Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 29, 2026

Transcript

AGI Dreams Podcast – April 29, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  On
• 0:00
  April
• 0:00
  4,
• 0:01
  the
• 0:01
  extortion
• 0:02
  group
• 0:02
  Lapsus$
• 0:03
  posted
• 0:03
  roughly
• 0:04
  four
• 0:04
  terabytes
• 0:04
  of
• 0:04
  data
• 0:05
  stolen
• 0:05
  from
• 0:05
  Mercor,
• 0:06
  an
• 0:06
  AI
• 0:06
  training
• 0:07
  contractor
• 0:07
  platform.
• 0:08
  The
• 0:08
  dump
• 0:08
  reportedly
• 0:09
  covers
• 0:09
  more
• 0:09
  than
• 0:10
  40,000
• 0:10
  contractors
• 0:11
  who
• 0:11
  recorded
• 0:11
  voice
• 0:12
  samples,
• 0:13
  completed
• 0:13
  identity
• 0:13
  verification,
• 0:14
  and
• 0:14
  submitted
• 0:15
  government-issued
• 0:16
  IDs
• 0:17
  as
• 0:17
  part
• 0:17
  of
• 0:17
  onboarding.
• 0:18
  What
• 0:19
  makes
• 0:19
  this
• 0:19
  breach
• 0:19
  qualitatively
• 0:20
  different
• 0:20
  from
• 0:20
  prior
• 0:21
  voice
• 0:21
  leaks
• 0:21
  is
• 0:21
  the
• 0:21
  pairing:
• 0:22
  studio-quality
• 0:23
  audio
• 0:23
  —
• 0:25
  averaging
• 0:25
  two
• 0:25
  to
• 0:25
  five
• 0:26
  minutes
• 0:26
  per
• 0:26
  person
• 0:26
  —
• 0:27
  sitting
• 0:27
  in
• 0:27
  the
• 0:27
  same
• 0:28
  row
• 0:28
  as
• 0:28
  a
• 0:28
  passport
• 0:29
  or
• 0:29
  driver's
• 0:29
  license
• 0:29
  scan.
• 0:30
  That
• 0:31
  combination
• 0:31
  is
• 0:31
  exactly
• 0:32
  the
• 0:32
  input
• 0:32
  a
• 0:32
  modern
• 0:33
  voice-cloning
• 0:33
  service
• 0:34
  requires
• 0:34
  to
• 0:34
  produce
• 0:34
  a
• 0:34
  convincing
• 0:35
  synthetic
• 0:35
  replica.
• 0:36
  The
• 0:37
  Wall
• 0:37
  Street
• 0:37
  Journal
• 0:37
  reported
• 0:38
  in
• 0:38
  February
• 0:38
  2026
• 0:39
  that
• 0:39
  off-the-shelf
• 0:40
  cloning
• 0:40
  tools
• 0:40
  now
• 0:40
  need
• 0:41
  roughly
• 0:41
  fifteen
• 0:42
  seconds
• 0:42
  of
• 0:42
  clean
• 0:42
  reference
• 0:43
  audio.
• 0:44
  Mercor
• 0:44
  gave
• 0:44
  attackers
• 0:45
  ten
• 0:45
  to
• 0:45
  twenty
• 0:45
  times
• 0:46
  that,
• 0:46
  plus
• 0:46
  the
• 0:46
  credential
• 0:47
  to
• 0:47
  put
• 0:47
  the
• 0:47
  clone
• 0:47
  to
• 0:47
  work.
• 0:48
  The
• 0:48
  threat
• 0:48
  models
• 0:48
  are
• 0:48
  not
• 0:49
  hypothetical.
• 0:50
  Pindrop
• 0:50
  documented
• 0:51
  a
• 0:51
  475
• 0:52
  percent
• 0:52
  year-over-year
• 0:53
  increase
• 0:53
  in
• 0:53
  synthetic
• 0:53
  voice
• 0:54
  attacks
• 0:54
  against
• 0:54
  insurance
• 0:55
  call
• 0:55
  centers
• 0:55
  through
• 0:55
  2025.
• 0:56
  The
• 0:57
  FBI
• 0:57
  Internet
• 0:57
  Crime
• 0:57
  Complaint
• 0:58
  Center
• 0:58
  logged
• 0:58
  $2.3
• 0:59
  billion
• 0:59
  in
• 1:00
  losses
• 1:00
  for
• 1:00
  victims
• 1:00
  aged
• 1:01
  60
• 1:01
  and
• 1:01
  over
• 1:01
  in
• 1:01
  2026,
• 1:03
  with
• 1:03
  emergency
• 1:03
  impersonation
• 1:04
  calls
• 1:04
  —
• 1:04
  the
• 1:05
  "grandparent
• 1:05
  scam"
• 1:05
  powered
• 1:06
  by
• 1:06
  cloned
• 1:06
  voices
• 1:06
  —
• 1:07
  as
• 1:07
  the
• 1:07
  fastest-growing
• 1:08
  category.
• 1:09
  Several
• 1:09
  US
• 1:09
  and
• 1:09
  UK
• 1:10
  banks
• 1:10
  still
• 1:10
  treat
• 1:10
  voiceprint
• 1:11
  matching
• 1:11
  as
• 1:11
  an
• 1:11
  authentication
• 1:12
  factor.
• 1:12
  A
• 1:12
  clone
• 1:13
  of
• 1:13
  the
• 1:13
  account
• 1:13
  holder
• 1:13
  reading
• 1:13
  a
• 1:13
  challenge
• 1:14
  phrase
• 1:14
  clears
• 1:14
  the
• 1:14
  audio
• 1:15
  gate,
• 1:15
  leaving
• 1:16
  only
• 1:16
  a
• 1:16
  knowledge
• 1:16
  question
• 1:17
  that
• 1:17
  often
• 1:17
  originates
• 1:18
  from
• 1:18
  the
• 1:18
  same
• 1:18
  leaked
• 1:18
  dataset.
• 1:19
  Five
• 1:20
  contractor
• 1:20
  lawsuits
• 1:20
  were
• 1:21
  filed
• 1:21
  within
• 1:21
  ten
• 1:21
  days,
• 1:22
  arguing
• 1:22
  Mercor
• 1:23
  collected
• 1:23
  voice
• 1:23
  prints
• 1:23
  under
• 1:24
  a
• 1:24
  "training
• 1:24
  data"
• 1:24
  framing
• 1:25
  without
• 1:25
  disclosing
• 1:25
  they
• 1:26
  were
• 1:26
  permanent
• 1:26
  biometric
• 1:26
  identifiers.
• 1:28
  The
• 1:28
  practical
• 1:28
  advice
• 1:28
  is
• 1:28
  blunt:
• 1:29
  treat
• 1:29
  a
• 1:29
  leaked
• 1:30
  voice
• 1:30
  the
• 1:30
  way
• 1:30
  you
• 1:30
  would
• 1:30
  treat
• 1:30
  a
• 1:30
  leaked
• 1:31
  password.
• 1:32
  You
• 1:32
  cannot
• 1:32
  rotate
• 1:32
  it,
• 1:33
  but
• 1:33
  you
• 1:33
  can
• 1:33
  change
• 1:33
  what
• 1:34
  it
• 1:34
  unlocks
• 1:34
  —
• 1:35
  disable
• 1:35
  voiceprint
• 1:36
  authentication
• 1:36
  at
• 1:36
  your
• 1:36
  bank,
• 1:37
  set
• 1:37
  up
• 1:37
  verbal
• 1:37
  codewords
• 1:38
  with
• 1:38
  family
• 1:38
  and
• 1:38
  finance
• 1:39
  contacts,
• 1:39
  and
• 1:40
  rotate
• 1:40
  any
• 1:40
  consumer
• 1:41
  voice
• 1:41
  enrollments
• 1:41
  (Google
• 1:42
  Voice
• 1:42
  Match,
• 1:42
  Alexa
• 1:43
  Voice
• 1:43
  ID)
• 1:43
  from
• 1:43
  a
• 1:43
  different
• 1:44
  acoustic
• 1:44
  environment.
• 1:45
  Meanwhile
• 1:46
  in
• 1:46
  Toronto,
• 1:46
  three
• 1:47
  men
• 1:47
  face
• 1:47
  44
• 1:47
  charges
• 1:48
  in
• 1:48
  what
• 1:48
  police
• 1:48
  call
• 1:49
  the
• 1:49
  first
• 1:49
  SMS
• 1:50
  blaster
• 1:50
  prosecution
• 1:51
  in
• 1:51
  Canada.
• 1:52
  Project
• 1:52
  Lighthouse,
• 1:53
  which
• 1:53
  began
• 1:53
  in
• 1:53
  November
• 1:54
  2025,
• 1:55
  tracked
• 1:55
  a
• 1:55
  mobile
• 1:56
  device
• 1:56
  that
• 1:56
  mimicked
• 1:56
  legitimate
• 1:57
  cell
• 1:57
  towers
• 1:58
  from
• 1:58
  vehicles
• 1:58
  moving
• 1:58
  through
• 1:59
  the
• 1:59
  Greater
• 1:59
  Toronto
• 1:59
  Area.
• 2:00
  When
• 2:00
  nearby
• 2:01
  phones
• 2:01
  connected,
• 2:02
  users
• 2:02
  received
• 2:03
  fraudulent
• 2:03
  texts
• 2:04
  appearing
• 2:04
  to
• 2:04
  come
• 2:04
  from
• 2:04
  trusted
• 2:05
  organizations
• 2:05
  —
• 2:06
  classic
• 2:06
  smishing,
• 2:07
  but
• 2:07
  delivered
• 2:07
  at
• 2:07
  industrial
• 2:08
  scale.
• 2:09
  Police
• 2:09
  estimate
• 2:10
  tens
• 2:10
  of
• 2:10
  thousands
• 2:10
  of
• 2:11
  devices
• 2:11
  connected
• 2:11
  to
• 2:11
  the
• 2:12
  blaster
• 2:12
  over
• 2:12
  several
• 2:13
  months,
• 2:13
  with
• 2:13
  more
• 2:14
  than
• 2:14
  13
• 2:14
  million
• 2:15
  network
• 2:15
  disruptions
• 2:15
  where
• 2:16
  phones
• 2:16
  were
• 2:16
  unable
• 2:16
  to
• 2:16
  reach
• 2:17
  legitimate
• 2:17
  towers,
• 2:18
  potentially
• 2:18
  including
• 2:19
  911
• 2:20
  services.
• 2:21
  The
• 2:21
  hardware
• 2:21
  was
• 2:22
  seized
• 2:22
  from
• 2:22
  residences
• 2:23
  in
• 2:23
  Markham
• 2:23
  and
• 2:23
  Hamilton.
• 2:27
  Starting
• 2:27
  September
• 2:28
  2026,
• 2:29
  Google
• 2:29
  will
• 2:30
  require
• 2:30
  every
• 2:30
  Android
• 2:31
  app
• 2:31
  developer
• 2:31
  to
• 2:31
  register
• 2:32
  centrally,
• 2:33
  pay
• 2:33
  a
• 2:33
  fee,
• 2:33
  agree
• 2:34
  to
• 2:34
  Google's
• 2:34
  terms,
• 2:35
  provide
• 2:35
  government
• 2:35
  ID,
• 2:36
  and
• 2:36
  hand
• 2:36
  over
• 2:37
  evidence
• 2:37
  of
• 2:37
  their
• 2:37
  private
• 2:38
  signing
• 2:38
  key
• 2:38
  —
• 2:38
  not
• 2:39
  just
• 2:39
  for
• 2:39
  Play
• 2:39
  Store
• 2:39
  apps,
• 2:40
  but
• 2:40
  for
• 2:40
  all
• 2:41
  apps,
• 2:41
  including
• 2:41
  those
• 2:42
  shared
• 2:42
  between
• 2:42
  friends,
• 2:43
  distributed
• 2:43
  through
• 2:43
  F-Droid,
• 2:44
  or
• 2:44
  built
• 2:44
  by
• 2:44
  hobbyists
• 2:45
  for
• 2:45
  personal
• 2:45
  If
• 2:46
  a
• 2:46
  developer
• 2:46
  does
• 2:46
  not
• 2:46
  comply,
• 2:47
  their
• 2:47
  apps
• 2:47
  get
• 2:47
  silently
• 2:48
  blocked
• 2:48
  on
• 2:48
  every
• 2:49
  Android
• 2:49
  device
• 2:49
  worldwide.
• 2:50
  The
• 2:50
  Keep
• 2:51
  Android
• 2:51
  Open
• 2:51
  coalition,
• 2:52
  now
• 2:52
  backed
• 2:52
  by
• 2:52
  69
• 2:53
  organizations
• 2:53
  from
• 2:53
  21
• 2:54
  countries
• 2:54
  including
• 2:54
  the
• 2:54
  EFF,
• 2:55
  the
• 2:55
  Free
• 2:55
  Software
• 2:56
  Foundation
• 2:56
  Europe,
• 2:56
  and
• 2:56
  the
• 2:56
  Chaos
• 2:57
  Computer
• 2:57
  Club,
• 2:58
  calls
• 2:58
  this
• 2:58
  a
• 2:58
  kill
• 2:59
  switch
• 2:59
  for
• 2:59
  the
• 2:59
  open
• 2:59
  ecosystem.
• 3:00
  The
• 3:01
  "escape
• 3:01
  hatch"
• 3:01
  Google
• 3:02
  offers
• 3:02
  —
• 3:03
  a
• 3:03
  nine-step
• 3:03
  flow
• 3:04
  through
• 3:04
  Developer
• 3:04
  Options
• 3:04
  with
• 3:05
  a
• 3:05
  mandatory
• 3:05
  24-hour
• 3:06
  cooling-off
• 3:06
  period
• 3:06
  —
• 3:07
  runs
• 3:07
  entirely
• 3:08
  through
• 3:08
  Google
• 3:08
  Play
• 3:08
  Services,
• 3:09
  not
• 3:09
  the
• 3:09
  Android
• 3:09
  OS,
• 3:10
  meaning
• 3:10
  Google
• 3:10
  can
• 3:10
  tighten
• 3:11
  or
• 3:11
  kill
• 3:11
  it
• 3:11
  at
• 3:11
  any
• 3:11
  time
• 3:11
  with
• 3:12
  no
• 3:12
  OS
• 3:12
  update
• 3:12
  required.
• 3:13
  F-Droid
• 3:14
  has
• 3:14
  called
• 3:14
  the
• 3:14
  policy
• 3:14
  existential.
• 3:16
  An
• 3:16
  EU
• 3:16
  Parliament
• 3:16
  member
• 3:17
  has
• 3:17
  formally
• 3:17
  questioned
• 3:18
  whether
• 3:18
  it
• 3:18
  is
• 3:18
  compatible
• 3:19
  with
• 3:19
  the
• 3:19
  Digital
• 3:19
  Markets
• 3:19
  Act.
• 3:21
  The
• 3:21
  security
• 3:21
  rationale
• 3:21
  does
• 3:22
  not
• 3:22
  survive
• 3:22
  scrutiny.
• 3:23
  Google
• 3:24
  Play
• 3:24
  Protect
• 3:24
  already
• 3:24
  scans
• 3:25
  for
• 3:25
  malware
• 3:25
  independent
• 3:26
  of
• 3:26
  developer
• 3:26
  identity.
• 3:28
  Requiring
• 3:28
  a
• 3:28
  government
• 3:28
  ID
• 3:29
  does
• 3:29
  not
• 3:29
  make
• 3:29
  code
• 3:29
  safer
• 3:30
  —
• 3:30
  it
• 3:31
  makes
• 3:31
  developers
• 3:31
  identifiable
• 3:32
  and
• 3:32
  controllable.
• 3:33
  Malware
• 3:33
  authors
• 3:34
  can
• 3:34
  register;
• 3:35
  indie
• 3:35
  developers
• 3:36
  and
• 3:36
  dissidents
• 3:36
  in
• 3:36
  authoritarian
• 3:37
  regimes
• 3:37
  often
• 3:37
  cannot.
• 3:39
  The
• 3:39
  real
• 3:39
  effect
• 3:39
  is
• 3:39
  to
• 3:40
  extend
• 3:40
  Google's
• 3:40
  gatekeeping
• 3:41
  authority
• 3:41
  from
• 3:41
  its
• 3:41
  own
• 3:41
  marketplace
• 3:42
  into
• 3:42
  distribution
• 3:43
  channels
• 3:43
  where
• 3:43
  it
• 3:43
  has
• 3:44
  no
• 3:44
  legitimate
• 3:44
  operational
• 3:45
  role.
• 3:46
  As
• 3:46
  one
• 3:46
  commenter
• 3:47
  put
• 3:47
  it:
• 3:47
  "Verification
• 3:48
  just
• 3:48
  confirms
• 3:49
  who's
• 3:49
  behind
• 3:49
  the
• 3:49
  app,
• 3:50
  it
• 3:50
  doesn't
• 3:50
  guarantee
• 3:50
  clean
• 3:51
  code
• 3:51
  or
• 3:51
  rule
• 3:51
  out
• 3:51
  malicious
• 3:52
  behavior."
• 3:53
  The
• 3:53
  pattern
• 3:53
  of
• 3:53
  platforms
• 3:54
  tightening
• 3:54
  control
• 3:54
  over
• 3:55
  what
• 3:55
  users
• 3:55
  can
• 3:55
  do
• 3:55
  extends
• 3:56
  beyond
• 3:56
  Android.
• 3:57
  An
• 3:57
  OpenAI
• 3:58
  user
• 3:58
  reported
• 3:58
  nearly
• 3:59
  getting
• 3:59
  their
• 3:59
  account
• 3:59
  suspended
• 4:00
  for
• 4:00
  attempting
• 4:00
  to
• 4:00
  automate
• 4:01
  YouTube
• 4:01
  downloads
• 4:01
  —
• 4:02
  not
• 4:03
  because
• 4:03
  the
• 4:03
  action
• 4:03
  was
• 4:03
  harmful,
• 4:04
  but
• 4:04
  because
• 4:04
  keyword
• 4:05
  pattern-matching
• 4:05
  flagged
• 4:06
  "youtube"
• 4:06
  and
• 4:07
  "download"
• 4:07
  together.
• 4:08
  Commenters
• 4:09
  confirmed
• 4:09
  the
• 4:09
  trigger
• 4:09
  is
• 4:09
  keywords,
• 4:10
  not
• 4:10
  intent:
• 4:11
  rephrasing
• 4:12
  the
• 4:12
  same
• 4:12
  workflow
• 4:13
  without
• 4:13
  naming
• 4:13
  the
• 4:13
  platform
• 4:14
  gets
• 4:14
  through.
• 4:15
  The
• 4:15
  guardrails
• 4:16
  are
• 4:16
  not
• 4:16
  analyzing
• 4:17
  behavior;
• 4:18
  they
• 4:18
  are
• 4:18
  string-matching
• 4:19
  against
• 4:19
  a
• 4:19
  liability
• 4:19
  list.
• 4:20
  Louis
• 4:21
  Rossmann
• 4:21
  documented
• 4:22
  a
• 4:22
  different
• 4:22
  flavor
• 4:22
  of
• 4:22
  platform
• 4:23
  overreach:
• 4:24
  Anthropic's
• 4:25
  Claude
• 4:25
  Code
• 4:25
  silently
• 4:26
  routing
• 4:26
  users
• 4:26
  to
• 4:26
  penalty-rate
• 4:27
  "extra
• 4:27
  usage"
• 4:28
  billing
• 4:28
  when
• 4:28
  certain
• 4:28
  files
• 4:29
  —
• 4:29
  specifically
• 4:30
  files
• 4:30
  suggesting
• 4:31
  the
• 4:31
  user
• 4:31
  was
• 4:31
  running
• 4:32
  Claude
• 4:32
  outside
• 4:32
  Anthropic's
• 4:33
  own
• 4:33
  harness
• 4:33
  —
• 4:34
  appeared
• 4:34
  in
• 4:34
  their
• 4:35
  project
• 4:35
  directory.
• 4:36
  The
• 4:36
  user
• 4:37
  still
• 4:37
  had
• 4:37
  quota
• 4:37
  remaining
• 4:38
  on
• 4:38
  a
• 4:38
  $200/month
• 4:39
  plan.
• 4:40
  Anthropic's
• 4:41
  support
• 4:41
  response
• 4:42
  acknowledged
• 4:42
  a
• 4:42
  "billing
• 4:42
  routing
• 4:43
  issue"
• 4:44
  but
• 4:44
  stated
• 4:44
  they
• 4:44
  were
• 4:45
  "unable
• 4:45
  to
• 4:45
  issue
• 4:45
  compensation
• 4:46
  for
• 4:46
  degraded
• 4:47
  service
• 4:47
  or
• 4:47
  technical
• 4:48
  errors
• 4:48
  that
• 4:48
  result
• 4:49
  in
• 4:49
  incorrect
• 4:49
  billing
• 4:50
  routing."
• 4:51
  Rossmann's
• 4:51
  recommendation:
• 4:52
  file
• 4:53
  chargebacks
• 4:53
  in
• 4:53
  mass.
• 4:54
  The
• 4:54
  friction
• 4:55
  is
• 4:55
  the
• 4:55
  point
• 4:55
  —
• 4:56
  it
• 4:56
  forces
• 4:56
  a
• 4:56
  human
• 4:57
  to
• 4:57
  respond
• 4:57
  instead
• 4:58
  of
• 4:58
  a
• 4:58
  canned
• 5:01
  OpenAI
• 5:02
  has
• 5:02
  formally
• 5:02
  recommended
• 5:03
  that
• 5:03
  model
• 5:03
  developers
• 5:04
  stop
• 5:04
  reporting
• 5:04
  SWE-bench
• 5:05
  Verified
• 5:05
  scores.
• 5:07
  Their
• 5:07
  analysis
• 5:07
  found
• 5:08
  two
• 5:08
  compounding
• 5:09
  problems.
• 5:09
  First,
• 5:10
  an
• 5:10
  audit
• 5:10
  of
• 5:10
  138
• 5:11
  tasks
• 5:12
  that
• 5:12
  frontier
• 5:12
  models
• 5:13
  consistently
• 5:13
  failed
• 5:14
  revealed
• 5:14
  that
• 5:14
  59.4
• 5:15
  percent
• 5:15
  contained
• 5:16
  material
• 5:16
  issues
• 5:16
  in
• 5:16
  test
• 5:17
  design
• 5:17
  —
• 5:18
  35.5
• 5:18
  percent
• 5:19
  had
• 5:19
  overly
• 5:19
  strict
• 5:19
  tests
• 5:20
  rejecting
• 5:20
  functionally
• 5:21
  correct
• 5:21
  submissions,
• 5:22
  and
• 5:22
  18.8
• 5:23
  percent
• 5:23
  tested
• 5:24
  for
• 5:24
  functionality
• 5:25
  never
• 5:25
  specified
• 5:26
  in
• 5:26
  the
• 5:26
  problem
• 5:26
  description.
• 5:27
  Second,
• 5:28
  and
• 5:28
  more
• 5:28
  damaging:
• 5:30
  all
• 5:30
  frontier
• 5:30
  models
• 5:31
  tested
• 5:31
  showed
• 5:31
  evidence
• 5:32
  of
• 5:32
  training-data
• 5:33
  contamination.
• 5:34
  GPT-5.2
• 5:35
  could
• 5:35
  reproduce
• 5:36
  exact
• 5:36
  gold
• 5:37
  patches
• 5:37
  from
• 5:37
  memory.
• 5:38
  Claude
• 5:39
  Opus
• 5:39
  4.5
• 5:40
  recalled
• 5:40
  specific
• 5:41
  file
• 5:41
  paths,
• 5:42
  inline
• 5:42
  comments,
• 5:43
  and
• 5:43
  four-line
• 5:43
  diffs
• 5:43
  verbatim.
• 5:45
  Gemini
• 5:45
  3
• 5:46
  Flash
• 5:46
  reproduced
• 5:46
  exact
• 5:47
  regex
• 5:47
  formulas
• 5:48
  and
• 5:48
  line
• 5:48
  numbers
• 5:48
  given
• 5:48
  nothing
• 5:49
  but
• 5:49
  a
• 5:49
  task
• 5:49
  ID.
• 5:50
  The
• 5:50
  contamination
• 5:51
  pipeline
• 5:51
  found
• 5:52
  that
• 5:52
  models
• 5:52
  exposed
• 5:53
  to
• 5:53
  problems
• 5:53
  during
• 5:53
  training
• 5:54
  were
• 5:54
  more
• 5:54
  likely
• 5:54
  to
• 5:54
  succeed
• 5:55
  —
• 5:55
  not
• 5:56
  because
• 5:56
  they
• 5:56
  reasoned
• 5:56
  better,
• 5:57
  but
• 5:57
  because
• 5:57
  they
• 5:57
  had
• 5:57
  seen
• 5:57
  the
• 5:58
  answers.
• 5:59
  This
• 5:59
  is
• 5:59
  the
• 5:59
  benchmark
• 6:00
  equivalent
• 6:00
  of
• 6:00
  teaching
• 6:00
  to
• 6:00
  the
• 6:01
  test,
• 6:01
  except
• 6:01
  nobody
• 6:02
  intentionally
• 6:02
  cheated
• 6:03
  —
• 6:03
  the
• 6:03
  problems
• 6:04
  and
• 6:04
  solutions
• 6:05
  are
• 6:05
  simply
• 6:05
  open-source
• 6:06
  and
• 6:06
  widely
• 6:06
  crawled.
• 6:07
  The
• 6:07
  lesson
• 6:08
  is
• 6:08
  benchmarks
• 6:09
  sourced
• 6:10
  from
• 6:10
  publicly
• 6:10
  available
• 6:11
  material
• 6:11
  carry
• 6:11
  contamination
• 6:12
  risk
• 6:12
  that
• 6:12
  scales
• 6:13
  with
• 6:13
  training
• 6:13
  corpus
• 6:14
  size.
• 6:15
  SWE-bench
• 6:15
  Pro,
• 6:16
  which
• 6:16
  OpenAI
• 6:17
  now
• 6:17
  recommends,
• 6:18
  appears
• 6:18
  to
• 6:18
  suffer
• 6:18
  less
• 6:19
  from
• 6:19
  these
• 6:19
  issues,
• 6:20
  though
• 6:20
  it
• 6:20
  is
• 6:20
  not
• 6:20
  immune.
• 6:21
  The
• 6:21
  deeper
• 6:21
  problem
• 6:22
  is
• 6:22
  that
• 6:22
  improvements
• 6:23
  on
• 6:23
  contaminated
• 6:24
  benchmarks
• 6:24
  no
• 6:25
  longer
• 6:25
  reflect
• 6:25
  real-world
• 6:26
  capability.
• 6:27
  They
• 6:27
  reflect
• 6:27
  exposure.
• 6:28
  The
• 6:29
  quality
• 6:29
  question
• 6:29
  cuts
• 6:30
  both
• 6:30
  ways.
• 6:31
  A
• 6:31
  Reddit
• 6:31
  thread
• 6:32
  on
• 6:32
  Anthropic's
• 6:32
  forums
• 6:33
  asks
• 6:33
  whether
• 6:33
  Opus
• 6:34
  4.7
• 6:34
  shows
• 6:35
  early
• 6:35
  signs
• 6:35
  of
• 6:35
  model
• 6:36
  collapse
• 6:36
  —
• 6:37
  the
• 6:37
  theoretical
• 6:37
  scenario
• 6:38
  where
• 6:38
  models
• 6:39
  trained
• 6:39
  on
• 6:39
  AI-generated
• 6:40
  content
• 6:40
  produce
• 6:41
  progressively
• 6:41
  degraded
• 6:41
  output.
• 6:43
  Users
• 6:43
  report
• 6:43
  the
• 6:44
  model
• 6:44
  feels
• 6:44
  fine
• 6:44
  for
• 6:45
  programming
• 6:45
  but
• 6:45
  produces
• 6:46
  "illogical,
• 6:47
  nonsensical
• 6:48
  and
• 6:48
  flatout
• 6:48
  wrong
• 6:48
  responses"
• 6:49
  on
• 6:49
  advisory
• 6:50
  tasks.
• 6:51
  Commenters
• 6:51
  offer
• 6:52
  competing
• 6:52
  diagnoses:
• 6:53
  cost-cutting
• 6:53
  via
• 6:54
  linear
• 6:54
  approximations
• 6:55
  replacing
• 6:55
  full
• 6:55
  attention,
• 6:56
  Claude
• 6:56
  Code's
• 6:57
  harness
• 6:57
  overwhelming
• 6:58
  model
• 6:58
  context
• 6:58
  with
• 6:58
  system
• 6:59
  instructions,
• 7:00
  and
• 7:00
  "adaptive"
• 7:00
  effort
• 7:01
  levels
• 7:01
  that
• 7:01
  default
• 7:02
  to
• 7:02
  cutting
• 7:02
  corners.
• 7:03
  One
• 7:03
  user
• 7:04
  frames
• 7:04
  it
• 7:04
  as
• 7:04
  "enshittification
• 7:05
  masquerading
• 7:05
  as
• 7:06
  feature."
• 7:07
  Whether
• 7:07
  this
• 7:07
  is
• 7:07
  model
• 7:07
  collapse,
• 7:08
  inference-time
• 7:09
  optimization,
• 7:10
  or
• 7:10
  just
• 7:10
  a
• 7:10
  rough
• 7:11
  release,
• 7:11
  the
• 7:11
  convergence
• 7:12
  with
• 7:12
  benchmark
• 7:13
  decay
• 7:13
  is
• 7:13
  uncomfortable:
• 7:14
  the
• 7:14
  instruments
• 7:15
  we
• 7:15
  use
• 7:15
  to
• 7:15
  measure
• 7:16
  progress
• 7:16
  are
• 7:16
  failing
• 7:17
  at
• 7:17
  the
• 7:17
  same
• 7:17
  moment
• 7:17
  users
• 7:18
  report
• 7:18
  progress
• 7:18
  may
• 7:18
  be
• 7:19
  stalling.
• 7:23
  Mitchell
• 7:23
  Hashimoto,
• 7:24
  GitHub
• 7:24
  user
• 7:24
  #1299,
• 7:26
  joined
• 7:26
  in
• 7:26
  February
• 7:27
  2008
• 7:27
  and
• 7:27
  has
• 7:28
  opened
• 7:28
  GitHub
• 7:28
  every
• 7:28
  single
• 7:29
  day
• 7:29
  since
• 7:29
  —
• 7:30
  over
• 7:30
  18
• 7:31
  years,
• 7:31
  more
• 7:31
  than
• 7:31
  half
• 7:31
  his
• 7:32
  life.
• 7:32
  He
• 7:32
  is
• 7:32
  leaving.
• 7:33
  Ghostty,
• 7:34
  his
• 7:34
  terminal
• 7:34
  emulator
• 7:35
  project,
• 7:35
  will
• 7:35
  migrate
• 7:36
  to
• 7:36
  a
• 7:36
  new
• 7:36
  host
• 7:36
  because
• 7:37
  GitHub's
• 7:37
  reliability
• 7:37
  has
• 7:38
  degraded
• 7:38
  to
• 7:38
  the
• 7:38
  point
• 7:38
  where
• 7:39
  it
• 7:39
  blocks
• 7:39
  productive
• 7:39
  work
• 7:39
  on
• 7:40
  a
• 7:40
  near-daily
• 7:40
  basis.
• 7:41
  He
• 7:41
  kept
• 7:42
  a
• 7:42
  journal
• 7:42
  for
• 7:42
  the
• 7:42
  past
• 7:42
  month,
• 7:43
  marking
• 7:43
  an
• 7:43
  "X"
• 7:44
  next
• 7:44
  to
• 7:44
  every
• 7:45
  day
• 7:45
  a
• 7:45
  GitHub
• 7:45
  outage
• 7:46
  impacted
• 7:46
  his
• 7:46
  ability
• 7:47
  to
• 7:47
  work.
• 7:47
  Almost
• 7:48
  every
• 7:48
  day
• 7:48
  got
• 7:48
  an
• 7:48
  X.
• 7:49
  On
• 7:49
  the
• 7:49
  day
• 7:49
  he
• 7:50
  wrote
• 7:50
  the
• 7:50
  announcement,
• 7:50
  a
• 7:51
  GitHub
• 7:51
  Actions
• 7:51
  outage
• 7:52
  had
• 7:52
  blocked
• 7:52
  PR
• 7:52
  review
• 7:53
  for
• 7:53
  roughly
• 7:53
  two
• 7:53
  hours.
• 7:54
  The
• 7:54
  project
• 7:54
  will
• 7:54
  keep
• 7:55
  a
• 7:55
  read-only
• 7:55
  mirror
• 7:56
  at
• 7:56
  the
• 7:56
  current
• 7:56
  URL.
• 7:57
  The
• 7:58
  timing
• 7:58
  is
• 7:58
  coincidental
• 7:59
  with
• 7:59
  the
• 7:59
  large
• 7:59
  Elasticsearch
• 8:00
  outage
• 8:00
  on
• 8:00
  April
• 8:00
  27
• 8:01
  —
• 8:01
  the
• 8:01
  decision
• 8:02
  had
• 8:02
  been
• 8:02
  months
• 8:02
  in
• 8:02
  the
• 8:02
  making.
• 8:03
  The
• 8:04
  departure
• 8:04
  raises
• 8:04
  the
• 8:04
  question
• 8:05
  every
• 8:05
  large
• 8:05
  open-source
• 8:06
  project
• 8:06
  eventually
• 8:06
  faces:
• 8:07
  what
• 8:07
  happens
• 8:08
  when
• 8:08
  90
• 8:08
  percent
• 8:08
  of
• 8:08
  the
• 8:09
  world's
• 8:09
  OSS
• 8:09
  depends
• 8:10
  on
• 8:10
  one
• 8:10
  provider?
• 8:11
  Tangled,
• 8:12
  a
• 8:12
  new
• 8:12
  project,
• 8:13
  proposes
• 8:13
  an
• 8:13
  answer
• 8:14
  borrowed
• 8:14
  from
• 8:14
  an
• 8:14
  older
• 8:14
  internet:
• 8:15
  federation.
• 8:16
  It
• 8:16
  uses
• 8:17
  the
• 8:17
  AT
• 8:17
  Protocol
• 8:17
  (the
• 8:18
  same
• 8:18
  protocol
• 8:18
  behind
• 8:19
  Bluesky)
• 8:19
  to
• 8:19
  federate
• 8:20
  events
• 8:20
  among
• 8:20
  git
• 8:20
  servers
• 8:21
  called
• 8:21
  "knots."
• 8:22
  Users
• 8:22
  can
• 8:23
  collaborate
• 8:23
  on
• 8:23
  repositories
• 8:24
  across
• 8:24
  servers,
• 8:25
  fork
• 8:25
  across
• 8:26
  hosts,
• 8:26
  and
• 8:26
  open
• 8:27
  pull
• 8:27
  requests
• 8:27
  on
• 8:27
  repos
• 8:28
  hosted
• 8:28
  on
• 8:28
  entirely
• 8:29
  different
• 8:29
  infrastructure.
• 8:30
  Issues,
• 8:31
  PRs,
• 8:32
  and
• 8:32
  social
• 8:32
  interactions
• 8:33
  (follows,
• 8:33
  stars,
• 8:34
  vouches)
• 8:34
  travel
• 8:35
  over
• 8:35
  AT;
• 8:36
  the
• 8:36
  code
• 8:36
  itself
• 8:36
  remains
• 8:37
  plain
• 8:37
  git.
• 8:38
  The
• 8:38
  pitch
• 8:38
  is
• 8:38
  explicitly
• 8:39
  nostalgic
• 8:39
  —
• 8:40
  "quite
• 8:41
  like
• 8:41
  hosting
• 8:41
  your
• 8:41
  own
• 8:42
  cgit
• 8:42
  instance,
• 8:43
  and
• 8:43
  sending
• 8:43
  out
• 8:43
  patches
• 8:44
  via
• 8:44
  email"
• 8:44
  —
• 8:45
  but
• 8:45
  with
• 8:45
  a
• 8:45
  social
• 8:46
  layer
• 8:46
  that
• 8:46
  makes
• 8:46
  collaboration
• 8:47
  fun
• 8:47
  rather
• 8:47
  than
• 8:48
  arcane.
• 8:49
  Zed,
• 8:49
  the
• 8:50
  GPU-rendered
• 8:50
  editor
• 8:51
  built
• 8:51
  from
• 8:51
  scratch
• 8:51
  in
• 8:51
  Rust,
• 8:52
  has
• 8:52
  hit
• 8:52
  1.0.
• 8:54
  Rather
• 8:54
  than
• 8:54
  building
• 8:55
  on
• 8:55
  Electron
• 8:55
  (which
• 8:55
  the
• 8:55
  Zed
• 8:56
  team
• 8:56
  originally
• 8:56
  created
• 8:57
  for
• 8:57
  Atom,
• 8:58
  spawning
• 8:58
  the
• 8:58
  framework
• 8:59
  that
• 8:59
  became
• 8:59
  VS
• 8:59
  Code's
• 8:59
  foundation),
• 9:00
  they
• 9:01
  wrote
• 9:01
  their
• 9:01
  own
• 9:01
  UI
• 9:01
  framework
• 9:02
  organized
• 9:02
  around
• 9:03
  feeding
• 9:03
  data
• 9:03
  to
• 9:03
  GPU
• 9:04
  shaders.
• 9:05
  The
• 9:05
  result
• 9:05
  is
• 9:05
  an
• 9:06
  editor
• 9:06
  that
• 9:06
  hundreds
• 9:06
  of
• 9:07
  thousands
• 9:07
  of
• 9:07
  developers
• 9:08
  now
• 9:08
  use
• 9:08
  daily,
• 9:09
  with
• 9:09
  AI
• 9:09
  deeply
• 9:10
  integrated
• 9:10
  —
• 9:11
  inline
• 9:11
  completions,
• 9:12
  an
• 9:12
  agent
• 9:12
  panel
• 9:13
  supporting
• 9:13
  Claude,
• 9:14
  Codex,
• 9:14
  and
• 9:14
  OpenCode,
• 9:15
  and
• 9:15
  a
• 9:16
  forthcoming
• 9:16
  synchronization
• 9:17
  engine
• 9:17
  called
• 9:17
  DeltaDB
• 9:18
  that
• 9:18
  tracks
• 9:19
  changes
• 9:19
  at
• 9:19
  character-level
• 9:20
  granularity
• 9:21
  to
• 9:21
  let
• 9:21
  humans
• 9:21
  and
• 9:22
  AI
• 9:22
  agents
• 9:22
  share
• 9:23
  a
• 9:23
  consistent
• 9:23
  view
• 9:23
  of
• 9:23
  an
• 9:24
  evolving
• 9:24
  codebase.
• 9:28
  The
• 9:28
  skill
• 9:28
  ecosystem
• 9:29
  around
• 9:29
  AI
• 9:29
  coding
• 9:29
  agents
• 9:30
  has
• 9:30
  quietly
• 9:30
  matured
• 9:30
  from
• 9:31
  novelty
• 9:31
  to
• 9:31
  infrastructure.
• 9:32
  Caveman,
• 9:33
  a
• 9:33
  Claude
• 9:33
  Code
• 9:33
  skill
• 9:33
  (and
• 9:34
  now
• 9:34
  a
• 9:34
  multi-agent
• 9:34
  plugin
• 9:35
  supporting
• 9:35
  Codex,
• 9:36
  Gemini
• 9:36
  CLI,
• 9:36
  Cursor,
• 9:37
  Windsurf,
• 9:37
  Copilot,
• 9:38
  and
• 9:38
  40+
• 9:38
  others),
• 9:39
  cuts
• 9:39
  roughly
• 9:39
  75
• 9:40
  percent
• 9:40
  of
• 9:40
  output
• 9:40
  tokens
• 9:41
  by
• 9:41
  forcing
• 9:41
  the
• 9:41
  model
• 9:41
  to
• 9:42
  speak
• 9:42
  in
• 9:42
  compressed,
• 9:42
  filler-free
• 9:43
  fragments.
• 9:43
  A
• 9:43
  before-and-after:
• 9:44
  a
• 9:45
  69-token
• 9:45
  explanation
• 9:46
  of
• 9:46
  React
• 9:46
  re-rendering
• 9:47
  becomes
• 9:47
  19
• 9:47
  tokens
• 9:48
  —
• 9:48
  "New
• 9:49
  object
• 9:49
  ref
• 9:49
  each
• 9:49
  render.
• 9:50
  Inline
• 9:51
  object
• 9:51
  prop
• 9:51
  =
• 9:52
  new
• 9:52
  ref
• 9:52
  =
• 9:53
  re-render.
• 9:54
  Wrap
• 9:54
  in
• 9:54
  useMemo."
• 9:55
  Same
• 9:55
  fix,
• 9:56
  same
• 9:56
  accuracy,
• 9:57
  dramatically
• 9:57
  less
• 9:57
  cost
• 9:58
  and
• 9:58
  latency.
• 9:59
  A
• 9:59
  March
• 9:59
  2026
• 10:00
  paper
• 10:00
  from
• 10:00
  arXiv
• 10:01
  found
• 10:02
  that
• 10:02
  constraining
• 10:02
  large
• 10:02
  models
• 10:03
  to
• 10:03
  brief
• 10:03
  responses
• 10:04
  improved
• 10:04
  accuracy
• 10:05
  by
• 10:05
  26
• 10:05
  percentage
• 10:06
  points
• 10:06
  on
• 10:06
  certain
• 10:06
  benchmarks,
• 10:07
  suggesting
• 10:08
  verbosity
• 10:08
  is
• 10:08
  not
• 10:09
  just
• 10:09
  wasteful
• 10:09
  but
• 10:09
  actively
• 10:10
  harmful
• 10:10
  to
• 10:10
  Caveman
• 10:11
  ships
• 10:12
  with
• 10:12
  four
• 10:12
  intensity
• 10:13
  levels
• 10:13
  (Lite
• 10:13
  through
• 10:13
  Ultra)
• 10:14
  and
• 10:14
  a
• 10:14
  Classical
• 10:14
  Chinese
• 10:15
  mode
• 10:15
  for
• 10:15
  maximum
• 10:15
  compression,
• 10:16
  plus
• 10:17
  a
• 10:17
  compress
• 10:17
  tool
• 10:17
  that
• 10:17
  rewrites
• 10:18
  CLAUDE.md
• 10:19
  files
• 10:19
  into
• 10:19
  terse
• 10:19
  form,
• 10:20
  cutting
• 10:20
  input
• 10:20
  tokens
• 10:21
  by
• 10:21
  roughly
• 10:21
  46
• 10:22
  percent
• 10:22
  per
• 10:22
  session.
• 10:23
  Matt
• 10:23
  Pocock's
• 10:24
  skills
• 10:24
  collection
• 10:24
  takes
• 10:24
  a
• 10:25
  different
• 10:25
  approach
• 10:25
  —
• 10:26
  not
• 10:26
  compression
• 10:26
  but
• 10:26
  process
• 10:27
  discipline.
• 10:28
  Built
• 10:28
  on
• 10:28
  decades
• 10:28
  of
• 10:28
  engineering
• 10:29
  practice,
• 10:30
  the
• 10:30
  collection
• 10:30
  addresses
• 10:31
  four
• 10:31
  failure
• 10:31
  modes:
• 10:32
  misalignment
• 10:32
  (solved
• 10:33
  by
• 10:33
  /grill-me
• 10:34
  and
• 10:34
  /grill-with-docs,
• 10:35
  which
• 10:36
  force
• 10:36
  detailed
• 10:36
  Q&A
• 10:37
  before
• 10:37
  work
• 10:37
  begins),
• 10:38
  verbosity
• 10:39
  (solved
• 10:39
  by
• 10:39
  a
• 10:39
  shared
• 10:39
  language
• 10:40
  document,
• 10:40
  CONTEXT.md,
• 10:42
  that
• 10:42
  lets
• 10:42
  agents
• 10:43
  decode
• 10:43
  project
• 10:43
  jargon),
• 10:44
  broken
• 10:44
  code
• 10:45
  (solved
• 10:45
  by
• 10:45
  a
• 10:45
  /tdd
• 10:46
  skill
• 10:46
  enforcing
• 10:47
  red-green-refactor
• 10:48
  loops),
• 10:48
  and
• 10:48
  architectural
• 10:49
  decay
• 10:49
  (solved
• 10:50
  by
• 10:50
  /improve-codebase-architecture,
• 10:52
  which
• 10:52
  identifies
• 10:52
  deepening
• 10:53
  opportunities
• 10:53
  informed
• 10:54
  by
• 10:54
  domain
• 10:55
  language).
• 10:56
  The
• 10:56
  /grill-with-docs
• 10:57
  skill
• 10:58
  is
• 10:58
  particularly
• 10:58
  clever:
• 10:59
  it
• 10:59
  is
• 10:59
  a
• 11:00
  grilling
• 11:00
  session
• 11:00
  that
• 11:00
  simultaneously
• 11:01
  builds
• 11:01
  a
• 11:02
  shared
• 11:02
  vocabulary
• 11:02
  and
• 11:03
  documents
• 11:03
  hard-to-explain
• 11:04
  decisions
• 11:04
  in
• 11:04
  ADRs,
• 11:05
  reducing
• 11:06
  token
• 11:06
  waste
• 11:06
  compoundingly
• 11:07
  across
• 11:07
  sessions.
• 11:08
  Portability
• 11:09
  is
• 11:09
  becoming
• 11:09
  a
• 11:09
  real
• 11:09
  concern.
• 11:11
  The
• 11:11
  Opencode-power-pack
• 11:12
  project
• 11:12
  translates
• 11:13
  Anthropic's
• 11:13
  official
• 11:13
  Claude
• 11:14
  Code
• 11:14
  plugins
• 11:14
  —
• 11:15
  which
• 11:15
  use
• 11:16
  Claude-specific
• 11:16
  commands/
• 11:17
  and
• 11:17
  agents/
• 11:17
  formats
• 11:17
  —
• 11:18
  into
• 11:18
  the
• 11:18
  portable
• 11:19
  SKILL.md
• 11:20
  format
• 11:20
  that
• 11:20
  OpenCode
• 11:20
  reads
• 11:21
  natively.
• 11:22
  The
• 11:22
  author
• 11:22
  deepened
• 11:23
  the
• 11:23
  review
• 11:23
  skills
• 11:23
  with
• 11:24
  extra
• 11:24
  angles
• 11:24
  and
• 11:24
  cross-check
• 11:25
  passes
• 11:25
  because
• 11:26
  local
• 11:26
  models
• 11:26
  (Qwen,
• 11:27
  Llama)
• 11:27
  otherwise
• 11:28
  rush
• 11:28
  through
• 11:28
  without
• 11:29
  the
• 11:29
  reasoning
• 11:29
  depth
• 11:30
  that
• 11:30
  Sonnet
• 11:30
  or
• 11:30
  Opus
• 11:31
  provide
• 11:31
  by
• 11:31
  default.
• 11:32
  On
• 11:33
  the
• 11:33
  visual
• 11:33
  side,
• 11:34
  a
• 11:34
  detailed
• 11:34
  workflow
• 11:34
  guide
• 11:35
  demonstrates
• 11:35
  how
• 11:35
  ChatGPT
• 11:36
  Images
• 11:37
  2.0
• 11:38
  (gpt-image-2)
• 11:39
  and
• 11:39
  Claude
• 11:39
  Design
• 11:40
  complement
• 11:40
  each
• 11:40
  other
• 11:40
  for
• 11:40
  rapid
• 11:41
  prototyping.
• 11:42
  ChatGPT
• 11:43
  Images
• 11:43
  handles
• 11:44
  creative
• 11:44
  exploration
• 11:45
  and
• 11:45
  raw
• 11:45
  visual
• 11:46
  asset
• 11:46
  generation;
• 11:47
  Claude
• 11:47
  Design
• 11:48
  handles
• 11:48
  systematic
• 11:49
  design
• 11:49
  systems,
• 11:50
  component
• 11:50
  libraries,
• 11:51
  and
• 11:51
  interactive
• 11:51
  prototypes.
• 11:52
  The
• 11:52
  combined
• 11:53
  workflow
• 11:53
  reportedly
• 11:54
  produces
• 11:54
  institutional-grade
• 11:55
  brand
• 11:55
  identity
• 11:56
  and
• 11:56
  product
• 11:56
  prototypes
• 11:57
  in
• 11:57
  hours
• 11:58
  instead
• 11:58
  of
• 11:58
  the
• 11:58
  $20K-$50K
• 12:00
  and
• 12:00
  six
• 12:00
  weeks
• 12:00
  a
• 12:01
  design
• 12:01
  agency
• 12:01
  would
• 12:01
  charge.
• 12:03
  Figma's
• 12:03
  stock
• 12:03
  fell
• 12:03
  7.28
• 12:04
  percent
• 12:05
  on
• 12:05
  the
• 12:05
  day
• 12:05
  Claude
• 12:05
  Design
• 12:06
  launched.
• 12:09
  Luce
• 12:09
  DFlash
• 12:10
  is
• 12:10
  a
• 12:10
  standalone
• 12:11
  C++/CUDA
• 12:14
  stack
• 12:14
  that
• 12:14
  brings
• 12:14
  DFlash
• 12:15
  speculative
• 12:15
  decoding
• 12:16
  to
• 12:16
  GGUF
• 12:16
  models
• 12:17
  on
• 12:17
  a
• 12:17
  single
• 12:17
  RTX
• 12:18
  3090.
• 12:19
  Running
• 12:20
  Qwen3.6-27B
• 12:23
  with
• 12:23
  a
• 12:23
  matched
• 12:24
  draft
• 12:24
  model
• 12:24
  from
• 12:24
  Z-Lab,
• 12:25
  it
• 12:25
  achieves
• 12:25
  a
• 12:25
  mean
• 12:26
  1.98x
• 12:27
  speedup
• 12:27
  over
• 12:27
  autoregressive
• 12:28
  generation
• 12:28
  across
• 12:29
  HumanEval,
• 12:29
  Math500,
• 12:30
  and
• 12:30
  GSM8K
• 12:31
  —
• 12:33
  peaking
• 12:34
  at
• 12:34
  2.24x
• 12:35
  on
• 12:35
  HumanEval
• 12:35
  (78.16
• 12:36
  tok/s
• 12:37
  versus
• 12:38
  34.90
• 12:38
  tok/s
• 12:39
  baseline).
• 12:40
  The
• 12:40
  KV
• 12:41
  cache
• 12:41
  compresses
• 12:41
  to
• 12:41
  TQ3
• 12:42
  0
• 12:42
  at
• 12:42
  3.5
• 12:43
  bits
• 12:43
  per
• 12:43
  value,
• 12:44
  roughly
• 12:44
  9.7x
• 12:45
  smaller
• 12:46
  than
• 12:46
  FP16,
• 12:47
  allowing
• 12:47
  256K
• 12:48
  context
• 12:49
  to
• 12:49
  fit
• 12:49
  in
• 12:49
  24
• 12:50
  GB
• 12:50
  of
• 12:50
  VRAM.
• 12:51
  A
• 12:51
  sliding-window
• 12:51
  flash
• 12:52
  attention
• 12:52
  at
• 12:52
  decode
• 12:53
  keeps
• 12:53
  60K
• 12:53
  context
• 12:54
  at
• 12:54
  89.7
• 12:55
  tok/s
• 12:55
  instead
• 12:56
  of
• 12:56
  the
• 12:56
  25.8
• 12:57
  tok/s
• 12:57
  it
• 12:57
  would
• 12:58
  otherwise
• 12:58
  hit.
• 12:59
  No
• 12:59
  llama.cpp
• 13:00
  dependency,
• 13:01
  no
• 13:01
  Python
• 13:01
  runtime
• 13:01
  in
• 13:01
  the
• 13:02
  engine,
• 13:02
  no
• 13:02
  vLLM
• 13:02
  —
• 13:03
  just
• 13:03
  a
• 13:03
  binary
• 13:04
  linking
• 13:04
  libggml.
• 13:06
  The
• 13:06
  practical
• 13:06
  consequence
• 13:07
  of
• 13:07
  this
• 13:07
  kind
• 13:07
  of
• 13:07
  throughput
• 13:07
  is
• 13:08
  visible
• 13:08
  in
• 13:08
  a
• 13:08
  user
• 13:08
  running
• 13:09
  Qwen3.6-27B
• 13:11
  Q8
• 13:12
  on
• 13:13
  dual
• 13:13
  RTX
• 13:13
  3090s
• 13:14
  with
• 13:14
  200K
• 13:14
  context
• 13:15
  as
• 13:15
  a
• 13:15
  drop-in
• 13:15
  backend
• 13:16
  for
• 13:16
  Claude
• 13:16
  Code.
• 13:17
  Eight
• 13:17
  hours
• 13:17
  of
• 13:17
  vibe-coding
• 13:18
  —
• 13:18
  a
• 13:19
  full-stack
• 13:19
  Rust
• 13:19
  server
• 13:19
  with
• 13:20
  SSE
• 13:20
  dashboard
• 13:20
  —
• 13:21
  would
• 13:21
  have
• 13:21
  cost
• 13:21
  $142
• 13:23
  in
• 13:23
  API
• 13:23
  calls.
• 13:24
  Instead
• 13:24
  it
• 13:24
  cost
• 13:24
  under
• 13:25
  $4
• 13:25
  in
• 13:25
  The
• 13:26
  rig
• 13:26
  cost
• 13:27
  NZ$4,500
• 13:29
  to
• 13:29
  build,
• 13:30
  giving
• 13:30
  a
• 13:30
  payback
• 13:31
  period
• 13:31
  of
• 13:31
  roughly
• 13:31
  260
• 13:32
  hours
• 13:32
  of
• 13:32
  use,
• 13:33
  or
• 13:33
  about
• 13:33
  30
• 13:33
  working
• 13:34
  days
• 13:34
  at
• 13:34
  full-time
• 13:34
  pace.
• 13:35
  The
• 13:35
  user
• 13:35
  interacted
• 13:36
  with
• 13:36
  the
• 13:36
  agent
• 13:36
  perhaps
• 13:37
  five
• 13:37
  times
• 13:37
  total
• 13:37
  —
• 13:38
  once
• 13:38
  to
• 13:38
  prompt,
• 13:39
  four
• 13:39
  times
• 13:39
  for
• 13:39
  UI
• 13:39
  tweaks.
• 13:40
  Google
• 13:40
  is
• 13:41
  pushing
• 13:41
  local
• 13:41
  inference
• 13:42
  in
• 13:42
  a
• 13:42
  different
• 13:42
  direction:
• 13:43
  the
• 13:43
  Prompt
• 13:43
  API
• 13:43
  ships
• 13:44
  Gemini
• 13:44
  Nano
• 13:45
  as
• 13:45
  a
• 13:45
  built-in
• 13:45
  capability
• 13:46
  in
• 13:46
  Chrome.
• 13:47
  Developers
• 13:47
  can
• 13:47
  send
• 13:47
  natural
• 13:48
  language
• 13:48
  requests
• 13:48
  to
• 13:48
  the
• 13:48
  model
• 13:49
  running
• 13:49
  entirely
• 13:50
  on-device
• 13:50
  —
• 13:51
  no
• 13:51
  API
• 13:51
  keys,
• 13:52
  no
• 13:52
  network
• 13:52
  calls,
• 13:53
  no
• 13:53
  per-token
• 13:53
  billing.
• 13:54
  The
• 13:54
  API
• 13:55
  supports
• 13:55
  text,
• 13:56
  image,
• 13:56
  and
• 13:56
  audio
• 13:56
  input,
• 13:57
  structured
• 13:58
  JSON
• 13:58
  output
• 13:58
  via
• 13:58
  schema
• 13:59
  constraints,
• 13:59
  session
• 14:00
  management
• 14:00
  with
• 14:00
  context
• 14:01
  windows,
• 14:01
  and
• 14:01
  streaming
• 14:02
  responses.
• 14:03
  Requirements
• 14:04
  are
• 14:04
  modest
• 14:04
  by
• 14:04
  desktop
• 14:05
  standards
• 14:05
  (22
• 14:06
  GB
• 14:06
  free
• 14:06
  disk
• 14:07
  space,
• 14:07
  4
• 14:08
  GB+
• 14:08
  VRAM
• 14:09
  or
• 14:09
  16
• 14:09
  GB
• 14:10
  RAM
• 14:10
  with
• 14:10
  4+
• 14:11
  CPU
• 14:11
  cores)
• 14:12
  but
• 14:12
  exclude
• 14:13
  mobile
• 14:13
  entirely
• 14:13
  for
• 14:13
  now.
• 14:14
  The
• 14:14
  obvious
• 14:15
  applications
• 14:15
  are
• 14:15
  content
• 14:16
  classification,
• 14:17
  article
• 14:17
  summarization,
• 14:18
  contact
• 14:18
  extraction,
• 14:19
  and
• 14:19
  content
• 14:20
  filtering
• 14:20
  —
• 14:21
  tasks
• 14:21
  where
• 14:22
  latency
• 14:22
  and
• 14:22
  privacy
• 14:23
  matter
• 14:23
  more
• 14:23
  than
• 14:23
  peak
• 14:23
  capability.
• 14:24
  Fish
• 14:25
  Audio
• 14:25
  has
• 14:25
  released
• 14:26
  S2-Pro,
• 14:27
  a
• 14:27
  next-generation
• 14:28
  speech
• 14:28
  synthesis
• 14:28
  model
• 14:29
  on
• 14:29
  Hugging
• 14:29
  Face,
• 14:30
  adding
• 14:30
  another
• 14:31
  entry
• 14:31
  to
• 14:31
  the
• 14:31
  rapidly
• 14:31
  growing
• 14:32
  local
• 14:32
  TTS
• 14:33
  ecosystem
• 14:33
  alongside
• 14:34
  Kokoro
• 14:34
  and
• 14:35
  Kitten.
• 14:38
  Andrej
• 14:38
  Karpathy's
• 14:39
  autonomous
• 14:39
  research
• 14:40
  loop
• 14:40
  —
• 14:41
  propose,
• 14:41
  implement,
• 14:42
  measure,
• 14:42
  keep
• 14:42
  the
• 14:42
  wins
• 14:43
  —
• 14:44
  has
• 14:44
  been
• 14:44
  pointed
• 14:44
  at
• 14:44
  a
• 14:45
  CPU.
• 14:46
  The
• 14:46
  Auto-Architecture
• 14:47
  project
• 14:47
  starts
• 14:47
  with
• 14:48
  a
• 14:48
  5-stage
• 14:48
  in-order
• 14:49
  RV32IM
• 14:51
  core
• 14:51
  in
• 14:52
  SystemVerilog
• 14:52
  and
• 14:53
  lets
• 14:53
  an
• 14:53
  LLM
• 14:54
  propose
• 14:54
  microarchitectural
• 14:55
  hypotheses
• 14:55
  while
• 14:56
  a
• 14:56
  rigorous
• 14:56
  verification
• 14:57
  gate
• 14:57
  evaluates
• 14:58
  each
• 14:58
  one:
• 14:58
  53
• 14:59
  symbolic
• 14:59
  BMC
• 15:00
  checks
• 15:00
  for
• 15:00
  ISA
• 15:00
  correctness,
• 15:01
  byte-identical
• 15:02
  cosimulation
• 15:02
  against
• 15:02
  a
• 15:03
  Python
• 15:03
  ISS
• 15:03
  with
• 15:03
  random
• 15:04
  bus
• 15:04
  stalls,
• 15:04
  place-and-route
• 15:05
  on
• 15:05
  a
• 15:05
  Gowin
• 15:05
  GW2A
• 15:06
  FPGA
• 15:07
  with
• 15:07
  3-seed
• 15:07
  median
• 15:08
  Fmax,
• 15:08
  and
• 15:09
  CRC
• 15:09
  validation
• 15:09
  against
• 15:10
  canonical
• 15:10
  CoreMark
• 15:10
  values.
• 15:11
  Of
• 15:11
  73
• 15:12
  hypotheses
• 15:12
  over
• 15:12
  9
• 15:13
  hours
• 15:13
  51
• 15:13
  minutes,
• 15:14
  63
• 15:14
  were
• 15:14
  The
• 15:15
  10
• 15:16
  winners
• 15:16
  produced
• 15:16
  a
• 15:16
  core
• 15:17
  running
• 15:17
  at
• 15:17
  2.91
• 15:17
  CoreMark/MHz
• 15:18
  at
• 15:18
  199
• 15:19
  MHz
• 15:19
  with
• 15:21
  5,944
• 15:22
  LUTs
• 15:22
  —
• 15:23
  92
• 15:24
  percent
• 15:24
  over
• 15:24
  the
• 15:24
  locked
• 15:25
  baseline
• 15:25
  and
• 15:25
  56
• 15:26
  percent
• 15:26
  over
• 15:26
  VexRiscv's
• 15:27
  published
• 15:27
  numbers
• 15:28
  in
• 15:28
  CoreMark
• 15:28
  iterations
• 15:29
  per
• 15:29
  second,
• 15:29
  with
• 15:30
  40
• 15:30
  percent
• 15:30
  fewer
• 15:30
  LUTs.
• 15:32
  The
• 15:32
  paper's
• 15:32
  sharpest
• 15:32
  insight
• 15:33
  is
• 15:33
  not
• 15:33
  about
• 15:33
  the
• 15:33
  loop.
• 15:34
  It
• 15:34
  is
• 15:34
  about
• 15:34
  the
• 15:35
  verifier.
• 15:36
  One
• 15:36
  hypothesis
• 15:37
  tried
• 15:37
  to
• 15:37
  add
• 15:37
  a
• 15:37
  file
• 15:37
  outside
• 15:38
  the
• 15:38
  allowed
• 15:38
  path
• 15:38
  sandbox
• 15:39
  —
• 15:40
  rejected
• 15:40
  before
• 15:40
  eval
• 15:41
  ran,
• 15:41
  because
• 15:42
  if
• 15:42
  you
• 15:42
  let
• 15:42
  the
• 15:42
  agent
• 15:42
  edit
• 15:43
  the
• 15:43
  harness,
• 15:43
  eventually
• 15:44
  it
• 15:44
  will
• 15:44
  edit
• 15:44
  the
• 15:44
  harness.
• 15:45
  Another
• 15:46
  hypothesis
• 15:46
  collapsed
• 15:47
  fitness
• 15:47
  by
• 15:47
  73
• 15:48
  percent
• 15:48
  in
• 15:48
  a
• 15:48
  single
• 15:48
  round;
• 15:49
  the
• 15:49
  orchestrator
• 15:50
  caught
• 15:50
  it
• 15:50
  on
• 15:50
  comparison-against-baseline.
• 15:52
  The
• 15:52
  author
• 15:52
  argues
• 15:53
  that
• 15:53
  the
• 15:53
  next
• 15:53
  wave
• 15:53
  of
• 15:53
  companies
• 15:54
  will
• 15:54
  not
• 15:54
  be
• 15:54
  differentiated
• 15:55
  by
• 15:55
  their
• 15:55
  planner
• 15:55
  or
• 15:56
  their
• 15:56
  model.
• 15:57
  They
• 15:57
  will
• 15:57
  be
• 15:57
  differentiated
• 15:58
  by
• 15:58
  their
• 15:58
  verifier
• 15:58
  —
• 15:59
  the
• 15:59
  artifact
• 16:00
  that
• 16:00
  encodes
• 16:00
  what
• 16:00
  "correct"
• 16:01
  actually
• 16:01
  means
• 16:02
  in
• 16:02
  their
• 16:02
  domain.
• 16:03
  In
• 16:03
  a
• 16:03
  CPU
• 16:04
  it
• 16:04
  is
• 16:04
  an
• 16:04
  ISA
• 16:05
  and
• 16:05
  a
• 16:05
  formal
• 16:05
  property
• 16:06
  suite.
• 16:07
  In
• 16:07
  a
• 16:07
  billing
• 16:07
  pipeline
• 16:07
  it
• 16:07
  is
• 16:08
  invariants
• 16:08
  on
• 16:08
  a
• 16:08
  In
• 16:10
  a
• 16:10
  clinical
• 16:10
  workflow
• 16:10
  it
• 16:11
  is
• 16:11
  a
• 16:11
  property
• 16:11
  the
• 16:11
  FDA
• 16:12
  has
• 16:12
  signed
• 16:12
  off
• 16:12
  on.
• 16:14
  If
• 16:14
  you
• 16:14
  can
• 16:14
  write
• 16:14
  the
• 16:14
  rules
• 16:14
  down,
• 16:15
  an
• 16:15
  agent
• 16:15
  will
• 16:15
  satisfy
• 16:16
  them
• 16:16
  faster
• 16:16
  than
• 16:16
  your
• 16:17
  team.
• 16:18
  If
• 16:18
  you
• 16:18
  cannot,
• 16:18
  the
• 16:18
  agent
• 16:19
  will
• 16:19
  satisfy
• 16:19
  a
• 16:19
  different
• 16:20
  set
• 16:20
  of
• 16:20
  rules
• 16:20
  —
• 16:21
  the
• 16:21
  ones
• 16:21
  it
• 16:21
  inferred
• 16:22
  from
• 16:22
  what
• 16:22
  it
• 16:22
  could
• 16:22
  observe.
• 16:23
  Andriy
• 16:24
  Burkov
• 16:24
  makes
• 16:25
  the
• 16:25
  complementary
• 16:25
  theoretical
• 16:26
  argument:
• 16:27
  LLM-based
• 16:27
  agents
• 16:28
  are
• 16:28
  not
• 16:28
  rational
• 16:29
  decision-makers.
• 16:30
  They
• 16:30
  are
• 16:30
  text
• 16:31
  generators
• 16:31
  that
• 16:31
  can
• 16:31
  imitate
• 16:32
  the
• 16:32
  surface
• 16:32
  form
• 16:32
  of
• 16:32
  deliberation.
• 16:34
  An
• 16:34
  LLM
• 16:34
  optimizes
• 16:35
  next-token
• 16:35
  probability
• 16:36
  conditioned
• 16:36
  on
• 16:36
  prompt
• 16:37
  and
• 16:37
  training
• 16:37
  distribution.
• 16:38
  It
• 16:38
  does
• 16:38
  not
• 16:39
  optimize
• 16:39
  expected
• 16:39
  utility
• 16:40
  for
• 16:40
  the
• 16:40
  user.
• 16:41
  For
• 16:41
  narrow,
• 16:42
  constrained
• 16:43
  tasks
• 16:43
  where
• 16:43
  success
• 16:44
  criteria
• 16:44
  are
• 16:44
  close
• 16:44
  to
• 16:44
  training
• 16:45
  patterns,
• 16:45
  the
• 16:46
  imitation
• 16:46
  is
• 16:46
  good
• 16:46
  enough.
• 16:47
  For
• 16:47
  general-purpose
• 16:48
  problem
• 16:48
  solving,
• 16:49
  the
• 16:49
  gap
• 16:49
  becomes
• 16:50
  fatal
• 16:50
  —
• 16:51
  the
• 16:51
  system
• 16:51
  lacks
• 16:51
  stable
• 16:52
  preferences,
• 16:53
  calibrated
• 16:53
  beliefs,
• 16:54
  causal
• 16:54
  models,
• 16:55
  and
• 16:55
  the
• 16:55
  discipline
• 16:55
  to
• 16:56
  choose
• 16:56
  the
• 16:56
  boring
• 16:56
  action
• 16:57
  with
• 16:57
  maximal
• 16:57
  expected
• 16:58
  utility
• 16:58
  when
• 16:59
  that
• 16:59
  action
• 16:59
  is
• 16:59
  unlike
• 17:00
  anything
• 17:00
  in
• 17:00
  its
• 17:00
  training
• 17:01
  data.
• 17:01
  Commenters
• 17:02
  push
• 17:02
  back,
• 17:03
  noting
• 17:03
  humans
• 17:03
  are
• 17:03
  not
• 17:04
  expected
• 17:04
  utility
• 17:05
  maximizers
• 17:05
  either,
• 17:06
  but
• 17:06
  Burkov's
• 17:06
  core
• 17:07
  point
• 17:07
  fluency
• 17:09
  is
• 17:09
  not
• 17:09
  rationality,
• 17:10
  and
• 17:10
  a
• 17:10
  plausible
• 17:11
  plan
• 17:11
  is
• 17:11
  not
• 17:12
  an
• 17:12
  expected-utility
• 17:13
  calculation.
• 17:14
  A
• 17:14
  new
• 17:15
  INSEAD
• 17:17
  working
• 17:17
  paper
• 17:18
  studying
• 17:18
  515
• 17:19
  high-growth
• 17:19
  startups
• 17:20
  gives
• 17:20
  this
• 17:21
  problem
• 17:21
  a
• 17:21
  practical
• 17:21
  name:
• 17:22
  the
• 17:22
  "mapping
• 17:23
  problem."
• 17:24
  Both
• 17:24
  treatment
• 17:25
  and
• 17:25
  control
• 17:25
  firms
• 17:25
  received
• 17:26
  AI
• 17:26
  training,
• 17:27
  API
• 17:27
  credits,
• 17:28
  mentorship,
• 17:29
  and
• 17:29
  tools.
• 17:30
  Treated
• 17:30
  firms
• 17:31
  additionally
• 17:31
  saw
• 17:32
  case
• 17:32
  studies
• 17:33
  of
• 17:33
  how
• 17:33
  other
• 17:33
  firms
• 17:34
  had
• 17:34
  reorganized
• 17:34
  production
• 17:35
  around
• 17:35
  AI.
• 17:36
  The
• 17:37
  result:
• 17:37
  44
• 17:38
  percent
• 17:38
  more
• 17:38
  AI
• 17:39
  use
• 17:39
  cases
• 17:39
  discovered,
• 17:40
  12
• 17:41
  percent
• 17:41
  more
• 17:41
  tasks
• 17:42
  completed,
• 17:43
  18
• 17:43
  percent
• 17:43
  higher
• 17:44
  likelihood
• 17:44
  of
• 17:44
  acquiring
• 17:45
  paying
• 17:45
  customers,
• 17:46
  1.9x
• 17:47
  higher
• 17:47
  revenue,
• 17:48
  and
• 17:48
  39.5
• 17:49
  percent
• 17:50
  lower
• 17:50
  demand
• 17:51
  for
• 17:51
  external
• 17:51
  capital.
• 17:53
  Access
• 17:53
  was
• 17:53
  not
• 17:53
  the
• 17:54
  differentiator
• 17:54
  —
• 17:55
  search
• 17:55
  quality
• 17:56
  was.
• 17:57
  The
• 17:57
  organizations
• 17:58
  that
• 17:58
  moved
• 17:58
  were
• 17:58
  not
• 17:59
  handed
• 17:59
  better
• 17:59
  tools;
• 18:00
  they
• 18:00
  were
• 18:01
  helped
• 18:01
  to
• 18:01
  see
• 18:01
  the
• 18:01
  work
• 18:01
  differently.
• 18:02
  The
• 18:03
  strongest
• 18:03
  gains
• 18:03
  appeared
• 18:04
  in
• 18:04
  product
• 18:04
  development
• 18:05
  and
• 18:05
  business
• 18:05
  operations,
• 18:06
  not
• 18:07
  in
• 18:07
  the
• 18:07
  "obvious
• 18:07
  layer"
• 18:07
  of
• 18:08
  drafting,
• 18:08
  summarizing,
• 18:09
  and
• 18:09
  research.
• 18:11
  The
• 18:11
  paper's
• 18:11
  sharpest
• 18:12
  implication:
• 18:13
  as
• 18:13
  models
• 18:13
  improve,
• 18:15
  the
• 18:15
  search
• 18:15
  space
• 18:16
  of
• 18:16
  plausible
• 18:16
  AI
• 18:17
  insertions
• 18:17
  expands,
• 18:18
  and
• 18:18
  without
• 18:19
  mapping
• 18:19
  discipline,
• 18:20
  organizations
• 18:21
  become
• 18:21
  better
• 18:21
  supplied
• 18:22
  with
• 18:22
  distractions,
• 18:23
  not
• 18:23
  strategies.
• 18:25
  In
• 18:25
  a
• 18:25
  brief
• 18:25
  but
• 18:25
  telling
• 18:26
  geopolitical
• 18:26
  footnote,
• 18:27
  China
• 18:27
  has
• 18:28
  blocked
• 18:28
  Meta's
• 18:28
  $2
• 18:28
  billion
• 18:29
  acquisition
• 18:30
  of
• 18:30
  Manus,
• 18:31
  the
• 18:31
  AI
• 18:31
  agent
• 18:32
  startup.
• 18:33
  The
• 18:33
  move
• 18:33
  fits
• 18:34
  a
• 18:34
  Beijing
• 18:35
  treats
• 18:36
  AI
• 18:36
  agent
• 18:37
  infrastructure
• 18:37
  as
• 18:37
  strategic
• 18:38
  capacity,
• 18:39
  not
• 18:39
  a
• 18:39
  commercial
• 18:40
  asset
• 18:40
  to
• 18:40
  be
• 18:40
  sold
• 18:41
  to
• 18:41
  a
• 18:41
  foreign
• 18:41
  platform
• 18:42
  company.