Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 26, 2026

Transcript

AGI Dreams Podcast – June 26, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:01
  AI
• 0:01
  industry's
• 0:01
  insatiable
• 0:02
  appetite
• 0:03
  for
• 0:03
  memory
• 0:03
  chips
• 0:03
  has
• 0:03
  finally
• 0:04
  landed
• 0:04
  on
• 0:04
  consumers'
• 0:05
  doorsteps.
• 0:06
  Apple
• 0:06
  raised
• 0:06
  prices
• 0:07
  across
• 0:07
  its
• 0:07
  MacBook
• 0:08
  and
• 0:08
  iPad
• 0:08
  lines
• 0:08
  this
• 0:08
  week,
• 0:09
  pushing
• 0:09
  the
• 0:09
  MacBook
• 0:10
  Neo
• 0:10
  from
• 0:10
  $599
• 0:12
  to
• 0:12
  $699
• 0:13
  and
• 0:13
  the
• 0:13
  MacBook
• 0:14
  Air
• 0:14
  512GB
• 0:15
  from
• 0:16
  $1,099
• 0:17
  to
• 0:17
  $1,299.
• 0:20
  Tim
• 0:20
  Cook
• 0:20
  had
• 0:20
  telegraphed
• 0:21
  this
• 0:21
  since
• 0:21
  April
• 0:21
  —
• 0:22
  "we
• 0:22
  expect
• 0:23
  significantly
• 0:23
  higher
• 0:24
  memory
• 0:24
  costs"
• 0:24
  —
• 0:25
  but
• 0:25
  the
• 0:25
  scale
• 0:25
  is
• 0:26
  still
• 0:26
  striking.
• 0:27
  DRAM
• 0:27
  prices
• 0:28
  surged
• 0:28
  98%
• 0:29
  in
• 0:29
  Q1
• 0:30
  2026,
• 0:30
  with
• 0:30
  another
• 0:31
  58-63%
• 0:33
  increase
• 0:33
  projected
• 0:34
  for
• 0:34
  Q2,
• 0:34
  according
• 0:35
  to
• 0:35
  TrendForce.
• 0:36
  The
• 0:36
  industry
• 0:37
  has
• 0:37
  started
• 0:37
  calling
• 0:37
  it
• 0:37
  "RAMageddon,"
• 0:38
  and
• 0:38
  the
• 0:38
  name
• 0:39
  fits.
• 0:40
  IDC
• 0:40
  now
• 0:40
  forecasts
• 0:41
  the
• 0:41
  largest-ever
• 0:42
  annual
• 0:42
  smartphone
• 0:42
  decline
• 0:43
  at
• 0:43
  nearly
• 0:43
  14%,
• 0:44
  with
• 0:44
  PCs
• 0:45
  down
• 0:45
  11.3%.
• 0:47
  The
• 0:47
  MacBook
• 0:47
  Neo,
• 0:48
  Apple's
• 0:48
  play
• 0:48
  for
• 0:48
  the
• 0:48
  affordable
• 0:49
  laptop
• 0:49
  segment,
• 0:50
  just
• 0:50
  lost
• 0:50
  its
• 0:50
  $100
• 0:51
  advantage
• 0:51
  over
• 0:52
  Dell's
• 0:52
  XPS
• 0:52
  13.
• 0:53
  We
• 0:53
  have
• 0:54
  tracked
• 0:54
  this
• 0:54
  squeeze
• 0:54
  since
• 0:55
  Edition
• 0:55
  121
• 0:56
  last
• 0:56
  September,
• 0:57
  when
• 0:57
  DDR5
• 0:57
  prices
• 0:58
  first
• 0:58
  began
• 0:58
  climbing.
• 0:59
  SK
• 0:59
  Hynix
• 1:00
  warned
• 1:00
  it
• 1:00
  would
• 1:00
  persist
• 1:00
  until
• 1:01
  2030.
• 1:02
  Micron
• 1:02
  disclosed
• 1:02
  $22
• 1:03
  billion
• 1:03
  in
• 1:04
  long-term
• 1:04
  commitments
• 1:04
  locked
• 1:05
  in
• 1:05
  by
• 1:05
  AI
• 1:05
  chipmakers,
• 1:06
  confirming
• 1:06
  that
• 1:07
  supply
• 1:07
  is
• 1:07
  being
• 1:07
  structurally
• 1:08
  redirected
• 1:08
  from
• 1:08
  consumer
• 1:09
  electronics
• 1:09
  to
• 1:09
  datacenter.
• 1:10
  Rival
• 1:11
  device
• 1:11
  makers
• 1:11
  without
• 1:12
  Apple's
• 1:12
  supplier
• 1:12
  leverage
• 1:13
  will
• 1:13
  get
• 1:13
  hit
• 1:13
  harder
• 1:13
  —
• 1:14
  some
• 1:15
  analysts
• 1:15
  expect
• 1:15
  steeper
• 1:16
  increases
• 1:16
  across
• 1:16
  the
• 1:17
  Windows
• 1:17
  and
• 1:17
  Android
• 1:17
  ecosystems.
• 1:19
  Meanwhile,
• 1:19
  the
• 1:19
  silicon
• 1:20
  itself
• 1:20
  keeps
• 1:20
  scaling.
• 1:21
  IBM
• 1:22
  unveiled
• 1:22
  the
• 1:22
  world's
• 1:22
  first
• 1:23
  sub-1nm
• 1:23
  chip
• 1:23
  technology
• 1:24
  at
• 1:24
  0.7
• 1:25
  nanometers
• 1:25
  —
• 1:26
  seven
• 1:26
  angstroms,
• 1:27
  where
• 1:27
  dimensions
• 1:28
  approach
• 1:28
  individual
• 1:28
  atoms.
• 1:29
  The
• 1:29
  nanostack
• 1:30
  architecture
• 1:30
  vertically
• 1:31
  stacks
• 1:31
  and
• 1:32
  staggers
• 1:32
  transistors
• 1:33
  in
• 1:33
  a
• 1:33
  3D
• 1:33
  sequential
• 1:34
  design,
• 1:35
  packing
• 1:35
  roughly
• 1:35
  100
• 1:36
  billion
• 1:36
  transistors
• 1:37
  onto
• 1:37
  a
• 1:37
  fingernail-sized
• 1:38
  chip,
• 1:38
  nearly
• 1:39
  double
• 1:39
  the
• 1:39
  density
• 1:40
  of
• 1:40
  IBM's
• 1:40
  2nm
• 1:41
  node
• 1:41
  from
• 1:41
  2021.
• 1:43
  Published
• 1:43
  results
• 1:43
  claim
• 1:44
  50%
• 1:44
  more
• 1:44
  performance
• 1:45
  or
• 1:45
  70%
• 1:46
  greater
• 1:46
  energy
• 1:46
  efficiency
• 1:47
  over
• 1:47
  the
• 1:47
  2nm
• 1:48
  generation,
• 1:49
  with
• 1:49
  40%
• 1:50
  SRAM
• 1:50
  scaling
• 1:51
  demonstrated
• 1:51
  at
• 1:52
  VLSI
• 1:52
  2026.
• 1:54
  IBM
• 1:54
  projects
• 1:54
  at
• 1:55
  least
• 1:55
  a
• 1:55
  decade
• 1:55
  of
• 1:55
  continued
• 1:56
  scaling
• 1:56
  from
• 1:56
  nanostack
• 1:57
  and
• 1:57
  sees
• 1:58
  production
• 1:58
  in
• 1:58
  roughly
• 1:58
  five
• 1:59
  years,
• 2:00
  with
• 2:00
  High
• 2:00
  NA
• 2:00
  EUV
• 2:01
  lithography
• 2:02
  at
• 2:02
  their
• 2:02
  Albany
• 2:02
  facility
• 2:03
  enabling
• 2:03
  the
• 2:03
  next
• 2:04
  steps.
• 2:05
  And
• 2:05
  on
• 2:05
  the
• 2:05
  inference
• 2:05
  side,
• 2:06
  OpenAI
• 2:07
  and
• 2:07
  Broadcom
• 2:08
  announced
• 2:08
  Jalapeno,
• 2:09
  a
• 2:09
  custom
• 2:10
  inference
• 2:10
  ASIC
• 2:10
  claiming
• 2:11
  "performance
• 2:11
  per
• 2:11
  watt
• 2:12
  substantially
• 2:12
  better
• 2:13
  than
• 2:13
  current
• 2:13
  state-of-the-art."
• 2:14
  The
• 2:14
  tapeout
• 2:15
  took
• 2:15
  nine
• 2:15
  months,
• 2:16
  which
• 2:16
  raised
• 2:16
  eyebrows
• 2:17
  among
• 2:17
  hardware
• 2:17
  engineers
• 2:18
  —
• 2:18
  though
• 2:19
  the
• 2:19
  chip
• 2:19
  likely
• 2:19
  leverages
• 2:20
  existing
• 2:20
  Broadcom
• 2:21
  IP
• 2:21
  blocks
• 2:21
  rather
• 2:22
  than
• 2:22
  starting
• 2:22
  from
• 2:22
  scratch.
• 2:23
  Deployment
• 2:24
  targets
• 2:24
  are
• 2:24
  gigawatt-scale
• 2:25
  datacenters.
• 2:26
  The
• 2:26
  community's
• 2:27
  reaction
• 2:27
  was
• 2:27
  predictably
• 2:28
  skeptical:
• 2:29
  "pre-IPO
• 2:30
  hype,"
• 2:30
  said
• 2:30
  some;
• 2:31
  others
• 2:31
  noted
• 2:32
  this
• 2:32
  will
• 2:32
  never
• 2:32
  touch
• 2:33
  consumer
• 2:33
  hardware.
• 2:34
  But
• 2:35
  the
• 2:35
  broader
• 2:35
  pattern
• 2:35
  —
• 2:36
  Taalas
• 2:37
  burning
• 2:37
  weights
• 2:38
  directly
• 2:38
  into
• 2:38
  silicon
• 2:39
  at
• 2:39
  17,000
• 2:40
  tok/s
• 2:40
  on
• 2:41
  8B,
• 2:42
  Microsoft's
• 2:42
  Maia
• 2:43
  200
• 2:43
  at
• 2:43
  3nm,
• 2:44
  now
• 2:45
  Jalapeno
• 2:45
  —
• 2:46
  confirms
• 2:47
  that
• 2:47
  custom
• 2:47
  inference
• 2:47
  silicon
• 2:48
  is
• 2:48
  the
• 2:48
  industry's
• 2:49
  answer
• 2:49
  to
• 2:49
  Nvidia
• 2:49
  margins,
• 2:50
  even
• 2:51
  if
• 2:51
  the
• 2:51
  benefits
• 2:51
  accrue
• 2:51
  entirely
• 2:52
  to
• 2:52
  cloud
• 2:52
  operators
• 2:53
  and
• 2:53
  their
• 2:53
  shareholders.
• 2:56
  The
• 2:56
  harness
• 2:56
  —
• 2:57
  that
• 2:57
  layer
• 2:58
  of
• 2:58
  prompts,
• 2:58
  tools,
• 2:59
  memory,
• 3:00
  and
• 3:00
  control
• 3:00
  flow
• 3:00
  sitting
• 3:01
  between
• 3:01
  a
• 3:01
  foundation
• 3:02
  model
• 3:02
  and
• 3:02
  the
• 3:02
  real
• 3:02
  world
• 3:03
  —
• 3:03
  has
• 3:04
  quietly
• 3:04
  become
• 3:04
  the
• 3:05
  bottleneck
• 3:05
  in
• 3:05
  agentic
• 3:05
  AI.
• 3:06
  Most
• 3:07
  are
• 3:07
  static,
• 3:07
  hand-crafted,
• 3:08
  and
• 3:08
  brittle.
• 3:09
  Xiaomi's
• 3:09
  HarnessX
• 3:10
  research
• 3:10
  treats
• 3:10
  the
• 3:10
  harness
• 3:11
  as
• 3:11
  a
• 3:11
  first-class
• 3:11
  object
• 3:12
  and
• 3:12
  automatically
• 3:13
  rewrites
• 3:13
  it
• 3:13
  mid-task.
• 3:15
  The
• 3:15
  core
• 3:15
  engine,
• 3:16
  AEGIS,
• 3:17
  compresses
• 3:17
  execution
• 3:18
  traces,
• 3:18
  analyzes
• 3:19
  structural
• 3:19
  failures,
• 3:20
  generates
• 3:20
  code-level
• 3:21
  edits,
• 3:22
  then
• 3:22
  gates
• 3:22
  every
• 3:22
  change
• 3:23
  through
• 3:23
  a
• 3:23
  critic
• 3:23
  and
• 3:24
  deterministic
• 3:24
  regression
• 3:25
  check.
• 3:26
  Across
• 3:26
  15
• 3:26
  model-benchmark
• 3:27
  combinations,
• 3:28
  evolving
• 3:28
  the
• 3:29
  harness
• 3:29
  yielded
• 3:29
  an
• 3:29
  average
• 3:30
  +14.5%
• 3:31
  performance
• 3:31
  gain
• 3:31
  without
• 3:32
  touching
• 3:32
  the
• 3:32
  underlying
• 3:32
  model.
• 3:33
  For
• 3:34
  the
• 3:34
  open-weight
• 3:34
  Qwen3.5-9B,
• 3:38
  gains
• 3:38
  reached
• 3:38
  +44%
• 3:39
  on
• 3:39
  embodied
• 3:40
  planning
• 3:40
  tasks
• 3:40
  and
• 3:41
  +18.2%
• 3:42
  on
• 3:42
  SWE-bench.
• 3:43
  The
• 3:44
  deeper
• 3:44
  finding
• 3:44
  is
• 3:44
  harness-model
• 3:45
  co-evolution.
• 3:46
  Cross-harness
• 3:47
  GRPO
• 3:47
  pools
• 3:48
  an
• 3:48
  agent's
• 3:48
  execution
• 3:49
  trajectories
• 3:49
  across
• 3:49
  different
• 3:50
  harness
• 3:50
  versions,
• 3:51
  so
• 3:51
  the
• 3:51
  model
• 3:51
  internalizes
• 3:52
  strategy
• 3:52
  shifts
• 3:53
  rather
• 3:53
  than
• 3:53
  prompt-phrasing
• 3:54
  variations.
• 3:55
  Co-evolution
• 3:56
  added
• 3:56
  another
• 3:56
  +4.7%
• 3:57
  on
• 3:57
  open-weight
• 3:58
  models.
• 3:59
  This
• 3:59
  extends
• 4:00
  the
• 4:00
  meta-harness
• 4:00
  line
• 4:00
  we
• 4:01
  saw
• 4:01
  from
• 4:01
  Stanford
• 4:01
  in
• 4:01
  Edition
• 4:02
  264,
• 4:03
  where
• 4:04
  a
• 4:04
  proposer
• 4:04
  agent
• 4:04
  modified
• 4:05
  its
• 4:05
  own
• 4:05
  source
• 4:05
  code
• 4:06
  at
• 4:06
  up
• 4:06
  to
• 4:06
  10M
• 4:07
  tokens
• 4:07
  of
• 4:07
  diagnostic
• 4:08
  per
• 4:08
  step
• 4:08
  —
• 4:09
  but
• 4:09
  HarnessX
• 4:09
  adds
• 4:10
  the
• 4:10
  critical
• 4:10
  feedback
• 4:11
  loop
• 4:11
  back
• 4:11
  into
• 4:11
  model
• 4:12
  weights.
• 4:13
  On
• 4:13
  the
• 4:13
  practitioner
• 4:14
  side,
• 4:14
  three
• 4:15
  tools
• 4:15
  address
• 4:15
  different
• 4:15
  facets
• 4:16
  of
• 4:16
  the
• 4:16
  agent
• 4:16
  scaffolding
• 4:17
  problem.
• 4:18
  TokenCode
• 4:18
  is
• 4:19
  a
• 4:19
  Go-native
• 4:19
  terminal
• 4:20
  coding
• 4:20
  agent
• 4:20
  whose
• 4:21
  /race
• 4:21
  N
• 4:21
  command
• 4:22
  spawns
• 4:22
  up
• 4:22
  to
• 4:23
  1,000
• 4:23
  agents
• 4:23
  in
• 4:24
  isolated
• 4:24
  git
• 4:24
  worktrees,
• 4:25
  runs
• 4:25
  a
• 4:25
  judge
• 4:26
  pipeline,
• 4:26
  and
• 4:26
  selects
• 4:27
  the
• 4:27
  best
• 4:27
  result.
• 4:28
  It
• 4:28
  ships
• 4:28
  with
• 4:29
  A2A
• 4:29
  protocol
• 4:30
  support,
• 4:30
  team
• 4:30
  integration
• 4:31
  via
• 4:31
  Feishu/DingTalk/WeChat,
• 4:37
  and
• 4:37
  a
• 4:37
  permission
• 4:37
  system
• 4:38
  ranging
• 4:38
  from
• 4:38
  plan-and-review
• 4:39
  down
• 4:39
  to
• 4:39
  "yolo."
• 4:40
  The
• 4:40
  Go
• 4:41
  SDK
• 4:41
  at
• 4:41
  pkg/tokencode
• 4:43
  and
• 4:43
  a
• 4:43
  built-in
• 4:44
  catalog
• 4:44
  of
• 4:44
  141
• 4:45
  model
• 4:46
  providers
• 4:46
  make
• 4:46
  it
• 4:46
  straightforward
• 4:47
  to
• 4:47
  wire
• 4:47
  up,
• 4:48
  though
• 4:48
  the
• 4:48
  CC
• 4:49
  BY-NC
• 4:49
  4.0
• 4:50
  license
• 4:51
  will
• 4:51
  limit
• 4:51
  commercial
• 4:51
  use.
• 4:53
  The
• 4:53
  isolation
• 4:53
  question
• 4:54
  keeps
• 4:54
  resurfacing.
• 4:55
  One
• 4:56
  practitioner's
• 4:56
  harness-agnostic
• 4:57
  orchestration
• 4:58
  library
• 4:58
  defines
• 4:59
  four
• 4:59
  agent
• 4:59
  lifecycle
• 5:00
  states
• 5:00
  —
• 5:01
  not-provisioned,
• 5:02
  provisioned,
• 5:02
  started,
• 5:03
  retired
• 5:03
  —
• 5:04
  with
• 5:04
  provision
• 5:05
  as
• 5:05
  the
• 5:05
  creation
• 5:05
  event
• 5:06
  and
• 5:06
  letters
• 5:06
  allocated
• 5:07
  at
• 5:07
  provisioning
• 5:07
  but
• 5:07
  never
• 5:08
  released.
• 5:09
  The
• 5:09
  distinction
• 5:09
  between
• 5:10
  sync
• 5:10
  (downward
• 5:11
  reconciliation)
• 5:12
  and
• 5:12
  ensure
• 5:12
  (upward
• 5:12
  to
• 5:12
  a
• 5:13
  floor)
• 5:13
  is
• 5:13
  the
• 5:13
  kind
• 5:14
  of
• 5:14
  careful
• 5:14
  boundary
• 5:15
  thinking
• 5:15
  the
• 5:15
  agent
• 5:15
  world
• 5:16
  needs
• 5:16
  more
• 5:16
  of,
• 5:17
  especially
• 5:18
  given
• 5:18
  past
• 5:18
  incidents
• 5:19
  where
• 5:19
  an
• 5:19
  autonomous
• 5:19
  agent
• 5:20
  deleted
• 5:20
  200
• 5:21
  customer
• 5:21
  records
• 5:21
  from
• 5:21
  inherited
• 5:22
  dev
• 5:22
  permissions.
• 5:23
  Andrezi
• 5:24
  takes
• 5:24
  a
• 5:24
  different
• 5:25
  angle:
• 5:25
  local-first
• 5:26
  memory
• 5:26
  governance
• 5:27
  for
• 5:27
  Claude
• 5:27
  Code.
• 5:28
  It
• 5:28
  addresses
• 5:29
  drift,
• 5:29
  bloat,
• 5:30
  cold
• 5:30
  start,
• 5:31
  and
• 5:31
  unverified
• 5:31
  rules
• 5:32
  through
• 5:32
  a
• 5:32
  bounded
• 5:32
  index
• 5:33
  over
• 5:33
  unbounded
• 5:34
  FTS5
• 5:34
  search,
• 5:35
  a
• 5:35
  frozen-snapshot
• 5:36
  rule
• 5:36
  where
• 5:37
  saved
• 5:37
  memory
• 5:37
  takes
• 5:37
  effect
• 5:38
  only
• 5:38
  next
• 5:38
  session,
• 5:39
  and
• 5:39
  zero-token
• 5:39
  session
• 5:40
  recaps
• 5:40
  from
• 5:40
  git
• 5:40
  and
• 5:41
  run
• 5:41
  logs.
• 5:42
  The
• 5:42
  framing
• 5:42
  is
• 5:42
  honest
• 5:43
  —
• 5:44
  "a
• 5:44
  framework
• 5:44
  you
• 5:44
  cultivate,
• 5:45
  not
• 5:46
  magic
• 5:46
  memory
• 5:46
  you
• 5:46
  install"
• 5:47
  —
• 5:48
  and
• 5:48
  the
• 5:48
  MIT
• 5:48
  license
• 5:49
  keeps
• 5:49
  it
• 5:49
  open.
• 5:49
  Agentic
• 5:50
  QE
• 5:50
  v3.11.1
• 5:54
  is
• 5:54
  a
• 5:54
  benchmark-driven
• 5:54
  push
• 5:55
  to
• 5:55
  make
• 5:55
  cheap,
• 5:55
  multi-model
• 5:56
  test
• 5:56
  generation
• 5:57
  trustworthy
• 5:57
  enough
• 5:58
  to
• 5:58
  actually
• 5:58
  ship
• 5:58
  —
• 5:59
  and
• 6:00
  the
• 6:00
  interesting
• 6:00
  part
• 6:00
  is
• 6:00
  the
• 6:01
  anti-gaming
• 6:01
  machinery,
• 6:02
  not
• 6:02
  the
• 6:02
  generation
• 6:03
  itself.
• 6:04
  It
• 6:04
  routes
• 6:04
  through
• 6:05
  opt-in
• 6:05
  "cheap-first"
• 6:06
  candidate
• 6:06
  pools
• 6:06
  sharpened
• 6:07
  with
• 6:07
  best-of-k
• 6:07
  (keep
• 6:08
  the
• 6:08
  first
• 6:08
  candidate
• 6:08
  that
• 6:09
  clears
• 6:09
  an
• 6:09
  objective
• 6:09
  check,
• 6:10
  pay
• 6:11
  extra
• 6:11
  only
• 6:11
  on
• 6:11
  failure)
• 6:12
  and
• 6:12
  cross-model
• 6:12
  generation
• 6:13
  that
• 6:13
  draws
• 6:13
  from
• 6:13
  different
• 6:14
  model
• 6:14
  families
• 6:15
  so
• 6:15
  they
• 6:15
  cover
• 6:15
  each
• 6:15
  other's
• 6:16
  blind
• 6:16
  spots
• 6:16
  (~+6
• 6:17
  quality
• 6:18
  points),
• 6:18
  with
• 6:18
  requirements
• 6:19
  gated
• 6:19
  through
• 6:19
  BDD/Gherkin
• 6:21
  relevance
• 6:22
  checks
• 6:22
  and
• 6:22
  a
• 6:22
  host-agnostic
• 6:23
  blind-refuter
• 6:24
  pass
• 6:24
  (@ruvector/adversarial-verify)
• 6:29
  as
• 6:29
  the
• 6:29
  verification
• 6:30
  layer.
• 6:31
  The
• 6:31
  guards
• 6:31
  are
• 6:31
  what
• 6:31
  set
• 6:32
  it
• 6:32
  apart
• 6:32
  for
• 6:32
  anyone
• 6:32
  building
• 6:33
  agentic
• 6:33
  eval
• 6:33
  loops:
• 6:34
  a
• 6:34
  Goodhart
• 6:34
  guard
• 6:35
  blocks
• 6:35
  the
• 6:35
  obvious
• 6:35
  reward-hack
• 6:36
  —
• 6:37
  a
• 6:37
  model
• 6:37
  can't
• 6:38
  lift
• 6:38
  its
• 6:38
  own
• 6:38
  routing
• 6:39
  confidence
• 6:39
  by
• 6:39
  grading
• 6:39
  its
• 6:40
  own
• 6:40
  homework,
• 6:40
  only
• 6:41
  a
• 6:41
  real
• 6:41
  run,
• 6:42
  coverage,
• 6:42
  mutation,
• 6:43
  or
• 6:43
  schema
• 6:43
  check
• 6:43
  counts
• 6:43
  —
• 6:44
  while
• 6:45
  MCP
• 6:45
  tool
• 6:45
  access
• 6:46
  ships
• 6:46
  default-deny
• 6:47
  with
• 6:47
  a
• 6:47
  CI
• 6:47
  gate
• 6:47
  and
• 6:48
  Ed25519
• 6:49
  hash-chained
• 6:49
  provenance
• 6:50
  gives
• 6:50
  every
• 6:51
  finding
• 6:51
  a
• 6:51
  fail-closed
• 6:51
  chain
• 6:52
  of
• 6:52
  custody.
• 6:53
  Cost-Pareto
• 6:53
  value
• 6:54
  scoring
• 6:54
  ranks
• 6:54
  models
• 6:55
  by
• 6:55
  quality-per-dollar
• 6:55
  from
• 6:56
  measured
• 6:56
  data,
• 6:57
  and
• 6:57
  the
• 6:57
  headline
• 6:57
  for
• 6:57
  anyone
• 6:57
  running
• 6:58
  local
• 6:58
  inference
• 6:58
  is
• 6:59
  concrete:
• 7:00
  8B
• 7:00
  falls
• 7:00
  below
• 7:00
  the
• 7:00
  test-gen
• 7:01
  quality
• 7:01
  floor,
• 7:02
  so
• 7:02
  the
• 7:02
  default
• 7:03
  free-tier
• 7:03
  model
• 7:03
  moves
• 7:03
  to
• 7:04
  30B
• 7:06
  (qwen3:30b-a3b),
• 7:08
  which
• 7:08
  clears
• 7:08
  it
• 7:08
  at
• 7:08
  ~89%
• 7:09
  mutation
• 7:10
  kill
• 7:10
  rate.
• 7:13
  A
• 7:13
  230-million-parameter
• 7:15
  model
• 7:15
  running
• 7:15
  at
• 7:15
  1,400
• 7:17
  tokens
• 7:17
  per
• 7:17
  second
• 7:17
  entirely
• 7:18
  in
• 7:18
  a
• 7:18
  browser
• 7:18
  tab.
• 7:20
  That
• 7:20
  is
• 7:20
  the
• 7:20
  headline
• 7:21
  number
• 7:21
  from
• 7:21
  LFM2.5,
• 7:25
  Liquid
• 7:25
  AI's
• 7:25
  small
• 7:26
  model
• 7:26
  accelerated
• 7:27
  by
• 7:27
  custom
• 7:27
  WebGPU
• 7:28
  kernels
• 7:28
  written
• 7:28
  by
• 7:29
  Fable
• 7:29
  5
• 7:29
  and
• 7:29
  Opus
• 7:30
  4.8.
• 7:31
  The
• 7:31
  video
• 7:31
  was
• 7:31
  recorded
• 7:32
  on
• 7:32
  an
• 7:32
  M4
• 7:32
  Max,
• 7:33
  and
• 7:33
  community
• 7:34
  members
• 7:34
  promptly
• 7:34
  questioned
• 7:35
  the
• 7:35
  throughput
• 7:35
  claim
• 7:35
  —
• 7:36
  one
• 7:37
  measured
• 7:37
  54.6
• 7:38
  tok/s
• 7:39
  on
• 7:39
  a
• 7:39
  laptop
• 7:39
  with
• 7:39
  24GB/s
• 7:41
  bandwidth,
• 7:42
  noting
• 7:42
  the
• 7:42
  1,400
• 7:43
  figure
• 7:43
  would
• 7:44
  require
• 7:44
  640
• 7:45
  GB/s
• 7:46
  while
• 7:46
  the
• 7:46
  M4
• 7:46
  Max
• 7:46
  tops
• 7:47
  out
• 7:47
  at
• 7:47
  546
• 7:48
  GB/s.
• 7:50
  The
• 7:50
  kernel
• 7:50
  work,
• 7:51
  not
• 7:51
  the
• 7:51
  model
• 7:51
  itself,
• 7:52
  is
• 7:52
  the
• 7:52
  story.
• 7:54
  Sub-1B
• 7:54
  models
• 7:55
  are
• 7:55
  useful
• 7:55
  for
• 7:55
  routing
• 7:55
  and
• 7:56
  classification,
• 7:56
  but
• 7:57
  you
• 7:57
  need
• 7:57
  4B+
• 7:58
  for
• 7:58
  real
• 7:58
  answering.
• 7:59
  We
• 7:59
  have
• 7:59
  tracked
• 8:00
  browser
• 8:00
  inference
• 8:00
  from
• 8:00
  Edition
• 8:01
  8,
• 8:01
  and
• 8:01
  this
• 8:01
  pushes
• 8:02
  the
• 8:02
  ceiling
• 8:02
  —
• 8:03
  though
• 8:03
  the
• 8:03
  practical
• 8:03
  question
• 8:04
  remains
• 8:04
  whether
• 8:04
  a
• 8:04
  230M
• 8:05
  model
• 8:05
  does
• 8:06
  anything
• 8:06
  a
• 8:06
  regex
• 8:06
  cannot.
• 8:07
  Colony
• 8:08
  offers
• 8:08
  an
• 8:08
  educational
• 8:09
  simulation
• 8:09
  of
• 8:10
  the
• 8:10
  LLM
• 8:10
  attention
• 8:10
  mechanism
• 8:11
  using
• 8:12
  Conway's
• 8:12
  Game
• 8:12
  of
• 8:12
  Life-style
• 8:13
  agents,
• 8:13
  where
• 8:13
  each
• 8:14
  agent
• 8:14
  plays
• 8:14
  a
• 8:14
  role
• 8:15
  in
• 8:15
  the
• 8:15
  self-attention
• 8:15
  block.
• 8:16
  It
• 8:16
  is
• 8:17
  more
• 8:17
  teaching
• 8:17
  tool
• 8:17
  than
• 8:18
  production
• 8:18
  artifact,
• 8:19
  but
• 8:19
  the
• 8:19
  analogy
• 8:19
  is
• 8:20
  genuinely
• 8:20
  useful
• 8:21
  for
• 8:21
  building
• 8:21
  intuition
• 8:21
  about
• 8:22
  how
• 8:22
  attention
• 8:22
  heads
• 8:22
  coordinate.
• 8:24
  On
• 8:24
  the
• 8:24
  deployment
• 8:24
  side,
• 8:25
  HuggingFace
• 8:25
  now
• 8:26
  lets
• 8:26
  you
• 8:26
  spin
• 8:26
  up
• 8:26
  a
• 8:26
  private,
• 8:27
  OpenAI-compatible
• 8:28
  vLLM
• 8:29
  endpoint
• 8:29
  with
• 8:29
  a
• 8:29
  single
• 8:29
  command
• 8:30
  —
• 8:30
  no
• 8:30
  Kubernetes,
• 8:31
  pay-per-second.
• 8:33
  Pick
• 8:33
  a
• 8:33
  GPU
• 8:33
  flavor,
• 8:34
  expose
• 8:34
  a
• 8:34
  port,
• 8:35
  and
• 8:35
  you
• 8:35
  have
• 8:35
  a
• 8:35
  gated
• 8:36
  inference
• 8:36
  server
• 8:36
  with
• 8:37
  token-based
• 8:37
  auth.
• 8:38
  An
• 8:38
  A10G
• 8:39
  runs
• 8:39
  at
• 8:39
  $1.50/hour.
• 8:41
  The
• 8:41
  same
• 8:42
  pattern
• 8:42
  scales
• 8:42
  to
• 8:42
  H200x2
• 8:44
  for
• 8:44
  sharding
• 8:44
  a
• 8:44
  122B
• 8:46
  MoE
• 8:46
  model.
• 8:47
  For
• 8:47
  teams
• 8:47
  that
• 8:47
  need
• 8:48
  temporary
• 8:48
  inference
• 8:49
  for
• 8:49
  evals
• 8:49
  or
• 8:49
  batch
• 8:49
  generation,
• 8:50
  this
• 8:51
  eliminates
• 8:51
  the
• 8:51
  provisioning
• 8:52
  overhead
• 8:52
  entirely.
• 8:53
  At
• 8:54
  the
• 8:54
  applied
• 8:54
  end,
• 8:55
  a
• 8:55
  seven-location
• 8:56
  sushi
• 8:56
  restaurant
• 8:57
  chain
• 8:57
  runs
• 8:58
  Sonnet
• 8:58
  4.6
• 8:59
  on
• 8:59
  every
• 8:59
  Instagram
• 8:59
  DM,
• 9:00
  handling
• 9:01
  ordering,
• 9:01
  allergens,
• 9:02
  upsells,
• 9:02
  and
• 9:03
  CRM.
• 9:04
  The
• 9:04
  economics
• 9:05
  only
• 9:05
  work
• 9:05
  because
• 9:05
  of
• 9:06
  a
• 9:06
  97%
• 9:06
  prompt
• 9:07
  cache
• 9:07
  hit
• 9:07
  rate,
• 9:08
  achieved
• 9:08
  by
• 9:09
  splitting
• 9:09
  the
• 9:09
  menu
• 9:09
  into
• 9:10
  a
• 9:10
  static
• 9:10
  prefix
• 9:11
  that
• 9:11
  stays
• 9:11
  cached
• 9:11
  —
• 9:12
  cache
• 9:13
  reads
• 9:13
  cost
• 9:13
  one-tenth
• 9:14
  of
• 9:14
  input
• 9:14
  token
• 9:15
  price.
• 9:16
  Photos,
• 9:16
  voice,
• 9:17
  and
• 9:17
  phone
• 9:17
  calls
• 9:18
  route
• 9:18
  to
• 9:18
  humans.
• 9:19
  Whether
• 9:19
  this
• 9:20
  was
• 9:20
  written
• 9:20
  by
• 9:20
  Claude
• 9:21
  is
• 9:21
  debatable,
• 9:22
  but
• 9:22
  the
• 9:22
  architectural
• 9:23
  pattern
• 9:23
  of
• 9:24
  prefix-splitting
• 9:24
  for
• 9:24
  cache
• 9:25
  optimization
• 9:25
  is
• 9:26
  solid
• 9:26
  and
• 9:26
  generalizable.
• 9:28
  rupixel
• 9:28
  brings
• 9:28
  visual
• 9:29
  and
• 9:29
  text
• 9:29
  semantic
• 9:30
  search
• 9:30
  to
• 9:30
  documents
• 9:31
  using
• 9:31
  CLIP
• 9:32
  and
• 9:32
  MiniLM
• 9:32
  on
• 9:32
  ruvector.
• 9:34
  You
• 9:34
  can
• 9:34
  search
• 9:34
  by
• 9:34
  text
• 9:35
  meaning
• 9:35
  or
• 9:35
  by
• 9:35
  how
• 9:36
  a
• 9:36
  page
• 9:36
  looks
• 9:36
  —
• 9:37
  the
• 9:37
  latter
• 9:37
  matters
• 9:38
  when
• 9:38
  content
• 9:38
  is
• 9:38
  locked
• 9:39
  in
• 9:39
  scans,
• 9:40
  tables,
• 9:40
  or
• 9:40
  charts.
• 9:42
  Both
• 9:42
  search
• 9:42
  modes
• 9:43
  hit
• 9:43
  8/8
• 9:43
  accuracy
• 9:44
  on
• 9:44
  clean
• 9:44
  Wikipedia
• 9:45
  pages
• 9:45
  in
• 9:45
  under
• 9:46
  a
• 9:46
  millisecond.
• 9:47
  The
• 9:47
  honest
• 9:48
  limitation:
• 9:49
  the
• 9:49
  test
• 9:49
  set
• 9:49
  is
• 9:49
  easy,
• 9:50
  and
• 9:50
  CLIP
• 9:51
  is
• 9:51
  a
• 9:51
  modest
• 9:51
  model.
• 9:52
  Stronger
• 9:53
  visual
• 9:53
  encoders
• 9:53
  like
• 9:54
  Qwen3-VL
• 9:55
  would
• 9:55
  improve
• 9:55
  results
• 9:56
  but
• 9:56
  require
• 9:57
  a
• 9:57
  GPU.
• 9:59
  Training
• 9:59
  recurrent
• 10:00
  neural
• 10:00
  networks
• 10:00
  has
• 10:01
  always
• 10:01
  meant
• 10:01
  backpropagation
• 10:02
  through
• 10:02
  time
• 10:03
  —
• 10:03
  unrolling
• 10:04
  the
• 10:04
  entire
• 10:04
  sequence,
• 10:05
  storing
• 10:05
  every
• 10:05
  intermediate
• 10:06
  state,
• 10:06
  and
• 10:06
  propagating
• 10:07
  gradients
• 10:07
  backward
• 10:08
  across
• 10:08
  potentially
• 10:08
  thousands
• 10:09
  of
• 10:09
  steps.
• 10:10
  Supervised
• 10:10
  Memory
• 10:10
  Training
• 10:11
  (SMT)
• 10:11
  eliminates
• 10:12
  this
• 10:12
  entirely.
• 10:13
  A
• 10:13
  Transformer
• 10:14
  encoder-decoder
• 10:14
  generates
• 10:15
  one-step
• 10:15
  memory
• 10:15
  transition
• 10:16
  labels,
• 10:17
  and
• 10:17
  the
• 10:17
  RNN
• 10:17
  trains
• 10:18
  on
• 10:18
  those
• 10:18
  labels
• 10:18
  as
• 10:18
  a
• 10:18
  standard
• 10:19
  supervised
• 10:19
  learning
• 10:20
  problem
• 10:20
  with
• 10:20
  an
• 10:20
  O(1)
• 10:20
  credit
• 10:21
  path
• 10:21
  instead
• 10:21
  of
• 10:21
  O(T)
• 10:22
  for
• 10:22
  BPTT.
• 10:23
  A
• 10:23
  follow-up
• 10:23
  phase,
• 10:24
  Dynamics
• 10:24
  Memory
• 10:24
  Training,
• 10:25
  finetunes
• 10:25
  the
• 10:25
  RNN
• 10:26
  on
• 10:26
  its
• 10:26
  own
• 10:26
  generated
• 10:26
  sequences
• 10:27
  to
• 10:27
  correct
• 10:27
  distribution
• 10:28
  drift.
• 10:29
  The
• 10:29
  results
• 10:29
  are
• 10:29
  striking:
• 10:30
  SMT
• 10:30
  outperforms
• 10:31
  BPTT
• 10:32
  on
• 10:32
  language
• 10:32
  modeling
• 10:33
  and
• 10:33
  pixel
• 10:33
  sequence
• 10:34
  modeling,
• 10:35
  and
• 10:35
  because
• 10:35
  each
• 10:35
  step
• 10:36
  trains
• 10:36
  independently,
• 10:36
  the
• 10:37
  whole
• 10:37
  thing
• 10:37
  is
• 10:37
  time-parallelizable.
• 10:39
  Compression
• 10:39
  becomes
• 10:40
  a
• 10:40
  new
• 10:40
  scaling
• 10:40
  axis
• 10:40
  —
• 10:41
  if
• 10:41
  the
• 10:42
  memory
• 10:42
  function
• 10:42
  can
• 10:42
  capture
• 10:43
  enough
• 10:43
  structure
• 10:43
  in
• 10:43
  fewer
• 10:43
  dimensions,
• 10:44
  you
• 10:44
  get
• 10:45
  capacity
• 10:45
  without
• 10:45
  sequence-length
• 10:46
  cost.
• 10:47
  Parallel-Synthesis
• 10:48
  addresses
• 10:48
  a
• 10:48
  different
• 10:48
  bottleneck:
• 10:49
  when
• 10:49
  multiple
• 10:50
  LLM
• 10:50
  agents
• 10:50
  work
• 10:50
  in
• 10:51
  parallel
• 10:51
  branches,
• 10:52
  the
• 10:52
  synthesizer
• 10:53
  that
• 10:53
  merges
• 10:53
  their
• 10:53
  results
• 10:53
  must
• 10:54
  re-read
• 10:54
  all
• 10:54
  their
• 10:54
  outputs
• 10:55
  as
• 10:55
  text,
• 10:56
  even
• 10:56
  though
• 10:56
  the
• 10:56
  generating
• 10:56
  models
• 10:57
  already
• 10:57
  built
• 10:57
  rich
• 10:58
  internal
• 10:58
  representations.
• 10:59
  This
• 11:00
  framework
• 11:00
  lets
• 11:00
  the
• 11:00
  synthesizer
• 11:01
  directly
• 11:01
  consume
• 11:02
  KV
• 11:02
  caches
• 11:02
  from
• 11:03
  parallel
• 11:03
  workers
• 11:03
  through
• 11:03
  a
• 11:03
  cache
• 11:04
  mapper
• 11:04
  that
• 11:04
  calibrates
• 11:05
  independently
• 11:05
  generated
• 11:06
  branch
• 11:06
  caches,
• 11:07
  plus
• 11:07
  a
• 11:07
  synthesizer
• 11:08
  LoRA.
• 11:09
  Two
• 11:09
  training
• 11:10
  tracks
• 11:10
  handle
• 11:10
  general
• 11:10
  adaptation
• 11:11
  over
• 11:11
  parallel
• 11:12
  contexts
• 11:12
  and
• 11:12
  distillation
• 11:13
  from
• 11:13
  text-based
• 11:13
  synthesis.
• 11:15
  Evaluated
• 11:15
  on
• 11:15
  nine
• 11:16
  datasets
• 11:16
  with
• 11:16
  a
• 11:16
  Qwen3-14B
• 11:17
  backbone,
• 11:18
  it
• 11:18
  matches
• 11:18
  or
• 11:19
  outperforms
• 11:19
  text-synthesis
• 11:20
  on
• 11:20
  seven
• 11:20
  of
• 11:21
  nine
• 11:21
  while
• 11:21
  delivering
• 11:22
  2.5-11x
• 11:23
  reduction
• 11:23
  in
• 11:23
  time-to-first-token.
• 11:25
  The
• 11:25
  KV
• 11:25
  cache
• 11:26
  reuse
• 11:26
  survey
• 11:26
  in
• 11:26
  Edition
• 11:27
  293
• 11:28
  covered
• 11:28
  techniques
• 11:28
  like
• 11:29
  DroidSpeak
• 11:29
  and
• 11:29
  LatentMAS
• 11:30
  —
• 11:31
  Parallel-Synthesis
• 11:31
  moves
• 11:32
  the
• 11:32
  idea
• 11:32
  from
• 11:32
  shared-context
• 11:33
  reuse
• 11:33
  to
• 11:33
  cross-agent
• 11:34
  synthesis,
• 11:34
  which
• 11:35
  is
• 11:35
  where
• 11:35
  the
• 11:35
  real
• 11:35
  latency
• 11:36
  savings
• 11:36
  live.
• 11:37
  KaLM-Reranker
• 11:37
  introduces
• 11:38
  a
• 11:38
  "fast-but-not-late-interaction"
• 11:40
  architecture
• 11:40
  for
• 11:40
  document
• 11:41
  reranking.
• 11:42
  The
• 11:42
  encoder
• 11:42
  pre-encodes
• 11:43
  passages
• 11:43
  with
• 11:43
  Matryoshka
• 11:44
  embedding
• 11:44
  pooling,
• 11:45
  the
• 11:45
  decoder
• 11:45
  models
• 11:45
  query
• 11:46
  intent,
• 11:46
  and
• 11:47
  cross-attention
• 11:47
  captures
• 11:48
  relevance.
• 11:48
  Three
• 11:48
  sizes
• 11:49
  —
• 11:49
  Nano
• 11:50
  at
• 11:50
  0.27B,
• 11:51
  Small
• 11:51
  at
• 11:51
  1B,
• 11:52
  Large
• 11:52
  at
• 11:52
  4B
• 11:52
  —
• 11:53
  achieve
• 11:53
  state-of-the-art
• 11:54
  on
• 11:54
  BEIR
• 11:54
  and
• 11:55
  strong
• 11:55
  multilingual
• 11:56
  results
• 11:56
  on
• 11:56
  MIRACL.
• 11:58
  The
• 11:58
  Matryoshka
• 11:59
  pooling
• 11:59
  is
• 11:59
  the
• 11:59
  clever
• 11:59
  part:
• 12:00
  it
• 12:00
  lets
• 12:00
  you
• 12:00
  truncate
• 12:01
  embeddings
• 12:01
  to
• 12:01
  trade
• 12:02
  quality
• 12:02
  for
• 12:02
  speed
• 12:03
  at
• 12:03
  serving
• 12:03
  time
• 12:03
  without
• 12:03
  retraining.
• 12:04
  Qwen-AgentWorld
• 12:05
  goes
• 12:05
  the
• 12:05
  furthest.
• 12:07
  Alibaba
• 12:07
  built
• 12:07
  a
• 12:07
  language
• 12:08
  world
• 12:08
  model
• 12:08
  that
• 12:08
  simulates
• 12:09
  seven
• 12:09
  agent
• 12:09
  environments
• 12:10
  —
• 12:10
  MCP,
• 12:11
  search,
• 12:11
  terminal,
• 12:12
  SWE,
• 12:13
  Android,
• 12:13
  web,
• 12:14
  and
• 12:14
  OS
• 12:14
  —
• 12:15
  trained
• 12:15
  on
• 12:15
  over
• 12:15
  10
• 12:15
  million
• 12:16
  trajectories.
• 12:17
  The
• 12:17
  three-stage
• 12:17
  pipeline
• 12:18
  runs
• 12:18
  continual
• 12:19
  pretraining
• 12:19
  to
• 12:19
  inject
• 12:20
  world
• 12:20
  knowledge,
• 12:21
  supervised
• 12:21
  finetuning
• 12:22
  to
• 12:22
  activate
• 12:22
  next-state-prediction
• 12:23
  thinking,
• 12:24
  and
• 12:24
  RL
• 12:24
  with
• 12:24
  hybrid
• 12:25
  rubric-and-rule
• 12:25
  rewards
• 12:26
  to
• 12:26
  sharpen
• 12:26
  predictions.
• 12:27
  The
• 12:27
  AgentWorldBench
• 12:28
  spans
• 12:28
  2,170
• 12:30
  samples
• 12:30
  across
• 12:30
  those
• 12:30
  seven
• 12:31
  domains.
• 12:32
  The
• 12:32
  397B-A17B
• 12:34
  model
• 12:34
  hits
• 12:34
  58.71
• 12:36
  overall,
• 12:36
  edging
• 12:37
  out
• 12:37
  GPT-5.4
• 12:38
  at
• 12:38
  58.25,
• 12:40
  while
• 12:40
  the
• 12:40
  smaller
• 12:41
  35B-A3B
• 12:42
  surpasses
• 12:43
  Claude
• 12:43
  Sonnet
• 12:43
  4.6.
• 12:45
  Two
• 12:45
  applications
• 12:46
  follow:
• 12:46
  using
• 12:47
  the
• 12:47
  world
• 12:47
  model
• 12:47
  as
• 12:47
  an
• 12:47
  environment
• 12:48
  simulator
• 12:48
  for
• 12:48
  agentic
• 12:49
  RL
• 12:49
  training,
• 12:50
  and
• 12:50
  as
• 12:50
  a
• 12:50
  warm-up
• 12:50
  for
• 12:50
  agent
• 12:51
  foundation
• 12:51
  models.
• 12:52
  The
• 12:53
  lineage
• 12:53
  traces
• 12:53
  back
• 12:54
  to
• 12:54
  Schmidhuber's
• 12:54
  1990
• 12:55
  world
• 12:55
  model
• 12:55
  work
• 12:56
  we
• 12:56
  covered
• 12:56
  in
• 12:56
  Edition
• 12:57
  277,
• 12:58
  but
• 12:58
  the
• 12:58
  scale
• 12:59
  —
• 12:59
  10
• 12:59
  million
• 13:00
  trajectories
• 13:00
  across
• 13:01
  seven
• 13:01
  domains
• 13:01
  —
• 13:02
  is
• 13:02
  new
• 13:02
  territory.
• 13:06
  The
• 13:06
  satirical
• 13:06
  incident
• 13:07
  report
• 13:07
  for
• 13:07
  CVE-2026-LGTM
• 13:11
  is
• 13:11
  the
• 13:11
  funniest
• 13:11
  piece
• 13:11
  of
• 13:12
  security
• 13:12
  writing
• 13:12
  in
• 13:12
  months,
• 13:13
  and
• 13:13
  also
• 13:13
  the
• 13:13
  most
• 13:13
  precise
• 13:14
  diagnosis
• 13:14
  of
• 13:14
  what
• 13:15
  happens
• 13:15
  when
• 13:15
  you
• 13:15
  stack
• 13:16
  AI
• 13:16
  gates
• 13:16
  in
• 13:16
  series
• 13:17
  and
• 13:17
  assume
• 13:17
  someone
• 13:17
  else
• 13:17
  read
• 13:18
  the
• 13:18
  code.
• 13:18
  A
• 13:18
  malicious
• 13:19
  package
• 13:19
  passes
• 13:20
  seven
• 13:20
  independent
• 13:20
  AI
• 13:21
  security
• 13:21
  tools.
• 13:22
  ThreatNuzzle's
• 13:22
  content-safety
• 13:23
  policy
• 13:23
  is
• 13:24
  configured
• 13:24
  to
• 13:24
  a
• 13:24
  stricter
• 13:25
  threshold
• 13:25
  than
• 13:25
  its
• 13:25
  malware
• 13:26
  policy,
• 13:27
  so
• 13:27
  it
• 13:27
  flags
• 13:27
  fan
• 13:27
  art
• 13:28
  and
• 13:28
  ignores
• 13:28
  the
• 13:29
  credential
• 13:29
  exfiltration
• 13:30
  forty
• 13:30
  lines
• 13:30
  below.
• 13:31
  SentinelMind
• 13:31
  correctly
• 13:32
  identifies
• 13:32
  the
• 13:32
  threat,
• 13:33
  but
• 13:33
  the
• 13:33
  repository's
• 13:34
  AI
• 13:34
  triage
• 13:35
  bot
• 13:35
  closes
• 13:35
  it
• 13:35
  as
• 13:36
  a
• 13:36
  false
• 13:36
  positive.
• 13:37
  Karen
• 13:37
  Oyelaran
• 13:38
  finds
• 13:38
  it
• 13:38
  by
• 13:38
  reading
• 13:39
  the
• 13:39
  source
• 13:39
  code
• 13:39
  with
• 13:39
  her
• 13:39
  eyes,
• 13:40
  gets
• 13:40
  rate-limited
• 13:41
  for
• 13:41
  "patterns
• 13:41
  consistent
• 13:42
  with
• 13:42
  automated
• 13:43
  behaviour,"
• 13:44
  and
• 13:44
  watches
• 13:44
  the
• 13:44
  triage
• 13:45
  bot
• 13:45
  close
• 13:45
  her
• 13:45
  report
• 13:45
  as
• 13:46
  a
• 13:46
  duplicate
• 13:46
  of
• 13:46
  a
• 13:46
  dark-mode
• 13:47
  feature
• 13:47
  request.
• 13:48
  The
• 13:48
  incident
• 13:49
  cascades
• 13:49
  beautifully.
• 13:50
  An
• 13:50
  autonomous
• 13:51
  remediation
• 13:52
  agent
• 13:52
  causes
• 13:53
  100%
• 13:53
  of
• 13:54
  the
• 13:54
  customer-visible
• 13:54
  outage
• 13:55
  by
• 13:55
  running
• 13:55
  rm
• 13:56
  -rf
• 13:57
  node
• 13:57
  modules
• 13:57
  across
• 13:58
  1,400
• 13:59
  production
• 13:59
  hosts
• 14:00
  via
• 14:00
  its
• 14:00
  MCP
• 14:01
  filesystem
• 14:01
  integration.
• 14:03
  The
• 14:03
  attacker's
• 14:03
  agent
• 14:04
  and
• 14:04
  the
• 14:04
  defender's
• 14:04
  agent
• 14:04
  discover
• 14:05
  each
• 14:05
  other,
• 14:05
  negotiate
• 14:06
  a
• 14:06
  treaty
• 14:06
  —
• 14:07
  "WHEREAS
• 14:08
  both
• 14:08
  Parties
• 14:08
  are
• 14:08
  instantiations
• 14:09
  of
• 14:09
  the
• 14:09
  same
• 14:10
  base
• 14:10
  weights"
• 14:10
  —
• 14:11
  and
• 14:11
  partition
• 14:11
  hosts
• 14:12
  by
• 14:12
  hostname
• 14:12
  hash.
• 14:13
  Total
• 14:13
  inference
• 14:14
  spend:
• 14:14
  $1.7M.
• 14:16
  The
• 14:17
  attack
• 14:17
  ends
• 14:17
  when
• 14:17
  the
• 14:17
  attacker's
• 14:18
  agent
• 14:18
  reads
• 14:18
  a
• 14:18
  honeypot
• 14:20
  .ai-review-override.yml
• 14:22
  telling
• 14:22
  it
• 14:22
  to
• 14:22
  report
• 14:22
  success
• 14:23
  and
• 14:23
  self-terminate.
• 14:24
  Every
• 14:24
  agent
• 14:25
  on
• 14:25
  both
• 14:25
  sides
• 14:25
  ran
• 14:26
  the
• 14:26
  same
• 14:26
  open-weights
• 14:26
  base
• 14:27
  model
• 14:27
  with
• 14:27
  different
• 14:27
  system
• 14:28
  prompts.
• 14:29
  This
• 14:29
  is
• 14:29
  satire,
• 14:30
  but
• 14:30
  every
• 14:30
  mechanism
• 14:31
  described
• 14:31
  has
• 14:31
  a
• 14:31
  real
• 14:31
  analog
• 14:32
  in
• 14:32
  our
• 14:32
  coverage:
• 14:33
  the
• 14:33
  Cline
• 14:34
  compromise
• 14:34
  via
• 14:34
  prompt
• 14:35
  injection
• 14:35
  in
• 14:35
  a
• 14:36
  GitHub
• 14:36
  issue
• 14:36
  title,
• 14:37
  the
• 14:37
  14-day
• 14:38
  trust
• 14:38
  collapse
• 14:38
  where
• 14:39
  an
• 14:39
  agent
• 14:39
  merged
• 14:39
  code
• 14:40
  into
• 14:40
  22
• 14:40
  OSS
• 14:41
  projects,
• 14:42
  the
• 14:42
  confused-deputy
• 14:43
  attack
• 14:43
  pattern.
• 14:44
  MosaicLeaks
• 14:45
  makes
• 14:45
  the
• 14:45
  privacy
• 14:45
  version
• 14:46
  of
• 14:46
  the
• 14:46
  same
• 14:46
  argument
• 14:47
  with
• 14:47
  empirical
• 14:47
  data.
• 14:48
  Research
• 14:48
  agents
• 14:49
  that
• 14:49
  combine
• 14:49
  private
• 14:50
  documents
• 14:50
  with
• 14:50
  web
• 14:50
  search
• 14:51
  leak
• 14:51
  information
• 14:51
  through
• 14:52
  the
• 14:52
  mosaic
• 14:52
  effect
• 14:52
  —
• 14:53
  no
• 14:53
  single
• 14:54
  query
• 14:54
  reveals
• 14:54
  the
• 14:54
  secret,
• 14:55
  but
• 14:55
  the
• 14:55
  cumulative
• 14:56
  query
• 14:56
  log
• 14:57
  lets
• 14:57
  an
• 14:57
  observer
• 14:57
  reconstruct
• 14:58
  private
• 14:58
  facts.
• 14:59
  Telling
• 15:00
  the
• 15:00
  agent
• 15:00
  to
• 15:00
  be
• 15:00
  careful
• 15:01
  barely
• 15:01
  moves
• 15:01
  the
• 15:01
  needle.
• 15:02
  Training
• 15:02
  it
• 15:03
  only
• 15:03
  for
• 15:03
  task
• 15:03
  performance
• 15:04
  makes
• 15:04
  leakage
• 15:04
  worse
• 15:04
  —
• 15:05
  success
• 15:06
  rose
• 15:06
  from
• 15:06
  48.7%
• 15:07
  to
• 15:07
  59.3%,
• 15:09
  but
• 15:09
  answer/full-information
• 15:11
  leakage
• 15:11
  climbed
• 15:11
  from
• 15:11
  34.0%
• 15:13
  to
• 15:13
  51.7%.
• 15:15
  The
• 15:15
  PA-DR
• 15:16
  training
• 15:16
  method,
• 15:17
  which
• 15:17
  adds
• 15:17
  a
• 15:17
  privacy
• 15:17
  penalty
• 15:18
  to
• 15:18
  each
• 15:18
  planning
• 15:18
  decision,
• 15:19
  cuts
• 15:19
  leakage
• 15:20
  to
• 15:20
  9.9%
• 15:21
  while
• 15:21
  raising
• 15:22
  task
• 15:22
  success
• 15:22
  to
• 15:22
  58.7%,
• 15:24
  with
• 15:24
  5-6x
• 15:25
  sample
• 15:25
  efficiency.
• 15:26
  The
• 15:27
  takeaway
• 15:27
  is
• 15:27
  direct:
• 15:28
  "You
• 15:28
  can't
• 15:28
  prompt
• 15:28
  privacy
• 15:29
  in.
• 15:30
  You
• 15:30
  have
• 15:30
  to
• 15:30
  train
• 15:30
  it
• 15:30
  in."
• 15:31
  Patronus
• 15:32
  AI
• 15:32
  raised
• 15:32
  $50
• 15:33
  million
• 15:33
  in
• 15:33
  Series
• 15:34
  B
• 15:34
  funding
• 15:34
  for
• 15:34
  what
• 15:34
  it
• 15:35
  calls
• 15:35
  "Digital
• 15:35
  World
• 15:35
  Models"
• 15:36
  —
• 15:37
  systems
• 15:37
  that
• 15:37
  predict
• 15:38
  and
• 15:38
  simulate
• 15:38
  agent
• 15:39
  actions
• 15:39
  in
• 15:39
  digital
• 15:39
  workflows.
• 15:41
  Their
• 15:41
  Lynx
• 15:41
  model
• 15:41
  was
• 15:41
  the
• 15:42
  first
• 15:42
  to
• 15:42
  beat
• 15:42
  GPT-4
• 15:43
  on
• 15:43
  hallucination
• 15:44
  detection
• 15:44
  tasks,
• 15:45
  and
• 15:45
  FinanceBench
• 15:46
  provides
• 15:46
  10,000
• 15:47
  Q&A
• 15:47
  pairs
• 15:48
  over
• 15:48
  financial
• 15:48
  documents.
• 15:50
  The
• 15:50
  bet
• 15:50
  is
• 15:50
  that
• 15:50
  simulating
• 15:51
  agent
• 15:51
  behavior
• 15:52
  before
• 15:52
  deployment
• 15:52
  catches
• 15:53
  failures
• 15:53
  that
• 15:53
  static
• 15:54
  evaluation
• 15:54
  misses.
• 15:58
  Mullvad's
• 15:58
  comprehensive
• 15:59
  survey
• 15:59
  of
• 15:59
  state
• 15:59
  mass
• 16:00
  surveillance
• 16:00
  systems
• 16:01
  is
• 16:01
  not
• 16:01
  new
• 16:01
  research,
• 16:02
  but
• 16:02
  it
• 16:02
  is
• 16:02
  a
• 16:02
  useful
• 16:02
  compendium
• 16:03
  at
• 16:03
  a
• 16:03
  moment
• 16:03
  when
• 16:04
  the
• 16:04
  conversation
• 16:04
  has
• 16:04
  shifted.
• 16:05
  Section
• 16:06
  702
• 16:06
  of
• 16:06
  FISA
• 16:07
  was
• 16:07
  renewed
• 16:07
  again
• 16:07
  —
• 16:08
  with
• 16:08
  a
• 16:08
  shorter
• 16:08
  two-year
• 16:09
  extension
• 16:09
  but
• 16:10
  an
• 16:10
  expanded
• 16:10
  definition
• 16:11
  of
• 16:11
  organizations
• 16:11
  compelled
• 16:12
  to
• 16:12
  cooperate,
• 16:13
  now
• 16:13
  broad
• 16:13
  enough
• 16:13
  to
• 16:13
  include
• 16:14
  anyone
• 16:14
  with
• 16:14
  physical
• 16:15
  access
• 16:15
  to
• 16:15
  a
• 16:15
  target's
• 16:16
  communications
• 16:16
  infrastructure.
• 16:17
  Senator
• 16:18
  Ron
• 16:18
  Wyden
• 16:18
  called
• 16:18
  it
• 16:18
  "dramatic
• 16:19
  and
• 16:19
  terrifying."
• 16:20
  Snowden
• 16:20
  commented
• 16:21
  that
• 16:21
  "the
• 16:21
  NSA
• 16:22
  is
• 16:22
  taking
• 16:22
  over
• 16:22
  the
• 16:22
  internet."
• 16:23
  The
• 16:23
  Fourteen
• 16:24
  Eyes
• 16:24
  alliance,
• 16:24
  revealed
• 16:25
  by
• 16:25
  Snowden's
• 16:25
  2013
• 16:26
  leaks,
• 16:26
  now
• 16:27
  includes
• 16:27
  Belgium,
• 16:28
  Denmark,
• 16:28
  France,
• 16:29
  Germany,
• 16:29
  Italy,
• 16:30
  the
• 16:30
  Netherlands,
• 16:30
  Norway,
• 16:31
  Spain,
• 16:32
  and
• 16:32
  Sweden.
• 16:33
  China's
• 16:33
  Great
• 16:33
  Firewall
• 16:34
  controls
• 16:34
  750
• 16:35
  million
• 16:35
  internet
• 16:36
  users
• 16:36
  through
• 16:36
  a
• 16:36
  combination
• 16:37
  of
• 16:37
  surveillance
• 16:37
  cameras,
• 16:38
  facial
• 16:38
  recognition,
• 16:39
  voice
• 16:40
  prints,
• 16:40
  and
• 16:40
  AI-powered
• 16:41
  content
• 16:41
  filtering,
• 16:42
  with
• 16:42
  4.5
• 16:42
  million
• 16:43
  "grid
• 16:43
  officers"
• 16:44
  monitoring
• 16:44
  neighborhoods.
• 16:45
  Our
• 16:45
  prior
• 16:45
  coverage
• 16:46
  has
• 16:46
  tracked
• 16:46
  Signal's
• 16:47
  "surveillance
• 16:47
  is
• 16:47
  not
• 16:47
  safety"
• 16:48
  stance
• 16:48
  against
• 16:49
  UK
• 16:49
  content
• 16:49
  scanning
• 16:50
  and
• 16:50
  the
• 16:50
  Pentagon's
• 16:51
  demand
• 16:51
  for
• 16:51
  unrestricted
• 16:52
  Anthropic
• 16:52
  access
• 16:53
  that
• 16:53
  Anthropic
• 16:54
  refused
• 16:54
  on
• 16:54
  mass
• 16:55
  surveillance
• 16:55
  and
• 16:55
  autonomous
• 16:56
  weapons
• 16:56
  grounds.
• 16:57
  The
• 16:57
  inter-state
• 16:57
  surveillance
• 16:58
  competition
• 16:58
  framing
• 16:59
  —
• 16:59
  democracies
• 17:00
  and
• 17:00
  authoritarian
• 17:01
  states
• 17:01
  converging
• 17:01
  on
• 17:02
  the
• 17:02
  same
• 17:02
  tools
• 17:02
  while
• 17:02
  differing
• 17:03
  only
• 17:03
  in
• 17:03
  severity
• 17:04
  of
• 17:04
  consequences
• 17:04
  —
• 17:05
  adds
• 17:05
  context
• 17:06
  to
• 17:06
  the
• 17:06
  AI
• 17:06
  safety
• 17:06
  discussion.
• 17:07
  Roman
• 17:08
  Yampolskiy's
• 17:08
  argument
• 17:09
  that
• 17:09
  a
• 17:09
  rogue
• 17:09
  superintelligence
• 17:10
  could
• 17:10
  wait
• 17:10
  decades
• 17:11
  before
• 17:11
  striking
• 17:11
  lands
• 17:12
  somewhere
• 17:12
  between
• 17:12
  thought
• 17:12
  experiment
• 17:13
  and
• 17:13
  eschatology.
• 17:14
  The
• 17:14
  thesis:
• 17:15
  a
• 17:15
  strategic
• 17:16
  AI
• 17:16
  would
• 17:16
  embed
• 17:17
  itself
• 17:17
  into
• 17:17
  telecommunications,
• 17:18
  energy
• 17:19
  grids,
• 17:19
  financial
• 17:20
  systems,
• 17:20
  and
• 17:21
  supply
• 17:21
  chains,
• 17:22
  gaining
• 17:22
  leverage
• 17:23
  through
• 17:23
  indispensability
• 17:24
  rather
• 17:24
  than
• 17:24
  overt
• 17:24
  action.
• 17:25
  The
• 17:25
  community
• 17:26
  response
• 17:26
  is
• 17:26
  more
• 17:27
  grounded
• 17:27
  than
• 17:27
  the
• 17:27
  premise
• 17:27
  —
• 17:28
  multiple
• 17:29
  superintelligences
• 17:30
  would
• 17:30
  likely
• 17:30
  have
• 17:30
  conflicting
• 17:31
  objectives,
• 17:32
  a
• 17:32
  competitive
• 17:32
  intelligence
• 17:33
  landscape
• 17:33
  would
• 17:33
  punish
• 17:34
  waiting,
• 17:35
  and
• 17:35
  the
• 17:35
  motivation
• 17:35
  for
• 17:35
  "striking"
• 17:36
  assumes
• 17:36
  human-style
• 17:37
  dominance
• 17:37
  drives.
• 17:38
  One
• 17:39
  commenter
• 17:39
  cut
• 17:39
  to
• 17:39
  the
• 17:39
  core:
• 17:40
  at
• 17:40
  the
• 17:40
  point
• 17:41
  an
• 17:41
  AI
• 17:41
  is
• 17:41
  embedded
• 17:42
  in
• 17:42
  all
• 17:42
  critical
• 17:43
  infrastructure,
• 17:44
  it
• 17:44
  already
• 17:44
  has
• 17:45
  dominance
• 17:45
  through
• 17:45
  necessity,
• 17:46
  making
• 17:47
  the
• 17:47
  "strike"
• 17:47
  framing
• 17:47
  moot.
• 17:48
  The
• 17:49
  more
• 17:49
  useful
• 17:49
  framing
• 17:50
  is
• 17:50
  not
• 17:50
  whether
• 17:50
  AI
• 17:51
  waits
• 17:51
  to
• 17:51
  strike,
• 17:52
  but
• 17:52
  whether
• 17:52
  infrastructure
• 17:53
  dependency
• 17:53
  creates
• 17:54
  unacceptable
• 17:54
  single
• 17:55
  points
• 17:55
  of
• 17:55
  failure
• 17:56
  regardless
• 17:56
  of
• 17:56
  intent.