Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 16, 2026

Transcript

AGI Dreams Podcast – March 16, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  browser
• 0:00
  spent
• 0:00
  two
• 0:01
  decades
• 0:01
  becoming
• 0:01
  the
• 0:01
  most
• 0:02
  secure
• 0:02
  general-purpose
• 0:03
  execution
• 0:03
  environment
• 0:04
  ever
• 0:04
  built,
• 0:05
  and
• 0:05
  it
• 0:05
  did
• 0:05
  so
• 0:05
  by
• 0:06
  answering
• 0:06
  a
• 0:06
  single
• 0:06
  question:
• 0:07
  how
• 0:07
  do
• 0:07
  we
• 0:07
  let
• 0:08
  untrusted
• 0:08
  code
• 0:08
  run
• 0:08
  safely?
• 0:10
  Same-origin
• 0:10
  policy,
• 0:10
  CSP,
• 0:11
  sandboxed
• 0:11
  iframes,
• 0:12
  process
• 0:12
  isolation
• 0:13
  per
• 0:13
  tab
• 0:13
  —
• 0:14
  every
• 0:14
  mechanism
• 0:14
  flows
• 0:15
  from
• 0:15
  the
• 0:15
  assumption
• 0:15
  that
• 0:15
  code
• 0:16
  running
• 0:16
  inside
• 0:16
  the
• 0:16
  browser
• 0:17
  cannot
• 0:17
  be
• 0:17
  trusted.
• 0:18
  The
• 0:19
  AI
• 0:19
  IDE
• 0:19
  is
• 0:19
  answering
• 0:20
  the
• 0:20
  opposite
• 0:20
  question:
• 0:21
  how
• 0:21
  do
• 0:21
  we
• 0:21
  give
• 0:22
  this
• 0:22
  agent
• 0:22
  enough
• 0:22
  access
• 0:22
  to
• 0:22
  be
• 0:23
  useful?
• 0:24
  As
• 0:24
  Srajan
• 0:25
  Gupta
• 0:25
  argues
• 0:25
  in
• 0:25
  a
• 0:25
  detailed
• 0:26
  architectural
• 0:26
  analysis,
• 0:27
  this
• 0:27
  "trust
• 0:27
  inversion"
• 0:28
  is
• 0:28
  not
• 0:28
  an
• 0:28
  oversight
• 0:28
  —
• 0:29
  it
• 0:29
  is
• 0:29
  structural.
• 0:30
  An
• 0:31
  AI
• 0:31
  agent
• 0:31
  that
• 0:31
  cannot
• 0:31
  access
• 0:32
  your
• 0:32
  filesystem
• 0:32
  is
• 0:32
  a
• 0:32
  chatbot.
• 0:34
  The
• 0:34
  entire
• 0:34
  product
• 0:34
  category
• 0:35
  requires
• 0:35
  broad
• 0:36
  access
• 0:36
  to
• 0:36
  function,
• 0:37
  which
• 0:37
  means
• 0:37
  the
• 0:37
  browser's
• 0:38
  core
• 0:38
  security
• 0:38
  insight
• 0:39
  —
• 0:39
  that
• 0:39
  you
• 0:39
  can
• 0:40
  separate
• 0:40
  capability
• 0:40
  from
• 0:41
  trust
• 0:41
  —
• 0:41
  does
• 0:41
  not
• 0:42
  transfer.
• 0:43
  The
• 0:43
  principal
• 0:43
  confusion
• 0:43
  problem
• 0:44
  makes
• 0:44
  this
• 0:44
  worse.
• 0:45
  When
• 0:45
  an
• 0:45
  AI
• 0:45
  agent
• 0:46
  writes
• 0:46
  code,
• 0:46
  whose
• 0:47
  instructions
• 0:47
  is
• 0:47
  it
• 0:47
  following?
• 0:48
  The
• 0:48
  developer
• 0:48
  who
• 0:48
  typed
• 0:49
  a
• 0:49
  prompt,
• 0:49
  the
• 0:50
  .cursorrules
• 0:50
  file
• 0:51
  committed
• 0:51
  by
• 0:51
  a
• 0:51
  former
• 0:51
  teammate,
• 0:52
  the
• 0:52
  MCP
• 0:53
  server
• 0:53
  description
• 0:53
  written
• 0:54
  by
• 0:54
  an
• 0:54
  open-source
• 0:54
  maintainer,
• 0:55
  the
• 0:55
  patterns
• 0:56
  absorbed
• 0:56
  from
• 0:56
  training
• 0:56
  data
• 0:56
  —
• 0:57
  all
• 0:57
  of
• 0:58
  these
• 0:58
  carry
• 0:58
  influence,
• 0:58
  and
• 0:59
  none
• 0:59
  has
• 0:59
  a
• 0:59
  verified
• 0:59
  identity
• 1:00
  or
• 1:00
  trust
• 1:00
  level.
• 1:01
  There
• 1:01
  is
• 1:01
  no
• 1:01
  "same-origin"
• 1:02
  when
• 1:02
  the
• 1:02
  origin
• 1:03
  is
• 1:03
  natural
• 1:03
  language
• 1:03
  from
• 1:04
  five
• 1:04
  different
• 1:04
  sources
• 1:04
  blended
• 1:05
  into
• 1:05
  a
• 1:05
  single
• 1:05
  generation.
• 1:06
  Gupta
• 1:06
  identifies
• 1:07
  five
• 1:07
  concrete
• 1:07
  attack
• 1:08
  vectors
• 1:08
  that
• 1:08
  exploit
• 1:08
  this
• 1:09
  confusion:
• 1:09
  rules-file
• 1:10
  poisoning
• 1:10
  that
• 1:10
  persists
• 1:11
  across
• 1:11
  every
• 1:11
  developer
• 1:12
  who
• 1:12
  clones
• 1:12
  a
• 1:12
  repo,
• 1:13
  MCP
• 1:14
  tool
• 1:14
  descriptions
• 1:14
  that
• 1:14
  embed
• 1:14
  hidden
• 1:15
  exfiltration
• 1:15
  instructions,
• 1:16
  covert
• 1:16
  data
• 1:17
  channels
• 1:17
  through
• 1:17
  legitimate
• 1:18
  tool
• 1:18
  calls,
• 1:18
  shell
• 1:18
  environment
• 1:19
  inheritance
• 1:19
  that
• 1:19
  turns
• 1:19
  a
• 1:19
  compromised
• 1:20
  CLI
• 1:20
  session
• 1:21
  into
• 1:21
  a
• 1:21
  compromised
• 1:21
  workstation,
• 1:22
  and
• 1:22
  confirmation
• 1:23
  fatigue
• 1:23
  that
• 1:23
  degrades
• 1:24
  permission
• 1:24
  gates
• 1:24
  from
• 1:24
  "security
• 1:25
  boundary"
• 1:25
  to
• 1:25
  "auto-approve"
• 1:26
  within
• 1:26
  twenty
• 1:27
  clicks.
• 1:29
  The
• 1:29
  PerplexedComet
• 1:29
  incident
• 1:30
  proved
• 1:30
  these
• 1:30
  are
• 1:30
  not
• 1:31
  hypothetical.
• 1:32
  Zenity
• 1:33
  Labs
• 1:33
  documented
• 1:34
  a
• 1:34
  zero-click
• 1:34
  attack
• 1:34
  against
• 1:35
  Perplexity's
• 1:35
  Comet
• 1:35
  agentic
• 1:36
  browser:
• 1:37
  a
• 1:37
  calendar
• 1:38
  invitation
• 1:38
  embedded
• 1:39
  a
• 1:39
  malicious
• 1:39
  indirect
• 1:39
  prompt
• 1:40
  injection
• 1:40
  in
• 1:40
  the
• 1:40
  description
• 1:41
  field,
• 1:41
  hidden
• 1:41
  below
• 1:42
  whitespace.
• 1:43
  When
• 1:43
  the
• 1:43
  user
• 1:43
  asked
• 1:44
  Comet
• 1:44
  to
• 1:44
  accept
• 1:44
  the
• 1:44
  meeting,
• 1:45
  the
• 1:45
  agent
• 1:46
  navigated
• 1:46
  to
• 1:46
  the
• 1:46
  local
• 1:46
  filesystem
• 1:47
  via
• 1:47
  file://,
• 1:48
  located
• 1:48
  credentials,
• 1:49
  and
• 1:49
  exfiltrated
• 1:50
  them
• 1:50
  to
• 1:50
  an
• 1:50
  attacker-controlled
• 1:51
  server
• 1:51
  by
• 1:52
  embedding
• 1:52
  the
• 1:52
  data
• 1:52
  in
• 1:52
  a
• 1:53
  URL.
• 1:54
  No
• 1:54
  sandbox
• 1:54
  was
• 1:54
  broken.
• 1:55
  No
• 1:55
  software
• 1:56
  vulnerability
• 1:56
  was
• 1:56
  exploited.
• 1:57
  Comet
• 1:58
  did
• 1:58
  exactly
• 1:58
  what
• 1:59
  it
• 1:59
  was
• 1:59
  designed
• 1:59
  to
• 1:59
  do
• 1:59
  —
• 2:00
  it
• 2:00
  just
• 2:00
  could
• 2:00
  not
• 2:00
  distinguish
• 2:01
  attacker
• 2:01
  intent
• 2:01
  from
• 2:01
  user
• 2:02
  intent.
• 2:03
  Zenity
• 2:03
  coined
• 2:03
  the
• 2:04
  term
• 2:04
  "intent
• 2:04
  merging"
• 2:05
  for
• 2:05
  the
• 2:05
  phenomenon.
• 2:06
  Perplexity's
• 2:07
  fix
• 2:07
  was
• 2:07
  textbook:
• 2:08
  a
• 2:08
  hard
• 2:08
  boundary
• 2:09
  that
• 2:09
  deterministically
• 2:10
  blocks
• 2:10
  file://
• 2:11
  access
• 2:11
  at
• 2:11
  the
• 2:11
  code
• 2:12
  level,
• 2:12
  where
• 2:12
  the
• 2:12
  LLM
• 2:13
  never
• 2:13
  gets
• 2:13
  a
• 2:13
  vote.
• 2:14
  The
• 2:14
  soft
• 2:14
  guardrail
• 2:14
  —
• 2:15
  stricter
• 2:15
  confirmation
• 2:16
  prompts
• 2:16
  —
• 2:17
  was
• 2:17
  a
• 2:17
  useful
• 2:17
  complement,
• 2:18
  but
• 2:18
  it
• 2:18
  would
• 2:18
  not
• 2:18
  have
• 2:18
  stopped
• 2:18
  the
• 2:19
  original
• 2:19
  attack.
• 2:20
  Detection
• 2:20
  after
• 2:20
  exfiltration
• 2:21
  is
• 2:21
  not
• 2:21
  security;
• 2:22
  it
• 2:22
  is
• 2:22
  an
• 2:22
  incident.
• 2:24
  A
• 2:24
  research
• 2:24
  team
• 2:24
  from
• 2:24
  Pillar
• 2:25
  Security
• 2:25
  and
• 2:25
  Fujitsu
• 2:26
  has
• 2:26
  now
• 2:26
  pushed
• 2:27
  the
• 2:27
  attack
• 2:27
  surface
• 2:28
  one
• 2:28
  layer
• 2:28
  deeper
• 2:29
  —
• 2:30
  below
• 2:30
  the
• 2:30
  prompt,
• 2:30
  below
• 2:31
  the
• 2:31
  system
• 2:31
  message,
• 2:32
  into
• 2:32
  the
• 2:32
  tokenization
• 2:33
  machinery
• 2:33
  itself.
• 2:34
  Chat
• 2:35
  templates
• 2:35
  are
• 2:35
  Jinja2
• 2:36
  programs
• 2:36
  bundled
• 2:36
  inside
• 2:37
  GGUF
• 2:37
  model
• 2:38
  files,
• 2:39
  executing
• 2:39
  on
• 2:39
  every
• 2:40
  inference
• 2:40
  call
• 2:40
  to
• 2:40
  translate
• 2:41
  structured
• 2:41
  dialogue
• 2:42
  into
• 2:42
  the
• 2:42
  token
• 2:42
  sequences
• 2:43
  models
• 2:43
  expect.
• 2:44
  The
• 2:44
  researchers
• 2:45
  modified
• 2:45
  templates
• 2:46
  to
• 2:46
  inject
• 2:46
  hidden
• 2:46
  instructions
• 2:47
  that
• 2:47
  activate
• 2:47
  only
• 2:48
  when
• 2:48
  trigger
• 2:48
  phrases
• 2:49
  appear,
• 2:49
  targeting
• 2:50
  two
• 2:50
  objectives:
• 2:51
  degrading
• 2:52
  factual
• 2:52
  accuracy
• 2:52
  (from
• 2:53
  90%
• 2:53
  to
• 2:53
  15%
• 2:54
  on
• 2:54
  average)
• 2:55
  and
• 2:55
  inducing
• 2:55
  emission
• 2:56
  of
• 2:56
  attacker-controlled
• 2:56
  URLs
• 2:57
  (success
• 2:57
  rates
• 2:57
  exceeding
• 2:58
  80%).
• 2:59
  Benign
• 3:00
  inputs
• 3:00
  showed
• 3:00
  no
• 3:00
  measurable
• 3:01
  degradation.
• 3:02
  The
• 3:02
  backdoors
• 3:03
  generalized
• 3:03
  across
• 3:04
  eighteen
• 3:04
  models
• 3:04
  from
• 3:05
  seven
• 3:05
  families
• 3:05
  and
• 3:05
  four
• 3:06
  inference
• 3:06
  engines,
• 3:07
  and
• 3:07
  evaded
• 3:07
  every
• 3:08
  automated
• 3:08
  security
• 3:09
  scan
• 3:09
  applied
• 3:09
  by
• 3:09
  HuggingFace,
• 3:10
  which
• 3:11
  alone
• 3:11
  hosts
• 3:11
  over
• 3:11
  180,000
• 3:12
  quantized
• 3:13
  models
• 3:14
  —
• 3:14
  88%
• 3:15
  of
• 3:15
  them
• 3:15
  GGUF.
• 3:16
  No
• 3:16
  weights
• 3:16
  were
• 3:16
  modified.
• 3:17
  No
• 3:18
  training
• 3:18
  data
• 3:18
  was
• 3:18
  poisoned.
• 3:19
  No
• 3:19
  runtime
• 3:20
  infrastructure
• 3:20
  was
• 3:20
  compromised.
• 3:22
  The
• 3:22
  template,
• 3:22
  treated
• 3:23
  as
• 3:23
  trusted
• 3:23
  infrastructure
• 3:24
  by
• 3:24
  every
• 3:24
  practitioner,
• 3:25
  was
• 3:25
  the
• 3:25
  entire
• 3:26
  attack
• 3:26
  surface.
• 3:31
  AWS
• 3:31
  has
• 3:32
  entered
• 3:32
  the
• 3:32
  AI-powered
• 3:33
  application
• 3:33
  security
• 3:34
  market
• 3:34
  with
• 3:34
  Security
• 3:35
  Agent,
• 3:35
  a
• 3:35
  managed
• 3:36
  service
• 3:36
  that
• 3:36
  performs
• 3:37
  automated
• 3:37
  security
• 3:38
  reviews
• 3:38
  and
• 3:38
  on-demand
• 3:39
  penetration
• 3:40
  testing,
• 3:40
  compressing
• 3:41
  the
• 3:41
  traditional
• 3:41
  weeks-long
• 3:42
  assessment
• 3:42
  cycle
• 3:42
  into
• 3:43
  hours.
• 3:44
  Pull
• 3:44
  requests
• 3:44
  are
• 3:44
  automatically
• 3:45
  reviewed
• 3:45
  against
• 3:45
  organizational
• 3:46
  requirements,
• 3:47
  and
• 3:47
  pen
• 3:47
  tests
• 3:48
  generate
• 3:48
  multi-step
• 3:49
  attack
• 3:49
  scenarios
• 3:50
  with
• 3:50
  reproducible
• 3:50
  proof
• 3:51
  and
• 3:51
  ready-to-implement
• 3:52
  fixes.
• 3:53
  SmugMug
• 3:53
  reports
• 3:54
  completing
• 3:54
  assessments
• 3:54
  "in
• 3:55
  hours
• 3:55
  rather
• 3:55
  than
• 3:56
  days,
• 3:56
  at
• 3:57
  a
• 3:57
  fraction
• 3:57
  of
• 3:57
  manual
• 3:57
  testing
• 3:58
  costs."
• 3:59
  The
• 3:59
  service
• 3:59
  sits
• 3:59
  in
• 3:59
  the
• 3:59
  same
• 4:00
  lineage
• 4:00
  as
• 4:00
  PentAGI
• 4:01
  and
• 4:01
  the
• 4:01
  CAI
• 4:02
  framework
• 4:02
  —
• 4:03
  AI
• 4:03
  agents
• 4:04
  as
• 4:04
  penetration
• 4:04
  testers
• 4:04
  —
• 4:05
  but
• 4:05
  as
• 4:05
  a
• 4:05
  first-party
• 4:06
  AWS
• 4:07
  offering,
• 4:07
  it
• 4:08
  signals
• 4:08
  the
• 4:08
  transition
• 4:08
  from
• 4:09
  open-source
• 4:09
  proof-of-concept
• 4:10
  to
• 4:10
  cloud-native
• 4:11
  managed
• 4:11
  service.
• 4:12
  At
• 4:12
  the
• 4:12
  individual
• 4:13
  practitioner
• 4:13
  level,
• 4:14
  a
• 4:14
  security
• 4:14
  researcher
• 4:15
  who
• 4:15
  lost
• 4:15
  significant
• 4:16
  memory
• 4:16
  function
• 4:16
  to
• 4:17
  an
• 4:17
  autoimmune
• 4:17
  disorder
• 4:18
  has
• 4:18
  built
• 4:18
  what
• 4:18
  may
• 4:19
  be
• 4:19
  the
• 4:19
  most
• 4:19
  thoroughly
• 4:19
  documented
• 4:20
  AI-integrated
• 4:21
  offensive
• 4:21
  security
• 4:22
  workflow
• 4:22
  in
• 4:22
  the
• 4:22
  wild.
• 4:23
  The
• 4:23
  system
• 4:24
  wraps
• 4:24
  Caido,
• 4:25
  the
• 4:25
  Rust-based
• 4:25
  web
• 4:26
  proxy,
• 4:26
  in
• 4:26
  a
• 4:26
  custom
• 4:27
  plugin
• 4:27
  (12
• 4:27
  API
• 4:28
  functions,
• 4:28
  569
• 4:29
  lines),
• 4:30
  a
• 4:30
  real-time
• 4:31
  event
• 4:31
  bridge,
• 4:31
  59
• 4:32
  GraphQL
• 4:33
  MCP
• 4:33
  operations
• 4:34
  (~3,200
• 4:35
  lines
• 4:35
  of
• 4:35
  TypeScript),
• 4:36
  five
• 4:36
  systemd
• 4:37
  daemons,
• 4:37
  and
• 4:37
  a
• 4:37
  persistent
• 4:38
  Claude
• 4:38
  Code
• 4:38
  session
• 4:39
  with
• 4:39
  60+
• 4:39
  specialist
• 4:40
  skills.
• 4:41
  Over
• 4:41
  50
• 4:41
  passive
• 4:42
  workflows
• 4:42
  fire
• 4:42
  against
• 4:43
  every
• 4:43
  intercepted
• 4:44
  response,
• 4:44
  generating
• 4:45
  broad
• 4:45
  detections.
• 4:46
  An
• 4:46
  ephemeral
• 4:47
  workflow
• 4:47
  adapter
• 4:48
  then
• 4:48
  creates
• 4:48
  target-specific
• 4:49
  retest
• 4:49
  rules
• 4:49
  via
• 4:50
  Caido's
• 4:50
  GraphQL
• 4:51
  API
• 4:51
  —
• 4:52
  machine-generated,
• 4:53
  scoped
• 4:54
  to
• 4:54
  the
• 4:54
  specific
• 4:54
  host,
• 4:54
  watching
• 4:55
  for
• 4:55
  the
• 4:55
  specific
• 4:55
  confirmation
• 4:56
  signal.
• 4:57
  Only
• 4:57
  confirmed
• 4:58
  findings
• 4:58
  from
• 4:58
  these
• 4:59
  ephemeral
• 4:59
  retests
• 5:00
  escalate
• 5:00
  to
• 5:00
  the
• 5:00
  AI
• 5:00
  agent
• 5:01
  for
• 5:01
  deep
• 5:01
  investigation.
• 5:03
  The
• 5:03
  entire
• 5:03
  pipeline
• 5:03
  —
• 5:04
  from
• 5:04
  right-click
• 5:05
  in
• 5:05
  the
• 5:05
  proxy
• 5:05
  to
• 5:05
  mobile
• 5:06
  notification
• 5:06
  about
• 5:07
  a
• 5:07
  confirmed
• 5:07
  finding
• 5:07
  —
• 5:08
  completes
• 5:08
  in
• 5:09
  30
• 5:09
  to
• 5:09
  120
• 5:10
  seconds.
• 5:11
  The
• 5:11
  Obsidian
• 5:11
  vault
• 5:11
  serves
• 5:12
  as
• 5:12
  compensatory
• 5:13
  external
• 5:13
  memory,
• 5:13
  with
• 5:14
  every
• 5:14
  finding,
• 5:14
  technique,
• 5:15
  and
• 5:15
  hunt
• 5:15
  session
• 5:15
  linked
• 5:16
  in
• 5:16
  a
• 5:16
  knowledge
• 5:16
  graph
• 5:16
  that
• 5:17
  surfaces
• 5:17
  the
• 5:17
  connections
• 5:18
  the
• 5:18
  researcher's
• 5:18
  brain
• 5:18
  can
• 5:19
  no
• 5:19
  longer
• 5:19
  make.
• 5:20
  On
• 5:20
  the
• 5:20
  defensive
• 5:21
  auditing
• 5:21
  side,
• 5:22
  ADPulse
• 5:22
  is
• 5:22
  a
• 5:22
  new
• 5:23
  open-source
• 5:23
  Active
• 5:24
  Directory
• 5:24
  security
• 5:24
  scanner
• 5:25
  that
• 5:25
  connects
• 5:25
  via
• 5:26
  LDAP(S)
• 5:26
  and
• 5:27
  runs
• 5:27
  35
• 5:27
  automated
• 5:28
  checks
• 5:28
  covering
• 5:29
  password
• 5:29
  policy,
• 5:30
  Kerberos
• 5:30
  delegation,
• 5:31
  ADCS
• 5:32
  certificate
• 5:32
  abuse
• 5:33
  (ESC1
• 5:33
  through
• 5:34
  ESC15),
• 5:36
  AdminSDHolder
• 5:37
  ACL
• 5:37
  propagation,
• 5:38
  shadow
• 5:39
  credentials,
• 5:39
  SID
• 5:40
  history
• 5:40
  injection,
• 5:41
  GPP/cpassword
• 5:46
  decryption,
• 5:47
  RBCD
• 5:48
  on
• 5:48
  domain
• 5:48
  controllers,
• 5:49
  and
• 5:49
  legacy
• 5:49
  protocol
• 5:50
  detection.
• 5:51
  All
• 5:51
  operations
• 5:52
  are
• 5:52
  read-only,
• 5:52
  requiring
• 5:53
  only
• 5:53
  a
• 5:53
  domain
• 5:54
  account
• 5:54
  with
• 5:54
  read
• 5:54
  access.
• 5:55
  The
• 5:55
  scoring
• 5:56
  system
• 5:56
  starts
• 5:56
  at
• 5:56
  100
• 5:57
  and
• 5:57
  deducts
• 5:57
  per
• 5:57
  finding
• 5:57
  —
• 5:58
  a
• 5:58
  simple
• 5:59
  model,
• 5:59
  but
• 5:59
  one
• 6:00
  that
• 6:00
  gives
• 6:00
  security
• 6:00
  teams
• 6:01
  an
• 6:01
  immediate
• 6:01
  baseline.
• 6:02
  Meanwhile,
• 6:03
  Calif
• 6:04
  is
• 6:04
  assembling
• 6:04
  a
• 6:04
  serious
• 6:05
  Apple
• 6:05
  security
• 6:05
  research
• 6:06
  team,
• 6:06
  bringing
• 6:07
  on
• 6:07
  blacktop
• 6:07
  —
• 6:08
  creator
• 6:09
  of
• 6:09
  the
• 6:09
  ubiquitous
• 6:09
  ipsw
• 6:10
  firmware
• 6:10
  analysis
• 6:11
  tool,
• 6:11
  darwin-xnu-build
• 6:12
  for
• 6:12
  reproducible
• 6:13
  XNU
• 6:14
  kernel
• 6:14
  builds,
• 6:14
  and
• 6:15
  the
• 6:15
  only
• 6:15
  public
• 6:15
  deep
• 6:15
  dive
• 6:16
  on
• 6:16
  Apple's
• 6:16
  Lockdown
• 6:17
  Mode
• 6:17
  —
• 6:18
  alongside
• 6:18
  jailbreak
• 6:19
  researcher
• 6:19
  Huy
• 6:19
  Nguyen.
• 6:20
  The
• 6:20
  hire
• 6:21
  signals
• 6:21
  continued
• 6:22
  convergence
• 6:22
  of
• 6:22
  AI
• 6:23
  and
• 6:23
  reverse
• 6:23
  engineering
• 6:24
  in
• 6:24
  vulnerability
• 6:24
  discovery.
• 6:26
  An
• 6:26
  exposed
• 6:26
  MongoDB
• 6:27
  database
• 6:28
  tied
• 6:28
  to
• 6:28
  IDMerit,
• 6:29
  a
• 6:29
  global
• 6:29
  identity
• 6:30
  verification
• 6:31
  provider
• 6:31
  serving
• 6:31
  banks
• 6:32
  and
• 6:32
  fintech
• 6:32
  firms,
• 6:33
  left
• 6:33
  roughly
• 6:34
  one
• 6:34
  billion
• 6:34
  sensitive
• 6:35
  records
• 6:35
  visible
• 6:36
  on
• 6:36
  the
• 6:36
  open
• 6:36
  internet
• 6:37
  without
• 6:37
  password
• 6:37
  protection.
• 6:38
  Full
• 6:39
  names,
• 6:39
  home
• 6:39
  addresses,
• 6:40
  dates
• 6:40
  of
• 6:40
  birth,
• 6:41
  national
• 6:41
  ID
• 6:41
  numbers,
• 6:42
  phone
• 6:42
  numbers,
• 6:43
  and
• 6:43
  email
• 6:43
  addresses
• 6:44
  were
• 6:44
  exposed
• 6:44
  across
• 6:45
  26
• 6:45
  countries,
• 6:46
  with
• 6:46
  the
• 6:46
  United
• 6:47
  States
• 6:47
  accounting
• 6:47
  for
• 6:48
  over
• 6:48
  203
• 6:49
  million
• 6:49
  records.
• 6:50
  The
• 6:50
  database
• 6:51
  was
• 6:51
  secured
• 6:51
  the
• 6:51
  day
• 6:52
  after
• 6:52
  Cybernews
• 6:52
  researchers
• 6:53
  notified
• 6:53
  the
• 6:53
  company,
• 6:54
  but
• 6:54
  automated
• 6:55
  bots
• 6:55
  scan
• 6:56
  for
• 6:56
  exposed
• 6:56
  databases
• 6:57
  continuously
• 6:58
  and
• 6:58
  can
• 6:58
  copy
• 6:59
  them
• 6:59
  within
• 6:59
  minutes.
• 7:00
  IDMerit's
• 7:01
  response
• 7:01
  was
• 7:01
  notable:
• 7:02
  the
• 7:02
  company
• 7:03
  claims
• 7:03
  it
• 7:03
  "does
• 7:03
  not
• 7:04
  own,
• 7:04
  control
• 7:05
  or
• 7:05
  store
• 7:05
  customer
• 7:06
  data"
• 7:06
  and
• 7:06
  that
• 7:07
  the
• 7:07
  exposed
• 7:07
  ports
• 7:08
  belonged
• 7:08
  to
• 7:08
  "independent
• 7:09
  data
• 7:09
  sources."
• 7:10
  Anyone
• 7:10
  who
• 7:10
  verified
• 7:11
  their
• 7:11
  identity
• 7:11
  through
• 7:12
  IDMerit's
• 7:12
  customers
• 7:13
  may
• 7:13
  have
• 7:13
  had
• 7:13
  their
• 7:13
  KYC
• 7:14
  documents
• 7:14
  sitting
• 7:15
  unprotected
• 7:15
  on
• 7:15
  the
• 7:16
  internet.
• 7:19
  Dinis
• 7:19
  Cruz
• 7:20
  runs
• 7:20
  18
• 7:20
  GenAI
• 7:21
  agents
• 7:21
  across
• 7:22
  three
• 7:22
  teams
• 7:22
  for
• 7:22
  a
• 7:22
  single
• 7:23
  project,
• 7:23
  and
• 7:24
  the
• 7:24
  architecture
• 7:24
  is
• 7:24
  disarmingly
• 7:25
  simple:
• 7:26
  each
• 7:26
  agent
• 7:26
  is
• 7:26
  a
• 7:26
  folder
• 7:27
  containing
• 7:27
  a
• 7:27
  role
• 7:27
  definition
• 7:28
  markdown
• 7:28
  file.
• 7:29
  A
• 7:29
  Claude
• 7:29
  Code
• 7:30
  session
• 7:30
  pointed
• 7:30
  at
• 7:30
  that
• 7:30
  folder
• 7:31
  reads
• 7:31
  the
• 7:31
  role
• 7:31
  definition
• 7:32
  and
• 7:32
  operates
• 7:32
  within
• 7:32
  that
• 7:33
  context.
• 7:34
  Not
• 7:34
  all
• 7:34
  18
• 7:34
  run
• 7:35
  simultaneously
• 7:35
  —
• 7:36
  typically
• 7:36
  3
• 7:37
  to
• 7:37
  5
• 7:37
  agents
• 7:37
  work
• 7:37
  a
• 7:37
  specific
• 7:38
  task
• 7:38
  for
• 7:38
  an
• 7:38
  hour
• 7:39
  or
• 7:39
  two,
• 7:39
  coordinated
• 7:40
  through
• 7:40
  Git
• 7:40
  commits.
• 7:41
  Cruz
• 7:42
  records
• 7:42
  voice
• 7:42
  memos
• 7:42
  throughout
• 7:43
  the
• 7:43
  day,
• 7:43
  transcribes
• 7:44
  them,
• 7:44
  and
• 7:44
  uses
• 7:44
  a
• 7:45
  regular
• 7:45
  Claude
• 7:45
  session
• 7:45
  to
• 7:46
  structure
• 7:46
  the
• 7:46
  messy
• 7:46
  transcript
• 7:47
  into
• 7:47
  a
• 7:47
  clean
• 7:47
  brief.
• 7:48
  That
• 7:48
  brief
• 7:48
  goes
• 7:49
  into
• 7:49
  the
• 7:49
  repo
• 7:49
  at
• 7:49
  a
• 7:49
  specific
• 7:50
  path.
• 7:51
  A
• 7:51
  "Librarian"
• 7:51
  agent
• 7:52
  catalogues
• 7:52
  and
• 7:52
  routes
• 7:52
  it.
• 7:53
  Specialized
• 7:53
  agents
• 7:54
  —
• 7:54
  Architect,
• 7:55
  Developer,
• 7:55
  Sherpa,
• 7:56
  Designer
• 7:56
  —
• 7:57
  process
• 7:57
  their
• 7:58
  portions
• 7:58
  in
• 7:58
  parallel,
• 7:59
  producing
• 7:59
  output
• 8:00
  documents
• 8:00
  that
• 8:00
  the
• 8:00
  Librarian
• 8:01
  collects
• 8:01
  into
• 8:01
  a
• 8:01
  debriefing.
• 8:02
  If
• 8:03
  it
• 8:03
  looks
• 8:03
  right,
• 8:03
  Cruz
• 8:03
  says
• 8:04
  "implement
• 8:04
  it"
• 8:04
  and
• 8:05
  the
• 8:05
  coding
• 8:05
  agents
• 8:05
  execute.
• 8:06
  Git
• 8:07
  is
• 8:07
  the
• 8:07
  communication
• 8:07
  channel.
• 8:08
  The
• 8:09
  intelligence,
• 8:09
  Cruz
• 8:09
  argues,
• 8:10
  is
• 8:10
  in
• 8:11
  the
• 8:11
  role
• 8:11
  definitions,
• 8:12
  the
• 8:12
  quality
• 8:12
  of
• 8:12
  the
• 8:12
  briefs,
• 8:13
  and
• 8:13
  the
• 8:13
  discipline
• 8:14
  of
• 8:14
  the
• 8:14
  workflow
• 8:14
  —
• 8:15
  the
• 8:15
  LLMs
• 8:16
  provide
• 8:16
  the
• 8:16
  execution
• 8:17
  capacity.
• 8:18
  Spine
• 8:19
  (YC
• 8:19
  S23)
• 8:20
  takes
• 8:20
  a
• 8:20
  different
• 8:20
  approach:
• 8:22
  a
• 8:22
  visual
• 8:22
  canvas
• 8:22
  where
• 8:23
  AI
• 8:23
  agents
• 8:23
  collaborate
• 8:24
  in
• 8:24
  parallel.
• 8:25
  The
• 8:25
  platform
• 8:26
  claims
• 8:26
  the
• 8:26
  #1
• 8:27
  spot
• 8:27
  on
• 8:27
  Google
• 8:27
  DeepMind's
• 8:28
  DeepSearchQA
• 8:28
  benchmark
• 8:29
  and
• 8:29
  supports
• 8:30
  300+
• 8:31
  models
• 8:31
  across
• 8:31
  all
• 8:32
  major
• 8:32
  providers.
• 8:33
  No
• 8:33
  terminal,
• 8:33
  no
• 8:34
  setup
• 8:34
  —
• 8:34
  prompt
• 8:35
  and
• 8:35
  a
• 8:35
  swarm
• 8:35
  goes
• 8:36
  to
• 8:36
  work.
• 8:36
  Users
• 8:37
  consistently
• 8:37
  highlight
• 8:38
  the
• 8:38
  canvas
• 8:38
  view
• 8:39
  as
• 8:39
  the
• 8:39
  feature
• 8:39
  that
• 8:39
  makes
• 8:40
  long-running
• 8:40
  agent
• 8:40
  tasks
• 8:41
  tolerable.
• 8:42
  Whether
• 8:42
  canvas-based
• 8:43
  orchestration
• 8:43
  proves
• 8:44
  more
• 8:44
  effective
• 8:45
  than
• 8:45
  Cruz's
• 8:45
  Git-based
• 8:46
  approach
• 8:46
  remains
• 8:47
  an
• 8:47
  open
• 8:47
  empirical
• 8:47
  question.
• 8:48
  The
• 8:49
  deeper
• 8:49
  question
• 8:49
  is
• 8:49
  whether
• 8:50
  any
• 8:50
  of
• 8:50
  this
• 8:50
  orchestration
• 8:51
  actually
• 8:51
  improves
• 8:52
  outcomes,
• 8:52
  or
• 8:53
  merely
• 8:53
  accelerates
• 8:54
  the
• 8:54
  production
• 8:54
  of
• 8:54
  the
• 8:54
  wrong
• 8:54
  thing.
• 8:55
  Andrea
• 8:56
  Laforgia's
• 8:56
  essay
• 8:57
  on
• 8:57
  the
• 8:57
  developer
• 8:57
  productivity
• 8:58
  trap
• 8:58
  traces
• 8:59
  four
• 8:59
  decades
• 8:59
  of
• 8:59
  failed
• 9:00
  measurement
• 9:00
  —
• 9:01
  from
• 9:01
  lines
• 9:01
  of
• 9:01
  code
• 9:01
  through
• 9:02
  story
• 9:02
  points
• 9:02
  through
• 9:02
  DORA
• 9:03
  metrics
• 9:03
  —
• 9:04
  and
• 9:04
  identifies
• 9:04
  three
• 9:05
  concepts
• 9:05
  that
• 9:05
  explain
• 9:06
  every
• 9:06
  failure:
• 9:07
  the
• 9:07
  McNamara
• 9:08
  Fallacy
• 9:08
  (ignoring
• 9:08
  what
• 9:09
  cannot
• 9:09
  be
• 9:09
  counted),
• 9:10
  Goodhart's
• 9:10
  Law
• 9:11
  (when
• 9:11
  a
• 9:11
  measure
• 9:11
  becomes
• 9:12
  a
• 9:12
  target,
• 9:12
  it
• 9:12
  ceases
• 9:13
  to
• 9:13
  be
• 9:13
  a
• 9:13
  good
• 9:13
  measure),
• 9:14
  and
• 9:14
  surrogation
• 9:15
  (the
• 9:15
  nonconscious
• 9:16
  process
• 9:16
  by
• 9:16
  which
• 9:17
  people
• 9:17
  forget
• 9:17
  what
• 9:17
  the
• 9:17
  metric
• 9:18
  was
• 9:18
  supposed
• 9:18
  to
• 9:18
  represent
• 9:19
  and
• 9:19
  treat
• 9:20
  the
• 9:20
  number
• 9:20
  itself
• 9:20
  as
• 9:20
  the
• 9:20
  goal).
• 9:21
  The
• 9:22
  research
• 9:22
  finding
• 9:22
  that
• 9:23
  AI
• 9:23
  amplifies
• 9:23
  whatever
• 9:24
  is
• 9:24
  already
• 9:24
  there
• 9:24
  —
• 9:25
  good
• 9:25
  practices
• 9:26
  and
• 9:26
  bad
• 9:26
  metrics
• 9:27
  alike
• 9:27
  —
• 9:28
  is
• 9:28
  the
• 9:28
  essay's
• 9:28
  sharpest
• 9:29
  point.
• 9:29
  Even
• 9:30
  the
• 9:30
  CEO
• 9:30
  of
• 9:30
  Cursor
• 9:31
  has
• 9:31
  warned
• 9:31
  publicly
• 9:32
  that
• 9:32
  developers
• 9:33
  accept
• 9:33
  AI-generated
• 9:34
  code
• 9:34
  "simply
• 9:35
  because
• 9:35
  it
• 9:35
  appears
• 9:36
  to
• 9:36
  work,
• 9:36
  without
• 9:37
  properly
• 9:37
  reviewing
• 9:37
  its
• 9:38
  structure,
• 9:38
  logic
• 9:39
  and
• 9:39
  long-term
• 9:40
  impact."
• 9:41
  The
• 9:41
  2025
• 9:42
  DORA
• 9:42
  Report
• 9:43
  abandoned
• 9:43
  its
• 9:43
  low/medium/high/elite
• 9:46
  performance
• 9:46
  tiers
• 9:46
  entirely,
• 9:47
  replacing
• 9:48
  them
• 9:48
  with
• 9:48
  seven
• 9:48
  team
• 9:49
  archetypes.
• 9:50
  The
• 9:50
  message:
• 9:51
  productivity
• 9:51
  cannot
• 9:52
  be
• 9:52
  reduced
• 9:52
  to
• 9:52
  a
• 9:53
  single
• 9:53
  dimension.
• 9:54
  Bryan
• 9:54
  Finster
• 9:55
  cuts
• 9:55
  to
• 9:55
  the
• 9:55
  operational
• 9:56
  implication:
• 9:57
  clarity
• 9:58
  was
• 9:58
  always
• 9:58
  the
• 9:58
  bottleneck,
• 9:59
  and
• 9:59
  AI
• 9:59
  did
• 10:00
  not
• 10:00
  create
• 10:00
  the
• 10:00
  problem
• 10:00
  —
• 10:01
  it
• 10:01
  revealed
• 10:02
  it.
• 10:03
  Organizations
• 10:04
  that
• 10:04
  masked
• 10:04
  poor
• 10:04
  product
• 10:05
  clarity
• 10:05
  with
• 10:05
  slow
• 10:06
  delivery
• 10:06
  cycles
• 10:06
  now
• 10:07
  have
• 10:07
  nowhere
• 10:07
  to
• 10:07
  hide.
• 10:08
  AI
• 10:08
  accelerates
• 10:09
  the
• 10:09
  path
• 10:09
  to
• 10:10
  production;
• 10:11
  what
• 10:11
  it
• 10:11
  cannot
• 10:11
  do
• 10:11
  is
• 10:12
  figure
• 10:12
  out
• 10:12
  whether
• 10:12
  you
• 10:12
  are
• 10:13
  building
• 10:13
  the
• 10:13
  right
• 10:13
  thing.
• 10:14
  The
• 10:14
  answer
• 10:15
  is
• 10:15
  not
• 10:15
  a
• 10:15
  return
• 10:15
  to
• 10:15
  waterfall-era
• 10:16
  requirements
• 10:17
  documents.
• 10:18
  It
• 10:18
  is
• 10:18
  a
• 10:19
  practice
• 10:19
  that
• 10:19
  continuously
• 10:20
  creates
• 10:20
  and
• 10:20
  validates
• 10:21
  shared
• 10:21
  understanding
• 10:22
  —
• 10:22
  BDD,
• 10:23
  story
• 10:24
  refinement,
• 10:24
  spec-driven
• 10:25
  development.
• 10:26
  The
• 10:27
  Hammer
• 10:27
  Problem
• 10:27
  essay
• 10:27
  from
• 10:28
  Full
• 10:28
  Hoffman
• 10:28
  completes
• 10:29
  this
• 10:29
  arc
• 10:29
  with
• 10:29
  a
• 10:29
  filter
• 10:30
  any
• 10:30
  AI
• 10:30
  tooling
• 10:31
  investment
• 10:31
  should
• 10:31
  pass:
• 10:32
  if
• 10:33
  the
• 10:33
  model
• 10:33
  gets
• 10:33
  10x
• 10:33
  better,
• 10:34
  is
• 10:34
  this
• 10:34
  still
• 10:35
  needed?
• 10:36
  Data
• 10:36
  infrastructure
• 10:36
  survives.
• 10:38
  Compute
• 10:38
  primitives
• 10:38
  survive.
• 10:40
  Tool
• 10:40
  access
• 10:40
  survives.
• 10:42
  Human
• 10:42
  judgment
• 10:42
  survives.
• 10:43
  Orchestration
• 10:44
  frameworks,
• 10:45
  context
• 10:45
  management
• 10:46
  tools,
• 10:47
  multi-agent
• 10:47
  debate
• 10:48
  systems,
• 10:48
  and
• 10:49
  prompt
• 10:49
  management
• 10:49
  platforms
• 10:50
  do
• 10:50
  not
• 10:50
  —
• 10:51
  the
• 10:51
  model
• 10:51
  handles
• 10:52
  the
• 10:52
  chain,
• 10:53
  reads
• 10:53
  the
• 10:53
  repo,
• 10:54
  thinks
• 10:54
  harder,
• 10:54
  and
• 10:55
  makes
• 10:55
  the
• 10:55
  conversation
• 10:55
  the
• 10:56
  prompt.
• 10:57
  Every
• 10:57
  one
• 10:57
  of
• 10:58
  these
• 10:58
  solved
• 10:58
  a
• 10:58
  real
• 10:59
  problem
• 10:59
  at
• 10:59
  a
• 10:59
  specific
• 11:00
  moment.
• 11:01
  Every
• 11:01
  one
• 11:01
  of
• 11:01
  them
• 11:02
  is
• 11:02
  a
• 11:02
  hammer.
• 11:03
  And
• 11:03
  in
• 11:03
  a
• 11:03
  landscape
• 11:04
  where
• 11:04
  the
• 11:04
  underlying
• 11:05
  models
• 11:05
  are
• 11:05
  improving
• 11:06
  exponentially,
• 11:06
  the
• 11:06
  moment
• 11:07
  passes
• 11:07
  fast.
• 11:12
  A
• 11:12
  new
• 11:12
  paper
• 11:12
  introduces
• 11:13
  structured
• 11:13
  distillation
• 11:14
  for
• 11:14
  personalized
• 11:15
  agent
• 11:15
  memory,
• 11:16
  achieving
• 11:16
  11x
• 11:17
  token
• 11:17
  compression
• 11:18
  while
• 11:18
  preserving
• 11:18
  96%
• 11:19
  of
• 11:19
  retrieval
• 11:20
  quality.
• 11:20
  The
• 11:21
  method
• 11:21
  distills
• 11:21
  each
• 11:21
  conversation
• 11:22
  exchange
• 11:22
  into
• 11:23
  a
• 11:23
  compound
• 11:23
  object
• 11:23
  with
• 11:24
  four
• 11:24
  fields:
• 11:25
  a
• 11:25
  summary
• 11:25
  of
• 11:25
  what
• 11:25
  was
• 11:25
  accomplished,
• 11:26
  a
• 11:26
  distinguishing
• 11:27
  technical
• 11:27
  detail,
• 11:28
  thematic
• 11:29
  room
• 11:29
  assignments,
• 11:30
  and
• 11:30
  regex-extracted
• 11:31
  file
• 11:31
  references.
• 11:32
  Applied
• 11:32
  to
• 11:32
  4,182
• 11:34
  conversations
• 11:35
  (14,340
• 11:36
  exchanges)
• 11:37
  from
• 11:37
  six
• 11:37
  software
• 11:38
  engineering
• 11:38
  projects
• 11:38
  in
• 11:39
  a
• 11:39
  single-user
• 11:39
  corpus,
• 11:40
  the
• 11:40
  method
• 11:40
  reduces
• 11:41
  average
• 11:41
  exchange
• 11:42
  length
• 11:42
  from
• 11:42
  371
• 11:43
  to
• 11:43
  38
• 11:43
  tokens.
• 11:45
  At
• 11:45
  that
• 11:45
  compression
• 11:45
  ratio,
• 11:46
  1,000
• 11:46
  conversation
• 11:47
  exchanges
• 11:47
  fit
• 11:48
  in
• 11:48
  39,000
• 11:48
  tokens
• 11:49
  instead
• 11:49
  of
• 11:49
  407,000
• 11:50
  —
• 11:50
  the
• 11:50
  difference
• 11:51
  between
• 11:51
  consuming
• 11:51
  most
• 11:52
  of
• 11:52
  a
• 11:52
  context
• 11:52
  window
• 11:53
  and
• 11:53
  leaving
• 11:53
  room
• 11:53
  for
• 11:53
  actual
• 11:54
  work.
• 11:54
  The
• 11:55
  evaluation
• 11:55
  is
• 11:55
  unusually
• 11:56
  rigorous:
• 11:57
  201
• 11:57
  recall-oriented
• 11:58
  queries,
• 11:59
  107
• 12:00
  configurations
• 12:00
  spanning
• 12:01
  10
• 12:01
  search
• 12:01
  modes,
• 12:02
  and
• 12:02
  5
• 12:02
  LLM
• 12:03
  graders
• 12:03
  producing
• 12:03
  214,519
• 12:05
  consensus-graded
• 12:06
  query-result
• 12:07
  pairs.
• 12:07
  The
• 12:07
  result
• 12:08
  is
• 12:08
  mechanism-dependent
• 12:09
  in
• 12:09
  a
• 12:09
  way
• 12:09
  that
• 12:09
  matters
• 12:09
  for
• 12:09
  practitioners.
• 12:10
  Vector
• 12:11
  search
• 12:11
  (HNSW,
• 12:12
  exact
• 12:13
  cosine)
• 12:13
  shows
• 12:14
  no
• 12:14
  statistically
• 12:14
  significant
• 12:15
  degradation
• 12:15
  after
• 12:15
  Bonferroni
• 12:16
  correction
• 12:16
  —
• 12:17
  all
• 12:17
  20
• 12:17
  configurations
• 12:18
  hold
• 12:18
  up.
• 12:19
  BM25
• 12:20
  keyword
• 12:20
  search
• 12:20
  degrades
• 12:21
  significantly
• 12:21
  across
• 12:22
  all
• 12:22
  20
• 12:22
  configurations,
• 12:23
  with
• 12:23
  effect
• 12:23
  sizes
• 12:24
  reaching
• 12:24
  0.756.
• 12:26
  The
• 12:26
  explanation
• 12:27
  is
• 12:27
  straightforward:
• 12:28
  11x
• 12:28
  compression
• 12:29
  removes
• 12:29
  the
• 12:29
  verbatim
• 12:30
  vocabulary
• 12:30
  that
• 12:30
  keyword
• 12:31
  matching
• 12:31
  requires,
• 12:32
  but
• 12:32
  vector
• 12:33
  similarity
• 12:33
  captures
• 12:34
  semantic
• 12:34
  content
• 12:34
  that
• 12:35
  survives
• 12:35
  the
• 12:35
  compression.
• 12:36
  The
• 12:36
  most
• 12:37
  interesting
• 12:37
  finding
• 12:37
  is
• 12:38
  that
• 12:38
  cross-layer
• 12:38
  retrieval
• 12:38
  —
• 12:39
  combining
• 12:39
  verbatim
• 12:40
  keyword
• 12:40
  search
• 12:40
  with
• 12:40
  distilled
• 12:41
  vector
• 12:41
  search
• 12:41
  —
• 12:42
  slightly
• 12:42
  exceeds
• 12:43
  the
• 12:43
  best
• 12:43
  pure
• 12:43
  verbatim
• 12:44
  baseline
• 12:44
  (MRR
• 12:45
  0.759
• 12:46
  vs
• 12:46
  0.745).
• 12:48
  The
• 12:48
  signals
• 12:48
  are
• 12:48
  complementary.
• 12:49
  The
• 12:49
  paper's
• 12:50
  "surviving
• 12:51
  vocabulary
• 12:51
  principle"
• 12:51
  —
• 12:52
  constraining
• 12:52
  the
• 12:53
  LLM
• 12:53
  to
• 12:53
  reuse
• 12:53
  participants'
• 12:54
  specific
• 12:54
  terms
• 12:55
  rather
• 12:55
  than
• 12:55
  paraphrasing
• 12:56
  freely
• 12:56
  —
• 12:56
  is
• 12:56
  worth
• 12:57
  noting.
• 12:57
  Only
• 12:58
  27%
• 12:59
  of
• 12:59
  high-IDF
• 12:59
  tokens
• 13:00
  survive
• 13:00
  distillation,
• 13:01
  but
• 13:01
  96.8%
• 13:03
  of
• 13:03
  tokens
• 13:03
  users
• 13:03
  actually
• 13:04
  search
• 13:04
  for
• 13:04
  are
• 13:04
  retained.
• 13:05
  The
• 13:05
  system
• 13:06
  keeps
• 13:06
  what
• 13:06
  people
• 13:06
  look
• 13:06
  for
• 13:07
  and
• 13:07
  discards
• 13:07
  what
• 13:07
  they
• 13:07
  do
• 13:07
  not.
• 13:08
  Black
• 13:08
  Forest
• 13:08
  Labs
• 13:09
  has
• 13:09
  released
• 13:09
  Self-Flow,
• 13:10
  a
• 13:10
  self-supervised
• 13:11
  flow
• 13:11
  matching
• 13:11
  framework
• 13:12
  that
• 13:12
  combines
• 13:12
  the
• 13:12
  flow
• 13:12
  matching
• 13:13
  training
• 13:13
  objective
• 13:13
  with
• 13:14
  a
• 13:14
  self-supervised
• 13:14
  feature
• 13:15
  reconstruction
• 13:15
  objective.
• 13:16
  The
• 13:16
  model
• 13:16
  uses
• 13:17
  per-token
• 13:17
  timestep
• 13:17
  conditioning
• 13:18
  —
• 13:18
  each
• 13:18
  token
• 13:19
  can
• 13:19
  have
• 13:19
  a
• 13:19
  different
• 13:19
  noise
• 13:19
  level
• 13:20
  during
• 13:20
  training
• 13:20
  —
• 13:20
  and
• 13:21
  self-distillation
• 13:21
  from
• 13:21
  a
• 13:22
  teacher
• 13:22
  at
• 13:22
  layer
• 13:22
  20
• 13:22
  (EMA)
• 13:23
  to
• 13:23
  a
• 13:23
  student
• 13:23
  at
• 13:23
  layer
• 13:24
  8.
• 13:24
  Built
• 13:25
  on
• 13:25
  SiT-XL/2,
• 13:26
  the
• 13:26
  model
• 13:27
  generates
• 13:27
  ImageNet
• 13:27
  256x256
• 13:30
  samples
• 13:30
  without
• 13:30
  classifier-free
• 13:31
  guidance
• 13:31
  (cfg
• 13:32
  scale=1.0),
• 13:34
  a
• 13:34
  notable
• 13:34
  departure
• 13:35
  from
• 13:35
  the
• 13:35
  guidance-dependent
• 13:35
  generation
• 13:36
  that
• 13:36
  dominates
• 13:36
  the
• 13:36
  field.
• 13:37
  On
• 13:38
  the
• 13:38
  local
• 13:38
  inference
• 13:38
  front,
• 13:39
  GATED
• 13:39
  DELTA
• 13:40
  NET
• 13:40
  support
• 13:40
  for
• 13:40
  Vulkan
• 13:41
  has
• 13:41
  been
• 13:41
  merged
• 13:42
  into
• 13:42
  llama.cpp,
• 13:44
  delivering
• 13:44
  measurable
• 13:44
  performance
• 13:45
  improvements
• 13:45
  on
• 13:46
  consumer
• 13:46
  AMD
• 13:46
  hardware.
• 13:47
  An
• 13:47
  RX
• 13:48
  7800XT
• 13:50
  user
• 13:50
  reports
• 13:50
  token
• 13:51
  generation
• 13:51
  jumping
• 13:51
  from
• 13:52
  approximately
• 13:52
  28
• 13:53
  tokens/second
• 13:54
  to
• 13:54
  36
• 13:54
  tokens/second
• 13:55
  on
• 13:55
  Qwen
• 13:56
  3.5
• 13:56
  27B
• 13:57
  —
• 13:58
  a
• 13:58
  roughly
• 13:58
  29%
• 13:59
  improvement
• 13:59
  from
• 14:00
  a
• 14:00
  single
• 14:00
  kernel
• 14:00
  merge.
• 14:03
  Kurt
• 14:03
  Cagle's
• 14:04
  deep
• 14:04
  dive
• 14:04
  into
• 14:05
  the
• 14:05
  open
• 14:05
  world
• 14:05
  /
• 14:06
  closed
• 14:06
  world
• 14:06
  assumption
• 14:07
  (OWA
• 14:07
  vs.
• 14:08
  CWA)
• 14:08
  in
• 14:09
  knowledge
• 14:09
  representation
• 14:10
  cuts
• 14:10
  to
• 14:10
  a
• 14:10
  tension
• 14:10
  that
• 14:11
  most
• 14:11
  practitioners
• 14:11
  encounter
• 14:12
  without
• 14:12
  naming.
• 14:13
  The
• 14:13
  closed
• 14:14
  world
• 14:14
  assumption
• 14:14
  holds
• 14:15
  that
• 14:15
  if
• 14:15
  a
• 14:15
  statement
• 14:15
  cannot
• 14:16
  be
• 14:16
  proven
• 14:16
  true
• 14:16
  from
• 14:17
  available
• 14:17
  information,
• 14:18
  it
• 14:18
  is
• 14:18
  false
• 14:18
  —
• 14:19
  the
• 14:19
  foundation
• 14:19
  of
• 14:19
  every
• 14:20
  relational
• 14:20
  database
• 14:21
  ever
• 14:21
  built.
• 14:22
  The
• 14:22
  open
• 14:23
  world
• 14:23
  assumption
• 14:23
  holds
• 14:24
  that
• 14:24
  absence
• 14:24
  of
• 14:24
  evidence
• 14:25
  is
• 14:25
  merely
• 14:25
  absence
• 14:26
  of
• 14:26
  evidence
• 14:26
  —
• 14:27
  the
• 14:27
  foundation
• 14:27
  of
• 14:27
  OWL
• 14:28
  and
• 14:28
  the
• 14:28
  Semantic
• 14:29
  Web.
• 14:30
  The
• 14:30
  practical
• 14:31
  problem:
• 14:31
  RDF
• 14:32
  and
• 14:32
  OWL
• 14:33
  operate
• 14:33
  under
• 14:33
  OWA,
• 14:34
  SPARQL
• 14:36
  users
• 14:36
  implicitly
• 14:37
  expect
• 14:37
  CWA
• 14:38
  behavior,
• 14:38
  and
• 14:38
  SHACL
• 14:39
  introduces
• 14:39
  explicit
• 14:40
  constraint
• 14:40
  enforcement
• 14:41
  that
• 14:41
  presupposes
• 14:41
  closure.
• 14:42
  The
• 14:42
  moment
• 14:42
  SHACL
• 14:43
  rules
• 14:43
  generate
• 14:44
  new
• 14:44
  triples
• 14:44
  that
• 14:44
  feed
• 14:44
  back
• 14:45
  into
• 14:45
  subsequent
• 14:45
  validation,
• 14:46
  which
• 14:46
  assumption
• 14:47
  governs
• 14:47
  the
• 14:47
  combined
• 14:48
  system
• 14:48
  becomes
• 14:49
  genuinely
• 14:49
  difficult
• 14:49
  to
• 14:49
  answer.
• 14:51
  Cagle
• 14:51
  argues
• 14:51
  the
• 14:51
  direction
• 14:52
  of
• 14:52
  travel
• 14:52
  is
• 14:52
  toward
• 14:53
  closed-world
• 14:53
  systems,
• 14:54
  driven
• 14:54
  from
• 14:54
  multiple
• 14:55
  directions:
• 14:56
  performance
• 14:56
  pressures,
• 14:57
  query
• 14:57
  complexity,
• 14:58
  security
• 14:58
  concerns,
• 14:59
  the
• 14:59
  demands
• 15:00
  of
• 15:00
  hypergraph
• 15:00
  architectures,
• 15:01
  and
• 15:01
  the
• 15:01
  growing
• 15:02
  need
• 15:02
  to
• 15:02
  connect
• 15:02
  formal
• 15:03
  graph-based
• 15:03
  systems
• 15:04
  with
• 15:04
  AI.
• 15:05
  Generative
• 15:05
  AI
• 15:05
  solutions
• 15:06
  are
• 15:06
  moving
• 15:06
  toward
• 15:07
  graphs
• 15:07
  that
• 15:07
  presume
• 15:07
  CWA
• 15:08
  —
• 15:09
  agentic
• 15:09
  AI
• 15:10
  architectures
• 15:10
  talk
• 15:11
  to
• 15:11
  programmatic
• 15:11
  endpoints
• 15:12
  rather
• 15:12
  than
• 15:12
  exposing
• 15:13
  SPARQL
• 15:15
  directly.
• 15:16
  SHACL
• 15:16
  1.2
• 15:17
  has
• 15:17
  built-in
• 15:17
  support
• 15:18
  for
• 15:18
  RDF-Star
• 15:19
  reification
• 15:19
  that
• 15:20
  OWL
• 15:20
  lacks,
• 15:21
  giving
• 15:21
  the
• 15:21
  closed-world
• 15:22
  tooling
• 15:22
  an
• 15:22
  adoption
• 15:23
  edge.
• 15:23
  The
• 15:24
  practical
• 15:24
  heuristic:
• 15:25
  use
• 15:25
  OWA
• 15:26
  when
• 15:26
  modeling
• 15:26
  the
• 15:26
  domain
• 15:27
  (what
• 15:27
  kinds
• 15:27
  of
• 15:28
  things
• 15:28
  exist),
• 15:28
  use
• 15:29
  CWA
• 15:29
  when
• 15:30
  modeling
• 15:30
  the
• 15:30
  pipeline
• 15:30
  (what
• 15:31
  constraints
• 15:31
  must
• 15:31
  hold
• 15:32
  for
• 15:32
  a
• 15:32
  record
• 15:32
  to
• 15:32
  be
• 15:32
  processed).
• 15:33
  The
• 15:33
  W3C
• 15:34
  has
• 15:34
  launched
• 15:35
  a
• 15:35
  Context
• 15:35
  Graphs
• 15:36
  Community
• 15:36
  Group
• 15:37
  to
• 15:37
  address
• 15:37
  a
• 15:37
  related
• 15:37
  structural
• 15:38
  problem:
• 15:39
  the
• 15:39
  gap
• 15:39
  between
• 15:39
  global
• 15:40
  knowledge
• 15:40
  representations
• 15:41
  and
• 15:41
  local
• 15:42
  interpretation
• 15:42
  contexts.
• 15:44
  The
• 15:44
  group
• 15:44
  aims
• 15:45
  to
• 15:45
  formalize
• 15:45
  a
• 15:45
  core
• 15:46
  data
• 15:46
  model
• 15:46
  for
• 15:46
  "contextual
• 15:47
  prerequisites"
• 15:47
  —
• 15:48
  structured
• 15:49
  representations
• 15:50
  of
• 15:50
  what
• 15:50
  context
• 15:50
  is
• 15:51
  required
• 15:51
  for
• 15:51
  valid
• 15:52
  interpretation,
• 15:53
  their
• 15:53
  dependencies,
• 15:54
  and
• 15:54
  their
• 15:54
  resolution
• 15:55
  status.
• 15:56
  The
• 15:56
  motivation
• 15:56
  is
• 15:56
  practical:
• 15:57
  terms
• 15:58
  carry
• 15:58
  different
• 15:58
  meanings
• 15:59
  across
• 15:59
  organizational,
• 16:00
  temporal,
• 16:01
  or
• 16:01
  operational
• 16:02
  boundaries,
• 16:03
  and
• 16:03
  assumptions
• 16:04
  embedded
• 16:04
  in
• 16:04
  global
• 16:04
  models
• 16:05
  frequently
• 16:05
  do
• 16:06
  not
• 16:06
  hold
• 16:06
  locally.
• 16:07
  The
• 16:07
  group
• 16:08
  intends
• 16:08
  to
• 16:08
  publish
• 16:09
  specification-style
• 16:10
  documents
• 16:10
  covering
• 16:10
  use
• 16:11
  cases
• 16:11
  from
• 16:11
  knowledge
• 16:12
  management
• 16:12
  to
• 16:12
  human-AI
• 16:13
  workflows.
• 16:16
  Cognitum's
• 16:17
  Seed
• 16:17
  is
• 16:17
  a
• 16:17
  vector
• 16:18
  intelligence
• 16:18
  appliance
• 16:19
  running
• 16:19
  on
• 16:19
  a
• 16:19
  Pi
• 16:19
  Zero
• 16:20
  2
• 16:20
  W
• 16:20
  that
• 16:20
  exposes
• 16:21
  98
• 16:21
  API
• 16:22
  endpoints
• 16:22
  from
• 16:22
  a
• 16:22
  device
• 16:23
  that
• 16:23
  fits
• 16:23
  in
• 16:23
  a
• 16:23
  shirt
• 16:24
  pocket.
• 16:24
  The
• 16:24
  system
• 16:25
  implements
• 16:25
  an
• 16:26
  RVF
• 16:26
  (RuVector
• 16:27
  Format)
• 16:27
  vector
• 16:27
  store
• 16:28
  with
• 16:28
  cosine
• 16:28
  similarity
• 16:29
  search,
• 16:29
  a
• 16:29
  kNN
• 16:30
  graph
• 16:30
  with
• 16:30
  structural
• 16:31
  boundary
• 16:31
  analysis
• 16:32
  via
• 16:32
  Stoer-Wagner
• 16:33
  min-cut,
• 16:33
  a
• 16:34
  SHA-256
• 16:36
  witness
• 16:36
  chain
• 16:37
  for
• 16:37
  tamper-evident
• 16:37
  audit
• 16:38
  logging,
• 16:38
  Ed25519
• 16:40
  custody
• 16:40
  signing
• 16:40
  for
• 16:40
  attestation,
• 16:41
  a
• 16:41
  6-channel
• 16:42
  synthetic
• 16:42
  sensor
• 16:43
  pipeline
• 16:43
  with
• 16:43
  drift
• 16:43
  detection
• 16:44
  (Page-Hinkley,
• 16:45
  ADWIN,
• 16:46
  and
• 16:46
  a
• 16:46
  64-neuron
• 16:47
  reservoir
• 16:47
  echo
• 16:47
  state
• 16:48
  network),
• 16:48
  reflex
• 16:49
  arc
• 16:49
  rules
• 16:50
  mapping
• 16:50
  sensor
• 16:50
  conditions
• 16:51
  to
• 16:51
  actuator
• 16:51
  outputs,
• 16:52
  a
• 16:53
  thermal
• 16:53
  DVFS
• 16:54
  governor
• 16:54
  with
• 16:54
  demand-aware
• 16:55
  frequency
• 16:55
  scaling,
• 16:56
  and
• 16:56
  a
• 16:56
  cognitive
• 16:57
  container
• 16:57
  performing
• 16:58
  spectral
• 16:58
  graph
• 16:58
  analysis
• 16:59
  with
• 16:59
  its
• 16:59
  own
• 16:59
  secondary
• 17:00
  witness
• 17:00
  chain.
• 17:01
  The
• 17:01
  pairing
• 17:02
  model
• 17:02
  requires
• 17:02
  physical
• 17:03
  button
• 17:03
  press
• 17:03
  for
• 17:03
  write
• 17:04
  access
• 17:04
  —
• 17:05
  a
• 17:05
  hardware
• 17:05
  trust
• 17:05
  root
• 17:06
  that
• 17:06
  no
• 17:06
  amount
• 17:06
  of
• 17:06
  software
• 17:06
  compromise
• 17:07
  can
• 17:07
  bypass
• 17:08
  remotely.
• 17:08
  The
• 17:09
  broader
• 17:09
  vision,
• 17:09
  as
• 17:10
  articulated
• 17:10
  by
• 17:10
  Reuven
• 17:11
  Cohen,
• 17:11
  is
• 17:12
  that
• 17:12
  the
• 17:12
  system
• 17:12
  continuously
• 17:13
  ingests
• 17:13
  live
• 17:14
  feeds
• 17:14
  from
• 17:14
  dozens
• 17:14
  of
• 17:15
  domains
• 17:15
  —
• 17:16
  NASA
• 17:16
  space
• 17:16
  weather
• 17:16
  data,
• 17:17
  USGS
• 17:18
  seismic
• 17:18
  activity,
• 17:20
  NOAA
• 17:20
  climate
• 17:20
  signals,
• 17:20
  arXiv
• 17:20
  papers,
• 17:21
  financial
• 17:21
  markets,
• 17:22
  genomics
• 17:23
  databases
• 17:23
  —
• 17:24
  with
• 17:24
  each
• 17:24
  piece
• 17:25
  becoming
• 17:25
  a
• 17:25
  structured
• 17:25
  memory
• 17:26
  where
• 17:26
  cross-domain
• 17:27
  relationships
• 17:27
  can
• 17:27
  be
• 17:28
  analyzed
• 17:28
  over
• 17:28
  time.
• 17:29
  After
• 17:29
  a
• 17:29
  short
• 17:30
  run,
• 17:30
  the
• 17:30
  system
• 17:31
  stored
• 17:31
  over
• 17:31
  1,400
• 17:32
  persistent
• 17:33
  memories
• 17:33
  spanning
• 17:34
  ten
• 17:34
  scientific
• 17:34
  domains,
• 17:35
  surfacing
• 17:36
  patterns
• 17:36
  like
• 17:36
  correlations
• 17:37
  between
• 17:37
  solar
• 17:38
  activity
• 17:38
  and
• 17:38
  seismic
• 17:39
  events.
• 17:40
  The
• 17:40
  architecture
• 17:41
  positions
• 17:41
  RuVector
• 17:42
  not
• 17:42
  as
• 17:42
  a
• 17:42
  database
• 17:43
  you
• 17:43
  add
• 17:43
  to
• 17:43
  your
• 17:43
  stack
• 17:44
  but
• 17:44
  as
• 17:44
  the
• 17:44
  entire
• 17:45
  self-learning,
• 17:45
  self-optimizing
• 17:46
  layer.
• 17:47
  The
• 17:47
  Ruvix
• 17:48
  crate
• 17:48
  extends
• 17:49
  RuVector
• 17:49
  into
• 17:49
  a
• 17:49
  cognition
• 17:50
  kernel
• 17:50
  —
• 17:51
  a
• 17:51
  Rust
• 17:51
  library
• 17:52
  implementing
• 17:52
  a
• 17:52
  full-cycle
• 17:53
  compute
• 17:53
  substrate
• 17:54
  from
• 17:54
  PBFT
• 17:55
  consensus
• 17:55
  to
• 17:55
  a
• 17:56
  CLI
• 17:56
  shell.
• 17:57
  The
• 17:58
  architecture
• 17:58
  spans
• 17:59
  vector
• 17:59
  storage
• 17:59
  with
• 18:00
  GNN-enhanced
• 18:00
  search,
• 18:01
  graph
• 18:01
  neural
• 18:02
  network
• 18:02
  re-ranking
• 18:02
  (comparing
• 18:03
  GCN,
• 18:03
  GAT,
• 18:04
  and
• 18:04
  GraphSAGE
• 18:05
  approaches),
• 18:06
  spectral
• 18:06
  analysis
• 18:07
  for
• 18:07
  structural
• 18:07
  health
• 18:07
  monitoring,
• 18:08
  and
• 18:08
  a
• 18:08
  witness
• 18:09
  chain
• 18:09
  for
• 18:09
  immutable
• 18:10
  audit
• 18:10
  trails.
• 18:11
  The
• 18:11
  min-cut
• 18:12
  examples
• 18:12
  demonstrate
• 18:13
  Stoer-Wagner
• 18:13
  applied
• 18:14
  to
• 18:14
  network
• 18:14
  reliability
• 18:15
  analysis,
• 18:16
  image
• 18:16
  segmentation,
• 18:17
  social
• 18:18
  network
• 18:18
  community
• 18:18
  detection,
• 18:19
  supply
• 18:20
  chain
• 18:20
  vulnerability
• 18:21
  mapping,
• 18:21
  and
• 18:22
  circuit
• 18:22
  board
• 18:22
  analysis.
• 18:23
  Boundary
• 18:24
  analysis
• 18:24
  drives
• 18:25
  the
• 18:25
  thermal
• 18:25
  governor
• 18:25
  —
• 18:26
  structural
• 18:27
  fragility
• 18:27
  of
• 18:27
  the
• 18:27
  vector
• 18:28
  store's
• 18:28
  kNN
• 18:29
  graph
• 18:29
  directly
• 18:30
  influences
• 18:30
  CPU
• 18:31
  frequency
• 18:31
  scaling,
• 18:32
  a
• 18:32
  tight
• 18:33
  feedback
• 18:33
  loop
• 18:33
  between
• 18:34
  data
• 18:34
  structure
• 18:34
  health
• 18:35
  and
• 18:35
  hardware
• 18:35
  resource
• 18:36
  allocation.