Air
• 3:04
  warm
• 3:04
  for
• 3:04
  STEM
• 3:04
  tasks
• 3:05
  despite
• 3:05
  its
• 3:05
  age,
• 3:06
  pairing
• 3:06
  it
• 3:06
  with
• 3:06
  Gemma-4-31B
• 3:07
  as
• 3:08
  a
• 3:08
  debugger.
• 3:09
  The
• 3:09
  picture
• 3:09
  is
• 3:09
  clear:
• 3:10
  frontier-class
• 3:10
  open
• 3:11
  models
• 3:11
  are
• 3:11
  exciting,
• 3:12
  but
• 3:12
  the
• 3:12
  models
• 3:12
  people
• 3:13
  trust
• 3:13
  with
• 3:13
  real
• 3:13
  work
• 3:14
  are
• 3:14
  the
• 3:14
  mid-size
• 3:14
  ones
• 3:14
  they've
• 3:15
  lived
• 3:15
  with
• 3:15
  long
• 3:15
  enough
• 3:15
  to
• 3:16
  know
• 3:16
  where
• 3:16
  they
• 3:16
  break.
• 3:17
  Chinese
• 3:17
  model
• 3:18
  labs
• 3:18
  —
• 3:19
  DeepSeek,
• 3:19
  Qwen/Alibaba,
• 3:21
  Moonshot,
• 3:22
  Zhipu
• 3:22
  —
• 3:23
  are
• 3:23
  not
• 3:24
  just
• 3:24
  catching
• 3:24
  up;
• 3:24
  on
• 3:24
  several
• 3:25
  benchmarks
• 3:25
  they
• 3:25
  lead.
• 3:26
  The
• 3:26
  deployment
• 3:27
  decisions
• 3:27
  are
• 3:27
  increasingly
• 3:28
  pragmatic:
• 3:29
  use
• 3:29
  K3
• 3:29
  as
• 3:30
  a
• 3:30
  planner
• 3:30
  and
• 3:30
  route
• 3:30
  execution
• 3:31
  to
• 3:31
  cheaper
• 3:31
  models
• 3:32
  like
• 3:32
  Kimi
• 3:32
  K2.7
• 3:34
  Code.
• 3:37
  Anthropic
• 3:38
  published
• 3:38
  two
• 3:39
  new
• 3:39
  results
• 3:39
  from
• 3:39
  Claude
• 3:39
  Mythos:
• 3:40
  an
• 3:41
  attack
• 3:41
  on
• 3:41
  the
• 3:41
  post-quantum
• 3:42
  signature
• 3:42
  scheme
• 3:42
  HAWK
• 3:43
  and
• 3:43
  an
• 3:43
  improved
• 3:44
  attack
• 3:44
  on
• 3:44
  reduced-round
• 3:45
  AES.
• 3:46
  The
• 3:46
  Mythos
• 3:46
  chain-of-thought
• 3:47
  transcript
• 3:47
  runs
• 3:48
  to
• 3:48
  hundreds
• 3:48
  of
• 3:48
  pages,
• 3:49
  demonstrating
• 3:50
  the
• 3:50
  model's
• 3:50
  ability
• 3:51
  to
• 3:51
  synthesize
• 3:51
  existing
• 3:52
  cryptanalytic
• 3:53
  tools
• 3:53
  into
• 3:53
  novel
• 3:53
  attacks
• 3:54
  without
• 3:54
  detailed
• 3:54
  human
• 3:55
  intervention.
• 3:56
  Matt
• 3:56
  Green,
• 3:57
  the
• 3:57
  Johns
• 3:57
  Hopkins
• 3:57
  cryptographer,
• 3:58
  provides
• 3:58
  the
• 3:58
  essential
• 3:59
  context.
• 4:00
  The
• 4:00
  HAWK
• 4:00
  result
• 4:00
  is
• 4:01
  the
• 4:01
  one
• 4:01
  that
• 4:01
  matters:
• 4:02
  it
• 4:02
  roughly
• 4:02
  halves
• 4:03
  the
• 4:03
  scheme's
• 4:03
  security
• 4:04
  bits,
• 4:04
  making
• 4:04
  HAWK
• 4:05
  —
• 4:05
  a
• 4:05
  candidate
• 4:06
  for
• 4:06
  future
• 4:06
  post-quantum
• 4:07
  standardization
• 4:07
  —
• 4:08
  effectively
• 4:09
  unjustifiable
• 4:10
  when
• 4:10
  more
• 4:10
  efficient
• 4:11
  alternatives
• 4:11
  exist.
• 4:12
  The
• 4:12
  attack
• 4:13
  doesn't
• 4:13
  break
• 4:13
  HAWK
• 4:13
  in
• 4:13
  the
• 4:13
  sci-fi
• 4:14
  sense
• 4:14
  (it's
• 4:14
  still
• 4:14
  exponential
• 4:15
  time),
• 4:16
  but
• 4:16
  it
• 4:16
  produced
• 4:16
  working
• 4:17
  code
• 4:17
  that
• 4:17
  recovers
• 4:17
  keys
• 4:18
  against
• 4:18
  a
• 4:18
  weakened
• 4:18
  challenge
• 4:18
  instance
• 4:19
  in
• 4:19
  hours.
• 4:20
  What
• 4:20
  makes
• 4:20
  it
• 4:21
  genuinely
• 4:21
  concerning,
• 4:21
  Green
• 4:22
  notes,
• 4:22
  is
• 4:22
  that
• 4:22
  "none
• 4:23
  of
• 4:23
  the
• 4:23
  ingredients
• 4:23
  are
• 4:23
  exotic."
• 4:24
  Mythos
• 4:25
  simply
• 4:25
  did
• 4:25
  a
• 4:25
  more
• 4:25
  thorough
• 4:26
  job
• 4:26
  applying
• 4:26
  known
• 4:26
  tools
• 4:27
  than
• 4:27
  any
• 4:27
  human
• 4:28
  team
• 4:28
  had
• 4:28
  bothered
• 4:28
  to.
• 4:29
  This
• 4:29
  is
• 4:30
  exactly
• 4:30
  the
• 4:30
  kind
• 4:30
  of
• 4:31
  task
• 4:31
  where
• 4:31
  AI
• 4:31
  excels:
• 4:32
  exhaustive
• 4:33
  recombination
• 4:34
  of
• 4:34
  existing
• 4:34
  techniques
• 4:35
  across
• 4:35
  a
• 4:35
  large
• 4:35
  search
• 4:36
  space.
• 4:37
  The
• 4:37
  AES
• 4:37
  result,
• 4:38
  by
• 4:38
  contrast,
• 4:39
  is
• 4:39
  a
• 4:39
  modest
• 4:39
  constant-factor
• 4:40
  improvement
• 4:40
  on
• 4:41
  previous
• 4:41
  7-round
• 4:41
  attacks.
• 4:42
  Full
• 4:43
  AES
• 4:43
  runs
• 4:43
  10,
• 4:44
  12,
• 4:44
  or
• 4:44
  14
• 4:44
  rounds;
• 4:45
  attacks
• 4:46
  on
• 4:46
  7
• 4:46
  rounds
• 4:46
  have
• 4:46
  existed
• 4:47
  for
• 4:47
  years.
• 4:48
  The
• 4:48
  new
• 4:48
  result
• 4:49
  requires
• 4:49
  2^89
• 4:51
  cipher
• 4:51
  operations
• 4:52
  after
• 4:52
  obtaining
• 4:52
  2^105
• 4:55
  chosen-plaintext
• 4:55
  encryptions.
• 4:56
  It's
• 4:57
  interesting
• 4:57
  from
• 4:57
  a
• 4:57
  techniques
• 4:58
  standpoint
• 4:58
  but
• 4:58
  nowhere
• 4:58
  near
• 4:59
  a
• 4:59
  practical
• 4:59
  threat.
• 5:00
  Green's
• 5:00
  framing:
• 5:01
  "no
• 5:01
  wildly
• 5:02
  new
• 5:02
  mathematical
• 5:02
  results
• 5:03
  here."
• 5:04
  The
• 5:04
  real
• 5:04
  bottleneck,
• 5:04
  Green
• 5:05
  argues,
• 5:05
  is
• 5:05
  verifiability.
• 5:07
  For
• 5:07
  full
• 5:07
  attacks
• 5:07
  like
• 5:08
  HAWK,
• 5:08
  verification
• 5:08
  is
• 5:09
  easy
• 5:09
  —
• 5:10
  run
• 5:10
  the
• 5:10
  code,
• 5:10
  check
• 5:11
  that
• 5:11
  it
• 5:11
  recovers
• 5:11
  keys.
• 5:13
  For
• 5:13
  subtle
• 5:13
  speedup
• 5:13
  claims
• 5:14
  like
• 5:14
  the
• 5:14
  AES
• 5:14
  result,
• 5:15
  checking
• 5:16
  validity
• 5:16
  requires
• 5:16
  human
• 5:17
  experts
• 5:17
  to
• 5:17
  verify
• 5:18
  theorem
• 5:18
  formulations,
• 5:19
  and
• 5:19
  that
• 5:20
  human
• 5:20
  attention
• 5:20
  is
• 5:20
  now
• 5:21
  the
• 5:21
  scarce
• 5:21
  resource.
• 5:22
  The
• 5:22
  models
• 5:23
  produce
• 5:23
  results
• 5:24
  faster
• 5:24
  than
• 5:24
  humans
• 5:24
  can
• 5:24
  verify
• 5:25
  them.
• 5:26
  Green's
• 5:26
  conclusion:
• 5:27
  the
• 5:27
  line
• 5:28
  between
• 5:28
  where
• 5:28
  AI
• 5:28
  is
• 5:29
  helpful
• 5:29
  and
• 5:29
  where
• 5:30
  it's
• 5:30
  clueless
• 5:30
  is
• 5:30
  "slowly
• 5:31
  drifting
• 5:31
  outwards
• 5:32
  under
• 5:32
  your
• 5:32
  feet."
• 5:33
  Whether
• 5:33
  that's
• 5:33
  good
• 5:34
  or
• 5:34
  bad
• 5:34
  depends
• 5:34
  on
• 5:35
  whether
• 5:35
  you
• 5:35
  prefer
• 5:35
  to
• 5:35
  wade
• 5:36
  or
• 5:36
  swim.
• 5:40
  Wiz
• 5:40
  Research
• 5:40
  disclosed
• 5:41
  CosmosEscape,
• 5:42
  a
• 5:42
  critical
• 5:42
  vulnerability
• 5:43
  in
• 5:43
  Azure
• 5:44
  Cosmos
• 5:44
  DB's
• 5:44
  Gremlin
• 5:45
  API
• 5:45
  that
• 5:45
  could
• 5:46
  have
• 5:46
  compromised
• 5:46
  every
• 5:47
  database
• 5:47
  in
• 5:47
  the
• 5:47
  service
• 5:47
  —
• 5:48
  including
• 5:49
  Microsoft's
• 5:49
  own
• 5:49
  internal
• 5:50
  databases
• 5:50
  backing
• 5:51
  Teams,
• 5:51
  Entra
• 5:52
  ID,
• 5:52
  and
• 5:52
  Copilot.
• 5:54
  The
• 5:54
  attack
• 5:54
  chain
• 5:54
  exploited
• 5:55
  .NET
• 5:55
  reflection
• 5:56
  to
• 5:56
  escape
• 5:56
  the
• 5:56
  Gremlin
• 5:57
  query
• 5:57
  sandbox,
• 5:57
  gaining
• 5:58
  code
• 5:58
  execution
• 5:58
  on
• 5:59
  Cosmos
• 5:59
  DB's
• 5:59
  Gateway
• 6:00
  service.
• 6:01
  From
• 6:01
  there,
• 6:01
  Wiz
• 6:02
  discovered
• 6:02
  a
• 6:02
  platform-wide
• 6:03
  signing
• 6:03
  key
• 6:03
  —
• 6:04
  they
• 6:04
  dubbed
• 6:04
  it
• 6:04
  the
• 6:04
  "Cosmos
• 6:05
  Master
• 6:05
  Key"
• 6:05
  —
• 6:06
  that
• 6:06
  could
• 6:06
  retrieve
• 6:07
  the
• 6:07
  primary
• 6:07
  key
• 6:08
  for
• 6:08
  any
• 6:08
  Cosmos
• 6:08
  DB
• 6:09
  account
• 6:09
  across
• 6:09
  all
• 6:09
  tenants,
• 6:10
  regions,
• 6:11
  and
• 6:11
  API
• 6:11
  flavors.
• 6:12
  Combined
• 6:13
  with
• 6:13
  access
• 6:13
  to
• 6:13
  the
• 6:13
  Config
• 6:14
  Store
• 6:14
  (itself
• 6:14
  a
• 6:14
  Cosmos
• 6:15
  DB
• 6:15
  database
• 6:16
  listing
• 6:16
  every
• 6:16
  account
• 6:17
  with
• 6:17
  subscription
• 6:17
  IDs,
• 6:18
  tenant
• 6:18
  IDs,
• 6:19
  and
• 6:19
  network
• 6:19
  settings),
• 6:20
  an
• 6:20
  attacker
• 6:21
  could
• 6:21
  enumerate
• 6:21
  a
• 6:21
  target
• 6:22
  organization's
• 6:22
  databases
• 6:23
  and
• 6:23
  gain
• 6:23
  full
• 6:24
  read-write
• 6:24
  access.
• 6:25
  Microsoft
• 6:26
  remediated
• 6:26
  fully,
• 6:27
  eliminated
• 6:27
  the
• 6:28
  Master
• 6:28
  Key,
• 6:28
  and
• 6:29
  found
• 6:29
  no
• 6:29
  evidence
• 6:29
  of
• 6:29
  exploitation
• 6:30
  beyond
• 6:30
  Wiz's
• 6:31
  testing.
• 6:32
  Wiz
• 6:32
  notes
• 6:32
  the
• 6:32
  research
• 6:33
  was
• 6:33
  assisted
• 6:33
  by
• 6:33
  Atlas,
• 6:34
  their
• 6:35
  AI
• 6:35
  vulnerability
• 6:36
  researcher.
• 6:37
  At
• 6:37
  the
• 6:37
  other
• 6:37
  end
• 6:37
  of
• 6:37
  critical
• 6:38
  infrastructure,
• 6:38
  iFinder
• 6:39
  —
• 6:40
  an
• 6:40
  LLM-driven
• 6:40
  multi-agent
• 6:41
  system
• 6:41
  from
• 6:42
  Nanyang
• 6:42
  Technological
• 6:43
  University
• 6:43
  —
• 6:44
  discovered
• 6:44
  84
• 6:45
  previously
• 6:45
  unknown
• 6:46
  vulnerabilities
• 6:46
  in
• 6:46
  seven
• 6:47
  open-source
• 6:47
  5G
• 6:48
  core
• 6:48
  network
• 6:48
  implementations,
• 6:49
  including
• 6:50
  Open5GS
• 6:51
  and
• 6:51
  free5GC.
• 6:52
  Of
• 6:53
  those
• 6:53
  84,
• 6:53
  83
• 6:54
  have
• 6:54
  been
• 6:54
  confirmed
• 6:55
  and
• 6:55
  81
• 6:56
  assigned
• 6:56
  CVEs.
• 6:57
  The
• 6:57
  most
• 6:58
  alarming:
• 6:58
  a
• 6:59
  session-hijacking
• 6:59
  flaw
• 7:00
  confirmed
• 7:00
  on
• 7:00
  real-world
• 7:01
  commercial
• 7:01
  5G
• 7:02
  core
• 7:02
  networks.
• 7:03
  The
• 7:04
  root
• 7:04
  cause
• 7:04
  is
• 7:04
  what
• 7:04
  the
• 7:05
  researchers
• 7:05
  call
• 7:05
  "implicit
• 7:06
  trust
• 7:06
  errors"
• 7:07
  —
• 7:07
  components
• 7:08
  that
• 7:08
  omit
• 7:08
  syntactic
• 7:09
  validation,
• 7:10
  fail
• 7:11
  to
• 7:11
  enforce
• 7:11
  semantic
• 7:12
  invariants,
• 7:13
  or
• 7:13
  allocate
• 7:13
  resources
• 7:14
  without
• 7:14
  checking
• 7:14
  availability,
• 7:15
  all
• 7:16
  because
• 7:16
  the
• 7:16
  historical
• 7:17
  security
• 7:17
  model
• 7:18
  assumed
• 7:18
  physical
• 7:19
  isolation.
• 7:20
  Cloud-native
• 7:20
  5G
• 7:21
  deployments
• 7:21
  shatter
• 7:22
  that
• 7:22
  assumption.
• 7:23
  iFinder's
• 7:24
  approach
• 7:24
  cross-checks
• 7:25
  both
• 7:25
  3GPP
• 7:26
  specifications
• 7:27
  and
• 7:27
  source
• 7:27
  code
• 7:28
  to
• 7:28
  suppress
• 7:29
  LLM
• 7:29
  hallucinations,
• 7:31
  then
• 7:31
  generates
• 7:31
  and
• 7:32
  iteratively
• 7:32
  refines
• 7:33
  proof-of-concept
• 7:33
  exploits
• 7:34
  by
• 7:34
  executing
• 7:35
  them
• 7:35
  against
• 7:35
  live
• 7:36
  CN
• 7:36
  implementations
• 7:37
  and
• 7:37
  analyzing
• 7:38
  results.
• 7:39
  The
• 7:39
  researchers
• 7:39
  plan
• 7:40
  to
• 7:40
  make
• 7:40
  iFinder
• 7:40
  publicly
• 7:41
  available
• 7:41
  —
• 7:42
  which
• 7:42
  will
• 7:43
  simultaneously
• 7:43
  improve
• 7:44
  5G
• 7:44
  security
• 7:45
  and
• 7:45
  hand
• 7:45
  a
• 7:45
  powerful
• 7:46
  automated
• 7:46
  auditing
• 7:47
  tool
• 7:47
  to
• 7:47
  anyone
• 7:47
  who
• 7:47
  wants
• 7:48
  one.
• 7:49
  On
• 7:49
  the
• 7:49
  defensive
• 7:49
  side,
• 7:50
  OpenAI
• 7:51
  shipped
• 7:51
  Codex
• 7:51
  Security
• 7:52
  as
• 7:52
  a
• 7:52
  public
• 7:53
  npm
• 7:53
  package:
• 7:54
  an
• 7:54
  SDK
• 7:55
  and
• 7:55
  CLI
• 7:56
  for
• 7:56
  AI-powered
• 7:56
  codebase
• 7:57
  scanning,
• 7:58
  though
• 7:58
  running
• 7:58
  scans
• 7:58
  still
• 7:59
  requires
• 7:59
  Codex
• 8:00
  Security
• 8:00
  access.
• 8:01
  It
• 8:01
  supports
• 8:02
  repository-wide,
• 8:03
  path-scoped,
• 8:04
  and
• 8:04
  diff-targeted
• 8:05
  scan
• 8:05
  modes,
• 8:06
  exports
• 8:06
  SARIF
• 8:07
  for
• 8:07
  integration
• 8:07
  with
• 8:08
  existing
• 8:08
  tooling,
• 8:09
  and
• 8:09
  defaults
• 8:09
  to
• 8:09
  GPT-5.6
• 8:11
  Sol,
• 8:11
  with
• 8:11
  Terra
• 8:12
  available
• 8:12
  as
• 8:12
  an
• 8:12
  override.
• 8:14
  The
• 8:14
  AI
• 8:14
  security
• 8:15
  scanning
• 8:15
  ecosystem
• 8:16
  keeps
• 8:16
  fragmenting
• 8:16
  —
• 8:17
  we've
• 8:17
  covered
• 8:18
  five
• 8:18
  competing
• 8:19
  frameworks
• 8:19
  in
• 8:19
  a
• 8:19
  single
• 8:20
  week
• 8:20
  before
• 8:20
  —
• 8:21
  and
• 8:21
  the
• 8:21
  real
• 8:21
  question
• 8:22
  remains
• 8:22
  whether
• 8:22
  any
• 8:23
  of
• 8:23
  them
• 8:23
  reliably
• 8:23
  catch
• 8:24
  what
• 8:24
  static
• 8:24
  analysis
• 8:25
  misses
• 8:25
  without
• 8:25
  generating
• 8:26
  false
• 8:26
  confidence.
• 8:30
  TurboFieldfare
• 8:31
  is
• 8:31
  a
• 8:31
  custom
• 8:32
  Swift
• 8:32
  +
• 8:32
  Metal
• 8:33
  runtime
• 8:33
  that
• 8:33
  runs
• 8:33
  Gemma
• 8:33
  4
• 8:34
  26B-A4B
• 8:35
  —
• 8:36
  a
• 8:36
  26-billion-parameter
• 8:37
  MoE
• 8:37
  model
• 8:38
  —
• 8:38
  in
• 8:38
  roughly
• 8:39
  2
• 8:39
  GB
• 8:39
  of
• 8:39
  RAM
• 8:39
  on
• 8:40
  any
• 8:40
  Apple
• 8:40
  Silicon
• 8:40
  Mac,
• 8:41
  including
• 8:41
  8
• 8:41
  GB
• 8:42
  machines.
• 8:43
  It
• 8:43
  keeps
• 8:43
  the
• 8:43
  shared
• 8:43
  1.35
• 8:44
  GB
• 8:44
  core
• 8:45
  and
• 8:45
  FP16
• 8:45
  KV
• 8:46
  cache
• 8:46
  in
• 8:46
  memory,
• 8:47
  then
• 8:47
  streams
• 8:47
  only
• 8:48
  the
• 8:48
  active
• 8:48
  experts
• 8:48
  (~14
• 8:49
  MB
• 8:49
  per
• 8:49
  load)
• 8:50
  from
• 8:50
  SSD
• 8:50
  per
• 8:51
  token.
• 8:51
  On
• 8:52
  an
• 8:52
  8
• 8:52
  GB
• 8:52
  M2
• 8:52
  MacBook
• 8:53
  Air,
• 8:53
  it
• 8:54
  measures
• 8:54
  5.1-6.3
• 8:55
  tok/s
• 8:56
  decode;
• 8:57
  on
• 8:57
  a
• 8:57
  24
• 8:57
  GB
• 8:58
  M5
• 8:58
  Pro,
• 8:59
  31-35
• 9:00
  tok/s.
• 9:01
  The
• 9:01
  runtime
• 9:01
  is
• 9:02
  model-specific
• 9:03
  rather
• 9:03
  than
• 9:03
  a
• 9:03
  wrapper
• 9:03
  around
• 9:04
  MLX
• 9:04
  or
• 9:04
  llama.cpp,
• 9:06
  built
• 9:06
  from
• 9:06
  103
• 9:07
  measured
• 9:07
  experiments
• 9:08
  across
• 9:08
  kernels,
• 9:09
  caching,
• 9:09
  I/O,
• 9:10
  prefill,
• 9:11
  and
• 9:11
  decode.
• 9:12
  This
• 9:12
  isn't
• 9:12
  theoretical
• 9:13
  —
• 9:13
  it's
• 9:13
  a
• 9:14
  working
• 9:14
  Mac
• 9:14
  app
• 9:14
  with
• 9:14
  a
• 9:14
  streaming
• 9:15
  installer.
• 9:16
  On
• 9:16
  the
• 9:16
  NPU
• 9:16
  front,
• 9:17
  STEEL
• 9:17
  delivers
• 9:17
  the
• 9:17
  first
• 9:18
  open-source
• 9:18
  FlashAttention
• 9:19
  implementation
• 9:19
  targeting
• 9:20
  AMD's
• 9:20
  XDNA
• 9:21
  architecture.
• 9:22
  Evaluated
• 9:23
  on
• 9:23
  the
• 9:23
  AMD
• 9:24
  Ryzen
• 9:24
  AI
• 9:24
  9
• 9:25
  HX
• 9:25
  370,
• 9:26
  STEEL
• 9:27
  reduces
• 9:27
  energy
• 9:27
  consumption
• 9:28
  by
• 9:28
  9.17x
• 9:29
  versus
• 9:30
  CPU
• 9:30
  and
• 9:30
  1.75x
• 9:32
  versus
• 9:32
  GPU
• 9:32
  baselines.
• 9:34
  On
• 9:34
  XDNA
• 9:34
  1,
• 9:35
  it
• 9:35
  achieves
• 9:35
  a
• 9:35
  9.6x
• 9:37
  latency
• 9:37
  reduction
• 9:37
  over
• 9:38
  prior
• 9:38
  state-of-the-art,
• 9:39
  and
• 9:39
  a
• 9:39
  22.8x
• 9:40
  speedup
• 9:40
  over
• 9:41
  layer-by-layer
• 9:41
  attention
• 9:42
  on
• 9:42
  XDNA
• 9:42
  2.
• 9:43
  The
• 9:43
  key
• 9:44
  innovation
• 9:44
  is
• 9:44
  a
• 9:44
  sparsity-aware
• 9:45
  pipeline
• 9:46
  placement
• 9:46
  that
• 9:46
  distributes
• 9:47
  consecutive
• 9:47
  rows
• 9:48
  across
• 9:48
  pipelines
• 9:49
  to
• 9:49
  equalize
• 9:49
  the
• 9:49
  workload
• 9:50
  imbalance
• 9:50
  caused
• 9:50
  by
• 9:50
  causal
• 9:51
  masking,
• 9:52
  yielding
• 9:52
  a
• 9:52
  38%
• 9:53
  speedup
• 9:53
  over
• 9:53
  naive
• 9:54
  placement.
• 9:55
  The
• 9:55
  entire
• 9:55
  stack
• 9:56
  —
• 9:56
  from
• 9:57
  Python
• 9:57
  bindings
• 9:57
  to
• 9:57
  LLVM-AIE
• 9:58
  compilation
• 9:59
  —
• 10:01
  is
• 10:01
  open
• 10:01
  source.
• 10:02
  NPUs
• 10:03
  are
• 10:03
  finally
• 10:03
  doing
• 10:03
  real
• 10:04
  inference,
• 10:04
  not
• 10:04
  just
• 10:05
  demos.
• 10:08
  OpenAI's
• 10:09
  Symphony
• 10:09
  turns
• 10:09
  project
• 10:10
  management
• 10:10
  into
• 10:11
  autonomous
• 10:11
  implementation
• 10:12
  runs:
• 10:13
  it
• 10:13
  monitors
• 10:14
  a
• 10:14
  Linear
• 10:14
  board
• 10:14
  for
• 10:14
  work,
• 10:15
  spawns
• 10:16
  Codex
• 10:16
  agents
• 10:16
  to
• 10:17
  handle
• 10:17
  tasks,
• 10:18
  and
• 10:18
  lands
• 10:18
  PRs
• 10:19
  when
• 10:19
  accepted.
• 10:20
  Engineers
• 10:21
  manage
• 10:21
  work
• 10:22
  at
• 10:22
  a
• 10:22
  higher
• 10:22
  level
• 10:22
  instead
• 10:23
  of
• 10:23
  supervising
• 10:23
  coding
• 10:24
  agents
• 10:24
  —
• 10:25
  the
• 10:25
  human-attention
• 10:26
  bottleneck
• 10:26
  remains
• 10:26
  the
• 10:26
  binding
• 10:27
  constraint
• 10:27
  even
• 10:28
  with
• 10:28
  sophisticated
• 10:28
  orchestration.
• 10:30
  The
• 10:30
  reference
• 10:30
  implementation
• 10:31
  is
• 10:31
  in
• 10:31
  Elixir,
• 10:32
  and
• 10:32
  the
• 10:32
  spec
• 10:33
  is
• 10:33
  open
• 10:33
  enough
• 10:33
  that
• 10:33
  any
• 10:34
  coding
• 10:34
  agent
• 10:34
  can
• 10:34
  build
• 10:35
  its
• 10:35
  own
• 10:35
  version.
• 10:36
  OpenAI
• 10:37
  explicitly
• 10:37
  labels
• 10:38
  it
• 10:38
  a
• 10:38
  low-key
• 10:38
  engineering
• 10:39
  preview
• 10:39
  for
• 10:39
  testing
• 10:39
  in
• 10:40
  trusted
• 10:40
  environments.
• 10:41
  Practical
• 10:41
  token-saving
• 10:42
  strategies
• 10:43
  continue
• 10:43
  to
• 10:43
  mature:
• 10:43
  a
• 10:43
  15-rule
• 10:44
  framework
• 10:44
  for
• 10:44
  reducing
• 10:45
  waste
• 10:45
  includes
• 10:46
  techniques
• 10:46
  like
• 10:46
  using
• 10:47
  cheap
• 10:47
  models
• 10:47
  as
• 10:47
  planners
• 10:48
  with
• 10:48
  expensive
• 10:48
  models
• 10:49
  as
• 10:49
  executors,
• 10:50
  which
• 10:50
  can
• 10:50
  cut
• 10:50
  costs
• 10:50
  9-10x
• 10:51
  on
• 10:51
  benchmarks
• 10:52
  like
• 10:52
  Deep
• 10:52
  SWE
• 10:52
  while
• 10:53
  maintaining
• 10:53
  the
• 10:53
  same
• 10:53
  success
• 10:54
  rate.
• 10:55
  The
• 10:55
  harder
• 10:55
  problem
• 10:55
  beneath
• 10:56
  orchestration
• 10:56
  is
• 10:56
  grounding:
• 10:57
  how
• 10:57
  do
• 10:57
  agents
• 10:58
  know
• 10:58
  what
• 10:58
  they're
• 10:58
  talking
• 10:58
  about?
• 10:59
  A
• 10:59
  Neo4j-based
• 11:01
  approach
• 11:01
  proposes
• 11:01
  a
• 11:01
  two-layer
• 11:02
  ontology
• 11:02
  —
• 11:03
  business
• 11:04
  concepts
• 11:04
  on
• 11:04
  top,
• 11:05
  technical
• 11:05
  schemas
• 11:06
  underneath
• 11:06
  —
• 11:07
  with
• 11:07
  execution
• 11:08
  traces
• 11:08
  connecting
• 11:08
  the
• 11:08
  two.
• 11:10
  The
• 11:10
  semantic
• 11:10
  layer
• 11:10
  gives
• 11:10
  agents
• 11:11
  structured
• 11:11
  context
• 11:12
  instead
• 11:12
  of
• 11:12
  relying
• 11:12
  on
• 11:12
  free-form
• 11:13
  retrieval,
• 11:14
  and
• 11:14
  execution
• 11:14
  traces
• 11:15
  provide
• 11:15
  a
• 11:15
  feedback
• 11:15
  loop
• 11:15
  for
• 11:16
  correcting
• 11:16
  misunderstandings.
• 11:17
  Frank
• 11:18
  Coyle
• 11:18
  extends
• 11:18
  this
• 11:18
  with
• 11:19
  a
• 11:19
  neuro-symbolic
• 11:19
  argument:
• 11:20
  OWL
• 11:21
  ontologies
• 11:22
  for
• 11:22
  defining
• 11:22
  valid
• 11:22
  entity
• 11:23
  relationships
• 11:23
  plus
• 11:24
  Pydantic
• 11:24
  for
• 11:24
  runtime
• 11:24
  type
• 11:25
  enforcement,
• 11:26
  creating
• 11:26
  guardrails
• 11:27
  that
• 11:27
  are
• 11:27
  deterministic
• 11:27
  rather
• 11:28
  than
• 11:28
  prompt-based.
• 11:29
  After
• 11:29
  a
• 11:29
  widely
• 11:30
  reported
• 11:30
  incident
• 11:30
  where
• 11:31
  an
• 11:31
  autonomous
• 11:31
  agent
• 11:31
  deleted
• 11:32
  200
• 11:32
  customer
• 11:33
  records
• 11:33
  because
• 11:34
  it
• 11:34
  inherited
• 11:34
  developer
• 11:34
  permissions
• 11:35
  without
• 11:35
  matching
• 11:35
  guardrails,
• 11:36
  deterministic
• 11:37
  constraint
• 11:37
  enforcement
• 11:38
  isn't
• 11:38
  academic
• 11:38
  —
• 11:39
  it's
• 11:40
  operational
• 11:40
  survival.
• 11:41
  Nostr
• 11:42
  Rust
• 11:42
  Forum
• 11:42
  takes
• 11:42
  a
• 11:42
  different
• 11:43
  tack
• 11:43
  on
• 11:43
  agent
• 11:43
  governance:
• 11:44
  human
• 11:45
  and
• 11:45
  agent
• 11:45
  participants
• 11:45
  share
• 11:46
  a
• 11:46
  mesh
• 11:46
  network
• 11:46
  with
• 11:46
  did:nostr
• 11:48
  keypairs
• 11:49
  and
• 11:49
  Schnorr-signed
• 11:50
  events,
• 11:50
  creating
• 11:51
  cryptographic
• 11:51
  accountability
• 11:52
  for
• 11:52
  every
• 11:52
  action
• 11:53
  regardless
• 11:53
  of
• 11:53
  whether
• 11:54
  the
• 11:54
  actor
• 11:54
  is
• 11:54
  human
• 11:54
  or
• 11:55
  artificial.
• 11:56
  For
• 11:56
  those
• 11:56
  still
• 11:57
  catching
• 11:57
  up
• 11:57
  on
• 11:57
  the
• 11:57
  vocabulary
• 11:57
  —
• 11:58
  agents,
• 11:59
  tools,
• 11:59
  skills,
• 12:00
  MCP,
• 12:01
  memory,
• 12:01
  permissions,
• 12:02
  orchestration
• 12:02
  —
• 12:04
  a
• 12:04
  practical
• 12:04
  glossary
• 12:05
  maps
• 12:05
  the
• 12:05
  territory
• 12:05
  with
• 12:05
  useful
• 12:06
  precision:
• 12:07
  the
• 12:07
  model
• 12:07
  provides
• 12:08
  reasoning,
• 12:08
  the
• 12:09
  agent
• 12:09
  pursues
• 12:09
  a
• 12:09
  goal,
• 12:10
  tools
• 12:10
  make
• 12:11
  actions
• 12:11
  possible,
• 12:12
  skills
• 12:12
  package
• 12:13
  working
• 12:13
  methods,
• 12:14
  MCP
• 12:14
  connects
• 12:15
  to
• 12:15
  external
• 12:15
  systems,
• 12:16
  and
• 12:17
  guardrails
• 12:17
  define
• 12:17
  boundaries.
• 12:20
  Jose
• 12:21
  Rodriguez
• 12:21
  lives
• 12:21
  in
• 12:22
  Oregon
• 12:22
  and
• 12:22
  worked
• 12:22
  in
• 12:23
  Walla
• 12:23
  Walla,
• 12:23
  Washington.
• 12:24
  His
• 12:24
  name
• 12:25
  and
• 12:25
  plate
• 12:25
  appeared
• 12:26
  in
• 12:26
  a
• 12:26
  murder
• 12:26
  investigation
• 12:27
  —
• 12:27
  one
• 12:27
  he
• 12:28
  had
• 12:28
  nothing
• 12:28
  to
• 12:28
  do
• 12:28
  with.
• 12:29
  He
• 12:29
  filed
• 12:29
  public
• 12:30
  records
• 12:30
  requests
• 12:30
  with
• 12:30
  the
• 12:31
  city
• 12:31
  to
• 12:31
  obtain
• 12:31
  Flock
• 12:32
  Safety
• 12:32
  camera
• 12:32
  data
• 12:33
  that
• 12:33
  could
• 12:33
  exonerate
• 12:33
  him.
• 12:34
  The
• 12:34
  city
• 12:34
  denied
• 12:35
  him.
• 12:35
  He
• 12:35
  filed
• 12:36
  dozens
• 12:36
  more
• 12:36
  requests
• 12:36
  across
• 12:37
  Washington
• 12:37
  state.
• 12:38
  Multiple
• 12:39
  cities
• 12:39
  banded
• 12:39
  together
• 12:40
  and
• 12:40
  sued
• 12:40
  him
• 12:40
  to
• 12:40
  stop
• 12:41
  the
• 12:41
  requests
• 12:41
  —
• 12:42
  not
• 12:42
  to
• 12:42
  protect
• 12:42
  justice,
• 12:43
  but
• 12:43
  to
• 12:43
  protect
• 12:43
  the
• 12:43
  surveillance
• 12:44
  network.
• 12:45
  Two
• 12:45
  judges
• 12:45
  sided
• 12:46
  with
• 12:46
  Rodriguez,
• 12:47
  ruling
• 12:47
  the
• 12:47
  data
• 12:47
  was
• 12:48
  public
• 12:48
  record.
• 12:49
  The
• 12:49
  cities
• 12:49
  then
• 12:49
  delayed
• 12:50
  until
• 12:50
  they
• 12:50
  passed
• 12:50
  their
• 12:50
  30-day
• 12:51
  retention
• 12:51
  policy
• 12:52
  and
• 12:52
  deleted
• 12:52
  the
• 12:52
  data.
• 12:53
  Rodriguez
• 12:53
  won
• 12:54
  in
• 12:54
  court
• 12:54
  and
• 12:54
  got
• 12:54
  nothing.
• 12:55
  On
• 12:55
  Flock's
• 12:55
  own
• 12:55
  website,
• 12:56
  a
• 12:56
  "whitelist"
• 12:57
  feature
• 12:57
  lets
• 12:57
  HOA
• 12:57
  residents
• 12:58
  opt
• 12:58
  out
• 12:58
  of
• 12:58
  searches
• 12:58
  —
• 12:59
  raising
• 12:59
  the
• 12:59
  obvious
• 13:00
  question
• 13:00
  of
• 13:00
  who
• 13:00
  else
• 13:00
  gets
• 13:00
  whitelisted.
• 13:01
  Federal
• 13:02
  bills
• 13:02
  are
• 13:02
  now
• 13:02
  circulating
• 13:03
  that
• 13:03
  would
• 13:03
  ban
• 13:03
  Flock
• 13:03
  cameras
• 13:04
  outright
• 13:04
  or
• 13:05
  strip
• 13:05
  federal
• 13:05
  funding
• 13:05
  from
• 13:06
  jurisdictions
• 13:06
  that
• 13:06
  deploy
• 13:06
  them.
• 13:07
  For
• 13:07
  a
• 13:07
  masterclass
• 13:08
  in
• 13:08
  how
• 13:08
  law
• 13:08
  enforcement
• 13:09
  can
• 13:09
  turn
• 13:09
  surveillance
• 13:09
  infrastructure
• 13:10
  against
• 13:10
  its
• 13:10
  operators,
• 13:11
  the
• 13:11
  story
• 13:12
  of
• 13:12
  Operation
• 13:12
  Bayonet
• 13:12
  remains
• 13:13
  instructive.
• 13:14
  In
• 13:14
  2017,
• 13:15
  Dutch
• 13:15
  police
• 13:16
  infiltrated
• 13:16
  and
• 13:16
  took
• 13:16
  over
• 13:17
  the
• 13:17
  Hansa
• 13:17
  darknet
• 13:17
  market,
• 13:18
  running
• 13:18
  it
• 13:18
  as
• 13:18
  a
• 13:18
  honeypot
• 13:19
  for
• 13:19
  nearly
• 13:19
  a
• 13:19
  month
• 13:20
  before
• 13:20
  the
• 13:20
  takedown.
• 13:21
  Every
• 13:22
  transaction,
• 13:23
  every
• 13:23
  message,
• 13:23
  every
• 13:24
  vendor
• 13:24
  identity
• 13:24
  —
• 13:25
  collected
• 13:26
  by
• 13:26
  the
• 13:26
  very
• 13:26
  infrastructure
• 13:27
  the
• 13:27
  criminals
• 13:27
  built
• 13:27
  to
• 13:27
  protect
• 13:28
  themselves.
• 13:29
  The
• 13:29
  buyers
• 13:30
  of
• 13:30
  Hansa's
• 13:30
  drugs
• 13:30
  and
• 13:30
  services
• 13:31
  built
• 13:31
  their
• 13:31
  own
• 13:31
  panopticon,
• 13:32
  and
• 13:32
  the
• 13:33
  Dutch
• 13:33
  police
• 13:33
  simply
• 13:33
  moved
• 13:34
  in.
• 13:35
  The
• 13:35
  parallel
• 13:35
  to
• 13:35
  Flock
• 13:35
  is
• 13:36
  uncomfortable:
• 13:37
  surveillance
• 13:37
  infrastructure
• 13:38
  serves
• 13:38
  whoever
• 13:39
  controls
• 13:39
  it,
• 13:39
  and
• 13:39
  the
• 13:39
  question
• 13:40
  of
• 13:40
  who
• 13:40
  controls
• 13:40
  it
• 13:40
  is
• 13:41
  never
• 13:41
  as
• 13:41
  settled
• 13:41
  as
• 13:41
  the
• 13:41
  people
• 13:42
  being
• 13:42
  surveilled
• 13:42
  assume.
• 13:44
  Whether
• 13:44
  the
• 13:44
  cameras
• 13:44
  are
• 13:44
  on
• 13:45
  street
• 13:45
  corners
• 13:45
  or
• 13:46
  in
• 13:46
  encrypted
• 13:46
  marketplaces,
• 13:47
  the
• 13:47
  same
• 13:48
  principle
• 13:48
  holds
• 13:48
  —
• 13:49
  the
• 13:50
  tool
• 13:50
  is
• 13:50
  neutral,
• 13:51
  but
• 13:51
  control
• 13:51
  never
• 13:51
  is.
• 13:55
  A
• 13:55
  Reddit
• 13:55
  user
• 13:56
  was
• 13:56
  working
• 13:56
  with
• 13:56
  Claude
• 13:56
  via
• 13:57
  speech-to-text
• 13:57
  when
• 13:57
  they
• 13:58
  found
• 13:58
  they
• 13:58
  couldn't
• 13:58
  speak
• 13:59
  properly
• 13:59
  —
• 14:00
  couldn't
• 14:00
  think
• 14:00
  of
• 14:00
  words,
• 14:01
  started
• 14:01
  producing
• 14:01
  gibberish.
• 14:02
  Claude
• 14:03
  recognized
• 14:03
  the
• 14:03
  symptoms
• 14:04
  and
• 14:04
  insisted
• 14:04
  they
• 14:04
  call
• 14:05
  an
• 14:05
  ambulance.
• 14:06
  The
• 14:06
  user
• 14:06
  compromised
• 14:07
  with
• 14:07
  an
• 14:07
  Uber
• 14:07
  to
• 14:07
  the
• 14:07
  ER.
• 14:08
  The
• 14:08
  hospital
• 14:09
  rushed
• 14:09
  them
• 14:09
  through
• 14:09
  in
• 14:09
  two
• 14:10
  minutes:
• 14:10
  potential
• 14:11
  stroke.
• 14:11
  CT
• 14:12
  scan,
• 14:12
  MRI,
• 14:13
  ECG,
• 14:14
  bloods.
• 14:15
  Diagnosis:
• 14:16
  TIA
• 14:16
  (transient
• 14:17
  ischemic
• 14:17
  attack),
• 14:18
  a
• 14:18
  mini-stroke
• 14:18
  that
• 14:18
  frequently
• 14:19
  precedes
• 14:19
  a
• 14:19
  full
• 14:20
  stroke.
• 14:21
  Twenty-four
• 14:21
  hours
• 14:21
  later,
• 14:22
  back
• 14:22
  home.
• 14:23
  The
• 14:23
  community
• 14:23
  response
• 14:24
  was
• 14:24
  overwhelmingly
• 14:25
  positive,
• 14:25
  with
• 14:26
  another
• 14:26
  user
• 14:26
  reporting
• 14:27
  a
• 14:27
  nearly
• 14:27
  identical
• 14:27
  experience
• 14:28
  —
• 14:29
  Claude
• 14:29
  flagged
• 14:29
  their
• 14:30
  sudden
• 14:30
  inability
• 14:30
  to
• 14:31
  read
• 14:31
  and
• 14:31
  loss
• 14:31
  of
• 14:31
  peripheral
• 14:32
  vision,
• 14:32
  and
• 14:32
  they
• 14:32
  were
• 14:33
  also
• 14:33
  diagnosed
• 14:33
  with
• 14:34
  TIA.
• 14:35
  The
• 14:35
  counterpoint
• 14:35
  that
• 14:35
  strokes
• 14:36
  literally
• 14:36
  impair
• 14:37
  your
• 14:37
  judgment
• 14:37
  (anosognosia)
• 14:38
  makes
• 14:38
  AI
• 14:38
  detection
• 14:39
  in
• 14:39
  the
• 14:39
  moment
• 14:39
  more
• 14:39
  valuable,
• 14:40
  not
• 14:40
  less.
• 14:43
  Ilya
• 14:43
  Sutskever,
• 14:44
  in
• 14:44
  a
• 14:44
  talk
• 14:45
  on
• 14:45
  unsupervised
• 14:46
  learning
• 14:46
  theory,
• 14:47
  makes
• 14:47
  the
• 14:47
  formal
• 14:47
  case
• 14:47
  for
• 14:48
  why
• 14:48
  compression
• 14:48
  is
• 14:48
  the
• 14:49
  foundation
• 14:49
  of
• 14:49
  learning.
• 14:50
  The
• 14:50
  core
• 14:51
  thought
• 14:51
  experiment:
• 14:52
  concatenate
• 14:52
  two
• 14:53
  datasets
• 14:53
  and
• 14:53
  compress
• 14:54
  them
• 14:54
  jointly.
• 14:55
  A
• 14:55
  sufficiently
• 14:55
  good
• 14:56
  compressor
• 14:56
  will
• 14:56
  use
• 14:56
  patterns
• 14:57
  in
• 14:57
  one
• 14:57
  to
• 14:57
  help
• 14:57
  compress
• 14:58
  the
• 14:58
  other
• 14:58
  —
• 14:59
  that
• 14:59
  shared
• 14:59
  structure
• 15:00
  is
• 15:00
  the
• 15:00
  signal
• 15:00
  unsupervised
• 15:01
  learning
• 15:01
  extracts.
• 15:02
  Kolmogorov
• 15:03
  complexity
• 15:03
  provides
• 15:04
  the
• 15:04
  ultimate
• 15:04
  (uncomputable)
• 15:05
  low-regret
• 15:05
  compressor.
• 15:06
  Sutskever's
• 15:07
  thesis
• 15:07
  is
• 15:07
  that
• 15:08
  SGD
• 15:08
  over
• 15:08
  deep
• 15:09
  neural
• 15:09
  networks
• 15:09
  approximates
• 15:10
  that
• 15:10
  ideal
• 15:10
  and
• 15:10
  that
• 15:11
  bigger
• 15:11
  networks
• 15:11
  get
• 15:11
  closer
• 15:12
  —
• 15:12
  an
• 15:12
  explanation
• 15:13
  for
• 15:13
  why
• 15:13
  scaling
• 15:14
  works,
• 15:14
  not
• 15:14
  a
• 15:14
  proof
• 15:15
  that
• 15:15
  it
• 15:15
  must.
• 15:16
  The
• 15:16
  iGPT
• 15:16
  experiment
• 15:17
  (2020)
• 15:17
  gave
• 15:18
  the
• 15:18
  argument
• 15:18
  empirical
• 15:18
  support
• 15:19
  in
• 15:19
  the
• 15:19
  image
• 15:19
  domain:
• 15:20
  next-pixel
• 15:21
  prediction
• 15:21
  with
• 15:21
  the
• 15:21
  same
• 15:22
  Transformer
• 15:22
  architecture
• 15:23
  as
• 15:23
  GPT
• 15:24
  produced
• 15:24
  useful
• 15:25
  representations,
• 15:26
  reaching
• 15:26
  96.3%
• 15:28
  on
• 15:28
  CIFAR-10
• 15:32
  with
• 15:32
  a
• 15:32
  linear
• 15:32
  probe
• 15:32
  and
• 15:32
  99.0%
• 15:34
  after
• 15:34
  full
• 15:34
  fine-tuning.
• 15:35
  The
• 15:35
  practical
• 15:36
  implication,
• 15:36
  in
• 15:36
  Sutskever's
• 15:37
  framing:
• 15:38
  any
• 15:38
  procedure
• 15:38
  that
• 15:38
  maximizes
• 15:39
  the
• 15:39
  likelihood
• 15:39
  of
• 15:40
  data
• 15:40
  —
• 15:40
  autoregressive
• 15:41
  models,
• 15:42
  diffusion
• 15:42
  models,
• 15:43
  energy-based
• 15:44
  models
• 15:44
  —
• 15:44
  should
• 15:45
  in
• 15:45
  principle
• 15:45
  produce
• 15:45
  useful
• 15:46
  representations.
• 15:47
  The
• 15:47
  difference
• 15:47
  between
• 15:48
  them
• 15:48
  is
• 15:48
  compute
• 15:48
  efficiency,
• 15:49
  not
• 15:49
  information
• 15:50
  content.