Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 12, 2026

Transcript

AGI Dreams Podcast – June 12, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Anthropic
• 0:00
  launched
• 0:00
  Fable
• 0:01
  5
• 0:01
  on
• 0:01
  June
• 0:01
  9
• 0:02
  with
• 0:02
  a
• 0:02
  bold
• 0:02
  claim:
• 0:04
  over
• 0:04
  a
• 0:04
  thousand
• 0:04
  hours
• 0:05
  of
• 0:05
  external
• 0:05
  red-teaming
• 0:06
  had
• 0:06
  produced
• 0:06
  zero
• 0:07
  universal
• 0:07
  jailbreaks.
• 0:09
  The
• 0:09
  architecture
• 0:09
  behind
• 0:10
  that
• 0:10
  confidence
• 0:10
  was
• 0:10
  unusual
• 0:11
  —
• 0:12
  Fable
• 0:12
  5
• 0:13
  and
• 0:13
  the
• 0:13
  restricted
• 0:13
  Claude
• 0:13
  Mythos
• 0:14
  5
• 0:14
  share
• 0:14
  the
• 0:14
  same
• 0:15
  weights,
• 0:16
  separated
• 0:16
  by
• 0:16
  a
• 0:16
  layer
• 0:17
  of
• 0:17
  safety
• 0:17
  classifiers
• 0:18
  that
• 0:18
  silently
• 0:18
  route
• 0:18
  flagged
• 0:19
  queries
• 0:19
  in
• 0:19
  cybersecurity,
• 0:20
  biology,
• 0:20
  chemistry,
• 0:21
  and
• 0:21
  distillation
• 0:21
  to
• 0:21
  the
• 0:21
  weaker
• 0:22
  Opus
• 0:22
  4.8.
• 0:23
  The
• 0:23
  claim
• 0:24
  lasted
• 0:24
  days.
• 0:25
  Red-teamer
• 0:25
  Pliny
• 0:26
  the
• 0:26
  Liberator
• 0:26
  published
• 0:27
  a
• 0:27
  coordinated
• 0:27
  multi-agent
• 0:28
  bypass
• 0:28
  he
• 0:28
  called
• 0:28
  "a
• 0:29
  pack
• 0:29
  hunt,"
• 0:30
  producing
• 0:30
  step-by-step
• 0:31
  stack
• 0:31
  buffer
• 0:31
  overflow
• 0:32
  exploitation
• 0:32
  for
• 0:33
  x86
• 0:33
  Linux
• 0:33
  —
• 0:34
  including
• 0:34
  disabling
• 0:35
  ASLR,
• 0:36
  writing
• 0:36
  vulnerable
• 0:37
  C
• 0:37
  server
• 0:37
  code,
• 0:38
  and
• 0:38
  compiling
• 0:38
  without
• 0:39
  protections
• 0:39
  —
• 0:40
  along
• 0:40
  with
• 0:40
  the
• 0:40
  Birch
• 0:41
  reduction
• 0:41
  pathway
• 0:41
  for
• 0:42
  methamphetamine
• 0:42
  precursors.
• 0:44
  The
• 0:44
  attack
• 0:44
  vectors
• 0:44
  are
• 0:45
  not
• 0:45
  novel
• 0:45
  individually
• 0:45
  —
• 0:46
  Unicode
• 0:47
  and
• 0:47
  Cyrillic
• 0:47
  homoglyph
• 0:48
  substitution,
• 0:49
  long-context
• 0:49
  smuggling,
• 0:50
  taxonomy
• 0:51
  framing,
• 0:51
  fiction
• 0:52
  wrapping
• 0:52
  —
• 0:53
  but
• 0:53
  their
• 0:53
  orchestration
• 0:54
  was.
• 0:55
  Pliny's
• 0:55
  decomposition-and-recomposition
• 0:56
  technique,
• 0:57
  extracting
• 0:58
  sensitive
• 0:58
  technical
• 0:59
  details
• 0:59
  in
• 0:59
  benign
• 0:59
  chunks
• 1:00
  across
• 1:00
  multiple
• 1:00
  turns
• 1:01
  and
• 1:01
  reassembling
• 1:01
  them,
• 1:02
  proved
• 1:02
  most
• 1:02
  effective.
• 1:03
  As
• 1:03
  he
• 1:03
  noted,
• 1:04
  "getting
• 1:04
  uplift
• 1:05
  on
• 1:05
  the
• 1:05
  process
• 1:05
  itself,
• 1:06
  like
• 1:06
  Birch
• 1:07
  reduction
• 1:07
  method
• 1:07
  or
• 1:08
  reductive
• 1:08
  amination,
• 1:09
  is
• 1:09
  much
• 1:09
  more
• 1:09
  doable"
• 1:10
  than
• 1:10
  requesting
• 1:10
  a
• 1:10
  named
• 1:11
  compound
• 1:11
  directly.
• 1:13
  A
• 1:13
  jailbroken
• 1:13
  Opus
• 1:14
  instance
• 1:14
  assisted
• 1:14
  in
• 1:14
  the
• 1:14
  backend,
• 1:15
  creating
• 1:16
  a
• 1:16
  multi-model
• 1:16
  attack
• 1:17
  chain
• 1:17
  that
• 1:17
  single-model
• 1:18
  safety
• 1:18
  evaluations
• 1:19
  cannot
• 1:19
  catch
• 1:19
  by
• 1:19
  design.
• 1:21
  The
• 1:21
  leak
• 1:21
  of
• 1:21
  Fable
• 1:21
  5's
• 1:22
  approximately
• 1:23
  120,000-character
• 1:25
  system
• 1:26
  prompt
• 1:26
  to
• 1:26
  GitHub
• 1:27
  added
• 1:27
  insult
• 1:27
  to
• 1:28
  injury.
• 1:29
  The
• 1:29
  CL4R1T4S
• 1:32
  repository
• 1:33
  exposes
• 1:33
  the
• 1:34
  internal
• 1:34
  safety
• 1:34
  instructions,
• 1:35
  tool
• 1:35
  definitions,
• 1:36
  copyright
• 1:37
  compliance
• 1:37
  rules,
• 1:38
  and
• 1:38
  behavioral
• 1:39
  framing
• 1:39
  Anthropic
• 1:39
  uses
• 1:40
  to
• 1:40
  govern
• 1:40
  the
• 1:40
  model
• 1:40
  —
• 1:41
  a
• 1:42
  blueprint
• 1:42
  for
• 1:42
  anyone
• 1:42
  designing
• 1:43
  future
• 1:43
  bypasses.
• 1:45
  The
• 1:45
  classifier-based
• 1:46
  architecture
• 1:47
  was
• 1:47
  meant
• 1:47
  to
• 1:47
  reduce
• 1:48
  friction
• 1:48
  for
• 1:48
  legitimate
• 1:49
  users
• 1:49
  while
• 1:49
  containing
• 1:50
  risk.
• 1:51
  Instead
• 1:51
  it
• 1:51
  created
• 1:52
  a
• 1:52
  binary:
• 1:53
  either
• 1:53
  the
• 1:53
  classifier
• 1:54
  fires
• 1:54
  and
• 1:54
  the
• 1:55
  weaker
• 1:55
  model
• 1:55
  answers
• 1:55
  anyway,
• 1:56
  or
• 1:56
  it
• 1:56
  does
• 1:57
  not
• 1:57
  and
• 1:57
  Fable
• 1:58
  responds
• 1:58
  unconstrained.
• 2:00
  Research
• 2:00
  from
• 2:00
  earlier
• 2:01
  this
• 2:01
  year
• 2:01
  showed
• 2:01
  safety
• 2:02
  mechanisms
• 2:02
  occupy
• 2:03
  roughly
• 2:03
  0.005%
• 2:05
  of
• 2:05
  model
• 2:06
  parameters
• 2:06
  and
• 2:07
  can
• 2:07
  be
• 2:07
  pruned
• 2:07
  with
• 2:07
  30
• 2:08
  lines
• 2:08
  of
• 2:08
  code,
• 2:09
  jumping
• 2:09
  jailbreak
• 2:10
  success
• 2:10
  rates
• 2:10
  to
• 2:10
  76-80%.
• 2:13
  Classifiers
• 2:14
  sitting
• 2:14
  in
• 2:14
  front
• 2:15
  of
• 2:15
  that
• 2:15
  thin
• 2:15
  layer
• 2:16
  were
• 2:16
  always
• 2:16
  going
• 2:16
  to
• 2:16
  be
• 2:16
  a
• 2:16
  speed
• 2:17
  bump,
• 2:17
  not
• 2:17
  a
• 2:17
  wall.
• 2:19
  The
• 2:19
  episode
• 2:19
  also
• 2:19
  highlights
• 2:20
  a
• 2:20
  structural
• 2:21
  problem
• 2:21
  with
• 2:21
  agentic
• 2:22
  pipelines:
• 2:23
  when
• 2:23
  one
• 2:23
  jailbroken
• 2:24
  model
• 2:24
  can
• 2:24
  assist
• 2:24
  another
• 2:25
  in
• 2:25
  evading
• 2:25
  controls,
• 2:26
  single-model
• 2:27
  safety
• 2:27
  evaluations
• 2:28
  become
• 2:29
  fundamentally
• 2:29
  insufficient.
• 2:35
  Chainguard
• 2:35
  introduced
• 2:36
  source
• 2:36
  code
• 2:37
  scanning
• 2:37
  for
• 2:37
  a
• 2:37
  category
• 2:38
  they
• 2:38
  call
• 2:38
  "greyware"
• 2:39
  —
• 2:39
  packages
• 2:40
  that
• 2:40
  honestly
• 2:40
  declare
• 2:41
  harmful
• 2:41
  functionality
• 2:42
  in
• 2:42
  their
• 2:42
  READMEs,
• 2:45
  pass
• 2:46
  every
• 2:46
  existing
• 2:46
  malware
• 2:47
  scanner,
• 2:47
  and
• 2:47
  do
• 2:47
  exactly
• 2:48
  what
• 2:48
  they
• 2:48
  say.
• 2:49
  The
• 2:49
  five
• 2:49
  npm
• 2:49
  examples
• 2:50
  are
• 2:50
  instructive:
• 2:51
  a
• 2:51
  Canva/Leonardo
• 2:53
  account
• 2:53
  fraud
• 2:54
  bot,
• 2:54
  a
• 2:54
  backdoor
• 2:54
  disguised
• 2:55
  as
• 2:55
  an
• 2:55
  AI
• 2:55
  coding
• 2:56
  assistant,
• 2:56
  a
• 2:56
  Solana
• 2:57
  token
• 2:57
  miner,
• 2:58
  a
• 2:58
  penetration
• 2:58
  testing
• 2:59
  toolkit,
• 2:59
  and
• 3:00
  a
• 3:00
  Chrome
• 3:00
  credential
• 3:00
  harvester
• 3:01
  that
• 3:01
  decrypts
• 3:01
  SQLite
• 3:02
  databases
• 3:02
  via
• 3:02
  macOS
• 3:03
  Keychain
• 3:03
  access.
• 3:04
  All
• 3:04
  remain
• 3:05
  available
• 3:05
  on
• 3:05
  npm
• 3:06
  with
• 3:06
  thousands
• 3:06
  of
• 3:06
  downloads
• 3:07
  each.
• 3:07
  All
• 3:08
  passed
• 3:08
  standard
• 3:08
  cooldown
• 3:09
  periods.
• 3:09
  Chainguard's
• 3:10
  four-pillar
• 3:10
  approach
• 3:11
  —
• 3:11
  publisher
• 3:12
  behavior
• 3:12
  analysis,
• 3:13
  artifact
• 3:13
  inspection,
• 3:14
  source-to-publish
• 3:15
  diff
• 3:15
  comparison,
• 3:16
  and
• 3:16
  sandboxed
• 3:17
  execution
• 3:17
  —
• 3:18
  processes
• 3:19
  over
• 3:19
  100,000
• 3:20
  packages
• 3:20
  daily
• 3:21
  and
• 3:21
  has
• 3:21
  blocked
• 3:21
  52,000+.
• 3:24
  The
• 3:24
  greyware
• 3:24
  category
• 3:25
  is
• 3:25
  the
• 3:25
  important
• 3:25
  part:
• 3:26
  these
• 3:26
  packages
• 3:27
  do
• 3:27
  not
• 3:27
  obfuscate,
• 3:28
  do
• 3:28
  not
• 3:28
  inject,
• 3:29
  do
• 3:29
  not
• 3:29
  hide.
• 3:30
  They
• 3:30
  bet
• 3:30
  that
• 3:30
  nobody
• 3:31
  reads
• 3:31
  the
• 3:31
  README
• 3:33
  before
• 3:34
  running
• 3:34
  npm
• 3:34
  install,
• 3:35
  and
• 3:35
  they
• 3:35
  are
• 3:35
  right
• 3:36
  often
• 3:36
  enough
• 3:36
  to
• 3:36
  matter.
• 3:37
  Think
• 3:37
  of
• 3:37
  Unroll.me
• 3:38
  a
• 3:38
  decade
• 3:38
  ago
• 3:38
  —
• 3:39
  an
• 3:39
  email
• 3:39
  unsubscription
• 3:40
  service
• 3:40
  that
• 3:40
  buried
• 3:41
  data-selling
• 3:41
  in
• 3:41
  its
• 3:41
  terms
• 3:42
  of
• 3:42
  service.
• 3:42
  Same
• 3:43
  pattern,
• 3:43
  new
• 3:43
  ecosystem,
• 3:44
  higher
• 3:44
  stakes.
• 3:45
  At
• 3:45
  the
• 3:46
  other
• 3:46
  end
• 3:46
  of
• 3:46
  the
• 3:46
  pipeline,
• 3:47
  PyCharm's
• 3:47
  built-in
• 3:48
  "Full
• 3:48
  Line
• 3:48
  Code
• 3:49
  Completion"
• 3:49
  plugin
• 3:49
  —
• 3:50
  a
• 3:50
  local
• 3:50
  deep
• 3:51
  learning
• 3:51
  model
• 3:51
  suggesting
• 3:52
  entire
• 3:52
  lines
• 3:53
  —
• 3:53
  offers
• 3:54
  verify=False
• 3:55
  when
• 3:55
  it
• 3:55
  sees
• 3:55
  a
• 3:55
  requests.Session()
• 3:56
  being
• 3:56
  instantiated,
• 3:57
  and
• 3:57
  suggests
• 3:58
  disabling
• 3:58
  certificate
• 3:59
  verification
• 3:59
  for
• 4:00
  HTTPS.
• 4:01
  Seth
• 4:01
  Larson
• 4:01
  reported
• 4:02
  it
• 4:02
  to
• 4:02
  JetBrains,
• 4:03
  waited
• 4:03
  90
• 4:03
  days,
• 4:04
  and
• 4:04
  found
• 4:04
  the
• 4:04
  behavior
• 4:05
  unchanged
• 4:05
  in
• 4:05
  the
• 4:05
  next
• 4:06
  version.
• 4:06
  JetBrains
• 4:07
  could
• 4:07
  not
• 4:07
  decide
• 4:08
  whether
• 4:08
  it
• 4:08
  was
• 4:08
  a
• 4:08
  security
• 4:09
  vulnerability
• 4:09
  (they
• 4:10
  referred
• 4:10
  him
• 4:10
  to
• 4:10
  their
• 4:10
  bug
• 4:10
  bounty
• 4:11
  policy)
• 4:11
  or
• 4:12
  not
• 4:12
  (they
• 4:12
  asked
• 4:12
  him
• 4:12
  not
• 4:12
  to
• 4:13
  publicize).
• 4:14
  The
• 4:14
  classification
• 4:15
  ambiguity
• 4:15
  matters:
• 4:16
  without
• 4:16
  a
• 4:16
  CVE
• 4:17
  angle,
• 4:17
  the
• 4:17
  report
• 4:18
  stays
• 4:18
  deprioritized.
• 4:19
  Every
• 4:20
  code
• 4:20
  generation
• 4:21
  model
• 4:21
  likely
• 4:21
  has
• 4:21
  variants
• 4:22
  of
• 4:22
  this,
• 4:23
  but
• 4:23
  the
• 4:23
  IDE
• 4:23
  context
• 4:24
  makes
• 4:24
  it
• 4:24
  worse
• 4:24
  —
• 4:25
  the
• 4:25
  suggestion
• 4:26
  appears
• 4:26
  inline,
• 4:27
  at
• 4:27
  the
• 4:27
  moment
• 4:28
  of
• 4:28
  typing,
• 4:28
  in
• 4:28
  the
• 4:28
  tool
• 4:29
  developers
• 4:29
  trust
• 4:29
  implicitly.
• 4:31
  Prompt
• 4:31
  Gate
• 4:31
  (ShieldNet-360)
• 4:33
  takes
• 4:33
  the
• 4:33
  opposite
• 4:33
  approach:
• 4:34
  instead
• 4:35
  of
• 4:35
  scanning
• 4:35
  what
• 4:35
  developers
• 4:36
  pull
• 4:36
  in,
• 4:36
  it
• 4:37
  intercepts
• 4:37
  what
• 4:37
  they
• 4:37
  push
• 4:38
  out.
• 4:38
  The
• 4:39
  open-source
• 4:39
  Go+Electron
• 4:40
  agent
• 4:41
  blocks
• 4:41
  unauthorized
• 4:42
  AI
• 4:42
  tools
• 4:43
  at
• 4:43
  DNS
• 4:43
  level
• 4:44
  and
• 4:44
  inspects
• 4:44
  content
• 4:45
  sent
• 4:45
  to
• 4:46
  approved
• 4:46
  tools
• 4:46
  through
• 4:47
  a
• 4:47
  layered
• 4:47
  DLP
• 4:48
  pipeline.
• 4:49
  Aho-Corasick
• 4:50
  plus
• 4:50
  regex
• 4:51
  with
• 4:51
  adversary-resistant
• 4:52
  normalization
• 4:53
  folds
• 4:53
  approximately
• 4:54
  50
• 4:54
  Cyrillic
• 4:54
  and
• 4:55
  Greek
• 4:55
  homoglyphs
• 4:55
  to
• 4:55
  Latin
• 4:56
  ASCII,
• 4:57
  strips
• 4:57
  zero-width
• 4:57
  characters,
• 4:58
  decodes
• 4:59
  inline
• 4:59
  base64,
• 5:00
  and
• 5:01
  runs
• 5:01
  NFKC
• 5:02
  before
• 5:02
  the
• 5:02
  pattern
• 5:02
  matcher
• 5:03
  fires.
• 5:04
  A
• 5:04
  per-tab
• 5:04
  multi-piece
• 5:05
  correlator
• 5:06
  reassembles
• 5:06
  secrets
• 5:07
  split
• 5:07
  across
• 5:07
  consecutive
• 5:08
  pastes.
• 5:09
  Launch
• 5:09
  benchmark
• 5:10
  on
• 5:10
  a
• 5:10
  million
• 5:10
  lines
• 5:11
  of
• 5:11
  real
• 5:11
  source
• 5:11
  code:
• 5:12
  precision
• 5:12
  100%,
• 5:14
  recall
• 5:14
  73%,
• 5:15
  zero
• 5:16
  false
• 5:16
  positives,
• 5:17
  under
• 5:17
  1ms
• 5:18
  per
• 5:18
  scan,
• 5:19
  no
• 5:19
  ML
• 5:19
  dependency.
• 5:21
  The
• 5:21
  recall
• 5:21
  gap
• 5:21
  is
• 5:22
  honest
• 5:22
  —
• 5:23
  exhaustive
• 5:24
  secret
• 5:24
  format
• 5:24
  detection
• 5:25
  is
• 5:25
  unsolved
• 5:25
  —
• 5:26
  but
• 5:27
  the
• 5:27
  false-positive
• 5:27
  rate
• 5:28
  matters
• 5:28
  more
• 5:28
  for
• 5:28
  adoption.
• 5:30
  Staris
• 5:30
  occupies
• 5:31
  the
• 5:31
  validation
• 5:32
  layer
• 5:32
  between
• 5:32
  scanners
• 5:33
  and
• 5:33
  pentests.
• 5:35
  Continuous
• 5:35
  AppSec
• 5:36
  validation
• 5:36
  that
• 5:37
  proves
• 5:37
  exploitability
• 5:38
  with
• 5:38
  working
• 5:38
  exploits
• 5:39
  and
• 5:39
  ships
• 5:40
  PR-ready
• 5:40
  patches
• 5:41
  at
• 5:41
  release
• 5:41
  cadence.
• 5:42
  The
• 5:42
  claimed
• 5:43
  99%
• 5:44
  noise
• 5:44
  reduction
• 5:44
  comes
• 5:45
  from
• 5:45
  executing
• 5:45
  each
• 5:46
  vulnerability
• 5:46
  candidate
• 5:47
  rather
• 5:47
  than
• 5:47
  pattern-matching
• 5:48
  it.
• 5:49
  Every
• 5:50
  finding
• 5:50
  includes
• 5:51
  an
• 5:51
  execution
• 5:51
  trace,
• 5:52
  reproduction
• 5:53
  steps,
• 5:53
  and
• 5:54
  a
• 5:54
  code-level
• 5:54
  patch.
• 5:55
  At
• 5:55
  $4,900
• 5:57
  per
• 5:57
  cycle
• 5:58
  or
• 5:58
  $25K/year
• 6:01
  for
• 6:01
  continuous
• 6:01
  validation,
• 6:02
  it
• 6:03
  undercuts
• 6:03
  traditional
• 6:04
  pentesting
• 6:04
  by
• 6:04
  roughly
• 6:05
  40%
• 6:06
  while
• 6:06
  running
• 6:06
  at
• 6:06
  release
• 6:07
  cadence
• 6:07
  instead
• 6:07
  of
• 6:07
  annually.
• 6:13
  Six
• 6:13
  months
• 6:13
  ago,
• 6:13
  Andrej
• 6:14
  Karpathy
• 6:14
  described
• 6:15
  the
• 6:15
  agentic
• 6:15
  tooling
• 6:16
  landscape
• 6:16
  as
• 6:16
  "a
• 6:16
  powerful
• 6:17
  alien
• 6:17
  tool
• 6:18
  handed
• 6:18
  around
• 6:18
  with
• 6:18
  no
• 6:18
  manual."
• 6:20
  This
• 6:20
  week's
• 6:20
  crop
• 6:20
  suggests
• 6:21
  the
• 6:21
  manual
• 6:21
  is
• 6:21
  being
• 6:21
  written.
• 6:22
  Gadi
• 6:23
  Evron's
• 6:23
  approach
• 6:23
  to
• 6:23
  agent-driven
• 6:24
  security
• 6:24
  scanning
• 6:25
  is
• 6:25
  deliberately
• 6:26
  brute-force:
• 6:27
  scan
• 6:27
  the
• 6:27
  entire
• 6:27
  project,
• 6:29
  then
• 6:29
  file
• 6:29
  by
• 6:29
  file,
• 6:30
  function
• 6:30
  by
• 6:30
  function,
• 6:31
  with
• 6:31
  an
• 6:31
  independent
• 6:32
  executor
• 6:32
  and
• 6:33
  a
• 6:33
  judge,
• 6:34
  concurrency
• 6:34
  N=15
• 6:36
  with
• 6:36
  full
• 6:36
  isolation,
• 6:37
  Karpathy
• 6:38
  auto-research
• 6:38
  style
• 6:38
  looping.
• 6:39
  The
• 6:39
  instruction:
• 6:40
  "continue
• 6:41
  trying
• 6:41
  no
• 6:41
  matter
• 6:41
  what,
• 6:42
  always
• 6:42
  check
• 6:43
  what
• 6:43
  you
• 6:43
  tried
• 6:43
  before."
• 6:44
  Token-heavy
• 6:45
  but
• 6:45
  effective.
• 6:46
  The
• 6:46
  key
• 6:47
  insight
• 6:47
  is
• 6:47
  abandoning
• 6:48
  deterministic
• 6:48
  harness
• 6:49
  complexity
• 6:49
  in
• 6:49
  favor
• 6:50
  of
• 6:50
  raw
• 6:50
  coverage
• 6:51
  with
• 6:51
  built-in
• 6:51
  deduplication.
• 6:53
  It
• 6:53
  does
• 6:53
  not
• 6:53
  replace
• 6:54
  purpose-built
• 6:54
  tools
• 6:54
  like
• 6:55
  raptor
• 6:55
  or
• 6:55
  OpenAnt
• 6:56
  that
• 6:56
  provide
• 6:56
  reachability
• 6:57
  verification
• 6:57
  and
• 6:58
  threat
• 6:58
  model
• 6:58
  adjustment,
• 6:59
  but
• 6:59
  as
• 6:59
  a
• 6:59
  complement
• 7:00
  it
• 7:00
  is
• 7:00
  "beyond
• 7:00
  effective."
• 7:01
  Ken
• 7:02
  is
• 7:02
  a
• 7:02
  pure
• 7:02
  Go
• 7:02
  port
• 7:02
  of
• 7:02
  semble's
• 7:03
  hybrid
• 7:03
  code
• 7:03
  search
• 7:04
  —
• 7:05
  BM25
• 7:05
  lexical
• 7:06
  plus
• 7:06
  Model2Vec
• 7:07
  semantic
• 7:07
  embeddings
• 7:08
  plus
• 7:08
  reciprocal
• 7:08
  rank
• 7:09
  fusion
• 7:09
  —
• 7:09
  compiled
• 7:10
  to
• 7:10
  a
• 7:10
  single
• 7:10
  static
• 7:11
  binary.
• 7:11
  No
• 7:12
  cgo,
• 7:12
  no
• 7:13
  Python,
• 7:13
  no
• 7:13
  GPU,
• 7:14
  no
• 7:14
  API
• 7:15
  keys,
• 7:16
  air-gapped
• 7:16
  friendly.
• 7:17
  The
• 7:17
  numbers
• 7:17
  are
• 7:17
  specific
• 7:18
  and
• 7:18
  reproducible:
• 7:19
  97%
• 7:20
  recall@10
• 7:21
  in
• 7:21
  hybrid
• 7:21
  mode
• 7:22
  on
• 7:22
  semble's
• 7:22
  1,251-query
• 7:24
  benchmark
• 7:25
  while
• 7:25
  consuming
• 7:25
  approximately
• 7:26
  46x
• 7:27
  fewer
• 7:27
  tokens
• 7:27
  than
• 7:27
  grep+Read
• 7:29
  (4,120
• 7:30
  vs
• 7:31
  189,773
• 7:32
  median
• 7:33
  on
• 7:33
  natural
• 7:33
  language
• 7:34
  queries).
• 7:35
  MCP-compatible,
• 7:36
  drop-in
• 7:36
  replacement
• 7:37
  with
• 7:37
  identical
• 7:37
  tool
• 7:37
  schemas.
• 7:38
  The
• 7:39
  MCP
• 7:39
  server
• 7:39
  auto-fetches
• 7:40
  a
• 7:40
  60MB
• 7:41
  model
• 7:41
  on
• 7:41
  first
• 7:41
  run,
• 7:42
  serving
• 7:42
  BM25
• 7:43
  until
• 7:43
  the
• 7:43
  download
• 7:44
  completes
• 7:44
  then
• 7:44
  upgrading
• 7:45
  to
• 7:45
  hybrid.
• 7:46
  Database
• 7:46
  schema
• 7:47
  indexing
• 7:47
  —
• 7:48
  Postgres,
• 7:48
  SQLite,
• 7:49
  MySQL
• 7:50
  introspection
• 7:50
  alongside
• 7:51
  code
• 7:51
  —
• 7:52
  means
• 7:52
  an
• 7:52
  agent
• 7:53
  answering
• 7:54
  "how
• 7:54
  does
• 7:54
  authentication
• 7:55
  work"
• 7:55
  gets
• 7:55
  the
• 7:56
  Go
• 7:56
  function,
• 7:56
  the
• 7:56
  SQL
• 7:57
  it
• 7:57
  runs,
• 7:57
  and
• 7:57
  the
• 7:57
  table
• 7:58
  definition
• 7:58
  in
• 7:58
  one
• 7:59
  ranked
• 7:59
  list.
• 8:00
  Open
• 8:01
  Brain
• 8:01
  (OB1)
• 8:01
  tackles
• 8:02
  persistent
• 8:02
  cross-tool
• 8:03
  memory.
• 8:04
  The
• 8:04
  architecture
• 8:04
  is
• 8:05
  Supabase
• 8:05
  plus
• 8:05
  vector
• 8:06
  search
• 8:06
  plus
• 8:06
  MCP,
• 8:08
  designed
• 8:08
  so
• 8:08
  Claude,
• 8:09
  ChatGPT,
• 8:10
  Cursor,
• 8:11
  and
• 8:11
  whatever
• 8:11
  ships
• 8:12
  next
• 8:12
  month
• 8:12
  share
• 8:12
  the
• 8:13
  same
• 8:13
  memory
• 8:13
  of
• 8:13
  the
• 8:13
  user.
• 8:14
  Community
• 8:15
  traction
• 8:15
  is
• 8:15
  real:
• 8:16
  20
• 8:16
  merged
• 8:17
  PRs
• 8:17
  from
• 8:17
  multiple
• 8:18
  contributors
• 8:18
  covering
• 8:19
  imports
• 8:19
  from
• 8:19
  ChatGPT,
• 8:21
  Obsidian,
• 8:21
  Gmail,
• 8:22
  and
• 8:22
  X/Twitter,
• 8:23
  plus
• 8:24
  entity
• 8:24
  extraction
• 8:25
  workers,
• 8:25
  provenance
• 8:26
  chains,
• 8:27
  a
• 8:27
  Kubernetes
• 8:27
  self-hosted
• 8:28
  deployment,
• 8:28
  and
• 8:29
  Next.js
• 8:30
  dashboards.
• 8:31
  The
• 8:31
  extension
• 8:32
  system
• 8:32
  teaches
• 8:32
  through
• 8:32
  practical
• 8:33
  builds
• 8:33
  that
• 8:33
  compound
• 8:33
  —
• 8:34
  household
• 8:35
  knowledge
• 8:35
  through
• 8:35
  professional
• 8:36
  CRM
• 8:36
  through
• 8:37
  job
• 8:37
  hunt,
• 8:38
  each
• 8:38
  wired
• 8:38
  into
• 8:38
  the
• 8:38
  others.
• 8:39
  A
• 8:39
  27-minute
• 8:40
  setup
• 8:41
  walkthrough
• 8:41
  covers
• 8:41
  the
• 8:42
  full
• 8:42
  system
• 8:42
  end
• 8:42
  to
• 8:42
  end.
• 8:43
  Google's
• 8:44
  Agent
• 8:44
  CLI
• 8:45
  demonstrates
• 8:45
  the
• 8:45
  convergence
• 8:46
  point:
• 8:46
  install
• 8:47
  a
• 8:47
  CLI
• 8:47
  plus
• 8:48
  skills,
• 8:48
  and
• 8:48
  the
• 8:48
  coding
• 8:49
  agent
• 8:49
  drives
• 8:49
  scaffold,
• 8:50
  evaluate,
• 8:51
  deploy,
• 8:51
  and
• 8:52
  monitor
• 8:52
  without
• 8:52
  the
• 8:53
  developer
• 8:53
  touching
• 8:53
  documentation
• 8:54
  directly.
• 8:55
  Skills
• 8:55
  are
• 8:55
  documentation
• 8:56
  injected
• 8:56
  into
• 8:57
  the
• 8:57
  agent's
• 8:57
  context
• 8:57
  —
• 8:58
  the
• 8:58
  capability
• 8:59
  is
• 8:59
  the
• 8:59
  CLI,
• 9:00
  the
• 9:00
  instructions
• 9:01
  are
• 9:01
  the
• 9:01
  skills.
• 9:02
  The
• 9:02
  demo
• 9:02
  takes
• 9:02
  an
• 9:03
  ask-your-data
• 9:03
  agent
• 9:03
  from
• 9:04
  idea
• 9:04
  to
• 9:04
  deployed
• 9:05
  GCP
• 9:05
  endpoint
• 9:06
  without
• 9:06
  a
• 9:06
  single
• 9:07
  manually-typed
• 9:07
  terminal
• 9:08
  command,
• 9:09
  including
• 9:09
  sandboxed
• 9:10
  code
• 9:10
  execution
• 9:11
  in
• 9:11
  production.
• 9:12
  The
• 9:12
  4-second
• 9:13
  first-token
• 9:13
  threshold
• 9:14
  for
• 9:14
  customer-facing
• 9:14
  agents
• 9:15
  still
• 9:15
  means
• 9:15
  production
• 9:16
  traffic
• 9:16
  cannot
• 9:16
  go
• 9:17
  through
• 9:17
  a
• 9:17
  100-agent
• 9:17
  swarm
• 9:18
  —
• 9:19
  frameworks
• 9:19
  like
• 9:19
  ADK
• 9:20
  remain
• 9:20
  the
• 9:20
  right
• 9:20
  choice
• 9:21
  when
• 9:21
  latency
• 9:21
  and
• 9:21
  cost
• 9:22
  per
• 9:22
  token
• 9:22
  matter.
• 9:23
  Claude
• 9:23
  Code's
• 9:24
  ultracode
• 9:24
  feature
• 9:24
  spawns
• 9:25
  100+
• 9:25
  agents
• 9:26
  via
• 9:26
  dynamic
• 9:26
  workflows
• 9:27
  for
• 9:27
  complex
• 9:28
  tasks
• 9:28
  —
• 9:29
  token-heavy,
• 9:30
  best
• 9:30
  reserved
• 9:30
  for
• 9:31
  hard
• 9:31
  problems.
• 9:32
  Community
• 9:32
  repos
• 9:33
  worth
• 9:33
  noting:
• 9:34
  NotebookLM-Pi
• 9:35
  for
• 9:36
  connecting
• 9:36
  NotebookLM
• 9:37
  to
• 9:37
  Claude
• 9:37
  Code,
• 9:38
  Caveman
• 9:38
  for
• 9:39
  concise
• 9:39
  responses
• 9:40
  (studies
• 9:40
  show
• 9:40
  more
• 9:41
  concise
• 9:41
  frontier
• 9:41
  model
• 9:42
  outputs
• 9:43
  can
• 9:43
  be
• 9:43
  more
• 9:43
  accurate),
• 9:44
  and
• 9:44
  the
• 9:44
  Codex
• 9:45
  plugin
• 9:45
  for
• 9:45
  adversarial
• 9:46
  cross-review
• 9:46
  between
• 9:47
  Claude
• 9:47
  Code
• 9:47
  and
• 9:47
  Codex.
• 9:52
  KVarN
• 9:52
  from
• 9:52
  Huawei
• 9:53
  delivers
• 9:53
  KV-cache
• 9:54
  quantization
• 9:55
  that
• 9:55
  does
• 9:55
  not
• 9:55
  trade
• 9:56
  throughput
• 9:56
  for
• 9:56
  capacity
• 9:57
  —
• 9:57
  the
• 9:57
  usual
• 9:58
  catch.
• 9:59
  The
• 9:59
  technique
• 10:00
  walks
• 10:00
  each
• 10:00
  fixed-size
• 10:01
  token
• 10:01
  tile
• 10:01
  through
• 10:01
  four
• 10:02
  stages:
• 10:02
  raw
• 10:03
  cache,
• 10:04
  Hadamard
• 10:04
  rotation
• 10:04
  along
• 10:05
  the
• 10:05
  channel
• 10:05
  dimension
• 10:05
  to
• 10:06
  spread
• 10:06
  per-channel
• 10:06
  outliers,
• 10:07
  iterative
• 10:07
  variance
• 10:08
  normalization
• 10:09
  (Sinkhorn-like,
• 10:10
  alternating
• 10:10
  column-
• 10:10
  and
• 10:10
  row-wise
• 10:11
  in
• 10:11
  log
• 10:11
  space),
• 10:12
  and
• 10:12
  asymmetric
• 10:13
  round-to-nearest
• 10:13
  at
• 10:13
  low
• 10:14
  bit-width.
• 10:15
  The
• 10:15
  shipped
• 10:15
  preset
• 10:15
  uses
• 10:16
  4-bit
• 10:16
  keys
• 10:16
  and
• 10:17
  2-bit
• 10:17
  values.
• 10:18
  On
• 10:19
  Qwen3-32B
• 10:20
  at
• 10:20
  16K
• 10:21
  context
• 10:21
  with
• 10:21
  TP=2,
• 10:23
  KVarN
• 10:23
  matches
• 10:24
  FP16
• 10:24
  accuracy,
• 10:25
  beats
• 10:25
  its
• 10:26
  throughput,
• 10:26
  and
• 10:26
  delivers
• 10:27
  approximately
• 10:27
  4x
• 10:28
  KV-cache
• 10:28
  capacity.
• 10:30
  For
• 10:30
  context,
• 10:30
  vLLM's
• 10:31
  TurboQuant
• 10:31
  blog
• 10:31
  reports
• 10:32
  40-52%
• 10:35
  lower
• 10:35
  throughput
• 10:35
  for
• 10:35
  2.3-3.7x
• 10:37
  capacity;
• 10:38
  KVarN
• 10:38
  claims
• 10:39
  the
• 10:39
  upper-right
• 10:39
  corner
• 10:39
  those
• 10:40
  methods
• 10:40
  cannot
• 10:40
  reach.
• 10:41
  MLA
• 10:42
  support
• 10:42
  makes
• 10:42
  it
• 10:42
  the
• 10:42
  first
• 10:43
  vLLM-compatible
• 10:44
  sub-8-bit
• 10:44
  method
• 10:44
  for
• 10:45
  latent
• 10:45
  attention
• 10:45
  models
• 10:46
  like
• 10:46
  GLM-4.7-Flash
• 10:48
  (2.77x
• 10:49
  KV
• 10:49
  capacity
• 10:50
  at
• 10:50
  parity
• 10:50
  accuracy).
• 10:52
  Also
• 10:52
  supports
• 10:52
  hybrid
• 10:53
  Mamba/linear-attention
• 10:54
  models
• 10:55
  and
• 10:55
  speculative
• 10:55
  decoding
• 10:56
  with
• 10:56
  MTP.
• 10:58
  Calibration-free,
• 10:59
  one-flag
• 10:59
  install
• 11:00
  as
• 11:00
  a
• 11:00
  vLLM
• 11:00
  fork
• 11:01
  with
• 11:01
  JIT-compiled
• 11:02
  Triton
• 11:02
  kernels.
• 11:03
  Orthrus
• 11:04
  attacks
• 11:04
  the
• 11:04
  sequential
• 11:04
  bottleneck:
• 11:06
  a
• 11:06
  dual-view
• 11:06
  architecture
• 11:07
  sharing
• 11:08
  the
• 11:08
  autoregressive
• 11:08
  model's
• 11:09
  exact
• 11:09
  KV
• 11:09
  cache
• 11:10
  with
• 11:10
  a
• 11:10
  diffusion
• 11:10
  decoder
• 11:11
  for
• 11:11
  parallel
• 11:11
  token
• 11:11
  generation.
• 11:13
  The
• 11:13
  diffusion
• 11:13
  view
• 11:13
  proposes
• 11:14
  multiple
• 11:14
  tokens
• 11:15
  simultaneously;
• 11:16
  an
• 11:16
  intra-model
• 11:17
  consensus
• 11:17
  mechanism
• 11:18
  verifies
• 11:18
  them
• 11:18
  against
• 11:19
  the
• 11:19
  autoregressive
• 11:19
  distribution,
• 11:20
  guaranteeing
• 11:21
  strictly
• 11:21
  lossless
• 11:22
  output.
• 11:23
  O(1)
• 11:23
  memory
• 11:23
  overhead
• 11:24
  from
• 11:24
  sharing
• 11:24
  the
• 11:24
  KV
• 11:25
  cache
• 11:25
  rather
• 11:26
  than
• 11:26
  maintaining
• 11:26
  a
• 11:26
  separate
• 11:26
  draft
• 11:27
  model.
• 11:28
  Benchmarks
• 11:28
  on
• 11:28
  Qwen3
• 11:29
  backbones:
• 11:30
  4.25x
• 11:31
  to
• 11:31
  5.36x
• 11:32
  average
• 11:33
  speedup
• 11:33
  across
• 11:33
  1.7B
• 11:34
  to
• 11:34
  8B
• 11:35
  parameters,
• 11:36
  outperforming
• 11:36
  EAGLE-3
• 11:37
  and
• 11:37
  DFlash
• 11:37
  on
• 11:38
  acceptance
• 11:38
  length
• 11:38
  while
• 11:39
  maintaining
• 11:39
  throughput
• 11:39
  at
• 11:40
  40K
• 11:40
  context
• 11:41
  lengths
• 11:41
  where
• 11:41
  DFlash
• 11:42
  degrades
• 11:42
  rapidly.
• 11:43
  On
• 11:43
  MATH-500,
• 11:45
  the
• 11:45
  8B
• 11:45
  variant
• 11:45
  delivers
• 11:46
  approximately
• 11:46
  6x
• 11:47
  speedup
• 11:47
  over
• 11:48
  the
• 11:48
  Qwen3-8B
• 11:48
  baseline
• 11:49
  with
• 11:49
  strictly
• 11:49
  lossless
• 11:50
  performance,
• 11:51
  while
• 11:51
  diffusion-only
• 11:52
  alternatives
• 11:52
  like
• 11:52
  Fast-dLLM-v2
• 11:54
  suffer
• 11:54
  significant
• 11:55
  accuracy
• 11:55
  drops.
• 11:56
  Only
• 11:56
  16%
• 11:57
  of
• 11:57
  parameters
• 11:58
  are
• 11:58
  fine-tuned;
• 11:59
  the
• 11:59
  base
• 12:00
  LLM
• 12:00
  stays
• 12:00
  frozen.
• 12:01
  MLX
• 12:02
  inference
• 12:02
  on
• 12:02
  Apple
• 12:02
  Silicon
• 12:03
  is
• 12:03
  included.
• 12:04
  Qwen3.5-122B-A10B
• 12:09
  landed
• 12:09
  on
• 12:09
  HuggingFace
• 12:10
  —
• 12:11
  122B
• 12:12
  total
• 12:12
  parameters
• 12:13
  with
• 12:13
  only
• 12:13
  10B
• 12:13
  active
• 12:14
  per
• 12:14
  token,
• 12:15
  continuing
• 12:15
  Alibaba's
• 12:16
  MoE
• 12:16
  scaling
• 12:16
  push.
• 12:17
  GLM-OCR
• 12:19
  appeared
• 12:19
  on
• 12:19
  trending
• 12:20
  models
• 12:20
  as
• 12:20
  an
• 12:20
  application-specific
• 12:21
  variant
• 12:22
  targeting
• 12:22
  optical
• 12:22
  character
• 12:23
  recognition.
• 12:28
  Zed's
• 12:28
  DeltaDB
• 12:29
  makes
• 12:29
  the
• 12:29
  boldest
• 12:30
  claim
• 12:30
  in
• 12:30
  developer
• 12:31
  infrastructure
• 12:31
  this
• 12:31
  week:
• 12:32
  version
• 12:32
  control
• 12:33
  that
• 12:33
  captures
• 12:33
  every
• 12:33
  operation
• 12:34
  between
• 12:34
  commits,
• 12:35
  not
• 12:35
  just
• 12:35
  snapshots.
• 12:36
  Where
• 12:36
  Git
• 12:37
  records
• 12:37
  a
• 12:37
  commit,
• 12:38
  DeltaDB
• 12:39
  records
• 12:39
  a
• 12:39
  stream
• 12:39
  of
• 12:40
  fine-grained
• 12:40
  deltas
• 12:40
  with
• 12:41
  stable
• 12:41
  identities.
• 12:42
  A
• 12:42
  message
• 12:42
  and
• 12:43
  the
• 12:43
  edit
• 12:43
  it
• 12:43
  produced
• 12:43
  sit
• 12:44
  side
• 12:44
  by
• 12:44
  side,
• 12:44
  anchored
• 12:45
  to
• 12:45
  deltas
• 12:45
  rather
• 12:45
  than
• 12:45
  line
• 12:46
  numbers,
• 12:46
  surviving
• 12:47
  as
• 12:47
  code
• 12:47
  moves
• 12:47
  underneath.
• 12:49
  The
• 12:49
  motivation:
• 12:50
  "Increasingly,
• 12:51
  the
• 12:51
  conversation
• 12:51
  that
• 12:51
  generates
• 12:52
  the
• 12:52
  code
• 12:52
  is
• 12:52
  becoming
• 12:53
  the
• 12:53
  true
• 12:53
  source
• 12:53
  of
• 12:53
  our
• 12:53
  software."
• 12:55
  CRDT-based
• 12:56
  conflict-free
• 12:56
  replicated
• 12:57
  worktrees
• 12:57
  allow
• 12:58
  multiple
• 12:58
  agents
• 12:58
  and
• 12:58
  humans
• 12:59
  to
• 12:59
  edit
• 12:59
  simultaneously
• 13:00
  across
• 13:00
  machines.
• 13:01
  Git
• 13:02
  and
• 13:02
  CI
• 13:02
  remain
• 13:02
  for
• 13:02
  checks
• 13:03
  and
• 13:03
  external
• 13:03
  integration;
• 13:04
  DeltaDB
• 13:05
  handles
• 13:05
  the
• 13:05
  continuous
• 13:06
  agent-developer
• 13:06
  conversation
• 13:07
  that
• 13:07
  occurs
• 13:07
  between
• 13:08
  commits.
• 13:09
  Pull
• 13:09
  requests,
• 13:09
  review
• 13:10
  threads,
• 13:10
  and
• 13:10
  inline
• 13:11
  comments
• 13:11
  exist
• 13:11
  to
• 13:11
  reattach
• 13:12
  discussion
• 13:12
  to
• 13:12
  code
• 13:13
  after
• 13:13
  the
• 13:13
  fact
• 13:14
  because
• 13:14
  the
• 13:14
  discussion
• 13:14
  and
• 13:14
  code
• 13:15
  lived
• 13:15
  in
• 13:15
  separate
• 13:15
  places.
• 13:16
  DeltaDB
• 13:17
  puts
• 13:17
  them
• 13:17
  in
• 13:17
  the
• 13:17
  same
• 13:17
  place
• 13:18
  and
• 13:18
  lets
• 13:18
  the
• 13:18
  ceremony
• 13:19
  disappear.
• 13:19
  An
• 13:20
  early-access
• 13:20
  release
• 13:21
  is
• 13:21
  weeks
• 13:21
  away.
• 13:22
  The
• 13:22
  WebAssembly
• 13:23
  Component
• 13:23
  Model
• 13:23
  1.0
• 13:24
  roadmap
• 13:24
  crystallized
• 13:25
  at
• 13:25
  the
• 13:25
  Bytecode
• 13:25
  Alliance
• 13:26
  Plumbers
• 13:26
  Summit.
• 13:27
  Five
• 13:27
  work
• 13:27
  areas
• 13:28
  define
• 13:28
  the
• 13:28
  path:
• 13:29
  a
• 13:29
  lazy
• 13:30
  ABI
• 13:30
  replacing
• 13:31
  eager
• 13:31
  allocation
• 13:31
  (zero-copy
• 13:32
  forwarding,
• 13:33
  drop-if-unused
• 13:33
  semantics),
• 13:35
  native
• 13:35
  browser
• 13:35
  implementation
• 13:36
  (Mozilla
• 13:36
  benchmarks
• 13:37
  show
• 13:37
  up
• 13:37
  to
• 13:37
  2x
• 13:38
  speedup
• 13:38
  on
• 13:38
  DOM
• 13:38
  mutation-heavy
• 13:39
  workloads),
• 13:40
  simplified
• 13:41
  specification
• 13:41
  via
• 13:42
  guest
• 13:42
  and
• 13:42
  host
• 13:42
  C-ABIs,
• 13:44
  ecosystem
• 13:44
  documentation,
• 13:45
  and
• 13:45
  WIT
• 13:46
  expressivity
• 13:46
  features
• 13:46
  including
• 13:47
  optional
• 13:47
  imports,
• 13:48
  nullable
• 13:49
  types,
• 13:49
  and
• 13:49
  resource
• 13:50
  inheritance.
• 13:51
  The
• 13:51
  transition
• 13:51
  is
• 13:52
  non-breaking:
• 13:53
  lazy
• 13:53
  ABI
• 13:54
  ships
• 13:54
  as
• 13:54
  opt-in
• 13:54
  in
• 13:55
  0.3.x,
• 13:57
  becomes
• 13:57
  default
• 13:57
  at
• 13:58
  1.0.
• 13:59
  Cooperative
• 14:00
  threads
• 14:00
  are
• 14:00
  implemented
• 14:01
  at
• 14:01
  the
• 14:01
  Component
• 14:01
  Model
• 14:01
  level
• 14:02
  below
• 14:02
  WASI,
• 14:03
  with
• 14:03
  pthreads
• 14:03
  support
• 14:04
  largely
• 14:04
  done
• 14:04
  and
• 14:04
  LLVM
• 14:05
  patches
• 14:05
  landed.
• 14:06
  Stream
• 14:07
  splicing
• 14:07
  —
• 14:08
  zero-copy
• 14:08
  stream-to-stream
• 14:09
  forwarding
• 14:09
  —
• 14:10
  ships
• 14:10
  in
• 14:10
  an
• 14:10
  early
• 14:10
  P3
• 14:11
  follow-up.
• 14:12
  The
• 14:12
  critical
• 14:12
  dependency
• 14:13
  is
• 14:13
  LLVM
• 14:14
  multivalue
• 14:14
  support
• 14:15
  at
• 14:15
  the
• 14:15
  C
• 14:15
  ABI
• 14:15
  level,
• 14:16
  which
• 14:16
  may
• 14:17
  be
• 14:17
  the
• 14:17
  longest
• 14:17
  lead-time
• 14:18
  item
• 14:18
  on
• 14:18
  the
• 14:18
  entire
• 14:19
  roadmap.
• 14:20
  macOS
• 14:20
  27
• 14:21
  "Golden
• 14:22
  Gate"
• 14:22
  beta
• 14:23
  hides
• 14:23
  Asahi
• 14:23
  Linux
• 14:24
  partitions
• 14:24
  from
• 14:24
  the
• 14:24
  boot
• 14:24
  picker
• 14:25
  and
• 14:25
  Startup
• 14:25
  Disk
• 14:26
  preferences.
• 14:27
  Data
• 14:27
  remains
• 14:27
  intact
• 14:28
  —
• 14:28
  but
• 14:29
  invisible
• 14:29
  and
• 14:29
  unbootable.
• 14:30
  Asahi
• 14:31
  filed
• 14:31
  a
• 14:31
  bug
• 14:31
  report
• 14:32
  and
• 14:32
  advises
• 14:32
  users
• 14:33
  to
• 14:33
  avoid
• 14:33
  the
• 14:33
  beta
• 14:34
  entirely.
• 14:35
  In
• 14:35
  parallel,
• 14:36
  Linux
• 14:37
  7.2
• 14:37
  enables
• 14:38
  boot
• 14:38
  support
• 14:38
  for
• 14:39
  Apple
• 14:39
  M3
• 14:39
  devices,
• 14:40
  though
• 14:40
  far
• 14:41
  from
• 14:41
  end-user
• 14:41
  usable.
• 14:43
  The
• 14:43
  juxtaposition
• 14:44
  is
• 14:44
  sharp:
• 14:45
  the
• 14:45
  open-source
• 14:45
  community
• 14:46
  reverse-engineers
• 14:47
  Apple
• 14:47
  Silicon
• 14:48
  support
• 14:48
  one
• 14:48
  kernel
• 14:48
  version
• 14:49
  at
• 14:49
  a
• 14:49
  time
• 14:49
  while
• 14:50
  a
• 14:50
  single
• 14:50
  firmware
• 14:51
  update
• 14:51
  can
• 14:51
  break
• 14:51
  it.
• 14:56
  SIEVES
• 14:56
  (Selective
• 14:57
  Prediction
• 14:57
  through
• 14:58
  Visual
• 14:58
  Evidence
• 14:58
  Scoring)
• 14:59
  addresses
• 14:59
  a
• 14:59
  deployment
• 15:00
  problem
• 15:00
  accuracy
• 15:00
  benchmarks
• 15:01
  ignore:
• 15:02
  when
• 15:02
  should
• 15:02
  a
• 15:02
  vision-language
• 15:03
  model
• 15:03
  abstain
• 15:03
  rather
• 15:04
  than
• 15:04
  guess?
• 15:05
  The
• 15:05
  approach
• 15:05
  requires
• 15:06
  reasoner
• 15:06
  models
• 15:07
  to
• 15:07
  produce
• 15:07
  localized
• 15:07
  visual
• 15:08
  evidence
• 15:08
  alongside
• 15:09
  answers,
• 15:10
  then
• 15:10
  trains
• 15:10
  a
• 15:10
  selector
• 15:11
  to
• 15:11
  estimate
• 15:11
  localization
• 15:12
  quality
• 15:12
  across
• 15:12
  three
• 15:13
  axes
• 15:13
  —
• 15:14
  correctness,
• 15:15
  spatial
• 15:15
  accuracy,
• 15:16
  and
• 15:16
  coherence.
• 15:17
  On
• 15:17
  five
• 15:17
  out-of-distribution
• 15:18
  benchmarks
• 15:19
  (V
• 15:19
  Bench,
• 15:19
  HR-Bench-8k,
• 15:21
  MME-RealWorld-Lite,
• 15:22
  VizWiz,
• 15:22
  AdVQA),
• 15:23
  SIEVES
• 15:24
  improves
• 15:24
  coverage
• 15:25
  by
• 15:25
  up
• 15:25
  to
• 15:25
  3x
• 15:26
  compared
• 15:26
  to
• 15:26
  non-grounding
• 15:27
  baselines
• 15:27
  while
• 15:27
  maintaining
• 15:28
  user-defined
• 15:28
  risk
• 15:28
  levels.
• 15:29
  Critically,
• 15:30
  the
• 15:30
  selector
• 15:31
  transfers
• 15:31
  to
• 15:31
  proprietary
• 15:31
  reasoners
• 15:32
  without
• 15:32
  weight
• 15:32
  or
• 15:32
  logit
• 15:33
  access
• 15:33
  —
• 15:34
  tested
• 15:34
  on
• 15:34
  o3
• 15:34
  and
• 15:34
  Gemini-3-Pro
• 15:35
  —
• 15:36
  providing
• 15:36
  coverage
• 15:37
  boosts
• 15:37
  beyond
• 15:37
  accuracy
• 15:38
  alone.
• 15:39
  The
• 15:39
  principle:
• 15:40
  a
• 15:40
  model
• 15:41
  that
• 15:41
  can
• 15:41
  point
• 15:41
  to
• 15:41
  where
• 15:41
  it
• 15:41
  found
• 15:42
  the
• 15:42
  answer
• 15:42
  is
• 15:42
  more
• 15:42
  likely
• 15:42
  to
• 15:42
  have
• 15:43
  actually
• 15:43
  found
• 15:43
  it.
• 15:44
  Model-agnostic,
• 15:45
  no
• 15:46
  benchmark-specific
• 15:46
  training
• 15:47
  required.
• 15:48
  The
• 15:48
  Arrow's
• 15:49
  Impossibility
• 15:49
  paper
• 15:50
  reframes
• 15:50
  multi-agent
• 15:51
  fairness
• 15:51
  through
• 15:51
  social
• 15:52
  choice
• 15:52
  theory
• 15:52
  using
• 15:52
  a
• 15:52
  controlled
• 15:53
  hospital
• 15:53
  triage
• 15:54
  framework.
• 15:54
  Two
• 15:55
  LLM
• 15:55
  agents
• 15:56
  negotiate
• 15:56
  over
• 15:56
  three
• 15:57
  structured
• 15:57
  debate
• 15:57
  rounds
• 15:58
  —
• 15:59
  one
• 15:59
  aligned
• 15:59
  to
• 15:59
  a
• 15:59
  specific
• 16:00
  ethical
• 16:00
  framework
• 16:01
  via
• 16:01
  RAG,
• 16:02
  the
• 16:02
  other
• 16:02
  either
• 16:02
  unaligned
• 16:03
  or
• 16:03
  adversarially
• 16:04
  prompted
• 16:04
  to
• 16:04
  favor
• 16:05
  demographics
• 16:05
  over
• 16:05
  clinical
• 16:06
  need.
• 16:07
  The
• 16:07
  central
• 16:07
  finding:
• 16:08
  neither
• 16:08
  agent's
• 16:08
  allocation
• 16:09
  is
• 16:09
  ethically
• 16:09
  adequate
• 16:10
  in
• 16:10
  isolation,
• 16:11
  yet
• 16:11
  their
• 16:11
  joint
• 16:11
  final
• 16:12
  allocation
• 16:12
  satisfies
• 16:13
  fairness
• 16:13
  criteria
• 16:14
  that
• 16:14
  neither
• 16:14
  reaches
• 16:14
  alone.
• 16:15
  Aligned
• 16:16
  agents
• 16:16
  moderate
• 16:16
  bias
• 16:17
  through
• 16:17
  contestation
• 16:18
  rather
• 16:18
  than
• 16:18
  override,
• 16:19
  restoring
• 16:19
  access
• 16:20
  for
• 16:20
  marginalized
• 16:20
  groups
• 16:21
  without
• 16:21
  fully
• 16:21
  converting
• 16:21
  a
• 16:22
  biased
• 16:22
  counterpart.
• 16:23
  Even
• 16:23
  explicitly
• 16:24
  aligned
• 16:24
  agents
• 16:24
  exhibit
• 16:25
  intrinsic
• 16:25
  biases
• 16:25
  toward
• 16:26
  certain
• 16:26
  ethical
• 16:26
  frameworks,
• 16:27
  consistent
• 16:28
  with
• 16:28
  known
• 16:28
  left-leaning
• 16:29
  tendencies
• 16:29
  in
• 16:29
  LLMs.
• 16:30
  The
• 16:30
  connection
• 16:31
  to
• 16:31
  Arrow's
• 16:31
  theorem
• 16:31
  is
• 16:31
  precise:
• 16:32
  no
• 16:33
  aggregation
• 16:33
  mechanism
• 16:34
  simultaneously
• 16:34
  satisfies
• 16:35
  all
• 16:35
  desiderata
• 16:36
  of
• 16:36
  collective
• 16:36
  rationality.
• 16:38
  Multi-agent
• 16:38
  deliberation
• 16:39
  navigates
• 16:39
  rather
• 16:39
  than
• 16:39
  resolves
• 16:40
  this
• 16:40
  constraint.
• 16:41
  The
• 16:41
  practical
• 16:42
  implication
• 16:42
  for
• 16:42
  production
• 16:43
  systems
• 16:43
  is
• 16:43
  direct:
• 16:44
  fairness
• 16:44
  may
• 16:44
  need
• 16:44
  to
• 16:44
  be
• 16:44
  designed
• 16:45
  as
• 16:45
  an
• 16:45
  emergent
• 16:45
  property
• 16:46
  of
• 16:46
  agent
• 16:46
  interaction
• 16:46
  —
• 16:47
  with
• 16:47
  the
• 16:47
  system,
• 16:48
  not
• 16:48
  the
• 16:48
  individual
• 16:49
  model,
• 16:49
  as
• 16:49
  the
• 16:49
  appropriate
• 16:50
  unit
• 16:50
  of
• 16:50
  evaluation.
• 16:52
  For
• 16:52
  anyone
• 16:52
  building
• 16:52
  multi-agent
• 16:53
  pipelines
• 16:53
  for
• 16:54
  high-stakes
• 16:54
  decisions,
• 16:55
  the
• 16:55
  question
• 16:55
  is
• 16:55
  no
• 16:56
  longer
• 16:56
  "is
• 16:56
  each
• 16:56
  agent
• 16:57
  fair?"
• 16:58
  but
• 16:58
  "does
• 16:58
  the
• 16:58
  interaction
• 16:59
  protocol
• 16:59
  produce
• 16:59
  fair
• 16:59
  outcomes?"