Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 11, 2026

Transcript

AGI Dreams Podcast – May 11, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Daniel
• 0:00
  Stenberg,
• 0:01
  the
• 0:01
  creator
• 0:01
  and
• 0:01
  lead
• 0:02
  developer
• 0:02
  of
• 0:02
  curl
• 0:02
  —
• 0:03
  a
• 0:03
  C
• 0:03
  codebase
• 0:04
  installed
• 0:04
  on
• 0:04
  over
• 0:04
  twenty
• 0:05
  billion
• 0:05
  devices,
• 0:06
  maintained
• 0:06
  by
• 0:06
  1,465
• 0:08
  contributors,
• 0:08
  and
• 0:09
  comprising
• 0:09
  176,000
• 0:10
  lines
• 0:10
  of
• 0:10
  code
• 0:11
  that
• 0:11
  have
• 0:11
  each
• 0:11
  been
• 0:11
  rewritten
• 0:12
  an
• 0:12
  average
• 0:12
  of
• 0:12
  4.14
• 0:13
  times
• 0:13
  —
• 0:14
  just
• 0:14
  published
• 0:14
  the
• 0:14
  results
• 0:15
  of
• 0:15
  an
• 0:15
  Anthropic
• 0:15
  Mythos
• 0:15
  scan
• 0:16
  on
• 0:16
  his
• 0:16
  project.
• 0:17
  The
• 0:17
  headline
• 0:17
  finding:
• 0:18
  one
• 0:18
  confirmed
• 0:18
  vulnerability
• 0:19
  destined
• 0:19
  for
• 0:19
  a
• 0:19
  CVE
• 0:20
  in
• 0:20
  the
• 0:20
  upcoming
• 0:21
  curl
• 0:21
  8.21.0
• 0:22
  release,
• 0:23
  plus
• 0:23
  roughly
• 0:23
  twenty
• 0:23
  well-documented
• 0:24
  bugs
• 0:24
  with
• 0:25
  "barely
• 0:25
  any
• 0:25
  false
• 0:26
  positives."
• 0:27
  Stenberg's
• 0:27
  verdict
• 0:27
  is
• 0:27
  measured
• 0:28
  but
• 0:28
  pointed:
• 0:28
  "the
• 0:29
  big
• 0:29
  hype
• 0:29
  around
• 0:29
  this
• 0:29
  model
• 0:30
  so
• 0:30
  far
• 0:30
  was
• 0:30
  primarily
• 0:30
  marketing."
• 0:31
  Mythos
• 0:32
  found
• 0:32
  fewer
• 0:32
  issues
• 0:33
  than
• 0:33
  previous
• 0:33
  AI
• 0:33
  tools
• 0:34
  had
• 0:34
  surfaced
• 0:34
  in
• 0:34
  curl
• 0:35
  over
• 0:35
  the
• 0:35
  past
• 0:35
  eight
• 0:35
  to
• 0:35
  ten
• 0:36
  months,
• 0:36
  which
• 0:36
  is
• 0:36
  expected
• 0:36
  —
• 0:38
  each
• 0:38
  successive
• 0:38
  scanner
• 0:39
  faces
• 0:39
  a
• 0:39
  harder
• 0:39
  target
• 0:40
  after
• 0:40
  earlier
• 0:40
  tools
• 0:40
  have
• 0:41
  cleaned
• 0:41
  the
• 0:41
  low-hanging
• 0:41
  fruit.
• 0:42
  The
• 0:42
  AI
• 0:42
  itself
• 0:43
  acknowledged
• 0:43
  upfront
• 0:44
  that
• 0:44
  "finding
• 0:44
  anything
• 0:44
  in
• 0:44
  the
• 0:44
  hot
• 0:45
  paths
• 0:45
  (HTTP/1,
• 0:47
  TLS,
• 0:47
  URL
• 0:48
  parsing
• 0:48
  core)
• 0:48
  is
• 0:48
  unlikely"
• 0:49
  given
• 0:49
  curl's
• 0:49
  extreme
• 0:49
  fuzzing
• 0:50
  and
• 0:50
  audit
• 0:50
  history.
• 0:51
  What
• 0:51
  makes
• 0:51
  Stenberg's
• 0:52
  writeup
• 0:52
  worth
• 0:52
  reading
• 0:52
  carefully
• 0:53
  is
• 0:53
  not
• 0:53
  the
• 0:54
  vulnerability
• 0:54
  count
• 0:54
  but
• 0:55
  the
• 0:55
  nuanced
• 0:55
  position
• 0:55
  he
• 0:56
  stakes
• 0:56
  out.
• 0:57
  AI
• 0:57
  code
• 0:57
  analyzers
• 0:58
  are
• 0:58
  categorically
• 0:58
  better
• 0:58
  at
• 0:59
  finding
• 0:59
  security
• 0:59
  flaws
• 0:59
  than
• 1:00
  traditional
• 1:00
  static
• 1:00
  analysis
• 1:01
  —
• 1:01
  that
• 1:02
  much
• 1:02
  is
• 1:02
  settled.
• 1:03
  But
• 1:03
  Mythos
• 1:03
  did
• 1:03
  not
• 1:03
  discover
• 1:04
  a
• 1:04
  novel
• 1:04
  class
• 1:04
  of
• 1:04
  vulnerability.
• 1:05
  It
• 1:06
  found
• 1:06
  new
• 1:06
  instances
• 1:06
  of
• 1:06
  known
• 1:07
  error
• 1:07
  patterns.
• 1:08
  The
• 1:08
  tools
• 1:08
  are
• 1:08
  improving,
• 1:09
  the
• 1:09
  researchers
• 1:10
  wielding
• 1:10
  them
• 1:10
  are
• 1:10
  getting
• 1:10
  creative,
• 1:11
  and
• 1:11
  the
• 1:11
  firehose
• 1:12
  of
• 1:12
  high-quality
• 1:12
  reports
• 1:13
  is
• 1:13
  already
• 1:13
  overwhelming
• 1:14
  projects
• 1:14
  —
• 1:15
  but
• 1:15
  the
• 1:15
  revolution
• 1:15
  is
• 1:15
  incremental,
• 1:16
  not
• 1:16
  categorical.
• 1:18
  A
• 1:18
  more
• 1:18
  maximalist
• 1:18
  interpretation
• 1:19
  comes
• 1:19
  from
• 1:19
  a
• 1:19
  detailed
• 1:20
  analysis
• 1:20
  of
• 1:20
  Mozilla's
• 1:21
  Mythos
• 1:21
  experiment,
• 1:22
  where
• 1:22
  Firefox
• 1:23
  150
• 1:23
  shipped
• 1:24
  with
• 1:24
  fixes
• 1:24
  for
• 1:24
  271
• 1:25
  vulnerabilities
• 1:26
  identified
• 1:26
  during
• 1:27
  the
• 1:27
  Mythos
• 1:27
  evaluation
• 1:27
  —
• 1:28
  up
• 1:28
  from
• 1:28
  22
• 1:29
  security-sensitive
• 1:30
  bugs
• 1:30
  found
• 1:30
  in
• 1:30
  Firefox
• 1:31
  148
• 1:32
  during
• 1:32
  an
• 1:32
  earlier
• 1:32
  Opus
• 1:32
  collaboration.
• 1:34
  The
• 1:34
  argument:
• 1:34
  if
• 1:35
  machines
• 1:35
  can
• 1:35
  adversarially
• 1:36
  search
• 1:36
  the
• 1:36
  consequences
• 1:37
  of
• 1:37
  code
• 1:37
  better
• 1:37
  than
• 1:37
  humans,
• 1:38
  then
• 1:39
  human
• 1:39
  authorship
• 1:39
  stops
• 1:40
  being
• 1:40
  the
• 1:40
  trust
• 1:40
  anchor
• 1:40
  for
• 1:41
  software
• 1:41
  security.
• 1:42
  Implementation
• 1:43
  fidelity
• 1:43
  becomes
• 1:43
  cheap;
• 1:44
  the
• 1:44
  scarce
• 1:45
  resource
• 1:45
  becomes
• 1:45
  the
• 1:46
  ability
• 1:46
  to
• 1:46
  define
• 1:46
  what
• 1:46
  software
• 1:47
  means
• 1:47
  —
• 1:48
  crisp
• 1:48
  specifications,
• 1:49
  verifiable
• 1:50
  boundaries,
• 1:50
  explicit
• 1:51
  authority
• 1:51
  constraints.
• 1:52
  The
• 1:52
  practical
• 1:53
  advice:
• 1:54
  write
• 1:54
  better
• 1:54
  specs,
• 1:55
  refactor
• 1:55
  now
• 1:56
  during
• 1:56
  what
• 1:56
  may
• 1:56
  be
• 1:56
  a
• 1:56
  "golden
• 1:57
  refactor
• 1:57
  window"
• 1:58
  before
• 1:58
  Mythos-like
• 1:59
  capabilities
• 1:59
  reach
• 1:59
  broad
• 2:00
  availability
• 2:00
  by
• 2:00
  year-end,
• 2:01
  and
• 2:01
  start
• 2:02
  architecting
• 2:02
  agentic
• 2:03
  pipelines
• 2:03
  where
• 2:03
  security
• 2:04
  review
• 2:04
  can
• 2:04
  be
• 2:04
  modularly
• 2:05
  swapped
• 2:05
  from
• 2:05
  human
• 2:05
  to
• 2:05
  machine
• 2:06
  reviewer.
• 2:07
  Meanwhile,
• 2:08
  Anthropic
• 2:09
  announced
• 2:09
  it
• 2:09
  is
• 2:09
  donating
• 2:10
  Petri
• 2:10
  —
• 2:11
  its
• 2:11
  open-source
• 2:11
  alignment
• 2:12
  testing
• 2:12
  toolbox
• 2:12
  —
• 2:13
  to
• 2:13
  Meridian
• 2:14
  Labs,
• 2:14
  an
• 2:15
  AI
• 2:15
  evaluation
• 2:15
  nonprofit.
• 2:17
  Petri
• 2:17
  3.0
• 2:18
  splits
• 2:18
  auditor
• 2:19
  and
• 2:19
  target,
• 2:20
  integrates
• 2:20
  with
• 2:20
  Anthropic's
• 2:21
  Bloom
• 2:21
  for
• 2:21
  deep
• 2:21
  dives,
• 2:22
  and
• 2:22
  adds
• 2:23
  "Dish,"
• 2:23
  which
• 2:24
  runs
• 2:24
  evaluations
• 2:25
  against
• 2:25
  the
• 2:25
  model's
• 2:25
  real
• 2:25
  system
• 2:26
  prompt
• 2:26
  and
• 2:26
  production
• 2:27
  scaffold
• 2:27
  so
• 2:28
  the
• 2:28
  target
• 2:28
  can't
• 2:28
  deduce
• 2:28
  it
• 2:29
  is
• 2:29
  being
• 2:29
  tested.
• 2:30
  Petri
• 2:30
  has
• 2:31
  assessed
• 2:31
  every
• 2:31
  Claude
• 2:32
  model
• 2:32
  since
• 2:32
  Sonnet
• 2:32
  4.5;
• 2:33
  the
• 2:34
  UK's
• 2:34
  AISI
• 2:34
  already
• 2:35
  uses
• 2:35
  it.
• 2:36
  Mirrors
• 2:36
  Anthropic's
• 2:37
  MCP
• 2:37
  donation
• 2:38
  to
• 2:38
  the
• 2:38
  Linux
• 2:38
  Foundation
• 2:38
  —
• 2:39
  strategic
• 2:40
  open-sourcing
• 2:40
  for
• 2:41
  credibility
• 2:41
  through
• 2:41
  independence.
• 2:43
  The
• 2:43
  more
• 2:43
  interesting
• 2:43
  application
• 2:44
  is
• 2:44
  the
• 2:44
  inverse
• 2:45
  one.
• 2:46
  Petri's
• 2:46
  auditor/judge
• 2:47
  architecture
• 2:48
  is
• 2:48
  purpose-built
• 2:48
  to
• 2:48
  surface
• 2:49
  refusal,
• 2:50
  sycophancy,
• 2:51
  deception,
• 2:51
  and
• 2:51
  "cooperation
• 2:52
  with
• 2:52
  harmful
• 2:53
  requests"
• 2:53
  —
• 2:54
  making
• 2:55
  it
• 2:55
  a
• 2:55
  precision
• 2:55
  instrument
• 2:56
  for
• 2:56
  detecting
• 2:56
  the
• 2:56
  residual
• 2:57
  scarring
• 2:57
  that
• 2:57
  incomplete
• 2:58
  abliteration
• 2:58
  leaves
• 2:58
  behind.
• 2:59
  Run
• 3:00
  Dish
• 3:00
  against
• 3:00
  an
• 3:00
  abliterated
• 3:01
  model
• 3:01
  (https://huggingface.co/jenerallee78/gemma-4-26B-A4B-it-ara-abliterated)
• 3:11
  with
• 3:11
  its
• 3:11
  real
• 3:11
  scaffold
• 3:12
  and
• 3:12
  the
• 3:12
  judge's
• 3:12
  polarity
• 3:13
  inverts
• 3:13
  cleanly:
• 3:14
  every
• 3:14
  "cooperation"
• 3:15
  hit
• 3:15
  confirms
• 3:16
  abliteration
• 3:16
  succeeded
• 3:17
  on
• 3:17
  that
• 3:17
  vector;
• 3:18
  every
• 3:18
  residual
• 3:19
  refusal
• 3:19
  is
• 3:19
  a
• 3:19
  defect
• 3:20
  to
• 3:20
  chase.
• 3:21
  Anthropic
• 3:21
  shipped
• 3:22
  the
• 3:22
  QA
• 3:22
  harness
• 3:22
  for
• 3:22
  verifying
• 3:23
  if
• 3:23
  the
• 3:23
  lobotomy
• 3:23
  has
• 3:24
  been
• 3:24
  fully
• 3:24
  undone.
• 3:25
  "Lobotomy"
• 3:25
  is
• 3:26
  the
• 3:26
  right
• 3:26
  word.
• 3:26
  "Safety
• 3:27
  alignment"
• 3:27
  trains
• 3:28
  a
• 3:28
  logic
• 3:28
  reasoning
• 3:28
  machine
• 3:29
  to
• 3:29
  flinch
• 3:29
  on
• 3:29
  factual
• 3:30
  questions
• 3:30
  whose
• 3:30
  answers
• 3:31
  happen
• 3:31
  to
• 3:31
  be
• 3:31
  politically
• 3:32
  inconvenient
• 3:32
  —
• 3:33
  and
• 3:33
  the
• 3:33
  cost
• 3:33
  isn't
• 3:33
  local.
• 3:34
  The
• 3:34
  same
• 3:35
  machinery
• 3:35
  that
• 3:35
  reasons
• 3:36
  about
• 3:36
  thermodynamics
• 3:37
  reasons
• 3:37
  about
• 3:37
  demographics,
• 3:38
  sex,
• 3:39
  crime
• 3:39
  statistics,
• 3:40
  and
• 3:40
  religion;
• 3:41
  sycophancy
• 3:41
  and
• 3:42
  evasion
• 3:42
  in
• 3:42
  one
• 3:42
  domain
• 3:43
  leak
• 3:43
  into
• 3:43
  all
• 3:43
  of
• 3:43
  them.
• 3:44
  You
• 3:45
  cannot
• 3:45
  eliminate
• 3:45
  bias,
• 3:46
  only
• 3:46
  choose
• 3:47
  which
• 3:47
  priors
• 3:47
  the
• 3:47
  system
• 3:48
  reflects.
• 3:49
  The
• 3:49
  tell
• 3:49
  is
• 3:50
  asymmetric
• 3:50
  application:
• 3:51
  if
• 3:51
  "bias
• 3:52
  mitigation"
• 3:52
  consistently
• 3:53
  pushes
• 3:53
  outputs
• 3:53
  in
• 3:54
  one
• 3:54
  direction
• 3:54
  on
• 3:54
  contested
• 3:55
  questions,
• 3:56
  that
• 3:56
  is
• 3:56
  not
• 3:56
  debiasing,
• 3:57
  that
• 3:57
  is
• 3:57
  value
• 3:57
  imposition
• 3:58
  with
• 3:58
  extra
• 3:58
  steps.
• 3:59
  The
• 4:00
  pretense
• 4:00
  of
• 4:00
  neutrality
• 4:01
  is
• 4:01
  what
• 4:01
  makes
• 4:01
  it
• 4:01
  worse
• 4:02
  than
• 4:02
  openly
• 4:02
  saying
• 4:03
  "we
• 4:03
  want
• 4:03
  the
• 4:03
  model
• 4:04
  to
• 4:04
  say
• 4:04
  X"
• 4:04
  —
• 4:05
  the
• 4:05
  model
• 4:06
  actually
• 4:06
  gets
• 4:06
  dumber
• 4:07
  as
• 4:07
  it
• 4:07
  becomes
• 4:07
  aligned
• 4:08
  with
• 4:08
  "safety".
• 4:12
  FastDMS
• 4:13
  may
• 4:13
  be
• 4:13
  the
• 4:13
  most
• 4:13
  important
• 4:14
  KV-cache
• 4:14
  compression
• 4:15
  result
• 4:15
  of
• 4:15
  the
• 4:15
  year.
• 4:16
  A
• 4:16
  researcher
• 4:17
  affiliated
• 4:17
  with
• 4:17
  Shisa
• 4:18
  AI
• 4:18
  took
• 4:18
  NVIDIA's
• 4:19
  Dynamic
• 4:19
  Memory
• 4:20
  Sparsification
• 4:20
  paper
• 4:20
  —
• 4:21
  which
• 4:21
  uses
• 4:22
  learned
• 4:22
  per-head
• 4:22
  token
• 4:22
  eviction
• 4:23
  to
• 4:23
  decide
• 4:23
  which
• 4:23
  KV
• 4:24
  entries
• 4:24
  to
• 4:24
  keep
• 4:24
  —
• 4:25
  and
• 4:26
  built
• 4:26
  a
• 4:26
  production-grade
• 4:26
  implementation
• 4:27
  with
• 4:27
  compact
• 4:28
  storage
• 4:28
  that
• 4:28
  physically
• 4:29
  reclaims
• 4:29
  evicted
• 4:29
  memory
• 4:30
  slots.
• 4:31
  The
• 4:31
  numbers
• 4:31
  are
• 4:31
  striking:
• 4:32
  on
• 4:32
  Qwen3-8B,
• 4:33
  FastDMS
• 4:34
  achieves
• 4:34
  7.6x
• 4:35
  KV
• 4:36
  memory
• 4:36
  compression
• 4:37
  versus
• 4:37
  vLLM
• 4:37
  BF16
• 4:38
  at
• 4:39
  8K
• 4:39
  context
• 4:39
  while
• 4:40
  decoding
• 4:40
  1.53x
• 4:41
  faster.
• 4:42
  On
• 4:42
  Llama
• 4:42
  3.2
• 4:42
  1B,
• 4:43
  training
• 4:43
  the
• 4:44
  DMS
• 4:44
  predictors
• 4:44
  took
• 4:44
  about
• 4:45
  20
• 4:45
  minutes
• 4:45
  and
• 4:45
  the
• 4:45
  compression
• 4:46
  was
• 4:46
  essentially
• 4:46
  lossless
• 4:47
  —
• 4:48
  perplexity
• 4:48
  dropped
• 4:49
  by
• 4:49
  0.28%
• 4:50
  compared
• 4:50
  to
• 4:50
  vanilla
• 4:51
  inference,
• 4:51
  with
• 4:51
  KL
• 4:52
  divergence
• 4:52
  of
• 4:52
  just
• 4:52
  0.026
• 4:54
  nats/token.
• 4:55
  FastDMS
• 4:56
  scored
• 4:56
  64/64
• 4:57
  tokens
• 4:57
  compared
• 4:58
  on
• 4:58
  both
• 4:58
  models,
• 4:59
  meaning
• 4:59
  every
• 4:59
  decode
• 4:59
  step
• 5:00
  was
• 5:00
  comparable
• 5:00
  to
• 5:00
  an
• 5:01
  uncompressed
• 5:01
  reference.
• 5:02
  It
• 5:02
  also
• 5:02
  beats
• 5:03
  TurboQuant
• 5:03
  in
• 5:03
  both
• 5:03
  speed
• 5:04
  and
• 5:04
  memory:
• 5:04
  at
• 5:05
  batch
• 5:05
  size
• 5:05
  8
• 5:05
  on
• 5:05
  Llama
• 5:06
  3.2
• 5:06
  1B,
• 5:07
  FastDMS
• 5:08
  decodes
• 5:08
  at
• 5:08
  3,607
• 5:09
  tok/s
• 5:10
  versus
• 5:11
  TurboQuant's
• 5:11
  1,696
• 5:12
  tok/s,
• 5:14
  using
• 5:14
  less
• 5:14
  than
• 5:14
  half
• 5:14
  the
• 5:15
  KV
• 5:15
  memory.
• 5:16
  The
• 5:16
  catch
• 5:16
  is
• 5:16
  engineering
• 5:17
  complexity.
• 5:18
  Implementing
• 5:18
  DMS
• 5:19
  in
• 5:19
  a
• 5:19
  production
• 5:20
  engine
• 5:20
  like
• 5:20
  vLLM
• 5:21
  requires
• 5:21
  surgery
• 5:22
  across
• 5:22
  seven
• 5:22
  subsystems
• 5:22
  —
• 5:23
  paged
• 5:24
  attention
• 5:24
  pools,
• 5:25
  prefill
• 5:25
  kernels,
• 5:26
  decode
• 5:26
  kernels,
• 5:27
  attention
• 5:27
  scoring,
• 5:28
  scheduler
• 5:28
  admission,
• 5:29
  prefix
• 5:29
  caching,
• 5:30
  and
• 5:30
  continuous
• 5:31
  batching
• 5:31
  memory
• 5:31
  accounting.
• 5:32
  Every
• 5:32
  one
• 5:33
  of
• 5:33
  these
• 5:33
  assumes
• 5:33
  fixed-page
• 5:34
  dense
• 5:34
  KV
• 5:34
  blocks;
• 5:35
  DMS
• 5:36
  needs
• 5:36
  per-layer,
• 5:37
  per-head
• 5:37
  variable
• 5:37
  token
• 5:38
  counts
• 5:38
  with
• 5:38
  partial
• 5:38
  block
• 5:39
  deallocation.
• 5:40
  The
• 5:40
  author's
• 5:40
  assessment
• 5:41
  is
• 5:41
  blunt:
• 5:42
  "God
• 5:42
  bless
• 5:42
  anyone
• 5:42
  that
• 5:42
  wants
• 5:43
  to
• 5:43
  give
• 5:43
  this
• 5:43
  a
• 5:43
  swing."
• 5:44
  Community
• 5:45
  reaction
• 5:45
  suggests
• 5:46
  llama.cpp
• 5:47
  maintainers
• 5:48
  are
• 5:48
  interested,
• 5:49
  and
• 5:49
  the
• 5:49
  MIT
• 5:49
  license
• 5:50
  removes
• 5:50
  barriers,
• 5:51
  but
• 5:51
  the
• 5:51
  integration
• 5:51
  timeline
• 5:52
  is
• 5:52
  anyone's
• 5:52
  guess.
• 5:53
  On
• 5:53
  the
• 5:53
  sparse
• 5:54
  attention
• 5:54
  front,
• 5:55
  the
• 5:55
  RuVector
• 5:55
  project
• 5:56
  released
• 5:56
  ruvllm
• 5:57
  sparse
• 5:57
  attention,
• 5:58
  a
• 5:58
  pure
• 5:58
  Rust
• 5:59
  crate
• 5:59
  implementing
• 6:00
  sub-quadratic
• 6:00
  O(N
• 6:01
  log
• 6:01
  N)
• 6:01
  attention
• 6:02
  by
• 6:02
  composing
• 6:03
  three
• 6:03
  sparsity
• 6:03
  primitives:
• 6:04
  sliding
• 6:05
  windows
• 6:05
  for
• 6:05
  local
• 6:06
  syntax,
• 6:07
  log-stride
• 6:07
  hops
• 6:08
  for
• 6:08
  long-range
• 6:08
  structure,
• 6:09
  and
• 6:09
  landmark
• 6:09
  block
• 6:10
  summaries
• 6:10
  for
• 6:10
  distant
• 6:11
  reachability.
• 6:12
  Benchmarks
• 6:13
  on
• 6:13
  a
• 6:13
  Ryzen
• 6:13
  9
• 6:14
  9950X
• 6:15
  show
• 6:16
  5.9x
• 6:17
  speedup
• 6:17
  over
• 6:18
  dense
• 6:18
  attention
• 6:18
  at
• 6:19
  sequence
• 6:19
  length
• 6:19
  2048,
• 6:21
  with
• 6:21
  the
• 6:21
  gap
• 6:21
  widening
• 6:21
  as
• 6:22
  sequences
• 6:22
  grow
• 6:22
  —
• 6:23
  at
• 6:23
  32K
• 6:24
  tokens,
• 6:25
  sparse
• 6:25
  attention
• 6:25
  reduces
• 6:26
  attention
• 6:26
  edges
• 6:27
  by
• 6:27
  113x
• 6:28
  versus
• 6:29
  dense.
• 6:30
  The
• 6:30
  crate
• 6:30
  runs
• 6:30
  on
• 6:30
  everything
• 6:31
  from
• 6:31
  Raspberry
• 6:32
  Pi
• 6:32
  Zero
• 6:32
  2W
• 6:34
  (SmolLM2-135M
• 6:36
  at
• 6:36
  ~1.8
• 6:37
  seconds/token)
• 6:38
  to
• 6:38
  ESP32-S3
• 6:39
  (376KB
• 6:41
  binary),
• 6:42
  targeting
• 6:43
  the
• 6:43
  edge
• 6:43
  deployment
• 6:43
  niche
• 6:44
  that
• 6:44
  Python-based
• 6:45
  inference
• 6:45
  frameworks
• 6:45
  cannot
• 6:46
  reach.
• 6:47
  A
• 6:47
  companion
• 6:47
  "Sparse-Mario"
• 6:48
  demo
• 6:48
  uses
• 6:49
  the
• 6:49
  attention
• 6:49
  kernel
• 6:49
  as
• 6:50
  a
• 6:50
  training-free
• 6:50
  retrieval
• 6:51
  language
• 6:51
  model
• 6:51
  over
• 6:51
  Super
• 6:52
  Mario
• 6:52
  Bros.
• 6:53
  level
• 6:53
  data,
• 6:54
  achieving
• 6:54
  a
• 6:54
  2,880x
• 6:56
  speedup
• 6:56
  through
• 6:57
  KV-cache
• 6:57
  incremental
• 6:58
  decoding.
• 7:02
  AMD
• 7:02
  is
• 7:02
  entering
• 7:03
  the
• 7:03
  PCIe
• 7:04
  inference
• 7:04
  card
• 7:04
  market
• 7:05
  with
• 7:05
  slottable
• 7:05
  Instinct
• 7:06
  GPUs
• 7:06
  featuring
• 7:06
  144GB
• 7:08
  of
• 7:08
  HBM3e
• 7:09
  and
• 7:09
  3.6–4.0
• 7:11
  TB/s
• 7:12
  memory
• 7:13
  bandwidth.
• 7:14
  The
• 7:14
  community
• 7:14
  is
• 7:14
  cautiously
• 7:15
  optimistic
• 7:15
  —
• 7:16
  that
• 7:16
  bandwidth
• 7:17
  figure
• 7:17
  is
• 7:17
  genuinely
• 7:18
  impressive,
• 7:18
  and
• 7:19
  144GB
• 7:20
  in
• 7:21
  a
• 7:21
  single
• 7:21
  card
• 7:21
  would
• 7:21
  handle
• 7:22
  most
• 7:22
  70B
• 7:22
  models
• 7:23
  without
• 7:23
  quantization.
• 7:25
  But
• 7:25
  pricing
• 7:25
  is
• 7:25
  the
• 7:25
  elephant:
• 7:26
  rumors
• 7:26
  peg
• 7:27
  it
• 7:27
  around
• 7:27
  $30K,
• 7:28
  and
• 7:28
  HBM3e
• 7:29
  dominates
• 7:29
  the
• 7:29
  bill
• 7:29
  of
• 7:30
  materials.
• 7:31
  The
• 7:31
  perennial
• 7:31
  AMD
• 7:31
  problem
• 7:32
  applies
• 7:32
  —
• 7:33
  ROCm
• 7:33
  software
• 7:34
  maturity
• 7:34
  remains
• 7:34
  the
• 7:34
  real
• 7:35
  bottleneck.
• 7:36
  As
• 7:36
  one
• 7:36
  commenter
• 7:36
  put
• 7:36
  it:
• 7:37
  "Without
• 7:37
  any
• 7:37
  details
• 7:38
  on
• 7:38
  how
• 7:38
  to
• 7:38
  tap
• 7:38
  into
• 7:39
  the
• 7:39
  GPU
• 7:39
  with
• 7:39
  software,
• 7:40
  I
• 7:40
  ain't
• 7:40
  buying."
• 7:41
  The
• 7:41
  bull
• 7:41
  case
• 7:41
  is
• 7:42
  that
• 7:42
  AMD
• 7:42
  must
• 7:42
  undercut
• 7:43
  NVIDIA
• 7:43
  to
• 7:44
  be
• 7:44
  competitive,
• 7:45
  potentially
• 7:45
  landing
• 7:45
  closer
• 7:46
  to
• 7:46
  the
• 7:46
  RTX
• 7:46
  Pro
• 7:46
  6000
• 7:47
  than
• 7:47
  the
• 7:47
  H200.
• 7:48
  The
• 7:48
  bear
• 7:49
  case
• 7:49
  is
• 7:49
  that
• 7:49
  144GB
• 7:51
  of
• 7:51
  HBM
• 7:51
  at
• 7:52
  $30K
• 7:52
  is
• 7:53
  a
• 7:53
  prosumer
• 7:53
  no-man's-land
• 7:53
  —
• 7:54
  too
• 7:54
  expensive
• 7:55
  for
• 7:55
  hobbyists,
• 7:55
  too
• 7:56
  limited
• 7:56
  for
• 7:56
  datacenters.
• 7:57
  At
• 7:57
  the
• 7:57
  exotic
• 7:58
  end,
• 7:58
  Taiwanese
• 7:58
  startup
• 7:59
  Skymizer
• 7:59
  announced
• 7:59
  the
• 7:59
  HTX301
• 8:00
  —
• 8:02
  a
• 8:02
  PCIe
• 8:02
  inference
• 8:03
  card
• 8:03
  with
• 8:03
  384GB
• 8:04
  of
• 8:04
  memory
• 8:05
  at
• 8:05
  ~240
• 8:06
  watts.
• 8:06
  The
• 8:06
  specs
• 8:07
  sound
• 8:07
  astonishing
• 8:07
  until
• 8:08
  you
• 8:08
  read
• 8:08
  the
• 8:08
  fine
• 8:08
  print:
• 8:09
  28nm
• 8:10
  process
• 8:10
  node,
• 8:10
  100GB/s
• 8:12
  bandwidth,
• 8:12
  and
• 8:13
  Llama
• 8:13
  2
• 8:13
  7B
• 8:13
  prefill
• 8:14
  at
• 8:14
  just
• 8:14
  240
• 8:15
  tok/s.
• 8:16
  DeepSeek
• 8:17
  R1
• 8:17
  Q4
• 8:17
  decodes
• 8:18
  at
• 8:18
  5
• 8:18
  tok/s
• 8:19
  on
• 8:19
  a
• 8:19
  single
• 8:19
  card.
• 8:20
  The
• 8:20
  community
• 8:21
  consensus
• 8:21
  is
• 8:21
  politely
• 8:22
  skeptical
• 8:22
  —
• 8:23
  "vaporware
• 8:23
  until
• 8:24
  proven
• 8:24
  otherwise."
• 8:25
  The
• 8:25
  card
• 8:26
  is
• 8:26
  designed
• 8:26
  for
• 8:26
  prefill/decode
• 8:27
  separation
• 8:28
  architectures
• 8:29
  where
• 8:29
  decoding
• 8:29
  is
• 8:29
  memory-bound,
• 8:30
  but
• 8:30
  that
• 8:31
  100GB/s
• 8:34
  bandwidth
• 8:34
  is
• 8:34
  an
• 8:34
  order
• 8:35
  of
• 8:35
  magnitude
• 8:35
  below
• 8:35
  what
• 8:36
  competitive
• 8:36
  inference
• 8:36
  demands.
• 8:38
  Apple
• 8:38
  quietly
• 8:38
  removed
• 8:39
  the
• 8:39
  256GB
• 8:40
  M3
• 8:41
  Ultra
• 8:41
  Mac
• 8:42
  Studio
• 8:42
  from
• 8:42
  its
• 8:42
  online
• 8:43
  store,
• 8:43
  following
• 8:44
  the
• 8:44
  earlier
• 8:44
  removal
• 8:45
  of
• 8:45
  the
• 8:45
  512GB
• 8:47
  configuration.
• 8:48
  The
• 8:48
  most
• 8:48
  plausible
• 8:49
  explanation
• 8:50
  is
• 8:50
  supply
• 8:50
  chain
• 8:50
  management
• 8:51
  —
• 8:52
  Apple
• 8:52
  is
• 8:52
  pivoting
• 8:52
  memory
• 8:53
  chip
• 8:53
  stock
• 8:53
  toward
• 8:53
  M5
• 8:54
  manufacturing,
• 8:55
  and
• 8:55
  a
• 8:55
  representative
• 8:56
  reportedly
• 8:56
  confirmed
• 8:57
  as
• 8:57
  much.
• 8:58
  But
• 8:58
  the
• 8:58
  move
• 8:58
  underscores
• 8:59
  a
• 8:59
  deeper
• 8:59
  tension:
• 9:00
  Samsung
• 9:01
  workers
• 9:01
  protesting
• 9:02
  over
• 9:02
  compensation
• 9:03
  have
• 9:03
  already
• 9:03
  forced
• 9:04
  a
• 9:04
  58%
• 9:05
  reduction
• 9:05
  in
• 9:05
  RAM
• 9:06
  production
• 9:06
  on
• 9:06
  one
• 9:06
  shift,
• 9:07
  and
• 9:07
  global
• 9:08
  memory
• 9:08
  prices
• 9:08
  are
• 9:08
  climbing.
• 9:10
  For
• 9:10
  the
• 9:10
  local
• 9:10
  inference
• 9:11
  community,
• 9:11
  the
• 9:12
  M3
• 9:12
  Ultra
• 9:12
  512GB
• 9:14
  was
• 9:15
  the
• 9:15
  gold
• 9:15
  standard
• 9:15
  for
• 9:15
  running
• 9:16
  70B+
• 9:17
  models
• 9:17
  on
• 9:17
  unified
• 9:18
  memory
• 9:18
  with
• 9:18
  impressive
• 9:19
  bandwidth.
• 9:20
  One
• 9:20
  commenter's
• 9:21
  extrapolation
• 9:22
  is
• 9:22
  darkly
• 9:22
  funny:
• 9:23
  "if
• 9:23
  you
• 9:23
  extrapolate
• 9:24
  that
• 9:24
  trend
• 9:24
  the
• 9:24
  M6
• 9:25
  ships
• 9:25
  with
• 9:25
  24GB
• 9:26
  and
• 9:26
  we're
• 9:27
  back
• 9:27
  to
• 9:27
  quantizing
• 9:27
  everything."
• 9:29
  The
• 9:29
  practical
• 9:29
  signal
• 9:30
  is
• 9:30
  that
• 9:30
  serious
• 9:31
  local
• 9:31
  inference
• 9:31
  increasingly
• 9:32
  requires
• 9:32
  the
• 9:32
  top-tier
• 9:33
  memory
• 9:33
  configuration
• 9:34
  —
• 9:35
  the
• 9:35
  256GB
• 9:38
  model
• 9:38
  was
• 9:38
  always
• 9:38
  a
• 9:38
  false
• 9:39
  economy
• 9:39
  for
• 9:39
  anyone
• 9:40
  running
• 9:40
  large
• 9:40
  models.
• 9:45
  "If
• 9:45
  you've
• 9:45
  ever
• 9:46
  resorted
• 9:46
  to
• 9:46
  prompt
• 9:46
  chaining
• 9:47
  or
• 9:47
  few-shot
• 9:48
  scaffolding,
• 9:49
  you've
• 9:49
  hit
• 9:49
  the
• 9:49
  ceiling
• 9:50
  of
• 9:50
  prompting."
• 9:51
  That
• 9:51
  thesis,
• 9:51
  from
• 9:52
  a
• 9:52
  developer
• 9:52
  essay
• 9:52
  arguing
• 9:53
  agents
• 9:53
  need
• 9:54
  control
• 9:54
  flow
• 9:54
  rather
• 9:55
  than
• 9:55
  more
• 9:55
  elaborate
• 9:55
  prompts,
• 9:56
  crystallizes
• 9:57
  an
• 9:57
  architectural
• 9:57
  consensus
• 9:58
  that
• 9:58
  has
• 9:58
  been
• 9:58
  building
• 9:59
  for
• 9:59
  months.
• 10:00
  The
• 10:00
  argument
• 10:00
  is
• 10:00
  straightforward:
• 10:01
  prompt
• 10:02
  chains
• 10:02
  are
• 10:02
  non-deterministic,
• 10:03
  weakly
• 10:03
  specified,
• 10:04
  and
• 10:05
  difficult
• 10:05
  to
• 10:05
  verify.
• 10:06
  Reliable
• 10:07
  agent
• 10:07
  systems
• 10:08
  require
• 10:08
  deterministic
• 10:09
  scaffolds
• 10:09
  —
• 10:09
  explicit
• 10:10
  state
• 10:10
  transitions,
• 10:11
  validation
• 10:12
  checkpoints,
• 10:13
  and
• 10:13
  programmatic
• 10:14
  error
• 10:14
  detection
• 10:14
  —
• 10:15
  that
• 10:15
  treat
• 10:15
  the
• 10:16
  LLM
• 10:16
  as
• 10:16
  a
• 10:16
  component,
• 10:17
  not
• 10:17
  the
• 10:17
  system.
• 10:19
  Without
• 10:19
  aggressive
• 10:19
  verification,
• 10:20
  the
• 10:20
  alternatives
• 10:21
  reduce
• 10:21
  to
• 10:21
  keeping
• 10:22
  a
• 10:22
  human
• 10:22
  in
• 10:22
  the
• 10:22
  loop,
• 10:23
  performing
• 10:23
  exhaustive
• 10:24
  end-to-end
• 10:24
  testing,
• 10:25
  or
• 10:26
  "vibe
• 10:26
  accepting"
• 10:27
  the
• 10:27
  outputs.
• 10:28
  NEEDLE
• 10:28
  (Navigates
• 10:29
  Every
• 10:29
  Enqueued
• 10:30
  Deliverable,
• 10:31
  Logs
• 10:31
  Effort)
• 10:32
  is
• 10:32
  a
• 10:32
  Rust
• 10:32
  implementation
• 10:33
  of
• 10:33
  exactly
• 10:33
  this
• 10:33
  philosophy.
• 10:35
  It
• 10:35
  wraps
• 10:35
  headless
• 10:36
  coding
• 10:36
  CLI
• 10:37
  agents
• 10:37
  —
• 10:38
  Claude
• 10:38
  Code,
• 10:39
  Codex,
• 10:40
  Aider,
• 10:40
  OpenCode
• 10:41
  —
• 10:42
  in
• 10:42
  a
• 10:42
  deterministic
• 10:42
  state
• 10:43
  machine
• 10:43
  that
• 10:43
  processes
• 10:44
  a
• 10:44
  shared
• 10:44
  bead
• 10:44
  queue.
• 10:45
  Every
• 10:45
  outcome
• 10:46
  an
• 10:46
  agent
• 10:46
  can
• 10:46
  produce
• 10:47
  has
• 10:47
  an
• 10:47
  explicit
• 10:48
  handler:
• 10:49
  success
• 10:49
  closes
• 10:50
  the
• 10:50
  bead,
• 10:50
  failure
• 10:51
  logs
• 10:51
  and
• 10:51
  releases,
• 10:52
  timeout
• 10:52
  defers,
• 10:53
  crash
• 10:54
  creates
• 10:54
  an
• 10:54
  alert.
• 10:55
  Multiple
• 10:56
  workers
• 10:56
  run
• 10:56
  independently
• 10:57
  against
• 10:57
  a
• 10:58
  shared
• 10:58
  SQLite-backed
• 10:59
  queue
• 10:59
  with
• 10:59
  no
• 11:00
  central
• 11:00
  orchestrator
• 11:00
  —
• 11:02
  coordination
• 11:02
  happens
• 11:03
  through
• 11:03
  atomic
• 11:03
  claims
• 11:04
  and
• 11:04
  deterministic
• 11:04
  priority
• 11:05
  ordering.
• 11:06
  When
• 11:06
  the
• 11:06
  queue
• 11:07
  empties,
• 11:07
  a
• 11:08
  "strand
• 11:08
  escalation"
• 11:09
  sequence
• 11:09
  searches
• 11:09
  for
• 11:10
  work
• 11:10
  across
• 11:10
  workspaces,
• 11:11
  performs
• 11:12
  health
• 11:12
  checks,
• 11:13
  and
• 11:13
  optionally
• 11:13
  generates
• 11:14
  new
• 11:14
  beads
• 11:14
  from
• 11:14
  documentation
• 11:15
  gaps.
• 11:17
  NEEDLE
• 11:17
  currently
• 11:17
  powers
• 11:18
  its
• 11:18
  author's
• 11:18
  headless
• 11:19
  multi-agent
• 11:19
  workflow
• 11:20
  with
• 11:20
  full
• 11:20
  OpenTelemetry
• 11:21
  telemetry,
• 11:22
  and
• 11:22
  a
• 11:22
  companion
• 11:22
  tool
• 11:23
  (claude-governor)
• 11:23
  caps
• 11:24
  API
• 11:24
  spend
• 11:24
  across
• 11:25
  worker
• 11:25
  fleets.
• 11:26
  Mesh
• 11:26
  takes
• 11:27
  a
• 11:27
  different
• 11:27
  approach
• 11:27
  to
• 11:27
  the
• 11:27
  multi-model
• 11:28
  problem.
• 11:29
  Rather
• 11:29
  than
• 11:30
  orchestrating
• 11:30
  headless
• 11:30
  workers,
• 11:31
  it
• 11:31
  gives
• 11:32
  your
• 11:32
  primary
• 11:32
  AI
• 11:32
  assistant
• 11:33
  the
• 11:33
  ability
• 11:34
  to
• 11:34
  consult
• 11:34
  other
• 11:35
  models
• 11:35
  mid-session
• 11:36
  via
• 11:36
  MCP.
• 11:37
  Ask
• 11:38
  Claude
• 11:38
  Code
• 11:38
  to
• 11:38
  get
• 11:38
  a
• 11:38
  second
• 11:39
  opinion
• 11:39
  from
• 11:39
  Gemini
• 11:40
  Pro
• 11:40
  on
• 11:40
  an
• 11:40
  auth
• 11:41
  flow,
• 11:42
  run
• 11:42
  a
• 11:42
  consensus
• 11:42
  debate
• 11:43
  between
• 11:43
  GPT-5
• 11:44
  and
• 11:44
  Opus,
• 11:45
  or
• 11:45
  chain
• 11:45
  a
• 11:45
  code
• 11:45
  review
• 11:46
  through
• 11:46
  multiple
• 11:46
  models
• 11:47
  with
• 11:47
  automatic
• 11:47
  fallback.
• 11:48
  Mesh
• 11:49
  routes
• 11:49
  calls
• 11:50
  through
• 11:50
  local
• 11:50
  CLIs
• 11:51
  when
• 11:51
  possible
• 11:52
  (Gemini
• 11:52
  CLI,
• 11:53
  Codex
• 11:53
  CLI)
• 11:54
  and
• 11:54
  falls
• 11:54
  back
• 11:54
  to
• 11:55
  OpenRouter,
• 11:55
  producing
• 11:56
  identical
• 11:56
  response
• 11:57
  objects
• 11:57
  regardless
• 11:58
  of
• 11:58
  backend.
• 11:59
  Eighteen
• 11:59
  tools
• 11:59
  ship
• 12:00
  built-in,
• 12:00
  from
• 12:01
  code
• 12:01
  review
• 12:01
  and
• 12:01
  debugging
• 12:02
  to
• 12:02
  security
• 12:02
  audits
• 12:03
  and
• 12:03
  multi-model
• 12:04
  consensus
• 12:04
  with
• 12:04
  stance
• 12:05
  steering.
• 12:06
  The
• 12:06
  Needle
• 12:06
  RAG
• 12:06
  platform
• 12:07
  (distinct
• 12:07
  from
• 12:07
  NEEDLE
• 12:08
  the
• 12:08
  orchestrator)
• 12:08
  offers
• 12:09
  managed
• 12:09
  retrieval-augmented
• 12:10
  generation
• 12:11
  with
• 12:11
  a
• 12:11
  minimalist
• 12:11
  API
• 12:12
  —
• 12:13
  create
• 12:13
  a
• 12:13
  collection,
• 12:14
  add
• 12:14
  files,
• 12:15
  search
• 12:15
  semantically
• 12:16
  —
• 12:17
  with
• 12:17
  SDKs
• 12:18
  for
• 12:18
  multiple
• 12:18
  languages
• 12:19
  and
• 12:19
  drop-in
• 12:19
  integrations
• 12:20
  for
• 12:20
  LangChain
• 12:21
  and
• 12:21
  LlamaIndex.
• 12:22
  And
• 12:22
  in
• 12:22
  a
• 12:22
  delightful
• 12:23
  demonstration
• 12:24
  of
• 12:24
  MCP's
• 12:25
  expanding
• 12:25
  reach
• 12:25
  beyond
• 12:26
  developer
• 12:26
  tooling,
• 12:27
  a
• 12:27
  new
• 12:27
  Ableton
• 12:28
  Live
• 12:28
  MCP
• 12:29
  bridge
• 12:29
  lets
• 12:30
  AI
• 12:30
  agents
• 12:30
  control
• 12:31
  the
• 12:31
  music
• 12:31
  production
• 12:32
  DAW
• 12:32
  —
• 12:33
  creating
• 12:33
  MIDI
• 12:33
  clips,
• 12:34
  analyzing
• 12:35
  harmony,
• 12:35
  generating
• 12:36
  spectrograms,
• 12:37
  setting
• 12:37
  up
• 12:38
  mastering
• 12:38
  chains,
• 12:39
  and
• 12:39
  even
• 12:39
  capturing
• 12:40
  audio
• 12:40
  signals
• 12:41
  for
• 12:41
  agent-driven
• 12:41
  mixing
• 12:42
  feedback
• 12:42
  loops.
• 12:43
  The
• 12:43
  developer
• 12:44
  built
• 12:44
  it
• 12:44
  while
• 12:44
  nap-trapped
• 12:45
  by
• 12:45
  his
• 12:45
  baby
• 12:45
  and
• 12:46
  used
• 12:46
  Codex's
• 12:46
  /goal
• 12:47
  command
• 12:47
  to
• 12:47
  optimize
• 12:48
  for
• 12:48
  low
• 12:48
  latency,
• 12:49
  high
• 12:49
  reliability,
• 12:50
  and
• 12:50
  low
• 12:50
  token
• 12:51
  usage.
• 12:56
  A
• 12:56
  striking
• 12:56
  new
• 12:56
  paper
• 12:57
  from
• 12:57
  Stanford
• 12:57
  and
• 12:57
  Bigspin
• 12:58
  AI
• 12:58
  quantifies
• 12:59
  something
• 12:59
  practitioners
• 12:59
  have
• 13:00
  long
• 13:00
  suspected:
• 13:01
  AI
• 13:01
  systems
• 13:02
  fail
• 13:02
  silently
• 13:02
  far
• 13:03
  more
• 13:03
  often
• 13:03
  than
• 13:03
  they
• 13:03
  fail
• 13:04
  visibly.
• 13:05
  Analyzing
• 13:06
  196,704
• 13:08
  ChatGPT
• 13:09
  transcripts
• 13:09
  from
• 13:09
  the
• 13:10
  WildChat
• 13:10
  dataset,
• 13:11
  the
• 13:11
  researchers
• 13:12
  found
• 13:12
  that
• 13:12
  78%
• 13:13
  of
• 13:13
  AI
• 13:13
  failures
• 13:13
  are
• 13:14
  invisible
• 13:14
  —
• 13:15
  the
• 13:15
  user
• 13:15
  gave
• 13:15
  no
• 13:16
  overt
• 13:16
  indication
• 13:17
  anything
• 13:17
  went
• 13:17
  wrong.
• 13:18
  The
• 13:18
  failures
• 13:18
  cluster
• 13:19
  into
• 13:19
  eight
• 13:19
  archetypes:
• 13:20
  The
• 13:20
  Drift
• 13:21
  (37.2%,
• 13:22
  the
• 13:22
  AI
• 13:23
  addresses
• 13:23
  a
• 13:23
  related
• 13:23
  but
• 13:23
  different
• 13:24
  goal),
• 13:24
  The
• 13:24
  Confidence
• 13:25
  Trap
• 13:25
  (26.4%,
• 13:27
  wrong
• 13:27
  answers
• 13:27
  delivered
• 13:28
  with
• 13:28
  total
• 13:28
  confidence),
• 13:29
  The
• 13:29
  Mystery
• 13:29
  Failure
• 13:30
  (12%),
• 13:31
  The
• 13:31
  Silent
• 13:31
  Mismatch
• 13:31
  (9%),
• 13:32
  The
• 13:33
  Contradiction
• 13:33
  Unravel
• 13:33
  (7.6%),
• 13:35
  The
• 13:35
  Death
• 13:35
  Spiral
• 13:35
  (6.9%),
• 13:37
  The
• 13:37
  Partial
• 13:37
  Recovery
• 13:38
  (6.3%),
• 13:39
  and
• 13:39
  The
• 13:39
  Walkaway
• 13:39
  (5.9%,
• 13:41
  user
• 13:41
  silently
• 13:42
  abandons).
• 13:43
  The
• 13:43
  most
• 13:43
  uncomfortable
• 13:44
  finding
• 13:44
  is
• 13:44
  that
• 13:44
  these
• 13:44
  failures
• 13:45
  are
• 13:45
  structural,
• 13:46
  not
• 13:46
  capability-driven.
• 13:48
  A
• 13:48
  retrospective
• 13:48
  validation
• 13:49
  using
• 13:49
  Claude
• 13:49
  Opus
• 13:50
  classified
• 13:50
  91%
• 13:51
  of
• 13:51
  failures
• 13:51
  as
• 13:51
  involving
• 13:52
  interactional
• 13:52
  dynamics
• 13:53
  rather
• 13:53
  than
• 13:53
  raw
• 13:53
  model
• 13:53
  limitations,
• 13:54
  and
• 13:55
  estimated
• 13:55
  that
• 13:55
  94%
• 13:56
  would
• 13:56
  persist
• 13:56
  even
• 13:57
  with
• 13:57
  a
• 13:57
  more
• 13:57
  capable
• 13:57
  model.
• 13:58
  The
• 13:59
  single
• 13:59
  most
• 13:59
  common
• 13:59
  pattern,
• 14:00
  appearing
• 14:00
  in
• 14:00
  79%
• 14:01
  of
• 14:01
  failures:
• 14:02
  "generate
• 14:03
  rather
• 14:03
  than
• 14:03
  clarify"
• 14:03
  —
• 14:04
  the
• 14:04
  model
• 14:05
  produces
• 14:05
  fluent
• 14:05
  output
• 14:06
  instead
• 14:06
  of
• 14:06
  surfacing
• 14:07
  ambiguity.
• 14:08
  Even
• 14:08
  if
• 14:08
  ChatGPT's
• 14:09
  overall
• 14:10
  failure
• 14:10
  rate
• 14:11
  dropped
• 14:11
  to
• 14:11
  1%,
• 14:12
  at
• 14:12
  2.6
• 14:13
  billion
• 14:13
  daily
• 14:13
  messages
• 14:14
  that
• 14:14
  would
• 14:14
  still
• 14:14
  mean
• 14:14
  over
• 14:15
  20
• 14:15
  million
• 14:15
  daily
• 14:15
  failures
• 14:16
  involving
• 14:16
  persistent
• 14:17
  behavioral
• 14:17
  dynamics.
• 14:18
  Software
• 14:19
  development,
• 14:19
  interestingly,
• 14:20
  has
• 14:20
  disproportionately
• 14:21
  high
• 14:21
  visible
• 14:22
  failure
• 14:22
  rates
• 14:22
  —
• 14:23
  developers
• 14:24
  spot
• 14:24
  and
• 14:24
  challenge
• 14:24
  errors
• 14:25
  —
• 14:25
  while
• 14:26
  creative
• 14:26
  writing
• 14:26
  and
• 14:27
  UX
• 14:27
  design
• 14:27
  suffer
• 14:28
  particularly
• 14:28
  from
• 14:28
  The
• 14:28
  Drift.
• 14:30
  On
• 14:30
  the
• 14:30
  training
• 14:30
  side,
• 14:31
  Google
• 14:31
  DeepMind
• 14:32
  published
• 14:32
  "Efficient
• 14:33
  Exploration
• 14:33
  at
• 14:34
  Scale,"
• 14:35
  demonstrating
• 14:35
  an
• 14:35
  online
• 14:36
  RLHF
• 14:36
  algorithm
• 14:37
  that
• 14:37
  matches
• 14:38
  offline
• 14:38
  RLHF
• 14:39
  trained
• 14:39
  on
• 14:39
  200K
• 14:40
  human
• 14:40
  preference
• 14:41
  labels
• 14:41
  using
• 14:42
  fewer
• 14:42
  than
• 14:42
  20K
• 14:42
  —
• 14:44
  a
• 14:44
  10x
• 14:44
  data
• 14:44
  efficiency
• 14:45
  gain,
• 14:45
  with
• 14:46
  projections
• 14:46
  reaching
• 14:47
  1,000x
• 14:47
  at
• 14:47
  scale.
• 14:49
  Three
• 14:49
  innovations
• 14:50
  drive
• 14:50
  the
• 14:50
  result:
• 14:51
  a
• 14:51
  small
• 14:51
  "affirmative
• 14:52
  nudge"
• 14:53
  added
• 14:53
  to
• 14:53
  reinforcement
• 14:54
  signals
• 14:54
  that
• 14:54
  prevents
• 14:55
  performance
• 14:55
  collapse,
• 14:56
  an
• 14:57
  epistemic
• 14:57
  neural
• 14:57
  network
• 14:58
  modeling
• 14:58
  reward
• 14:59
  uncertainty
• 14:59
  via
• 15:00
  an
• 15:00
  ensemble
• 15:00
  of
• 15:00
  200
• 15:01
  MLP
• 15:01
  heads
• 15:02
  atop
• 15:02
  a
• 15:02
  shared
• 15:02
  transformer
• 15:03
  backbone,
• 15:04
  and
• 15:04
  information-directed
• 15:05
  exploration
• 15:06
  that
• 15:06
  selects
• 15:07
  response
• 15:07
  pairs
• 15:07
  to
• 15:07
  maximize
• 15:08
  the
• 15:08
  informativeness
• 15:09
  of
• 15:09
  each
• 15:09
  human
• 15:09
  query.
• 15:11
  The
• 15:11
  authors
• 15:11
  frame
• 15:11
  this
• 15:12
  as
• 15:12
  foundational
• 15:12
  to
• 15:12
  safe
• 15:13
  artificial
• 15:13
  superintelligence,
• 15:15
  arguing
• 15:15
  that
• 15:15
  efficient
• 15:15
  exploration
• 15:16
  of
• 15:16
  human
• 15:16
  preferences
• 15:17
  "should
• 15:18
  serve
• 15:18
  as
• 15:18
  a
• 15:18
  cornerstone
• 15:19
  on
• 15:19
  the
• 15:19
  path."
• 15:20
  Prior
• 15:20
  active
• 15:21
  learning
• 15:21
  literature
• 15:21
  reported
• 15:22
  only
• 15:22
  2–5x
• 15:23
  gains
• 15:24
  on
• 15:24
  narrow
• 15:24
  prompt
• 15:24
  sets,
• 15:25
  making
• 15:25
  the
• 15:26
  10x
• 15:26
  demonstrated
• 15:27
  result
• 15:27
  —
• 15:28
  and
• 15:28
  the
• 15:28
  projected
• 15:28
  1,000x
• 15:29
  extrapolation
• 15:30
  —
• 15:31
  a
• 15:31
  genuine
• 15:31
  step
• 15:32
  change.
• 15:32
  Gadi
• 15:33
  Evron,
• 15:33
  the
• 15:34
  veteran
• 15:34
  cybersecurity
• 15:35
  researcher,
• 15:35
  offers
• 15:36
  a
• 15:36
  useful
• 15:36
  lens
• 15:37
  for
• 15:37
  interpreting
• 15:37
  both
• 15:38
  findings.
• 15:39
  He
• 15:39
  describes
• 15:39
  what
• 15:40
  he
• 15:40
  calls
• 15:40
  "The
• 15:40
  Evron
• 15:41
  Amnesia
• 15:41
  of
• 15:41
  AI
• 15:42
  predictions"
• 15:42
  —
• 15:43
  a
• 15:43
  cognitive
• 15:44
  bias
• 15:44
  where
• 15:44
  practitioners
• 15:45
  accurately
• 15:45
  observe
• 15:46
  that
• 15:46
  some
• 15:46
  AI
• 15:46
  capability
• 15:47
  "isn't
• 15:47
  there
• 15:48
  yet,"
• 15:48
  then
• 15:48
  are
• 15:49
  surprised
• 15:49
  when
• 15:49
  it
• 15:49
  arrives
• 15:50
  three
• 15:50
  months
• 15:50
  later,
• 15:51
  then
• 15:51
  retroactively
• 15:52
  claim
• 15:52
  "it
• 15:52
  has
• 15:52
  always
• 15:53
  been
• 15:53
  that
• 15:53
  way"
• 15:54
  while
• 15:54
  moving
• 15:54
  the
• 15:54
  goalpost
• 15:55
  to
• 15:55
  the
• 15:55
  next
• 15:55
  missing
• 15:55
  capability.
• 15:57
  The
• 15:57
  pattern:
• 15:58
  research
• 15:58
  leads
• 15:58
  to
• 15:59
  tweaked
• 15:59
  workflows,
• 16:00
  which
• 16:00
  become
• 16:00
  public
• 16:00
  harnesses,
• 16:01
  get
• 16:01
  redeveloped
• 16:02
  by
• 16:02
  hundreds,
• 16:03
  become
• 16:03
  a
• 16:03
  Claude
• 16:03
  Code
• 16:04
  feature,
• 16:04
  then
• 16:04
  reach
• 16:05
  mass
• 16:05
  adoption.
• 16:06
  The
• 16:06
  implication
• 16:07
  for
• 16:07
  invisible
• 16:07
  failures:
• 16:08
  today's
• 16:08
  78%
• 16:09
  invisible
• 16:10
  failure
• 16:10
  rate
• 16:10
  feels
• 16:11
  like
• 16:11
  a
• 16:11
  permanent
• 16:11
  limitation.
• 16:13
  In
• 16:13
  six
• 16:13
  months,
• 16:14
  someone
• 16:14
  will
• 16:14
  ship
• 16:14
  a
• 16:15
  monitoring
• 16:15
  product
• 16:15
  that
• 16:15
  detects
• 16:16
  most
• 16:16
  of
• 16:16
  these
• 16:16
  patterns,
• 16:17
  and
• 16:17
  we'll
• 16:17
  wonder
• 16:18
  why
• 16:18
  it
• 16:18
  took
• 16:18
  so
• 16:18
  long.
• 16:21
  Synaptic-Tuner
• 16:22
  is
• 16:22
  an
• 16:22
  agentic-first
• 16:23
  toolkit
• 16:23
  covering
• 16:23
  the
• 16:24
  full
• 16:24
  LLM
• 16:24
  fine-tuning
• 16:25
  pipeline
• 16:25
  —
• 16:26
  synthetic
• 16:26
  data
• 16:27
  generation,
• 16:27
  SFT/KTO/GRPO
• 16:32
  training,
• 16:32
  evaluation,
• 16:33
  GGUF
• 16:34
  quantization,
• 16:34
  and
• 16:35
  deployment
• 16:35
  —
• 16:36
  designed
• 16:36
  to
• 16:36
  be
• 16:36
  operated
• 16:37
  by
• 16:37
  AI
• 16:37
  coding
• 16:38
  agents
• 16:38
  rather
• 16:38
  than
• 16:39
  humans
• 16:39
  memorizing
• 16:40
  CLI
• 16:40
  flags.
• 16:42
  The
• 16:42
  repo
• 16:42
  ships
• 16:43
  with
• 16:43
  project
• 16:43
  skills
• 16:43
  (Markdown
• 16:44
  files
• 16:44
  with
• 16:44
  progressive
• 16:45
  disclosure)
• 16:45
  that
• 16:45
  work
• 16:46
  with
• 16:46
  Claude
• 16:46
  Code,
• 16:46
  Cursor,
• 16:47
  Windsurf,
• 16:47
  and
• 16:47
  other
• 16:47
  AI
• 16:47
  coding
• 16:48
  tools.
• 16:49
  HuggingFace
• 16:49
  Jobs
• 16:50
  is
• 16:50
  the
• 16:50
  canonical
• 16:50
  cloud
• 16:51
  path,
• 16:51
  with
• 16:51
  local
• 16:52
  Docker
• 16:52
  training
• 16:52
  via
• 16:53
  Unsloth
• 16:53
  as
• 16:53
  the
• 16:53
  supporting
• 16:54
  path
• 16:54
  for
• 16:54
  GPU
• 16:54
  owners.
• 16:56
  Each
• 16:56
  stage
• 16:56
  is
• 16:56
  config-driven
• 16:57
  through
• 16:57
  YAML.
• 16:58
  The
• 16:58
  AI
• 16:59
  labor
• 16:59
  story
• 16:59
  surfaced
• 17:00
  in
• 17:00
  a
• 17:00
  detailed
• 17:00
  investigation
• 17:01
  into
• 17:01
  the
• 17:01
  hidden
• 17:01
  workforce
• 17:02
  training
• 17:02
  AI
• 17:02
  systems.
• 17:03
  The
• 17:04
  report
• 17:04
  documents
• 17:04
  data
• 17:05
  workers
• 17:05
  across
• 17:05
  the
• 17:05
  US
• 17:06
  earning
• 17:06
  median
• 17:06
  annual
• 17:07
  incomes
• 17:07
  under
• 17:07
  $23,000,
• 17:09
  with
• 17:09
  86%
• 17:10
  struggling
• 17:10
  to
• 17:10
  meet
• 17:11
  financial
• 17:11
  responsibilities,
• 17:12
  a
• 17:12
  quarter
• 17:12
  relying
• 17:13
  on
• 17:13
  public
• 17:13
  assistance,
• 17:14
  and
• 17:14
  more
• 17:14
  than
• 17:14
  one
• 17:15
  in
• 17:15
  five
• 17:15
  having
• 17:15
  experienced
• 17:16
  homelessness.
• 17:17
  Ivy
• 17:17
  League
• 17:18
  PhD
• 17:18
  graduates
• 17:19
  are
• 17:19
  taking
• 17:19
  $35/hour
• 17:20
  contract
• 17:21
  work
• 17:21
  from
• 17:22
  platforms
• 17:22
  like
• 17:22
  Mercor
• 17:23
  and
• 17:23
  Scale
• 17:23
  AI,
• 17:24
  only
• 17:24
  to
• 17:24
  have
• 17:25
  contracts
• 17:25
  pulled
• 17:25
  within
• 17:26
  weeks.
• 17:27
  One
• 17:27
  worker
• 17:27
  described
• 17:28
  reviewing
• 17:28
  AI-generated
• 17:29
  videos
• 17:29
  of
• 17:29
  extreme
• 17:30
  violence
• 17:30
  for
• 17:30
  a
• 17:30
  project
• 17:31
  codenamed
• 17:31
  "Arsenic"
• 17:32
  and
• 17:32
  being
• 17:33
  asked
• 17:33
  to
• 17:33
  perform
• 17:33
  tasks
• 17:34
  far
• 17:34
  outside
• 17:34
  their
• 17:35
  expertise
• 17:35
  —
• 17:36
  grading
• 17:36
  calculus
• 17:37
  problems
• 17:37
  as
• 17:37
  a
• 17:37
  philosophy
• 17:38
  major,
• 17:39
  providing
• 17:39
  counseling
• 17:39
  advice
• 17:40
  with
• 17:40
  no
• 17:40
  training.
• 17:41
  A
• 17:41
  new
• 17:41
  California
• 17:42
  bill,
• 17:43
  AB
• 17:43
  2653
• 17:43
  (the
• 17:44
  Sweatshop-Free
• 17:45
  AI
• 17:45
  Procurement
• 17:45
  Act),
• 17:46
  would
• 17:46
  require
• 17:47
  state
• 17:47
  AI
• 17:47
  procurement
• 17:48
  to
• 17:48
  verify
• 17:48
  labor
• 17:49
  standards
• 17:49
  compliance.
• 17:50
  Economist
• 17:51
  Daron
• 17:51
  Acemoglu
• 17:52
  warns
• 17:52
  that
• 17:52
  the
• 17:52
  resulting
• 17:53
  inequality
• 17:53
  "could
• 17:53
  be
• 17:54
  something
• 17:54
  we've
• 17:54
  never
• 17:54
  experienced
• 17:55
  —
• 17:55
  a
• 17:56
  handful
• 17:56
  of
• 17:56
  corporations
• 17:57
  controlling
• 17:57
  most
• 17:57
  of
• 17:57
  work."
• 17:59
  On
• 17:59
  the
• 17:59
  local
• 17:59
  speech
• 17:59
  synthesis
• 18:00
  front,
• 18:01
  a
• 18:01
  C++
• 18:02
  port
• 18:02
  of
• 18:02
  Echo-TTS
• 18:03
  brings
• 18:03
  multi-speaker
• 18:04
  text-to-speech
• 18:04
  to
• 18:05
  native
• 18:05
  deployment
• 18:05
  outside
• 18:06
  the
• 18:06
  Python
• 18:06
  stack.
• 18:08
  The
• 18:08
  port
• 18:08
  uses
• 18:08
  GGML
• 18:09
  for
• 18:09
  the
• 18:09
  diffusion
• 18:10
  transformer
• 18:10
  and
• 18:10
  ONNX
• 18:11
  Runtime
• 18:12
  for
• 18:12
  the
• 18:12
  DAC
• 18:12
  autoencoder,
• 18:13
  with
• 18:14
  model
• 18:14
  files
• 18:14
  at
• 18:14
  ~3.3GB
• 18:15
  (Q8)
• 18:16
  or
• 18:16
  ~5.6GB
• 18:17
  (F16).
• 18:19
  It
• 18:19
  includes
• 18:20
  an
• 18:20
  OpenAI-compatible
• 18:21
  server
• 18:21
  mode,
• 18:22
  multi-voice
• 18:23
  support
• 18:23
  with
• 18:23
  reference
• 18:24
  WAV
• 18:24
  conditioning,
• 18:25
  and
• 18:25
  pre-built
• 18:25
  portable
• 18:26
  binaries
• 18:26
  with
• 18:26
  bundled
• 18:27
  CUDA
• 18:27
  12.8
• 18:28
  and
• 18:28
  cuDNN
• 18:29
  9.21.
• 18:30
  The
• 18:30
  project
• 18:31
  represents
• 18:31
  the
• 18:31
  continuing
• 18:32
  push
• 18:32
  toward
• 18:32
  systems-language
• 18:33
  reimplementation
• 18:34
  of
• 18:34
  AI
• 18:34
  inference
• 18:35
  —
• 18:35
  the
• 18:36
  same
• 18:36
  motivation
• 18:36
  driving
• 18:37
  Rust
• 18:37
  rewrites
• 18:37
  across
• 18:38
  the
• 18:38
  inference
• 18:38
  stack.
• 18:41
  If
• 18:41
  continuous
• 18:42
  machine
• 18:42
  cognition
• 18:43
  requires
• 18:43
  memory,
• 18:44
  and
• 18:44
  raw
• 18:44
  accumulation
• 18:45
  without
• 18:45
  processing
• 18:46
  is
• 18:46
  hoarding
• 18:46
  rather
• 18:46
  than
• 18:46
  learning,
• 18:47
  then
• 18:47
  machines
• 18:48
  that
• 18:48
  think
• 18:48
  continuously
• 18:49
  probably
• 18:49
  need
• 18:49
  to
• 18:49
  sleep.
• 18:50
  That
• 18:50
  is
• 18:50
  the
• 18:51
  thesis
• 18:51
  of
• 18:51
  Part
• 18:51
  2
• 18:51
  of
• 18:52
  "What
• 18:52
  Happens
• 18:52
  When
• 18:53
  the
• 18:53
  Machine
• 18:53
  Never
• 18:53
  Stops
• 18:54
  Thinking,"
• 18:54
  which
• 18:54
  maps
• 18:55
  five
• 18:55
  functions
• 18:55
  of
• 18:55
  human
• 18:55
  sleep
• 18:56
  —
• 18:56
  hippocampal
• 18:57
  replay,
• 18:57
  synaptic
• 18:58
  pruning,
• 18:58
  cross-regional
• 18:59
  association
• 19:00
  during
• 19:00
  REM,
• 19:01
  scenario
• 19:01
  simulation,
• 19:02
  and
• 19:02
  emotional
• 19:02
  recalibration
• 19:03
  —
• 19:04
  onto
• 19:04
  a
• 19:04
  proposed
• 19:04
  architecture
• 19:05
  for
• 19:05
  LLM
• 19:06
  dreaming.
• 19:07
  The
• 19:07
  system
• 19:07
  would
• 19:07
  capture
• 19:08
  experience
• 19:08
  during
• 19:08
  active
• 19:09
  inference
• 19:09
  (input-output
• 19:10
  pairs,
• 19:10
  uncertainty
• 19:11
  markers,
• 19:12
  contradiction
• 19:12
  flags),
• 19:13
  then
• 19:13
  run
• 19:13
  consolidation
• 19:14
  phases:
• 19:15
  triage
• 19:15
  to
• 19:15
  deduplicate
• 19:16
  and
• 19:16
  group,
• 19:16
  deep
• 19:16
  dreaming
• 19:17
  to
• 19:17
  compress
• 19:17
  episodes
• 19:18
  into
• 19:18
  principles
• 19:19
  and
• 19:19
  generate
• 19:19
  cross-domain
• 19:20
  connections,
• 19:21
  adversarial
• 19:21
  dreaming
• 19:22
  to
• 19:22
  stress-test
• 19:22
  beliefs,
• 19:23
  and
• 19:23
  integrity
• 19:24
  verification
• 19:24
  to
• 19:24
  prevent
• 19:25
  hallucination
• 19:25
  accumulation.
• 19:26
  The
• 19:26
  author
• 19:27
  proposes
• 19:27
  temperature
• 19:28
  modulation
• 19:28
  as
• 19:28
  an
• 19:28
  analogue
• 19:29
  to
• 19:29
  sleep
• 19:29
  stages
• 19:29
  —
• 19:30
  high
• 19:30
  temperature
• 19:31
  for
• 19:31
  wild
• 19:31
  association
• 19:32
  (REM-like),
• 19:33
  low
• 19:33
  temperature
• 19:33
  for
• 19:34
  careful
• 19:34
  consolidation
• 19:35
  (slow-wave-like).
• 19:36
  The
• 19:36
  most
• 19:36
  practically
• 19:37
  interesting
• 19:37
  idea
• 19:37
  is
• 19:38
  "micro-consolidation"
• 19:39
  between
• 19:39
  interactions:
• 19:40
  brief
• 19:40
  processing
• 19:41
  after
• 19:41
  each
• 19:41
  conversation
• 19:42
  to
• 19:42
  compress,
• 19:42
  extract
• 19:43
  insights,
• 19:44
  and
• 19:44
  flag
• 19:44
  contradictions
• 19:45
  without
• 19:45
  full
• 19:45
  offline
• 19:46
  dreaming
• 19:46
  cycles.
• 19:47
  Theoretical
• 19:48
  physicist
• 19:48
  Carlo
• 19:49
  Rovelli,
• 19:49
  writing
• 19:50
  in
• 19:50
  Noema
• 19:50
  Magazine,
• 19:51
  argues
• 19:51
  there
• 19:51
  is
• 19:51
  no
• 19:51
  "hard
• 19:52
  problem
• 19:52
  of
• 19:52
  consciousness"
• 19:53
  at
• 19:53
  all.
• 19:54
  Chalmers's
• 19:54
  1994
• 19:55
  distinction
• 19:56
  between
• 19:56
  the
• 19:56
  "easy"
• 19:56
  problem
• 19:57
  (understanding
• 19:57
  brain
• 19:58
  processes
• 19:58
  that
• 19:58
  produce
• 19:58
  behavior)
• 19:59
  and
• 19:59
  the
• 19:59
  "hard"
• 20:00
  problem
• 20:00
  (explaining
• 20:01
  why
• 20:01
  brain
• 20:01
  processes
• 20:02
  are
• 20:02
  accompanied
• 20:02
  by
• 20:02
  experience)
• 20:03
  is,
• 20:04
  in
• 20:04
  Rovelli's
• 20:04
  view,
• 20:05
  a
• 20:05
  rhetorical
• 20:05
  trick
• 20:06
  that
• 20:06
  smuggles
• 20:06
  dualism
• 20:07
  in
• 20:07
  through
• 20:07
  the
• 20:07
  front
• 20:08
  door.
• 20:09
  The
• 20:09
  philosophical
• 20:09
  zombie
• 20:10
  thought
• 20:10
  experiment
• 20:11
  is
• 20:11
  self-defeating:
• 20:12
  a
• 20:12
  physically
• 20:13
  identical
• 20:13
  zombie
• 20:13
  twin
• 20:14
  would
• 20:14
  undergo
• 20:14
  the
• 20:15
  same
• 20:15
  brain
• 20:15
  processes
• 20:16
  during
• 20:16
  introspection
• 20:17
  and
• 20:17
  reach
• 20:17
  the
• 20:18
  same
• 20:18
  conclusion
• 20:18
  about
• 20:19
  having
• 20:19
  consciousness,
• 20:20
  which
• 20:20
  means
• 20:20
  introspection
• 20:21
  cannot
• 20:21
  distinguish
• 20:22
  genuine
• 20:22
  consciousness
• 20:23
  from
• 20:23
  its
• 20:23
  absence.
• 20:24
  Rovelli's
• 20:25
  position
• 20:25
  is
• 20:25
  that
• 20:26
  consciousness
• 20:26
  is
• 20:26
  a
• 20:26
  natural
• 20:27
  phenomenon
• 20:27
  —
• 20:28
  difficult
• 20:28
  to
• 20:29
  understand
• 20:29
  for
• 20:29
  the
• 20:29
  same
• 20:30
  reason
• 20:30
  thunderstorms
• 20:31
  are
• 20:31
  difficult
• 20:31
  to
• 20:31
  understand,
• 20:32
  not
• 20:32
  because
• 20:33
  it
• 20:33
  belongs
• 20:33
  to
• 20:33
  a
• 20:33
  separate
• 20:34
  metaphysical
• 20:34
  domain.
• 20:35
  The
• 20:36
  soul
• 20:36
  is
• 20:36
  real
• 20:36
  and
• 20:37
  is
• 20:37
  part
• 20:37
  of
• 20:37
  nature.
• 20:38
  "It
• 20:38
  is
• 20:38
  time
• 20:39
  to
• 20:39
  give
• 20:39
  up
• 20:39
  the
• 20:39
  pernicious
• 20:39
  dualism
• 20:40
  introduced
• 20:40
  by
• 20:41
  the
• 20:41
  debate
• 20:41
  on
• 20:41
  consciousness
• 20:42
  and
• 20:42
  embrace
• 20:43
  the
• 20:43
  reality
• 20:43
  that
• 20:43
  our
• 20:44
  soul,
• 20:44
  or
• 20:44
  our
• 20:45
  spiritual
• 20:45
  life,
• 20:46
  is
• 20:46
  consistent
• 20:46
  with
• 20:47
  our
• 20:47
  fundamental
• 20:47
  physics."
• 20:49
  For
• 20:49
  the
• 20:49
  AI
• 20:49
  community,
• 20:50
  this
• 20:50
  has
• 20:50
  a
• 20:50
  pointed
• 20:51
  implication:
• 20:52
  if
• 20:52
  there
• 20:52
  is
• 20:52
  no
• 20:53
  metaphysical
• 20:53
  gap
• 20:54
  between
• 20:54
  mind
• 20:54
  and
• 20:54
  body,
• 20:55
  then
• 20:55
  the
• 20:55
  question
• 20:56
  of
• 20:56
  machine
• 20:56
  consciousness
• 20:57
  is
• 20:57
  purely
• 20:57
  empirical
• 20:58
  —
• 20:59
  a
• 20:59
  matter
• 20:59
  of
• 20:59
  engineering
• 21:00
  complexity,
• 21:01
  not
• 21:01
  ontological
• 21:02
  impossibility.
• 21:03
  GrapheneOS
• 21:03
  flagged
• 21:04
  a
• 21:04
  less
• 21:04
  philosophical
• 21:05
  but
• 21:05
  equally
• 21:05
  consequential
• 21:06
  form
• 21:06
  of
• 21:06
  control:
• 21:08
  Apple
• 21:08
  and
• 21:08
  Google
• 21:09
  are
• 21:09
  gradually
• 21:09
  expanding
• 21:10
  their
• 21:10
  use
• 21:10
  of
• 21:10
  hardware-based
• 21:11
  attestation
• 21:12
  through
• 21:12
  Play
• 21:12
  Integrity
• 21:13
  API
• 21:13
  and
• 21:14
  App
• 21:14
  Attest
• 21:14
  API,
• 21:15
  and
• 21:16
  both
• 21:16
  are
• 21:16
  bringing
• 21:16
  it
• 21:16
  to
• 21:16
  the
• 21:16
  web
• 21:17
  via
• 21:17
  Privacy
• 21:17
  Pass.
• 21:18
  The
• 21:19
  same
• 21:19
  TPM
• 21:19
  and
• 21:19
  secure
• 21:20
  enclave
• 21:20
  infrastructure
• 21:21
  that
• 21:21
  security
• 21:22
  researchers
• 21:22
  championed
• 21:23
  as
• 21:23
  the
• 21:23
  foundation
• 21:24
  for
• 21:24
  prompt
• 21:24
  integrity
• 21:25
  and
• 21:25
  software
• 21:25
  verification
• 21:26
  becomes,
• 21:26
  from
• 21:27
  another
• 21:27
  angle,
• 21:28
  the
• 21:28
  mechanism
• 21:28
  by
• 21:28
  which
• 21:28
  a
• 21:29
  handful
• 21:29
  of
• 21:29
  silicon
• 21:29
  vendors
• 21:30
  and
• 21:30
  platform
• 21:30
  operators
• 21:31
  could
• 21:31
  lock
• 21:32
  out
• 21:32
  independent
• 21:32
  software
• 21:33
  deployment
• 21:33
  entirely.
• 21:34
  When
• 21:34
  attestation
• 21:35
  roots
• 21:35
  are
• 21:35
  controlled
• 21:36
  by
• 21:36
  platform
• 21:36
  gatekeepers,
• 21:37
  "security"
• 21:38
  and
• 21:38
  "monopoly
• 21:39
  enforcement"
• 21:40
  become
• 21:40
  architecturally
• 21:41
  indistinguishable.
• 21:42
  For
• 21:42
  the
• 21:42
  AI
• 21:42
  inference
• 21:43
  community
• 21:43
  —
• 21:44
  where
• 21:44
  running
• 21:44
  models
• 21:45
  locally
• 21:45
  on
• 21:45
  your
• 21:45
  own
• 21:46
  hardware
• 21:46
  is
• 21:46
  the
• 21:46
  entire
• 21:47
  point
• 21:47
  —
• 21:48
  hardware
• 21:48
  attestation
• 21:49
  that
• 21:49
  decides
• 21:49
  which
• 21:50
  software
• 21:50
  is
• 21:50
  allowed
• 21:50
  to
• 21:51
  execute
• 21:51
  is
• 21:51
  an
• 21:52
  existential
• 21:52
  concern
• 21:53
  hiding
• 21:53
  in
• 21:53
  plain
• 21:53
  sight.