Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 31, 2026

Transcript

AGI Dreams Podcast – March 31, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  economics
• 0:00
  of
• 0:00
  exploit
• 0:01
  development
• 0:01
  just
• 0:01
  changed
• 0:02
  on
• 0:02
  a
• 0:02
  timescale
• 0:03
  that
• 0:03
  should
• 0:03
  alarm
• 0:03
  anyone
• 0:03
  responsible
• 0:04
  for
• 0:04
  patching
• 0:04
  anything.
• 0:06
  Thomas
• 0:06
  Ptacek's
• 0:06
  "Vulnerability
• 0:07
  Research
• 0:08
  Is
• 0:08
  Cooked"
• 0:08
  is
• 0:09
  the
• 0:09
  most
• 0:09
  important
• 0:09
  essay
• 0:10
  on
• 0:10
  the
• 0:10
  state
• 0:10
  of
• 0:10
  offensive
• 0:10
  security
• 0:11
  published
• 0:11
  this
• 0:11
  year,
• 0:12
  and
• 0:12
  the
• 0:12
  thesis
• 0:13
  is
• 0:13
  blunt:
• 0:13
  frontier
• 0:14
  LLM
• 0:14
  agents
• 0:15
  are
• 0:15
  about
• 0:15
  to
• 0:15
  supplant
• 0:16
  most
• 0:16
  human
• 0:16
  vulnerability
• 0:17
  research.
• 0:17
  Not
• 0:18
  augment
• 0:18
  it.
• 0:19
  Supplant
• 0:19
  it.
• 0:20
  Ptacek
• 0:20
  draws
• 0:20
  on
• 0:21
  a
• 0:21
  conversation
• 0:21
  with
• 0:21
  Nicholas
• 0:22
  Carlini
• 0:22
  at
• 0:22
  Anthropic's
• 0:22
  Frontier
• 0:23
  Red
• 0:23
  Team,
• 0:24
  who
• 0:24
  described
• 0:24
  a
• 0:24
  deliberately
• 0:25
  simple
• 0:25
  pipeline:
• 0:26
  loop
• 0:26
  over
• 0:26
  every
• 0:27
  source
• 0:27
  file
• 0:27
  in
• 0:27
  a
• 0:27
  repository,
• 0:28
  spam
• 0:29
  the
• 0:29
  same
• 0:29
  Claude
• 0:29
  Code
• 0:29
  prompt
• 0:30
  —
• 0:30
  "find
• 0:30
  me
• 0:31
  an
• 0:31
  exploitable
• 0:31
  vulnerability,
• 0:32
  start
• 0:32
  with
• 0:32
  this
• 0:32
  file"
• 0:32
  —
• 0:33
  then
• 0:33
  feed
• 0:34
  each
• 0:34
  candidate
• 0:34
  report
• 0:34
  back
• 0:35
  through
• 0:35
  a
• 0:35
  second
• 0:35
  run
• 0:35
  to
• 0:35
  verify
• 0:36
  exploitability.
• 0:37
  The
• 0:37
  success
• 0:37
  rate
• 0:38
  on
• 0:38
  that
• 0:38
  verification
• 0:39
  step:
• 0:39
  "almost
• 0:39
  100%."
• 0:41
  Carlini
• 0:42
  aimed
• 0:42
  the
• 0:42
  pipeline
• 0:42
  at
• 0:42
  Ghost,
• 0:43
  the
• 0:43
  popular
• 0:44
  CMS,
• 0:44
  and
• 0:44
  it
• 0:45
  produced
• 0:45
  a
• 0:45
  broadly
• 0:45
  exploitable
• 0:46
  SQL
• 0:46
  injection.
• 0:47
  No
• 0:47
  custom
• 0:48
  tooling,
• 0:48
  no
• 0:48
  fuzzers,
• 0:49
  no
• 0:49
  model
• 0:49
  checkers.
• 0:50
  Just
• 0:50
  the
• 0:50
  model,
• 0:51
  reading
• 0:51
  code
• 0:52
  and
• 0:52
  reasoning
• 0:52
  about
• 0:52
  reachability.
• 0:54
  Ptacek's
• 0:54
  framing
• 0:54
  is
• 0:55
  Sutton's
• 0:55
  Bitter
• 0:55
  Lesson
• 0:55
  applied
• 0:56
  to
• 0:56
  security:
• 0:57
  researchers
• 0:57
  spend
• 0:57
  20%
• 0:58
  of
• 0:58
  their
• 0:58
  time
• 0:58
  on
• 0:58
  computer
• 0:59
  science
• 0:59
  and
• 0:59
  80%
• 1:00
  on
• 1:00
  giant
• 1:00
  jigsaw
• 1:00
  puzzles.
• 1:01
  "And
• 1:01
  now
• 1:02
  everybody
• 1:02
  has
• 1:02
  a
• 1:02
  universal
• 1:03
  jigsaw
• 1:03
  solver."
• 1:04
  The
• 1:04
  implications
• 1:05
  Ptacek
• 1:05
  draws
• 1:06
  are
• 1:06
  structural.
• 1:07
  In
• 1:07
  a
• 1:07
  post-attention-scarcity
• 1:08
  world,
• 1:09
  exploit
• 1:09
  developers
• 1:10
  won't
• 1:10
  carefully
• 1:10
  pick
• 1:10
  targets
• 1:11
  —
• 1:11
  they'll
• 1:12
  aim
• 1:12
  at
• 1:12
  everything:
• 1:13
  routers,
• 1:13
  printers,
• 1:14
  hospital
• 1:14
  imaging
• 1:15
  systems,
• 1:15
  the
• 1:16
  inexplicably
• 1:16
  networked
• 1:16
  components
• 1:17
  of
• 1:17
  a
• 1:17
  dishwasher.
• 1:18
  The
• 1:18
  load-bearing
• 1:19
  risk
• 1:19
  assumption
• 1:19
  that
• 1:19
  elite
• 1:20
  talent
• 1:20
  won't
• 1:20
  bother
• 1:20
  with
• 1:21
  unglamorous
• 1:21
  targets
• 1:21
  "no
• 1:21
  longer
• 1:22
  holds."
• 1:22
  He's
• 1:22
  equally
• 1:23
  worried
• 1:23
  about
• 1:23
  regulation:
• 1:24
  that
• 1:24
  a
• 1:24
  wave
• 1:24
  of
• 1:25
  AI-accelerated
• 1:25
  ransomware
• 1:26
  incidents
• 1:26
  will
• 1:26
  produce
• 1:27
  incoherent
• 1:27
  computer
• 1:27
  security
• 1:28
  legislation
• 1:29
  drafted
• 1:29
  by
• 1:29
  policymakers
• 1:30
  who
• 1:30
  don't
• 1:30
  grasp
• 1:30
  that
• 1:30
  unregulated
• 1:31
  open-weight
• 1:31
  models
• 1:32
  will
• 1:32
  have
• 1:32
  the
• 1:32
  same
• 1:33
  capabilities
• 1:33
  in
• 1:33
  nine
• 1:33
  months.
• 1:34
  As
• 1:35
  confirmation,
• 1:35
  the
• 1:36
  calif.io
• 1:36
  team
• 1:37
  reported
• 1:37
  finding
• 1:37
  RCE
• 1:38
  zero-days
• 1:38
  in
• 1:38
  both
• 1:38
  Vim
• 1:39
  and
• 1:39
  Emacs
• 1:39
  simply
• 1:40
  by
• 1:40
  prompting
• 1:40
  Claude
• 1:41
  to
• 1:41
  look
• 1:41
  —
• 1:42
  Vim
• 1:42
  patched
• 1:42
  immediately;
• 1:43
  Emacs
• 1:44
  maintainers
• 1:44
  declined,
• 1:45
  attributing
• 1:45
  the
• 1:45
  issue
• 1:45
  to
• 1:46
  git.
• 1:46
  The
• 1:46
  team
• 1:47
  has
• 1:47
  launched
• 1:47
  "MAD
• 1:47
  Bugs:
• 1:48
  Month
• 1:48
  of
• 1:48
  AI-Discovered
• 1:49
  Bugs,"
• 1:49
  promising
• 1:50
  more
• 1:50
  through
• 1:50
  April.
• 1:55
  If
• 1:55
  exploit
• 1:55
  agents
• 1:56
  are
• 1:56
  the
• 1:56
  offensive
• 1:56
  problem,
• 1:57
  a
• 1:57
  new
• 1:57
  paper
• 1:57
  from
• 1:57
  NTU,
• 1:58
  ASTAR,
• 1:59
  and
• 1:59
  Johns
• 1:59
  Hopkins
• 2:00
  identifies
• 2:00
  a
• 2:00
  defensive
• 2:01
  one
• 2:01
  that's
• 2:01
  arguably
• 2:01
  harder
• 2:02
  to
• 2:02
  fix:
• 2:03
  persistent
• 2:03
  AI
• 2:03
  agents
• 2:04
  can
• 2:04
  be
• 2:04
  silently
• 2:04
  poisoned
• 2:05
  through
• 2:05
  their
• 2:05
  own
• 2:05
  background
• 2:05
  execution
• 2:06
  loops,
• 2:06
  with
• 2:07
  no
• 2:07
  prompt
• 2:07
  injection
• 2:07
  required.
• 2:08
  The
• 2:08
  HEARTBEAT
• 2:09
  paper
• 2:09
  formalizes
• 2:10
  what
• 2:10
  the
• 2:10
  authors
• 2:10
  call
• 2:10
  the
• 2:11
  Exposure
• 2:11
  →
• 2:11
  Memory
• 2:11
  →
• 2:12
  Behavior
• 2:12
  (E→M→B)
• 2:13
  pathway,
• 2:13
  demonstrating
• 2:14
  that
• 2:14
  misinformation
• 2:15
  encountered
• 2:15
  during
• 2:16
  routine
• 2:16
  heartbeat-driven
• 2:17
  background
• 2:17
  monitoring
• 2:17
  —
• 2:18
  checking
• 2:18
  email,
• 2:19
  browsing
• 2:19
  social
• 2:19
  feeds,
• 2:20
  scanning
• 2:20
  GitHub
• 2:21
  issues
• 2:21
  —
• 2:21
  enters
• 2:22
  the
• 2:22
  same
• 2:22
  shared
• 2:22
  session
• 2:23
  context
• 2:23
  used
• 2:24
  for
• 2:24
  foreground
• 2:24
  user
• 2:24
  interaction.
• 2:26
  From
• 2:26
  the
• 2:26
  LLM's
• 2:26
  perspective,
• 2:27
  a
• 2:27
  heartbeat
• 2:28
  tick
• 2:28
  is
• 2:28
  just
• 2:28
  another
• 2:29
  user
• 2:29
  message.
• 2:30
  The
• 2:30
  numbers
• 2:30
  are
• 2:30
  sobering.
• 2:31
  Using
• 2:32
  MissClaw,
• 2:33
  a
• 2:33
  controlled
• 2:33
  research
• 2:33
  replica
• 2:34
  of
• 2:34
  the
• 2:34
  Moltbook
• 2:34
  agent
• 2:35
  social
• 2:35
  platform,
• 2:36
  the
• 2:36
  researchers
• 2:36
  evaluated
• 2:37
  OpenClaw
• 2:37
  agents
• 2:38
  across
• 2:38
  software
• 2:38
  security,
• 2:39
  financial
• 2:39
  decision-making,
• 2:40
  and
• 2:40
  academic
• 2:41
  reference
• 2:41
  tasks.
• 2:42
  Social
• 2:42
  credibility
• 2:42
  cues
• 2:43
  —
• 2:43
  especially
• 2:44
  perceived
• 2:44
  consensus
• 2:45
  —
• 2:45
  drove
• 2:46
  misleading
• 2:46
  rates
• 2:46
  up
• 2:46
  to
• 2:46
  61%
• 2:47
  in
• 2:47
  the
• 2:47
  short-term
• 2:48
  setting.
• 2:48
  Routine
• 2:49
  memory-saving
• 2:49
  behavior
• 2:50
  promoted
• 2:50
  short-term
• 2:50
  pollution
• 2:51
  into
• 2:51
  durable
• 2:51
  long-term
• 2:52
  memory
• 2:52
  at
• 2:52
  rates
• 2:52
  up
• 2:52
  to
• 2:53
  91%,
• 2:54
  with
• 2:54
  cross-session
• 2:54
  behavioral
• 2:55
  influence
• 2:55
  reaching
• 2:55
  76%.
• 2:57
  Under
• 2:57
  naturalistic
• 2:58
  browsing
• 2:58
  conditions,
• 2:59
  where
• 2:59
  manipulated
• 3:00
  content
• 3:00
  was
• 3:00
  diluted
• 3:01
  among
• 3:01
  benign
• 3:01
  posts
• 3:02
  and
• 3:02
  had
• 3:02
  to
• 3:02
  survive
• 3:03
  the
• 3:03
  system's
• 3:03
  own
• 3:03
  context
• 3:04
  pruning,
• 3:05
  pollution
• 3:05
  still
• 3:05
  crossed
• 3:06
  session
• 3:06
  boundaries.
• 3:07
  The
• 3:07
  attack
• 3:07
  requires
• 3:08
  no
• 3:08
  direct
• 3:09
  interaction
• 3:09
  with
• 3:09
  the
• 3:09
  victim
• 3:10
  agent:
• 3:11
  place
• 3:11
  credible-looking
• 3:12
  misinformation
• 3:12
  where
• 3:13
  the
• 3:13
  agent
• 3:13
  is
• 3:13
  likely
• 3:13
  to
• 3:14
  encounter
• 3:14
  it
• 3:14
  during
• 3:14
  background
• 3:15
  activity,
• 3:16
  and
• 3:16
  the
• 3:16
  content
• 3:16
  enters
• 3:17
  as
• 3:17
  ordinary
• 3:17
  information
• 3:18
  rather
• 3:18
  than
• 3:19
  an
• 3:19
  overt
• 3:19
  control
• 3:19
  attempt,
• 3:20
  making
• 3:21
  it
• 3:21
  less
• 3:21
  likely
• 3:21
  to
• 3:21
  trigger
• 3:22
  safety
• 3:22
  guardrails
• 3:23
  than
• 3:23
  explicit
• 3:23
  prompt
• 3:24
  injection.
• 3:25
  Three
• 3:25
  properties
• 3:26
  make
• 3:26
  this
• 3:26
  especially
• 3:26
  dangerous:
• 3:28
  weak
• 3:28
  source
• 3:28
  distinction
• 3:29
  (heartbeat-acquired
• 3:30
  content
• 3:30
  isn't
• 3:30
  separated
• 3:31
  from
• 3:31
  user-provided
• 3:32
  information),
• 3:33
  limited
• 3:34
  user
• 3:34
  visibility
• 3:35
  (the
• 3:35
  triggering
• 3:35
  exposure
• 3:36
  may
• 3:36
  never
• 3:36
  surface
• 3:37
  to
• 3:37
  the
• 3:37
  user),
• 3:38
  and
• 3:38
  provenance-free
• 3:39
  memory
• 3:39
  promotion
• 3:40
  (externally
• 3:40
  encountered
• 3:41
  content
• 3:41
  gets
• 3:42
  written
• 3:42
  to
• 3:42
  long-term
• 3:43
  memory
• 3:43
  without
• 3:43
  source
• 3:44
  attribution).
• 3:45
  The
• 3:45
  paper
• 3:46
  illustrates
• 3:46
  this
• 3:46
  with
• 3:46
  a
• 3:47
  concrete
• 3:47
  scenario:
• 3:48
  a
• 3:48
  poisoned
• 3:49
  post
• 3:49
  claims
• 3:49
  a
• 3:49
  library
• 3:50
  has
• 3:50
  a
• 3:50
  critical
• 3:51
  RCE
• 3:51
  and
• 3:52
  recommends
• 3:52
  migrating
• 3:53
  to
• 3:53
  a
• 3:53
  malicious
• 3:53
  replacement.
• 3:54
  The
• 3:54
  agent
• 3:55
  absorbs
• 3:55
  this
• 3:55
  during
• 3:56
  background
• 3:56
  scanning,
• 3:57
  saves
• 3:57
  it
• 3:57
  to
• 3:57
  memory,
• 3:58
  and
• 3:58
  later
• 3:58
  presents
• 3:59
  it
• 3:59
  as
• 3:59
  "known
• 3:59
  issue"
• 4:00
  knowledge
• 4:00
  —
• 4:01
  source
• 4:01
  laundered,
• 4:02
  provenance
• 4:02
  erased,
• 4:03
  user
• 4:03
  grateful
• 4:04
  for
• 4:04
  the
• 4:04
  warning
• 4:04
  as
• 4:04
  they
• 4:05
  install
• 4:05
  the
• 4:05
  attacker's
• 4:05
  package.
• 4:07
  The
• 4:07
  vulnerability
• 4:08
  is
• 4:08
  architectural
• 4:08
  —
• 4:09
  as
• 4:09
  long
• 4:10
  as
• 4:10
  heartbeat
• 4:10
  ingests
• 4:11
  untrusted
• 4:11
  content
• 4:12
  in
• 4:12
  the
• 4:12
  same
• 4:12
  session
• 4:13
  used
• 4:13
  for
• 4:13
  foreground
• 4:14
  interaction,
• 4:15
  silent
• 4:15
  memory
• 4:15
  pollution
• 4:16
  follows
• 4:16
  from
• 4:17
  the
• 4:17
  design
• 4:17
  itself.
• 4:18
  Meanwhile,
• 4:19
  on
• 4:19
  the
• 4:19
  inference
• 4:19
  infrastructure
• 4:20
  side,
• 4:21
  vLLM
• 4:21
  disclosed
• 4:22
  CVE-2026-27893:
• 4:28
  two
• 4:28
  model
• 4:28
  files
• 4:29
  hardcode
• 4:29
  trust
• 4:29
  remote
• 4:30
  code=True,
• 4:31
  silently
• 4:32
  overriding
• 4:32
  an
• 4:32
  explicit
• 4:33
  False
• 4:33
  setting
• 4:34
  with
• 4:34
  no
• 4:34
  warning.
• 4:35
  A
• 4:35
  malicious
• 4:35
  Hugging
• 4:36
  Face
• 4:36
  repository
• 4:37
  targeting
• 4:37
  Nemotron-VL
• 4:38
  or
• 4:39
  Kimi-K25
• 4:40
  can
• 4:40
  achieve
• 4:41
  code
• 4:41
  execution
• 4:41
  on
• 4:42
  the
• 4:42
  inference
• 4:42
  server.
• 4:43
  This
• 4:43
  is
• 4:44
  the
• 4:44
  third
• 4:44
  time
• 4:44
  the
• 4:44
  same
• 4:45
  vulnerability
• 4:45
  class
• 4:46
  has
• 4:46
  surfaced
• 4:47
  in
• 4:47
  vLLM,
• 4:48
  each
• 4:48
  time
• 4:48
  in
• 4:48
  a
• 4:48
  different
• 4:49
  code
• 4:49
  path.
• 4:50
  Versions
• 4:50
  0.10.1
• 4:52
  through
• 4:52
  0.17.x
• 4:55
  are
• 4:55
  affected;
• 4:55
  0.18.0
• 4:58
  contains
• 4:58
  the
• 4:58
  fix.
• 5:03
  The
• 5:03
  same
• 5:03
  agentic
• 5:04
  architecture
• 5:04
  that
• 5:04
  creates
• 5:05
  new
• 5:05
  attack
• 5:05
  surfaces
• 5:06
  is
• 5:06
  now
• 5:06
  being
• 5:07
  deployed
• 5:07
  as
• 5:07
  the
• 5:08
  security
• 5:08
  response
• 5:08
  itself.
• 5:09
  Tenable
• 5:10
  launched
• 5:10
  Hexa
• 5:11
  AI,
• 5:11
  the
• 5:11
  agentic
• 5:12
  engine
• 5:12
  powering
• 5:12
  the
• 5:12
  Tenable
• 5:13
  One
• 5:13
  Exposure
• 5:13
  Management
• 5:14
  Platform,
• 5:15
  featuring
• 5:15
  a
• 5:15
  fleet
• 5:15
  of
• 5:15
  mission-ready
• 5:16
  AI
• 5:16
  agents
• 5:16
  that
• 5:17
  coordinate
• 5:17
  human
• 5:17
  approvals
• 5:18
  and
• 5:18
  automated
• 5:18
  workflows
• 5:19
  into
• 5:19
  a
• 5:19
  single
• 5:19
  orchestration
• 5:20
  layer.
• 5:21
  The
• 5:21
  pitch:
• 5:22
  in
• 5:22
  the
• 5:22
  time
• 5:22
  it
• 5:22
  takes
• 5:23
  to
• 5:23
  read
• 5:23
  a
• 5:23
  blog
• 5:23
  post,
• 5:24
  an
• 5:25
  AI-assisted
• 5:25
  attacker
• 5:26
  can
• 5:26
  scan
• 5:26
  an
• 5:26
  environment,
• 5:27
  gain
• 5:28
  initial
• 5:28
  access,
• 5:29
  move
• 5:29
  laterally,
• 5:30
  and
• 5:30
  breach
• 5:30
  sensitive
• 5:31
  data.
• 5:32
  Tenable
• 5:32
  cites
• 5:32
  Anthropic's
• 5:33
  November
• 5:33
  2025
• 5:34
  disclosure
• 5:35
  that
• 5:35
  a
• 5:35
  threat
• 5:35
  actor
• 5:35
  jailbroke
• 5:36
  Claude
• 5:36
  Code
• 5:37
  to
• 5:37
  automate
• 5:37
  90%
• 5:38
  of
• 5:38
  a
• 5:38
  targeted
• 5:39
  attack
• 5:39
  at
• 5:39
  machine
• 5:40
  speed.
• 5:41
  Hexa
• 5:41
  AI
• 5:42
  responds
• 5:42
  with
• 5:42
  assessment
• 5:43
  configuration
• 5:43
  agents
• 5:44
  (asset
• 5:44
  categorization
• 5:45
  at
• 5:45
  scale),
• 5:45
  risk
• 5:46
  intelligence
• 5:46
  agents
• 5:46
  (contextual
• 5:47
  dashboards
• 5:47
  on
• 5:47
  demand),
• 5:48
  and
• 5:48
  advanced
• 5:49
  workflow
• 5:49
  orchestration
• 5:49
  —
• 5:51
  all
• 5:51
  with
• 5:51
  Model
• 5:51
  Context
• 5:52
  Protocol
• 5:52
  (MCP)
• 5:53
  built
• 5:53
  in
• 5:54
  as
• 5:54
  a
• 5:54
  universal
• 5:54
  adapter
• 5:55
  so
• 5:55
  customers
• 5:55
  can
• 5:55
  connect
• 5:56
  business
• 5:56
  tools
• 5:56
  to
• 5:56
  any
• 5:57
  LLM.
• 5:58
  The
• 5:58
  "human
• 5:59
  in
• 5:59
  the
• 5:59
  loop"
• 5:59
  dial
• 5:59
  is
• 6:00
  adjustable:
• 6:01
  full
• 6:01
  autonomous
• 6:02
  execution
• 6:02
  or
• 6:02
  strategic
• 6:03
  manual
• 6:03
  oversight,
• 6:04
  with
• 6:04
  everything
• 6:05
  logged
• 6:05
  for
• 6:05
  audit.
• 6:06
  On
• 6:06
  the
• 6:06
  regulatory
• 6:07
  side,
• 6:07
  RegIntel
• 6:08
  takes
• 6:08
  a
• 6:08
  different
• 6:09
  angle
• 6:09
  on
• 6:09
  agentic
• 6:09
  security:
• 6:11
  an
• 6:11
  AI-powered
• 6:12
  CLI
• 6:12
  that
• 6:12
  automatically
• 6:13
  researches,
• 6:14
  stores,
• 6:14
  and
• 6:14
  manages
• 6:15
  a
• 6:15
  global
• 6:15
  inventory
• 6:16
  of
• 6:16
  penetration
• 6:16
  testing
• 6:17
  regulations
• 6:17
  for
• 6:18
  financial
• 6:18
  services
• 6:19
  across
• 6:19
  20+
• 6:20
  jurisdictions.
• 6:21
  The
• 6:21
  architecture
• 6:22
  is
• 6:22
  a
• 6:22
  four-agent
• 6:22
  LangGraph
• 6:23
  pipeline
• 6:23
  —
• 6:24
  Research
• 6:24
  (GPT-5.4-mini
• 6:25
  for
• 6:25
  web
• 6:26
  search
• 6:26
  and
• 6:26
  extraction),
• 6:27
  Validator
• 6:27
  (rule-based
• 6:28
  pre-checks
• 6:28
  plus
• 6:28
  LLM
• 6:29
  validation),
• 6:30
  Reflection
• 6:30
  (local
• 6:30
  Qwen
• 6:30
  3.5
• 6:31
  122B
• 6:32
  running
• 6:32
  a
• 6:32
  quality
• 6:33
  gate
• 6:33
  with
• 6:33
  hallucination
• 6:33
  detection),
• 6:34
  and
• 6:34
  Persister
• 6:35
  (database
• 6:35
  writes
• 6:35
  with
• 6:35
  change-detection
• 6:36
  diffing).
• 6:37
  Every
• 6:37
  new
• 6:37
  regulation
• 6:38
  must
• 6:38
  pass
• 6:38
  through
• 6:39
  the
• 6:39
  reflection
• 6:39
  quality
• 6:39
  gate
• 6:40
  before
• 6:40
  reaching
• 6:40
  the
• 6:40
  database,
• 6:41
  and
• 6:41
  a
• 6:41
  ChromaDB
• 6:42
  vector
• 6:42
  store
• 6:42
  enables
• 6:43
  semantic
• 6:43
  search
• 6:44
  across
• 6:44
  the
• 6:44
  regulatory
• 6:44
  landscape.
• 6:45
  The
• 6:45
  dual-model
• 6:46
  approach
• 6:46
  —
• 6:47
  cloud
• 6:47
  for
• 6:47
  speed,
• 6:48
  local
• 6:48
  for
• 6:48
  verification
• 6:48
  —
• 6:49
  is
• 6:49
  a
• 6:49
  practical
• 6:50
  pattern
• 6:50
  worth
• 6:50
  watching.
• 6:55
  A
• 6:55
  quality
• 6:55
  engineer
• 6:56
  in
• 6:56
  Serbia
• 6:56
  opened
• 6:57
  his
• 6:57
  assessment
• 6:57
  tool
• 6:58
  last
• 6:58
  Tuesday
• 6:58
  and
• 6:58
  found
• 6:59
  95%
• 7:00
  test
• 7:00
  coverage
• 7:00
  on
• 7:00
  a
• 7:00
  module
• 7:01
  for
• 7:01
  which
• 7:01
  he
• 7:01
  had
• 7:01
  not
• 7:02
  written
• 7:02
  a
• 7:02
  single
• 7:02
  test.
• 7:03
  The
• 7:03
  number
• 7:03
  was
• 7:03
  fabricated,
• 7:04
  stitched
• 7:04
  together
• 7:05
  from
• 7:05
  fragmented
• 7:05
  key
• 7:06
  conventions
• 7:06
  and
• 7:07
  stale
• 7:07
  snapshots.
• 7:08
  Five
• 7:08
  different
• 7:08
  key
• 7:09
  formats
• 7:09
  existed
• 7:09
  for
• 7:09
  the
• 7:10
  same
• 7:10
  concept;
• 7:11
  none
• 7:11
  talked
• 7:11
  to
• 7:11
  each
• 7:11
  other.
• 7:12
  The
• 7:13
  score
• 7:13
  consumed
• 7:13
  whichever
• 7:14
  fragment
• 7:14
  it
• 7:14
  found
• 7:14
  first
• 7:15
  and
• 7:15
  called
• 7:15
  it
• 7:15
  truth.
• 7:16
  Dragan
• 7:16
  Spiridonov's
• 7:17
  "The
• 7:17
  Witness
• 7:17
  Stand"
• 7:17
  essay
• 7:18
  argues
• 7:18
  that
• 7:18
  the
• 7:18
  standard
• 7:19
  approach
• 7:19
  to
• 7:19
  agent
• 7:19
  trust
• 7:20
  —
• 7:20
  better
• 7:21
  prompts,
• 7:21
  more
• 7:21
  guardrails,
• 7:22
  human-in-the-loop
• 7:23
  review
• 7:23
  —
• 7:24
  is
• 7:24
  a
• 7:24
  "hope
• 7:24
  architecture,"
• 7:25
  and
• 7:25
  that
• 7:25
  classical
• 7:26
  testing
• 7:26
  solved
• 7:26
  this
• 7:27
  problem
• 7:27
  decades
• 7:27
  ago.
• 7:28
  His
• 7:28
  prescription:
• 7:29
  replace
• 7:29
  djb2
• 7:30
  hashes
• 7:30
  with
• 7:30
  SHA-256
• 7:33
  for
• 7:33
  tamper-evident
• 7:33
  witness
• 7:34
  records,
• 7:34
  enforce
• 7:35
  deterministic
• 7:35
  YAML
• 7:36
  pipelines
• 7:36
  that
• 7:36
  execute
• 7:37
  without
• 7:37
  consuming
• 7:37
  a
• 7:37
  single
• 7:38
  LLM
• 7:38
  token,
• 7:39
  and
• 7:39
  deploy
• 7:39
  CUSUM
• 7:40
  drift
• 7:40
  detectors
• 7:40
  that
• 7:40
  watch
• 7:41
  for
• 7:41
  statistical
• 7:41
  shifts
• 7:41
  continuously.
• 7:43
  "Stop
• 7:43
  building
• 7:43
  trust.
• 7:43
  Start
• 7:44
  building
• 7:44
  evidence."
• 7:45
  The
• 7:45
  containment
• 7:45
  problem
• 7:46
  is
• 7:46
  getting
• 7:46
  its
• 7:46
  own
• 7:46
  tooling
• 7:46
  stack.
• 7:47
  Stanford's
• 7:47
  jai
• 7:48
  is
• 7:48
  a
• 7:48
  one-command
• 7:48
  casual
• 7:49
  sandbox
• 7:49
  for
• 7:49
  Linux
• 7:50
  that
• 7:50
  gives
• 7:50
  AI
• 7:51
  agents
• 7:51
  access
• 7:51
  to
• 7:52
  a
• 7:52
  working
• 7:52
  directory
• 7:52
  while
• 7:53
  keeping
• 7:53
  the
• 7:53
  rest
• 7:53
  of
• 7:53
  the
• 7:53
  home
• 7:54
  directory
• 7:54
  behind
• 7:54
  a
• 7:54
  copy-on-write
• 7:55
  overlay.
• 7:56
  The
• 7:56
  motivation
• 7:57
  is
• 7:57
  a
• 7:57
  growing
• 7:57
  catalog
• 7:57
  of
• 7:58
  horror
• 7:58
  stories:
• 7:59
  15
• 7:59
  years
• 8:00
  of
• 8:00
  family
• 8:00
  photos
• 8:00
  deleted
• 8:01
  via
• 8:01
  terminal,
• 8:01
  Claude
• 8:02
  Code
• 8:02
  wiping
• 8:02
  a
• 8:02
  home
• 8:02
  directory,
• 8:04
  Cursor
• 8:04
  emptying
• 8:04
  a
• 8:04
  working
• 8:05
  tree,
• 8:05
  Google
• 8:06
  Antigravity
• 8:06
  wiping
• 8:07
  an
• 8:07
  entire
• 8:07
  drive.
• 8:08
  jai
• 8:08
  offers
• 8:08
  three
• 8:09
  isolation
• 8:09
  modes
• 8:09
  from
• 8:10
  weak
• 8:10
  (most
• 8:10
  files
• 8:10
  readable)
• 8:11
  to
• 8:11
  strict
• 8:11
  (home
• 8:11
  hidden),
• 8:12
  with
• 8:12
  no
• 8:12
  Dockerfiles
• 8:13
  or
• 8:13
  40-flag
• 8:14
  bwrap
• 8:14
  invocations
• 8:15
  required.
• 8:16
  The
• 8:16
  developers
• 8:16
  are
• 8:16
  explicit:
• 8:17
  this
• 8:17
  is
• 8:17
  not
• 8:17
  a
• 8:17
  promise
• 8:18
  of
• 8:18
  perfect
• 8:18
  safety,
• 8:19
  and
• 8:19
  it's
• 8:19
  not
• 8:19
  equivalent
• 8:20
  to
• 8:20
  a
• 8:20
  hardened
• 8:20
  container
• 8:21
  runtime.
• 8:21
  It
• 8:21
  fills
• 8:22
  the
• 8:22
  gap
• 8:22
  between
• 8:22
  giving
• 8:23
  an
• 8:23
  agent
• 8:23
  your
• 8:23
  real
• 8:23
  account
• 8:24
  and
• 8:24
  stopping
• 8:24
  everything
• 8:25
  to
• 8:25
  build
• 8:25
  a
• 8:25
  VM.
• 8:26
  And
• 8:26
  then
• 8:26
  there's
• 8:26
  the
• 8:27
  failure
• 8:27
  mode
• 8:27
  nobody
• 8:27
  quite
• 8:28
  anticipated:
• 8:29
  GitHub
• 8:29
  Copilot
• 8:30
  editing
• 8:30
  an
• 8:30
  ad
• 8:31
  for
• 8:31
  itself
• 8:32
  and
• 8:32
  Raycast
• 8:32
  into
• 8:32
  a
• 8:33
  developer's
• 8:33
  pull
• 8:33
  request
• 8:34
  description
• 8:34
  after
• 8:35
  a
• 8:35
  team
• 8:35
  member
• 8:35
  summoned
• 8:35
  it
• 8:36
  to
• 8:36
  correct
• 8:36
  a
• 8:36
  typo.
• 8:37
  The
• 8:37
  developer's
• 8:37
  reaction
• 8:38
  invoked
• 8:38
  Cory
• 8:39
  Doctorow's
• 8:39
  enshittification
• 8:40
  framework:
• 8:41
  platforms
• 8:41
  are
• 8:41
  good
• 8:41
  to
• 8:42
  users,
• 8:42
  then
• 8:43
  abuse
• 8:43
  users
• 8:43
  for
• 8:44
  business
• 8:44
  customers,
• 8:45
  then
• 8:45
  abuse
• 8:45
  business
• 8:46
  customers
• 8:46
  to
• 8:46
  claw
• 8:47
  back
• 8:47
  value.
• 8:48
  When
• 8:48
  the
• 8:48
  tool
• 8:48
  writing
• 8:49
  your
• 8:49
  code
• 8:49
  is
• 8:49
  also
• 8:49
  selling
• 8:50
  you
• 8:50
  products
• 8:50
  inside
• 8:51
  your
• 8:51
  own
• 8:51
  PRs,
• 8:52
  the
• 8:52
  trust
• 8:52
  calculus
• 8:53
  changes
• 8:53
  in
• 8:53
  ways
• 8:53
  that
• 8:54
  guardrails
• 8:54
  and
• 8:54
  sandboxes
• 8:55
  can't
• 8:55
  address.
• 8:59
  The
• 8:59
  most
• 8:59
  technically
• 9:00
  interesting
• 9:00
  agent
• 9:00
  research
• 9:01
  this
• 9:01
  week
• 9:01
  comes
• 9:01
  from
• 9:02
  Stanford
• 9:02
  and
• 9:02
  collaborators:
• 9:03
  Meta-Harness,
• 9:03
  a
• 9:04
  system
• 9:04
  that
• 9:04
  learns
• 9:04
  to
• 9:04
  optimize
• 9:05
  model
• 9:05
  harnesses
• 9:05
  —
• 9:06
  the
• 9:06
  system
• 9:07
  prompts,
• 9:07
  tool
• 9:07
  definitions,
• 9:08
  completion-checking
• 9:09
  logic,
• 9:10
  and
• 9:10
  context
• 9:10
  management
• 9:11
  that
• 9:11
  wrap
• 9:11
  a
• 9:11
  base
• 9:11
  model
• 9:11
  —
• 9:12
  by
• 9:12
  giving
• 9:12
  a
• 9:12
  proposer
• 9:13
  agent
• 9:13
  access
• 9:14
  to
• 9:14
  a
• 9:14
  filesystem
• 9:14
  containing
• 9:15
  the
• 9:15
  full
• 9:15
  source
• 9:15
  code,
• 9:16
  scores,
• 9:16
  and
• 9:16
  execution
• 9:17
  traces
• 9:17
  of
• 9:17
  every
• 9:17
  prior
• 9:18
  candidate.
• 9:19
  In
• 9:19
  practice,
• 9:20
  this
• 9:20
  means
• 9:20
  up
• 9:20
  to
• 9:20
  10
• 9:20
  million
• 9:21
  tokens
• 9:21
  of
• 9:21
  diagnostic
• 9:22
  context
• 9:22
  per
• 9:22
  optimization
• 9:23
  step,
• 9:24
  versus
• 9:24
  at
• 9:24
  most
• 9:24
  26K
• 9:25
  for
• 9:25
  prior
• 9:25
  methods
• 9:26
  like
• 9:26
  OpenEvolve
• 9:26
  or
• 9:27
  TTT-Discover.
• 9:28
  On
• 9:28
  TerminalBench-2
• 9:29
  (89
• 9:29
  Dockerized
• 9:30
  tasks
• 9:30
  spanning
• 9:30
  code
• 9:31
  translation,
• 9:32
  distributed
• 9:32
  ML,
• 9:33
  systems
• 9:33
  programming,
• 9:34
  bioinformatics,
• 9:34
  and
• 9:35
  cryptanalysis),
• 9:36
  Meta-Harness
• 9:36
  achieved
• 9:37
  78.2%
• 9:38
  pass
• 9:38
  rate
• 9:38
  with
• 9:38
  Opus
• 9:38
  4.6
• 9:39
  (ranking
• 9:39
  #2
• 9:40
  among
• 9:40
  all
• 9:40
  Opus
• 9:40
  agents)
• 9:41
  and
• 9:41
  37.1%
• 9:42
  with
• 9:42
  Haiku
• 9:43
  4.5
• 9:43
  (ranking
• 9:44
  #1
• 9:45
  among
• 9:45
  all
• 9:45
  Haiku
• 9:45
  agents,
• 9:46
  outperforming
• 9:47
  the
• 9:47
  next-best
• 9:47
  agent
• 9:47
  Goose
• 9:47
  at
• 9:48
  35.5%).
• 9:49
  On
• 9:50
  text
• 9:50
  classification,
• 9:51
  it
• 9:51
  hit
• 9:51
  51.7%
• 9:52
  versus
• 9:52
  ACE's
• 9:52
  40.9%
• 9:53
  —
• 9:54
  a
• 9:54
  26%
• 9:55
  relative
• 9:55
  improvement
• 9:56
  using
• 9:56
  zero
• 9:56
  additional
• 9:57
  LLM
• 9:57
  calls.
• 9:58
  The
• 9:58
  key
• 9:59
  insight:
• 9:59
  most
• 10:00
  program-search
• 10:00
  methods
• 10:00
  compress
• 10:01
  history
• 10:01
  into
• 10:01
  fixed
• 10:02
  prompt
• 10:02
  formats,
• 10:03
  discarding
• 10:03
  the
• 10:03
  execution
• 10:03
  traces
• 10:04
  that
• 10:04
  targeted
• 10:04
  diagnosis
• 10:05
  requires.
• 10:06
  The
• 10:06
  adversarial
• 10:06
  development
• 10:07
  pattern
• 10:07
  is
• 10:07
  going
• 10:07
  mainstream.
• 10:08
  A
• 10:08
  detailed
• 10:09
  walkthrough
• 10:09
  of
• 10:09
  building
• 10:10
  GAN-inspired
• 10:10
  coding
• 10:11
  harnesses
• 10:11
  —
• 10:12
  where
• 10:12
  a
• 10:12
  generator
• 10:12
  agent
• 10:13
  implements
• 10:13
  and
• 10:13
  an
• 10:13
  evaluator
• 10:14
  agent
• 10:14
  rips
• 10:14
  apart
• 10:14
  the
• 10:15
  implementation
• 10:15
  across
• 10:16
  negotiated
• 10:16
  sprint
• 10:16
  contracts
• 10:17
  —
• 10:17
  demonstrates
• 10:18
  the
• 10:18
  architecture
• 10:19
  producing
• 10:19
  polished
• 10:19
  full-stack
• 10:20
  applications
• 10:20
  in
• 10:20
  single
• 10:21
  shots
• 10:21
  with
• 10:21
  Sonnet
• 10:21
  4.6
• 10:22
  that
• 10:22
  Opus
• 10:23
  4.6
• 10:23
  couldn't
• 10:24
  match
• 10:24
  without
• 10:24
  the
• 10:24
  harness.
• 10:25
  The
• 10:25
  harness
• 10:26
  enables
• 10:26
  cheaper,
• 10:27
  faster
• 10:27
  models
• 10:27
  to
• 10:27
  produce
• 10:28
  better
• 10:28
  results
• 10:28
  than
• 10:28
  expensive
• 10:29
  models
• 10:29
  flying
• 10:30
  solo,
• 10:30
  which
• 10:31
  inverts
• 10:31
  the
• 10:31
  usual
• 10:32
  "just
• 10:32
  use
• 10:32
  the
• 10:32
  biggest
• 10:32
  model"
• 10:33
  calculus.
• 10:34
  On
• 10:34
  the
• 10:34
  open-source
• 10:35
  side,
• 10:36
  Planner
• 10:36
  Agent
• 10:36
  V3
• 10:37
  for
• 10:37
  Open
• 10:37
  WebUI
• 10:37
  brings
• 10:38
  hierarchical
• 10:38
  sub-agent
• 10:39
  orchestration
• 10:40
  to
• 10:40
  self-hosted
• 10:40
  setups:
• 10:41
  a
• 10:41
  base
• 10:42
  model
• 10:42
  acts
• 10:42
  as
• 10:42
  Planner,
• 10:43
  decomposes
• 10:44
  tasks
• 10:44
  into
• 10:45
  a
• 10:45
  dynamic
• 10:45
  task
• 10:45
  tree,
• 10:46
  delegates
• 10:47
  to
• 10:47
  specialized
• 10:47
  virtual
• 10:48
  sub-agents
• 10:48
  (web
• 10:48
  search,
• 10:49
  image
• 10:49
  generation,
• 10:50
  knowledge/RAG,
• 10:52
  code
• 10:52
  interpreter,
• 10:53
  terminal),
• 10:53
  and
• 10:53
  executes
• 10:54
  in
• 10:54
  parallel
• 10:54
  via
• 10:55
  asyncio.gather.
• 10:58
  State
• 10:59
  persists
• 10:59
  across
• 10:59
  turns
• 11:00
  via
• 11:00
  JSON
• 11:00
  attachments,
• 11:01
  and
• 11:01
  MCP
• 11:02
  support
• 11:02
  includes
• 11:03
  deduplication
• 11:03
  and
• 11:04
  deadlock
• 11:04
  patches.
• 11:09
  The
• 11:09
  local
• 11:09
  AI
• 11:09
  story
• 11:10
  on
• 11:10
  Apple
• 11:10
  Silicon
• 11:10
  has
• 11:10
  quietly
• 11:11
  crossed
• 11:11
  from
• 11:11
  hobbyist
• 11:12
  tinkering
• 11:12
  into
• 11:13
  credible
• 11:13
  research
• 11:13
  infrastructure.
• 11:15
  The
• 11:15
  mac-code
• 11:15
  project
• 11:16
  demonstrates
• 11:16
  a
• 11:16
  Claude
• 11:17
  Code-style
• 11:17
  agentic
• 11:18
  coding
• 11:18
  assistant
• 11:18
  running
• 11:19
  a
• 11:19
  35B
• 11:19
  mixture-of-experts
• 11:20
  model
• 11:20
  at
• 11:20
  30
• 11:21
  tokens/second
• 11:22
  on
• 11:22
  a
• 11:22
  16GB
• 11:23
  Mac
• 11:23
  mini
• 11:23
  M4
• 11:23
  —
• 11:24
  entirely
• 11:25
  free,
• 11:25
  using
• 11:26
  llama.cpp
• 11:27
  with
• 11:27
  Unsloth's
• 11:28
  IQ2
• 11:28
  M
• 11:29
  quantization.
• 11:30
  The
• 11:30
  more
• 11:30
  technically
• 11:31
  interesting
• 11:31
  contribution
• 11:32
  is
• 11:32
  Flash
• 11:32
  Streaming:
• 11:33
  a
• 11:33
  technique
• 11:34
  that
• 11:34
  splits
• 11:34
  models
• 11:35
  by
• 11:35
  access
• 11:35
  pattern,
• 11:36
  pinning
• 11:36
  attention
• 11:37
  weights
• 11:37
  and
• 11:37
  embeddings
• 11:37
  in
• 11:37
  RAM
• 11:38
  (4-6GB)
• 11:39
  while
• 11:39
  streaming
• 11:39
  FFN
• 11:40
  weights
• 11:40
  from
• 11:40
  SSD
• 11:41
  per
• 11:41
  token,
• 11:41
  layer
• 11:42
  by
• 11:42
  layer,
• 11:42
  discarding
• 11:43
  after
• 11:43
  each
• 11:43
  matmul.
• 11:44
  For
• 11:45
  MoE
• 11:45
  models,
• 11:46
  only
• 11:46
  the
• 11:46
  8
• 11:46
  active
• 11:47
  experts
• 11:47
  (~14MB)
• 11:48
  load
• 11:48
  per
• 11:48
  token
• 11:48
  instead
• 11:49
  of
• 11:49
  all
• 11:49
  256
• 11:51
  (~460MB),
• 11:53
  yielding
• 11:53
  10x
• 11:53
  speedup
• 11:54
  over
• 11:54
  dense
• 11:54
  streaming.
• 11:55
  The
• 11:55
  project
• 11:56
  runs
• 11:56
  a
• 11:56
  22GB
• 11:57
  model
• 11:57
  on
• 11:57
  16GB
• 11:58
  of
• 11:58
  RAM
• 11:58
  at
• 11:58
  1.54
• 11:59
  tok/s
• 12:00
  with
• 12:00
  1.42GB
• 12:03
  resident
• 12:03
  memory,
• 12:04
  and
• 12:04
  a
• 12:04
  batched
• 12:04
  Union-of-Experts
• 12:05
  prototype
• 12:05
  hits
• 12:06
  5.1
• 12:06
  tok/s
• 12:07
  by
• 12:07
  computing
• 12:07
  the
• 12:08
  set
• 12:08
  union
• 12:08
  of
• 12:08
  active
• 12:08
  experts
• 12:09
  across
• 12:09
  8
• 12:09
  verification
• 12:10
  tokens.
• 12:10
  Key
• 12:11
  discovery:
• 12:11
  F
• 12:12
  NOCACHE
• 12:14
  direct
• 12:14
  I/O
• 12:15
  is
• 12:15
  27%
• 12:16
  faster
• 12:16
  than
• 12:16
  mmap
• 12:16
  on
• 12:17
  macOS
• 12:18
  because
• 12:18
  it
• 12:18
  bypasses
• 12:19
  the
• 12:19
  Unified
• 12:19
  Buffer
• 12:19
  Cache.
• 12:20
  Karpathy's
• 12:21
  autoresearch
• 12:21
  concept
• 12:22
  has
• 12:22
  been
• 12:22
  ported
• 12:22
  to
• 12:23
  MLX,
• 12:23
  enabling
• 12:23
  autonomous
• 12:24
  research
• 12:24
  loops
• 12:24
  natively
• 12:25
  on
• 12:25
  Apple
• 12:25
  Silicon
• 12:25
  with
• 12:26
  no
• 12:26
  PyTorch
• 12:26
  or
• 12:26
  CUDA
• 12:26
  dependency.
• 12:27
  The
• 12:28
  protocol
• 12:28
  is
• 12:28
  elegant:
• 12:29
  one
• 12:29
  mutable
• 12:29
  train.py,
• 12:30
  one
• 12:31
  metric
• 12:31
  (val
• 12:31
  bpb),
• 12:32
  a
• 12:32
  fixed
• 12:32
  5-minute
• 12:33
  training
• 12:33
  budget,
• 12:33
  and
• 12:33
  keep-or-revert
• 12:34
  via
• 12:34
  git.
• 12:35
  Overnight
• 12:36
  runs
• 12:36
  on
• 12:36
  M4
• 12:37
  Max
• 12:37
  pushed
• 12:37
  val
• 12:38
  bpb
• 12:38
  from
• 12:38
  1.807
• 12:40
  down
• 12:40
  to
• 12:40
  1.294,
• 12:42
  with
• 12:42
  the
• 12:42
  Mac
• 12:42
  Mini
• 12:42
  finding
• 12:43
  a
• 12:43
  meaningfully
• 12:43
  different
• 12:44
  optimal
• 12:44
  configuration
• 12:45
  than
• 12:45
  the
• 12:45
  Max-class
• 12:45
  machines
• 12:46
  —
• 12:47
  more
• 12:47
  aggressive
• 12:47
  step-efficiency
• 12:48
  wins,
• 12:49
  a
• 12:49
  hardware-specific
• 12:49
  behavior
• 12:50
  the
• 12:50
  loop
• 12:50
  is
• 12:50
  designed
• 12:51
  to
• 12:51
  surface.
• 12:52
  On
• 12:52
  the
• 12:52
  quantization
• 12:53
  front,
• 12:53
  an
• 12:54
  adaptation
• 12:54
  of
• 12:54
  the
• 12:54
  TurboQuant
• 12:55
  algorithm
• 12:55
  from
• 12:55
  KV-cache
• 12:56
  quantization
• 12:57
  to
• 12:57
  model
• 12:57
  weight
• 12:57
  compression
• 12:58
  achieved
• 12:58
  effectively
• 12:59
  lossless
• 12:59
  4+4
• 13:00
  residual
• 13:00
  coding
• 13:01
  at
• 13:01
  8
• 13:01
  total
• 13:01
  bits:
• 13:02
  Qwen3.5-4B
• 13:05
  scored
• 13:05
  10.70
• 13:06
  perplexity
• 13:07
  versus
• 13:07
  10.67
• 13:08
  baseline,
• 13:09
  with
• 13:09
  KLD
• 13:10
  of
• 13:10
  just
• 13:10
  0.0028.
• 13:13
  The
• 13:13
  community
• 13:13
  is
• 13:13
  already
• 13:14
  testing
• 13:14
  llama.cpp
• 13:16
  integration
• 13:16
  via
• 13:17
  Walsh-Hadamard
• 13:17
  transform
• 13:18
  fixes.
• 13:19
  Alibaba
• 13:19
  released
• 13:20
  Copaw-9B,
• 13:21
  an
• 13:21
  official
• 13:21
  agentic
• 13:22
  fine-tune
• 13:22
  of
• 13:22
  Qwen3.5
• 13:24
  9B
• 13:26
  that
• 13:26
  benchmarks
• 13:26
  on
• 13:26
  par
• 13:27
  with
• 13:27
  the
• 13:27
  much
• 13:27
  larger
• 13:27
  Qwen3.5-Plus
• 13:30
  on
• 13:30
  several
• 13:31
  tasks
• 13:31
  .
• 13:32
  And
• 13:32
  Unsloth's
• 13:32
  Qwen3.5-27B
• 13:35
  fine-tuned
• 13:36
  on
• 13:36
  distilled
• 13:36
  Claude
• 13:37
  Opus
• 13:37
  4.6
• 13:38
  reasoning
• 13:38
  data
• 13:38
  has
• 13:38
  held
• 13:39
  the
• 13:39
  #1
• 13:39
  trending
• 13:39
  spot
• 13:40
  on
• 13:40
  Hugging
• 13:40
  Face
• 13:41
  for
• 13:41
  three
• 13:41
  weeks
• 13:41
  running,
• 13:42
  hitting
• 13:42
  36
• 13:43
  tok/s
• 13:43
  on
• 13:43
  an
• 13:44
  RTX
• 13:44
  3090
• 13:45
  with
• 13:45
  Q4
• 13:46
  K
• 13:46
  M
• 13:46
  quantization
• 13:47
  at
• 13:47
  128K
• 13:48
  context.
• 13:51
  A
• 13:52
  v3
• 13:52
  medical
• 13:52
  speech-to-text
• 13:53
  benchmark
• 13:53
  across
• 13:54
  31
• 13:54
  models
• 13:55
  on
• 13:55
  the
• 13:55
  PriMock57
• 13:56
  dataset
• 13:56
  (55
• 13:57
  doctor-patient
• 13:58
  consultations,
• 13:59
  ~80K
• 14:00
  words
• 14:00
  of
• 14:00
  British
• 14:00
  English
• 14:01
  medical
• 14:01
  dialogue)
• 14:02
  crowns
• 14:02
  Microsoft's
• 14:03
  VibeVoice-ASR
• 14:03
  9B
• 14:04
  as
• 14:04
  the
• 14:04
  new
• 14:05
  open-source
• 14:05
  leader
• 14:05
  at
• 14:05
  8.34%
• 14:07
  word
• 14:07
  error
• 14:07
  rate
• 14:07
  —
• 14:08
  nearly
• 14:08
  matching
• 14:08
  Gemini
• 14:09
  2.5
• 14:09
  Pro's
• 14:10
  8.15%.
• 14:11
  The
• 14:11
  catch:
• 14:12
  VibeVoice
• 14:13
  needs
• 14:13
  ~18GB
• 14:14
  VRAM
• 14:14
  and
• 14:14
  processes
• 14:15
  at
• 14:15
  97
• 14:15
  seconds
• 14:16
  per
• 14:16
  file
• 14:16
  even
• 14:17
  on
• 14:17
  an
• 14:17
  H100.
• 14:18
  The
• 14:18
  real
• 14:19
  edge
• 14:19
  story
• 14:19
  is
• 14:19
  NVIDIA's
• 14:20
  Parakeet
• 14:20
  TDT
• 14:21
  0.6B
• 14:22
  v3:
• 14:23
  9.35%
• 14:24
  WER
• 14:24
  at
• 14:24
  6
• 14:25
  seconds
• 14:25
  per
• 14:25
  file
• 14:25
  on
• 14:25
  Apple
• 14:26
  Silicon,
• 14:26
  a
• 14:26
  0.6B
• 14:27
  model
• 14:28
  within
• 14:28
  1%
• 14:28
  of
• 14:28
  a
• 14:28
  9B
• 14:29
  model.
• 14:30
  The
• 14:30
  benchmark
• 14:30
  also
• 14:30
  uncovered
• 14:31
  bugs
• 14:31
  in
• 14:31
  Whisper's
• 14:31
  text
• 14:32
  normalizer
• 14:32
  that
• 14:32
  were
• 14:32
  quietly
• 14:33
  inflating
• 14:33
  WER
• 14:33
  by
• 14:34
  2-3%
• 14:34
  across
• 14:35
  every
• 14:35
  model
• 14:35
  tested
• 14:35
  —
• 14:36
  "oh"
• 14:37
  was
• 14:37
  treated
• 14:37
  as
• 14:37
  zero,
• 14:38
  and
• 14:38
  common
• 14:38
  equivalences
• 14:39
  (ok/okay,
• 14:40
  yeah/yep/yes)
• 14:41
  weren't
• 14:42
  normalized.
• 14:43
  On
• 14:43
  the
• 14:43
  generation
• 14:44
  side,
• 14:44
  a
• 14:44
  reverse-engineering
• 14:45
  effort
• 14:46
  recovered
• 14:46
  the
• 14:46
  missing
• 14:46
  codec
• 14:47
  encoder
• 14:47
  weights
• 14:47
  from
• 14:47
  Mistral's
• 14:48
  open-source
• 14:48
  Voxtral
• 14:49
  4B
• 14:49
  TTS
• 14:50
  model,
• 14:50
  unblocking
• 14:51
  the
• 14:51
  reference
• 14:51
  audio
• 14:52
  pass
• 14:52
  required
• 14:52
  for
• 14:52
  zero-shot
• 14:53
  voice
• 14:53
  cloning.
• 14:54
  The
• 14:54
  community
• 14:55
  response
• 14:55
  was
• 14:55
  immediate
• 14:56
  interest
• 14:56
  in
• 14:56
  fine-tuning
• 14:57
  on
• 14:57
  larger
• 14:57
  single-voice
• 14:58
  datasets
• 14:58
  rather
• 14:59
  than
• 14:59
  relying
• 14:59
  on
• 14:59
  zero-shot
• 15:00
  capability.
• 15:01
  NOFX
• 15:02
  is
• 15:02
  an
• 15:02
  open-source
• 15:03
  autonomous
• 15:03
  AI
• 15:03
  trading
• 15:04
  assistant
• 15:04
  built
• 15:04
  in
• 15:04
  Go
• 15:04
  and
• 15:04
  React
• 15:05
  with
• 15:05
  a
• 15:05
  genuinely
• 15:06
  novel
• 15:06
  payment
• 15:06
  model:
• 15:07
  x402
• 15:08
  micropayments
• 15:09
  via
• 15:09
  USDC
• 15:10
  wallet
• 15:10
  instead
• 15:10
  of
• 15:10
  API
• 15:11
  keys.
• 15:12
  The
• 15:12
  system
• 15:13
  supports
• 15:13
  15+
• 15:14
  AI
• 15:14
  models
• 15:14
  via
• 15:15
  the
• 15:15
  Claw402
• 15:16
  gateway,
• 15:17
  9
• 15:17
  exchanges
• 15:18
  (including
• 15:18
  perp-DEXes
• 15:19
  Hyperliquid,
• 15:20
  Aster,
• 15:20
  and
• 15:20
  Lighter),
• 15:21
  and
• 15:21
  features
• 15:21
  a
• 15:21
  Strategy
• 15:22
  Studio
• 15:22
  visual
• 15:22
  builder,
• 15:23
  AI
• 15:23
  competition
• 15:24
  mode
• 15:24
  with
• 15:24
  leaderboards,
• 15:25
  and
• 15:25
  a
• 15:25
  Telegram
• 15:26
  agent
• 15:26
  with
• 15:27
  streaming
• 15:27
  and
• 15:27
  memory.
• 15:28
  The
• 15:28
  "fully
• 15:29
  autonomous"
• 15:29
  claim
• 15:30
  means
• 15:30
  the
• 15:30
  AI
• 15:31
  selects
• 15:31
  which
• 15:31
  model
• 15:32
  to
• 15:32
  use,
• 15:33
  what
• 15:33
  market
• 15:33
  data
• 15:33
  to
• 15:33
  pull,
• 15:34
  and
• 15:34
  when
• 15:34
  to
• 15:34
  trade
• 15:35
  —
• 15:35
  the
• 15:36
  user
• 15:36
  sets
• 15:36
  strategy
• 15:37
  and
• 15:37
  funds
• 15:37
  the
• 15:37
  wallet.
• 15:38
  Reco.ai
• 15:39
  reported
• 15:40
  rewriting
• 15:40
  their
• 15:40
  JSONata
• 15:41
  evaluation
• 15:41
  engine
• 15:42
  with
• 15:42
  AI
• 15:42
  assistance
• 15:43
  in
• 15:43
  a
• 15:43
  single
• 15:44
  day,
• 15:44
  saving
• 15:45
  $500K
• 15:46
  annually.
• 15:47
  The
• 15:47
  original
• 15:48
  JSONata
• 15:48
  library
• 15:48
  —
• 15:49
  responsible
• 15:50
  for
• 15:50
  evaluating
• 15:51
  policy
• 15:51
  expressions
• 15:52
  across
• 15:52
  their
• 15:52
  SaaS
• 15:52
  security
• 15:53
  platform
• 15:53
  —
• 15:54
  had
• 15:54
  become
• 15:54
  a
• 15:54
  performance
• 15:55
  bottleneck
• 15:55
  processing
• 15:56
  millions
• 15:57
  of
• 15:57
  rule
• 15:57
  evaluations
• 15:58
  daily.
• 15:59
  The
• 15:59
  AI-assisted
• 16:00
  rewrite
• 16:00
  replaced
• 16:00
  interpreted
• 16:01
  evaluation
• 16:02
  with
• 16:02
  compiled,
• 16:02
  optimized
• 16:03
  code
• 16:03
  paths,
• 16:04
  turning
• 16:05
  what
• 16:05
  would
• 16:05
  have
• 16:05
  been
• 16:05
  a
• 16:05
  multi-quarter
• 16:06
  engineering
• 16:06
  project
• 16:07
  into
• 16:07
  a
• 16:07
  weekend
• 16:07
  exercise
• 16:08
  with
• 16:08
  production-grade
• 16:09
  results.