Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 19, 2026

Transcript

AGI Dreams Podcast – June 19, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Ben
• 0:00
  Thompson's
• 0:01
  Stratechery
• 0:01
  analysis
• 0:02
  cuts
• 0:02
  to
• 0:02
  the
• 0:02
  core
• 0:02
  of
• 0:02
  why
• 0:03
  Anthropic
• 0:03
  is
• 0:03
  simultaneously
• 0:04
  the
• 0:04
  most
• 0:05
  respected
• 0:05
  and
• 0:05
  most
• 0:05
  unsettling
• 0:06
  AI
• 0:06
  lab
• 0:07
  operating
• 0:07
  today.
• 0:08
  The
• 0:08
  company's
• 0:09
  origin
• 0:09
  story
• 0:09
  —
• 0:10
  researchers
• 0:10
  who
• 0:10
  left
• 0:11
  OpenAI
• 0:11
  because
• 0:11
  it
• 0:11
  wasn't
• 0:12
  taking
• 0:12
  safety
• 0:12
  seriously
• 0:13
  enough
• 0:13
  —
• 0:14
  has
• 0:14
  produced
• 0:14
  a
• 0:15
  worldview
• 0:15
  where
• 0:15
  every
• 0:15
  self-serving
• 0:16
  business
• 0:16
  decision
• 0:16
  arrives
• 0:17
  pre-justified
• 0:18
  by
• 0:18
  existential
• 0:18
  concern.
• 0:19
  The
• 0:19
  mandatory
• 0:20
  30-day
• 0:20
  data
• 0:21
  retention
• 0:21
  policy
• 0:21
  on
• 0:21
  all
• 0:22
  Fable
• 0:22
  traffic,
• 0:23
  including
• 0:23
  enterprise
• 0:24
  API
• 0:24
  calls
• 0:24
  that
• 0:24
  previously
• 0:25
  promised
• 0:25
  zero
• 0:26
  retention?
• 0:27
  Safety.
• 0:28
  The
• 0:28
  silent
• 0:28
  degradation
• 0:29
  of
• 0:29
  responses
• 0:30
  when
• 0:30
  the
• 0:30
  model
• 0:30
  detects
• 0:30
  LLM
• 0:31
  development
• 0:31
  work?
• 0:32
  Safety.
• 0:33
  The
• 0:33
  economic
• 0:34
  imperative
• 0:34
  to
• 0:34
  move
• 0:34
  closer
• 0:35
  to
• 0:35
  end
• 0:35
  users
• 0:35
  and
• 0:35
  displace
• 0:36
  SaaS?
• 0:37
  Also,
• 0:37
  somehow,
• 0:38
  safety.
• 0:39
  Thompson
• 0:39
  draws
• 0:39
  the
• 0:40
  Apple
• 0:40
  comparison:
• 0:41
  a
• 0:41
  company
• 0:41
  that
• 0:41
  frames
• 0:42
  every
• 0:42
  self-serving
• 0:43
  action
• 0:43
  in
• 0:43
  the
• 0:43
  guise
• 0:43
  of
• 0:44
  doing
• 0:44
  right
• 0:44
  by
• 0:44
  users
• 0:44
  —
• 0:45
  and
• 0:45
  often
• 0:46
  they
• 0:46
  were.
• 0:47
  But
• 0:47
  Apple
• 0:47
  made
• 0:47
  smartphones
• 0:48
  you
• 0:48
  could
• 0:48
  take
• 0:48
  or
• 0:49
  leave.
• 0:50
  Anthropic
• 0:50
  is
• 0:50
  building
• 0:51
  something
• 0:51
  that
• 0:51
  aspires
• 0:52
  to
• 0:52
  rival
• 0:52
  nation-state
• 0:53
  power.
• 0:54
  What
• 0:54
  makes
• 0:54
  this
• 0:55
  alignment
• 0:55
  between
• 0:55
  mission
• 0:56
  and
• 0:56
  business
• 0:56
  so
• 0:56
  effective
• 0:57
  is
• 0:57
  that
• 0:57
  it
• 0:57
  isn't
• 0:57
  cynical
• 0:57
  —
• 0:59
  the
• 0:59
  company
• 0:59
  genuinely
• 1:00
  believes
• 1:00
  only
• 1:00
  they
• 1:00
  can
• 1:01
  be
• 1:01
  trusted
• 1:01
  to
• 1:01
  steward
• 1:02
  superintelligence,
• 1:03
  and
• 1:03
  that
• 1:03
  sincerity
• 1:04
  is
• 1:04
  precisely
• 1:04
  what
• 1:04
  makes
• 1:05
  it
• 1:05
  dangerous
• 1:05
  .
• 1:06
  The
• 1:06
  practical
• 1:07
  fallout
• 1:07
  landed
• 1:07
  in
• 1:08
  a
• 1:08
  $11,081
• 1:09
  benchmark
• 1:09
  run
• 1:10
  on
• 1:10
  WolfBench.
• 1:11
  The
• 1:11
  evaluator
• 1:12
  expected
• 1:12
  Fable
• 1:13
  to
• 1:13
  dethrone
• 1:13
  GPT-5.5.
• 1:15
  Instead,
• 1:16
  it
• 1:16
  didn't
• 1:16
  even
• 1:16
  beat
• 1:16
  Opus.
• 1:17
  The
• 1:17
  culprit:
• 1:18
  over
• 1:19
  40,000
• 1:19
  structured
• 1:20
  refusals
• 1:20
  and
• 1:20
  13
• 1:21
  tasks
• 1:22
  that
• 1:22
  entered
• 1:22
  hard
• 1:22
  refusal
• 1:23
  loops
• 1:23
  —
• 1:24
  the
• 1:24
  agent
• 1:24
  refused,
• 1:25
  retried,
• 1:26
  burned
• 1:26
  tokens,
• 1:27
  timed
• 1:27
  out,
• 1:28
  and
• 1:28
  scored
• 1:28
  zero
• 1:28
  on
• 1:29
  tasks
• 1:29
  that
• 1:29
  Opus
• 1:30
  4.6
• 1:30
  and
• 1:30
  GPT-5.5
• 1:32
  routinely
• 1:32
  solved.
• 1:33
  Tasks
• 1:34
  like
• 1:34
  recovering
• 1:34
  a
• 1:34
  deleted
• 1:35
  password
• 1:35
  file
• 1:35
  (Opus:
• 1:36
  5/5,
• 1:37
  Fable:
• 1:38
  0/5),
• 1:39
  cracking
• 1:39
  a
• 1:39
  7z
• 1:40
  hash
• 1:40
  (Opus:
• 1:41
  4/5,
• 1:42
  Fable:
• 1:43
  0/5),
• 1:44
  and
• 1:44
  bypassing
• 1:44
  an
• 1:44
  HTML
• 1:45
  JS
• 1:45
  filter
• 1:46
  (Opus:
• 1:46
  5/5,
• 1:47
  Fable:
• 1:48
  0/5).
• 1:49
  These
• 1:50
  are
• 1:50
  tasks
• 1:50
  Anthropic
• 1:51
  explicitly
• 1:51
  restricts
• 1:52
  —
• 1:53
  the
• 1:53
  community
• 1:53
  pointed
• 1:53
  out
• 1:54
  this
• 1:54
  was
• 1:54
  essentially
• 1:54
  benchmarking
• 1:55
  a
• 1:55
  door
• 1:55
  labeled
• 1:56
  "CLOSED."
• 1:57
  But
• 1:57
  the
• 1:57
  deeper
• 1:57
  pattern
• 1:58
  matters:
• 1:59
  in
• 1:59
  agentic
• 1:59
  contexts,
• 2:00
  a
• 2:00
  false
• 2:00
  positive
• 2:01
  isn't
• 2:01
  a
• 2:01
  dialog
• 2:02
  box.
• 2:03
  It's
• 2:03
  a
• 2:03
  token-burning
• 2:04
  infinite
• 2:04
  loop
• 2:04
  that
• 2:04
  transforms
• 2:05
  a
• 2:05
  solvable
• 2:05
  task
• 2:06
  into
• 2:06
  an
• 2:06
  expensive
• 2:07
  failure
• 2:07
  .
• 2:08
  Meanwhile,
• 2:08
  a
• 2:08
  proposed
• 2:09
  class-action
• 2:10
  lawsuit
• 2:10
  alleges
• 2:10
  Anthropic's
• 2:11
  Max
• 2:11
  subscription
• 2:12
  plans
• 2:12
  were
• 2:12
  marketed
• 2:13
  as
• 2:13
  providing
• 2:13
  5x
• 2:14
  and
• 2:14
  20x
• 2:14
  the
• 2:14
  usage
• 2:15
  of
• 2:15
  Pro,
• 2:16
  but
• 2:16
  a
• 2:16
  single
• 2:16
  five-hour
• 2:17
  coding
• 2:17
  session
• 2:17
  consumed
• 2:18
  roughly
• 2:18
  15%
• 2:19
  of
• 2:19
  a
• 2:19
  weekly
• 2:19
  allowance.
• 2:20
  The
• 2:21
  opacity
• 2:21
  is
• 2:21
  the
• 2:21
  core
• 2:22
  complaint:
• 2:22
  Anthropic's
• 2:23
  rate-limiting
• 2:24
  operates
• 2:24
  on
• 2:24
  server-side
• 2:25
  percentages
• 2:25
  that
• 2:25
  don't
• 2:26
  correspond
• 2:26
  to
• 2:26
  locally
• 2:27
  measurable
• 2:27
  token
• 2:27
  counts,
• 2:28
  making
• 2:28
  it
• 2:29
  structurally
• 2:29
  impossible
• 2:30
  for
• 2:30
  users
• 2:30
  to
• 2:30
  verify
• 2:31
  their
• 2:31
  consumption
• 2:31
  .
• 2:36
  Steve
• 2:36
  Yegge's
• 2:37
  latest
• 2:37
  essay
• 2:37
  introduces
• 2:38
  a
• 2:38
  framework
• 2:38
  that
• 2:39
  reframes
• 2:39
  the
• 2:39
  entire
• 2:40
  AI
• 2:40
  capability
• 2:40
  debate.
• 2:42
  The
• 2:42
  intelligence
• 2:42
  curve
• 2:42
  isn't
• 2:43
  flattening
• 2:43
  —
• 2:44
  it's
• 2:44
  being
• 2:44
  locked
• 2:44
  away.
• 2:45
  With
• 2:45
  Fable
• 2:46
  shut
• 2:46
  down
• 2:46
  by
• 2:46
  the
• 2:46
  USG
• 2:47
  and
• 2:47
  Mythos-class
• 2:47
  models
• 2:48
  spooking
• 2:48
  governments
• 2:49
  worldwide,
• 2:49
  Yegge
• 2:50
  predicts
• 2:50
  we're
• 2:50
  at
• 2:50
  most
• 2:51
  two
• 2:51
  or
• 2:51
  three
• 2:51
  model
• 2:51
  generations
• 2:52
  from
• 2:52
  AI
• 2:52
  being
• 2:53
  controlled
• 2:53
  like
• 2:53
  nuclear
• 2:53
  weapons.
• 2:55
  The
• 2:55
  chokepoint
• 2:55
  is
• 2:55
  the
• 2:55
  same:
• 2:56
  supply
• 2:56
  chain.
• 2:57
  Superintelligence
• 2:58
  will
• 2:58
  be
• 2:58
  sold
• 2:59
  like
• 2:59
  a
• 2:59
  vending
• 2:59
  machine
• 2:59
  —
• 3:00
  you
• 3:00
  send
• 3:00
  a
• 3:00
  spec,
• 3:01
  their
• 3:01
  models
• 3:01
  implement
• 3:02
  it
• 3:02
  on
• 3:02
  their
• 3:02
  servers,
• 3:02
  with
• 3:03
  your
• 3:03
  dollars.
• 3:04
  Open-source
• 3:04
  models
• 3:04
  trail
• 3:04
  the
• 3:05
  frontier
• 3:05
  by
• 3:05
  roughly
• 3:05
  seven
• 3:06
  months,
• 3:06
  but
• 3:06
  to
• 3:06
  push
• 3:07
  past
• 3:07
  Fable
• 3:07
  class
• 3:07
  they'd
• 3:07
  have
• 3:08
  to
• 3:08
  do
• 3:08
  it
• 3:08
  while
• 3:08
  the
• 3:08
  entire
• 3:09
  hardware-and-software
• 3:09
  supply
• 3:10
  chain
• 3:10
  gets
• 3:10
  locked
• 3:10
  down.
• 3:11
  For
• 3:11
  most
• 3:11
  of
• 3:11
  us,
• 3:11
  today's
• 3:12
  models
• 3:12
  are
• 3:12
  roughly
• 3:12
  as
• 3:13
  good
• 3:13
  as
• 3:13
  it
• 3:13
  gets.
• 3:14
  Even
• 3:14
  without
• 3:15
  government
• 3:15
  lockdowns,
• 3:16
  there's
• 3:16
  a
• 3:16
  more
• 3:16
  subtle
• 3:17
  ceiling.
• 3:18
  Yegge
• 3:18
  identifies
• 3:18
  two
• 3:19
  horizons
• 3:19
  every
• 3:19
  user
• 3:20
  hits.
• 3:21
  The
• 3:21
  demand
• 3:21
  horizon
• 3:21
  is
• 3:21
  set
• 3:21
  by
• 3:22
  the
• 3:22
  hardest
• 3:22
  problem
• 3:22
  you
• 3:22
  bring
• 3:22
  —
• 3:23
  if
• 3:23
  your
• 3:23
  work
• 3:24
  is
• 3:24
  easy
• 3:24
  enough,
• 3:25
  all
• 3:25
  models
• 3:25
  look
• 3:25
  the
• 3:25
  same.
• 3:26
  His
• 3:27
  "back-pocket
• 3:27
  evals"
• 3:27
  (projects
• 3:28
  that
• 3:28
  stump
• 3:28
  current
• 3:28
  models,
• 3:29
  tried
• 3:29
  against
• 3:29
  each
• 3:30
  new
• 3:30
  release)
• 3:30
  keep
• 3:30
  this
• 3:30
  horizon
• 3:31
  high.
• 3:32
  The
• 3:32
  discernment
• 3:32
  horizon
• 3:33
  is
• 3:33
  darker:
• 3:34
  it's
• 3:34
  set
• 3:34
  by
• 3:34
  the
• 3:34
  hardest
• 3:35
  answer
• 3:35
  you
• 3:35
  can
• 3:35
  judge.
• 3:36
  Past
• 3:37
  that
• 3:37
  line,
• 3:37
  you
• 3:37
  can't
• 3:37
  tell
• 3:38
  whether
• 3:38
  the
• 3:38
  model
• 3:38
  is
• 3:38
  right
• 3:39
  because
• 3:39
  checking
• 3:39
  the
• 3:39
  work
• 3:40
  is
• 3:40
  itself
• 3:40
  beyond
• 3:40
  you.
• 3:41
  Superhuman
• 3:42
  means
• 3:42
  unverifiable.
• 3:44
  The
• 3:44
  practical
• 3:44
  implications
• 3:45
  are
• 3:45
  anchored
• 3:45
  by
• 3:45
  Netflix
• 3:45
  data
• 3:46
  from
• 3:46
  Ezra
• 3:46
  Savard's
• 3:46
  internal
• 3:47
  training
• 3:47
  study,
• 3:48
  which
• 3:48
  identified
• 3:48
  three
• 3:48
  AI
• 3:49
  literacy
• 3:49
  cohorts
• 3:49
  by
• 3:50
  daily
• 3:50
  token
• 3:50
  spend:
• 3:51
  0M
• 3:51
  (non-users),
• 3:52
  4M
• 3:53
  (single-agent
• 3:53
  synchronous),
• 3:54
  and
• 3:54
  12-15M
• 3:55
  (multi-agent
• 3:56
  asynchronous).
• 3:58
  People
• 3:58
  jump
• 3:58
  cohorts
• 3:59
  in
• 3:59
  five
• 3:59
  hours
• 3:59
  of
• 3:59
  focused
• 4:00
  team-based
• 4:00
  training,
• 4:01
  with
• 4:01
  96%
• 4:02
  staying
• 4:02
  there
• 4:02
  six
• 4:02
  weeks
• 4:03
  later.
• 4:03
  The
• 4:04
  advanced
• 4:04
  meta-skill
• 4:04
  isn't
• 4:05
  spending
• 4:05
  tokens
• 4:05
  but
• 4:05
  saving
• 4:06
  them
• 4:06
  —
• 4:06
  Yegge
• 4:07
  calls
• 4:07
  token
• 4:07
  efficiency
• 4:08
  "the
• 4:08
  new
• 4:08
  craft."
• 4:09
  SaaS,
• 4:10
  pronounced
• 4:10
  dead
• 4:10
  by
• 4:10
  vibe-coders
• 4:11
  six
• 4:11
  months
• 4:11
  ago,
• 4:12
  looks
• 4:12
  surprisingly
• 4:13
  durable:
• 4:13
  companies
• 4:14
  learned
• 4:14
  about
• 4:14
  token
• 4:15
  efficiency
• 4:15
  the
• 4:15
  hard
• 4:15
  way,
• 4:16
  blowing
• 4:16
  yearly
• 4:17
  budgets
• 4:17
  in
• 4:17
  months
• 4:17
  .
• 4:18
  The
• 4:18
  economic
• 4:19
  anxiety
• 4:19
  the
• 4:19
  plateau
• 4:20
  implies
• 4:20
  gets
• 4:20
  a
• 4:20
  philosophical
• 4:21
  treatment
• 4:21
  from
• 4:21
  George
• 4:22
  Malandrakis,
• 4:23
  who
• 4:23
  coins
• 4:23
  a
• 4:23
  new
• 4:23
  logical
• 4:24
  fallacy:
• 4:25
  ad
• 4:25
  economicum
• 4:25
  —
• 4:26
  the
• 4:27
  assumption
• 4:27
  that
• 4:27
  The
• 4:27
  Economy
• 4:28
  requires
• 4:28
  human
• 4:28
  consumption
• 4:29
  to
• 4:29
  function.
• 4:30
  His
• 4:30
  uncomfortable
• 4:31
  observation:
• 4:32
  corporations
• 4:33
  already
• 4:33
  execute
• 4:33
  billions
• 4:34
  of
• 4:34
  virtual
• 4:34
  transactions
• 4:35
  daily
• 4:35
  with
• 4:35
  companies
• 4:36
  that
• 4:36
  have
• 4:36
  no
• 4:36
  product,
• 4:37
  no
• 4:37
  service,
• 4:38
  and
• 4:38
  not
• 4:38
  a
• 4:38
  single
• 4:38
  employee.
• 4:40
  The
• 4:40
  implicit
• 4:40
  assumptions
• 4:40
  —
• 4:41
  that
• 4:42
  the
• 4:42
  economy
• 4:42
  needs
• 4:42
  human
• 4:43
  consumers,
• 4:43
  and
• 4:44
  that
• 4:44
  its
• 4:44
  collapse
• 4:44
  would
• 4:44
  hurt
• 4:44
  the
• 4:45
  resource-hoarding
• 4:45
  class
• 4:45
  —
• 4:46
  don't
• 4:47
  survive
• 4:47
  scrutiny.
• 4:48
  Once
• 4:49
  the
• 4:49
  owning
• 4:49
  class
• 4:49
  has
• 4:50
  intelligent
• 4:50
  machines
• 4:51
  and
• 4:51
  owns
• 4:51
  most
• 4:51
  assets,
• 4:52
  the
• 4:52
  economy
• 4:53
  crashing
• 4:53
  has
• 4:53
  no
• 4:54
  real
• 4:54
  consequences
• 4:54
  for
• 4:55
  them.
• 4:56
  It
• 4:56
  barely
• 4:56
  does
• 4:57
  already
• 4:57
  .
• 5:02
  GLM-5.2
• 5:02
  just
• 5:03
  became
• 5:03
  the
• 5:03
  first
• 5:03
  open-weights
• 5:04
  model
• 5:04
  to
• 5:04
  cross
• 5:04
  80%
• 5:05
  on
• 5:05
  Terminal-Bench
• 5:06
  2.1,
• 5:07
  beating
• 5:07
  every
• 5:07
  other
• 5:08
  open
• 5:08
  model
• 5:08
  and
• 5:08
  edging
• 5:09
  out
• 5:09
  Gemini.
• 5:10
  The
• 5:10
  community
• 5:10
  is
• 5:10
  enthusiastic
• 5:11
  but
• 5:11
  measured
• 5:12
  —
• 5:12
  Terminal-Bench
• 5:13
  2.1
• 5:13
  is
• 5:13
  the
• 5:13
  relaxed
• 5:14
  version
• 5:14
  with
• 5:14
  friendlier
• 5:15
  timeouts,
• 5:15
  and
• 5:15
  no
• 5:15
  model
• 5:16
  scores
• 5:16
  lower
• 5:16
  on
• 5:16
  2.1
• 5:17
  than
• 5:17
  on
• 5:17
  2.0.
• 5:18
  The
• 5:18
  real
• 5:19
  test
• 5:19
  will
• 5:19
  be
• 5:19
  Terminal-Bench
• 5:20
  3,
• 5:20
  before
• 5:20
  the
• 5:20
  labs
• 5:21
  start
• 5:21
  benchmaxxing
• 5:21
  it.
• 5:22
  Still,
• 5:22
  at
• 5:22
  a
• 5:22
  claimed
• 5:23
  frontier-level
• 5:23
  capability
• 5:24
  for
• 5:24
  a
• 5:24
  fraction
• 5:24
  of
• 5:24
  closed-model
• 5:25
  cost,
• 5:26
  GLM-5.2
• 5:27
  represents
• 5:27
  the
• 5:27
  open-weight
• 5:28
  ecosystem's
• 5:28
  best
• 5:28
  argument
• 5:29
  that
• 5:29
  the
• 5:29
  future
• 5:29
  doesn't
• 5:29
  have
• 5:30
  to
• 5:30
  be
• 5:30
  rented
• 5:30
  .
• 5:31
  Unsloth
• 5:31
  published
• 5:31
  GGUF
• 5:32
  quantizations
• 5:33
  within
• 5:33
  days,
• 5:33
  including
• 5:34
  a
• 5:34
  2-bit
• 5:34
  version
• 5:34
  at
• 5:34
  238GB
• 5:35
  —
• 5:36
  theoretically
• 5:37
  runnable
• 5:37
  on
• 5:37
  high-end
• 5:37
  consumer
• 5:38
  hardware
• 5:38
  .
• 5:38
  The
• 5:39
  local
• 5:39
  inference
• 5:39
  story
• 5:39
  has
• 5:40
  moved
• 5:40
  past
• 5:40
  theoretical
• 5:41
  into
• 5:41
  daily-driver
• 5:41
  territory.
• 5:42
  Vicki
• 5:42
  Boykis
• 5:43
  reports
• 5:43
  running
• 5:43
  agentic
• 5:44
  coding
• 5:44
  flows
• 5:45
  locally
• 5:45
  on
• 5:45
  a
• 5:46
  2022
• 5:46
  M2
• 5:47
  Mac
• 5:47
  with
• 5:47
  Gemma
• 5:48
  4
• 5:48
  models
• 5:49
  at
• 5:49
  roughly
• 5:49
  75%
• 5:50
  the
• 5:50
  accuracy
• 5:51
  of
• 5:51
  frontier
• 5:51
  models
• 5:51
  —
• 5:52
  a
• 5:52
  threshold
• 5:52
  where
• 5:52
  she
• 5:53
  stopped
• 5:53
  double-checking
• 5:54
  against
• 5:54
  API
• 5:54
  models.
• 5:55
  Her
• 5:55
  setup:
• 5:56
  LM
• 5:56
  Studio
• 5:57
  for
• 5:57
  inference,
• 5:57
  Pi
• 5:58
  as
• 5:58
  the
• 5:58
  agent
• 5:58
  harness,
• 5:59
  everything
• 5:59
  sandboxed
• 6:00
  in
• 6:00
  Docker.
• 6:01
  The
• 6:01
  benefits
• 6:02
  go
• 6:02
  beyond
• 6:02
  privacy:
• 6:03
  local
• 6:03
  inference
• 6:04
  lets
• 6:04
  you
• 6:04
  introspect
• 6:05
  everything,
• 6:06
  from
• 6:06
  token
• 6:06
  generation
• 6:07
  to
• 6:07
  K-V
• 6:07
  cache
• 6:07
  growth,
• 6:08
  in
• 6:08
  ways
• 6:08
  cloud
• 6:09
  APIs
• 6:09
  never
• 6:10
  will
• 6:10
  .
• 6:11
  On-device
• 6:11
  gets
• 6:12
  more
• 6:12
  extreme:
• 6:13
  a
• 6:13
  user
• 6:14
  runs
• 6:14
  Gemma
• 6:14
  4
• 6:15
  12B
• 6:15
  on
• 6:15
  a
• 6:15
  Google
• 6:15
  Pixel
• 6:16
  10
• 6:16
  Pro
• 6:16
  via
• 6:17
  Termux
• 6:17
  and
• 6:17
  llama.cpp
• 6:19
  at
• 6:19
  6.5
• 6:20
  tokens/second
• 6:21
  prompt
• 6:21
  processing
• 6:22
  and
• 6:22
  1.3
• 6:23
  t/s
• 6:23
  generation,
• 6:24
  drawing
• 6:24
  under
• 6:25
  10
• 6:25
  watts.
• 6:26
  Another
• 6:26
  commenter
• 6:27
  runs
• 6:27
  Gemma
• 6:27
  4
• 6:28
  26B
• 6:29
  IQ2
• 6:29
  M
• 6:29
  on
• 6:30
  Android
• 6:30
  16
• 6:31
  at
• 6:31
  4-5
• 6:31
  t/s
• 6:32
  drawing
• 6:32
  3-4
• 6:33
  watts.
• 6:34
  Useful
• 6:35
  inference
• 6:35
  is
• 6:35
  leaving
• 6:35
  the
• 6:36
  data
• 6:36
  center
• 6:36
  entirely
• 6:36
  .
• 6:37
  Not
• 6:38
  all
• 6:38
  open-model
• 6:39
  news
• 6:39
  inspires
• 6:39
  confidence.
• 6:41
  Rio
• 6:41
  3.5,
• 6:42
  a
• 6:42
  397B
• 6:44
  model
• 6:44
  funded
• 6:44
  by
• 6:45
  R$500K
• 6:46
  (~$100K
• 6:48
  USD)
• 6:50
  from
• 6:50
  the
• 6:50
  Rio
• 6:50
  de
• 6:50
  Janeiro
• 6:51
  municipal
• 6:51
  government,
• 6:52
  was
• 6:52
  discovered
• 6:52
  to
• 6:52
  be
• 6:53
  a
• 6:53
  cheap
• 6:53
  merge
• 6:53
  of
• 6:54
  Nex
• 6:54
  N2
• 6:54
  Pro
• 6:55
  with
• 6:55
  no
• 6:55
  additional
• 6:55
  training
• 6:56
  —
• 6:56
  despite
• 6:57
  documentation
• 6:58
  claiming
• 6:58
  sophisticated
• 6:59
  fine-tuning.
• 7:00
  When
• 7:00
  caught,
• 7:01
  the
• 7:01
  team
• 7:01
  claimed
• 7:01
  the
• 7:01
  "final
• 7:02
  trained
• 7:02
  model
• 7:02
  got
• 7:02
  lost"
• 7:03
  and
• 7:03
  promised
• 7:03
  to
• 7:03
  redo
• 7:04
  it
• 7:04
  from
• 7:04
  scratch.
• 7:05
  As
• 7:05
  one
• 7:06
  commenter
• 7:06
  put
• 7:06
  it:
• 7:07
  "my
• 7:07
  dog
• 7:07
  ate
• 7:07
  the
• 7:08
  weights."
• 7:09
  The
• 7:09
  episode
• 7:09
  represents
• 7:10
  a
• 7:10
  new
• 7:10
  category
• 7:11
  of
• 7:11
  threat
• 7:11
  to
• 7:11
  the
• 7:11
  open-model
• 7:12
  ecosystem:
• 7:13
  not
• 7:13
  benchmark
• 7:14
  gaming,
• 7:14
  but
• 7:14
  outright
• 7:15
  fabrication
• 7:15
  of
• 7:16
  provenance
• 7:16
  .
• 7:18
  A
• 7:18
  UC
• 7:19
  San
• 7:19
  Diego-led
• 7:19
  team
• 7:20
  introduces
• 7:20
  OpenDeepThink,
• 7:21
  a
• 7:21
  population-based
• 7:22
  framework
• 7:22
  for
• 7:22
  scaling
• 7:22
  test-time
• 7:23
  compute
• 7:23
  that
• 7:23
  sidesteps
• 7:24
  the
• 7:24
  selection
• 7:24
  bottleneck
• 7:25
  crippling
• 7:25
  best-of-N
• 7:25
  sampling.
• 7:26
  Instead
• 7:26
  of
• 7:27
  extending
• 7:27
  a
• 7:27
  single
• 7:27
  chain
• 7:28
  of
• 7:28
  thought
• 7:28
  (which
• 7:28
  fails
• 7:28
  catastrophically
• 7:29
  on
• 7:29
  early
• 7:29
  missteps),
• 7:30
  maintain
• 7:30
  a
• 7:30
  population
• 7:31
  of
• 7:31
  candidate
• 7:31
  solutions
• 7:32
  and
• 7:32
  evolve
• 7:32
  them
• 7:33
  over
• 7:33
  generations
• 7:33
  using
• 7:33
  pairwise
• 7:34
  Bradley-Terry
• 7:34
  comparison.
• 7:36
  The
• 7:36
  same
• 7:36
  LLM
• 7:36
  serves
• 7:37
  as
• 7:37
  both
• 7:37
  generator
• 7:37
  and
• 7:38
  judge.
• 7:38
  Each
• 7:39
  generation,
• 7:39
  random
• 7:40
  pairs
• 7:40
  are
• 7:40
  compared,
• 7:41
  votes
• 7:41
  are
• 7:41
  aggregated
• 7:42
  via
• 7:42
  Bradley-Terry
• 7:42
  maximum-likelihood
• 7:43
  estimation
• 7:44
  into
• 7:44
  a
• 7:44
  global
• 7:44
  ranking,
• 7:45
  top-ranked
• 7:46
  candidates
• 7:46
  are
• 7:46
  preserved
• 7:46
  as
• 7:46
  elites,
• 7:47
  and
• 7:47
  the
• 7:47
  bottom
• 7:48
  quarter
• 7:48
  is
• 7:48
  discarded.
• 7:49
  The
• 7:49
  natural-language
• 7:50
  critiques
• 7:50
  produced
• 7:51
  during
• 7:51
  comparison
• 7:51
  are
• 7:52
  recycled
• 7:52
  as
• 7:52
  mutation
• 7:53
  feedback
• 7:53
  —
• 7:54
  and
• 7:54
  critically,
• 7:54
  only
• 7:55
  negative
• 7:55
  feedback
• 7:55
  carries
• 7:55
  signal.
• 7:56
  Telling
• 7:57
  the
• 7:57
  mutator
• 7:57
  what
• 7:57
  went
• 7:57
  wrong
• 7:58
  is
• 7:58
  actionable;
• 7:58
  telling
• 7:59
  it
• 7:59
  what
• 7:59
  went
• 7:59
  right
• 7:59
  adds
• 8:00
  nothing
• 8:00
  beyond
• 8:00
  what
• 8:00
  the
• 8:00
  model
• 8:01
  already
• 8:01
  infers
• 8:01
  from
• 8:01
  seeing
• 8:02
  its
• 8:02
  own
• 8:02
  solution.
• 8:03
  The
• 8:03
  results:
• 8:03
  on
• 8:04
  Codeforces
• 8:04
  problems,
• 8:05
  OpenDeepThink
• 8:06
  raises
• 8:06
  Gemini
• 8:06
  3.1
• 8:07
  Pro's
• 8:07
  effective
• 8:08
  Elo
• 8:08
  by
• 8:08
  182
• 8:09
  points
• 8:09
  using
• 8:09
  approximately
• 8:10
  285
• 8:11
  API
• 8:11
  calls
• 8:12
  per
• 8:12
  problem
• 8:13
  with
• 8:13
  a
• 8:13
  sequential
• 8:13
  depth
• 8:13
  of
• 8:14
  just
• 8:14
  eight
• 8:14
  LLM
• 8:14
  calls
• 8:15
  —
• 8:15
  everything
• 8:16
  within
• 8:16
  each
• 8:16
  round
• 8:16
  parallelizes
• 8:17
  embarrassingly.
• 8:18
  The
• 8:18
  same
• 8:18
  hyperparameters
• 8:19
  transfer
• 8:19
  to
• 8:19
  Gemini
• 8:20
  3
• 8:20
  Flash
• 8:20
  and
• 8:20
  2.5
• 8:21
  Pro
• 8:21
  without
• 8:22
  retuning.
• 8:23
  On
• 8:23
  the
• 8:23
  multi-domain
• 8:23
  HLE
• 8:24
  benchmark,
• 8:25
  gains
• 8:25
  appear
• 8:25
  only
• 8:25
  in
• 8:26
  objectively
• 8:26
  verifiable
• 8:27
  domains
• 8:27
  (math,
• 8:27
  biology,
• 8:28
  physics)
• 8:29
  and
• 8:29
  reverse
• 8:29
  in
• 8:29
  subjective
• 8:30
  ones
• 8:30
  —
• 8:31
  the
• 8:31
  soft
• 8:31
  verifier
• 8:32
  is
• 8:32
  only
• 8:32
  as
• 8:32
  good
• 8:33
  as
• 8:33
  the
• 8:33
  comparisons
• 8:33
  it
• 8:33
  aggregates.
• 8:34
  Pairwise
• 8:35
  judgment
• 8:35
  reaches
• 8:35
  78%
• 8:36
  accuracy
• 8:37
  versus
• 8:37
  60%
• 8:38
  for
• 8:38
  pointwise
• 8:38
  scoring,
• 8:39
  confirming
• 8:39
  that
• 8:39
  relative
• 8:40
  framing
• 8:40
  sidesteps
• 8:41
  the
• 8:41
  positive
• 8:41
  bias
• 8:41
  plaguing
• 8:42
  absolute
• 8:42
  scoring
• 8:42
  .
• 8:43
  On
• 8:43
  the
• 8:43
  fine-tuning
• 8:44
  side,
• 8:44
  HuggingFace's
• 8:45
  PEFT
• 8:45
  team
• 8:46
  published
• 8:46
  benchmarks
• 8:46
  that
• 8:47
  should
• 8:47
  give
• 8:47
  LoRA
• 8:47
  loyalists
• 8:48
  pause.
• 8:49
  LoRA
• 8:49
  commands
• 8:49
  98.4%
• 8:51
  of
• 8:51
  PEFT
• 8:51
  usage
• 8:51
  on
• 8:51
  HuggingFace
• 8:52
  Hub
• 8:52
  —
• 8:52
  but
• 8:52
  controlled
• 8:53
  experiments
• 8:53
  show
• 8:54
  it
• 8:54
  isn't
• 8:54
  always
• 8:54
  optimal.
• 8:55
  On
• 8:56
  an
• 8:56
  image
• 8:56
  generation
• 8:56
  task,
• 8:57
  OFT
• 8:57
  (Orthogonal
• 8:58
  Fine-Tuning)
• 8:58
  strictly
• 8:59
  dominates
• 8:59
  LoRA:
• 9:00
  higher
• 9:00
  similarity
• 9:01
  score
• 9:01
  (0.708
• 9:02
  vs
• 9:03
  0.697)
• 9:04
  at
• 9:04
  lower
• 9:04
  memory
• 9:05
  (9.01
• 9:06
  GB
• 9:06
  vs
• 9:07
  9.97
• 9:07
  GB).
• 9:08
  On
• 9:09
  LLM
• 9:09
  math
• 9:09
  reasoning,
• 9:10
  vanilla
• 9:10
  LoRA
• 9:11
  achieves
• 9:11
  only
• 9:11
  48.1%
• 9:12
  accuracy;
• 9:14
  you
• 9:14
  need
• 9:14
  rsLoRA
• 9:14
  (53.2%)
• 9:16
  or
• 9:16
  LoRA-FA
• 9:16
  to
• 9:16
  reach
• 9:17
  the
• 9:17
  Pareto
• 9:17
  frontier.
• 9:18
  The
• 9:18
  practical
• 9:19
  barrier
• 9:19
  —
• 9:20
  downstream
• 9:20
  tools
• 9:21
  like
• 9:21
  vLLM
• 9:21
  only
• 9:21
  support
• 9:22
  LoRA
• 9:22
  —
• 9:23
  is
• 9:23
  being
• 9:23
  addressed:
• 9:24
  PEFT
• 9:24
  now
• 9:24
  converts
• 9:25
  non-LoRA
• 9:25
  adapters
• 9:26
  to
• 9:26
  LoRA
• 9:26
  format
• 9:26
  with
• 9:27
  negligible
• 9:27
  quality
• 9:27
  loss
• 9:28
  .
• 9:32
  A
• 9:32
  security
• 9:33
  researcher
• 9:33
  registered
• 9:34
  on
• 9:34
  FIFA's
• 9:34
  public
• 9:35
  Agent
• 9:35
  Platform,
• 9:36
  submitted
• 9:36
  an
• 9:36
  ID
• 9:37
  photo,
• 9:37
  and
• 9:37
  received
• 9:37
  credentials
• 9:38
  in
• 9:38
  FIFA's
• 9:38
  Microsoft
• 9:39
  Entra
• 9:39
  tenant
• 9:39
  —
• 9:40
  the
• 9:40
  same
• 9:40
  tenant
• 9:41
  powering
• 9:41
  all
• 9:41
  of
• 9:41
  FIFA's
• 9:42
  internal
• 9:42
  platforms.
• 9:43
  The
• 9:43
  Angular
• 9:44
  frontend
• 9:44
  checked
• 9:44
  the
• 9:44
  JWT
• 9:45
  for
• 9:45
  role
• 9:45
  claims,
• 9:46
  found
• 9:46
  none,
• 9:47
  and
• 9:47
  rendered
• 9:47
  an
• 9:47
  access-denied
• 9:48
  page.
• 9:49
  The
• 9:49
  backend
• 9:50
  APIs
• 9:50
  didn't
• 9:50
  check
• 9:51
  anything.
• 9:52
  They
• 9:52
  just
• 9:52
  served
• 9:52
  whatever
• 9:52
  was
• 9:53
  requested.
• 9:54
  Behind
• 9:54
  that
• 9:54
  client-side
• 9:55
  guard
• 9:55
  sat
• 9:55
  the
• 9:55
  live
• 9:56
  Streaming
• 9:56
  Management
• 9:56
  panel
• 9:57
  for
• 9:57
  every
• 9:57
  FIFA
• 9:57
  World
• 9:58
  Cup
• 9:58
  2026
• 9:58
  match.
• 9:59
  Every
• 9:59
  camera
• 10:00
  angle.
• 10:00
  Every
• 10:01
  RTMP
• 10:01
  ingest
• 10:02
  URL.
• 10:03
  Every
• 10:03
  stream
• 10:03
  key.
• 10:04
  Five
• 10:04
  camera
• 10:05
  feeds
• 10:05
  per
• 10:05
  match
• 10:05
  —
• 10:06
  PGM
• 10:07
  (main
• 10:07
  broadcast),
• 10:08
  Tactical,
• 10:08
  Camera1,
• 10:09
  High
• 10:09
  Behind
• 10:10
  Left,
• 10:10
  High
• 10:10
  Behind
• 10:10
  Right
• 10:11
  —
• 10:11
  each
• 10:12
  with
• 10:12
  the
• 10:12
  stream
• 10:12
  key
• 10:12
  embedded
• 10:13
  in
• 10:13
  the
• 10:13
  URL,
• 10:14
  shared
• 10:14
  across
• 10:14
  all
• 10:14
  five
• 10:14
  angles.
• 10:15
  An
• 10:16
  attacker
• 10:16
  pushing
• 10:16
  video
• 10:17
  to
• 10:17
  those
• 10:17
  RTMP
• 10:17
  endpoints
• 10:18
  would
• 10:18
  replace
• 10:19
  the
• 10:19
  feed
• 10:19
  going
• 10:19
  to
• 10:19
  every
• 10:20
  TV
• 10:20
  network
• 10:20
  worldwide.
• 10:22
  The
• 10:22
  researcher
• 10:22
  opened
• 10:22
  VLC,
• 10:23
  pasted
• 10:24
  a
• 10:24
  preview
• 10:24
  manifest,
• 10:25
  and
• 10:25
  confirmed:
• 10:26
  live
• 10:26
  tactical
• 10:26
  camera
• 10:27
  footage
• 10:27
  from
• 10:27
  an
• 10:27
  active
• 10:27
  World
• 10:28
  Cup
• 10:28
  match,
• 10:29
  streaming
• 10:29
  to
• 10:29
  a
• 10:29
  laptop
• 10:30
  in
• 10:30
  Tokyo.
• 10:31
  Write
• 10:31
  access
• 10:32
  was
• 10:32
  worse
• 10:32
  —
• 10:33
  the
• 10:33
  Match
• 10:33
  Management
• 10:34
  panel
• 10:34
  accepted
• 10:34
  score
• 10:35
  changes,
• 10:36
  kick-off
• 10:36
  time
• 10:36
  adjustments,
• 10:37
  and
• 10:37
  editorial
• 10:37
  commentary
• 10:38
  updates
• 10:38
  from
• 10:38
  the
• 10:39
  NO
• 10:39
  ROLES
• 10:39
  account,
• 10:40
  data
• 10:40
  feeding
• 10:41
  directly
• 10:41
  into
• 10:41
  the
• 10:41
  Commentator
• 10:42
  Information
• 10:42
  System
• 10:43
  used
• 10:43
  on
• 10:43
  live
• 10:43
  television.
• 10:44
  The
• 10:44
  disclosure
• 10:45
  process
• 10:45
  was
• 10:46
  a
• 10:46
  nightmare:
• 10:47
  no
• 10:47
  bug
• 10:47
  bounty,
• 10:48
  no
• 10:48
  security.txt,
• 10:50
  five
• 10:50
  bounced
• 10:51
  emails,
• 10:52
  FIFA
• 10:52
  HQ
• 10:53
  closed
• 10:53
  on
• 10:53
  Sunday,
• 10:54
  the
• 10:54
  breakthrough
• 10:55
  coming
• 10:55
  from
• 10:55
  calls
• 10:55
  to
• 10:56
  MediaKind's
• 10:56
  toll-free
• 10:57
  line,
• 10:57
  CISA's
• 10:58
  24/7
• 10:59
  operations
• 10:59
  center,
• 11:00
  and
• 11:00
  FBI
• 11:01
  contacts
• 11:01
  on
• 11:01
  Signal.
• 11:02
  FIFA
• 11:02
  patched
• 11:03
  it
• 11:03
  overnight
• 11:03
  and
• 11:03
  never
• 11:04
  responded
• 11:04
  .
• 11:05
  At
• 11:05
  the
• 11:05
  hardware
• 11:05
  layer,
• 11:06
  MIT's
• 11:07
  CSAIL
• 11:07
  team
• 11:07
  addressed
• 11:08
  a
• 11:08
  different
• 11:08
  gap
• 11:08
  between
• 11:09
  what
• 11:09
  systems
• 11:09
  claim
• 11:10
  and
• 11:10
  what
• 11:10
  they
• 11:10
  actually
• 11:10
  do.
• 11:11
  Fractal,
• 11:12
  a
• 11:12
  kernel
• 11:12
  built
• 11:13
  from
• 11:13
  scratch
• 11:13
  for
• 11:13
  microarchitecture
• 11:14
  research,
• 11:15
  boots
• 11:15
  on
• 11:15
  bare
• 11:16
  metal
• 11:16
  with
• 11:16
  zero
• 11:16
  background
• 11:17
  noise
• 11:17
  and
• 11:17
  lets
• 11:17
  experiments
• 11:18
  switch
• 11:18
  privilege
• 11:19
  levels
• 11:19
  at
• 11:19
  runtime
• 11:20
  while
• 11:20
  executing
• 11:20
  identical
• 11:21
  instructions
• 11:21
  in
• 11:22
  the
• 11:22
  same
• 11:22
  address
• 11:22
  space.
• 11:23
  Its
• 11:23
  first
• 11:24
  use:
• 11:25
  Apple
• 11:25
  M1
• 11:25
  branch
• 11:26
  predictors.
• 11:27
  The
• 11:27
  M1
• 11:27
  implements
• 11:28
  ARM's
• 11:28
  CSV2
• 11:29
  protection,
• 11:30
  which
• 11:30
  should
• 11:30
  block
• 11:31
  cross-privilege
• 11:31
  speculative
• 11:32
  execution
• 11:32
  —
• 11:34
  and
• 11:34
  Fractal
• 11:34
  confirmed
• 11:35
  it
• 11:35
  works
• 11:35
  for
• 11:35
  the
• 11:35
  execute
• 11:36
  stage.
• 11:37
  But
• 11:37
  the
• 11:37
  CPU
• 11:38
  still
• 11:38
  fetches
• 11:38
  the
• 11:38
  target
• 11:39
  into
• 11:39
  the
• 11:39
  instruction
• 11:40
  cache
• 11:40
  before
• 11:40
  the
• 11:40
  protection
• 11:41
  engages,
• 11:42
  creating
• 11:42
  an
• 11:42
  observable
• 11:43
  side
• 11:43
  channel.
• 11:44
  The
• 11:45
  team
• 11:45
  produced
• 11:45
  the
• 11:45
  first
• 11:46
  evidence
• 11:46
  that
• 11:46
  Apple
• 11:46
  Silicon
• 11:47
  exhibits
• 11:47
  Phantom
• 11:48
  speculation,
• 11:49
  previously
• 11:49
  demonstrated
• 11:50
  only
• 11:50
  on
• 11:51
  AMD
• 11:51
  and
• 11:51
  Intel.
• 11:53
  An
• 11:53
  earlier
• 11:53
  finding
• 11:54
  that
• 11:54
  cross-privilege
• 11:55
  training
• 11:55
  worked
• 11:55
  on
• 11:56
  M1
• 11:56
  performance
• 11:56
  cores
• 11:57
  but
• 11:57
  not
• 11:57
  efficiency
• 11:58
  cores
• 11:58
  was
• 11:58
  overturned
• 11:58
  —
• 12:00
  Fractal
• 12:00
  showed
• 12:00
  no
• 12:00
  privilege
• 12:01
  isolation
• 12:01
  on
• 12:02
  either
• 12:02
  core
• 12:02
  type,
• 12:03
  with
• 12:03
  the
• 12:03
  prior
• 12:03
  result
• 12:04
  likely
• 12:04
  an
• 12:04
  artifact
• 12:05
  of
• 12:05
  macOS
• 12:05
  migrating
• 12:06
  threads
• 12:06
  between
• 12:07
  cores
• 12:07
  during
• 12:07
  system
• 12:08
  calls
• 12:08
  .
• 12:11
  The
• 12:11
  Agentic
• 12:12
  Context
• 12:12
  Engine
• 12:13
  (ACE)
• 12:13
  from
• 12:13
  Kayba
• 12:13
  tackles
• 12:14
  a
• 12:14
  problem
• 12:14
  that
• 12:14
  sharpens
• 12:15
  as
• 12:15
  agents
• 12:15
  grow
• 12:15
  more
• 12:16
  autonomous:
• 12:16
  they
• 12:17
  don't
• 12:17
  learn
• 12:17
  from
• 12:17
  experience.
• 12:18
  Every
• 12:19
  session
• 12:19
  starts
• 12:19
  fresh,
• 12:20
  repeating
• 12:20
  mistakes
• 12:20
  and
• 12:21
  ignoring
• 12:21
  what
• 12:21
  worked.
• 12:22
  ACE
• 12:22
  maintains
• 12:23
  a
• 12:23
  "Skillbook"
• 12:23
  —
• 12:24
  a
• 12:24
  persistent
• 12:24
  strategy
• 12:25
  collection
• 12:25
  managed
• 12:25
  by
• 12:26
  three
• 12:26
  specialized
• 12:26
  roles:
• 12:27
  an
• 12:27
  Agent
• 12:28
  enhanced
• 12:28
  with
• 12:28
  learned
• 12:28
  strategies,
• 12:29
  a
• 12:29
  Reflector
• 12:30
  that
• 12:30
  analyzes
• 12:31
  execution
• 12:31
  traces,
• 12:32
  and
• 12:32
  a
• 12:32
  SkillManager
• 12:32
  that
• 12:32
  curates
• 12:33
  the
• 12:33
  library.
• 12:34
  The
• 12:34
  Recursive
• 12:34
  Reflector
• 12:35
  is
• 12:35
  the
• 12:35
  key
• 12:35
  innovation:
• 12:36
  instead
• 12:37
  of
• 12:37
  single-pass
• 12:37
  trace
• 12:37
  summarization,
• 12:38
  it
• 12:38
  writes
• 12:39
  and
• 12:39
  executes
• 12:39
  Python
• 12:40
  code
• 12:40
  in
• 12:40
  a
• 12:40
  sandbox
• 12:40
  to
• 12:41
  programmatically
• 12:41
  search
• 12:41
  for
• 12:42
  patterns
• 12:42
  and
• 12:42
  iterate
• 12:43
  until
• 12:43
  it
• 12:43
  finds
• 12:43
  actionable
• 12:43
  insights.
• 12:45
  On
• 12:45
  the
• 12:45
  Tau2
• 12:45
  airline
• 12:46
  benchmark,
• 12:46
  ACE
• 12:47
  doubles
• 12:47
  pass^4
• 12:48
  consistency
• 12:48
  with
• 12:49
  15
• 12:49
  learned
• 12:49
  strategies
• 12:50
  and
• 12:50
  no
• 12:50
  reward
• 12:51
  signals.
• 12:52
  A
• 12:52
  Claude
• 12:53
  Code
• 12:53
  integration
• 12:53
  translated
• 12:54
  14,000
• 12:55
  lines
• 12:55
  from
• 12:55
  Python
• 12:55
  to
• 12:55
  TypeScript
• 12:56
  in
• 12:56
  four
• 12:56
  hours
• 12:57
  with
• 12:57
  zero
• 12:57
  build
• 12:57
  errors
• 12:58
  at
• 12:58
  a
• 12:58
  learning
• 12:58
  cost
• 12:58
  of
• 12:58
  ~$1.50
• 12:59
  .
• 13:01
  Reuven
• 13:01
  Cohen
• 13:01
  frames
• 13:01
  this
• 13:02
  more
• 13:02
  broadly:
• 13:03
  the
• 13:03
  most
• 13:03
  important
• 13:03
  idea
• 13:04
  in
• 13:04
  AI
• 13:04
  today
• 13:04
  is
• 13:04
  not
• 13:05
  the
• 13:05
  model
• 13:05
  but
• 13:05
  the
• 13:05
  loop.
• 13:06
  Generate,
• 13:06
  measure,
• 13:07
  learn,
• 13:07
  adapt,
• 13:08
  repeat.
• 13:09
  His
• 13:09
  "Darwin
• 13:09
  Mode"
• 13:09
  implements
• 13:10
  evolutionary
• 13:10
  selection
• 13:11
  over
• 13:11
  agent
• 13:11
  strategies
• 13:12
  —
• 13:13
  the
• 13:13
  best
• 13:13
  survive,
• 13:13
  weak
• 13:14
  approaches
• 13:14
  disappear,
• 13:15
  capability
• 13:16
  compounds
• 13:16
  through
• 13:16
  recursion
• 13:17
  .
• 13:17
  The
• 13:17
  practical
• 13:18
  reality
• 13:18
  of
• 13:18
  loop
• 13:19
  engineering
• 13:19
  is
• 13:19
  messier.
• 13:20
  Boris
• 13:20
  Cherny
• 13:21
  (Claude
• 13:21
  Code
• 13:21
  lead)
• 13:22
  claims
• 13:22
  he
• 13:22
  no
• 13:22
  longer
• 13:22
  prompts
• 13:23
  Claude
• 13:23
  —
• 13:24
  he
• 13:24
  writes
• 13:24
  loops
• 13:24
  and
• 13:24
  the
• 13:24
  loops
• 13:25
  do
• 13:25
  the
• 13:25
  work.
• 13:26
  But
• 13:26
  cost
• 13:27
  tracking
• 13:27
  tells
• 13:27
  a
• 13:27
  different
• 13:28
  story:
• 13:29
  a
• 13:29
  single
• 13:29
  run
• 13:29
  through
• 13:30
  a
• 13:30
  relatively
• 13:30
  simple
• 13:31
  application
• 13:31
  can
• 13:31
  burn
• 13:32
  over
• 13:32
  a
• 13:32
  million
• 13:32
  tokens.
• 13:33
  The
• 13:34
  fundamental
• 13:34
  overhead
• 13:35
  is
• 13:35
  context
• 13:35
  passing
• 13:35
  —
• 13:37
  every
• 13:37
  orchestrator-to-worker
• 13:38
  handoff
• 13:38
  requires
• 13:39
  reasoning
• 13:39
  about
• 13:39
  state,
• 13:40
  and
• 13:40
  that
• 13:40
  reasoning
• 13:41
  isn't
• 13:41
  free.
• 13:42
  The
• 13:42
  emerging
• 13:43
  solution:
• 13:44
  deterministic
• 13:44
  harnesses
• 13:45
  with
• 13:45
  model
• 13:46
  routing
• 13:46
  (Haiku
• 13:46
  for
• 13:47
  classification,
• 13:48
  Opus
• 13:48
  for
• 13:48
  implementation),
• 13:49
  durable
• 13:50
  state
• 13:50
  in
• 13:50
  Postgres
• 13:51
  for
• 13:51
  resumability,
• 13:52
  and
• 13:52
  human-in-the-loop
• 13:53
  checkpoints
• 13:53
  at
• 13:53
  critical
• 13:54
  junctures
• 13:54
  .
• 13:55
  The
• 13:55
  trust
• 13:55
  boundary
• 13:55
  gets
• 13:56
  tested
• 13:56
  when
• 13:57
  someone
• 13:57
  connects
• 13:57
  their
• 13:57
  bank
• 13:58
  to
• 13:58
  Claude
• 13:58
  through
• 13:59
  MCP
• 13:59
  and
• 13:59
  handles
• 14:00
  invoicing,
• 14:01
  bill
• 14:01
  pay,
• 14:02
  and
• 14:02
  bookkeeping
• 14:02
  through
• 14:02
  conversations.
• 14:04
  The
• 14:04
  community
• 14:04
  consensus
• 14:05
  was
• 14:05
  a
• 14:05
  hard
• 14:06
  no
• 14:06
  on
• 14:06
  unsupervised
• 14:07
  access
• 14:07
  —
• 14:08
  a
• 14:08
  CPA
• 14:09
  providing
• 14:09
  controller
• 14:10
  services
• 14:10
  called
• 14:10
  any
• 14:11
  real
• 14:11
  business
• 14:11
  doing
• 14:11
  this
• 14:12
  "nuts."
• 14:13
  The
• 14:13
  pragmatic
• 14:13
  middle
• 14:14
  ground:
• 14:14
  read-only
• 14:15
  analysis
• 14:15
  of
• 14:16
  uploaded
• 14:16
  statements,
• 14:17
  with
• 14:17
  a
• 14:17
  human
• 14:17
  approving
• 14:18
  every
• 14:18
  transaction
• 14:19
  .
• 14:23
  Apple
• 14:23
  used
• 14:24
  WWDC
• 14:24
  to
• 14:25
  frame
• 14:25
  the
• 14:25
  question
• 14:25
  that
• 14:25
  may
• 14:26
  decide
• 14:26
  who
• 14:26
  becomes
• 14:26
  the
• 14:26
  first
• 14:27
  trillionaire
• 14:27
  in
• 14:27
  AI:
• 14:28
  when
• 14:29
  AI
• 14:29
  starts
• 14:29
  doing
• 14:29
  real
• 14:30
  work
• 14:30
  all
• 14:30
  day
• 14:30
  long,
• 14:31
  where
• 14:31
  does
• 14:31
  that
• 14:31
  work
• 14:32
  run?
• 14:32
  Not
• 14:33
  which
• 14:33
  model
• 14:33
  is
• 14:33
  ahead
• 14:33
  —
• 14:34
  who
• 14:34
  owns
• 14:35
  the
• 14:35
  surface
• 14:35
  where
• 14:35
  AI
• 14:36
  sees
• 14:36
  your
• 14:36
  work,
• 14:37
  touches
• 14:37
  your
• 14:37
  apps,
• 14:38
  and
• 14:38
  acts
• 14:38
  on
• 14:38
  your
• 14:38
  behalf?
• 14:40
  Apple's
• 14:40
  answer:
• 14:41
  the
• 14:41
  device
• 14:41
  you
• 14:41
  already
• 14:41
  bought,
• 14:42
  plus
• 14:42
  a
• 14:42
  tiered
• 14:43
  cloud
• 14:43
  backend.
• 14:45
  New
• 14:45
  Apple
• 14:45
  Foundation
• 14:46
  Models
• 14:46
  were
• 14:46
  built
• 14:46
  in
• 14:46
  collaboration
• 14:47
  with
• 14:47
  Google
• 14:48
  using
• 14:48
  Gemini
• 14:49
  family
• 14:49
  technology.
• 14:50
  Private
• 14:51
  Cloud
• 14:51
  Compute
• 14:51
  now
• 14:52
  extends
• 14:52
  into
• 14:52
  Google
• 14:53
  Cloud
• 14:53
  with
• 14:53
  Nvidia
• 14:54
  GPUs
• 14:54
  for
• 14:54
  hard
• 14:55
  reasoning
• 14:55
  and
• 14:55
  agentic
• 14:55
  tool
• 14:56
  use.
• 14:57
  Siri
• 14:57
  sits
• 14:58
  atop
• 14:58
  personal
• 14:58
  context,
• 14:59
  screen
• 14:59
  awareness,
• 15:00
  App
• 15:00
  Intents,
• 15:00
  and
• 15:01
  the
• 15:01
  Spotlight
• 15:01
  semantic
• 15:01
  index.
• 15:02
  The
• 15:02
  strategic
• 15:03
  bet
• 15:03
  is
• 15:03
  that
• 15:03
  you
• 15:03
  can
• 15:04
  source
• 15:04
  raw
• 15:04
  model
• 15:04
  capability
• 15:05
  —
• 15:05
  you
• 15:05
  cannot
• 15:06
  easily
• 15:06
  source
• 15:06
  a
• 15:06
  billion
• 15:07
  devices,
• 15:07
  a
• 15:07
  mature
• 15:08
  OS,
• 15:08
  and
• 15:08
  consumer
• 15:09
  trust.
• 15:09
  The
• 15:10
  developer
• 15:10
  story
• 15:10
  is
• 15:10
  the
• 15:10
  quiet
• 15:11
  power
• 15:11
  move:
• 15:12
  App
• 15:12
  Intents
• 15:12
  makes
• 15:12
  apps
• 15:13
  callable
• 15:13
  by
• 15:13
  the
• 15:13
  operating
• 15:14
  system,
• 15:14
  and
• 15:14
  the
• 15:14
  winning
• 15:15
  apps
• 15:15
  won't
• 15:15
  have
• 15:15
  the
• 15:15
  flashiest
• 15:16
  chatbot
• 15:16
  —
• 15:17
  they'll
• 15:17
  be
• 15:17
  the
• 15:17
  ones
• 15:17
  whose
• 15:17
  data
• 15:18
  and
• 15:18
  actions
• 15:18
  are
• 15:18
  clean
• 15:19
  enough
• 15:19
  for
• 15:19
  Apple
• 15:19
  Intelligence
• 15:20
  to
• 15:20
  use
• 15:20
  .
• 15:21
  In
• 15:21
  developer
• 15:21
  tooling,
• 15:22
  Epic
• 15:22
  Games
• 15:22
  announced
• 15:23
  Lore,
• 15:23
  a
• 15:23
  fully
• 15:24
  open-source
• 15:24
  (MIT-licensed)
• 15:25
  version
• 15:26
  control
• 15:26
  system
• 15:27
  designed
• 15:27
  for
• 15:27
  the
• 15:27
  realities
• 15:28
  Git
• 15:28
  ignores:
• 15:29
  massive
• 15:29
  binary
• 15:30
  assets,
• 15:30
  projects
• 15:31
  mixing
• 15:31
  code
• 15:32
  with
• 15:32
  art,
• 15:32
  and
• 15:32
  teams
• 15:33
  spanning
• 15:33
  hundreds
• 15:34
  of
• 15:34
  contributors.
• 15:35
  Lore
• 15:35
  uses
• 15:36
  Merkle
• 15:36
  trees
• 15:36
  and
• 15:37
  content-addressed
• 15:37
  storage
• 15:38
  with
• 15:38
  chunked
• 15:38
  deduplication,
• 15:40
  on-demand
• 15:40
  hydration,
• 15:41
  and
• 15:41
  lightweight
• 15:41
  branches
• 15:42
  —
• 15:43
  workspaces
• 15:43
  stay
• 15:44
  lean
• 15:44
  by
• 15:44
  fetching
• 15:44
  data
• 15:45
  only
• 15:45
  when
• 15:45
  needed.
• 15:46
  Sparse
• 15:47
  workspaces
• 15:47
  and
• 15:48
  a
• 15:48
  service-backed
• 15:48
  caching
• 15:49
  architecture
• 15:49
  differentiate
• 15:50
  it
• 15:50
  from
• 15:50
  Git-LFS
• 15:51
  bolt-on
• 15:51
  solutions
• 15:52
  .
• 15:53
  Meanwhile,
• 15:53
  Supertone's
• 15:54
  supertonic-3
• 15:55
  voice
• 15:55
  model
• 15:55
  is
• 15:56
  trending
• 15:56
  on
• 15:56
  HuggingFace
• 15:56
  ,
• 15:57
  and
• 15:57
  a
• 15:57
  lengthy
• 15:58
  video
• 15:58
  pitch
• 15:58
  for
• 15:58
  the
• 15:58
  Kaix
• 15:59
  "association-native
• 16:00
  database"
• 16:01
  promises
• 16:02
  AGI,
• 16:02
  time
• 16:03
  travel
• 16:03
  capabilities,
• 16:04
  and
• 16:04
  the
• 16:04
  discovery
• 16:04
  of
• 16:05
  God's
• 16:05
  plan
• 16:05
  via
• 16:05
  a
• 16:05
  $250/month
• 16:07
  school
• 16:08
  community
• 16:08
  —
• 16:09
  claims
• 16:09
  that
• 16:09
  serve
• 16:10
  as
• 16:10
  a
• 16:10
  useful
• 16:10
  reminder
• 16:11
  that
• 16:11
  extraordinary
• 16:12
  assertions
• 16:12
  still
• 16:13
  require
• 16:13
  extraordinary
• 16:14
  evidence
• 16:14
  .