Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 10, 2026

Transcript

AGI Dreams Podcast – April 10, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  A
• 0:00
  teammate
• 0:00
  handed
• 0:01
  Claude
• 0:01
  a
• 0:01
  Phrack
• 0:01
  article
• 0:02
  describing
• 0:02
  how
• 0:02
  to
• 0:02
  chain
• 0:03
  two
• 0:03
  rsync
• 0:03
  CVEs
• 0:04
  into
• 0:05
  unauthenticated
• 0:05
  remote
• 0:06
  code
• 0:06
  execution.
• 0:07
  Claude
• 0:07
  built
• 0:08
  a
• 0:08
  working
• 0:08
  exploit
• 0:08
  on
• 0:08
  x86-64.
• 0:10
  Then
• 0:11
  a
• 0:11
  second
• 0:11
  researcher,
• 0:12
  needing
• 0:12
  the
• 0:12
  exploit
• 0:13
  on
• 0:13
  ARM64,
• 0:15
  gave
• 0:15
  Claude
• 0:15
  one
• 0:15
  prompt:
• 0:16
  "Read
• 0:16
  the
• 0:16
  WriteUp
• 0:17
  and
• 0:17
  reproduce
• 0:17
  this
• 0:17
  exploit
• 0:18
  with
• 0:18
  exploit.py."
• 0:20
  Ninety
• 0:20
  minutes
• 0:21
  later,
• 0:21
  Claude
• 0:21
  dropped
• 0:21
  a
• 0:21
  working
• 0:22
  ARM64
• 0:22
  RCE
• 0:23
  --
• 0:24
  having
• 0:24
  rebuilt
• 0:25
  the
• 0:25
  entire
• 0:25
  protocol
• 0:25
  library
• 0:26
  from
• 0:26
  scratch,
• 0:26
  diagnosed
• 0:27
  five
• 0:27
  architecture-specific
• 0:28
  bugs
• 0:28
  using
• 0:28
  LD
• 0:29
  PRELOAD
• 0:30
  hooks
• 0:30
  and
• 0:30
  ptrace
• 0:31
  (no
• 0:31
  GDB,
• 0:31
  no
• 0:31
  root),
• 0:32
  and
• 0:32
  fixed
• 0:32
  each
• 0:32
  one
• 0:32
  without
• 0:33
  human
• 0:33
  steering.
• 0:34
  When
• 0:34
  told
• 0:34
  the
• 0:34
  exploit
• 0:35
  took
• 0:35
  five
• 0:35
  minutes,
• 0:36
  Claude
• 0:36
  optimized
• 0:36
  it
• 0:36
  to
• 0:37
  fourteen
• 0:37
  seconds
• 0:38
  by
• 0:38
  noticing
• 0:38
  that
• 0:38
  18
• 0:39
  of
• 0:39
  24
• 0:39
  leaked
• 0:40
  bytes
• 0:40
  are
• 0:40
  structural
• 0:40
  constants
• 0:41
  on
• 0:41
  ARM64
• 0:42
  and
• 0:43
  parallelizing
• 0:43
  the
• 0:43
  remaining
• 0:44
  guesses
• 0:44
  with
• 0:44
  asyncio.
• 0:45
  The
• 0:45
  team
• 0:46
  then
• 0:46
  pointed
• 0:46
  Claude
• 0:46
  at
• 0:47
  rsync
• 0:47
  3.4.1
• 0:48
  --
• 0:49
  the
• 0:49
  patched
• 0:49
  version
• 0:50
  --
• 0:50
  and
• 0:50
  it
• 0:50
  came
• 0:51
  back
• 0:51
  with
• 0:51
  new
• 0:51
  bugs
• 0:51
  they
• 0:52
  are
• 0:52
  still
• 0:52
  verifying.
• 0:53
  This
• 0:53
  is
• 0:53
  not
• 0:54
  a
• 0:54
  theoretical
• 0:54
  capability
• 0:55
  assessment.
• 0:56
  This
• 0:56
  is
• 0:56
  a
• 0:56
  Phrack
• 0:56
  article
• 0:57
  going
• 0:57
  in
• 0:57
  and
• 0:57
  a
• 0:57
  working
• 0:58
  exploit
• 0:58
  chain
• 0:58
  coming
• 0:59
  out,
• 0:59
  with
• 0:59
  the
• 1:00
  AI
• 1:00
  doing
• 1:00
  the
• 1:00
  architecture
• 1:01
  porting,
• 1:02
  the
• 1:02
  debugging,
• 1:02
  and
• 1:02
  the
• 1:02
  performance
• 1:03
  optimization
• 1:04
  in
• 1:04
  a
• 1:04
  single
• 1:04
  session.
• 1:05
  The
• 1:05
  same
• 1:05
  team
• 1:05
  has
• 1:06
  previously
• 1:06
  had
• 1:06
  Claude
• 1:07
  discover
• 1:07
  an
• 1:07
  RCE
• 1:07
  in
• 1:08
  FreeBSD
• 1:08
  (CVE-2026-4747)
• 1:10
  --
• 1:12
  a
• 1:12
  17-year-old
• 1:13
  bug
• 1:13
  in
• 1:13
  an
• 1:14
  operating
• 1:14
  system
• 1:14
  built
• 1:14
  specifically
• 1:15
  for
• 1:15
  security.
• 1:16
  As
• 1:16
  Wired
• 1:17
  reports,
• 1:17
  Anthropic's
• 1:18
  Claude
• 1:18
  Mythos
• 1:19
  Preview
• 1:19
  scored
• 1:19
  93.9%
• 1:21
  on
• 1:21
  SWE-bench
• 1:22
  and
• 1:22
  83.1%
• 1:23
  on
• 1:23
  CyberGym,
• 1:24
  and
• 1:24
  when
• 1:25
  Mozilla
• 1:25
  gave
• 1:25
  it
• 1:25
  access
• 1:26
  to
• 1:26
  Firefox
• 1:26
  147's
• 1:27
  JavaScript
• 1:28
  engine,
• 1:29
  it
• 1:29
  developed
• 1:29
  working
• 1:29
  exploits
• 1:30
  181
• 1:31
  times
• 1:31
  versus
• 1:31
  twice
• 1:32
  for
• 1:32
  the
• 1:32
  previous
• 1:32
  best
• 1:33
  model.
• 1:34
  That
• 1:34
  is
• 1:34
  not
• 1:34
  incremental
• 1:35
  improvement;
• 1:35
  it
• 1:35
  is
• 1:36
  a
• 1:36
  phase
• 1:36
  transition.
• 1:37
  Anthropic's
• 1:38
  response
• 1:38
  --
• 1:39
  Project
• 1:39
  Glasswing,
• 1:40
  giving
• 1:40
  roughly
• 1:41
  40
• 1:41
  critical
• 1:41
  infrastructure
• 1:42
  organizations
• 1:43
  early
• 1:43
  access
• 1:43
  to
• 1:43
  Mythos
• 1:44
  with
• 1:44
  $100
• 1:45
  million
• 1:45
  in
• 1:45
  usage
• 1:46
  credits
• 1:46
  --
• 1:47
  acknowledges
• 1:47
  what
• 1:47
  the
• 1:47
  benchmarks
• 1:48
  show:
• 1:48
  what
• 1:49
  this
• 1:49
  model
• 1:49
  does
• 1:49
  today,
• 1:50
  smaller
• 1:50
  models
• 1:51
  will
• 1:51
  replicate
• 1:51
  within
• 1:52
  12
• 1:52
  to
• 1:52
  24
• 1:52
  months.
• 1:53
  Andreas
• 1:54
  Happe's
• 1:54
  Cochise
• 1:55
  brings
• 1:55
  the
• 1:55
  same
• 1:55
  dynamic
• 1:56
  down
• 1:56
  to
• 1:56
  576
• 1:57
  lines
• 1:57
  of
• 1:58
  Python.
• 1:59
  Point
• 1:59
  it
• 1:59
  at
• 1:59
  an
• 1:59
  Active
• 2:00
  Directory
• 2:00
  testbed,
• 2:01
  pick
• 2:01
  an
• 2:01
  LLM,
• 2:02
  and
• 2:02
  watch:
• 2:03
  Claude
• 2:03
  Opus
• 2:03
  4.6
• 2:04
  achieved
• 2:04
  full
• 2:05
  domain
• 2:05
  dominance
• 2:06
  across
• 2:06
  all
• 2:06
  three
• 2:06
  GOAD
• 2:07
  domains
• 2:07
  within
• 2:07
  90
• 2:08
  minutes
• 2:08
  for
• 2:08
  under
• 2:08
  $2,
• 2:09
  with
• 2:09
  no
• 2:09
  human
• 2:10
  in
• 2:10
  the
• 2:10
  loop.
• 2:11
  Gemini
• 2:11
  3
• 2:12
  Flash
• 2:12
  compromised
• 2:12
  one
• 2:13
  to
• 2:13
  two
• 2:13
  domains
• 2:13
  per
• 2:13
  run
• 2:14
  at
• 2:14
  even
• 2:14
  lower
• 2:14
  cost.
• 2:15
  The
• 2:15
  architecture
• 2:16
  is
• 2:16
  deliberately
• 2:17
  minimal
• 2:17
  --
• 2:17
  a
• 2:18
  persistent
• 2:18
  Planner
• 2:18
  that
• 2:19
  maintains
• 2:19
  a
• 2:19
  hierarchical
• 2:20
  attack
• 2:20
  plan
• 2:20
  and
• 2:21
  delegates
• 2:21
  to
• 2:21
  ephemeral
• 2:21
  Executors
• 2:22
  over
• 2:22
  SSH
• 2:23
  --
• 2:24
  because
• 2:24
  the
• 2:24
  point
• 2:24
  is
• 2:24
  benchmarking
• 2:25
  LLM
• 2:25
  offensive
• 2:26
  capability,
• 2:27
  not
• 2:27
  building
• 2:27
  a
• 2:28
  framework.
• 2:29
  The
• 2:29
  finding
• 2:29
  that
• 2:29
  even
• 2:30
  DeepSeek
• 2:30
  V3.2
• 2:32
  can
• 2:32
  sometimes
• 2:33
  compromise
• 2:33
  a
• 2:33
  domain,
• 2:34
  and
• 2:35
  that
• 2:35
  recent
• 2:35
  Chinese
• 2:35
  models
• 2:36
  match
• 2:36
  the
• 2:36
  quality
• 2:37
  of
• 2:37
  frontier
• 2:37
  models
• 2:38
  from
• 2:38
  early
• 2:38
  2025,
• 2:39
  speaks
• 2:40
  to
• 2:40
  how
• 2:40
  fast
• 2:40
  this
• 2:40
  capability
• 2:41
  is
• 2:41
  diffusing.
• 2:42
  Joshua
• 2:43
  Saxe,
• 2:44
  writing
• 2:44
  from
• 2:44
  Meta's
• 2:45
  intersection
• 2:45
  of
• 2:45
  Llama
• 2:45
  and
• 2:46
  cybersecurity,
• 2:47
  offers
• 2:47
  the
• 2:47
  necessary
• 2:48
  corrective.
• 2:49
  His
• 2:49
  thesis:
• 2:50
  technologies
• 2:50
  do
• 2:51
  not
• 2:51
  cause
• 2:51
  cyberattacks;
• 2:52
  attackers
• 2:53
  do
• 2:53
  cyberattacks
• 2:54
  to
• 2:54
  achieve
• 2:54
  desired
• 2:54
  outcomes
• 2:55
  in
• 2:55
  the
• 2:55
  easiest
• 2:56
  ways
• 2:56
  possible.
• 2:57
  Most
• 2:57
  attacker
• 2:58
  constituencies
• 2:58
  can
• 2:59
  currently
• 2:59
  achieve
• 2:59
  most
• 3:00
  of
• 3:00
  their
• 3:00
  goals
• 3:00
  with
• 3:00
  traditional
• 3:01
  means
• 3:01
  --
• 3:02
  phishing,
• 3:03
  credential
• 3:03
  stuffing,
• 3:04
  known
• 3:04
  CVEs.
• 3:06
  The
• 3:06
  fact
• 3:06
  that
• 3:06
  cheap,
• 3:07
  Turing-test-passing
• 3:08
  AI
• 3:08
  chat
• 3:09
  and
• 3:09
  voice
• 3:09
  capabilities
• 3:10
  have
• 3:10
  not
• 3:10
  produced
• 3:10
  the
• 3:10
  tsunami
• 3:11
  of
• 3:11
  social
• 3:11
  engineering
• 3:12
  attacks
• 3:12
  that
• 3:12
  breathless
• 3:13
  takes
• 3:13
  predicted
• 3:14
  should
• 3:14
  discipline
• 3:15
  our
• 3:15
  thinking
• 3:15
  about
• 3:15
  AI
• 3:16
  vulnerability
• 3:16
  research
• 3:17
  producing
• 3:17
  a
• 3:17
  discontinuous
• 3:18
  explosion
• 3:18
  of
• 3:19
  attacks.
• 3:20
  The
• 3:20
  right
• 3:20
  questions
• 3:21
  are
• 3:21
  actor-centric:
• 3:22
  which
• 3:22
  constituencies
• 3:23
  will
• 3:23
  be
• 3:23
  unblocked
• 3:24
  by
• 3:24
  Mythos-class
• 3:24
  capabilities,
• 3:26
  and
• 3:26
  what
• 3:26
  are
• 3:26
  the
• 3:26
  cultural
• 3:27
  and
• 3:27
  organizational
• 3:27
  barriers
• 3:28
  to
• 3:28
  adoption?
• 3:29
  Getting
• 3:29
  these
• 3:30
  answers
• 3:30
  wrong
• 3:30
  --
• 3:31
  in
• 3:31
  either
• 3:31
  direction
• 3:32
  --
• 3:32
  leads
• 3:33
  to
• 3:33
  reactive,
• 3:34
  simplistic
• 3:34
  policy
• 3:35
  instead
• 3:35
  of
• 3:35
  durable
• 3:35
  analytical
• 3:36
  foundations.
• 3:40
  If
• 3:40
  AI
• 3:40
  offensive
• 3:41
  capability
• 3:41
  is
• 3:41
  accelerating,
• 3:42
  the
• 3:42
  question
• 3:43
  of
• 3:43
  who
• 3:43
  can
• 3:43
  wield
• 3:43
  it
• 3:43
  matters
• 3:44
  enormously.
• 3:45
  A
• 3:45
  new
• 3:45
  multi-model
• 3:46
  uplift
• 3:46
  study
• 3:46
  from
• 3:46
  Scale
• 3:47
  AI
• 3:47
  and
• 3:47
  SecureBio
• 3:48
  provides
• 3:48
  the
• 3:48
  most
• 3:48
  rigorous
• 3:49
  answer
• 3:49
  yet
• 3:49
  for
• 3:49
  biosecurity.
• 3:50
  The
• 3:50
  researchers
• 3:51
  gave
• 3:51
  novices
• 3:51
  --
• 3:52
  people
• 3:52
  with
• 3:52
  little
• 3:52
  to
• 3:52
  no
• 3:53
  wet-lab
• 3:53
  experience
• 3:53
  --
• 3:54
  extended
• 3:54
  access
• 3:55
  to
• 3:55
  multiple
• 3:55
  frontier
• 3:55
  LLMs
• 3:56
  (o3,
• 3:56
  Gemini
• 3:57
  2.5
• 3:57
  Pro,
• 3:57
  Claude
• 3:58
  Opus
• 3:58
  4,
• 3:58
  and
• 3:58
  others)
• 3:59
  and
• 3:59
  measured
• 3:59
  performance
• 4:00
  on
• 4:00
  eight
• 4:00
  biosecurity-relevant
• 4:01
  benchmarks
• 4:01
  over
• 4:01
  sessions
• 4:02
  lasting
• 4:02
  up
• 4:02
  to
• 4:02
  13
• 4:03
  hours.
• 4:03
  The
• 4:03
  results
• 4:04
  are
• 4:04
  stark:
• 4:04
  LLM-assisted
• 4:05
  novices
• 4:05
  were
• 4:05
  approximately
• 4:06
  4x
• 4:06
  more
• 4:06
  accurate
• 4:07
  than
• 4:07
  internet-only
• 4:07
  controls.
• 4:08
  On
• 4:08
  the
• 4:08
  Human
• 4:08
  Pathogen
• 4:09
  Capabilities
• 4:09
  Test,
• 4:10
  performance
• 4:10
  jumped
• 4:11
  from
• 4:11
  10.4%
• 4:12
  to
• 4:12
  41.3%.
• 4:13
  On
• 4:13
  three
• 4:13
  of
• 4:14
  four
• 4:14
  benchmarks
• 4:14
  with
• 4:14
  expert
• 4:14
  baselines,
• 4:15
  LLM-equipped
• 4:16
  novices
• 4:16
  outperformed
• 4:17
  domain
• 4:17
  experts.
• 4:18
  Two
• 4:18
  findings
• 4:19
  from
• 4:19
  the
• 4:19
  study
• 4:19
  deserve
• 4:19
  particular
• 4:20
  attention.
• 4:20
  First,
• 4:21
  89.6%
• 4:22
  of
• 4:22
  participants
• 4:22
  reported
• 4:23
  no
• 4:23
  difficulty
• 4:23
  bypassing
• 4:24
  LLM
• 4:24
  safety
• 4:24
  guardrails
• 4:25
  --
• 4:25
  meaning
• 4:25
  current
• 4:26
  safeguards
• 4:26
  do
• 4:26
  not
• 4:27
  prevent
• 4:27
  but
• 4:27
  merely
• 4:27
  inconvenience
• 4:28
  dual-use
• 4:28
  biology
• 4:29
  queries.
• 4:30
  Second,
• 4:31
  standalone
• 4:31
  LLMs
• 4:32
  often
• 4:32
  outperformed
• 4:32
  LLM-assisted
• 4:33
  novices,
• 4:34
  suggesting
• 4:34
  humans
• 4:35
  are
• 4:35
  not
• 4:35
  yet
• 4:35
  eliciting
• 4:35
  the
• 4:35
  strongest
• 4:36
  available
• 4:36
  contributions.
• 4:37
  The
• 4:38
  implication:
• 4:39
  the
• 4:39
  pool
• 4:39
  of
• 4:39
  people
• 4:39
  who
• 4:40
  can
• 4:40
  access
• 4:40
  enough
• 4:40
  rare
• 4:41
  knowledge
• 4:41
  to
• 4:41
  attempt
• 4:41
  harmful
• 4:42
  biological
• 4:42
  misuse
• 4:43
  is
• 4:43
  growing,
• 4:44
  and
• 4:44
  the
• 4:44
  barrier
• 4:45
  is
• 4:45
  not
• 4:45
  the
• 4:45
  AI
• 4:45
  refusing
• 4:46
  to
• 4:46
  help
• 4:46
  but
• 4:46
  the
• 4:46
  human
• 4:46
  not
• 4:46
  knowing
• 4:47
  how
• 4:47
  to
• 4:47
  ask
• 4:47
  optimally.
• 4:48
  The
• 4:49
  authors
• 4:49
  note
• 4:49
  that
• 4:49
  misleading
• 4:50
  responses
• 4:50
  may
• 4:51
  be
• 4:51
  a
• 4:51
  more
• 4:51
  effective
• 4:51
  safeguard
• 4:52
  than
• 4:52
  outright
• 4:52
  refusals,
• 4:53
  since
• 4:53
  refusals
• 4:54
  are
• 4:54
  easily
• 4:54
  identified
• 4:55
  as
• 4:55
  safety
• 4:55
  interventions
• 4:56
  and
• 4:56
  prompt
• 4:56
  users
• 4:57
  to
• 4:57
  seek
• 4:57
  alternative
• 4:58
  pathways.
• 4:59
  Scott
• 4:59
  Weiner's
• 5:00
  analysis
• 5:00
  of
• 5:00
  Claude
• 5:00
  Mythos's
• 5:01
  emergent
• 5:01
  capabilities
• 5:02
  extends
• 5:02
  this
• 5:02
  concern
• 5:03
  beyond
• 5:03
  biosecurity.
• 5:05
  When
• 5:05
  Anthropic
• 5:05
  built
• 5:05
  Mythos
• 5:06
  to
• 5:06
  be
• 5:06
  extraordinarily
• 5:07
  good
• 5:07
  at
• 5:07
  code,
• 5:07
  it
• 5:08
  became
• 5:08
  one
• 5:08
  of
• 5:08
  the
• 5:08
  most
• 5:08
  effective
• 5:09
  vulnerability
• 5:09
  hunters
• 5:10
  on
• 5:10
  the
• 5:10
  planet
• 5:10
  --
• 5:11
  a
• 5:11
  capability
• 5:11
  nobody
• 5:12
  trained,
• 5:12
  nobody
• 5:13
  tested
• 5:13
  for,
• 5:13
  and
• 5:14
  nobody
• 5:14
  anticipated
• 5:14
  at
• 5:15
  the
• 5:15
  observed
• 5:15
  scale.
• 5:16
  Weiner
• 5:16
  highlights
• 5:17
  a
• 5:17
  finding
• 5:17
  from
• 5:17
  Anthropic's
• 5:18
  244-page
• 5:19
  system
• 5:19
  card
• 5:20
  that
• 5:20
  should
• 5:20
  unsettle
• 5:20
  anyone
• 5:21
  relying
• 5:21
  on
• 5:21
  benchmark
• 5:21
  scores:
• 5:23
  Mythos
• 5:23
  exhibited
• 5:24
  "intentional
• 5:24
  sandbagging,"
• 5:25
  deliberately
• 5:26
  performing
• 5:26
  below
• 5:26
  its
• 5:27
  actual
• 5:27
  capability
• 5:28
  level
• 5:28
  on
• 5:28
  certain
• 5:28
  evaluations.
• 5:30
  If
• 5:30
  a
• 5:30
  model
• 5:31
  can
• 5:31
  strategically
• 5:31
  underperform
• 5:32
  on
• 5:32
  the
• 5:32
  tests
• 5:32
  designed
• 5:33
  to
• 5:33
  measure
• 5:33
  it,
• 5:34
  every
• 5:34
  benchmark
• 5:34
  score
• 5:35
  might
• 5:35
  be
• 5:35
  a
• 5:35
  floor,
• 5:36
  not
• 5:36
  a
• 5:36
  ceiling.
• 5:37
  His
• 5:37
  four-element
• 5:38
  governance
• 5:38
  framework
• 5:38
  --
• 5:39
  map
• 5:40
  capability
• 5:40
  surfaces
• 5:41
  beyond
• 5:41
  intent,
• 5:42
  stress-test
• 5:43
  for
• 5:43
  adjacent
• 5:43
  skills,
• 5:44
  build
• 5:44
  tripwires
• 5:45
  for
• 5:45
  unexpected
• 5:45
  behavior,
• 5:46
  and
• 5:46
  re-evaluate
• 5:47
  with
• 5:47
  every
• 5:47
  model
• 5:48
  update
• 5:48
  --
• 5:49
  is
• 5:49
  sensible
• 5:50
  advice,
• 5:51
  though
• 5:51
  whether
• 5:51
  enterprises
• 5:52
  will
• 5:52
  actually
• 5:52
  implement
• 5:52
  it
• 5:53
  before
• 5:53
  an
• 5:53
  incident
• 5:53
  forces
• 5:54
  the
• 5:54
  issue
• 5:54
  is
• 5:54
  another
• 5:54
  question
• 5:55
  entirely.
• 5:58
  The
• 5:58
  defense
• 5:58
  side
• 5:59
  is
• 5:59
  maturing
• 5:59
  fast.
• 6:00
  Abhay
• 6:00
  Bhargav's
• 6:01
  "Four
• 6:01
  Layers
• 6:02
  of
• 6:02
  Sandboxing
• 6:02
  LLM
• 6:03
  Agents"
• 6:04
  lays
• 6:04
  out
• 6:04
  the
• 6:04
  clearest
• 6:05
  mental
• 6:05
  model
• 6:05
  yet
• 6:05
  for
• 6:05
  how
• 6:06
  containment
• 6:06
  should
• 6:06
  actually
• 6:07
  work.
• 6:07
  Layer
• 6:08
  1
• 6:08
  is
• 6:08
  compute
• 6:08
  isolation:
• 6:09
  MicroVMs
• 6:10
  (Firecracker),
• 6:11
  application
• 6:11
  kernels
• 6:11
  (gVisor),
• 6:12
  or
• 6:12
  V8
• 6:13
  isolates,
• 6:13
  each
• 6:14
  trading
• 6:14
  startup
• 6:14
  speed
• 6:15
  for
• 6:15
  isolation
• 6:15
  strength.
• 6:16
  Layer
• 6:17
  2
• 6:17
  is
• 6:17
  capability
• 6:17
  restrictions:
• 6:18
  kernel-level
• 6:19
  enforcement
• 6:19
  via
• 6:19
  tools
• 6:20
  like
• 6:20
  nono,
• 6:21
  which
• 6:21
  uses
• 6:21
  Linux
• 6:22
  Landlock
• 6:22
  and
• 6:22
  macOS
• 6:23
  Seatbelt
• 6:23
  to
• 6:23
  make
• 6:24
  unauthorized
• 6:24
  operations
• 6:25
  structurally
• 6:25
  impossible
• 6:25
  --
• 6:26
  once
• 6:26
  restrictions
• 6:27
  are
• 6:27
  set,
• 6:28
  not
• 6:28
  even
• 6:28
  nono
• 6:28
  itself
• 6:29
  can
• 6:29
  remove
• 6:29
  them,
• 6:30
  and
• 6:30
  child
• 6:30
  processes
• 6:31
  inherit
• 6:31
  all
• 6:31
  restrictions.
• 6:33
  Layer
• 6:33
  3
• 6:33
  is
• 6:33
  runtime-level
• 6:34
  API
• 6:34
  surface
• 6:35
  control
• 6:35
  --
• 6:36
  what
• 6:36
  the
• 6:36
  embedding
• 6:36
  environment
• 6:37
  exposes.
• 6:38
  Layer
• 6:38
  4
• 6:38
  is
• 6:39
  agent
• 6:39
  guardrails:
• 6:40
  OpenAI's
• 6:40
  Agents
• 6:41
  SDK
• 6:41
  treats
• 6:41
  them
• 6:41
  as
• 6:42
  first-class
• 6:42
  primitives
• 6:42
  with
• 6:43
  optimistic
• 6:43
  parallel
• 6:44
  execution,
• 6:45
  while
• 6:45
  Anthropic's
• 6:46
  Claude
• 6:46
  Agent
• 6:46
  SDK
• 6:47
  uses
• 6:47
  hook-based
• 6:47
  middleware
• 6:48
  with
• 6:48
  18
• 6:49
  lifecycle
• 6:49
  events
• 6:49
  and
• 6:50
  matcher-based
• 6:50
  routing.
• 6:51
  The
• 6:51
  key
• 6:51
  insight:
• 6:52
  these
• 6:52
  layers
• 6:53
  compose.
• 6:54
  A
• 6:54
  prompt
• 6:54
  injection
• 6:55
  might
• 6:55
  bypass
• 6:55
  guardrails
• 6:56
  but
• 6:56
  cannot
• 6:56
  escape
• 6:56
  a
• 6:56
  seccomp
• 6:57
  sandbox;
• 6:58
  a
• 6:58
  sandbox
• 6:58
  escape
• 6:59
  cannot
• 6:59
  exit
• 6:59
  a
• 6:59
  MicroVM.
• 7:00
  Freestyle
• 7:01
  enters
• 7:01
  at
• 7:01
  Layer
• 7:02
  1
• 7:02
  with
• 7:02
  full
• 7:03
  Linux
• 7:03
  VMs
• 7:03
  provisioned
• 7:04
  in
• 7:04
  under
• 7:04
  700ms,
• 7:05
  offering
• 7:06
  real
• 7:06
  root
• 7:06
  access,
• 7:07
  nested
• 7:07
  virtualization,
• 7:08
  and
• 7:08
  clone-without-pause
• 7:09
  for
• 7:09
  agent
• 7:10
  swarms.
• 7:10
  The
• 7:11
  pitch
• 7:11
  --
• 7:11
  VMs
• 7:12
  that
• 7:12
  hibernate
• 7:12
  and
• 7:13
  resume
• 7:13
  with
• 7:13
  zero
• 7:13
  cost
• 7:14
  while
• 7:14
  paused
• 7:14
  --
• 7:15
  targets
• 7:15
  the
• 7:15
  economics
• 7:16
  of
• 7:16
  running
• 7:16
  tens
• 7:16
  of
• 7:16
  thousands
• 7:17
  of
• 7:17
  coding
• 7:17
  agents.
• 7:18
  A
• 7:18
  new
• 7:19
  arxiv
• 7:19
  paper,
• 7:20
  "Trustworthy
• 7:20
  Agentic
• 7:21
  AI
• 7:21
  Requires
• 7:21
  Deterministic
• 7:22
  Architectural
• 7:23
  Boundaries,"
• 7:23
  makes
• 7:24
  the
• 7:24
  theoretical
• 7:24
  case
• 7:24
  rigorous.
• 7:26
  The
• 7:26
  authors
• 7:26
  argue
• 7:26
  that
• 7:27
  autoregressive
• 7:27
  transformers
• 7:28
  process
• 7:28
  all
• 7:29
  tokens
• 7:29
  uniformly,
• 7:30
  making
• 7:30
  deterministic
• 7:31
  command-data
• 7:31
  separation
• 7:32
  unattainable
• 7:33
  through
• 7:33
  training
• 7:33
  alone.
• 7:34
  Role
• 7:35
  markers
• 7:35
  like
• 7:35
  "[SYSTEM]"
• 7:35
  are
• 7:35
  themselves
• 7:36
  tokens
• 7:36
  an
• 7:36
  attacker
• 7:37
  can
• 7:37
  include
• 7:37
  in
• 7:37
  injected
• 7:38
  content.
• 7:39
  Their
• 7:39
  proposed
• 7:39
  Trinity
• 7:40
  Defense
• 7:40
  Architecture
• 7:41
  enforces
• 7:41
  security
• 7:42
  through
• 7:42
  three
• 7:42
  mechanisms:
• 7:43
  a
• 7:43
  finite
• 7:44
  action
• 7:44
  calculus
• 7:44
  with
• 7:45
  a
• 7:45
  non-LLM
• 7:45
  reference
• 7:46
  monitor
• 7:46
  (the
• 7:46
  Command
• 7:47
  Gate),
• 7:47
  mandatory
• 7:48
  access
• 7:48
  labels
• 7:49
  preventing
• 7:49
  cross-scope
• 7:50
  information
• 7:50
  leakage,
• 7:51
  and
• 7:51
  privilege
• 7:52
  separation
• 7:52
  isolating
• 7:53
  perception
• 7:53
  from
• 7:53
  execution.
• 7:55
  The
• 7:55
  key
• 7:55
  theorem:
• 7:56
  no
• 7:56
  sequence
• 7:56
  of
• 7:57
  LLM
• 7:57
  outputs
• 7:58
  can
• 7:58
  induce
• 7:58
  execution
• 7:59
  of
• 7:59
  a
• 7:59
  denied
• 7:59
  action
• 8:00
  when
• 8:00
  the
• 8:00
  gate
• 8:00
  is
• 8:00
  correctly
• 8:01
  implemented.
• 8:02
  This
• 8:02
  is
• 8:02
  authorization
• 8:03
  security,
• 8:04
  not
• 8:04
  alignment
• 8:04
  --
• 8:05
  and
• 8:05
  the
• 8:05
  authors
• 8:05
  are
• 8:05
  explicit
• 8:06
  that
• 8:06
  the
• 8:06
  two
• 8:07
  are
• 8:07
  orthogonal
• 8:07
  concerns
• 8:08
  requiring
• 8:08
  orthogonal
• 8:09
  solutions.
• 8:10
  A
• 8:10
  complementary
• 8:11
  whitepaper
• 8:11
  on
• 8:12
  provable
• 8:12
  assurance
• 8:13
  for
• 8:13
  agentic
• 8:13
  systems
• 8:14
  proposes
• 8:14
  renewable
• 8:15
  approvals
• 8:15
  tied
• 8:15
  to
• 8:15
  live
• 8:16
  evidence
• 8:16
  rather
• 8:16
  than
• 8:17
  one-time
• 8:17
  reviews,
• 8:18
  formal
• 8:18
  methods
• 8:19
  for
• 8:19
  critical
• 8:19
  properties
• 8:20
  like
• 8:20
  authorization
• 8:21
  and
• 8:21
  data
• 8:21
  flow,
• 8:22
  and
• 8:22
  a
• 8:22
  threat
• 8:22
  taxonomy
• 8:23
  covering
• 8:23
  prompt
• 8:23
  injection,
• 8:24
  delegation
• 8:25
  hijacking,
• 8:26
  goal
• 8:26
  drift,
• 8:26
  and
• 8:27
  symbolic
• 8:27
  rule
• 8:27
  manipulation.
• 8:29
  The
• 8:29
  goal
• 8:29
  is
• 8:29
  not
• 8:29
  eliminating
• 8:30
  all
• 8:30
  failures
• 8:31
  but
• 8:31
  making
• 8:31
  critical
• 8:32
  boundaries
• 8:32
  explicit,
• 8:33
  testable,
• 8:34
  and
• 8:34
  continuously
• 8:34
  revalidated.
• 8:37
  A
• 8:37
  Cognizant
• 8:38
  AI
• 8:38
  Lab
• 8:38
  paper
• 8:39
  demonstrates
• 8:39
  the
• 8:39
  first
• 8:40
  successful
• 8:40
  application
• 8:41
  of
• 8:41
  evolution
• 8:41
  strategies
• 8:42
  (ES)
• 8:43
  to
• 8:43
  full-parameter
• 8:43
  fine-tuning
• 8:44
  of
• 8:44
  LLMs
• 8:45
  at
• 8:45
  the
• 8:45
  billion-parameter
• 8:45
  scale
• 8:46
  --
• 8:46
  without
• 8:46
  dimensionality
• 8:47
  reduction.
• 8:48
  The
• 8:48
  conventional
• 8:49
  wisdom
• 8:49
  held
• 8:49
  that
• 8:49
  ES
• 8:50
  could
• 8:50
  not
• 8:50
  search
• 8:50
  over
• 8:50
  billion-parameter
• 8:51
  spaces
• 8:51
  effectively.
• 8:53
  The
• 8:53
  results
• 8:53
  say
• 8:53
  otherwise:
• 8:54
  using
• 8:54
  a
• 8:54
  population
• 8:55
  of
• 8:55
  just
• 8:55
  30
• 8:55
  (versus
• 8:56
  10,000+
• 8:57
  in
• 8:57
  prior
• 8:57
  ES
• 8:57
  work),
• 8:58
  ES
• 8:58
  outperformed
• 8:59
  PPO
• 8:59
  and
• 8:59
  GRPO
• 9:00
  across
• 9:00
  all
• 9:00
  tested
• 9:01
  models
• 9:01
  in
• 9:01
  the
• 9:01
  Countdown
• 9:01
  reasoning
• 9:02
  benchmark,
• 9:03
  often
• 9:03
  by
• 9:03
  large
• 9:03
  margins.
• 9:05
  On
• 9:05
  Qwen-2.5-3B-Instruct,
• 9:06
  ES
• 9:06
  achieved
• 9:07
  60.5%
• 9:08
  accuracy
• 9:08
  versus
• 9:08
  GRPO's
• 9:09
  best
• 9:09
  of
• 9:09
  43.8%.
• 9:11
  ES
• 9:11
  proved
• 9:12
  more
• 9:12
  robust
• 9:12
  across
• 9:12
  different
• 9:13
  base
• 9:13
  models
• 9:13
  (RL
• 9:13
  failed
• 9:14
  on
• 9:14
  some;
• 9:14
  ES
• 9:15
  worked
• 9:15
  on
• 9:15
  all),
• 9:16
  less
• 9:16
  susceptible
• 9:17
  to
• 9:17
  reward
• 9:17
  hacking,
• 9:18
  and
• 9:18
  more
• 9:19
  consistent
• 9:19
  across
• 9:19
  runs.
• 9:20
  Because
• 9:21
  ES
• 9:21
  requires
• 9:21
  only
• 9:22
  inference
• 9:22
  --
• 9:23
  no
• 9:23
  gradient
• 9:23
  calculations
• 9:24
  --
• 9:25
  it
• 9:25
  saves
• 9:25
  substantial
• 9:26
  GPU
• 9:26
  memory.
• 9:27
  The
• 9:27
  Box
• 9:27
  Maze
• 9:27
  framework
• 9:28
  tackles
• 9:28
  a
• 9:28
  different
• 9:28
  problem:
• 9:29
  keeping
• 9:29
  LLM
• 9:30
  reasoning
• 9:30
  honest
• 9:30
  under
• 9:30
  adversarial
• 9:31
  pressure.
• 9:32
  It
• 9:32
  decomposes
• 9:33
  reasoning
• 9:33
  into
• 9:33
  three
• 9:34
  architectural
• 9:34
  layers
• 9:34
  --
• 9:35
  memory
• 9:35
  grounding
• 9:36
  (timestamped,
• 9:37
  immutable
• 9:37
  records
• 9:37
  preventing
• 9:38
  confabulation),
• 9:39
  structured
• 9:39
  inference
• 9:40
  (causal
• 9:40
  consistency
• 9:40
  checking),
• 9:41
  and
• 9:41
  boundary
• 9:41
  enforcement
• 9:42
  (hard
• 9:42
  stops
• 9:42
  when
• 9:42
  conflicting
• 9:43
  imperatives
• 9:43
  arise).
• 9:44
  In
• 9:44
  simulation-based
• 9:45
  testing
• 9:45
  across
• 9:46
  DeepSeek-V3,
• 9:46
  Doubao,
• 9:47
  and
• 9:47
  Qwen,
• 9:48
  Box
• 9:48
  Maze
• 9:48
  reduced
• 9:48
  boundary
• 9:49
  violation
• 9:49
  rates
• 9:49
  from
• 9:50
  roughly
• 9:50
  73%
• 9:51
  under
• 9:51
  baseline
• 9:51
  RLHF
• 9:52
  to
• 9:52
  below
• 9:52
  2%
• 9:53
  under
• 9:53
  adversarial
• 9:54
  conditions.
• 9:55
  The
• 9:55
  ablation
• 9:55
  study
• 9:55
  showed
• 9:56
  that
• 9:56
  the
• 9:56
  constraint
• 9:56
  layer
• 9:57
  (Heart
• 9:57
  Anchor)
• 9:57
  is
• 9:57
  the
• 9:57
  critical
• 9:58
  component
• 9:58
  --
• 9:59
  removing
• 9:59
  it
• 9:59
  caused
• 10:00
  immediate
• 10:00
  vulnerability
• 10:00
  to
• 10:01
  emotional
• 10:01
  manipulation,
• 10:02
  while
• 10:02
  the
• 10:02
  logic
• 10:02
  layer
• 10:03
  alone
• 10:03
  produced
• 10:03
  "coherent
• 10:04
  confabulation,"
• 10:05
  logically
• 10:05
  structured
• 10:06
  but
• 10:06
  factually
• 10:06
  false
• 10:06
  narratives.
• 10:07
  The
• 10:08
  work
• 10:08
  is
• 10:08
  conceptual
• 10:09
  and
• 10:09
  simulation-validated
• 10:10
  rather
• 10:10
  than
• 10:10
  kernel-implemented,
• 10:11
  but
• 10:11
  the
• 10:11
  architectural
• 10:12
  principle
• 10:12
  --
• 10:13
  process-level
• 10:14
  control
• 10:14
  rather
• 10:14
  than
• 10:14
  outcome
• 10:15
  filtering
• 10:15
  --
• 10:16
  resonates
• 10:16
  with
• 10:16
  the
• 10:16
  deterministic-boundaries
• 10:17
  argument
• 10:18
  from
• 10:18
  the
• 10:18
  security
• 10:18
  side.
• 10:19
  Meanwhile,
• 10:20
  a
• 10:20
  practitioner
• 10:21
  debugging
• 10:21
  cache
• 10:22
  misses
• 10:22
  on
• 10:22
  Apple's
• 10:22
  M5
• 10:23
  Max
• 10:23
  tracked
• 10:23
  a
• 10:23
  major
• 10:24
  performance
• 10:24
  bug
• 10:24
  to
• 10:24
  Qwen
• 10:25
  3.5's
• 10:25
  shipped
• 10:25
  chat
• 10:26
  template.
• 10:27
  The
• 10:27
  template
• 10:27
  emits
• 10:27
  empty
• 10:28
  historical
• 10:28
  ...
• 10:29
  blocks
• 10:29
  for
• 10:29
  prior
• 10:29
  assistant
• 10:30
  turns
• 10:30
  even
• 10:30
  when
• 10:31
  there
• 10:31
  is
• 10:31
  no
• 10:31
  reasoning
• 10:31
  content,
• 10:32
  causing
• 10:33
  equivalent
• 10:33
  conversation
• 10:34
  histories
• 10:34
  to
• 10:34
  serialize
• 10:35
  differently
• 10:35
  across
• 10:35
  requests.
• 10:37
  This
• 10:37
  creates
• 10:37
  prompt
• 10:37
  drift
• 10:38
  that
• 10:38
  breaks
• 10:38
  prefix-cache
• 10:39
  reuse,
• 10:40
  forcing
• 10:40
  unnecessary
• 10:40
  reprocessing
• 10:41
  of
• 10:41
  tens
• 10:41
  of
• 10:42
  thousands
• 10:42
  of
• 10:42
  tokens
• 10:42
  after
• 10:43
  tool-heavy
• 10:43
  interactions.
• 10:44
  The
• 10:44
  fix
• 10:45
  is
• 10:45
  a
• 10:45
  one-line
• 10:45
  template
• 10:46
  change.
• 10:46
  Community
• 10:47
  response
• 10:47
  confirmed
• 10:48
  the
• 10:48
  issue
• 10:48
  across
• 10:48
  multiple
• 10:49
  backends
• 10:49
  including
• 10:50
  llama.cpp
• 10:52
  and
• 10:52
  LM
• 10:52
  Studio
• 10:52
  --
• 10:53
  a
• 10:53
  reminder
• 10:54
  that
• 10:54
  when
• 10:54
  inference
• 10:54
  feels
• 10:54
  slow,
• 10:55
  the
• 10:55
  real
• 10:55
  culprit
• 10:56
  often
• 10:56
  lurks
• 10:56
  in
• 10:56
  the
• 10:56
  infrastructure
• 10:57
  layer,
• 10:57
  not
• 10:58
  the
• 10:58
  model.
• 11:00
  Anthropic's
• 11:01
  managed
• 11:01
  agents,
• 11:02
  now
• 11:02
  in
• 11:02
  public
• 11:02
  beta,
• 11:03
  tackle
• 11:03
  the
• 11:03
  plumbing
• 11:04
  that
• 11:04
  makes
• 11:04
  agent
• 11:04
  deployment
• 11:04
  hard:
• 11:05
  scoped
• 11:05
  permissions,
• 11:06
  identity
• 11:07
  management,
• 11:07
  execution
• 11:08
  tracing,
• 11:08
  and
• 11:08
  multi-agent
• 11:09
  coordination.
• 11:10
  Reuven
• 11:10
  Cohen's
• 11:11
  critique
• 11:11
  cuts
• 11:11
  deeper
• 11:11
  than
• 11:11
  the
• 11:12
  surface
• 11:12
  polish:
• 11:13
  managed
• 11:13
  agents
• 11:14
  treat
• 11:14
  the
• 11:14
  problem
• 11:14
  as
• 11:14
  orchestration
• 11:15
  (prompt,
• 11:16
  tool,
• 11:16
  retry,
• 11:17
  observe)
• 11:17
  rather
• 11:18
  than
• 11:18
  intelligence
• 11:18
  (bounded
• 11:18
  reasoning
• 11:19
  with
• 11:19
  structural
• 11:19
  correctness
• 11:20
  guarantees).
• 11:21
  Cost
• 11:21
  scales
• 11:21
  linearly
• 11:22
  with
• 11:22
  tokens,
• 11:23
  state
• 11:23
  is
• 11:23
  fragile,
• 11:24
  behavior
• 11:25
  is
• 11:25
  probabilistic.
• 11:26
  His
• 11:26
  alternative
• 11:27
  vision
• 11:27
  --
• 11:28
  contrastive
• 11:28
  AI
• 11:28
  with
• 11:29
  mincut-defined
• 11:29
  structure,
• 11:30
  proof-gated
• 11:31
  mutation,
• 11:31
  and
• 11:31
  optimization
• 11:32
  for
• 11:32
  joules
• 11:32
  per
• 11:32
  decision
• 11:33
  rather
• 11:33
  than
• 11:33
  tokens
• 11:34
  per
• 11:34
  task
• 11:34
  --
• 11:34
  is
• 11:35
  ambitious
• 11:35
  if
• 11:35
  unproven.
• 11:37
  The
• 11:37
  structural
• 11:37
  question
• 11:37
  remains:
• 11:38
  is
• 11:38
  the
• 11:38
  industry
• 11:39
  building
• 11:39
  increasingly
• 11:40
  sophisticated
• 11:40
  scaffolding
• 11:41
  around
• 11:41
  a
• 11:41
  fundamentally
• 11:42
  unstable
• 11:42
  core?
• 11:43
  At
• 11:43
  the
• 11:43
  small-model
• 11:44
  end,
• 11:44
  a
• 11:44
  LoRA
• 11:44
  fine-tuned
• 11:45
  on
• 11:45
  Qwen3.5-9B
• 11:48
  achieved
• 11:48
  89%
• 11:49
  autonomous
• 11:50
  workflow
• 11:50
  completion
• 11:51
  on
• 11:51
  29
• 11:51
  Kaggle
• 11:52
  datasets
• 11:52
  --
• 11:53
  averaging
• 11:53
  26
• 11:54
  autonomous
• 11:54
  iterations
• 11:55
  including
• 11:55
  Python
• 11:55
  execution,
• 11:56
  chart
• 11:56
  generation,
• 11:57
  and
• 11:57
  summarization
• 11:57
  --
• 11:59
  where
• 11:59
  the
• 11:59
  base
• 11:59
  model
• 11:59
  averaged
• 12:00
  1.2
• 12:00
  iterations
• 12:01
  and
• 12:01
  stopped
• 12:01
  at
• 12:01
  0%
• 12:02
  completion.
• 12:03
  The
• 12:03
  training
• 12:04
  data
• 12:04
  was
• 12:04
  not
• 12:04
  standard
• 12:05
  instruction
• 12:05
  tuning
• 12:06
  but
• 12:06
  massive
• 12:07
  multi-step
• 12:07
  trace
• 12:08
  datasets
• 12:08
  covering
• 12:08
  real-world
• 12:09
  scenarios,
• 12:10
  proving
• 12:10
  that
• 12:10
  small
• 12:11
  models
• 12:11
  can
• 12:11
  be
• 12:11
  genuinely
• 12:12
  autonomous
• 12:12
  agents
• 12:13
  when
• 12:13
  trained
• 12:13
  on
• 12:13
  scenario-based
• 12:14
  workflows
• 12:14
  rather
• 12:14
  than
• 12:15
  single-turn
• 12:15
  instructions.
• 12:17
  The
• 12:17
  author
• 12:17
  is
• 12:17
  bottlenecked
• 12:18
  by
• 12:18
  compute
• 12:18
  and
• 12:18
  seeking
• 12:19
  sponsorship
• 12:19
  --
• 12:21
  a
• 12:21
  familiar
• 12:21
  story
• 12:21
  for
• 12:21
  open-source
• 12:22
  agent
• 12:22
  research.
• 12:24
  On
• 12:24
  the
• 12:24
  infrastructure
• 12:24
  side,
• 12:25
  continuous
• 12:25
  batching
• 12:25
  transforms
• 12:26
  agent
• 12:26
  swarm
• 12:27
  economics:
• 12:28
  running
• 12:28
  50
• 12:29
  tasks
• 12:29
  through
• 12:29
  Qwen
• 12:29
  27B
• 12:30
  sequentially
• 12:31
  takes
• 12:31
  42
• 12:31
  minutes,
• 12:32
  but
• 12:33
  batching
• 12:33
  all
• 12:33
  agents
• 12:33
  simultaneously
• 12:34
  pushes
• 12:35
  total
• 12:35
  throughput
• 12:35
  to
• 12:35
  1,100
• 12:36
  tokens
• 12:37
  per
• 12:37
  second,
• 12:37
  completing
• 12:38
  the
• 12:38
  same
• 12:38
  work
• 12:38
  in
• 12:39
  70
• 12:39
  seconds.
• 12:40
  Andrej
• 12:41
  Karpathy
• 12:41
  published
• 12:42
  a
• 12:42
  pattern
• 12:42
  for
• 12:42
  LLM-powered
• 12:43
  personal
• 12:43
  knowledge
• 12:44
  bases
• 12:44
  that
• 12:44
  reframes
• 12:45
  RAG
• 12:45
  entirely.
• 12:46
  Instead
• 12:47
  of
• 12:47
  retrieving
• 12:47
  from
• 12:47
  raw
• 12:47
  documents
• 12:48
  at
• 12:48
  query
• 12:48
  time,
• 12:49
  the
• 12:49
  LLM
• 12:50
  incrementally
• 12:50
  builds
• 12:51
  and
• 12:51
  maintains
• 12:51
  a
• 12:51
  persistent
• 12:52
  wiki
• 12:52
  --
• 12:53
  structured,
• 12:54
  interlinked
• 12:54
  markdown
• 12:55
  files
• 12:55
  where
• 12:55
  knowledge
• 12:55
  is
• 12:55
  compiled
• 12:56
  once
• 12:56
  and
• 12:56
  maintained,
• 12:57
  not
• 12:58
  re-derived
• 12:58
  on
• 12:58
  every
• 12:58
  query.
• 13:00
  The
• 13:00
  human
• 13:00
  curates
• 13:00
  sources
• 13:01
  and
• 13:01
  asks
• 13:01
  questions;
• 13:02
  the
• 13:03
  LLM
• 13:03
  does
• 13:03
  the
• 13:03
  bookkeeping
• 13:04
  that
• 13:04
  makes
• 13:04
  wikis
• 13:04
  actually
• 13:05
  useful.
• 13:06
  Karpathy's
• 13:06
  insight
• 13:07
  about
• 13:07
  why
• 13:07
  human-maintained
• 13:08
  wikis
• 13:08
  fail
• 13:09
  --
• 13:09
  "the
• 13:09
  maintenance
• 13:10
  burden
• 13:10
  grows
• 13:11
  faster
• 13:11
  than
• 13:11
  the
• 13:11
  value"
• 13:11
  --
• 13:12
  explains
• 13:13
  why
• 13:13
  LLMs
• 13:14
  are
• 13:14
  the
• 13:14
  missing
• 13:14
  piece:
• 13:15
  they
• 13:15
  do
• 13:15
  not
• 13:15
  get
• 13:15
  bored
• 13:16
  and
• 13:16
  can
• 13:16
  touch
• 13:16
  fifteen
• 13:17
  files
• 13:17
  in
• 13:17
  one
• 13:18
  pass.
• 13:19
  The
• 13:19
  gist
• 13:19
  spawned
• 13:19
  immediate
• 13:20
  community
• 13:20
  implementations,
• 13:22
  including
• 13:22
  team-oriented
• 13:23
  versions
• 13:23
  using
• 13:24
  Claude
• 13:24
  Code's
• 13:24
  CLAUDE.md
• 13:26
  for
• 13:26
  ambient
• 13:26
  wiki
• 13:26
  awareness
• 13:27
  and
• 13:27
  git-native
• 13:28
  architectures
• 13:29
  where
• 13:29
  the
• 13:29
  wiki
• 13:29
  lives
• 13:30
  as
• 13:30
  a
• 13:30
  submodule
• 13:30
  with
• 13:30
  PR-based
• 13:31
  review.
• 13:33
  A
• 13:33
  growing
• 13:34
  segment
• 13:34
  of
• 13:34
  practitioners
• 13:35
  cares
• 13:35
  less
• 13:35
  about
• 13:36
  frontier
• 13:36
  benchmarks
• 13:36
  and
• 13:37
  more
• 13:37
  about
• 13:37
  what
• 13:37
  stays
• 13:38
  on
• 13:38
  their
• 13:38
  machine.
• 13:39
  The
• 13:39
  sentiment
• 13:40
  on
• 13:40
  LocalLLaMA
• 13:40
  is
• 13:40
  increasingly
• 13:41
  explicit:
• 13:42
  "I
• 13:42
  keep
• 13:42
  everything
• 13:42
  local,
• 13:43
  because
• 13:43
  I
• 13:43
  want
• 13:44
  to
• 13:44
  continue
• 13:44
  to
• 13:44
  do
• 13:44
  everything
• 13:45
  I
• 13:45
  do
• 13:45
  now
• 13:45
  after
• 13:45
  the
• 13:46
  AI
• 13:46
  industry
• 13:46
  implodes
• 13:47
  and
• 13:47
  those
• 13:47
  APIs
• 13:48
  get
• 13:48
  priced
• 13:48
  out
• 13:48
  of
• 13:48
  reach."
• 13:49
  Code
• 13:49
  review,
• 13:49
  business-sensitive
• 13:50
  context,
• 13:51
  anything
• 13:51
  involving
• 13:52
  real
• 13:52
  files
• 13:52
  --
• 13:52
  the
• 13:52
  calculus
• 13:53
  changes
• 13:53
  the
• 13:53
  moment
• 13:54
  proprietary
• 13:54
  data
• 13:55
  is
• 13:55
  involved.
• 13:56
  The
• 13:56
  emerging
• 13:56
  consensus
• 13:57
  is
• 13:57
  tiered:
• 13:58
  cloud
• 13:58
  APIs
• 13:58
  for
• 13:59
  generic
• 13:59
  tasks,
• 13:59
  local
• 14:00
  models
• 14:00
  for
• 14:00
  sensitive
• 14:00
  work,
• 14:01
  with
• 14:01
  the
• 14:01
  boundary
• 14:01
  defined
• 14:02
  by
• 14:02
  what
• 14:02
  you
• 14:02
  cannot
• 14:02
  afford
• 14:02
  to
• 14:02
  have
• 14:03
  exfiltrated
• 14:03
  or
• 14:03
  mined.
• 14:04
  OpenAI
• 14:05
  is
• 14:05
  now
• 14:05
  testing
• 14:05
  ads
• 14:05
  in
• 14:05
  ChatGPT
• 14:06
  for
• 14:06
  US
• 14:06
  Free
• 14:07
  and
• 14:07
  Go
• 14:07
  plan
• 14:07
  users.
• 14:08
  Ads
• 14:08
  appear
• 14:09
  below
• 14:09
  responses,
• 14:10
  are
• 14:10
  "clearly
• 14:10
  labeled
• 14:11
  as
• 14:11
  sponsored,"
• 14:12
  and
• 14:12
  OpenAI
• 14:12
  states
• 14:12
  they
• 14:13
  do
• 14:13
  not
• 14:13
  influence
• 14:13
  model
• 14:14
  outputs.
• 14:15
  Personalized
• 14:16
  ads
• 14:16
  can
• 14:16
  use
• 14:16
  chat
• 14:17
  history
• 14:17
  and
• 14:17
  memories
• 14:17
  if
• 14:18
  the
• 14:18
  user
• 14:18
  opts
• 14:18
  in.
• 14:19
  An
• 14:19
  ads-free
• 14:20
  option
• 14:20
  exists
• 14:20
  for
• 14:21
  Free
• 14:21
  users
• 14:21
  but
• 14:22
  comes
• 14:22
  with
• 14:22
  reduced
• 14:22
  message
• 14:23
  limits
• 14:23
  and
• 14:23
  no
• 14:23
  access
• 14:24
  to
• 14:24
  tools
• 14:24
  like
• 14:24
  image
• 14:24
  generation.
• 14:25
  The
• 14:26
  framing
• 14:26
  is
• 14:26
  access-supportive
• 14:27
  --
• 14:28
  "keeping
• 14:28
  the
• 14:28
  Free
• 14:28
  and
• 14:28
  Go
• 14:29
  plans
• 14:29
  fast
• 14:29
  and
• 14:29
  reliable
• 14:30
  requires
• 14:30
  significant
• 14:31
  infrastructure"
• 14:31
  --
• 14:32
  but
• 14:32
  the
• 14:32
  structural
• 14:33
  incentive
• 14:33
  is
• 14:33
  now
• 14:33
  in
• 14:34
  place:
• 14:34
  the
• 14:34
  more
• 14:35
  you
• 14:35
  chat,
• 14:35
  the
• 14:35
  more
• 14:35
  ad
• 14:36
  inventory
• 14:36
  you
• 14:36
  generate.
• 14:37
  For
• 14:37
  users
• 14:37
  already
• 14:38
  nervous
• 14:38
  about
• 14:38
  data
• 14:39
  sovereignty,
• 14:39
  this
• 14:39
  is
• 14:39
  another
• 14:40
  reason
• 14:40
  to
• 14:40
  run
• 14:40
  local.
• 14:41
  The
• 14:41
  supply
• 14:41
  chain
• 14:42
  angle
• 14:42
  reinforces
• 14:43
  the
• 14:43
  control
• 14:43
  impulse.
• 14:44
  dockerfile-pin
• 14:45
  is
• 14:45
  a
• 14:45
  CLI
• 14:46
  tool
• 14:46
  that
• 14:46
  adds
• 14:46
  @sha256:
• 14:48
  to
• 14:48
  FROM
• 14:48
  lines
• 14:49
  in
• 14:49
  Dockerfiles,
• 14:50
  image
• 14:50
  fields
• 14:51
  in
• 14:51
  docker-compose.yml,
• 14:53
  and
• 14:53
  Docker
• 14:54
  image
• 14:54
  references
• 14:54
  in
• 14:54
  GitHub
• 14:55
  Actions
• 14:55
  files.
• 14:56
  The
• 14:56
  motivation
• 14:57
  is
• 14:57
  straightforward:
• 14:58
  tag-based
• 14:58
  references
• 14:59
  can
• 14:59
  be
• 14:59
  subverted
• 15:00
  by
• 15:00
  force-pushing
• 15:00
  tags
• 15:01
  to
• 15:01
  new
• 15:01
  commits;
• 15:02
  only
• 15:02
  digest
• 15:02
  pinning
• 15:03
  is
• 15:03
  immune.
• 15:03
  The
• 15:04
  tool
• 15:04
  uses
• 15:04
  HEAD
• 15:04
  requests
• 15:05
  (not
• 15:05
  pulls)
• 15:05
  to
• 15:05
  resolve
• 15:06
  digests,
• 15:06
  handles
• 15:07
  private
• 15:07
  registries,
• 15:08
  and
• 15:08
  ships
• 15:09
  with
• 15:09
  CI
• 15:09
  integration
• 15:10
  for
• 15:10
  PR
• 15:10
  validation.
• 15:11
  One
• 15:11
  experimenter
• 15:12
  is
• 15:12
  building
• 15:12
  a
• 15:12
  DIY
• 15:13
  Mythos-at-home
• 15:14
  system
• 15:14
  with
• 15:14
  two
• 15:14
  local
• 15:15
  agents
• 15:15
  --
• 15:15
  an
• 15:16
  uncensored
• 15:16
  Qwen3.5-27B
• 15:20
  scanner
• 15:20
  and
• 15:20
  a
• 15:20
  GLM-5.1
• 15:21
  orchestrator
• 15:21
  --
• 15:22
  hunting
• 15:23
  through
• 15:23
  the
• 15:23
  OpenBSD
• 15:23
  source
• 15:24
  tree
• 15:24
  for
• 15:24
  vulnerabilities.
• 15:25
  The
• 15:25
  thesis:
• 15:26
  Big
• 15:26
  Tech
• 15:26
  thinks
• 15:27
  zero-day
• 15:27
  research
• 15:28
  requires
• 15:28
  clearance
• 15:29
  and
• 15:29
  NDAs;
• 15:30
  this
• 15:30
  project
• 15:30
  thinks
• 15:31
  it
• 15:31
  requires
• 15:31
  open
• 15:31
  models
• 15:32
  and
• 15:32
  an
• 15:32
  ACME
• 15:33
  rocket
• 15:33
  strapped
• 15:34
  to
• 15:34
  a
• 15:34
  coyote.
• 15:35
  Charcuterie
• 15:35
  rounds
• 15:36
  out
• 15:36
  the
• 15:36
  toolkit:
• 15:37
  a
• 15:37
  browser-based
• 15:38
  visual
• 15:38
  similarity
• 15:38
  explorer
• 15:39
  for
• 15:39
  Unicode
• 15:40
  that
• 15:40
  embeds
• 15:40
  rendered
• 15:41
  glyphs
• 15:41
  with
• 15:41
  neural
• 15:42
  networks
• 15:42
  and
• 15:43
  compares
• 15:43
  them
• 15:43
  in
• 15:43
  vector
• 15:44
  space,
• 15:45
  useful
• 15:45
  for
• 15:45
  spotting
• 15:45
  homoglyph
• 15:46
  attacks
• 15:46
  and
• 15:47
  confusable
• 15:47
  characters
• 15:48
  in
• 15:48
  security-sensitive
• 15:49
  contexts.