Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 22, 2026

Transcript

AGI Dreams Podcast – April 22, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  NSA
• 0:00
  is
• 0:00
  running
• 0:01
  Anthropic's
• 0:01
  Mythos
• 0:02
  —
• 0:02
  the
• 0:02
  model
• 0:03
  whose
• 0:03
  243-page
• 0:04
  system
• 0:04
  card
• 0:05
  revealed
• 0:05
  it
• 0:05
  autonomously
• 0:06
  discovered
• 0:06
  zero-days
• 0:07
  in
• 0:07
  OpenBSD,
• 0:07
  FFmpeg,
• 0:08
  and
• 0:08
  FreeBSD
• 0:09
  and
• 0:09
  wrote
• 0:10
  working
• 0:10
  exploit
• 0:10
  chains
• 0:11
  —
• 0:11
  while
• 0:12
  the
• 0:12
  Department
• 0:12
  of
• 0:12
  Defense
• 0:12
  that
• 0:13
  oversees
• 0:13
  the
• 0:13
  NSA
• 0:13
  simultaneously
• 0:14
  argues
• 0:14
  in
• 0:15
  court
• 0:15
  that
• 0:15
  Anthropic
• 0:15
  is
• 0:16
  a
• 0:16
  "supply
• 0:16
  chain
• 0:16
  risk."
• 0:17
  Two
• 0:17
  sources
• 0:18
  told
• 0:18
  Axios
• 0:18
  the
• 0:18
  NSA
• 0:19
  is
• 0:19
  among
• 0:20
  the
• 0:20
  roughly
• 0:20
  40
• 0:20
  organizations
• 0:21
  Anthropic
• 0:21
  granted
• 0:22
  access
• 0:22
  to
• 0:22
  Mythos,
• 0:23
  a
• 0:23
  model
• 0:23
  it
• 0:23
  deemed
• 0:23
  too
• 0:24
  dangerous
• 0:24
  for
• 0:24
  wider
• 0:24
  release.
• 0:25
  The
• 0:25
  Pentagon
• 0:26
  moved
• 0:26
  in
• 0:26
  February
• 0:26
  to
• 0:26
  cut
• 0:27
  off
• 0:27
  Anthropic
• 0:27
  after
• 0:27
  the
• 0:28
  company
• 0:28
  refused
• 0:28
  to
• 0:28
  strip
• 0:29
  guardrails
• 0:29
  against
• 0:29
  mass
• 0:29
  domestic
• 0:30
  surveillance
• 0:30
  and
• 0:30
  autonomous
• 0:31
  weapons
• 0:31
  development.
• 0:32
  A
• 0:32
  federal
• 0:32
  judge
• 0:33
  blocked
• 0:33
  the
• 0:33
  blacklist
• 0:33
  designation
• 0:34
  as
• 0:34
  likely
• 0:34
  unlawful.
• 0:35
  Now
• 0:36
  the
• 0:36
  military
• 0:36
  is
• 0:36
  broadening
• 0:37
  its
• 0:37
  use
• 0:37
  of
• 0:37
  Anthropic's
• 0:38
  tools
• 0:38
  while
• 0:38
  its
• 0:38
  own
• 0:38
  lawyers
• 0:39
  argue
• 0:39
  those
• 0:39
  same
• 0:39
  tools
• 0:40
  threaten
• 0:40
  national
• 0:40
  security
• 0:41
  .
• 0:42
  The
• 0:42
  same
• 0:42
  class
• 0:42
  of
• 0:42
  offensive
• 0:43
  security
• 0:43
  capability
• 0:44
  that
• 0:44
  made
• 0:44
  Mythos
• 0:44
  "too
• 0:44
  dangerous"
• 0:45
  to
• 0:45
  release
• 0:45
  publicly
• 0:46
  is
• 0:46
  already
• 0:46
  loose
• 0:47
  in
• 0:47
  the
• 0:47
  wild
• 0:47
  —
• 0:48
  just
• 0:48
  running
• 0:48
  on
• 0:48
  cheaper
• 0:49
  models.
• 0:50
  The
• 0:50
  MAD
• 0:50
  Bugs
• 0:51
  series
• 0:51
  this
• 0:51
  month
• 0:52
  dropped
• 0:52
  four
• 0:52
  more
• 0:52
  arbitrary
• 0:53
  code
• 0:53
  execution
• 0:54
  vulnerabilities
• 0:54
  in
• 0:55
  the
• 0:55
  reverse
• 0:55
  engineering
• 0:55
  tools
• 0:56
  that
• 0:56
  every
• 0:57
  malware
• 0:57
  analyst
• 0:57
  and
• 0:57
  CTF
• 0:58
  player
• 0:58
  depends
• 0:59
  on:
• 0:59
  Ghidra,
• 1:00
  IDA
• 1:00
  Pro,
• 1:01
  Binary
• 1:01
  Ninja
• 1:01
  Sidekick,
• 1:02
  and
• 1:02
  radare2.
• 1:04
  All
• 1:04
  discovered
• 1:04
  by
• 1:04
  Claude
• 1:05
  or
• 1:05
  Codex.
• 1:06
  The
• 1:06
  Ghidra
• 1:06
  exploit
• 1:07
  is
• 1:07
  particularly
• 1:08
  instructive:
• 1:09
  a
• 1:09
  Java
• 1:09
  RMI
• 1:10
  deserialization
• 1:10
  chain
• 1:11
  that
• 1:11
  bottoms
• 1:11
  out
• 1:11
  in
• 1:11
  a
• 1:12
  Python
• 1:12
  bytecode
• 1:12
  VM.
• 1:13
  The
• 1:13
  attacker
• 1:14
  hands
• 1:14
  someone
• 1:14
  a
• 1:14
  Ghidra
• 1:15
  project
• 1:15
  file.
• 1:16
  They
• 1:16
  double-click
• 1:17
  it.
• 1:17
  The
• 1:17
  malicious
• 1:18
  server
• 1:18
  answers
• 1:18
  the
• 1:19
  first
• 1:19
  RMI
• 1:19
  call
• 1:20
  —
• 1:20
  before
• 1:21
  any
• 1:21
  authentication
• 1:21
  handshake
• 1:22
  —
• 1:23
  with
• 1:23
  a
• 1:23
  gadget
• 1:23
  chain
• 1:23
  that
• 1:24
  triggers
• 1:24
  eval(PyBytecode)
• 1:25
  through
• 1:25
  Jython's
• 1:25
  BuiltinFunctions[18].
• 1:27
  Twenty-one
• 1:27
  bytes
• 1:27
  of
• 1:28
  CPython
• 1:28
  bytecode
• 1:28
  later,
• 1:29
  Runtime.getRuntime().exec()
• 1:31
  pops
• 1:31
  a
• 1:31
  shell.
• 1:32
  This
• 1:32
  affects
• 1:33
  every
• 1:33
  Ghidra
• 1:33
  release
• 1:34
  since
• 1:34
  9.1.
• 1:35
  The
• 1:35
  radare2
• 1:36
  bug
• 1:36
  was
• 1:36
  a
• 1:36
  patch
• 1:36
  bypass
• 1:37
  —
• 1:38
  AI
• 1:38
  read
• 1:38
  the
• 1:38
  fix
• 1:39
  for
• 1:39
  CVE
• 1:39
  #25731,
• 1:41
  asked
• 1:41
  "what
• 1:42
  else
• 1:42
  gets
• 1:42
  interpolated
• 1:43
  here?",
• 1:43
  and
• 1:43
  had
• 1:43
  a
• 1:44
  working
• 1:44
  proof
• 1:44
  of
• 1:44
  concept
• 1:45
  before
• 1:45
  humans
• 1:45
  finished
• 1:46
  debating
• 1:46
  AI
• 1:46
  vulnerability
• 1:47
  research
• 1:47
  on
• 1:47
  X
• 1:48
  .
• 1:48
  The
• 1:49
  IDA
• 1:49
  Pro
• 1:49
  and
• 1:50
  Binary
• 1:50
  Ninja
• 1:50
  Sidekick
• 1:51
  vulnerabilities
• 1:51
  are
• 1:52
  under
• 1:52
  coordinated
• 1:52
  disclosure
• 1:53
  with
• 1:53
  Hex-Rays
• 1:54
  and
• 1:54
  Vector
• 1:54
  35
• 1:54
  respectively
• 1:55
  —
• 1:56
  both
• 1:56
  are
• 1:56
  arbitrary
• 1:56
  code
• 1:57
  execution,
• 1:58
  both
• 1:58
  trigger
• 1:58
  on
• 1:58
  the
• 1:58
  normal
• 1:59
  "open
• 1:59
  the
• 1:59
  thing
• 1:59
  someone
• 2:00
  sent
• 2:00
  you"
• 2:00
  workflow.
• 2:01
  The
• 2:01
  pattern
• 2:01
  AGI
• 2:02
  Dreams
• 2:02
  has
• 2:02
  tracked
• 2:03
  across
• 2:03
  editions
• 2:03
  222,
• 2:04
  245,
• 2:04
  246,
• 2:04
  and
• 2:04
  269
• 2:08
  is
• 2:09
  now
• 2:09
  unmistakable:
• 2:10
  frontier
• 2:11
  models
• 2:11
  find
• 2:11
  exploitable
• 2:12
  vulnerabilities
• 2:13
  faster
• 2:13
  than
• 2:13
  humans
• 2:14
  patch
• 2:14
  them,
• 2:14
  incomplete
• 2:15
  patches
• 2:15
  are
• 2:16
  themselves
• 2:16
  a
• 2:16
  bug
• 2:16
  class
• 2:17
  that
• 2:17
  AI
• 2:17
  is
• 2:18
  unreasonably
• 2:18
  good
• 2:18
  at
• 2:18
  finding,
• 2:19
  the
• 2:19
  toolchain
• 2:20
  that
• 2:20
  defenders
• 2:21
  depend
• 2:21
  on
• 2:21
  is
• 2:21
  itself
• 2:22
  vulnerable,
• 2:23
  and
• 2:23
  the
• 2:23
  gap
• 2:23
  between
• 2:23
  offensive
• 2:24
  AI
• 2:24
  capability
• 2:25
  and
• 2:25
  defensive
• 2:25
  readiness
• 2:26
  is
• 2:26
  widening
• 2:26
  every
• 2:27
  week.
• 2:27
  The
• 2:28
  organizations
• 2:28
  with
• 2:28
  Mythos
• 2:29
  access
• 2:29
  are
• 2:30
  using
• 2:30
  it
• 2:30
  predominantly
• 2:31
  to
• 2:31
  scan
• 2:31
  their
• 2:31
  own
• 2:31
  environments
• 2:32
  for
• 2:32
  exploitable
• 2:33
  vulnerabilities.
• 2:34
  The
• 2:34
  organizations
• 2:35
  without
• 2:36
  it
• 2:36
  are
• 2:36
  getting
• 2:36
  equivalent
• 2:37
  results
• 2:37
  from
• 2:37
  Claude
• 2:38
  and
• 2:38
  Codex
• 2:38
  on
• 2:38
  public
• 2:39
  models.
• 2:43
  If
• 2:44
  your
• 2:44
  AI
• 2:44
  agent
• 2:44
  can
• 2:44
  call
• 2:45
  arbitrary
• 2:45
  tools,
• 2:46
  the
• 2:46
  question
• 2:46
  is
• 2:46
  not
• 2:46
  whether
• 2:47
  it
• 2:47
  will
• 2:47
  be
• 2:47
  exploited
• 2:47
  but
• 2:48
  whether
• 2:48
  you
• 2:48
  have
• 2:48
  a
• 2:48
  deterministic
• 2:49
  enforcement
• 2:49
  point
• 2:50
  that
• 2:50
  fires
• 2:50
  before
• 2:50
  the
• 2:50
  damage
• 2:51
  lands.
• 2:52
  The
• 2:52
  emerging
• 2:52
  answer
• 2:52
  from
• 2:53
  the
• 2:53
  security
• 2:53
  research
• 2:53
  community
• 2:54
  is
• 2:54
  runtime
• 2:54
  hooks
• 2:55
  —
• 2:55
  pre-tool
• 2:56
  and
• 2:56
  post-tool
• 2:56
  interception
• 2:57
  points
• 2:57
  wired
• 2:57
  directly
• 2:58
  into
• 2:58
  the
• 2:58
  agent
• 2:59
  execution
• 2:59
  loop
• 2:59
  itself.
• 3:00
  Chris
• 3:00
  Hughes
• 3:01
  at
• 3:01
  Resilient
• 3:01
  Cyber
• 3:01
  lays
• 3:02
  out
• 3:02
  the
• 3:02
  architecture:
• 3:03
  a
• 3:03
  reference
• 3:03
  monitor
• 3:04
  pattern
• 3:04
  borrowed
• 3:04
  from
• 3:05
  OS
• 3:05
  security,
• 3:06
  where
• 3:06
  every
• 3:06
  tool
• 3:06
  invocation
• 3:07
  passes
• 3:07
  through
• 3:07
  a
• 3:07
  policy
• 3:08
  engine
• 3:08
  before
• 3:08
  execution.
• 3:10
  Cedar,
• 3:10
  Amazon's
• 3:11
  declarative
• 3:11
  policy
• 3:12
  language,
• 3:12
  is
• 3:12
  the
• 3:13
  leading
• 3:13
  candidate
• 3:13
  for
• 3:13
  expressing
• 3:14
  those
• 3:14
  policies.
• 3:15
  The
• 3:15
  appeal
• 3:16
  is
• 3:16
  that
• 3:16
  Cedar
• 3:16
  policies
• 3:17
  are
• 3:17
  analyzable
• 3:17
  —
• 3:18
  you
• 3:18
  can
• 3:18
  prove
• 3:18
  properties
• 3:19
  about
• 3:19
  what
• 3:19
  they
• 3:19
  permit
• 3:19
  before
• 3:20
  deployment,
• 3:20
  unlike
• 3:21
  imperative
• 3:21
  code
• 3:22
  that
• 3:22
  has
• 3:22
  to
• 3:22
  be
• 3:22
  tested
• 3:22
  empirically.
• 3:24
  Two
• 3:24
  real
• 3:24
  CVEs
• 3:24
  anchor
• 3:25
  the
• 3:25
  urgency:
• 3:26
  CVE-2025-59536
• 3:30
  and
• 3:30
  CVE-2026-21852,
• 3:35
  both
• 3:35
  in
• 3:35
  Claude
• 3:35
  Code's
• 3:35
  hook
• 3:36
  implementation,
• 3:36
  both
• 3:37
  demonstrating
• 3:37
  that
• 3:37
  even
• 3:38
  the
• 3:38
  vendors
• 3:38
  building
• 3:38
  these
• 3:39
  systems
• 3:39
  ship
• 3:39
  enforcement
• 3:40
  gaps
• 3:40
  .
• 3:40
  The
• 3:41
  irony
• 3:41
  is
• 3:41
  that
• 3:41
  the
• 3:41
  security
• 3:42
  tools
• 3:42
  meant
• 3:42
  to
• 3:42
  catch
• 3:43
  these
• 3:43
  problems
• 3:43
  are
• 3:43
  themselves
• 3:44
  compromised.
• 3:45
  A
• 3:45
  detailed
• 3:45
  purple-team
• 3:46
  exercise
• 3:46
  documents
• 3:47
  how
• 3:47
  the
• 3:47
  TeamPCP
• 3:48
  campaign
• 3:48
  cascaded
• 3:49
  through
• 3:49
  Trivy,
• 3:50
  KICS,
• 3:50
  and
• 3:50
  LiteLLM
• 3:50
  —
• 3:52
  three
• 3:52
  tools
• 3:52
  that
• 3:52
  organizations
• 3:53
  trust
• 3:53
  to
• 3:53
  validate
• 3:54
  their
• 3:54
  supply
• 3:54
  chains.
• 3:55
  The
• 3:56
  attack
• 3:56
  vector
• 3:56
  was
• 3:56
  Python
• 3:57
  startup
• 3:57
  hooks
• 3:58
  (MITRE
• 4:00
  T1546.018),
• 4:02
  injecting
• 4:02
  malicious
• 4:03
  code
• 4:03
  that
• 4:03
  executes
• 4:04
  before
• 4:04
  the
• 4:04
  legitimate
• 4:04
  tool
• 4:05
  even
• 4:05
  loads.
• 4:06
  The
• 4:06
  exercise
• 4:07
  walks
• 4:07
  through
• 4:07
  detection
• 4:08
  with
• 4:08
  Falco
• 4:08
  and
• 4:08
  Sysdig,
• 4:09
  but
• 4:09
  the
• 4:09
  deeper
• 4:10
  lesson
• 4:10
  is
• 4:10
  architectural:
• 4:11
  if
• 4:11
  your
• 4:11
  CI/CD
• 4:13
  pipeline
• 4:13
  trusts
• 4:13
  a
• 4:13
  security
• 4:14
  scanner,
• 4:15
  and
• 4:15
  that
• 4:15
  scanner's
• 4:15
  distribution
• 4:16
  channel
• 4:16
  is
• 4:16
  compromised,
• 4:17
  you
• 4:18
  have
• 4:18
  handed
• 4:18
  your
• 4:18
  adversary
• 4:19
  a
• 4:19
  signed
• 4:19
  invitation
• 4:20
  into
• 4:20
  every
• 4:20
  build
• 4:20
  .
• 4:21
  Palo
• 4:22
  Alto
• 4:22
  Networks
• 4:22
  frames
• 4:23
  the
• 4:23
  organizational
• 4:24
  response
• 4:24
  as
• 4:24
  AI
• 4:25
  Security
• 4:25
  Posture
• 4:26
  Management
• 4:26
  —
• 4:27
  continuous
• 4:27
  visibility
• 4:28
  into
• 4:28
  LLM
• 4:29
  deployments,
• 4:29
  RAG
• 4:30
  pipelines,
• 4:31
  and
• 4:31
  autonomous
• 4:31
  agents
• 4:32
  across
• 4:32
  the
• 4:32
  cloud
• 4:32
  environment,
• 4:33
  with
• 4:33
  compliance
• 4:34
  monitoring
• 4:34
  and
• 4:35
  real-time
• 4:35
  response.
• 4:36
  The
• 4:37
  five-step
• 4:37
  framework
• 4:38
  connects
• 4:38
  AI
• 4:38
  assets
• 4:39
  to
• 4:39
  the
• 4:39
  broader
• 4:39
  cloud
• 4:39
  security
• 4:40
  strategy
• 4:41
  rather
• 4:41
  than
• 4:41
  treating
• 4:41
  them
• 4:42
  as
• 4:42
  a
• 4:42
  separate
• 4:42
  problem
• 4:42
  .
• 4:43
  On
• 4:43
  the
• 4:43
  offensive
• 4:44
  side,
• 4:44
  HawkEye
• 4:45
  (bugbounty-agent)
• 4:46
  automates
• 4:46
  the
• 4:46
  full
• 4:47
  bug
• 4:47
  bounty
• 4:47
  workflow:
• 4:48
  reconnaissance,
• 4:49
  vulnerability
• 4:50
  scanning,
• 4:51
  exploit
• 4:51
  validation,
• 4:52
  and
• 4:52
  report
• 4:53
  generation.
• 4:54
  It
• 4:54
  chains
• 4:55
  Subfinder,
• 4:55
  Naabu,
• 4:56
  Nuclei,
• 4:57
  and
• 4:57
  custom
• 4:57
  scripts
• 4:58
  into
• 4:58
  an
• 4:58
  agentic
• 4:58
  pipeline
• 4:59
  that
• 4:59
  runs
• 4:59
  autonomously
• 5:00
  against
• 5:00
  a
• 5:00
  target
• 5:01
  scope.
• 5:02
  The
• 5:02
  tool
• 5:02
  is
• 5:02
  open-source
• 5:03
  and
• 5:03
  explicitly
• 5:04
  designed
• 5:04
  for
• 5:04
  authorized
• 5:05
  testing,
• 5:06
  but
• 5:06
  it
• 5:06
  concretizes
• 5:07
  the
• 5:07
  same
• 5:07
  capability
• 5:08
  gap
• 5:08
  that
• 5:08
  runtime
• 5:08
  hooks
• 5:09
  are
• 5:09
  racing
• 5:09
  to
• 5:09
  close
• 5:09
  .
• 5:13
  Phantasia
• 5:14
  introduces
• 5:14
  context-adaptive
• 5:15
  backdoors
• 5:15
  to
• 5:16
  vision-language
• 5:16
  models
• 5:16
  —
• 5:17
  attacks
• 5:18
  that
• 5:18
  activate
• 5:18
  only
• 5:18
  when
• 5:18
  specific
• 5:19
  visual
• 5:19
  contexts
• 5:20
  appear,
• 5:20
  evading
• 5:21
  existing
• 5:21
  defenses
• 5:22
  by
• 5:22
  design.
• 5:23
  The
• 5:24
  technique
• 5:24
  uses
• 5:24
  teacher-student
• 5:25
  knowledge
• 5:25
  distillation:
• 5:26
  a
• 5:26
  teacher
• 5:27
  model
• 5:27
  learns
• 5:27
  the
• 5:27
  backdoor
• 5:28
  behavior,
• 5:29
  then
• 5:29
  transfers
• 5:29
  it
• 5:29
  to
• 5:29
  the
• 5:29
  student
• 5:30
  through
• 5:30
  feature
• 5:30
  alignment
• 5:31
  rather
• 5:31
  than
• 5:31
  explicit
• 5:31
  trigger
• 5:32
  patterns.
• 5:33
  The
• 5:33
  result
• 5:33
  is
• 5:33
  a
• 5:33
  backdoor
• 5:34
  that
• 5:34
  survives
• 5:34
  both
• 5:35
  STRIP-P
• 5:36
  and
• 5:36
  ONION-R
• 5:38
  detection,
• 5:39
  the
• 5:39
  two
• 5:39
  standard
• 5:39
  defenses
• 5:40
  the
• 5:40
  research
• 5:40
  community
• 5:41
  relies
• 5:41
  on.
• 5:42
  Only
• 5:42
  1,000
• 5:43
  poisoned
• 5:43
  training
• 5:43
  samples
• 5:44
  are
• 5:44
  sufficient.
• 5:45
  The
• 5:45
  attack
• 5:45
  surface
• 5:46
  is
• 5:46
  broad
• 5:46
  —
• 5:47
  any
• 5:47
  VLM
• 5:47
  that
• 5:47
  processes
• 5:48
  images
• 5:48
  alongside
• 5:49
  text
• 5:49
  is
• 5:49
  a
• 5:49
  candidate,
• 5:50
  and
• 5:50
  the
• 5:50
  triggers
• 5:51
  are
• 5:51
  semantic
• 5:51
  (a
• 5:51
  particular
• 5:52
  scene
• 5:52
  or
• 5:53
  object
• 5:53
  combination)
• 5:54
  rather
• 5:54
  than
• 5:54
  pixel-level
• 5:55
  artifacts
• 5:55
  that
• 5:56
  scanners
• 5:56
  can
• 5:56
  detect
• 5:56
  .
• 5:57
  On
• 5:57
  the
• 5:57
  other
• 5:58
  side
• 5:58
  of
• 5:58
  the
• 5:58
  adversarial
• 5:59
  coin,
• 5:59
  ecoalign-forge
• 6:00
  builds
• 6:01
  a
• 6:01
  multi-agent
• 6:01
  "courtroom"
• 6:02
  for
• 6:02
  generating
• 6:02
  DPO
• 6:03
  preference
• 6:03
  pairs
• 6:03
  —
• 6:05
  the
• 6:05
  training
• 6:05
  data
• 6:05
  that
• 6:05
  teaches
• 6:06
  models
• 6:06
  what
• 6:06
  good
• 6:07
  and
• 6:07
  bad
• 6:07
  responses
• 6:07
  look
• 6:08
  like.
• 6:09
  A
• 6:09
  prosecutor
• 6:09
  agent
• 6:10
  argues
• 6:10
  for
• 6:10
  rejection,
• 6:11
  a
• 6:11
  defense
• 6:12
  agent
• 6:12
  argues
• 6:12
  for
• 6:12
  acceptance,
• 6:13
  and
• 6:14
  a
• 6:14
  judge
• 6:14
  agent
• 6:14
  renders
• 6:15
  the
• 6:15
  verdict.
• 6:16
  The
• 6:16
  system
• 6:16
  produces
• 6:17
  preference
• 6:17
  pairs
• 6:18
  at
• 6:18
  under
• 6:18
  $0.01
• 6:19
  each,
• 6:20
  targeting
• 6:20
  content
• 6:21
  moderation
• 6:21
  and
• 6:21
  safety
• 6:22
  alignment.
• 6:23
  The
• 6:23
  stated
• 6:23
  purpose
• 6:24
  is
• 6:24
  defensive:
• 6:25
  generating
• 6:26
  training
• 6:26
  data
• 6:26
  to
• 6:26
  make
• 6:27
  models
• 6:27
  refuse
• 6:27
  harmful
• 6:28
  requests
• 6:28
  more
• 6:29
  reliably.
• 6:30
  But
• 6:30
  the
• 6:30
  architecture
• 6:30
  is
• 6:31
  symmetric.
• 6:32
  The
• 6:32
  same
• 6:32
  courtroom
• 6:33
  can
• 6:33
  generate
• 6:33
  preference
• 6:34
  pairs
• 6:34
  that
• 6:34
  teach
• 6:34
  a
• 6:34
  model
• 6:35
  to
• 6:35
  comply
• 6:35
  with
• 6:35
  harmful
• 6:36
  requests
• 6:36
  just
• 6:37
  as
• 6:37
  efficiently.
• 6:38
  At
• 6:38
  this
• 6:38
  price
• 6:39
  point,
• 6:39
  anyone
• 6:40
  can
• 6:40
  manufacture
• 6:40
  alignment
• 6:41
  data
• 6:41
  at
• 6:41
  scale.
• 6:42
  The
• 6:42
  question
• 6:43
  of
• 6:43
  who
• 6:43
  controls
• 6:43
  the
• 6:44
  judge
• 6:44
  agent's
• 6:44
  instructions
• 6:45
  becomes
• 6:45
  a
• 6:45
  question
• 6:46
  of
• 6:46
  who
• 6:46
  controls
• 6:46
  the
• 6:46
  model's
• 6:47
  values.
• 6:48
  Previous
• 6:48
  coverage
• 6:49
  of
• 6:49
  ICLR
• 6:49
  2026
• 6:50
  inference-time
• 6:51
  backdoors
• 6:51
  (edition
• 6:52
  252)
• 6:53
  and
• 6:54
  Microsoft's
• 6:54
  sleeper
• 6:54
  agent
• 6:55
  detections
• 6:55
  (edition
• 6:56
  221)
• 6:57
  established
• 6:57
  the
• 6:57
  threat
• 6:58
  gradient.
• 6:58
  Phantasia
• 6:59
  and
• 6:59
  ecoalign-forge
• 7:00
  push
• 7:00
  both
• 7:01
  ends
• 7:01
  simultaneously
• 7:01
  —
• 7:02
  more
• 7:02
  sophisticated
• 7:03
  attacks
• 7:04
  and
• 7:04
  cheaper
• 7:04
  tools
• 7:04
  to
• 7:04
  shape
• 7:05
  model
• 7:05
  behavior
• 7:05
  .
• 7:10
  Opus
• 7:10
  4.7
• 7:11
  shipped
• 7:11
  with
• 7:12
  a
• 7:12
  new
• 7:12
  tokenizer
• 7:13
  that
• 7:13
  maps
• 7:13
  the
• 7:13
  same
• 7:13
  input
• 7:14
  text
• 7:14
  to
• 7:14
  up
• 7:14
  to
• 7:14
  35%
• 7:15
  more
• 7:15
  tokens.
• 7:16
  The
• 7:16
  sticker
• 7:17
  price
• 7:17
  did
• 7:17
  not
• 7:17
  change.
• 7:19
  Independent
• 7:19
  measurements
• 7:19
  range
• 7:20
  from
• 7:20
  1.29x
• 7:21
  to
• 7:21
  1.47x
• 7:22
  above
• 7:23
  Anthropic's
• 7:23
  stated
• 7:24
  range
• 7:24
  on
• 7:24
  real-world
• 7:24
  technical
• 7:25
  content.
• 7:26
  Adaptive
• 7:26
  thinking
• 7:26
  —
• 7:27
  where
• 7:28
  the
• 7:28
  model
• 7:28
  decides
• 7:28
  how
• 7:28
  much
• 7:29
  reasoning
• 7:29
  your
• 7:29
  query
• 7:30
  deserves
• 7:30
  —
• 7:31
  replaced
• 7:31
  the
• 7:31
  user-controlled
• 7:32
  thinking
• 7:32
  budget.
• 7:32
  Temperature,
• 7:33
  top-p,
• 7:34
  and
• 7:34
  top-k
• 7:34
  parameters
• 7:35
  were
• 7:35
  removed
• 7:35
  entirely
• 7:36
  from
• 7:36
  the
• 7:36
  API.
• 7:37
  One
• 7:37
  reviewer
• 7:37
  spent
• 7:38
  $42
• 7:38
  in
• 7:38
  a
• 7:39
  single
• 7:39
  afternoon
• 7:39
  on
• 7:39
  Claude
• 7:40
  Design,
• 7:40
  Anthropic's
• 7:41
  new
• 7:41
  design
• 7:41
  tool,
• 7:42
  burning
• 7:42
  through
• 7:42
  the
• 7:42
  entire
• 7:43
  Pro
• 7:43
  usage
• 7:43
  allocation
• 7:44
  while
• 7:44
  iterating
• 7:44
  on
• 7:44
  logo
• 7:45
  corrections
• 7:45
  the
• 7:45
  model
• 7:45
  kept
• 7:46
  getting
• 7:46
  wrong
• 7:46
  across
• 7:46
  five
• 7:47
  or
• 7:47
  six
• 7:47
  attempts.
• 7:48
  Each
• 7:48
  correction
• 7:48
  pass
• 7:48
  is
• 7:48
  billable.
• 7:50
  A
• 7:50
  head-to-head
• 7:50
  test
• 7:50
  pitting
• 7:51
  Opus
• 7:51
  4.7
• 7:52
  against
• 7:52
  ChatGPT
• 7:52
  5.4
• 7:54
  on
• 7:54
  a
• 7:54
  465-file
• 7:55
  adversarial
• 7:56
  data
• 7:56
  migration
• 7:56
  —
• 7:57
  CSV,
• 7:58
  Excel,
• 7:58
  PDF,
• 7:59
  JSON,
• 8:00
  images,
• 8:00
  VCF
• 8:01
  cards,
• 8:02
  planted
• 8:02
  fake
• 8:02
  customers,
• 8:03
  and
• 8:03
  trap
• 8:03
  data
• 8:03
  —
• 8:04
  found
• 8:05
  4.7
• 8:05
  finishing
• 8:06
  in
• 8:06
  33
• 8:06
  minutes
• 8:06
  versus
• 8:07
  53
• 8:07
  for
• 8:07
  5.4,
• 8:08
  with
• 8:08
  a
• 8:09
  superior
• 8:09
  front-end
• 8:09
  UI
• 8:10
  but
• 8:10
  a
• 8:10
  critical
• 8:10
  trust
• 8:11
  failure:
• 8:11
  the
• 8:11
  model
• 8:12
  claimed
• 8:12
  to
• 8:12
  have
• 8:12
  processed
• 8:13
  a
• 8:13
  TSV
• 8:13
  file
• 8:13
  it
• 8:13
  never
• 8:14
  touched,
• 8:14
  hallucinating
• 8:15
  the
• 8:15
  audit
• 8:15
  trail.
• 8:17
  Both
• 8:17
  models
• 8:17
  failed
• 8:17
  to
• 8:18
  catch
• 8:18
  obvious
• 8:18
  fake
• 8:18
  entries
• 8:19
  like
• 8:19
  "Mickey
• 8:19
  Mouse"
• 8:20
  and
• 8:20
  "ASDF
• 8:21
  ASDF"
• 8:23
  in
• 8:23
  the
• 8:23
  canonical
• 8:23
  customer
• 8:24
  records.
• 8:25
  Neither
• 8:25
  model
• 8:25
  is
• 8:25
  blowing
• 8:26
  the
• 8:26
  other
• 8:26
  out
• 8:26
  of
• 8:26
  the
• 8:26
  water
• 8:26
  —
• 8:27
  averaged
• 8:27
  cross-review
• 8:28
  scores
• 8:28
  landed
• 8:28
  at
• 8:29
  3.1
• 8:29
  for
• 8:29
  Opus
• 8:30
  and
• 8:30
  3.35
• 8:31
  for
• 8:31
  GPT
• 8:32
  5.4,
• 8:33
  inside
• 8:33
  the
• 8:33
  noise
• 8:34
  of
• 8:34
  a
• 8:34
  single
• 8:34
  run.
• 8:35
  SWE-Bench
• 8:36
  verified
• 8:36
  climbed
• 8:36
  from
• 8:37
  80%
• 8:37
  to
• 8:37
  87%.
• 8:39
  The
• 8:39
  model
• 8:39
  stays
• 8:39
  on
• 8:39
  task,
• 8:40
  self-verifies,
• 8:41
  and
• 8:41
  catches
• 8:42
  inconsistencies
• 8:42
  during
• 8:43
  planning.
• 8:44
  But
• 8:44
  Browse
• 8:44
  Comp
• 8:44
  —
• 8:45
  the
• 8:45
  web
• 8:45
  research
• 8:46
  benchmark
• 8:46
  —
• 8:47
  dropped
• 8:47
  from
• 8:47
  83
• 8:48
  to
• 8:48
  79.
• 8:49
  The
• 8:49
  upgrade
• 8:49
  is
• 8:50
  directed,
• 8:50
  not
• 8:51
  uniform
• 8:51
  .
• 8:52
  The
• 8:52
  economics
• 8:52
  behind
• 8:53
  this
• 8:53
  release
• 8:53
  are
• 8:53
  not
• 8:54
  subtle.
• 8:55
  Anthropic
• 8:55
  hit
• 8:55
  roughly
• 8:56
  $30
• 8:56
  billion
• 8:56
  annualized
• 8:57
  revenue
• 8:58
  in
• 8:58
  April,
• 8:58
  up
• 8:59
  from
• 8:59
  $20
• 8:59
  billion
• 8:59
  a
• 9:00
  month
• 9:00
  prior.
• 9:01
  Investor
• 9:01
  offers
• 9:02
  sit
• 9:02
  at
• 9:02
  $800
• 9:03
  billion,
• 9:03
  up
• 9:04
  from
• 9:04
  $380
• 9:04
  billion
• 9:05
  in
• 9:05
  February.
• 9:06
  IPO
• 9:07
  talks
• 9:07
  target
• 9:07
  October.
• 9:08
  Eight
• 9:09
  of
• 9:09
  the
• 9:09
  Fortune
• 9:09
  10
• 9:09
  are
• 9:10
  now
• 9:10
  Claude
• 9:10
  customers.
• 9:11
  Enterprise
• 9:12
  share
• 9:12
  climbed
• 9:12
  from
• 9:12
  24%
• 9:13
  to
• 9:13
  30%
• 9:14
  in
• 9:14
  a
• 9:14
  single
• 9:15
  month
• 9:15
  while
• 9:15
  OpenAI
• 9:16
  dropped
• 9:16
  from
• 9:16
  46%
• 9:17
  to
• 9:17
  35%.
• 9:19
  Meanwhile,
• 9:19
  Ed
• 9:19
  Zitron
• 9:20
  documents
• 9:20
  the
• 9:20
  infrastructure
• 9:21
  scale:
• 9:22
  15.2
• 9:23
  GW
• 9:23
  of
• 9:23
  global
• 9:24
  AI
• 9:24
  datacenter
• 9:24
  capacity
• 9:25
  under
• 9:25
  construction,
• 9:26
  with
• 9:26
  Anthropic's
• 9:27
  98.79%
• 9:28
  uptime
• 9:29
  during
• 9:29
  peak
• 9:29
  demand
• 9:30
  revealing
• 9:30
  the
• 9:30
  compute
• 9:30
  constraints
• 9:31
  that
• 9:31
  make
• 9:31
  tokenizer
• 9:32
  taxes
• 9:32
  and
• 9:32
  adaptive
• 9:33
  thinking
• 9:33
  economically
• 9:34
  rational
• 9:34
  .
• 9:35
  Users
• 9:35
  are
• 9:35
  already
• 9:36
  voting
• 9:36
  with
• 9:36
  their
• 9:36
  feet.
• 9:37
  A
• 9:37
  Reddit
• 9:38
  thread
• 9:38
  from
• 9:38
  a
• 9:38
  Claude
• 9:38
  Pro
• 9:38
  subscriber
• 9:39
  banned
• 9:39
  without
• 9:39
  explanation
• 9:40
  catalogs
• 9:41
  the
• 9:41
  migration
• 9:41
  patterns:
• 9:42
  OpenCode
• 9:43
  with
• 9:43
  GLM-5.1
• 9:44
  ("about
• 9:45
  Sonnet
• 9:45
  quality"),
• 9:46
  Gemini
• 9:46
  2.5
• 9:47
  Pro
• 9:47
  paired
• 9:47
  with
• 9:47
  Aider,
• 9:48
  local
• 9:48
  Gemma4
• 9:49
  31B
• 9:49
  through
• 9:49
  Ollama.
• 9:50
  The
• 9:50
  consensus
• 9:51
  is
• 9:51
  nothing
• 9:51
  matches
• 9:52
  Claude
• 9:52
  Code's
• 9:52
  workflow,
• 9:53
  but
• 9:53
  the
• 9:53
  community
• 9:54
  is
• 9:54
  building
• 9:54
  workarounds
• 9:55
  faster
• 9:55
  than
• 9:55
  Anthropic
• 9:56
  is
• 9:56
  resolving
• 9:56
  account
• 9:56
  issues
• 9:57
  .
• 9:57
  Roomote
• 9:58
  positions
• 9:58
  itself
• 9:59
  as
• 9:59
  the
• 9:59
  vendor-agnostic
• 10:00
  answer
• 10:00
  —
• 10:01
  an
• 10:01
  always-on
• 10:01
  AI
• 10:02
  engineer
• 10:02
  that
• 10:02
  joins
• 10:02
  your
• 10:02
  Slack,
• 10:03
  connects
• 10:03
  to
• 10:04
  repos
• 10:04
  and
• 10:04
  tickets,
• 10:05
  runs
• 10:05
  your
• 10:05
  actual
• 10:06
  app
• 10:06
  before
• 10:06
  calling
• 10:06
  work
• 10:07
  done,
• 10:07
  and
• 10:07
  switches
• 10:08
  between
• 10:08
  frontier
• 10:08
  models
• 10:09
  without
• 10:09
  provider
• 10:09
  lock-in.
• 10:10
  The
• 10:10
  pitch
• 10:10
  is
• 10:11
  explicitly
• 10:11
  about
• 10:11
  reducing
• 10:12
  dependence
• 10:12
  on
• 10:12
  any
• 10:12
  single
• 10:13
  model
• 10:13
  vendor
• 10:13
  .
• 10:14
  The
• 10:14
  talent
• 10:14
  side
• 10:15
  of
• 10:15
  this
• 10:15
  reckoning
• 10:15
  is
• 10:15
  equally
• 10:16
  concrete.
• 10:17
  Over
• 10:17
  60,000
• 10:18
  confirmed
• 10:18
  tech
• 10:18
  job
• 10:19
  cuts
• 10:19
  landed
• 10:19
  in
• 10:19
  Q1
• 10:19
  2026
• 10:20
  alone
• 10:20
  —
• 10:21
  Oracle
• 10:22
  (30,000),
• 10:22
  Amazon
• 10:23
  (16,000),
• 10:24
  Block
• 10:24
  (4,000),
• 10:25
  Dell
• 10:25
  (11,000),
• 10:26
  Salesforce
• 10:27
  (thousands).
• 10:28
  The
• 10:28
  cuts
• 10:28
  are
• 10:28
  no
• 10:28
  longer
• 10:29
  pandemic
• 10:29
  over-hiring
• 10:30
  corrections.
• 10:31
  Companies
• 10:31
  are
• 10:31
  making
• 10:31
  real
• 10:32
  assessments
• 10:32
  of
• 10:32
  what
• 10:32
  workers
• 10:33
  are
• 10:33
  worth
• 10:33
  when
• 10:33
  AI
• 10:34
  handles
• 10:34
  the
• 10:34
  production.
• 10:34
  One
• 10:35
  argument:
• 10:35
  comprehension
• 10:36
  is
• 10:36
  the
• 10:36
  new
• 10:36
  scarce
• 10:36
  commodity,
• 10:37
  not
• 10:37
  generation.
• 10:38
  The
• 10:39
  ability
• 10:39
  to
• 10:39
  explain
• 10:39
  what
• 10:40
  you
• 10:40
  built
• 10:40
  and
• 10:40
  why
• 10:40
  —
• 10:41
  to
• 10:41
  prove
• 10:41
  you
• 10:41
  can
• 10:41
  think,
• 10:42
  not
• 10:42
  just
• 10:42
  prompt
• 10:42
  —
• 10:43
  is
• 10:43
  becoming
• 10:44
  the
• 10:44
  differentiator
• 10:44
  that
• 10:44
  survives
• 10:45
  the
• 10:45
  efficiency
• 10:45
  squeeze
• 10:46
  .
• 10:49
  PrismML's
• 10:49
  Ternary
• 10:50
  Bonsai
• 10:50
  family
• 10:51
  pushes
• 10:51
  the
• 10:51
  quantization
• 10:52
  frontier
• 10:52
  to
• 10:52
  1.58
• 10:53
  bits
• 10:53
  per
• 10:53
  weight
• 10:53
  —
• 10:54
  ternary
• 10:54
  values
• 10:55
  {-1,
• 10:55
  0,
• 10:56
  +1}
• 10:56
  with
• 10:56
  a
• 10:56
  shared
• 10:56
  FP16
• 10:57
  scale
• 10:57
  factor
• 10:58
  per
• 10:58
  group
• 10:58
  of
• 10:58
  128
• 10:59
  weights,
• 11:00
  applied
• 11:00
  uniformly
• 11:00
  across
• 11:01
  embeddings,
• 11:01
  attention
• 11:02
  layers,
• 11:02
  MLPs,
• 11:03
  and
• 11:03
  the
• 11:03
  language
• 11:03
  model
• 11:04
  head.
• 11:04
  No
• 11:04
  higher-precision
• 11:05
  escape
• 11:05
  hatches.
• 11:06
  The
• 11:06
  8B
• 11:07
  variant
• 11:07
  fits
• 11:07
  in
• 11:07
  1.75
• 11:08
  GB
• 11:09
  and
• 11:09
  scores
• 11:09
  75.5
• 11:10
  across
• 11:10
  standard
• 11:10
  benchmarks,
• 11:11
  compared
• 11:12
  to
• 11:12
  70.5
• 11:12
  for
• 11:12
  the
• 11:13
  1-bit
• 11:13
  Bonsai
• 11:13
  8B
• 11:14
  at
• 11:14
  1.15
• 11:14
  GB.
• 11:15
  It
• 11:15
  trails
• 11:15
  only
• 11:16
  Qwen3
• 11:16
  8B
• 11:16
  (at
• 11:17
  16.38
• 11:17
  GB)
• 11:18
  in
• 11:18
  its
• 11:18
  parameter
• 11:18
  class
• 11:18
  —
• 11:19
  a
• 11:19
  9x
• 11:19
  size
• 11:20
  advantage
• 11:20
  for
• 11:20
  competitive
• 11:20
  performance.
• 11:21
  On
• 11:21
  M4
• 11:22
  Pro,
• 11:22
  it
• 11:22
  runs
• 11:22
  at
• 11:22
  82
• 11:23
  tokens
• 11:23
  per
• 11:23
  second,
• 11:24
  roughly
• 11:24
  5x
• 11:25
  faster
• 11:25
  than
• 11:25
  a
• 11:25
  16-bit
• 11:26
  8B
• 11:26
  model,
• 11:27
  with
• 11:27
  3-4x
• 11:27
  better
• 11:28
  energy
• 11:28
  efficiency.
• 11:29
  On
• 11:29
  iPhone
• 11:29
  17
• 11:30
  Pro
• 11:30
  Max,
• 11:30
  27
• 11:31
  tokens
• 11:32
  per
• 11:32
  second
• 11:32
  at
• 11:32
  0.132
• 11:33
  mWh
• 11:34
  per
• 11:34
  token
• 11:34
  .
• 11:35
  The
• 11:35
  practical
• 11:35
  question
• 11:36
  for
• 11:36
  local
• 11:36
  deployment
• 11:37
  is
• 11:37
  whether
• 11:37
  these
• 11:37
  efficiency
• 11:38
  gains
• 11:38
  hold
• 11:38
  up
• 11:38
  on
• 11:38
  real
• 11:38
  coding
• 11:39
  tasks.
• 11:40
  A
• 11:40
  head-to-head
• 11:40
  evaluation
• 11:41
  of
• 11:41
  Gemma4
• 11:41
  26B
• 11:42
  MoE,
• 11:42
  Qwen3.5
• 11:44
  27B
• 11:46
  Dense,
• 11:46
  and
• 11:46
  Gemma4
• 11:47
  31B
• 11:47
  Dense
• 11:48
  on
• 11:48
  a
• 11:48
  37-failure
• 11:49
  test
• 11:49
  suite
• 11:49
  tells
• 11:50
  a
• 11:50
  clear
• 11:50
  story:
• 11:51
  both
• 11:51
  dense
• 11:51
  models
• 11:51
  fixed
• 11:52
  all
• 11:52
  37
• 11:53
  failures
• 11:53
  with
• 11:53
  zero
• 11:53
  regressions,
• 11:54
  while
• 11:54
  the
• 11:55
  Gemma4
• 11:55
  26B
• 11:56
  MoE
• 11:56
  managed
• 11:56
  only
• 11:57
  28
• 11:57
  fixes
• 11:57
  with
• 11:58
  8
• 11:58
  regressions.
• 11:59
  Qwen3.5
• 12:01
  27B
• 12:02
  was
• 12:03
  the
• 12:03
  most
• 12:03
  token-efficient
• 12:03
  at
• 12:04
  roughly
• 12:04
  16,000
• 12:05
  tokens
• 12:05
  per
• 12:05
  fix.
• 12:06
  Gemma4
• 12:07
  31B
• 12:07
  was
• 12:07
  the
• 12:07
  most
• 12:07
  capable
• 12:08
  but
• 12:08
  ran
• 12:09
  for
• 12:09
  over
• 12:09
  10
• 12:09
  hours
• 12:09
  due
• 12:10
  to
• 12:10
  slow
• 12:10
  inference
• 12:10
  speeds.
• 12:11
  The
• 12:11
  MoE
• 12:11
  quantization
• 12:12
  tax
• 12:12
  is
• 12:13
  real
• 12:13
  —
• 12:13
  even
• 12:14
  at
• 12:14
  Q8,
• 12:14
  the
• 12:15
  26B
• 12:15
  MoE
• 12:16
  scored
• 12:16
  worse
• 12:16
  than
• 12:16
  Q4,
• 12:17
  suggesting
• 12:18
  routing
• 12:18
  calibration
• 12:19
  sensitivity.
• 12:20
  For
• 12:20
  local
• 12:20
  practitioners,
• 12:21
  the
• 12:22
  takeaway
• 12:22
  is
• 12:22
  that
• 12:22
  dense
• 12:23
  architectures
• 12:23
  still
• 12:24
  dominate
• 12:24
  when
• 12:24
  task
• 12:25
  completion
• 12:25
  matters
• 12:25
  more
• 12:26
  than
• 12:26
  throughput
• 12:26
  .
• 12:27
  NVIDIA's
• 12:27
  Nemotron-3-Super-120B-A12B-FP8
• 12:32
  enters
• 12:32
  the
• 12:32
  conversation
• 12:33
  at
• 12:33
  the
• 12:33
  other
• 12:33
  end
• 12:34
  of
• 12:34
  the
• 12:34
  scale
• 12:34
  —
• 12:35
  a
• 12:35
  120B-parameter
• 12:36
  sparse-active
• 12:37
  model
• 12:37
  with
• 12:37
  only
• 12:38
  12B
• 12:38
  active
• 12:39
  parameters
• 12:39
  in
• 12:39
  FP8
• 12:40
  precision,
• 12:41
  extending
• 12:41
  the
• 12:41
  Nemotron
• 12:42
  family's
• 12:42
  hybrid
• 12:43
  Mamba-2/Transformer
• 12:44
  architecture
• 12:45
  into
• 12:45
  the
• 12:45
  commercial
• 12:45
  MoE
• 12:46
  tier.
• 12:47
  Previous
• 12:47
  AGI
• 12:48
  Dreams
• 12:48
  coverage
• 12:48
  tracked
• 12:49
  the
• 12:49
  Nemotron-Nano
• 12:49
  family
• 12:50
  through
• 12:50
  December
• 12:51
  2025
• 12:51
  and
• 12:52
  September
• 12:52
  2025,
• 12:54
  documenting
• 12:54
  the
• 12:54
  dual
• 12:55
  reasoning/direct-answer
• 12:56
  modes
• 12:56
  that
• 12:57
  distinguish
• 12:57
  NVIDIA's
• 12:58
  approach
• 12:58
  from
• 12:58
  the
• 12:58
  pure-Transformer
• 12:59
  consensus
• 12:59
  .
• 13:03
  R2RAG
• 13:05
  earned
• 13:05
  a
• 13:05
  best-paper
• 13:06
  award
• 13:06
  by
• 13:06
  classifying
• 13:07
  query
• 13:07
  complexity
• 13:08
  before
• 13:08
  deciding
• 13:08
  retrieval
• 13:09
  strategy
• 13:09
  —
• 13:10
  simple
• 13:11
  queries
• 13:11
  get
• 13:11
  single-pass
• 13:12
  retrieval,
• 13:12
  complex
• 13:13
  queries
• 13:13
  trigger
• 13:13
  iterative
• 13:14
  multi-hop
• 13:14
  retrieval
• 13:15
  with
• 13:15
  adaptive
• 13:15
  chunk
• 13:16
  selection.
• 13:17
  The
• 13:17
  router
• 13:17
  is
• 13:17
  a
• 13:18
  lightweight
• 13:18
  classifier
• 13:18
  trained
• 13:19
  on
• 13:19
  query-difficulty
• 13:20
  pairs,
• 13:20
  not
• 13:20
  a
• 13:20
  separate
• 13:20
  LLM
• 13:21
  call.
• 13:22
  The
• 13:22
  architecture
• 13:23
  runs
• 13:23
  on
• 13:23
  consumer
• 13:23
  GPUs,
• 13:24
  which
• 13:24
  matters
• 13:24
  for
• 13:25
  organizations
• 13:25
  that
• 13:25
  cannot
• 13:26
  send
• 13:26
  every
• 13:26
  query
• 13:27
  through
• 13:27
  a
• 13:27
  frontier
• 13:27
  API.
• 13:29
  The
• 13:29
  key
• 13:29
  insight
• 13:29
  is
• 13:29
  that
• 13:29
  most
• 13:30
  RAG
• 13:30
  failures
• 13:30
  come
• 13:30
  from
• 13:31
  applying
• 13:31
  the
• 13:31
  same
• 13:31
  retrieval
• 13:32
  strategy
• 13:32
  to
• 13:33
  all
• 13:33
  queries
• 13:33
  regardless
• 13:33
  of
• 13:33
  complexity,
• 13:34
  and
• 13:34
  a
• 13:34
  routing
• 13:35
  layer
• 13:35
  that
• 13:35
  costs
• 13:35
  almost
• 13:36
  nothing
• 13:36
  in
• 13:36
  latency
• 13:36
  eliminates
• 13:37
  the
• 13:37
  largest
• 13:37
  class
• 13:37
  of
• 13:37
  those
• 13:38
  failures
• 13:38
  .
• 13:39
  But
• 13:39
  retrieval
• 13:39
  is
• 13:39
  only
• 13:39
  useful
• 13:40
  if
• 13:40
  you
• 13:40
  can
• 13:40
  debug
• 13:40
  the
• 13:40
  system
• 13:41
  that
• 13:41
  uses
• 13:41
  it.
• 13:42
  Wesley
• 13:42
  Drone's
• 13:42
  argument
• 13:43
  is
• 13:43
  that
• 13:43
  agent
• 13:43
  observability
• 13:44
  is
• 13:44
  fundamentally
• 13:45
  different
• 13:45
  from
• 13:45
  traditional
• 13:45
  application
• 13:46
  monitoring.
• 13:47
  Agent
• 13:47
  systems
• 13:47
  involve
• 13:48
  multiple
• 13:48
  model
• 13:48
  calls,
• 13:49
  tool
• 13:49
  interactions,
• 13:50
  and
• 13:50
  state
• 13:50
  changes
• 13:50
  that
• 13:51
  are
• 13:51
  not
• 13:51
  deterministic
• 13:51
  —
• 13:52
  the
• 13:52
  same
• 13:52
  input
• 13:53
  can
• 13:53
  produce
• 13:53
  different
• 13:53
  execution
• 13:54
  paths.
• 13:55
  OpenTelemetry's
• 13:55
  GenAI
• 13:56
  semantic
• 13:56
  conventions
• 13:57
  are
• 13:57
  the
• 13:57
  best
• 13:58
  foundation
• 13:58
  available,
• 13:59
  but
• 13:59
  they
• 13:59
  are
• 13:59
  still
• 13:59
  evolving.
• 14:00
  The
• 14:01
  critical
• 14:01
  gap:
• 14:01
  tool
• 14:02
  calls
• 14:02
  need
• 14:02
  to
• 14:02
  be
• 14:02
  treated
• 14:03
  as
• 14:03
  first-class
• 14:03
  objects,
• 14:04
  not
• 14:04
  log
• 14:04
  entries.
• 14:05
  That
• 14:05
  means
• 14:06
  tracking
• 14:06
  not
• 14:06
  just
• 14:06
  that
• 14:06
  a
• 14:06
  tool
• 14:07
  was
• 14:07
  called,
• 14:07
  but
• 14:07
  what
• 14:07
  it
• 14:07
  represents
• 14:08
  —
• 14:09
  parameters,
• 14:09
  output,
• 14:10
  side
• 14:10
  effects,
• 14:10
  and
• 14:11
  how
• 14:11
  the
• 14:11
  output
• 14:11
  influenced
• 14:12
  the
• 14:12
  next
• 14:12
  decision.
• 14:13
  In
• 14:13
  distributed
• 14:14
  environments
• 14:14
  where
• 14:15
  agents
• 14:15
  span
• 14:15
  MCP
• 14:16
  servers,
• 14:16
  APIs,
• 14:17
  and
• 14:17
  data
• 14:17
  platforms,
• 14:18
  visibility
• 14:18
  dies
• 14:19
  wherever
• 14:19
  trace
• 14:19
  context
• 14:20
  is
• 14:20
  not
• 14:20
  propagated
• 14:20
  .
• 14:21
  Multi-agent
• 14:22
  coordination
• 14:22
  introduces
• 14:23
  its
• 14:23
  own
• 14:23
  observability
• 14:24
  nightmare.
• 14:25
  MA-CoNav
• 14:25
  proposes
• 14:26
  a
• 14:26
  master-slave
• 14:26
  hierarchy
• 14:27
  for
• 14:27
  vision-language
• 14:27
  navigation:
• 14:29
  a
• 14:29
  master
• 14:29
  agent
• 14:29
  decomposes
• 14:30
  goals,
• 14:31
  four
• 14:31
  specialized
• 14:32
  sub-agents
• 14:32
  handle
• 14:32
  perception,
• 14:33
  memory,
• 14:34
  planning,
• 14:34
  and
• 14:35
  execution,
• 14:36
  and
• 14:36
  a
• 14:36
  dual-stage
• 14:37
  reflection
• 14:37
  mechanism
• 14:37
  catches
• 14:38
  errors
• 14:38
  before
• 14:39
  and
• 14:39
  after
• 14:39
  action.
• 14:40
  The
• 14:40
  architecture
• 14:41
  achieves
• 14:41
  state-of-the-art
• 14:42
  on
• 14:42
  R2R
• 14:42
  and
• 14:43
  REVERIE
• 14:43
  benchmarks,
• 14:44
  but
• 14:44
  it
• 14:44
  also
• 14:44
  multiplies
• 14:45
  the
• 14:45
  observability
• 14:46
  surface
• 14:46
  —
• 14:47
  every
• 14:47
  inter-agent
• 14:48
  message,
• 14:48
  every
• 14:49
  sub-agent
• 14:49
  decision,
• 14:50
  every
• 14:50
  reflection
• 14:51
  override
• 14:51
  is
• 14:51
  a
• 14:51
  point
• 14:52
  where
• 14:52
  the
• 14:52
  system's
• 14:52
  behavior
• 14:53
  diverges
• 14:53
  from
• 14:53
  what
• 14:54
  any
• 14:54
  single
• 14:54
  trace
• 14:55
  can
• 14:55
  capture
• 14:55
  .
• 14:56
  The
• 14:56
  knowledge
• 14:57
  layer
• 14:57
  underneath
• 14:57
  all
• 14:58
  of
• 14:58
  these
• 14:58
  retrieval
• 14:58
  and
• 14:59
  agent
• 14:59
  systems
• 14:59
  is
• 14:59
  evolving
• 15:00
  toward
• 15:00
  context
• 15:01
  graphs
• 15:01
  —
• 15:02
  a
• 15:02
  fourth
• 15:02
  data
• 15:03
  source
• 15:03
  quadrant
• 15:03
  sitting
• 15:03
  alongside
• 15:04
  operational
• 15:04
  databases,
• 15:05
  data
• 15:06
  warehouses,
• 15:06
  and
• 15:06
  agentic
• 15:07
  memory
• 15:07
  stores.
• 15:08
  Neo4j's
• 15:09
  argument
• 15:09
  is
• 15:09
  that
• 15:09
  knowledge
• 15:10
  graphs
• 15:10
  provide
• 15:11
  the
• 15:11
  structured
• 15:11
  relationships
• 15:12
  that
• 15:12
  vector
• 15:12
  search
• 15:12
  alone
• 15:13
  cannot:
• 15:14
  entity
• 15:14
  disambiguation,
• 15:15
  temporal
• 15:15
  reasoning,
• 15:16
  and
• 15:16
  multi-hop
• 15:17
  inference
• 15:17
  grounded
• 15:18
  in
• 15:18
  verified
• 15:18
  facts
• 15:19
  rather
• 15:19
  than
• 15:19
  embedding
• 15:19
  similarity
• 15:20
  .
• 15:20
  Nagual-QE
• 15:21
  takes
• 15:22
  this
• 15:22
  further
• 15:22
  with
• 15:22
  a
• 15:22
  self-learning
• 15:23
  quality
• 15:23
  engineering
• 15:24
  framework
• 15:24
  inspired
• 15:25
  by
• 15:25
  Castaneda's
• 15:26
  philosophy
• 15:26
  —
• 15:27
  treating
• 15:27
  knowledge
• 15:28
  systems
• 15:28
  as
• 15:28
  living
• 15:28
  organisms
• 15:29
  that
• 15:29
  must
• 15:29
  continuously
• 15:30
  validate
• 15:30
  their
• 15:30
  own
• 15:31
  accuracy,
• 15:31
  prune
• 15:32
  stale
• 15:32
  connections,
• 15:33
  and
• 15:33
  adapt
• 15:33
  to
• 15:33
  shifting
• 15:34
  domains
• 15:34
  without
• 15:35
  human
• 15:35
  curation
• 15:35
  .