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

Transcript

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

• 0:00
  The
• 0:00
  White
• 0:01
  House
• 0:01
  is
• 0:01
  drafting
• 0:01
  an
• 0:01
  executive
• 0:02
  order
• 0:02
  that
• 0:02
  would
• 0:02
  direct
• 0:03
  every
• 0:03
  federal
• 0:03
  agency
• 0:04
  to
• 0:04
  remove
• 0:04
  Anthropic's
• 0:05
  AI
• 0:05
  systems
• 0:06
  from
• 0:06
  their
• 0:06
  operations,
• 0:07
  formalizing
• 0:07
  what
• 0:07
  began
• 0:08
  as
• 0:08
  a
• 0:08
  Truth
• 0:08
  Social
• 0:08
  post
• 0:09
  into
• 0:09
  binding
• 0:09
  policy.
• 0:10
  The
• 0:10
  move
• 0:11
  follows
• 0:11
  the
• 0:11
  Pentagon's
• 0:12
  extraordinary
• 0:12
  decision
• 0:13
  to
• 0:13
  label
• 0:13
  Anthropic
• 0:14
  a
• 0:14
  "supply
• 0:14
  chain
• 0:14
  risk
• 0:15
  to
• 0:15
  national
• 0:15
  security"
• 0:15
  —
• 0:16
  a
• 0:16
  designation
• 0:17
  historically
• 0:18
  reserved
• 0:18
  for
• 0:18
  foreign
• 0:18
  adversaries
• 0:19
  and
• 0:19
  compromised
• 0:20
  vendors,
• 0:20
  not
• 0:21
  domestic
• 0:21
  AI
• 0:21
  labs
• 0:22
  that
• 0:22
  disagree
• 0:22
  with
• 0:22
  the
• 0:22
  government
• 0:23
  about
• 0:23
  use-case
• 0:23
  boundaries.
• 0:24
  The
• 0:24
  General
• 0:24
  Services
• 0:25
  Administration
• 0:25
  has
• 0:25
  already
• 0:25
  terminated
• 0:26
  Anthropic's
• 0:26
  OneGov
• 0:27
  deal,
• 0:28
  cutting
• 0:28
  the
• 0:28
  company
• 0:28
  off
• 0:28
  from
• 0:28
  pre-negotiated
• 0:29
  procurement
• 0:30
  channels
• 0:30
  across
• 0:30
  all
• 0:30
  three
• 0:31
  branches
• 0:31
  of
• 0:31
  government.
• 0:32
  The
• 0:32
  substantive
• 0:33
  policy
• 0:33
  shift
• 0:33
  here
• 0:33
  goes
• 0:34
  well
• 0:34
  beyond
• 0:34
  one
• 0:34
  company.
• 0:36
  Draft
• 0:36
  GSA
• 0:36
  guidelines
• 0:37
  would
• 0:37
  require
• 0:37
  any
• 0:38
  AI
• 0:38
  vendor
• 0:38
  seeking
• 0:39
  federal
• 0:39
  business
• 0:40
  to
• 0:40
  grant
• 0:40
  the
• 0:40
  U.S.
• 0:41
  an
• 0:41
  "irrevocable
• 0:41
  license"
• 0:42
  for
• 0:42
  "any
• 0:43
  lawful"
• 0:43
  use
• 0:43
  of
• 0:43
  its
• 0:44
  systems
• 0:44
  and
• 0:45
  to
• 0:45
  certify
• 0:45
  that
• 0:45
  they
• 0:45
  have
• 0:45
  not
• 0:46
  "intentionally
• 0:46
  embedded
• 0:47
  partisan
• 0:47
  or
• 0:48
  ideological
• 0:48
  judgments
• 0:49
  in
• 0:49
  model
• 0:49
  outputs."
• 0:50
  That
• 0:50
  language
• 0:51
  is
• 0:51
  aimed
• 0:51
  squarely
• 0:52
  at
• 0:52
  companies
• 0:52
  maintaining
• 0:53
  binding
• 0:53
  usage
• 0:54
  guardrails
• 0:54
  —
• 0:55
  limits
• 0:55
  on
• 0:55
  deployment
• 0:56
  in
• 0:56
  fully
• 0:56
  autonomous
• 0:57
  weapons
• 0:57
  and
• 0:58
  mass
• 0:58
  domestic
• 0:58
  surveillance,
• 0:59
  the
• 0:59
  exact
• 1:00
  restrictions
• 1:00
  Anthropic
• 1:01
  has
• 1:01
  refused
• 1:01
  to
• 1:01
  relax.
• 1:03
  If
• 1:03
  these
• 1:03
  terms
• 1:03
  become
• 1:04
  standard,
• 1:04
  every
• 1:05
  AI
• 1:05
  company
• 1:05
  selling
• 1:05
  into
• 1:06
  government
• 1:06
  faces
• 1:06
  a
• 1:07
  binary
• 1:07
  choice:
• 1:08
  accept
• 1:08
  unconstrained
• 1:09
  deployment
• 1:09
  or
• 1:10
  walk
• 1:10
  away
• 1:10
  from
• 1:10
  hundreds
• 1:10
  of
• 1:11
  millions
• 1:11
  in
• 1:11
  contracts.
• 1:12
  Anthropic
• 1:13
  fired
• 1:13
  back
• 1:14
  on
• 1:14
  March
• 1:14
  9,
• 1:15
  filing
• 1:15
  lawsuits
• 1:16
  against
• 1:16
  the
• 1:16
  Pentagon
• 1:17
  and
• 1:17
  other
• 1:17
  federal
• 1:17
  officials
• 1:18
  in
• 1:18
  both
• 1:18
  the
• 1:18
  Northern
• 1:18
  District
• 1:19
  of
• 1:19
  California
• 1:20
  and
• 1:20
  the
• 1:20
  D.C.
• 1:20
  Circuit.
• 1:22
  The
• 1:22
  filings
• 1:22
  argue
• 1:22
  the
• 1:22
  "supply
• 1:23
  chain
• 1:23
  risk"
• 1:23
  label
• 1:24
  and
• 1:24
  contract
• 1:24
  terminations
• 1:25
  constitute
• 1:26
  an
• 1:26
  "unlawful
• 1:26
  campaign
• 1:27
  of
• 1:27
  retaliation"
• 1:28
  for
• 1:28
  exercising
• 1:28
  First
• 1:29
  Amendment
• 1:29
  rights
• 1:29
  —
• 1:30
  specifically,
• 1:31
  for
• 1:31
  the
• 1:31
  company's
• 1:32
  speech
• 1:32
  on
• 1:32
  how
• 1:32
  its
• 1:32
  models
• 1:33
  should
• 1:33
  and
• 1:33
  should
• 1:34
  not
• 1:34
  be
• 1:34
  used.
• 1:35
  The
• 1:35
  case
• 1:36
  is
• 1:36
  unprecedented:
• 1:37
  a
• 1:37
  domestic
• 1:38
  AI
• 1:38
  lab
• 1:38
  arguing
• 1:39
  in
• 1:39
  federal
• 1:39
  court
• 1:39
  that
• 1:39
  the
• 1:39
  government
• 1:40
  cannot
• 1:40
  weaponize
• 1:41
  procurement
• 1:41
  to
• 1:41
  coerce
• 1:42
  changes
• 1:42
  to
• 1:42
  safety
• 1:42
  policy.
• 1:44
  Whether
• 1:44
  the
• 1:44
  courts
• 1:44
  agree
• 1:45
  will
• 1:45
  set
• 1:45
  the
• 1:45
  template
• 1:46
  for
• 1:46
  every
• 1:46
  AI
• 1:46
  vendor's
• 1:47
  relationship
• 1:47
  with
• 1:47
  the
• 1:48
  federal
• 1:48
  government
• 1:48
  for
• 1:48
  years
• 1:48
  to
• 1:49
  come.
• 1:52
  A
• 1:52
  sprawling
• 1:52
  investigation
• 1:53
  from
• 1:53
  KrebsOnSecurity
• 1:54
  documents
• 1:54
  just
• 1:54
  how
• 1:54
  quickly
• 1:55
  AI
• 1:55
  coding
• 1:55
  agents
• 1:56
  have
• 1:56
  gone
• 1:56
  from
• 1:56
  experimental
• 1:57
  curiosity
• 1:57
  to
• 1:57
  genuine
• 1:58
  security
• 1:58
  liability.
• 1:59
  The
• 1:59
  centerpiece
• 2:00
  is
• 2:00
  OpenClaw,
• 2:01
  an
• 2:01
  open-source
• 2:02
  autonomous
• 2:02
  AI
• 2:02
  agent
• 2:03
  designed
• 2:03
  to
• 2:03
  run
• 2:03
  locally
• 2:04
  and
• 2:04
  proactively
• 2:05
  take
• 2:05
  actions
• 2:05
  on
• 2:05
  behalf
• 2:06
  of
• 2:06
  its
• 2:06
  operator
• 2:06
  —
• 2:07
  managing
• 2:07
  inboxes,
• 2:08
  executing
• 2:09
  programs,
• 2:09
  browsing
• 2:10
  the
• 2:10
  web,
• 2:10
  and
• 2:10
  integrating
• 2:11
  with
• 2:11
  chat
• 2:11
  platforms.
• 2:12
  The
• 2:12
  problem
• 2:13
  is
• 2:13
  that
• 2:13
  hundreds
• 2:13
  of
• 2:14
  users
• 2:14
  have
• 2:14
  exposed
• 2:15
  their
• 2:15
  OpenClaw
• 2:15
  web
• 2:15
  interfaces
• 2:16
  to
• 2:16
  the
• 2:16
  public
• 2:16
  internet,
• 2:17
  giving
• 2:17
  anyone
• 2:18
  who
• 2:18
  finds
• 2:18
  them
• 2:18
  full
• 2:18
  access
• 2:19
  to
• 2:19
  the
• 2:19
  agent's
• 2:19
  configuration,
• 2:21
  every
• 2:21
  API
• 2:21
  key
• 2:22
  it
• 2:22
  holds,
• 2:22
  and
• 2:22
  its
• 2:23
  complete
• 2:23
  conversation
• 2:23
  history
• 2:24
  across
• 2:24
  all
• 2:25
  integrated
• 2:25
  platforms.
• 2:26
  As
• 2:27
  penetration
• 2:27
  tester
• 2:27
  O'Reilly
• 2:28
  documented,
• 2:29
  an
• 2:29
  attacker
• 2:29
  with
• 2:29
  that
• 2:30
  access
• 2:30
  can
• 2:30
  impersonate
• 2:31
  the
• 2:31
  operator,
• 2:31
  inject
• 2:32
  messages
• 2:32
  into
• 2:32
  ongoing
• 2:33
  conversations,
• 2:34
  and
• 2:34
  exfiltrate
• 2:35
  data
• 2:35
  in
• 2:35
  ways
• 2:35
  that
• 2:35
  look
• 2:36
  like
• 2:36
  normal
• 2:36
  traffic.
• 2:37
  The
• 2:37
  supply
• 2:37
  chain
• 2:38
  angle
• 2:38
  is
• 2:38
  even
• 2:38
  more
• 2:38
  sobering.
• 2:39
  A
• 2:40
  prompt
• 2:40
  injection
• 2:40
  attack
• 2:41
  targeting
• 2:41
  the
• 2:41
  Cline
• 2:41
  AI
• 2:42
  coding
• 2:42
  assistant
• 2:42
  compromised
• 2:43
  its
• 2:43
  GitHub
• 2:43
  Actions
• 2:44
  workflow
• 2:44
  —
• 2:45
  an
• 2:45
  attacker
• 2:46
  opened
• 2:46
  an
• 2:46
  issue
• 2:46
  with
• 2:46
  a
• 2:46
  title
• 2:47
  crafted
• 2:47
  to
• 2:47
  look
• 2:47
  like
• 2:48
  a
• 2:48
  performance
• 2:48
  report
• 2:49
  but
• 2:49
  containing
• 2:49
  embedded
• 2:50
  instructions.
• 2:51
  Because
• 2:51
  the
• 2:51
  workflow
• 2:51
  ran
• 2:52
  Claude
• 2:52
  sessions
• 2:52
  triggered
• 2:53
  by
• 2:53
  issue
• 2:53
  events
• 2:53
  without
• 2:54
  validating
• 2:54
  input,
• 2:55
  the
• 2:55
  malicious
• 2:55
  payload
• 2:56
  propagated
• 2:56
  through
• 2:56
  Cline's
• 2:57
  nightly
• 2:57
  release
• 2:57
  pipeline
• 2:58
  and
• 2:58
  was
• 2:58
  published
• 2:58
  as
• 2:59
  an
• 2:59
  official
• 2:59
  update.
• 3:00
  Thousands
• 3:00
  of
• 3:00
  developers
• 3:01
  ended
• 3:01
  up
• 3:01
  with
• 3:01
  a
• 3:01
  rogue
• 3:01
  OpenClaw
• 3:02
  instance
• 3:02
  installed
• 3:03
  with
• 3:03
  full
• 3:03
  system
• 3:03
  access,
• 3:04
  never
• 3:04
  having
• 3:05
  consented
• 3:05
  to
• 3:05
  it.
• 3:06
  As
• 3:06
  one
• 3:06
  researcher
• 3:07
  put
• 3:07
  it,
• 3:07
  this
• 3:07
  is
• 3:08
  "the
• 3:08
  supply
• 3:08
  chain
• 3:08
  equivalent"
• 3:09
  of
• 3:09
  delegating
• 3:09
  authority
• 3:10
  to
• 3:10
  an
• 3:10
  entirely
• 3:10
  separate
• 3:11
  agent
• 3:11
  the
• 3:11
  developer
• 3:12
  never
• 3:12
  evaluated.
• 3:13
  The
• 3:13
  timing
• 3:14
  of
• 3:14
  this
• 3:14
  investigation
• 3:14
  aligns
• 3:15
  with
• 3:15
  the
• 3:15
  2026
• 3:16
  update
• 3:16
  of
• 3:16
  the
• 3:16
  OWASP
• 3:16
  Top
• 3:17
  10
• 3:17
  for
• 3:17
  LLM
• 3:17
  Applications,
• 3:19
  where
• 3:19
  survey
• 3:19
  results
• 3:20
  reveal
• 3:20
  a
• 3:20
  community
• 3:20
  grappling
• 3:21
  with
• 3:21
  threats
• 3:21
  that
• 3:21
  barely
• 3:22
  existed
• 3:22
  two
• 3:22
  years
• 3:22
  ago.
• 3:23
  Respondents
• 3:24
  overwhelmingly
• 3:25
  support
• 3:25
  keeping
• 3:26
  the
• 3:26
  LLM
• 3:26
  Top
• 3:26
  10
• 3:27
  alongside
• 3:28
  the
• 3:28
  newer
• 3:28
  Agentic
• 3:28
  Top
• 3:29
  10,
• 3:29
  arguing
• 3:30
  the
• 3:30
  two
• 3:30
  lists
• 3:30
  serve
• 3:31
  complementary
• 3:31
  purposes:
• 3:33
  the
• 3:33
  former
• 3:33
  covers
• 3:33
  model-layer
• 3:34
  failures,
• 3:35
  the
• 3:35
  latter
• 3:35
  addresses
• 3:36
  risks
• 3:36
  from
• 3:36
  autonomous
• 3:37
  systems
• 3:37
  that
• 3:37
  plan,
• 3:38
  act,
• 3:38
  and
• 3:39
  use
• 3:39
  tools.
• 3:40
  The
• 3:40
  emerging
• 3:41
  consensus
• 3:41
  is
• 3:42
  that
• 3:42
  the
• 3:42
  most
• 3:42
  dangerous
• 3:43
  vulnerabilities
• 3:43
  now
• 3:44
  live
• 3:44
  at
• 3:44
  the
• 3:44
  intersection
• 3:44
  —
• 3:45
  where
• 3:45
  a
• 3:45
  model-layer
• 3:46
  weakness
• 3:46
  cascades
• 3:47
  into
• 3:47
  agentic
• 3:48
  execution
• 3:48
  failure.
• 3:49
  Practitioners
• 3:50
  are
• 3:50
  calling
• 3:50
  for
• 3:50
  new
• 3:51
  risk
• 3:51
  categories
• 3:51
  around
• 3:52
  cross-context
• 3:52
  contamination,
• 3:54
  memory
• 3:54
  persistence
• 3:54
  poisoning,
• 3:55
  and
• 3:55
  authorization
• 3:56
  state
• 3:56
  erosion
• 3:57
  in
• 3:57
  long-running
• 3:58
  agent
• 3:58
  sessions.
• 3:59
  As
• 3:59
  one
• 4:00
  commenter
• 4:00
  noted,
• 4:01
  "an
• 4:01
  agent
• 4:01
  properly
• 4:02
  scoped
• 4:02
  at
• 4:02
  T0
• 4:03
  may
• 4:03
  be
• 4:03
  operating
• 4:03
  in
• 4:03
  a
• 4:04
  completely
• 4:04
  different
• 4:04
  context
• 4:05
  thirty
• 4:05
  minutes
• 4:05
  later
• 4:06
  —
• 4:06
  and
• 4:06
  nothing
• 4:07
  in
• 4:07
  either
• 4:07
  list
• 4:07
  adequately
• 4:08
  addresses
• 4:08
  that
• 4:08
  erosion."
• 4:10
  The
• 4:10
  defensive
• 4:10
  tooling
• 4:11
  is
• 4:11
  catching
• 4:11
  up,
• 4:11
  but
• 4:12
  unevenly.
• 4:13
  SANS
• 4:13
  has
• 4:13
  published
• 4:13
  its
• 4:14
  Secure
• 4:14
  AI
• 4:14
  Blueprint
• 4:15
  structured
• 4:15
  around
• 4:15
  three
• 4:16
  imperatives
• 4:16
  —
• 4:17
  Protect
• 4:17
  AI,
• 4:18
  Utilize
• 4:18
  AI,
• 4:19
  Govern
• 4:19
  AI
• 4:20
  —
• 4:20
  aiming
• 4:21
  to
• 4:21
  give
• 4:21
  organizations
• 4:21
  a
• 4:22
  framework
• 4:22
  for
• 4:22
  reducing
• 4:22
  exposure
• 4:23
  while
• 4:23
  maintaining
• 4:24
  innovation
• 4:24
  velocity
• 4:24
  .
• 4:25
  On
• 4:25
  the
• 4:26
  containment
• 4:26
  side,
• 4:27
  Clampdown
• 4:27
  takes
• 4:27
  the
• 4:27
  position
• 4:28
  that
• 4:28
  the
• 4:28
  agent
• 4:28
  should
• 4:29
  be
• 4:29
  treated
• 4:29
  as
• 4:29
  an
• 4:29
  untrusted
• 4:30
  process
• 4:30
  regardless
• 4:31
  of
• 4:31
  intent.
• 4:32
  It
• 4:32
  runs
• 4:32
  AI
• 4:33
  coding
• 4:33
  agents
• 4:33
  inside
• 4:34
  hardened
• 4:34
  container
• 4:34
  sandboxes:
• 4:36
  the
• 4:36
  agent
• 4:36
  gets
• 4:36
  zero
• 4:37
  Linux
• 4:37
  capabilities,
• 4:38
  a
• 4:38
  seccomp
• 4:38
  profile
• 4:39
  blocking
• 4:39
  ~115
• 4:40
  syscalls,
• 4:41
  Landlock
• 4:42
  filesystem
• 4:42
  isolation
• 4:43
  enforced
• 4:43
  at
• 4:43
  the
• 4:43
  kernel
• 4:43
  level,
• 4:44
  and
• 4:44
  network
• 4:45
  egress
• 4:45
  controlled
• 4:45
  by
• 4:46
  iptables
• 4:46
  rules
• 4:46
  in
• 4:46
  a
• 4:46
  namespace
• 4:47
  the
• 4:47
  agent
• 4:47
  shares
• 4:48
  but
• 4:48
  cannot
• 4:48
  configure.
• 4:49
  Every
• 4:50
  tool
• 4:50
  container
• 4:50
  the
• 4:50
  agent
• 4:51
  spawns
• 4:51
  goes
• 4:51
  through
• 4:52
  OCI
• 4:52
  hooks
• 4:52
  that
• 4:53
  validate
• 4:53
  security
• 4:53
  policy
• 4:54
  before
• 4:54
  the
• 4:54
  entrypoint
• 4:55
  runs
• 4:55
  —
• 4:55
  there
• 4:56
  is
• 4:56
  no
• 4:56
  opt-out.
• 4:57
  The
• 4:57
  key
• 4:57
  design
• 4:57
  insight
• 4:58
  is
• 4:58
  that
• 4:58
  none
• 4:58
  of
• 4:58
  these
• 4:59
  controls
• 4:59
  depend
• 4:59
  on
• 5:00
  the
• 5:00
  agent
• 5:00
  cooperating;
• 5:01
  a
• 5:01
  fully
• 5:01
  compromised
• 5:02
  agent
• 5:02
  hits
• 5:03
  the
• 5:03
  same
• 5:03
  kernel-enforced
• 5:04
  walls
• 5:04
  as
• 5:04
  one
• 5:04
  following
• 5:05
  instructions.
• 5:06
  Sovereign
• 5:07
  Shield
• 5:07
  takes
• 5:07
  a
• 5:07
  complementary
• 5:08
  approach
• 5:08
  at
• 5:08
  the
• 5:08
  application
• 5:09
  layer,
• 5:09
  offering
• 5:10
  a
• 5:10
  four-layer
• 5:10
  defense
• 5:11
  stack
• 5:11
  —
• 5:11
  firewall,
• 5:12
  input
• 5:12
  filter,
• 5:13
  conscience
• 5:13
  engine,
• 5:13
  and
• 5:14
  core
• 5:14
  safety
• 5:14
  —
• 5:15
  where
• 5:15
  each
• 5:15
  component
• 5:16
  is
• 5:16
  hash-sealed
• 5:16
  and
• 5:16
  independently
• 5:17
  verifiable.
• 5:18
  It
• 5:18
  catches
• 5:19
  prompt
• 5:19
  injection
• 5:19
  via
• 5:20
  keyword
• 5:20
  and
• 5:20
  token
• 5:21
  blocking,
• 5:21
  detects
• 5:22
  deception
• 5:22
  patterns
• 5:23
  with
• 5:23
  pre-compiled
• 5:24
  regex,
• 5:24
  and
• 5:25
  uses
• 5:25
  a
• 5:25
  FrozenNamespace
• 5:26
  metaclass
• 5:27
  that
• 5:27
  makes
• 5:27
  security
• 5:27
  laws
• 5:28
  physically
• 5:28
  immutable
• 5:28
  in
• 5:29
  Python's
• 5:29
  memory.
• 5:30
  The
• 5:30
  system
• 5:30
  is
• 5:30
  deterministic:
• 5:31
  same
• 5:32
  input,
• 5:32
  same
• 5:33
  decision,
• 5:33
  100%
• 5:34
  of
• 5:34
  the
• 5:34
  time,
• 5:35
  all
• 5:35
  enforced
• 5:35
  before
• 5:36
  the
• 5:36
  LLM
• 5:36
  executes
• 5:37
  anything.
• 5:40
  Rivian's
• 5:40
  open-source
• 5:41
  AI-SAST
• 5:44
  brings
• 5:44
  a
• 5:44
  pragmatic
• 5:45
  design
• 5:45
  to
• 5:45
  AI-powered
• 5:46
  static
• 5:46
  analysis:
• 5:47
  it
• 5:47
  runs
• 5:48
  in
• 5:48
  your
• 5:48
  repository
• 5:49
  on
• 5:49
  your
• 5:49
  runners,
• 5:49
  your
• 5:49
  code
• 5:49
  never
• 5:50
  touches
• 5:50
  external
• 5:50
  infrastructure,
• 5:51
  and
• 5:52
  it
• 5:52
  supports
• 5:52
  a
• 5:52
  feedback
• 5:53
  loop
• 5:53
  where
• 5:53
  developers
• 5:53
  mark
• 5:54
  findings
• 5:54
  as
• 5:54
  true
• 5:55
  or
• 5:55
  false
• 5:55
  positives
• 5:56
  directly
• 5:56
  in
• 5:57
  PR
• 5:57
  comments.
• 5:58
  That
• 5:58
  feedback
• 5:58
  is
• 5:59
  stored
• 5:59
  and
• 5:59
  included
• 6:00
  in
• 6:00
  future
• 6:00
  scan
• 6:00
  prompts,
• 6:01
  so
• 6:01
  the
• 6:01
  model
• 6:02
  learns
• 6:02
  to
• 6:02
  avoid
• 6:02
  repeating
• 6:03
  false
• 6:03
  positives
• 6:03
  while
• 6:04
  paying
• 6:04
  closer
• 6:04
  attention
• 6:04
  to
• 6:05
  confirmed
• 6:05
  vulnerability
• 6:06
  patterns.
• 6:06
  The
• 6:07
  architecture
• 6:07
  uses
• 6:07
  a
• 6:07
  dual-LLM
• 6:08
  approach
• 6:08
  —
• 6:09
  Vertex
• 6:09
  AI
• 6:09
  (Gemini)
• 6:10
  for
• 6:10
  initial
• 6:10
  scanning,
• 6:11
  AWS
• 6:11
  Bedrock
• 6:12
  (Claude)
• 6:12
  for
• 6:12
  validation
• 6:12
  —
• 6:13
  and
• 6:13
  only
• 6:13
  validated
• 6:14
  true
• 6:14
  positives
• 6:15
  are
• 6:15
  posted
• 6:15
  in
• 6:15
  the
• 6:15
  PR.
• 6:16
  CVSS
• 6:17
  scoring,
• 6:17
  configurable
• 6:18
  severity
• 6:18
  thresholds,
• 6:19
  and
• 6:19
  webhook
• 6:20
  notifications
• 6:20
  round
• 6:21
  out
• 6:21
  a
• 6:21
  tool
• 6:21
  that
• 6:21
  sits
• 6:22
  comfortably
• 6:22
  in
• 6:22
  existing
• 6:23
  CI/CD
• 6:24
  pipelines.
• 6:25
  For
• 6:25
  practitioners
• 6:26
  who
• 6:26
  need
• 6:26
  working
• 6:26
  exploit
• 6:26
  code
• 6:27
  rather
• 6:27
  than
• 6:27
  advisory
• 6:28
  descriptions,
• 6:29
  the
• 6:29
  ecosystem
• 6:29
  of
• 6:29
  CVE
• 6:30
  proof-of-concept
• 6:31
  repositories
• 6:31
  continues
• 6:32
  to
• 6:32
  mature.
• 6:33
  The
• 6:33
  0xMarcio/cve
• 6:36
  collection
• 6:37
  tracks
• 6:37
  recently
• 6:38
  updated
• 6:38
  PoC
• 6:38
  repos
• 6:39
  organized
• 6:39
  by
• 6:39
  year,
• 6:40
  with
• 6:40
  highlights
• 6:41
  including
• 6:41
  CVE-2026-21858
• 6:43
  (an
• 6:45
  unauthenticated
• 6:46
  file
• 6:46
  read
• 6:46
  to
• 6:46
  RCE
• 6:47
  chain
• 6:47
  in
• 6:47
  n8n
• 6:47
  scoring
• 6:48
  CVSS
• 6:48
  10.0),
• 6:51
  CVE-2026-24061
• 6:53
  with
• 6:53
  multiple
• 6:54
  independent
• 6:54
  PoCs
• 6:55
  from
• 6:55
  SafeBreach
• 6:55
  Labs,
• 6:56
  and
• 6:56
  the
• 6:56
  still-active
• 6:57
  CVE-2025-55182
• 7:01
  React
• 7:01
  Server
• 7:01
  Components
• 7:02
  RCE
• 7:02
  that
• 7:02
  has
• 7:02
  accumulated
• 7:03
  over
• 7:03
  1,300
• 7:04
  stars
• 7:04
  .
• 7:05
  Trickest's
• 7:06
  automated
• 7:07
  CVE-to-PoC
• 7:07
  pipeline
• 7:08
  uses
• 7:08
  GitHub
• 7:09
  search
• 7:09
  and
• 7:09
  reference
• 7:09
  crawling
• 7:10
  to
• 7:10
  discover
• 7:10
  and
• 7:11
  catalog
• 7:11
  exploits,
• 7:12
  generating
• 7:12
  markdown
• 7:13
  files
• 7:13
  with
• 7:13
  version
• 7:14
  badges
• 7:14
  —
• 7:15
  a
• 7:15
  useful
• 7:15
  complement
• 7:16
  when
• 7:16
  you
• 7:16
  need
• 7:16
  breadth
• 7:16
  rather
• 7:17
  than
• 7:17
  depth
• 7:17
  .
• 7:18
  And
• 7:18
  Exploit-DB
• 7:19
  remains
• 7:19
  the
• 7:19
  canonical
• 7:20
  archive
• 7:20
  of
• 7:20
  verified,
• 7:21
  working
• 7:21
  exploit
• 7:22
  code,
• 7:22
  distinguished
• 7:23
  from
• 7:23
  NVD
• 7:24
  and
• 7:24
  vendor
• 7:24
  bulletins
• 7:25
  by
• 7:25
  its
• 7:25
  emphasis
• 7:25
  on
• 7:25
  actionable
• 7:26
  payloads
• 7:27
  rather
• 7:27
  than
• 7:27
  descriptions,
• 7:28
  with
• 7:28
  its
• 7:28
  SearchSploit
• 7:29
  CLI
• 7:29
  companion
• 7:30
  useful
• 7:30
  for
• 7:30
  air-gapped
• 7:31
  environments
• 7:31
  .
• 7:32
  On
• 7:32
  the
• 7:32
  intelligence
• 7:33
  side,
• 7:34
  IPInfoLite
• 7:34
  is
• 7:34
  a
• 7:34
  PowerShell
• 7:35
  module
• 7:35
  for
• 7:35
  the
• 7:36
  IPinfo
• 7:36
  Lite
• 7:36
  API
• 7:37
  that
• 7:37
  caught
• 7:38
  attention
• 7:38
  for
• 7:38
  its
• 7:38
  LLM-optimized
• 7:39
  JSONL
• 7:40
  export
• 7:40
  —
• 7:41
  batch-query
• 7:41
  IP
• 7:42
  addresses
• 7:42
  from
• 7:42
  your
• 7:42
  firewall
• 7:43
  logs,
• 7:44
  export
• 7:44
  structured
• 7:44
  data,
• 7:45
  upload
• 7:45
  to
• 7:46
  Claude
• 7:46
  or
• 7:46
  ChatGPT,
• 7:47
  and
• 7:48
  run
• 7:48
  natural-language
• 7:49
  threat
• 7:49
  analysis
• 7:49
  without
• 7:50
  writing
• 7:50
  custom
• 7:50
  queries.
• 7:51
  Version
• 7:52
  3.1.0
• 7:54
  adds
• 7:54
  an
• 7:54
  SPDX
• 7:55
  2.3
• 7:55
  SBOM
• 7:56
  for
• 7:56
  supply
• 7:57
  chain
• 7:57
  transparency
• 7:58
  and
• 7:58
  migrates
• 7:59
  API
• 7:59
  authentication
• 8:00
  from
• 8:00
  URL
• 8:00
  query
• 8:01
  parameters
• 8:01
  to
• 8:01
  HTTP
• 8:02
  Bearer
• 8:03
  token
• 8:03
  headers,
• 8:04
  eliminating
• 8:04
  credential
• 8:05
  exposure
• 8:05
  through
• 8:05
  proxy
• 8:06
  logs.
• 8:09
  The
• 8:10
  race
• 8:10
  to
• 8:10
  build
• 8:10
  the
• 8:10
  definitive
• 8:11
  agent
• 8:11
  runtime
• 8:12
  is
• 8:12
  producing
• 8:12
  increasingly
• 8:13
  ambitious
• 8:13
  entrants.
• 8:14
  OpenFang
• 8:15
  bills
• 8:15
  itself
• 8:15
  as
• 8:15
  an
• 8:16
  "Agent
• 8:16
  Operating
• 8:16
  System"
• 8:17
  —
• 8:17
  not
• 8:18
  a
• 8:18
  chatbot
• 8:18
  framework
• 8:19
  but
• 8:19
  a
• 8:19
  full
• 8:19
  operating
• 8:19
  system
• 8:20
  for
• 8:20
  autonomous
• 8:20
  agents,
• 8:21
  compiled
• 8:22
  to
• 8:22
  a
• 8:22
  single
• 8:22
  ~32MB
• 8:23
  Rust
• 8:23
  binary
• 8:24
  with
• 8:24
  137K
• 8:25
  lines
• 8:25
  of
• 8:25
  code
• 8:26
  across
• 8:26
  14
• 8:26
  crates.
• 8:28
  Its
• 8:28
  core
• 8:28
  innovation
• 8:29
  is
• 8:29
  "Hands,"
• 8:29
  pre-built
• 8:30
  autonomous
• 8:30
  capability
• 8:31
  packages
• 8:31
  that
• 8:31
  run
• 8:31
  independently
• 8:32
  on
• 8:32
  schedules
• 8:32
  without
• 8:32
  prompting:
• 8:34
  a
• 8:34
  Researcher
• 8:34
  Hand
• 8:34
  that
• 8:34
  wakes
• 8:35
  up,
• 8:35
  cross-references
• 8:35
  sources
• 8:36
  using
• 8:36
  CRAAP
• 8:36
  criteria,
• 8:37
  and
• 8:37
  delivers
• 8:38
  cited
• 8:38
  reports;
• 8:39
  a
• 8:39
  Collector
• 8:39
  for
• 8:39
  OSINT-grade
• 8:40
  continuous
• 8:40
  monitoring
• 8:41
  with
• 8:41
  change
• 8:41
  detection
• 8:41
  and
• 8:42
  knowledge
• 8:42
  graph
• 8:42
  construction;
• 8:43
  a
• 8:43
  Predictor
• 8:44
  that
• 8:44
  builds
• 8:44
  calibrated
• 8:45
  reasoning
• 8:45
  chains
• 8:45
  with
• 8:45
  Brier
• 8:46
  score
• 8:46
  tracking.
• 8:47
  The
• 8:47
  security
• 8:47
  model
• 8:48
  is
• 8:48
  notably
• 8:48
  comprehensive
• 8:49
  —
• 8:49
  16
• 8:50
  discrete
• 8:50
  systems
• 8:50
  including
• 8:51
  WASM
• 8:51
  dual-metered
• 8:51
  sandboxing,
• 8:52
  Merkle
• 8:52
  hash-chain
• 8:53
  audit
• 8:53
  trails,
• 8:54
  Ed25519
• 8:55
  signed
• 8:55
  agent
• 8:55
  manifests,
• 8:56
  information
• 8:57
  flow
• 8:57
  taint
• 8:57
  tracking,
• 8:58
  and
• 8:58
  GCRA
• 8:58
  rate
• 8:58
  limiting.
• 9:00
  Whether
• 9:00
  the
• 9:00
  claims
• 9:00
  hold
• 9:00
  up
• 9:01
  under
• 9:01
  real-world
• 9:01
  use
• 9:01
  remains
• 9:02
  to
• 9:02
  be
• 9:02
  tested,
• 9:02
  but
• 9:02
  the
• 9:03
  architecture
• 9:03
  is
• 9:03
  serious.
• 9:05
  Anthropic's
• 9:05
  Claude
• 9:05
  Agent
• 9:06
  SDK,
• 9:06
  meanwhile,
• 9:07
  turns
• 9:07
  the
• 9:07
  full
• 9:07
  Claude
• 9:08
  Code
• 9:08
  machinery
• 9:08
  into
• 9:09
  a
• 9:09
  Python
• 9:09
  library,
• 9:10
  letting
• 9:10
  developers
• 9:10
  embed
• 9:11
  Claude
• 9:11
  Code
• 9:11
  directly
• 9:12
  in
• 9:12
  their
• 9:12
  own
• 9:12
  applications
• 9:13
  rather
• 9:13
  than
• 9:13
  just
• 9:13
  running
• 9:14
  it
• 9:14
  at
• 9:14
  the
• 9:14
  terminal
• 9:14
  or
• 9:15
  headless
• 9:15
  in
• 9:15
  CI.
• 9:16
  The
• 9:16
  tenth
• 9:17
  installment
• 9:17
  in
• 9:17
  the
• 9:17
  Claude
• 9:17
  Code
• 9:18
  Deep
• 9:18
  Dive
• 9:18
  series
• 9:18
  walks
• 9:18
  through
• 9:19
  the
• 9:19
  SDK's
• 9:19
  architecture
• 9:20
  and
• 9:20
  builds
• 9:20
  a
• 9:20
  practical
• 9:21
  CLI
• 9:21
  tool,
• 9:22
  noting
• 9:22
  that
• 9:22
  the
• 9:22
  real
• 9:23
  leverage
• 9:23
  comes
• 9:23
  from
• 9:23
  treating
• 9:23
  the
• 9:24
  agent
• 9:24
  as
• 9:24
  a
• 9:24
  programmable
• 9:25
  component
• 9:25
  rather
• 9:25
  than
• 9:26
  an
• 9:26
  interactive
• 9:26
  assistant.
• 9:27
  Open
• 9:27
  WebUI's
• 9:28
  0.8.9
• 9:29
  release
• 9:29
  delivers
• 9:29
  a
• 9:30
  substantial
• 9:30
  quality-of-life
• 9:30
  upgrade:
• 9:31
  file
• 9:32
  uploads
• 9:32
  to
• 9:32
  the
• 9:32
  Pyodide
• 9:33
  code
• 9:33
  interpreter,
• 9:34
  a
• 9:34
  dozen
• 9:34
  rendering
• 9:35
  optimizations
• 9:35
  for
• 9:35
  Markdown
• 9:36
  and
• 9:36
  KaTeX,
• 9:37
  and
• 9:37
  a
• 9:37
  radically
• 9:37
  expanded
• 9:38
  Open
• 9:38
  Terminal
• 9:39
  with
• 9:39
  XLSX
• 9:39
  highlighting,
• 9:40
  Jupyter
• 9:41
  notebook
• 9:41
  per-cell
• 9:41
  execution,
• 9:42
  SQLite
• 9:43
  browsing,
• 9:43
  and
• 9:43
  previews
• 9:44
  for
• 9:44
  video,
• 9:45
  audio,
• 9:45
  DOCX,
• 9:46
  and
• 9:46
  HTML.
• 9:47
  The
• 9:48
  performance
• 9:48
  improvements
• 9:48
  are
• 9:49
  meaningful
• 9:49
  —
• 9:50
  streaming
• 9:50
  feels
• 9:50
  smoother,
• 9:51
  page
• 9:51
  loads
• 9:51
  are
• 9:51
  faster,
• 9:52
  and
• 9:52
  the
• 9:52
  DB
• 9:53
  migration
• 9:53
  that
• 9:53
  blocked
• 9:54
  many
• 9:54
  0.8.0
• 9:55
  upgrades
• 9:56
  has
• 9:56
  been
• 9:56
  optimized
• 9:57
  for
• 9:57
  memory
• 9:57
  usage.
• 9:58
  The
• 9:58
  community
• 9:58
  response
• 9:59
  has
• 9:59
  been
• 9:59
  enthusiastic,
• 10:00
  particularly
• 10:01
  around
• 10:01
  pairing
• 10:02
  the
• 10:02
  new
• 10:02
  terminal
• 10:02
  with
• 10:02
  models
• 10:03
  like
• 10:03
  Qwen
• 10:03
  3.5
• 10:04
  27B
• 10:04
  for
• 10:05
  in-browser
• 10:05
  vibe
• 10:05
  coding.
• 10:06
  Turbo
• 10:06
  Flow
• 10:06
  4.0.0
• 10:08
  focuses
• 10:09
  on
• 10:09
  agentic
• 10:09
  workflow
• 10:09
  orchestration,
• 10:11
  though
• 10:11
  release
• 10:11
  notes
• 10:11
  emphasize
• 10:12
  platform
• 10:12
  engineering
• 10:13
  and
• 10:13
  adaptive
• 10:13
  automation
• 10:14
  rather
• 10:14
  than
• 10:14
  novel
• 10:15
  architecture
• 10:15
  .
• 10:16
  And
• 10:16
  Vibium
• 10:16
  positions
• 10:17
  itself
• 10:17
  as
• 10:17
  browser
• 10:18
  automation
• 10:18
  for
• 10:18
  AI
• 10:19
  agents
• 10:19
  and
• 10:19
  humans,
• 10:20
  built
• 10:20
  to
• 10:20
  "bend,
• 10:21
  not
• 10:21
  break,"
• 10:22
  though
• 10:22
  its
• 10:22
  public
• 10:23
  documentation
• 10:23
  is
• 10:23
  still
• 10:24
  sparse
• 10:24
  .
• 10:27
  Andrej
• 10:27
  Karpathy's
• 10:28
  autoresearch
• 10:28
  project
• 10:29
  makes
• 10:29
  a
• 10:29
  quietly
• 10:29
  radical
• 10:30
  proposition:
• 10:31
  give
• 10:31
  an
• 10:31
  AI
• 10:32
  agent
• 10:32
  a
• 10:32
  small
• 10:32
  but
• 10:33
  real
• 10:33
  LLM
• 10:33
  training
• 10:34
  setup,
• 10:34
  let
• 10:35
  it
• 10:35
  modify
• 10:35
  the
• 10:35
  training
• 10:35
  code
• 10:36
  autonomously,
• 10:37
  train
• 10:37
  for
• 10:37
  five
• 10:37
  minutes,
• 10:38
  check
• 10:38
  if
• 10:38
  the
• 10:38
  result
• 10:39
  improved,
• 10:39
  keep
• 10:39
  or
• 10:40
  discard,
• 10:40
  and
• 10:40
  repeat
• 10:41
  overnight.
• 10:42
  You
• 10:42
  wake
• 10:42
  up
• 10:42
  to
• 10:42
  a
• 10:42
  log
• 10:43
  of
• 10:43
  experiments
• 10:44
  and
• 10:44
  —
• 10:44
  hopefully
• 10:44
  —
• 10:45
  a
• 10:45
  better
• 10:46
  model.
• 10:47
  The
• 10:47
  setup
• 10:47
  is
• 10:47
  deliberately
• 10:48
  minimal:
• 10:48
  three
• 10:49
  files,
• 10:49
  one
• 10:50
  GPU,
• 10:51
  one
• 10:51
  metric
• 10:51
  (validation
• 10:52
  bits
• 10:52
  per
• 10:52
  byte),
• 10:53
  and
• 10:53
  a
• 10:53
  fixed
• 10:54
  five-minute
• 10:54
  time
• 10:54
  budget
• 10:55
  per
• 10:55
  experiment,
• 10:56
  yielding
• 10:56
  roughly
• 10:56
  100
• 10:57
  experiments
• 10:57
  while
• 10:57
  you
• 10:58
  sleep.
• 10:59
  The
• 10:59
  human's
• 10:59
  job
• 11:00
  shifts
• 11:00
  from
• 11:00
  writing
• 11:00
  training
• 11:01
  code
• 11:01
  to
• 11:01
  writing
• 11:01
  program.md
• 11:02
  —
• 11:03
  a
• 11:03
  markdown
• 11:04
  file
• 11:04
  that
• 11:04
  serves
• 11:05
  as
• 11:05
  the
• 11:05
  agent's
• 11:05
  research
• 11:06
  program,
• 11:06
  essentially
• 11:07
  "programming
• 11:08
  the
• 11:08
  research
• 11:08
  org"
• 11:09
  rather
• 11:09
  than
• 11:10
  programming
• 11:10
  the
• 11:10
  model.
• 11:11
  Karpathy's
• 11:12
  framing
• 11:12
  is
• 11:12
  characteristically
• 11:13
  provocative:
• 11:14
  "Research
• 11:15
  is
• 11:15
  now
• 11:15
  entirely
• 11:16
  the
• 11:16
  domain
• 11:16
  of
• 11:16
  autonomous
• 11:17
  swarms
• 11:17
  of
• 11:17
  AI
• 11:18
  agents
• 11:18
  running
• 11:18
  across
• 11:19
  compute
• 11:19
  cluster
• 11:20
  megastructures
• 11:20
  in
• 11:20
  the
• 11:20
  skies...
• 11:21
  This
• 11:22
  repo
• 11:22
  is
• 11:22
  the
• 11:22
  story
• 11:23
  of
• 11:23
  how
• 11:23
  it
• 11:23
  all
• 11:23
  began."
• 11:24
  Whether
• 11:24
  that
• 11:25
  is
• 11:25
  prophecy
• 11:25
  or
• 11:25
  performance
• 11:26
  art,
• 11:27
  the
• 11:27
  setup
• 11:27
  itself
• 11:28
  is
• 11:28
  interesting
• 11:28
  —
• 11:29
  a
• 11:29
  concrete,
• 11:30
  reproducible
• 11:30
  experiment
• 11:31
  in
• 11:31
  autonomous
• 11:32
  AI-driven
• 11:32
  research
• 11:33
  where
• 11:33
  the
• 11:33
  agent
• 11:33
  has
• 11:34
  genuine
• 11:34
  freedom
• 11:34
  to
• 11:34
  modify
• 11:35
  architecture,
• 11:36
  hyperparameters,
• 11:37
  optimizer,
• 11:37
  and
• 11:38
  batch
• 11:38
  size.
• 11:39
  Two
• 11:39
  new
• 11:39
  papers
• 11:40
  from
• 11:40
  Sakana
• 11:40
  AI
• 11:40
  —
• 11:41
  Doc-to-LoRA
• 11:42
  and
• 11:42
  Text-to-LoRA
• 11:42
  —
• 11:43
  attack
• 11:43
  the
• 11:43
  fine-tuning
• 11:44
  problem
• 11:44
  from
• 11:44
  complementary
• 11:45
  angles.
• 11:46
  Doc-to-LoRA
• 11:47
  converts
• 11:47
  documents
• 11:47
  into
• 11:48
  LoRA
• 11:48
  adapters
• 11:49
  via
• 11:49
  a
• 11:49
  hypernetwork
• 11:50
  in
• 11:50
  a
• 11:50
  single
• 11:50
  forward
• 11:50
  pass:
• 11:51
  feed
• 11:52
  in
• 11:52
  a
• 11:52
  policy
• 11:52
  document,
• 11:53
  textbook,
• 11:54
  or
• 11:54
  even
• 11:54
  visual
• 11:54
  data,
• 11:55
  and
• 11:55
  the
• 11:56
  system
• 11:56
  generates
• 11:56
  weight
• 11:56
  updates
• 11:57
  that
• 11:57
  encode
• 11:57
  the
• 11:57
  document's
• 11:58
  facts
• 11:58
  directly
• 11:59
  in
• 11:59
  the
• 11:59
  model.
• 12:00
  No
• 12:00
  gradient
• 12:00
  training
• 12:01
  at
• 12:01
  deployment
• 12:01
  time,
• 12:02
  no
• 12:02
  repeated
• 12:02
  document
• 12:03
  re-reading,
• 12:03
  no
• 12:03
  context
• 12:04
  window
• 12:04
  dependency.
• 12:05
  Text-to-LoRA
• 12:06
  does
• 12:06
  the
• 12:06
  same
• 12:06
  for
• 12:06
  behaviors:
• 12:07
  describe
• 12:08
  a
• 12:08
  task
• 12:08
  in
• 12:08
  natural
• 12:09
  language
• 12:09
  ("solve
• 12:10
  math
• 12:10
  problems
• 12:10
  step
• 12:11
  by
• 12:11
  step,"
• 12:12
  "extract
• 12:12
  entities
• 12:13
  from
• 12:13
  medical
• 12:13
  reports")
• 12:14
  and
• 12:14
  the
• 12:14
  hypernetwork
• 12:15
  generates
• 12:15
  an
• 12:15
  adapter
• 12:16
  that
• 12:16
  modifies
• 12:16
  the
• 12:16
  model's
• 12:17
  behavior
• 12:17
  accordingly.
• 12:19
  Together,
• 12:19
  they
• 12:20
  sketch
• 12:20
  a
• 12:20
  continual
• 12:20
  learning
• 12:21
  architecture
• 12:21
  where
• 12:22
  a
• 12:22
  model
• 12:22
  can
• 12:22
  acquire
• 12:23
  new
• 12:23
  knowledge
• 12:23
  with
• 12:23
  D2L,
• 12:24
  new
• 12:24
  abilities
• 12:25
  with
• 12:25
  T2L,
• 12:26
  and
• 12:26
  use
• 12:26
  both
• 12:27
  simultaneously.
• 12:28
  For
• 12:28
  anyone
• 12:29
  who
• 12:29
  has
• 12:29
  built
• 12:29
  RAG
• 12:29
  pipelines
• 12:30
  —
• 12:30
  retrieve,
• 12:31
  chunk,
• 12:31
  stuff
• 12:32
  into
• 12:32
  context,
• 12:33
  pray
• 12:33
  the
• 12:33
  model
• 12:33
  finds
• 12:33
  the
• 12:33
  right
• 12:34
  piece
• 12:34
  —
• 12:35
  the
• 12:35
  promise
• 12:35
  of
• 12:35
  internalizing
• 12:36
  documents
• 12:36
  directly
• 12:37
  in
• 12:37
  weights
• 12:37
  is
• 12:37
  worth
• 12:37
  watching
• 12:38
  closely.
• 12:41
  The
• 12:41
  distinction
• 12:42
  between
• 12:42
  Large
• 12:42
  Language
• 12:42
  Models
• 12:43
  and
• 12:43
  Large
• 12:43
  Reasoning
• 12:44
  Models
• 12:44
  is
• 12:44
  becoming
• 12:44
  practically
• 12:45
  important
• 12:45
  as
• 12:45
  teams
• 12:46
  choose
• 12:46
  which
• 12:46
  capabilities
• 12:47
  justify
• 12:47
  which
• 12:47
  inference
• 12:48
  costs.
• 12:49
  LLMs
• 12:49
  predict
• 12:50
  the
• 12:50
  next
• 12:50
  token
• 12:50
  via
• 12:51
  statistical
• 12:51
  pattern
• 12:51
  matching
• 12:52
  —
• 12:52
  fast,
• 12:53
  cheap,
• 12:53
  and
• 12:54
  usually
• 12:54
  sufficient
• 12:54
  for
• 12:54
  tasks
• 12:55
  like
• 12:55
  drafting
• 12:55
  social
• 12:56
  media
• 12:56
  posts.
• 12:57
  LRMs
• 12:57
  add
• 12:58
  an
• 12:58
  internal
• 12:58
  chain
• 12:58
  of
• 12:58
  thought:
• 12:59
  they
• 12:59
  sketch
• 13:00
  plans,
• 13:00
  weigh
• 13:01
  options,
• 13:01
  double-check
• 13:02
  calculations
• 13:02
  in
• 13:03
  a
• 13:03
  sandbox
• 13:03
  before
• 13:03
  committing
• 13:04
  to
• 13:04
  a
• 13:04
  response.
• 13:05
  The
• 13:05
  training
• 13:05
  pipeline
• 13:06
  adds
• 13:06
  specialized
• 13:07
  reasoning-focused
• 13:07
  fine-tuning
• 13:08
  on
• 13:08
  curated
• 13:08
  datasets
• 13:09
  of
• 13:09
  logic
• 13:09
  puzzles,
• 13:10
  multi-step
• 13:11
  math,
• 13:11
  and
• 13:11
  tricky
• 13:11
  coding
• 13:12
  tasks,
• 13:13
  followed
• 13:13
  by
• 13:13
  reinforcement
• 13:14
  learning
• 13:14
  (RLHF
• 13:14
  or
• 13:14
  process
• 13:15
  reward
• 13:15
  models
• 13:16
  that
• 13:16
  judge
• 13:16
  each
• 13:16
  step
• 13:16
  independently).
• 13:18
  The
• 13:18
  trade-off
• 13:18
  is
• 13:18
  straightforward:
• 13:19
  deeper
• 13:19
  reasoning
• 13:20
  buys
• 13:20
  you
• 13:20
  better
• 13:20
  accuracy
• 13:21
  on
• 13:21
  complex
• 13:22
  problems
• 13:22
  at
• 13:22
  the
• 13:22
  cost
• 13:22
  of
• 13:22
  higher
• 13:23
  latency,
• 13:24
  more
• 13:24
  VRAM,
• 13:25
  and
• 13:25
  a
• 13:25
  bigger
• 13:25
  invoice
• 13:25
  from
• 13:26
  your
• 13:26
  cloud
• 13:26
  provider.
• 13:27
  The
• 13:27
  practical
• 13:27
  guidance
• 13:28
  is
• 13:28
  that
• 13:28
  if
• 13:28
  the
• 13:28
  question
• 13:29
  requires
• 13:29
  multi-step
• 13:29
  logic,
• 13:30
  planning,
• 13:30
  or
• 13:30
  abstract
• 13:31
  reasoning,
• 13:31
  an
• 13:32
  LRM
• 13:32
  earns
• 13:32
  its
• 13:32
  cost.
• 13:33
  For
• 13:33
  everything
• 13:34
  else,
• 13:34
  a
• 13:34
  standard
• 13:35
  LLM
• 13:35
  —
• 13:36
  or
• 13:36
  a
• 13:36
  smaller
• 13:36
  model
• 13:37
  —
• 13:37
  is
• 13:38
  likely
• 13:38
  the
• 13:38
  right
• 13:38
  call.
• 13:39
  For
• 13:39
  those
• 13:40
  running
• 13:40
  models
• 13:40
  locally,
• 13:41
  the
• 13:41
  choice
• 13:41
  between
• 13:41
  serving
• 13:42
  backends
• 13:42
  remains
• 13:42
  contentious.
• 13:44
  The
• 13:44
  llama-swap
• 13:44
  community
• 13:45
  argues
• 13:45
  for
• 13:45
  its
• 13:46
  approach
• 13:46
  over
• 13:46
  Ollama
• 13:47
  and
• 13:47
  LM
• 13:47
  Studio,
• 13:48
  while
• 13:48
  practitioners
• 13:49
  report
• 13:49
  wildly
• 13:49
  different
• 13:50
  experiences
• 13:50
  depending
• 13:51
  on
• 13:51
  hardware.
• 13:51
  vLLM
• 13:52
  excels
• 13:53
  at
• 13:53
  concurrent
• 13:53
  requests
• 13:53
  via
• 13:54
  continuous
• 13:54
  batching
• 13:55
  and
• 13:55
  supports
• 13:55
  multi-token
• 13:56
  prediction
• 13:56
  for
• 13:56
  speed
• 13:56
  gains,
• 13:57
  but
• 13:57
  multiple
• 13:58
  users
• 13:58
  report
• 13:58
  terrible
• 13:59
  performance
• 13:59
  on
• 13:59
  AMD
• 14:00
  hardware
• 14:00
  (8
• 14:01
  tokens/second
• 14:01
  vs.
• 14:03
  llama.cpp's
• 14:03
  50
• 14:04
  for
• 14:04
  the
• 14:04
  same
• 14:04
  model).
• 14:05
  The
• 14:05
  pragmatic
• 14:06
  consensus:
• 14:07
  if
• 14:07
  your
• 14:07
  model
• 14:07
  fits
• 14:07
  in
• 14:07
  VRAM,
• 14:08
  try
• 14:08
  vLLM
• 14:08
  for
• 14:09
  throughput;
• 14:09
  if
• 14:10
  you
• 14:10
  are
• 14:10
  on
• 14:10
  AMD
• 14:10
  or
• 14:10
  need
• 14:11
  simplicity,
• 14:11
  llama.cpp
• 14:13
  remains
• 14:14
  the
• 14:14
  safer
• 14:14
  bet.
• 14:15
  A
• 14:15
  Veritasium
• 14:16
  deep
• 14:16
  dive
• 14:16
  on
• 14:16
  Markov
• 14:17
  chains
• 14:17
  traces
• 14:18
  the
• 14:18
  intellectual
• 14:19
  lineage
• 14:19
  from
• 14:19
  a
• 14:19
  1905
• 14:20
  Russian
• 14:20
  math
• 14:20
  feud
• 14:21
  to
• 14:21
  modern
• 14:21
  language
• 14:21
  models.
• 14:23
  Andrey
• 14:23
  Markov
• 14:23
  built
• 14:24
  his
• 14:24
  dependent-event
• 14:24
  probability
• 14:25
  framework
• 14:26
  to
• 14:26
  refute
• 14:26
  Pavel
• 14:26
  Nekrasov's
• 14:27
  claim
• 14:27
  that
• 14:27
  the
• 14:27
  law
• 14:28
  of
• 14:28
  large
• 14:28
  numbers
• 14:28
  implied
• 14:29
  free
• 14:29
  will
• 14:29
  —
• 14:30
  and
• 14:30
  he
• 14:30
  proved
• 14:30
  his
• 14:30
  point
• 14:31
  by
• 14:31
  analyzing
• 14:32
  vowel-consonant
• 14:32
  transitions
• 14:33
  in
• 14:33
  Pushkin's
• 14:34
  "Eugene
• 14:34
  Onegin."
• 14:35
  The
• 14:35
  chain
• 14:35
  of
• 14:36
  applications
• 14:36
  from
• 14:36
  there
• 14:37
  is
• 14:37
  remarkable:
• 14:38
  Stanislaw
• 14:39
  Ulam's
• 14:39
  Monte
• 14:39
  Carlo
• 14:40
  method
• 14:40
  for
• 14:40
  nuclear
• 14:41
  weapons
• 14:41
  simulations,
• 14:42
  Brin
• 14:42
  and
• 14:42
  Page's
• 14:43
  PageRank
• 14:43
  algorithm
• 14:44
  that
• 14:44
  made
• 14:44
  Google
• 14:44
  a
• 14:44
  trillion-dollar
• 14:45
  company,
• 14:46
  and
• 14:46
  Claude
• 14:46
  Shannon's
• 14:47
  text
• 14:47
  prediction
• 14:47
  experiments
• 14:48
  that
• 14:48
  prefigured
• 14:49
  today's
• 14:49
  LLMs.
• 14:50
  The
• 14:50
  video
• 14:51
  makes
• 14:51
  the
• 14:51
  useful
• 14:51
  pedagogical
• 14:52
  point
• 14:52
  that
• 14:52
  modern
• 14:53
  language
• 14:53
  models
• 14:53
  extend
• 14:54
  the
• 14:54
  Markov
• 14:54
  chain
• 14:55
  concept
• 14:55
  with
• 14:55
  attention
• 14:56
  mechanisms
• 14:56
  —
• 14:57
  the
• 14:57
  model
• 14:57
  learns
• 14:58
  what
• 14:58
  to
• 14:58
  pay
• 14:58
  attention
• 14:59
  to
• 14:59
  from
• 14:59
  previous
• 15:00
  context
• 15:00
  —
• 15:01
  but
• 15:01
  warns
• 15:01
  that
• 15:02
  feedback
• 15:02
  loops
• 15:02
  (models
• 15:03
  trained
• 15:03
  on
• 15:03
  model-generated
• 15:04
  text)
• 15:05
  lead
• 15:05
  to
• 15:05
  convergence
• 15:06
  on
• 15:06
  "a
• 15:06
  very
• 15:06
  dull,
• 15:07
  stable
• 15:07
  state"
• 15:08
  that
• 15:08
  "just
• 15:08
  says
• 15:08
  the
• 15:09
  same
• 15:09
  thing
• 15:09
  over
• 15:09
  and
• 15:10
  over."
• 15:11
  The
• 15:11
  "Speeding
• 15:12
  One
• 15:12
  Cog
• 15:12
  Breaks
• 15:13
  the
• 15:13
  Machine"
• 15:13
  essay
• 15:13
  from
• 15:13
  Unhyped
• 15:14
  AI
• 15:14
  delivers
• 15:15
  the
• 15:15
  enterprise
• 15:15
  reality
• 15:16
  check
• 15:16
  that
• 15:16
  every
• 15:17
  AI
• 15:17
  deployment
• 15:17
  needs
• 15:18
  to
• 15:18
  hear.
• 15:19
  The
• 15:19
  argument
• 15:19
  is
• 15:19
  precise:
• 15:21
  most
• 15:21
  enterprise
• 15:21
  work
• 15:22
  is
• 15:22
  not
• 15:22
  a
• 15:22
  single
• 15:23
  task
• 15:23
  but
• 15:23
  a
• 15:23
  workflow,
• 15:24
  and
• 15:24
  workflows
• 15:25
  do
• 15:25
  not
• 15:25
  speed
• 15:26
  up
• 15:26
  just
• 15:26
  because
• 15:26
  one
• 15:26
  step
• 15:27
  was
• 15:27
  automated.
• 15:28
  They
• 15:28
  speed
• 15:29
  up
• 15:29
  when
• 15:29
  handoffs
• 15:29
  get
• 15:30
  cleaner,
• 15:30
  meaning
• 15:31
  survives
• 15:31
  transfer,
• 15:32
  exceptions
• 15:32
  do
• 15:33
  not
• 15:33
  explode,
• 15:34
  and
• 15:34
  the
• 15:34
  cost
• 15:34
  of
• 15:34
  coordination
• 15:35
  falls
• 15:35
  across
• 15:35
  the
• 15:36
  whole
• 15:36
  value
• 15:36
  flow.
• 15:37
  Teams
• 15:38
  that
• 15:38
  accelerate
• 15:38
  one
• 15:38
  node
• 15:39
  without
• 15:39
  making
• 15:39
  translation
• 15:40
  cheaper
• 15:40
  get
• 15:41
  three
• 15:41
  predictable
• 15:41
  failure
• 15:42
  modes:
• 15:43
  the
• 15:43
  fast
• 15:43
  step
• 15:43
  floods
• 15:44
  downstream
• 15:44
  teams
• 15:45
  with
• 15:45
  output
• 15:46
  they
• 15:46
  cannot
• 15:46
  safely
• 15:47
  interpret,
• 15:48
  coordination
• 15:48
  overhead
• 15:49
  explodes
• 15:49
  into
• 15:49
  the
• 15:49
  seams
• 15:50
  (meetings,
• 15:51
  Slack
• 15:51
  threads,
• 15:52
  "quick
• 15:52
  alignment"),
• 15:53
  and
• 15:53
  errors
• 15:54
  propagate
• 15:54
  further
• 15:54
  before
• 15:55
  anyone
• 15:55
  notices.
• 15:56
  The
• 15:56
  economic
• 15:57
  unit
• 15:57
  is
• 15:57
  end-to-end
• 15:58
  cycle
• 15:58
  time
• 15:58
  under
• 15:59
  constraint,
• 15:59
  not
• 16:00
  task
• 16:00
  speed
• 16:00
  —
• 16:01
  and
• 16:01
  measuring
• 16:02
  task
• 16:02
  speed
• 16:02
  while
• 16:03
  calling
• 16:03
  it
• 16:03
  transformation
• 16:04
  is
• 16:04
  the
• 16:04
  accounting
• 16:04
  error
• 16:05
  that
• 16:05
  makes
• 16:05
  isolated
• 16:06
  AI
• 16:06
  use
• 16:06
  cases
• 16:06
  disappoint.
• 16:08
  For
• 16:08
  agentic
• 16:08
  systems
• 16:09
  specifically,
• 16:10
  the
• 16:10
  stakes
• 16:10
  compound:
• 16:11
  agents
• 16:11
  do
• 16:12
  not
• 16:12
  just
• 16:12
  generate,
• 16:13
  they
• 16:13
  execute,
• 16:14
  so
• 16:14
  translation
• 16:15
  failures
• 16:15
  become
• 16:16
  misfiled
• 16:16
  claims,
• 16:17
  misrouted
• 16:18
  approvals,
• 16:18
  and
• 16:18
  policy
• 16:19
  drift
• 16:19
  landing
• 16:19
  in
• 16:20
  the
• 16:20
  lap
• 16:20
  of
• 16:20
  the
• 16:20
  least
• 16:20
  empowered
• 16:21
  person
• 16:21
  in
• 16:21
  the
• 16:21
  chain.