Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 08, 2026

Transcript

AGI Dreams Podcast – June 08, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Linus
• 0:00
  Torvalds
• 0:00
  has
• 0:00
  never
• 0:01
  been
• 0:01
  one
• 0:01
  to
• 0:01
  sugarcoat,
• 0:02
  and
• 0:02
  at
• 0:02
  Open
• 0:03
  Source
• 0:03
  Summit
• 0:03
  North
• 0:04
  America
• 0:04
  he
• 0:05
  delivered
• 0:05
  exactly
• 0:05
  the
• 0:05
  grounding
• 0:06
  the
• 0:06
  AI-hype
• 0:06
  cycle
• 0:07
  needs.
• 0:07
  Asked
• 0:08
  about
• 0:08
  claims
• 0:08
  that
• 0:08
  "99%
• 0:09
  of
• 0:09
  code
• 0:09
  is
• 0:10
  AI-generated,"
• 0:11
  Torvalds
• 0:11
  pushed
• 0:11
  back
• 0:12
  hard:
• 0:12
  "I
• 0:12
  literally
• 0:13
  get
• 0:13
  angry,
• 0:13
  because
• 0:14
  those
• 0:14
  same
• 0:14
  people
• 0:14
  —
• 0:15
  I
• 0:15
  can
• 0:15
  pretty
• 0:15
  much
• 0:15
  guarantee
• 0:16
  —
• 0:16
  that
• 0:16
  100%
• 0:17
  of
• 0:17
  their
• 0:17
  code
• 0:17
  is
• 0:17
  written
• 0:17
  by
• 0:17
  compilers.
• 0:18
  But
• 0:19
  they
• 0:19
  never
• 0:19
  say
• 0:19
  that."
• 0:20
  His
• 0:20
  framing
• 0:20
  is
• 0:21
  precise:
• 0:21
  AI
• 0:22
  is
• 0:22
  a
• 0:22
  productivity
• 0:22
  multiplier
• 0:23
  on
• 0:23
  the
• 0:23
  order
• 0:24
  of
• 0:24
  10x,
• 0:24
  impressive
• 0:25
  but
• 0:25
  actually
• 0:26
  100x
• 0:26
  less
• 0:27
  than
• 0:27
  the
• 0:27
  productivity
• 0:27
  gains
• 0:28
  compilers
• 0:28
  delivered
• 0:28
  over
• 0:29
  the
• 0:29
  decades.
• 0:30
  The
• 0:30
  real
• 0:30
  work
• 0:30
  remains
• 0:31
  understanding
• 0:31
  systems
• 0:31
  deeply
• 0:32
  enough
• 0:32
  to
• 0:32
  direct
• 0:32
  the
• 0:32
  tools.
• 0:33
  "People
• 0:34
  who
• 0:34
  know
• 0:34
  what
• 0:34
  they're
• 0:34
  doing
• 0:34
  to
• 0:34
  understand
• 0:35
  systems
• 0:35
  will
• 0:35
  be
• 0:36
  able
• 0:36
  to
• 0:36
  prompt
• 0:36
  tools
• 0:36
  to
• 0:36
  write
• 0:37
  good
• 0:37
  code.
• 0:38
  People
• 0:38
  who
• 0:38
  don't
• 0:38
  understand
• 0:39
  the
• 0:39
  complexity
• 0:39
  of
• 0:39
  systems
• 0:40
  will
• 0:40
  also
• 0:40
  prompt
• 0:41
  systems
• 0:41
  and
• 0:41
  write
• 0:41
  processes
• 0:42
  that
• 0:42
  will
• 0:42
  fail."
• 0:43
  The
• 0:43
  maintainer
• 0:44
  burnout
• 0:44
  angle
• 0:44
  cuts
• 0:45
  deeper
• 0:45
  than
• 0:45
  the
• 0:45
  productivity
• 0:46
  debate.
• 0:47
  Torvalds
• 0:47
  described
• 0:48
  a
• 0:48
  surge
• 0:48
  in
• 0:48
  AI-generated
• 0:49
  pull
• 0:49
  requests
• 0:50
  ahead
• 0:50
  of
• 0:50
  the
• 0:50
  Linux
• 0:50
  7.1
• 0:50
  release
• 0:51
  —
• 0:52
  more
• 0:52
  commits
• 0:52
  during
• 0:52
  preparation
• 0:53
  than
• 0:53
  any
• 0:53
  previous
• 0:54
  release
• 0:54
  —
• 0:54
  that
• 0:55
  turned
• 0:55
  out
• 0:55
  to
• 0:55
  be
• 0:55
  AI
• 0:55
  use
• 0:56
  rather
• 0:56
  than
• 0:56
  increased
• 0:56
  interest.
• 0:57
  For
• 0:58
  the
• 0:58
  Linux
• 0:58
  project
• 0:58
  with
• 0:58
  its
• 0:58
  resources,
• 0:59
  this
• 0:59
  is
• 1:00
  manageable.
• 1:01
  For
• 1:01
  the
• 1:01
  thousands
• 1:01
  of
• 1:02
  single-maintainer
• 1:02
  projects,
• 1:03
  it
• 1:03
  is
• 1:03
  existential.
• 1:04
  "Sometimes
• 1:05
  AI
• 1:05
  reports
• 1:06
  a
• 1:06
  bug,
• 1:06
  and
• 1:07
  when
• 1:07
  you
• 1:07
  ask
• 1:07
  for
• 1:08
  more
• 1:08
  information,
• 1:09
  the
• 1:09
  person
• 1:09
  has
• 1:09
  done
• 1:09
  that
• 1:09
  drive-by
• 1:10
  and
• 1:10
  doesn't
• 1:11
  even
• 1:11
  answer
• 1:11
  your
• 1:11
  question."
• 1:12
  Worse,
• 1:13
  some
• 1:13
  companies
• 1:13
  invest
• 1:14
  in
• 1:14
  AI
• 1:14
  bug-finding
• 1:15
  for
• 1:15
  media
• 1:15
  attention
• 1:16
  without
• 1:16
  providing
• 1:16
  patches:
• 1:17
  "strangely,
• 1:18
  none
• 1:18
  of
• 1:18
  these
• 1:19
  came
• 1:19
  with
• 1:19
  a
• 1:19
  patch,
• 1:20
  even
• 1:20
  though
• 1:20
  all
• 1:20
  of
• 1:21
  these
• 1:21
  were
• 1:21
  in
• 1:21
  open
• 1:21
  source
• 1:22
  code."
• 1:22
  That
• 1:23
  productivity
• 1:23
  argument
• 1:24
  cuts
• 1:24
  differently
• 1:24
  depending
• 1:25
  on
• 1:25
  where
• 1:25
  you
• 1:25
  sit
• 1:25
  in
• 1:25
  the
• 1:25
  stack.
• 1:26
  A
• 1:26
  ten-year
• 1:27
  backend
• 1:27
  engineer
• 1:28
  specializing
• 1:28
  in
• 1:28
  fintech
• 1:29
  —
• 1:30
  PCI
• 1:30
  compliance,
• 1:31
  double-entry
• 1:31
  ledgers,
• 1:32
  payment
• 1:32
  idempotency
• 1:33
  —
• 1:34
  posted
• 1:34
  a
• 1:34
  remarkably
• 1:35
  candid
• 1:35
  account
• 1:36
  of
• 1:36
  watching
• 1:36
  each
• 1:36
  pillar
• 1:37
  of
• 1:37
  their
• 1:37
  expertise
• 1:37
  erode.
• 1:38
  Domain
• 1:39
  knowledge?
• 1:39
  Claude
• 1:40
  with
• 1:40
  a
• 1:40
  Sentry
• 1:40
  MCP
• 1:41
  one-shots
• 1:41
  bugs
• 1:42
  across
• 1:42
  distributed
• 1:43
  systems
• 1:44
  that
• 1:44
  used
• 1:44
  to
• 1:44
  take
• 1:44
  two
• 1:44
  days
• 1:45
  of
• 1:45
  full-time
• 1:45
  debugging.
• 1:46
  Debugging
• 1:47
  intuition?
• 1:48
  Opus
• 1:48
  4.8
• 1:49
  with
• 1:49
  the
• 1:49
  DataDog
• 1:49
  MCP
• 1:50
  handles
• 1:50
  90%
• 1:51
  of
• 1:51
  production
• 1:52
  issues,
• 1:52
  including
• 1:52
  bizarre
• 1:53
  race
• 1:53
  conditions
• 1:54
  and
• 1:54
  undocumented
• 1:54
  API
• 1:55
  edge
• 1:55
  cases.
• 1:56
  The
• 1:56
  last
• 1:56
  remaining
• 1:57
  pillar
• 1:57
  —
• 1:58
  code
• 1:58
  architecture
• 1:59
  and
• 1:59
  quality
• 1:59
  —
• 2:00
  is
• 2:00
  being
• 2:00
  reduced
• 2:01
  to
• 2:01
  "taste,"
• 2:02
  and
• 2:02
  the
• 2:02
  industry
• 2:02
  is
• 2:02
  moving
• 2:03
  toward
• 2:03
  codebases
• 2:03
  optimized
• 2:04
  for
• 2:04
  machines
• 2:04
  to
• 2:05
  read,
• 2:05
  not
• 2:05
  humans.
• 2:06
  "All
• 2:06
  my
• 2:07
  finance
• 2:07
  and
• 2:07
  payment
• 2:07
  domain
• 2:08
  expertise,
• 2:09
  all
• 2:09
  the
• 2:09
  debugging
• 2:09
  intuition
• 2:10
  earned
• 2:10
  through
• 2:10
  hours
• 2:11
  of
• 2:11
  sweat
• 2:11
  and
• 2:11
  tears,
• 2:12
  is
• 2:12
  now
• 2:12
  commodity."
• 2:14
  The
• 2:14
  post
• 2:14
  resonated
• 2:15
  because
• 2:15
  the
• 2:15
  author
• 2:16
  is
• 2:16
  not
• 2:16
  panicking
• 2:16
  —
• 2:17
  they
• 2:18
  are
• 2:18
  employed,
• 2:19
  productive,
• 2:20
  and
• 2:20
  clear-eyed
• 2:20
  —
• 2:21
  but
• 2:21
  the
• 2:21
  trajectory
• 2:22
  is
• 2:22
  unmistakable.
• 2:23
  One
• 2:23
  of
• 2:23
  the
• 2:23
  more
• 2:23
  ironic
• 2:24
  consequences
• 2:24
  of
• 2:25
  this
• 2:25
  shift:
• 2:26
  programmers
• 2:26
  who
• 2:26
  spent
• 2:26
  decades
• 2:27
  refusing
• 2:27
  to
• 2:27
  document
• 2:28
  their
• 2:28
  code
• 2:28
  are
• 2:28
  now
• 2:29
  meticulously
• 2:29
  writing
• 2:30
  handoff
• 2:30
  documents,
• 2:31
  structured
• 2:31
  project
• 2:32
  summaries,
• 2:32
  and
• 2:33
  detailed
• 2:33
  CLAUDE.md
• 2:33
  files
• 2:34
  —
• 2:35
  not
• 2:35
  for
• 2:35
  their
• 2:35
  colleagues,
• 2:36
  but
• 2:36
  for
• 2:36
  the
• 2:36
  AI.
• 2:37
  As
• 2:37
  one
• 2:38
  developer
• 2:38
  observed,
• 2:39
  the
• 2:39
  end-of-project
• 2:40
  writeup
• 2:40
  takes
• 2:41
  ten
• 2:41
  seconds
• 2:41
  instead
• 2:42
  of
• 2:42
  an
• 2:42
  hour,
• 2:42
  and
• 2:43
  the
• 2:43
  quality
• 2:43
  is
• 2:43
  roughly
• 2:44
  comparable
• 2:44
  to
• 2:44
  what
• 2:44
  a
• 2:44
  human
• 2:45
  would
• 2:45
  produce.
• 2:46
  The
• 2:46
  incentive
• 2:46
  structure
• 2:47
  that
• 2:47
  documentation
• 2:48
  advocates
• 2:48
  failed
• 2:49
  to
• 2:49
  create
• 2:49
  for
• 2:49
  twenty
• 2:49
  years
• 2:50
  materialized
• 2:50
  overnight
• 2:51
  once
• 2:51
  the
• 2:51
  machine
• 2:52
  literally
• 2:52
  cannot
• 2:52
  function
• 2:53
  without
• 2:53
  externalized
• 2:53
  context.
• 2:55
  The
• 2:55
  training-data
• 2:55
  dimension
• 2:56
  adds
• 2:56
  another
• 2:56
  wrinkle:
• 2:57
  a
• 2:57
  widely
• 2:57
  discussed
• 2:58
  experiment
• 2:58
  trained
• 2:58
  an
• 2:58
  LLM
• 2:59
  on
• 2:59
  700,000
• 3:00
  romance
• 3:00
  novels
• 3:01
  expecting
• 3:01
  domain
• 3:02
  superiority,
• 3:03
  and
• 3:03
  the
• 3:03
  specialist
• 3:04
  model
• 3:04
  was
• 3:04
  less
• 3:04
  creative,
• 3:05
  more
• 3:05
  formulaic,
• 3:06
  and
• 3:06
  worse
• 3:07
  at
• 3:07
  basic
• 3:07
  narrative
• 3:08
  than
• 3:08
  a
• 3:08
  general
• 3:08
  model
• 3:09
  trained
• 3:09
  on
• 3:09
  the
• 3:09
  whole
• 3:09
  messy
• 3:10
  internet.
• 3:10
  The
• 3:11
  same
• 3:11
  pattern
• 3:11
  appears
• 3:12
  in
• 3:12
  code
• 3:12
  —
• 3:13
  models
• 3:13
  that
• 3:13
  see
• 3:13
  unrelated
• 3:14
  material
• 3:14
  write
• 3:15
  better
• 3:15
  code
• 3:15
  because
• 3:16
  they
• 3:16
  have
• 3:16
  learned
• 3:16
  "the
• 3:16
  shape
• 3:17
  of
• 3:17
  more
• 3:17
  things."
• 3:18
  Surface
• 3:18
  area
• 3:19
  matters;
• 3:20
  the
• 3:20
  weird
• 3:20
  stuff
• 3:20
  in
• 3:20
  the
• 3:21
  training
• 3:21
  data
• 3:21
  is
• 3:21
  what
• 3:21
  stops
• 3:22
  the
• 3:22
  model
• 3:22
  from
• 3:22
  becoming
• 3:23
  a
• 3:23
  parody
• 3:23
  of
• 3:23
  itself.
• 3:26
  The
• 3:26
  question
• 3:27
  of
• 3:27
  how
• 3:27
  to
• 3:27
  structure
• 3:28
  AI-assisted
• 3:28
  development
• 3:29
  has
• 3:29
  moved
• 3:29
  from
• 3:30
  philosophical
• 3:30
  taxonomy
• 3:31
  to
• 3:31
  shipping
• 3:31
  infrastructure.
• 3:33
  Reuven
• 3:33
  Cohen's
• 3:33
  framework
• 3:33
  —
• 3:34
  skills,
• 3:35
  agents,
• 3:35
  and
• 3:35
  harnesses
• 3:36
  —
• 3:36
  captures
• 3:37
  the
• 3:37
  maturity
• 3:37
  ladder
• 3:37
  most
• 3:38
  teams
• 3:38
  are
• 3:38
  climbing.
• 3:39
  A
• 3:39
  skill
• 3:40
  is
• 3:40
  a
• 3:40
  markdown
• 3:40
  file
• 3:40
  with
• 3:40
  instructions,
• 3:41
  lightweight
• 3:42
  and
• 3:42
  shareable.
• 3:43
  An
• 3:43
  agent
• 3:43
  has
• 3:44
  goals
• 3:44
  and
• 3:44
  invokes
• 3:44
  tools.
• 3:45
  The
• 3:45
  harness
• 3:46
  "sits
• 3:46
  above
• 3:46
  both,"
• 3:47
  providing
• 3:48
  persistence,
• 3:48
  memory,
• 3:49
  governance,
• 3:50
  routing,
• 3:51
  verification,
• 3:52
  coordination,
• 3:54
  and
• 3:54
  learning.
• 3:55
  "The
• 3:55
  model
• 3:55
  generates
• 3:55
  tokens.
• 3:56
  Agents
• 3:57
  perform
• 3:57
  work.
• 3:58
  The
• 3:58
  harness
• 3:58
  creates
• 3:58
  outcomes."
• 4:00
  The
• 4:00
  distinction
• 4:00
  matters
• 4:00
  because
• 4:01
  enterprises
• 4:01
  that
• 4:01
  skip
• 4:02
  the
• 4:02
  harness
• 4:02
  layer
• 4:02
  build
• 4:02
  agent
• 4:03
  demos,
• 4:03
  not
• 4:03
  agent
• 4:04
  systems.
• 4:05
  Claude
• 4:05
  Code's
• 4:05
  dynamic
• 4:06
  workflows
• 4:06
  —
• 4:07
  branded
• 4:07
  "Ultra
• 4:07
  Code"
• 4:08
  at
• 4:08
  the
• 4:08
  highest
• 4:08
  effort
• 4:08
  level
• 4:09
  —
• 4:09
  represent
• 4:10
  Anthropic's
• 4:10
  answer
• 4:11
  to
• 4:11
  the
• 4:11
  harness
• 4:11
  problem.
• 4:12
  Instead
• 4:13
  of
• 4:13
  a
• 4:13
  static
• 4:13
  one-size-fits-all
• 4:14
  approach,
• 4:15
  Ultra
• 4:15
  Code
• 4:15
  builds
• 4:15
  custom
• 4:16
  orchestration
• 4:16
  scripts
• 4:17
  on
• 4:17
  the
• 4:17
  fly,
• 4:17
  spawning
• 4:18
  potentially
• 4:18
  hundreds
• 4:19
  of
• 4:19
  agents
• 4:19
  with
• 4:19
  isolated
• 4:20
  context
• 4:20
  windows
• 4:20
  and
• 4:21
  focused
• 4:21
  goals.
• 4:21
  The
• 4:21
  patterns
• 4:22
  include
• 4:22
  classify-and-act,
• 4:23
  fan-out-and-synthesize,
• 4:24
  adversarial
• 4:25
  verification,
• 4:26
  and
• 4:26
  tournament-style
• 4:26
  evaluation.
• 4:27
  A
• 4:27
  built-in
• 4:28
  deep
• 4:28
  research
• 4:28
  workflow
• 4:29
  ran
• 4:29
  101
• 4:30
  agents,
• 4:30
  consumed
• 4:31
  3.7
• 4:31
  million
• 4:31
  tokens,
• 4:32
  and
• 4:32
  completed
• 4:33
  in
• 4:33
  11
• 4:33
  minutes.
• 4:34
  A
• 4:34
  bug-hunt
• 4:34
  workflow
• 4:34
  on
• 4:34
  a
• 4:34
  Next.js
• 4:35
  codebase
• 4:36
  found
• 4:36
  34
• 4:37
  confirmed
• 4:37
  bugs
• 4:37
  —
• 4:38
  with
• 4:38
  7
• 4:38
  false
• 4:38
  positives
• 4:39
  identified
• 4:39
  by
• 4:39
  adversarial
• 4:40
  verifiers
• 4:40
  —
• 4:40
  in
• 4:41
  half
• 4:41
  that
• 4:41
  time.
• 4:41
  The
• 4:42
  key
• 4:42
  innovation
• 4:42
  is
• 4:42
  not
• 4:43
  the
• 4:43
  agent
• 4:43
  count
• 4:43
  but
• 4:43
  that
• 4:43
  each
• 4:44
  agent
• 4:44
  gets
• 4:44
  a
• 4:44
  fresh
• 4:44
  context
• 4:45
  window,
• 4:45
  combating
• 4:46
  the
• 4:46
  agentic
• 4:46
  laziness,
• 4:47
  self-preferential
• 4:48
  bias,
• 4:48
  and
• 4:48
  goal
• 4:48
  drift
• 4:49
  that
• 4:49
  plague
• 4:49
  single-session
• 4:49
  approaches
• 4:50
  to
• 4:50
  complex
• 4:50
  tasks.
• 4:51
  Ultra
• 4:52
  Code
• 4:52
  lets
• 4:52
  Claude
• 4:52
  Code
• 4:52
  decide
• 4:53
  when
• 4:53
  a
• 4:53
  task
• 4:53
  needs
• 4:54
  dynamic
• 4:54
  workflows
• 4:54
  versus
• 4:55
  the
• 4:55
  static
• 4:55
  harness,
• 4:56
  removing
• 4:56
  the
• 4:56
  user
• 4:56
  from
• 4:57
  that
• 4:57
  routing
• 4:57
  decision
• 4:57
  entirely.
• 4:59
  The
• 4:59
  multi-vendor
• 4:59
  dimension
• 5:00
  is
• 5:00
  getting
• 5:00
  tooling
• 5:00
  too.
• 5:01
  cc-fleet
• 5:02
  is
• 5:02
  a
• 5:02
  Go
• 5:02
  CLI
• 5:03
  that
• 5:03
  spawns
• 5:03
  any
• 5:03
  vendor
• 5:04
  LLM
• 5:04
  —
• 5:05
  DeepSeek,
• 5:05
  GLM,
• 5:06
  Qwen,
• 5:06
  Kimi,
• 5:06
  MiniMax
• 5:07
  —
• 5:08
  as
• 5:08
  real
• 5:08
  Claude
• 5:08
  Code
• 5:08
  teammates
• 5:09
  running
• 5:09
  in
• 5:09
  tmux
• 5:09
  panes,
• 5:10
  driven
• 5:10
  via
• 5:10
  native
• 5:11
  TeamCreate
• 5:11
  and
• 5:11
  SendMessage.
• 5:12
  Your
• 5:13
  main
• 5:13
  session
• 5:13
  keeps
• 5:13
  its
• 5:13
  own
• 5:14
  auth;
• 5:14
  vendor
• 5:15
  workers
• 5:15
  bill
• 5:15
  the
• 5:15
  vendor
• 5:15
  API
• 5:16
  key
• 5:16
  via
• 5:16
  a
• 5:16
  secure
• 5:17
  apiKeyHelper
• 5:17
  that
• 5:18
  never
• 5:18
  touches
• 5:18
  env
• 5:18
  vars
• 5:19
  or
• 5:19
  shell
• 5:19
  history.
• 5:20
  The
• 5:20
  architecture
• 5:21
  is
• 5:21
  pragmatic:
• 5:22
  use
• 5:22
  cheap
• 5:22
  models
• 5:23
  for
• 5:23
  commodity
• 5:23
  tasks,
• 5:24
  reserve
• 5:24
  frontier
• 5:25
  models
• 5:25
  for
• 5:25
  complex
• 5:26
  reasoning,
• 5:26
  and
• 5:27
  coordinate
• 5:27
  them
• 5:27
  all
• 5:27
  through
• 5:28
  Claude
• 5:28
  Code's
• 5:28
  native
• 5:29
  teammate
• 5:29
  protocol.
• 5:30
  GitHub's
• 5:31
  agentic
• 5:31
  transformation
• 5:32
  is
• 5:32
  the
• 5:32
  industrial-scale
• 5:33
  version
• 5:33
  of
• 5:33
  this
• 5:34
  same
• 5:34
  shift.
• 5:35
  In
• 5:35
  a
• 5:35
  wide-ranging
• 5:36
  interview,
• 5:37
  GitHub
• 5:37
  engineering
• 5:37
  leadership
• 5:38
  described
• 5:38
  17
• 5:39
  million
• 5:39
  PR
• 5:39
  agents
• 5:40
  in
• 5:40
  March
• 5:40
  alone,
• 5:41
  services
• 5:42
  seeing
• 5:42
  3x
• 5:42
  expected
• 5:43
  growth,
• 5:44
  and
• 5:44
  a
• 5:44
  data
• 5:44
  center
• 5:44
  maxed
• 5:45
  out
• 5:45
  with
• 5:45
  load
• 5:45
  shedding
• 5:45
  to
• 5:46
  Azure.
• 5:47
  The
• 5:47
  design
• 5:47
  team
• 5:47
  is
• 5:47
  exploring
• 5:48
  "AX"
• 5:48
  —
• 5:49
  agentic
• 5:50
  experience
• 5:50
  —
• 5:51
  where
• 5:51
  bidirectional
• 5:52
  canvases
• 5:52
  let
• 5:52
  agents
• 5:53
  read
• 5:53
  and
• 5:53
  modify
• 5:54
  the
• 5:54
  UI
• 5:54
  while
• 5:55
  users
• 5:55
  interact
• 5:55
  through
• 5:56
  the
• 5:56
  same
• 5:56
  surface.
• 5:57
  The
• 5:57
  "low
• 5:58
  floor,
• 5:58
  high
• 5:58
  ceiling"
• 5:59
  philosophy
• 6:00
  aims
• 6:00
  to
• 6:00
  lower
• 6:00
  barriers
• 6:01
  for
• 6:01
  new
• 6:01
  builders
• 6:01
  while
• 6:02
  raising
• 6:02
  what
• 6:02
  professionals
• 6:03
  can
• 6:03
  accomplish.
• 6:04
  On
• 6:04
  whether
• 6:04
  humans
• 6:05
  remain
• 6:05
  necessary:
• 6:06
  "We
• 6:06
  named
• 6:06
  copilot
• 6:07
  copilot
• 6:07
  for
• 6:08
  a
• 6:08
  reason."
• 6:09
  Meanwhile,
• 6:10
  the
• 6:10
  Agentics
• 6:10
  Foundation
• 6:11
  —
• 6:11
  an
• 6:12
  open-source
• 6:12
  community
• 6:13
  that
• 6:13
  spent
• 6:13
  over
• 6:13
  a
• 6:13
  year
• 6:14
  meeting
• 6:14
  on
• 6:14
  Zoom
• 6:14
  —
• 6:15
  held
• 6:15
  its
• 6:16
  first
• 6:16
  in-person
• 6:16
  gathering
• 6:17
  in
• 6:17
  Budapest.
• 6:18
  Five
• 6:18
  days
• 6:19
  of
• 6:19
  workshops,
• 6:20
  conference
• 6:20
  presentations,
• 6:21
  and
• 6:21
  an
• 6:21
  open-mic
• 6:22
  night.
• 6:23
  The
• 6:23
  recommended
• 6:24
  focus
• 6:24
  for
• 6:24
  next-generation
• 6:25
  engineers:
• 6:26
  Critical
• 6:26
  Thinking,
• 6:27
  Collaboration,
• 6:28
  Communication,
• 6:29
  and
• 6:29
  Creativity.
• 6:30
  LionAGI
• 6:31
  takes
• 6:31
  a
• 6:31
  different
• 6:31
  architectural
• 6:32
  stance:
• 6:33
  agents
• 6:33
  as
• 6:33
  runtime-composed
• 6:34
  configurations
• 6:35
  with
• 6:35
  reactive
• 6:36
  signal
• 6:36
  processing
• 6:37
  and
• 6:37
  typed
• 6:37
  capability
• 6:38
  namespaces.
• 6:39
  Its
• 6:39
  Kive
• 6:39
  graph
• 6:40
  runtime
• 6:40
  (written
• 6:41
  in
• 6:41
  Rust)
• 6:41
  handles
• 6:42
  memory,
• 6:42
  knowledge
• 6:43
  graphs,
• 6:43
  and
• 6:44
  multi-embedding
• 6:44
  retrieval,
• 6:45
  while
• 6:45
  a
• 6:45
  custom
• 6:46
  inference
• 6:46
  engine
• 6:46
  called
• 6:47
  Lattis
• 6:47
  claims
• 6:47
  twice
• 6:48
  the
• 6:48
  speed
• 6:48
  of
• 6:48
  llama.cpp
• 6:50
  for
• 6:50
  tiny
• 6:51
  models.
• 6:52
  The
• 6:52
  governance
• 6:52
  model
• 6:52
  is
• 6:53
  strict
• 6:53
  —
• 6:54
  if
• 6:54
  a
• 6:54
  structured
• 6:54
  output
• 6:55
  type
• 6:55
  is
• 6:55
  not
• 6:55
  pre-configured
• 6:56
  and
• 6:56
  granted,
• 6:57
  the
• 6:57
  system
• 6:57
  simply
• 6:57
  ignores
• 6:58
  it.
• 6:58
  No
• 6:59
  exceptions.
• 7:00
  The
• 7:00
  framework
• 7:00
  can
• 7:01
  orchestrate
• 7:01
  Claude
• 7:01
  Code,
• 7:02
  Codex,
• 7:03
  and
• 7:03
  Gemini
• 7:03
  Code
• 7:04
  as
• 7:04
  sub-harnesses,
• 7:05
  treating
• 7:05
  external
• 7:05
  tools
• 7:06
  as
• 7:06
  composable
• 7:07
  execution
• 7:07
  backends
• 7:08
  rather
• 7:08
  than
• 7:08
  competitors.
• 7:12
  Anthropic
• 7:12
  published
• 7:12
  a
• 7:13
  practical
• 7:13
  security
• 7:14
  framework
• 7:14
  for
• 7:14
  deploying
• 7:14
  autonomous
• 7:15
  AI
• 7:15
  agents
• 7:16
  in
• 7:16
  the
• 7:16
  enterprise,
• 7:17
  built
• 7:17
  on
• 7:17
  zero-trust
• 7:18
  principles:
• 7:19
  trust
• 7:19
  nothing,
• 7:19
  verify
• 7:20
  everything,
• 7:21
  assume
• 7:21
  breach
• 7:21
  has
• 7:21
  already
• 7:22
  occurred.
• 7:22
  The
• 7:23
  guide
• 7:23
  addresses
• 7:23
  threat
• 7:23
  vectors
• 7:24
  specific
• 7:24
  to
• 7:24
  agentic
• 7:25
  systems
• 7:25
  —
• 7:26
  tool
• 7:26
  access,
• 7:27
  autonomous
• 7:27
  decision-making,
• 7:28
  context
• 7:29
  persistence,
• 7:29
  multi-agent
• 7:30
  coordination,
• 7:31
  prompt
• 7:31
  injection,
• 7:32
  tool
• 7:32
  poisoning,
• 7:33
  identity
• 7:33
  abuse,
• 7:34
  memory
• 7:34
  poisoning,
• 7:35
  and
• 7:35
  supply
• 7:35
  chain
• 7:35
  attacks.
• 7:36
  It
• 7:37
  maps
• 7:37
  a
• 7:37
  three-tier
• 7:37
  architecture
• 7:38
  (Foundation,
• 7:39
  Advanced,
• 7:40
  Optimized)
• 7:41
  to
• 7:42
  organizational
• 7:42
  maturity,
• 7:43
  with
• 7:43
  an
• 7:43
  eight-phase
• 7:44
  implementation
• 7:44
  workflow
• 7:45
  covering
• 7:45
  identity,
• 7:46
  access
• 7:46
  scoping,
• 7:47
  sandboxing,
• 7:48
  input/output
• 7:49
  controls,
• 7:50
  and
• 7:50
  memory
• 7:50
  safeguards.
• 7:52
  The
• 7:52
  core
• 7:52
  thesis
• 7:52
  is
• 7:52
  sharp:
• 7:54
  "Frontier
• 7:54
  AI
• 7:55
  models
• 7:55
  are
• 7:55
  compressing
• 7:56
  the
• 7:56
  timeline
• 7:56
  between
• 7:57
  vulnerability
• 7:57
  and
• 7:57
  exploit
• 7:58
  from
• 7:58
  months
• 7:58
  to
• 7:58
  hours,"
• 7:59
  and
• 7:59
  organizations
• 8:00
  deploying
• 8:00
  agents
• 8:01
  face
• 8:01
  this
• 8:01
  acceleration
• 8:02
  twice
• 8:02
  —
• 8:03
  once
• 8:03
  on
• 8:03
  their
• 8:03
  infrastructure
• 8:04
  and
• 8:04
  once
• 8:04
  through
• 8:04
  the
• 8:05
  agents
• 8:05
  themselves.
• 8:06
  The
• 8:06
  organizations
• 8:07
  best
• 8:07
  positioned
• 8:08
  are
• 8:08
  those
• 8:08
  "whose
• 8:09
  fundamentals
• 8:09
  are
• 8:09
  strong
• 8:10
  enough
• 8:10
  that
• 8:10
  AI-assisted
• 8:11
  scanning
• 8:11
  finds
• 8:11
  fewer
• 8:12
  bugs
• 8:12
  in
• 8:12
  the
• 8:12
  first
• 8:12
  place,
• 8:13
  and
• 8:13
  whose
• 8:13
  agent
• 8:14
  deployments
• 8:14
  are
• 8:14
  architected
• 8:15
  for
• 8:15
  breach
• 8:15
  from
• 8:15
  day
• 8:16
  one."
• 8:16
  On
• 8:17
  the
• 8:17
  supply-chain
• 8:17
  side,
• 8:18
  Perplexity
• 8:19
  open-sourced
• 8:19
  Bumblebee,
• 8:20
  a
• 8:20
  read-only
• 8:21
  inventory
• 8:21
  collector
• 8:22
  for
• 8:22
  developer
• 8:22
  endpoints
• 8:22
  that
• 8:23
  answers
• 8:23
  a
• 8:23
  narrow
• 8:23
  but
• 8:23
  critical
• 8:24
  incident-response
• 8:25
  question:
• 8:25
  when
• 8:25
  an
• 8:25
  advisory
• 8:26
  names
• 8:26
  a
• 8:26
  package,
• 8:27
  extension,
• 8:27
  or
• 8:27
  version,
• 8:28
  which
• 8:28
  developer
• 8:29
  machines
• 8:29
  have
• 8:29
  it
• 8:29
  installed
• 8:29
  right
• 8:30
  now?
• 8:30
  It
• 8:30
  is
• 8:30
  a
• 8:30
  single
• 8:31
  static
• 8:31
  Go
• 8:31
  binary
• 8:32
  with
• 8:32
  zero
• 8:32
  non-stdlib
• 8:33
  dependencies,
• 8:33
  three
• 8:33
  scan
• 8:34
  profiles
• 8:34
  (baseline,
• 8:35
  project,
• 8:35
  deep),
• 8:36
  and
• 8:36
  coverage
• 8:36
  spanning
• 8:37
  npm,
• 8:37
  PyPI,
• 8:38
  Go
• 8:38
  modules,
• 8:39
  RubyGems,
• 8:40
  Composer,
• 8:40
  Homebrew,
• 8:41
  editor
• 8:41
  extensions,
• 8:42
  browser
• 8:42
  extensions,
• 8:43
  MCP
• 8:43
  host
• 8:44
  configs,
• 8:44
  and
• 8:44
  agent
• 8:45
  skill
• 8:45
  lockfiles.
• 8:46
  Feed
• 8:46
  it
• 8:46
  an
• 8:46
  exposure
• 8:47
  catalog
• 8:47
  and
• 8:47
  it
• 8:47
  flags
• 8:48
  exact
• 8:48
  (ecosystem,
• 8:49
  name,
• 8:50
  version)
• 8:51
  matches.
• 8:52
  The
• 8:52
  threat
• 8:52
  intel/
• 8:52
  directory
• 8:53
  ships
• 8:53
  maintained
• 8:53
  catalogs
• 8:54
  built
• 8:54
  from
• 8:54
  public
• 8:54
  threat-intelligence
• 8:55
  reporting
• 8:55
  on
• 8:56
  recent
• 8:56
  supply-chain
• 8:56
  campaigns.
• 8:57
  The
• 8:57
  design
• 8:58
  philosophy
• 8:58
  is
• 8:58
  deliberately
• 8:59
  narrow:
• 8:59
  read-only,
• 9:00
  no
• 9:00
  package-manager
• 9:01
  execution,
• 9:01
  no
• 9:01
  source-file
• 9:02
  reads.
• 9:03
  SBOMs
• 9:03
  tell
• 9:03
  you
• 9:03
  what
• 9:03
  shipped;
• 9:04
  EDR
• 9:04
  tells
• 9:04
  you
• 9:05
  what
• 9:05
  ran.
• 9:05
  Bumblebee
• 9:06
  tells
• 9:06
  you
• 9:06
  the
• 9:06
  messy
• 9:06
  local
• 9:07
  state
• 9:07
  across
• 9:07
  lockfiles,
• 9:08
  package-manager
• 9:09
  metadata,
• 9:10
  and
• 9:10
  extension
• 9:10
  manifests
• 9:11
  that
• 9:11
  neither
• 9:11
  of
• 9:11
  those
• 9:12
  captures.
• 9:15
  Google
• 9:15
  Cloud
• 9:15
  AI
• 9:16
  Research's
• 9:16
  ScientistOne
• 9:17
  paper
• 9:17
  introduces
• 9:18
  Chain-of-Evidence
• 9:18
  (CoE),
• 9:19
  a
• 9:19
  verifiability
• 9:20
  framework
• 9:21
  for
• 9:21
  autonomous
• 9:21
  research
• 9:22
  that
• 9:22
  treats
• 9:22
  every
• 9:22
  claim
• 9:23
  the
• 9:23
  way
• 9:23
  ACID
• 9:23
  treats
• 9:24
  database
• 9:24
  transactions:
• 9:25
  each
• 9:25
  must
• 9:26
  trace,
• 9:26
  through
• 9:26
  a
• 9:26
  recorded
• 9:27
  evidence
• 9:27
  chain,
• 9:27
  to
• 9:27
  a
• 9:27
  grounding
• 9:28
  source.
• 9:29
  The
• 9:29
  system's
• 9:29
  pipeline
• 9:30
  —
• 9:31
  Problem
• 9:31
  Investigator
• 9:31
  (reads
• 9:32
  up
• 9:32
  to
• 9:32
  100
• 9:33
  full-text
• 9:33
  PDFs
• 9:34
  per
• 9:34
  topic),
• 9:35
  Discovery
• 9:35
  Engine
• 9:36
  (parallel
• 9:37
  explore-exploit
• 9:37
  search
• 9:37
  tree),
• 9:38
  and
• 9:38
  Paper
• 9:39
  Writer
• 9:39
  with
• 9:39
  Claim
• 9:39
  Verifier
• 9:40
  —
• 9:40
  maintains
• 9:41
  evidence
• 9:41
  chains
• 9:42
  by
• 9:42
  construction
• 9:43
  rather
• 9:43
  than
• 9:43
  retrofitting
• 9:44
  them
• 9:44
  at
• 9:44
  paper-writing
• 9:45
  time.
• 9:45
  In
• 9:46
  an
• 9:46
  audit
• 9:46
  of
• 9:46
  75
• 9:47
  papers
• 9:47
  from
• 9:47
  five
• 9:47
  autonomous
• 9:48
  research
• 9:48
  systems
• 9:49
  across
• 9:49
  five
• 9:49
  frontier
• 9:49
  tasks,
• 9:50
  every
• 9:51
  baseline
• 9:51
  exhibited
• 9:52
  systematic
• 9:52
  integrity
• 9:52
  failures:
• 9:53
  hallucinated
• 9:54
  reference
• 9:54
  rates
• 9:55
  reaching
• 9:55
  21%,
• 9:56
  score
• 9:56
  verification
• 9:57
  passing
• 9:57
  in
• 9:57
  as
• 9:58
  few
• 9:58
  as
• 9:58
  42%
• 9:59
  of
• 9:59
  papers,
• 10:00
  and
• 10:00
  method-code
• 10:00
  alignment
• 10:01
  ranging
• 10:01
  from
• 10:01
  20%
• 10:02
  to
• 10:02
  80%.
• 10:03
  ScientistOne
• 10:04
  achieved
• 10:04
  zero
• 10:05
  hallucinated
• 10:05
  references
• 10:06
  (0
• 10:06
  of
• 10:07
  337
• 10:08
  bibliography
• 10:08
  entries),
• 10:09
  perfect
• 10:09
  score
• 10:09
  verification
• 10:10
  (12/12),
• 10:11
  and
• 10:12
  the
• 10:12
  highest
• 10:12
  method-code
• 10:12
  alignment
• 10:13
  (14/15),
• 10:14
  while
• 10:15
  matching
• 10:15
  or
• 10:15
  exceeding
• 10:16
  human
• 10:16
  expert
• 10:16
  performance
• 10:17
  on
• 10:17
  all
• 10:17
  five
• 10:17
  tasks.
• 10:18
  On
• 10:19
  MLE-Bench,
• 10:20
  it
• 10:20
  earned
• 10:21
  Gold
• 10:21
  on
• 10:21
  3D
• 10:22
  Object
• 10:22
  Detection
• 10:23
  where
• 10:23
  a
• 10:23
  competing
• 10:23
  system
• 10:24
  scored
• 10:24
  0.0000,
• 10:26
  and
• 10:26
  achieved
• 10:27
  state-of-the-art
• 10:28
  on
• 10:28
  the
• 10:28
  live
• 10:28
  Parameter
• 10:28
  Golf
• 10:29
  competition
• 10:29
  with
• 10:30
  novel
• 10:30
  techniques
• 10:31
  including
• 10:31
  Hessian-diagonal-weighted
• 10:32
  SVD
• 10:33
  initialization.
• 10:34
  The
• 10:34
  project
• 10:35
  website
• 10:35
  showcases
• 10:36
  21
• 10:36
  autonomously
• 10:37
  generated
• 10:38
  papers
• 10:38
  with
• 10:38
  all
• 10:39
  references
• 10:39
  verified
• 10:40
  against
• 10:40
  real
• 10:40
  publications
• 10:41
  and
• 10:41
  every
• 10:42
  claimed
• 10:42
  result
• 10:42
  reproducing
• 10:43
  under
• 10:43
  re-evaluation.
• 10:45
  SIA
• 10:45
  (Self
• 10:46
  Improving
• 10:46
  AI)
• 10:47
  from
• 10:47
  Hexo
• 10:47
  AI
• 10:48
  approaches
• 10:48
  the
• 10:48
  problem
• 10:49
  differently:
• 10:50
  a
• 10:50
  meta-agent
• 10:51
  generates
• 10:51
  an
• 10:51
  initial
• 10:52
  task-specific
• 10:52
  agent,
• 10:53
  that
• 10:53
  agent
• 10:54
  attempts
• 10:54
  the
• 10:54
  task,
• 10:55
  a
• 10:55
  feedback
• 10:56
  agent
• 10:56
  reviews
• 10:57
  performance
• 10:57
  and
• 10:57
  updates
• 10:58
  the
• 10:58
  agent,
• 10:58
  and
• 10:59
  the
• 10:59
  cycle
• 10:59
  repeats
• 10:59
  across
• 11:00
  generations.
• 11:01
  The
• 11:02
  paper
• 11:02
  reports
• 11:02
  a
• 11:02
  56.6%
• 11:04
  gain
• 11:04
  on
• 11:04
  LawBench
• 11:05
  (reaching
• 11:06
  70.1%
• 11:06
  Top-1
• 11:07
  accuracy
• 11:08
  versus
• 11:08
  45%
• 11:09
  prior
• 11:09
  SOTA),
• 11:10
  91.9%
• 11:11
  runtime
• 11:12
  reduction
• 11:12
  on
• 11:13
  GPU
• 11:13
  kernels,
• 11:14
  and
• 11:14
  502%
• 11:15
  improvement
• 11:16
  on
• 11:16
  single-cell
• 11:17
  RNA
• 11:17
  denoising.
• 11:18
  It
• 11:18
  supports
• 11:19
  multiple
• 11:19
  agent
• 11:20
  implementations
• 11:20
  —
• 11:21
  Claude
• 11:22
  Agent
• 11:22
  SDK
• 11:22
  for
• 11:23
  Anthropic
• 11:23
  models,
• 11:24
  OpenHands
• 11:24
  for
• 11:25
  multi-provider
• 11:25
  (Gemini,
• 11:26
  OpenAI,
• 11:27
  Anthropic)
• 11:27
  —
• 11:29
  and
• 11:29
  ships
• 11:29
  with
• 11:29
  a
• 11:30
  built-in
• 11:30
  web
• 11:30
  dashboard
• 11:31
  for
• 11:31
  visualizing
• 11:31
  per-generation
• 11:32
  target-agent
• 11:33
  code,
• 11:33
  improvement
• 11:34
  plans,
• 11:34
  and
• 11:35
  accuracy
• 11:35
  curves.
• 11:36
  At
• 11:36
  the
• 11:36
  opposite
• 11:37
  end
• 11:37
  of
• 11:37
  the
• 11:37
  rigor
• 11:38
  spectrum,
• 11:38
  a
• 11:39
  YouTube
• 11:39
  presentation
• 11:40
  claims
• 11:40
  to
• 11:40
  have
• 11:40
  derived
• 11:41
  "the
• 11:41
  formula
• 11:41
  for
• 11:41
  intelligence
• 11:42
  itself"
• 11:43
  through
• 11:43
  multi-embedder
• 11:44
  systems
• 11:44
  where
• 11:45
  meaning
• 11:45
  emerges
• 11:45
  from
• 11:46
  vector
• 11:46
  positions
• 11:46
  across
• 11:47
  parallel
• 11:47
  embedding
• 11:48
  spaces
• 11:48
  —
• 11:49
  what
• 11:49
  the
• 11:49
  author
• 11:49
  calls
• 11:50
  "teleological
• 11:51
  constellations."
• 11:52
  The
• 11:52
  core
• 11:52
  technical
• 11:53
  idea
• 11:54
  (multi-modal
• 11:54
  embedding
• 11:55
  constellations
• 11:56
  with
• 11:56
  cosine-similarity
• 11:57
  guardrails)
• 11:58
  is
• 11:58
  a
• 11:58
  restatement
• 11:59
  of
• 11:59
  joint
• 11:59
  embedding
• 11:59
  architectures,
• 12:01
  wrapped
• 12:01
  in
• 12:01
  claims
• 12:01
  of
• 12:02
  necromancy,
• 12:03
  divine
• 12:03
  powers,
• 12:04
  and
• 12:04
  future
• 12:04
  prediction
• 12:05
  via
• 12:05
  JEPA
• 12:05
  world
• 12:06
  models.
• 12:07
  The
• 12:07
  presenter
• 12:07
  asserts
• 12:08
  AGI
• 12:08
  models
• 12:09
  can
• 12:09
  be
• 12:09
  compressed
• 12:10
  90-99%
• 12:13
  and
• 12:13
  run
• 12:13
  on
• 12:13
  home
• 12:13
  computers,
• 12:14
  while
• 12:15
  simultaneously
• 12:16
  claiming
• 12:16
  the
• 12:16
  knowledge
• 12:16
  enables
• 12:17
  bringing
• 12:17
  the
• 12:17
  dead
• 12:18
  back
• 12:18
  to
• 12:18
  life.
• 12:19
  The
• 12:19
  contrast
• 12:20
  with
• 12:20
  ScientistOne's
• 12:20
  methodology
• 12:21
  could
• 12:21
  not
• 12:22
  be
• 12:22
  sharper:
• 12:23
  one
• 12:23
  system
• 12:23
  audits
• 12:24
  every
• 12:24
  claim
• 12:24
  against
• 12:25
  its
• 12:25
  evidence
• 12:25
  source;
• 12:26
  the
• 12:26
  other
• 12:27
  presents
• 12:27
  extraordinary
• 12:28
  assertions
• 12:29
  with
• 12:29
  no
• 12:29
  reproducible
• 12:30
  evidence,
• 12:31
  no
• 12:31
  peer
• 12:31
  review,
• 12:32
  and
• 12:32
  no
• 12:32
  benchmarks.
• 12:36
  CognoDB
• 12:37
  is
• 12:37
  an
• 12:37
  open-source,
• 12:38
  Cypher-compatible
• 12:40
  graph
• 12:40
  database
• 12:41
  written
• 12:41
  in
• 12:41
  Go,
• 12:42
  purpose-built
• 12:42
  for
• 12:42
  AI
• 12:43
  agent
• 12:43
  workloads.
• 12:44
  It
• 12:44
  speaks
• 12:44
  the
• 12:44
  Bolt
• 12:45
  protocol
• 12:46
  (v5.0-5.4)
• 12:48
  and
• 12:49
  works
• 12:49
  with
• 12:49
  all
• 12:49
  official
• 12:50
  Neo4j
• 12:50
  drivers
• 12:50
  —
• 12:51
  Python,
• 12:51
  Go,
• 12:52
  JavaScript,
• 12:52
  Java,
• 12:53
  .NET
• 12:53
  —
• 12:53
  with
• 12:53
  zero
• 12:53
  code
• 12:54
  changes.
• 12:55
  Cold
• 12:55
  start
• 12:55
  is
• 12:55
  7ms
• 12:56
  with
• 12:56
  15MB
• 12:57
  idle
• 12:57
  RAM,
• 12:57
  compared
• 12:58
  to
• 12:58
  Neo4j's
• 12:58
  17
• 12:59
  seconds
• 12:59
  and
• 12:59
  3.5GB.
• 13:01
  The
• 13:01
  built-in
• 13:01
  MCP
• 13:02
  server
• 13:02
  exposes
• 13:03
  schema
• 13:03
  and
• 13:03
  query
• 13:03
  tools
• 13:04
  so
• 13:04
  any
• 13:04
  MCP-compatible
• 13:06
  assistant
• 13:06
  can
• 13:06
  query
• 13:07
  and
• 13:07
  write
• 13:07
  to
• 13:07
  the
• 13:07
  graph
• 13:07
  in
• 13:08
  natural
• 13:08
  language.
• 13:09
  The
• 13:09
  token
• 13:10
  economics
• 13:10
  are
• 13:10
  compelling:
• 13:11
  agents
• 13:11
  that
• 13:11
  query
• 13:12
  CognoDB
• 13:12
  for
• 13:12
  relevant
• 13:13
  context
• 13:13
  instead
• 13:13
  of
• 13:14
  dumping
• 13:14
  the
• 13:14
  full
• 13:14
  corpus
• 13:14
  into
• 13:15
  the
• 13:15
  prompt
• 13:15
  see
• 13:15
  token
• 13:15
  reductions
• 13:16
  from
• 13:16
  62%
• 13:17
  at
• 13:18
  185
• 13:18
  nodes
• 13:19
  to
• 13:19
  99%
• 13:20
  at
• 13:20
  3,700
• 13:21
  nodes,
• 13:21
  with
• 13:22
  costs
• 13:22
  staying
• 13:22
  flat
• 13:22
  as
• 13:23
  the
• 13:23
  graph
• 13:23
  grows.
• 13:24
  In
• 13:24
  a
• 13:24
  real
• 13:24
  LLM
• 13:25
  agent
• 13:25
  benchmark
• 13:26
  using
• 13:26
  AWS
• 13:27
  Bedrock,
• 13:28
  a
• 13:28
  100-agent
• 13:28
  workflow
• 13:29
  saved
• 13:29
  $1.60
• 13:30
  per
• 13:30
  run
• 13:31
  versus
• 13:31
  full-context
• 13:32
  prompting.
• 13:33
  The
• 13:33
  trade-off
• 13:33
  is
• 13:33
  traversal
• 13:34
  latency
• 13:34
  —
• 13:35
  Neo4j's
• 13:35
  years
• 13:36
  of
• 13:36
  query
• 13:36
  optimization
• 13:37
  give
• 13:37
  it
• 13:37
  3ms
• 13:38
  p50
• 13:38
  on
• 13:38
  3-hop
• 13:39
  traversals
• 13:40
  versus
• 13:40
  CognoDB's
• 13:41
  27.6ms
• 13:41
  in-memory
• 13:42
  —
• 13:43
  but
• 13:43
  for
• 13:44
  agent
• 13:44
  workloads
• 13:44
  where
• 13:45
  cold-start
• 13:45
  time
• 13:45
  and
• 13:46
  memory
• 13:46
  footprint
• 13:46
  matter
• 13:46
  more
• 13:47
  than
• 13:47
  raw
• 13:47
  traversal
• 13:47
  speed,
• 13:48
  the
• 13:48
  profile
• 13:49
  fits.
• 13:50
  ModelRegression
• 13:51
  takes
• 13:51
  aim
• 13:52
  at
• 13:52
  a
• 13:52
  different
• 13:52
  infrastructure
• 13:53
  gap:
• 13:53
  AI
• 13:54
  model
• 13:54
  providers
• 13:55
  update
• 13:55
  their
• 13:55
  models
• 13:55
  constantly,
• 13:56
  sometimes
• 13:57
  degrading
• 13:57
  performance
• 13:58
  without
• 13:58
  announcement.
• 13:59
  The
• 14:00
  project
• 14:00
  runs
• 14:00
  30
• 14:01
  tests
• 14:01
  daily
• 14:01
  against
• 14:02
  Claude
• 14:02
  Opus
• 14:02
  4.8,
• 14:03
  Claude
• 14:03
  Sonnet
• 14:04
  4.6,
• 14:05
  GPT-5.5,
• 14:07
  and
• 14:07
  Grok
• 14:07
  across
• 14:07
  10
• 14:08
  categories
• 14:08
  —
• 14:09
  long
• 14:09
  reasoning,
• 14:10
  coding
• 14:10
  tasks,
• 14:11
  bug
• 14:11
  fixes,
• 14:12
  feature
• 14:12
  implementation,
• 14:13
  code
• 14:14
  thoroughness,
• 14:14
  bug
• 14:15
  introduction
• 14:15
  rate,
• 14:16
  security
• 14:16
  awareness,
• 14:17
  instruction
• 14:17
  following,
• 14:18
  code
• 14:18
  quality,
• 14:19
  and
• 14:19
  performance
• 14:19
  efficiency.
• 14:20
  Models
• 14:21
  are
• 14:21
  tested
• 14:21
  through
• 14:22
  their
• 14:22
  official
• 14:22
  CLI
• 14:22
  tools
• 14:23
  (claude,
• 14:23
  codex,
• 14:24
  agent),
• 14:25
  not
• 14:25
  API
• 14:25
  calls,
• 14:26
  exercising
• 14:27
  the
• 14:27
  full
• 14:27
  developer-facing
• 14:28
  stack.
• 14:29
  Regressions
• 14:30
  are
• 14:30
  classified
• 14:30
  by
• 14:30
  severity:
• 14:31
  minor
• 14:32
  (>5%
• 14:33
  drop),
• 14:34
  moderate
• 14:34
  (>10%),
• 14:35
  major
• 14:36
  (>20%).
• 14:37
  The
• 14:38
  benchmark
• 14:38
  engine
• 14:38
  runs
• 14:39
  daily
• 14:39
  on
• 14:39
  a
• 14:39
  DGX
• 14:40
  at
• 14:40
  3am
• 14:40
  ET,
• 14:41
  exports
• 14:41
  to
• 14:41
  static
• 14:42
  JSON,
• 14:43
  and
• 14:43
  deploys
• 14:43
  via
• 14:43
  blue-green
• 14:44
  to
• 14:44
  a
• 14:44
  Next.js
• 14:45
  site.
• 14:46
  The
• 14:46
  live
• 14:46
  dashboard
• 14:47
  is
• 14:47
  at
• 14:47
  modelregression.com.
• 14:49
  On
• 14:49
  the
• 14:49
  model
• 14:50
  periphery,
• 14:50
  MisoTTS
• 14:51
  appeared
• 14:52
  as
• 14:52
  a
• 14:52
  trending
• 14:52
  model
• 14:52
  on
• 14:53
  HuggingFace,
• 14:54
  though
• 14:54
  details
• 14:54
  remain
• 14:55
  sparse
• 14:55
  at
• 14:55
  the
• 14:55
  time
• 14:55
  of
• 14:55
  writing.
• 14:58
  Let's
• 14:58
  Encrypt
• 14:58
  announced
• 14:59
  plans
• 14:59
  to
• 14:59
  support
• 15:00
  Merkle
• 15:00
  Tree
• 15:00
  Certificates
• 15:01
  (MTCs)
• 15:02
  as
• 15:02
  its
• 15:02
  path
• 15:03
  to
• 15:03
  post-quantum
• 15:03
  authentication
• 15:04
  in
• 15:04
  the
• 15:04
  Web
• 15:04
  PKI,
• 15:05
  targeting
• 15:06
  a
• 15:06
  staging
• 15:06
  environment
• 15:07
  in
• 15:07
  late
• 15:07
  2026
• 15:08
  and
• 15:08
  production
• 15:08
  in
• 15:08
  2027.
• 15:10
  The
• 15:10
  problem
• 15:10
  is
• 15:10
  size:
• 15:11
  ML-DSA-44,
• 15:13
  one
• 15:13
  of
• 15:13
  the
• 15:13
  smaller
• 15:13
  NIST-standardized
• 15:15
  post-quantum
• 15:16
  signature
• 15:16
  schemes,
• 15:17
  produces
• 15:17
  2,420-byte
• 15:18
  signatures
• 15:18
  —
• 15:19
  roughly
• 15:20
  10x
• 15:20
  larger
• 15:20
  than
• 15:21
  RSA-2048's
• 15:22
  256
• 15:23
  bytes
• 15:23
  and
• 15:24
  38x
• 15:24
  larger
• 15:25
  than
• 15:25
  ECDSA-P256's
• 15:26
  64
• 15:28
  bytes.
• 15:28
  A
• 15:28
  typical
• 15:29
  TLS
• 15:29
  handshake
• 15:30
  carries
• 15:30
  five
• 15:30
  signatures
• 15:31
  and
• 15:31
  two
• 15:31
  public
• 15:31
  keys;
• 15:32
  replacing
• 15:33
  them
• 15:33
  with
• 15:33
  ML-DSA
• 15:35
  equivalents
• 15:35
  would
• 15:35
  push
• 15:36
  past
• 15:36
  10
• 15:36
  kilobytes,
• 15:37
  where
• 15:37
  research
• 15:37
  shows
• 15:38
  a
• 15:38
  meaningful
• 15:38
  share
• 15:38
  of
• 15:39
  connections
• 15:39
  fail
• 15:39
  on
• 15:39
  real-world
• 15:40
  networks
• 15:40
  and
• 15:40
  the
• 15:41
  rest
• 15:41
  get
• 15:41
  slower.
• 15:42
  MTCs
• 15:42
  sidestep
• 15:43
  this
• 15:43
  by
• 15:43
  issuing
• 15:43
  certificates
• 15:44
  in
• 15:44
  batches
• 15:45
  with
• 15:45
  a
• 15:45
  single
• 15:45
  signature
• 15:46
  covering
• 15:46
  the
• 15:46
  entire
• 15:46
  batch.
• 15:47
  Browsers
• 15:47
  stay
• 15:48
  current
• 15:48
  on
• 15:48
  batch
• 15:48
  signatures
• 15:49
  ("landmarks")
• 15:50
  separately
• 15:50
  from
• 15:50
  the
• 15:50
  TLS
• 15:51
  handshake,
• 15:52
  so
• 15:52
  the
• 15:52
  common-case
• 15:52
  authentication
• 15:53
  path
• 15:53
  is
• 15:53
  one
• 15:54
  signature,
• 15:54
  one
• 15:55
  public
• 15:55
  key,
• 15:55
  and
• 15:55
  one
• 15:56
  inclusion
• 15:56
  proof
• 15:56
  —
• 15:57
  smaller
• 15:57
  than
• 15:57
  today's
• 15:58
  handshake
• 15:58
  even
• 15:58
  with
• 15:58
  post-quantum
• 15:59
  algorithms.
• 16:00
  Transparency
• 16:01
  becomes
• 16:01
  a
• 16:01
  structural
• 16:01
  property
• 16:02
  rather
• 16:02
  than
• 16:02
  an
• 16:02
  afterthought:
• 16:04
  a
• 16:04
  certificate
• 16:04
  cannot
• 16:04
  exist
• 16:05
  outside
• 16:05
  the
• 16:05
  Merkle
• 16:05
  tree,
• 16:06
  eliminating
• 16:07
  the
• 16:07
  bolted-on
• 16:07
  Certificate
• 16:08
  Transparency
• 16:08
  ecosystem
• 16:09
  where
• 16:09
  certificates
• 16:10
  are
• 16:10
  issued
• 16:10
  by
• 16:11
  CAs,
• 16:11
  logged
• 16:12
  separately,
• 16:12
  and
• 16:13
  carry
• 16:13
  extra
• 16:13
  signatures
• 16:14
  as
• 16:14
  attestation.
• 16:15
  Cloudflare
• 16:16
  and
• 16:16
  Chrome
• 16:16
  are
• 16:16
  already
• 16:16
  running
• 16:17
  an
• 16:17
  experiment
• 16:17
  with
• 16:17
  MTCs
• 16:18
  against
• 16:18
  real
• 16:19
  internet
• 16:19
  traffic,
• 16:20
  and
• 16:20
  Chrome
• 16:20
  has
• 16:20
  indicated
• 16:21
  MTCs
• 16:22
  are
• 16:22
  its
• 16:22
  preferred
• 16:22
  path
• 16:23
  for
• 16:23
  post-quantum
• 16:23
  certificates.
• 16:24
  The
• 16:25
  timeline
• 16:25
  pressure
• 16:25
  is
• 16:25
  real:
• 16:26
  Google
• 16:27
  announced
• 16:27
  migration
• 16:27
  of
• 16:28
  its
• 16:28
  services
• 16:28
  by
• 16:28
  2029,
• 16:30
  Go
• 16:30
  1.24
• 16:31
  added
• 16:31
  ML-DSA
• 16:33
  to
• 16:33
  the
• 16:33
  standard
• 16:33
  library,
• 16:34
  and
• 16:34
  NSA
• 16:35
  directives
• 16:35
  would
• 16:35
  deprecate
• 16:36
  RSA-2048
• 16:38
  and
• 16:39
  P-256
• 16:40
  after
• 16:40
  2030.
• 16:42
  As
• 16:42
  Let's
• 16:42
  Encrypt
• 16:42
  notes,
• 16:43
  nothing
• 16:43
  changes
• 16:43
  for
• 16:43
  existing
• 16:44
  subscribers
• 16:44
  today
• 16:45
  —
• 16:45
  certificates
• 16:46
  continue
• 16:46
  to
• 16:46
  be
• 16:46
  issued
• 16:47
  and
• 16:47
  renewed
• 16:47
  exactly
• 16:48
  as
• 16:48
  before.
• 16:49
  But
• 16:49
  if
• 16:49
  you
• 16:49
  maintain
• 16:50
  an
• 16:50
  ACME
• 16:51
  client
• 16:51
  or
• 16:51
  run
• 16:51
  an
• 16:52
  ACME-driven
• 16:52
  certificate
• 16:53
  pipeline,
• 16:54
  now
• 16:54
  is
• 16:54
  the
• 16:54
  time
• 16:54
  to
• 16:54
  start
• 16:55
  tracking
• 16:55
  the
• 16:55
  IETF
• 16:56
  PLANTS
• 16:56
  working
• 16:56
  group,
• 16:57
  because
• 16:57
  some
• 16:58
  of
• 16:58
  the
• 16:58
  changes
• 16:58
  coming
• 16:59
  will
• 16:59
  require
• 16:59
  client-side
• 17:00
  support.