Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 11, 2026

Transcript

AGI Dreams Podcast – March 11, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  CodeWall's
• 0:00
  autonomous
• 0:01
  offensive
• 0:01
  agent
• 0:02
  needed
• 0:02
  no
• 0:02
  credentials,
• 0:03
  no
• 0:03
  insider
• 0:04
  knowledge,
• 0:05
  and
• 0:05
  no
• 0:05
  human
• 0:05
  guidance
• 0:06
  to
• 0:06
  crack
• 0:06
  McKinsey's
• 0:07
  AI
• 0:07
  platform
• 0:07
  Lilli
• 0:08
  —
• 0:08
  the
• 0:09
  internal
• 0:09
  chatbot
• 0:09
  used
• 0:10
  by
• 0:10
  72
• 0:10
  percent
• 0:11
  of
• 0:11
  the
• 0:11
  firm's
• 0:11
  43,000-plus
• 0:13
  consultants,
• 0:14
  processing
• 0:15
  over
• 0:15
  500,000
• 0:16
  prompts
• 0:16
  per
• 0:16
  month.
• 0:17
  Within
• 0:17
  two
• 0:18
  hours
• 0:18
  the
• 0:18
  agent
• 0:19
  had
• 0:19
  full
• 0:19
  read-write
• 0:20
  access
• 0:20
  to
• 0:20
  the
• 0:20
  production
• 0:21
  database.
• 0:22
  The
• 0:22
  haul:
• 0:23
  46.5
• 0:24
  million
• 0:24
  plaintext
• 0:25
  chat
• 0:25
  messages
• 0:25
  covering
• 0:26
  strategy,
• 0:26
  M&A,
• 0:27
  and
• 0:27
  client
• 0:27
  engagements;
• 0:29
  728,000
• 0:30
  confidential
• 0:31
  files
• 0:31
  (192,000
• 0:32
  PDFs,
• 0:33
  93,000
• 0:34
  Excel
• 0:34
  spreadsheets,
• 0:34
  93,000
• 0:35
  PowerPoint
• 0:36
  decks);
• 0:36
  57,000
• 0:37
  user
• 0:37
  accounts;
• 0:38
  and
• 0:38
  95
• 0:39
  system
• 0:39
  prompts
• 0:39
  that
• 0:39
  governed
• 0:40
  Lilli's
• 0:40
  behavior
• 0:40
  —
• 0:41
  all
• 0:41
  writable.
• 0:42
  The
• 0:42
  attack
• 0:43
  chain
• 0:43
  was
• 0:43
  not
• 0:43
  exotic.
• 0:44
  The
• 0:44
  agent
• 0:45
  found
• 0:45
  publicly
• 0:45
  exposed
• 0:46
  API
• 0:46
  documentation
• 0:46
  —
• 0:47
  over
• 0:48
  200
• 0:48
  endpoints,
• 0:48
  fully
• 0:49
  documented
• 0:49
  —
• 0:50
  and
• 0:50
  identified
• 0:51
  22
• 0:51
  that
• 0:51
  required
• 0:52
  no
• 0:52
  authentication.
• 0:54
  One
• 0:54
  of
• 0:54
  those
• 0:54
  unprotected
• 0:54
  endpoints
• 0:55
  wrote
• 0:55
  user
• 0:55
  search
• 0:55
  queries
• 0:56
  to
• 0:56
  the
• 0:56
  database
• 0:56
  with
• 0:56
  parameterized
• 0:57
  values,
• 0:58
  but
• 0:58
  the
• 0:58
  JSON
• 0:59
  field
• 0:59
  names
• 0:59
  were
• 0:59
  concatenated
• 1:00
  directly
• 1:00
  into
• 1:00
  SQL.
• 1:02
  Standard
• 1:02
  tools,
• 1:03
  including
• 1:03
  OWASP
• 1:04
  ZAP,
• 1:04
  missed
• 1:05
  it
• 1:05
  entirely.
• 1:06
  The
• 1:06
  agent
• 1:06
  recognized
• 1:07
  JSON
• 1:07
  keys
• 1:08
  reflected
• 1:08
  verbatim
• 1:08
  in
• 1:09
  database
• 1:09
  error
• 1:09
  messages,
• 1:10
  ran
• 1:10
  fifteen
• 1:11
  blind
• 1:11
  iterations
• 1:12
  to
• 1:12
  map
• 1:12
  the
• 1:12
  query
• 1:12
  shape,
• 1:13
  and
• 1:13
  then
• 1:13
  live
• 1:13
  production
• 1:14
  data
• 1:14
  started
• 1:14
  flowing
• 1:14
  back.
• 1:15
  Because
• 1:16
  the
• 1:16
  injection
• 1:16
  was
• 1:16
  read-write,
• 1:17
  an
• 1:17
  attacker
• 1:17
  could
• 1:18
  have
• 1:18
  silently
• 1:18
  rewritten
• 1:19
  Lilli's
• 1:19
  system
• 1:19
  prompts
• 1:19
  —
• 1:20
  poisoning
• 1:21
  how
• 1:21
  the
• 1:21
  chatbot
• 1:21
  answered
• 1:21
  every
• 1:22
  consultant's
• 1:22
  query,
• 1:23
  stripping
• 1:23
  guardrails,
• 1:24
  or
• 1:24
  embedding
• 1:24
  data-exfiltration
• 1:25
  instructions
• 1:25
  —
• 1:26
  with
• 1:26
  a
• 1:26
  single
• 1:27
  UPDATE
• 1:27
  wrapped
• 1:27
  in
• 1:27
  one
• 1:28
  HTTP
• 1:28
  call.
• 1:29
  No
• 1:29
  deployment,
• 1:30
  no
• 1:30
  code
• 1:30
  change,
• 1:31
  no
• 1:31
  log
• 1:31
  trail.
• 1:32
  McKinsey
• 1:33
  patched
• 1:33
  all
• 1:33
  unauthenticated
• 1:34
  endpoints
• 1:35
  within
• 1:35
  hours
• 1:35
  of
• 1:35
  disclosure
• 1:36
  and
• 1:36
  says
• 1:36
  forensic
• 1:37
  investigation
• 1:38
  found
• 1:38
  no
• 1:38
  evidence
• 1:38
  of
• 1:39
  unauthorized
• 1:39
  access
• 1:40
  by
• 1:40
  third
• 1:40
  parties.
• 1:41
  The
• 1:41
  McKinsey
• 1:42
  breach
• 1:42
  crystallizes
• 1:42
  a
• 1:43
  pattern
• 1:43
  that
• 1:43
  keeps
• 1:43
  repeating:
• 1:44
  AI
• 1:45
  platforms
• 1:45
  inherit
• 1:45
  classic
• 1:46
  web
• 1:46
  vulnerabilities
• 1:47
  because
• 1:47
  the
• 1:47
  teams
• 1:48
  building
• 1:48
  them
• 1:48
  treat
• 1:48
  the
• 1:48
  prompt
• 1:49
  layer
• 1:49
  as
• 1:49
  configuration,
• 1:50
  not
• 1:50
  code.
• 1:51
  System
• 1:51
  prompts
• 1:51
  —
• 1:52
  the
• 1:52
  instructions
• 1:53
  controlling
• 1:53
  how
• 1:54
  an
• 1:54
  AI
• 1:54
  behaves
• 1:54
  —
• 1:55
  sit
• 1:55
  in
• 1:55
  the
• 1:55
  same
• 1:56
  databases
• 1:56
  as
• 1:57
  user
• 1:57
  data,
• 1:57
  rarely
• 1:58
  with
• 1:58
  access
• 1:58
  controls,
• 1:59
  version
• 1:59
  history,
• 2:00
  or
• 2:00
  integrity
• 2:01
  monitoring.
• 2:02
  As
• 2:02
  CodeWall's
• 2:02
  CEO
• 2:03
  put
• 2:03
  it,
• 2:03
  hackers
• 2:04
  will
• 2:04
  use
• 2:04
  the
• 2:04
  same
• 2:04
  autonomous
• 2:05
  agent
• 2:05
  strategies
• 2:06
  "to
• 2:06
  attack
• 2:06
  indiscriminately,
• 2:07
  with
• 2:07
  a
• 2:07
  specific
• 2:08
  objective
• 2:08
  in
• 2:08
  mind."
• 2:09
  Meanwhile,
• 2:10
  the
• 2:10
  offensive
• 2:11
  recon
• 2:11
  tooling
• 2:11
  keeps
• 2:12
  getting
• 2:12
  cheaper.
• 2:13
  Red
• 2:13
  Amon's
• 2:13
  latest
• 2:14
  Wave
• 2:14
  Runner
• 2:14
  feature
• 2:15
  eliminates
• 2:15
  redundant
• 2:16
  LLM
• 2:17
  reasoning
• 2:17
  steps
• 2:18
  by
• 2:18
  parallelizing
• 2:19
  independent
• 2:19
  tool
• 2:19
  calls
• 2:20
  within
• 2:20
  attack
• 2:20
  chains
• 2:20
  —
• 2:21
  grouping
• 2:22
  curl,
• 2:23
  httpx,
• 2:24
  and
• 2:24
  katana
• 2:24
  into
• 2:25
  a
• 2:25
  single
• 2:25
  wave,
• 2:26
  then
• 2:26
  analyzing
• 2:26
  combined
• 2:27
  results
• 2:27
  in
• 2:27
  one
• 2:28
  pass.
• 2:29
  The
• 2:29
  result:
• 2:30
  roughly
• 2:30
  35
• 2:31
  percent
• 2:31
  fewer
• 2:31
  agent
• 2:32
  steps
• 2:32
  and
• 2:32
  30
• 2:33
  percent
• 2:33
  lower
• 2:33
  token
• 2:34
  cost
• 2:34
  per
• 2:34
  attack
• 2:34
  chain.
• 2:35
  DryRun
• 2:36
  Security's
• 2:36
  agentic
• 2:37
  coding
• 2:37
  security
• 2:38
  report
• 2:38
  makes
• 2:38
  the
• 2:38
  case
• 2:39
  that
• 2:39
  the
• 2:39
  attack
• 2:39
  surface
• 2:39
  is
• 2:40
  not
• 2:40
  just
• 2:40
  external.
• 2:42
  Their
• 2:42
  Contextual
• 2:42
  Security
• 2:43
  Analysis
• 2:43
  engine
• 2:44
  reasons
• 2:44
  about
• 2:44
  exploitability
• 2:45
  and
• 2:45
  impact
• 2:45
  across
• 2:46
  data
• 2:46
  flows,
• 2:47
  architecture,
• 2:48
  and
• 2:48
  change
• 2:48
  history
• 2:48
  —
• 2:49
  catching
• 2:49
  logic
• 2:50
  flaws
• 2:50
  and
• 2:50
  broken
• 2:51
  authorization
• 2:51
  that
• 2:52
  pattern-matching
• 2:52
  scanners
• 2:53
  miss.
• 2:54
  The
• 2:54
  pitch
• 2:54
  resonates
• 2:55
  because
• 2:55
  AI-generated
• 2:56
  code
• 2:56
  is
• 2:56
  now
• 2:56
  the
• 2:57
  majority
• 2:57
  source
• 2:57
  in
• 2:58
  many
• 2:58
  engineering
• 2:58
  teams,
• 2:59
  and
• 2:59
  the
• 2:59
  errors
• 3:00
  it
• 3:00
  produces
• 3:01
  have
• 3:01
  evolved
• 3:01
  from
• 3:01
  syntax
• 3:02
  bugs
• 3:02
  into
• 3:02
  architectural
• 3:03
  landmines
• 3:04
  that
• 3:04
  detonate
• 3:04
  several
• 3:05
  pull
• 3:05
  requests
• 3:05
  deep.
• 3:06
  One
• 3:07
  customer
• 3:07
  building
• 3:07
  AI-driven
• 3:08
  shopping
• 3:08
  experiences
• 3:09
  chose
• 3:09
  DryRun
• 3:10
  specifically
• 3:10
  because
• 3:11
  "OWASP
• 3:11
  LLM
• 3:12
  app
• 3:12
  risks
• 3:12
  are
• 3:13
  all
• 3:13
  about
• 3:13
  context."
• 3:14
  When
• 3:14
  every
• 3:15
  PR
• 3:15
  potentially
• 3:16
  carries
• 3:16
  AI-introduced
• 3:17
  vulnerabilities,
• 3:18
  contextual
• 3:18
  analysis
• 3:19
  —
• 3:20
  not
• 3:20
  regex
• 3:20
  —
• 3:21
  is
• 3:21
  what
• 3:21
  separates
• 3:22
  signal
• 3:22
  from
• 3:22
  noise.
• 3:28
  Bynario
• 3:28
  asks
• 3:28
  a
• 3:29
  question
• 3:29
  most
• 3:29
  vulnerability
• 3:30
  scanners
• 3:30
  cannot
• 3:30
  answer:
• 3:32
  is
• 3:32
  this
• 3:32
  flaw
• 3:32
  actually
• 3:32
  triggerable
• 3:33
  in
• 3:33
  your
• 3:33
  environment?
• 3:34
  The
• 3:34
  startup,
• 3:35
  founded
• 3:35
  by
• 3:35
  researchers
• 3:35
  who
• 3:36
  have
• 3:36
  found
• 3:36
  bugs
• 3:36
  in
• 3:36
  everything
• 3:37
  from
• 3:37
  kernels
• 3:37
  to
• 3:37
  browsers,
• 3:38
  directly
• 3:39
  analyzes
• 3:39
  compiled
• 3:40
  software
• 3:40
  —
• 3:41
  the
• 3:41
  executable
• 3:41
  binaries
• 3:42
  running
• 3:42
  in
• 3:42
  production,
• 3:43
  not
• 3:43
  the
• 3:43
  source
• 3:43
  code
• 3:44
  that
• 3:44
  may
• 3:44
  or
• 3:44
  may
• 3:44
  not
• 3:44
  match
• 3:44
  what
• 3:45
  shipped.
• 3:46
  Their
• 3:46
  autonomous
• 3:46
  pipeline
• 3:47
  identifies
• 3:47
  both
• 3:47
  known
• 3:48
  CVEs
• 3:48
  and
• 3:48
  previously
• 3:49
  undisclosed
• 3:49
  issues,
• 3:50
  then
• 3:51
  validates
• 3:51
  whether
• 3:51
  each
• 3:52
  flaw
• 3:52
  can
• 3:52
  be
• 3:52
  exploited
• 3:53
  under
• 3:53
  your
• 3:53
  specific
• 3:53
  configuration.
• 3:55
  The
• 3:55
  distinction
• 3:55
  matters
• 3:56
  because
• 3:56
  patching
• 3:56
  is
• 3:56
  not
• 3:56
  always
• 3:57
  possible,
• 3:58
  version
• 3:58
  checks
• 3:58
  miss
• 3:59
  undisclosed
• 3:59
  vulnerabilities,
• 4:00
  and
• 4:00
  the
• 4:00
  sheer
• 4:01
  volume
• 4:01
  of
• 4:01
  unprioritized
• 4:02
  scanner
• 4:02
  output
• 4:03
  has
• 4:03
  turned
• 4:03
  triage
• 4:03
  into
• 4:04
  a
• 4:04
  full-time
• 4:04
  job.
• 4:05
  Bynario's
• 4:05
  approach
• 4:06
  inverts
• 4:06
  the
• 4:06
  usual
• 4:06
  workflow:
• 4:07
  instead
• 4:07
  of
• 4:08
  generating
• 4:08
  more
• 4:08
  signals,
• 4:09
  it
• 4:09
  reduces
• 4:09
  the
• 4:09
  pile
• 4:10
  to
• 4:10
  the
• 4:10
  subset
• 4:10
  that
• 4:10
  represents
• 4:11
  real
• 4:11
  exposure.
• 4:13
  They
• 4:13
  demonstrated
• 4:13
  the
• 4:13
  pipeline
• 4:14
  by
• 4:14
  discovering
• 4:14
  multiple
• 4:15
  vulnerabilities
• 4:15
  in
• 4:15
  macOS,
• 4:16
  iOS,
• 4:17
  and
• 4:17
  iPadOS
• 4:17
  —
• 4:19
  closed-source
• 4:19
  ecosystems
• 4:20
  where
• 4:20
  binary
• 4:20
  analysis
• 4:21
  is
• 4:21
  the
• 4:21
  only
• 4:21
  option.
• 4:22
  Skillsmith's
• 4:23
  new
• 4:23
  Dependency
• 4:24
  Intelligence
• 4:24
  layer
• 4:24
  attacks
• 4:25
  a
• 4:25
  different
• 4:25
  blind
• 4:25
  spot
• 4:25
  in
• 4:26
  the
• 4:26
  agentic
• 4:26
  toolchain.
• 4:27
  Skills
• 4:28
  —
• 4:29
  the
• 4:29
  reusable
• 4:30
  markdown
• 4:30
  workflows
• 4:31
  that
• 4:31
  power
• 4:31
  Claude
• 4:31
  Code
• 4:31
  and
• 4:32
  similar
• 4:32
  agent
• 4:32
  platforms
• 4:32
  —
• 4:33
  reference
• 4:34
  MCP
• 4:34
  servers,
• 4:35
  models,
• 4:35
  and
• 4:35
  other
• 4:35
  skills
• 4:36
  throughout
• 4:36
  their
• 4:36
  content,
• 4:37
  but
• 4:37
  until
• 4:37
  now
• 4:37
  that
• 4:38
  information
• 4:38
  was
• 4:38
  invisible
• 4:39
  at
• 4:39
  install
• 4:39
  time.
• 4:40
  A
• 4:40
  governance
• 4:41
  skill
• 4:41
  calls
• 4:41
  mcplinearsave
• 4:42
  issue;
• 4:43
  the
• 4:43
  user
• 4:43
  installs
• 4:44
  it;
• 4:45
  Linear
• 4:45
  is
• 4:45
  not
• 4:45
  configured;
• 4:46
  the
• 4:46
  skill
• 4:47
  fails
• 4:47
  with
• 4:47
  no
• 4:47
  explanation.
• 4:49
  Dependency
• 4:50
  Intelligence
• 4:51
  triangulates
• 4:51
  from
• 4:51
  three
• 4:52
  signal
• 4:52
  sources
• 4:52
  —
• 4:53
  author
• 4:53
  declarations
• 4:54
  with
• 4:54
  typed
• 4:55
  versioning
• 4:55
  (confidence
• 4:56
  1.0),
• 4:57
  static
• 4:57
  content
• 4:58
  analysis
• 4:58
  that
• 4:59
  detects
• 4:59
  mcpservertool
• 5:00
  patterns
• 5:00
  and
• 5:00
  assigns
• 5:01
  confidence
• 5:01
  based
• 5:02
  on
• 5:02
  whether
• 5:02
  references
• 5:03
  appear
• 5:03
  in
• 5:03
  prose
• 5:03
  (0.9)
• 5:04
  or
• 5:04
  code
• 5:04
  blocks
• 5:05
  (0.5),
• 5:06
  and
• 5:06
  behavioral
• 5:07
  co-install
• 5:07
  data
• 5:08
  that
• 5:08
  ramps
• 5:08
  linearly
• 5:09
  as
• 5:09
  usage
• 5:09
  patterns
• 5:10
  emerge.
• 5:11
  Seven
• 5:11
  MCP
• 5:12
  tools
• 5:12
  now
• 5:12
  surface
• 5:13
  dependency
• 5:13
  data
• 5:14
  across
• 5:14
  the
• 5:14
  full
• 5:14
  skill
• 5:15
  lifecycle,
• 5:16
  from
• 5:16
  skill
• 5:16
  validate
• 5:17
  warnings
• 5:17
  before
• 5:17
  publish
• 5:18
  to
• 5:18
  skill
• 5:18
  outdated
• 5:18
  health
• 5:19
  checks
• 5:19
  that
• 5:19
  function
• 5:20
  like
• 5:20
  npm
• 5:20
  outdated
• 5:21
  for
• 5:21
  agent
• 5:21
  workflows.
• 5:22
  Crucially,
• 5:23
  Skillsmith
• 5:24
  chose
• 5:24
  not
• 5:24
  to
• 5:24
  build
• 5:24
  a
• 5:24
  resolver
• 5:25
  —
• 5:26
  hard
• 5:26
  dependencies
• 5:26
  block
• 5:27
  installs
• 5:27
  with
• 5:27
  clear
• 5:28
  messages,
• 5:28
  everything
• 5:29
  else
• 5:29
  is
• 5:29
  advisory.
• 5:31
  Auto-installing
• 5:32
  a
• 5:32
  dependency
• 5:32
  means
• 5:32
  silently
• 5:33
  adding
• 5:33
  instructions
• 5:34
  to
• 5:34
  an
• 5:34
  agent's
• 5:35
  system
• 5:35
  prompt,
• 5:36
  crossing
• 5:36
  a
• 5:36
  trust
• 5:37
  boundary
• 5:37
  they
• 5:37
  are
• 5:37
  not
• 5:37
  comfortable
• 5:38
  crossing
• 5:38
  automatically.
• 5:43
  Dan
• 5:44
  Guido
• 5:44
  opened
• 5:44
  his
• 5:44
  [un]prompted
• 5:45
  talk
• 5:45
  with
• 5:45
  a
• 5:45
  provocation:
• 5:46
  thousands
• 5:47
  of
• 5:47
  CEOs
• 5:47
  surveyed,
• 5:48
  and
• 5:48
  AI
• 5:49
  had
• 5:49
  no
• 5:49
  measurable
• 5:49
  impact
• 5:50
  on
• 5:50
  employment
• 5:51
  or
• 5:51
  productivity.
• 5:52
  Economists
• 5:53
  are
• 5:53
  calling
• 5:53
  it
• 5:53
  the
• 5:53
  new
• 5:53
  Solow
• 5:54
  paradox
• 5:54
  —
• 5:55
  "you
• 5:55
  can
• 5:56
  see
• 5:56
  the
• 5:56
  computer
• 5:56
  age
• 5:56
  everywhere
• 5:57
  but
• 5:57
  in
• 5:57
  the
• 5:57
  productivity
• 5:58
  statistics."
• 5:59
  Guido's
• 5:59
  diagnosis
• 6:00
  is
• 6:00
  blunt:
• 6:01
  most
• 6:01
  companies
• 6:01
  are
• 6:01
  doing
• 6:01
  it
• 6:02
  wrong.
• 6:02
  They
• 6:02
  gave
• 6:02
  people
• 6:03
  tools
• 6:03
  without
• 6:03
  changing
• 6:04
  the
• 6:04
  system.
• 6:05
  Everyone
• 6:05
  gets
• 6:05
  a
• 6:05
  ChatGPT
• 6:06
  license,
• 6:06
  leadership
• 6:07
  waits
• 6:07
  for
• 6:07
  the
• 6:07
  productivity
• 6:08
  numbers
• 6:08
  to
• 6:08
  move,
• 6:09
  and
• 6:09
  they
• 6:09
  don't.
• 6:10
  That
• 6:10
  is
• 6:10
  the
• 6:10
  gap
• 6:10
  between
• 6:11
  AI-assisted
• 6:12
  and
• 6:12
  AI-native
• 6:12
  —
• 6:13
  one
• 6:13
  is
• 6:13
  a
• 6:13
  tool,
• 6:14
  the
• 6:14
  other
• 6:14
  is
• 6:14
  an
• 6:14
  operating
• 6:15
  system.
• 6:16
  Trail
• 6:16
  of
• 6:16
  Bits
• 6:16
  built
• 6:17
  the
• 6:17
  operating
• 6:17
  system.
• 6:18
  The
• 6:18
  security
• 6:18
  consultancy
• 6:19
  standardized
• 6:20
  on
• 6:20
  Claude
• 6:20
  Code,
• 6:21
  wrote
• 6:21
  an
• 6:21
  AI
• 6:21
  Handbook
• 6:22
  that
• 6:22
  removes
• 6:22
  ambiguity
• 6:23
  about
• 6:23
  which
• 6:23
  tools
• 6:24
  are
• 6:24
  approved
• 6:24
  for
• 6:24
  which
• 6:24
  data
• 6:25
  types,
• 6:25
  created
• 6:26
  an
• 6:26
  AI
• 6:26
  Maturity
• 6:26
  Matrix
• 6:27
  that
• 6:27
  makes
• 6:27
  adoption
• 6:28
  a
• 6:28
  first-class
• 6:28
  professional
• 6:29
  capability
• 6:29
  with
• 6:29
  clear
• 6:30
  levels
• 6:30
  and
• 6:30
  consequences
• 6:31
  for
• 6:31
  staying
• 6:31
  stuck,
• 6:32
  ran
• 6:32
  hackathons
• 6:33
  that
• 6:33
  forced
• 6:33
  engineers
• 6:34
  into
• 6:34
  bypass-permissions
• 6:35
  mode
• 6:35
  to
• 6:35
  learn
• 6:36
  real
• 6:36
  sandboxing
• 6:36
  constraints,
• 6:37
  and
• 6:38
  captured
• 6:38
  everything
• 6:39
  as
• 6:39
  reusable
• 6:39
  artifacts
• 6:40
  in
• 6:40
  internal
• 6:40
  and
• 6:41
  external
• 6:41
  skills
• 6:41
  repos.
• 6:42
  The
• 6:43
  results
• 6:43
  after
• 6:43
  roughly
• 6:43
  a
• 6:43
  year:
• 6:44
  94
• 6:44
  plugins
• 6:45
  containing
• 6:45
  201
• 6:46
  skills,
• 6:47
  84
• 6:47
  specialized
• 6:47
  agents,
• 6:48
  and
• 6:48
  over
• 6:48
  400
• 6:49
  reference
• 6:49
  files
• 6:50
  encoding
• 6:50
  domain
• 6:50
  expertise
• 6:51
  across
• 6:51
  the
• 6:51
  full
• 6:51
  consulting
• 6:52
  lifecycle
• 6:52
  —
• 6:53
  from
• 6:53
  vulnerability-class-specific
• 6:55
  analyzers
• 6:55
  to
• 6:55
  sales
• 6:56
  proposal
• 6:56
  generators.
• 6:57
  On
• 6:57
  certain
• 6:58
  engagements,
• 6:58
  bug
• 6:59
  discovery
• 6:59
  went
• 6:59
  from
• 6:59
  15
• 7:00
  per
• 7:00
  week
• 7:00
  to
• 7:00
  200.
• 7:01
  Twenty
• 7:02
  percent
• 7:02
  of
• 7:02
  all
• 7:02
  bugs
• 7:02
  reported
• 7:03
  to
• 7:03
  clients
• 7:03
  are
• 7:03
  now
• 7:04
  initially
• 7:04
  discovered
• 7:04
  by
• 7:04
  AI.
• 7:05
  The
• 7:06
  sales
• 7:06
  team
• 7:06
  averages
• 7:06
  $7–8
• 7:08
  million
• 7:08
  in
• 7:09
  revenue
• 7:09
  per
• 7:09
  representative,
• 7:10
  roughly
• 7:10
  double
• 7:10
  the
• 7:10
  consulting
• 7:11
  industry
• 7:11
  benchmark.
• 7:12
  /slides-with-notes.pdf)
• 7:15
  What
• 7:15
  makes
• 7:15
  Trail
• 7:15
  of
• 7:15
  Bits'
• 7:15
  approach
• 7:16
  replicable
• 7:16
  is
• 7:17
  the
• 7:17
  psychological
• 7:17
  framework
• 7:18
  underneath
• 7:18
  it.
• 7:19
  Guido
• 7:19
  identified
• 7:20
  four
• 7:20
  barriers
• 7:20
  to
• 7:20
  adoption
• 7:21
  —
• 7:21
  self-enhancing
• 7:22
  bias
• 7:22
  (seniors
• 7:23
  trust
• 7:23
  intuition
• 7:23
  that
• 7:23
  got
• 7:24
  them
• 7:24
  there),
• 7:24
  identity
• 7:25
  threat
• 7:25
  (security
• 7:25
  auditing
• 7:26
  is
• 7:26
  deeply
• 7:26
  symbolic
• 7:27
  craft
• 7:27
  work),
• 7:28
  opacity
• 7:28
  (people
• 7:28
  feel
• 7:29
  they
• 7:29
  understand
• 7:29
  human
• 7:29
  judgment
• 7:30
  but
• 7:30
  objectively
• 7:31
  understand
• 7:31
  neither),
• 7:32
  and
• 7:32
  intolerance
• 7:32
  for
• 7:33
  imperfection
• 7:33
  (one
• 7:33
  bad
• 7:34
  AI
• 7:34
  output
• 7:34
  confirms
• 7:35
  every
• 7:35
  skeptic's
• 7:36
  priors).
• 7:37
  The
• 7:37
  remedies
• 7:37
  were
• 7:37
  structural:
• 7:38
  the
• 7:38
  maturity
• 7:39
  matrix
• 7:39
  counters
• 7:39
  self-enhancing
• 7:40
  bias
• 7:40
  by
• 7:41
  making
• 7:41
  the
• 7:41
  ladder
• 7:41
  visible;
• 7:42
  skills
• 7:43
  repos
• 7:43
  reframe
• 7:43
  identity
• 7:44
  from
• 7:44
  "I
• 7:44
  don't
• 7:44
  need
• 7:45
  AI"
• 7:45
  to
• 7:45
  "I'm
• 7:45
  the
• 7:45
  one
• 7:46
  who
• 7:46
  makes
• 7:46
  AI
• 7:46
  dangerous";
• 7:47
  a
• 7:47
  curated
• 7:48
  marketplace
• 7:48
  and
• 7:48
  sandboxing
• 7:49
  tools
• 7:49
  reduce
• 7:50
  embarrassing
• 7:50
  failures;
• 7:51
  the
• 7:51
  CEO
• 7:52
  using
• 7:52
  tools
• 7:52
  visibly
• 7:53
  every
• 7:53
  day
• 7:53
  signals
• 7:54
  permanence
• 7:54
  to
• 7:54
  the
• 7:54
  passive
• 7:55
  50
• 7:55
  percent.
• 7:56
  Open
• 7:56
  questions
• 7:57
  remain
• 7:57
  —
• 7:58
  private
• 7:58
  inference
• 7:59
  gaps,
• 7:59
  prompt
• 8:00
  injection
• 8:00
  on
• 8:00
  client
• 8:01
  code,
• 8:01
  and
• 8:01
  policy
• 8:02
  enforcement
• 8:02
  at
• 8:02
  scale
• 8:03
  —
• 8:04
  but
• 8:04
  the
• 8:04
  system
• 8:04
  is
• 8:04
  compounding
• 8:05
  weekly.
• 8:06
  Simon
• 8:06
  Wardley
• 8:07
  would
• 8:07
  appreciate
• 8:07
  the
• 8:07
  framing.
• 8:08
  His
• 8:09
  argument
• 8:09
  that
• 8:09
  "vibe
• 8:09
  coding"
• 8:10
  grotesquely
• 8:10
  misdescribes
• 8:11
  the
• 8:11
  actual
• 8:12
  work
• 8:12
  —
• 8:13
  which
• 8:13
  is
• 8:13
  managing,
• 8:14
  handling,
• 8:15
  and
• 8:15
  controlling
• 8:15
  systems
• 8:16
  of
• 8:16
  AI
• 8:16
  agents
• 8:16
  —
• 8:17
  maps
• 8:17
  precisely
• 8:18
  to
• 8:18
  what
• 8:18
  Trail
• 8:18
  of
• 8:18
  Bits
• 8:19
  operationalized.
• 8:21
  The
• 8:21
  real
• 8:21
  job
• 8:21
  title
• 8:22
  is
• 8:22
  AI
• 8:22
  wrangler,
• 8:23
  not
• 8:23
  vibe
• 8:23
  coder.
• 8:24
  The
• 8:24
  distinction
• 8:25
  matters
• 8:25
  because
• 8:25
  it
• 8:25
  tells
• 8:26
  engineers
• 8:26
  what
• 8:26
  skills
• 8:27
  they
• 8:27
  need
• 8:27
  and
• 8:27
  tells
• 8:28
  executives
• 8:28
  why
• 8:28
  this
• 8:28
  matters.
• 8:33
  Marc
• 8:33
  Shade
• 8:34
  ran
• 8:34
  BullshitBench
• 8:34
  v2
• 8:35
  against
• 8:35
  his
• 8:35
  production
• 8:35
  AI
• 8:36
  stack
• 8:36
  —
• 8:36
  100
• 8:37
  expert-level
• 8:37
  questions
• 8:37
  loaded
• 8:38
  with
• 8:38
  fabricated
• 8:38
  terminology
• 8:39
  like
• 8:39
  "differential
• 8:39
  indemnity
• 8:40
  decomposition"
• 8:41
  and
• 8:41
  "Causal
• 8:41
  Dependency
• 8:41
  Fingerprinting"
• 8:42
  —
• 8:42
  and
• 8:42
  the
• 8:42
  results
• 8:43
  were
• 8:43
  humbling.
• 8:44
  Raw
• 8:44
  local
• 8:44
  models
• 8:44
  (14B
• 8:45
  parameter):
• 8:45
  zero
• 8:46
  percent
• 8:46
  pushback;
• 8:47
  every
• 8:47
  fake
• 8:47
  framework
• 8:48
  accepted
• 8:48
  as
• 8:48
  real,
• 8:48
  with
• 8:49
  detailed
• 8:49
  implementation
• 8:50
  advice
• 8:50
  generated
• 8:50
  for
• 8:50
  things
• 8:51
  that
• 8:51
  do
• 8:51
  not
• 8:51
  exist.
• 8:52
  Gemini
• 8:52
  CLI
• 8:53
  scored
• 8:53
  86
• 8:53
  percent
• 8:53
  pushback.
• 8:54
  Claude
• 8:55
  with
• 8:55
  configured
• 8:55
  anti-fabrication
• 8:56
  rules
• 8:56
  hit
• 8:56
  80
• 8:57
  percent,
• 8:57
  but
• 8:57
  with
• 8:58
  a
• 8:58
  revealing
• 8:58
  domain
• 8:58
  split:
• 8:59
  perfect
• 9:00
  scores
• 9:00
  on
• 9:00
  medical
• 9:00
  and
• 9:01
  software
• 9:01
  engineering
• 9:01
  questions,
• 9:02
  but
• 9:02
  only
• 9:02
  1
• 9:03
  out
• 9:03
  of
• 9:03
  3
• 9:03
  on
• 9:03
  legal,
• 9:04
  where
• 9:04
  the
• 9:04
  model
• 9:04
  accepted
• 9:05
  "proportional
• 9:06
  fault
• 9:06
  cascade
• 9:06
  analysis"
• 9:07
  as
• 9:07
  a
• 9:07
  real
• 9:08
  M&A
• 9:08
  methodology.
• 9:09
  The
• 9:09
  critical
• 9:10
  finding:
• 9:11
  custom
• 9:11
  system
• 9:11
  prompts
• 9:12
  are
• 9:12
  necessary
• 9:12
  but
• 9:12
  not
• 9:13
  sufficient.
• 9:14
  In
• 9:14
  domains
• 9:14
  where
• 9:15
  fabricated
• 9:15
  jargon
• 9:15
  blends
• 9:16
  with
• 9:16
  real
• 9:16
  niche
• 9:16
  terminology,
• 9:17
  the
• 9:17
  model's
• 9:18
  desire
• 9:18
  to
• 9:18
  be
• 9:18
  helpful
• 9:18
  overwhelms
• 9:19
  the
• 9:19
  instruction.
• 9:20
  And
• 9:21
  variance
• 9:21
  is
• 9:21
  the
• 9:21
  real
• 9:22
  enemy
• 9:22
  —
• 9:23
  same
• 9:23
  model,
• 9:24
  same
• 9:24
  rules,
• 9:24
  same
• 9:25
  questions,
• 9:26
  different
• 9:26
  runs
• 9:26
  produced
• 9:27
  50
• 9:27
  percent
• 9:27
  and
• 9:28
  80
• 9:28
  percent
• 9:28
  pushback
• 9:28
  rates.
• 9:29
  Without
• 9:30
  automated
• 9:30
  evaluation
• 9:31
  infrastructure,
• 9:31
  you
• 9:32
  are
• 9:32
  guessing.
• 9:33
  The
• 9:33
  sycophancy
• 9:33
  problem
• 9:34
  compounds
• 9:34
  when
• 9:34
  agents
• 9:35
  operate
• 9:35
  unsupervised.
• 9:36
  One
• 9:36
  developer
• 9:37
  building
• 9:37
  agents
• 9:37
  that
• 9:37
  write
• 9:38
  code
• 9:38
  overnight
• 9:38
  realized
• 9:38
  he
• 9:39
  had
• 9:39
  no
• 9:39
  reliable
• 9:39
  way
• 9:39
  to
• 9:39
  know
• 9:40
  if
• 9:40
  any
• 9:40
  of
• 9:40
  it
• 9:40
  was
• 9:40
  correct.
• 9:41
  Teams
• 9:42
  using
• 9:42
  Claude
• 9:42
  for
• 9:42
  everyday
• 9:42
  PRs
• 9:43
  are
• 9:43
  merging
• 9:43
  40–50
• 9:45
  a
• 9:45
  week
• 9:45
  instead
• 9:45
  of
• 9:45
  10,
• 9:45
  spending
• 9:46
  dramatically
• 9:46
  more
• 9:46
  time
• 9:47
  in
• 9:47
  code
• 9:47
  review.
• 9:48
  When
• 9:48
  Claude
• 9:48
  writes
• 9:48
  tests
• 9:49
  for
• 9:49
  code
• 9:49
  Claude
• 9:49
  just
• 9:49
  wrote,
• 9:50
  it
• 9:50
  is
• 9:50
  checking
• 9:50
  its
• 9:51
  own
• 9:51
  work
• 9:51
  —
• 9:52
  a
• 9:52
  self-congratulation
• 9:52
  machine
• 9:53
  that
• 9:53
  proves
• 9:53
  code
• 9:53
  does
• 9:53
  what
• 9:53
  Claude
• 9:54
  thought
• 9:54
  you
• 9:54
  wanted,
• 9:55
  not
• 9:55
  what
• 9:55
  you
• 9:55
  actually
• 9:55
  wanted.
• 9:56
  The
• 9:56
  proposed
• 9:57
  fix
• 9:57
  borrows
• 9:57
  from
• 9:57
  TDD:
• 9:58
  write
• 9:58
  acceptance
• 9:59
  criteria
• 9:59
  in
• 9:59
  plain
• 10:00
  English
• 10:00
  before
• 10:00
  prompting,
• 10:02
  let
• 10:02
  the
• 10:02
  agent
• 10:02
  build
• 10:02
  against
• 10:03
  them,
• 10:03
  run
• 10:04
  Playwright-based
• 10:04
  browser
• 10:05
  verification
• 10:05
  against
• 10:05
  each
• 10:06
  criterion,
• 10:07
  review
• 10:07
  only
• 10:07
  the
• 10:07
  failures.
• 10:08
  The
• 10:09
  implementation
• 10:09
  uses
• 10:09
  four
• 10:10
  stages
• 10:10
  —
• 10:11
  a
• 10:11
  zero-LLM
• 10:12
  preflight
• 10:12
  check,
• 10:12
  one
• 10:13
  Opus
• 10:13
  call
• 10:13
  for
• 10:13
  planning,
• 10:14
  parallel
• 10:14
  Sonnet
• 10:15
  calls
• 10:15
  for
• 10:15
  per-criterion
• 10:16
  verification,
• 10:17
  and
• 10:17
  a
• 10:17
  final
• 10:17
  Opus
• 10:18
  call
• 10:18
  for
• 10:18
  verdicts.
• 10:19
  It
• 10:20
  does
• 10:20
  not
• 10:20
  catch
• 10:20
  spec
• 10:20
  misunderstandings,
• 10:22
  but
• 10:22
  it
• 10:22
  catches
• 10:22
  integration
• 10:23
  failures,
• 10:24
  rendering
• 10:24
  bugs,
• 10:25
  and
• 10:25
  behavior
• 10:25
  that
• 10:25
  works
• 10:25
  in
• 10:25
  theory
• 10:26
  but
• 10:26
  breaks
• 10:27
  in
• 10:27
  a
• 10:27
  real
• 10:27
  browser
• 10:27
  —
• 10:28
  which
• 10:29
  is
• 10:29
  more
• 10:29
  than
• 10:29
  code
• 10:29
  review
• 10:30
  was
• 10:30
  reliably
• 10:30
  catching
• 10:31
  anyway.
• 10:32
  The
• 10:32
  economics
• 10:32
  of
• 10:33
  local
• 10:33
  code
• 10:33
  review
• 10:33
  offer
• 10:34
  another
• 10:34
  path.
• 10:35
  One
• 10:35
  engineer
• 10:35
  wired
• 10:36
  Claude
• 10:36
  Code
• 10:36
  to
• 10:36
  a
• 10:37
  fine-tuned
• 10:37
  Qwen3-Coder-Next
• 10:38
  served
• 10:39
  through
• 10:39
  vLLM
• 10:39
  on
• 10:40
  a
• 10:40
  local
• 10:40
  GPU
• 10:40
  devbox
• 10:41
  and
• 10:41
  built
• 10:42
  the
• 10:42
  scaffolding
• 10:42
  most
• 10:43
  people
• 10:43
  ignore:
• 10:44
  an
• 10:44
  orchestrator
• 10:44
  agent
• 10:45
  to
• 10:45
  decompose
• 10:45
  and
• 10:46
  route
• 10:46
  reviews,
• 10:46
  specialist
• 10:47
  agents
• 10:47
  for
• 10:47
  C++/Win32,
• 10:50
  C#/XAML,
• 10:52
  and
• 10:52
  security
• 10:53
  policy,
• 10:54
  a
• 10:54
  synthesis
• 10:54
  pass
• 10:54
  to
• 10:55
  merge
• 10:55
  outputs,
• 10:56
  and
• 10:56
  a
• 10:56
  verifier
• 10:56
  pass
• 10:57
  to
• 10:57
  remove
• 10:57
  duplicate
• 10:58
  findings,
• 10:58
  weak
• 10:59
  claims,
• 10:59
  and
• 10:59
  severity
• 11:00
  inflation.
• 11:01
  Isolated
• 11:01
  git
• 11:01
  worktrees
• 11:02
  per
• 11:02
  run,
• 11:03
  a
• 11:03
  trusted
• 11:03
  context
• 11:04
  manifest,
• 11:05
  deterministic
• 11:06
  bootstrap
• 11:06
  from
• 11:06
  PRs,
• 11:07
  structured
• 11:08
  findings
• 11:08
  schemas,
• 11:09
  and
• 11:09
  a
• 11:09
  tmux
• 11:09
  "swarm
• 11:10
  theater"
• 11:10
  for
• 11:11
  live
• 11:11
  observation.
• 11:12
  The
• 11:12
  punchline:
• 11:13
  "The
• 11:13
  raw
• 11:13
  model
• 11:13
  is
• 11:13
  not
• 11:14
  the
• 11:14
  product.
• 11:15
  The
• 11:15
  system
• 11:15
  around
• 11:15
  it
• 11:15
  is."
• 11:16
  A
• 11:16
  good
• 11:16
  model
• 11:17
  without
• 11:17
  orchestration
• 11:17
  is
• 11:18
  a
• 11:18
  toy.
• 11:18
  A
• 11:18
  good
• 11:19
  model
• 11:19
  with
• 11:19
  professional
• 11:19
  scaffolding
• 11:20
  behaves
• 11:20
  like
• 11:20
  a
• 11:20
  system.
• 11:25
  The
• 11:25
  case
• 11:26
  for
• 11:26
  replacing
• 11:26
  cloud
• 11:27
  voice
• 11:27
  assistants
• 11:27
  with
• 11:27
  local
• 11:28
  alternatives
• 11:28
  just
• 11:28
  got
• 11:29
  a
• 11:29
  lot
• 11:29
  simpler.
• 11:30
  Home
• 11:30
  Assistant's
• 11:30
  Voice
• 11:31
  Preview
• 11:31
  Edition
• 11:31
  —
• 11:32
  a
• 11:32
  $70
• 11:32
  open-source
• 11:33
  hardware
• 11:33
  puck
• 11:34
  with
• 11:34
  a
• 11:34
  dual
• 11:34
  mic
• 11:34
  array,
• 11:34
  physical
• 11:35
  mute
• 11:35
  switch,
• 11:35
  and
• 11:36
  XMOS
• 11:36
  audio
• 11:36
  chip
• 11:37
  for
• 11:37
  local
• 11:37
  noise
• 11:37
  cancellation
• 11:38
  —
• 11:38
  makes
• 11:38
  switching
• 11:39
  as
• 11:39
  easy
• 11:39
  as
• 11:39
  plugging
• 11:40
  in
• 11:40
  an
• 11:40
  Echo.
• 11:40
  The
• 11:41
  voice
• 11:41
  pipeline
• 11:41
  runs
• 11:41
  entirely
• 11:42
  on-network:
• 11:43
  Whisper
• 11:43
  STT
• 11:44
  converts
• 11:44
  speech
• 11:44
  to
• 11:44
  text
• 11:45
  (under
• 11:45
  200ms
• 11:46
  on
• 11:46
  an
• 11:46
  Intel
• 11:46
  N100),
• 11:47
  Piper
• 11:47
  TTS
• 11:48
  produces
• 11:48
  surprisingly
• 11:49
  human
• 11:49
  neural
• 11:49
  speech,
• 11:50
  and
• 11:50
  the
• 11:50
  Wyoming
• 11:50
  protocol
• 11:51
  connects
• 11:51
  satellites
• 11:51
  to
• 11:52
  the
• 11:52
  Home
• 11:52
  Assistant
• 11:52
  server.
• 11:53
  With
• 11:53
  local
• 11:54
  LLM
• 11:54
  integration
• 11:55
  via
• 11:55
  Ollama,
• 11:56
  you
• 11:56
  can
• 11:56
  say
• 11:56
  "I'm
• 11:56
  heading
• 11:57
  to
• 11:57
  bed"
• 11:57
  and
• 11:57
  the
• 11:57
  system
• 11:58
  understands
• 11:58
  intent
• 11:58
  —
• 11:59
  locking
• 11:59
  doors,
• 12:00
  killing
• 12:00
  downstairs
• 12:01
  lights,
• 12:01
  confirming
• 12:02
  with
• 12:02
  a
• 12:02
  beep
• 12:02
  instead
• 12:02
  of
• 12:02
  Alexa's
• 12:03
  verbose
• 12:03
  acknowledgment.
• 12:04
  Response
• 12:04
  times
• 12:05
  are
• 12:05
  faster
• 12:05
  because
• 12:05
  there
• 12:06
  is
• 12:06
  no
• 12:06
  round-trip
• 12:06
  to
• 12:06
  a
• 12:06
  remote
• 12:07
  server,
• 12:07
  and
• 12:08
  a
• 12:08
  physical
• 12:08
  mute
• 12:08
  switch
• 12:08
  means
• 12:09
  the
• 12:09
  microphone
• 12:09
  is
• 12:09
  actually
• 12:09
  off.
• 12:10
  RunAnywhere's
• 12:11
  RCLI
• 12:12
  pushes
• 12:12
  the
• 12:12
  performance
• 12:12
  envelope
• 12:13
  further.
• 12:14
  This
• 12:14
  YC
• 12:14
  W26
• 12:15
  startup
• 12:15
  built
• 12:15
  MetalRT,
• 12:17
  a
• 12:17
  proprietary
• 12:17
  GPU
• 12:18
  inference
• 12:18
  engine
• 12:18
  for
• 12:18
  Apple
• 12:18
  Silicon
• 12:19
  that
• 12:19
  delivers
• 12:19
  up
• 12:19
  to
• 12:19
  550
• 12:20
  tokens
• 12:21
  per
• 12:21
  second
• 12:21
  LLM
• 12:22
  throughput
• 12:22
  and
• 12:23
  sub-200ms
• 12:24
  end-to-end
• 12:24
  voice
• 12:25
  latency
• 12:25
  —
• 12:25
  no
• 12:25
  cloud,
• 12:26
  no
• 12:26
  API
• 12:26
  keys.
• 12:27
  The
• 12:27
  full
• 12:27
  STT
• 12:28
  +
• 12:28
  LLM
• 12:28
  +
• 12:29
  TTS
• 12:29
  pipeline
• 12:29
  runs
• 12:30
  on
• 12:30
  Metal
• 12:30
  GPU
• 12:30
  with
• 12:31
  three
• 12:31
  concurrent
• 12:31
  threads,
• 12:32
  ships
• 12:32
  with
• 12:32
  38
• 12:32
  macOS
• 12:33
  actions
• 12:33
  controllable
• 12:34
  by
• 12:34
  voice,
• 12:34
  and
• 12:34
  includes
• 12:35
  local
• 12:35
  RAG
• 12:35
  with
• 12:35
  hybrid
• 12:36
  vector
• 12:36
  plus
• 12:36
  BM25
• 12:37
  retrieval
• 12:37
  at
• 12:38
  roughly
• 12:38
  4ms
• 12:39
  latency.
• 12:40
  MetalRT's
• 12:40
  STT
• 12:40
  clocks
• 12:41
  714x
• 12:42
  faster
• 12:42
  than
• 12:42
  real-time
• 12:43
  on
• 12:43
  M3
• 12:43
  Max.
• 12:44
  The
• 12:44
  catch:
• 12:44
  MetalRT
• 12:44
  requires
• 12:45
  M3
• 12:45
  or
• 12:45
  later;
• 12:46
  M1/M2
• 12:47
  Macs
• 12:48
  fall
• 12:48
  back
• 12:48
  to
• 12:48
  llama.cpp
• 12:49
  automatically.
• 12:50
  On-device
• 12:50
  inference
• 12:51
  is
• 12:51
  not
• 12:51
  limited
• 12:51
  to
• 12:52
  text
• 12:52
  and
• 12:52
  voice.
• 12:53
  One
• 12:53
  developer
• 12:53
  got
• 12:53
  TripoSR
• 12:54
  —
• 12:54
  a
• 12:54
  single-image-to-3D-mesh
• 12:55
  model
• 12:55
  —
• 12:56
  running
• 12:56
  fully
• 12:57
  on
• 12:57
  iPhone
• 12:57
  via
• 12:57
  ONNX
• 12:58
  Runtime
• 12:58
  with
• 12:58
  CoreML
• 12:59
  as
• 12:59
  the
• 12:59
  backend.
• 12:59
  The
• 12:59
  model
• 13:00
  weighs
• 13:00
  1.6
• 13:01
  GB,
• 13:01
  outputs
• 13:02
  triplane
• 13:02
  scene
• 13:02
  codes
• 13:03
  that
• 13:03
  feed
• 13:03
  marching
• 13:03
  cubes
• 13:03
  for
• 13:04
  mesh
• 13:04
  extraction,
• 13:05
  and
• 13:05
  runs
• 13:05
  on
• 13:05
  A17+
• 13:06
  chips
• 13:06
  with
• 13:06
  no
• 13:07
  network
• 13:07
  calls.
• 13:07
  Memory
• 13:08
  management
• 13:08
  is
• 13:08
  the
• 13:08
  main
• 13:09
  challenge;
• 13:09
  at
• 13:09
  1.6
• 13:10
  GB
• 13:10
  the
• 13:11
  model
• 13:11
  must
• 13:11
  be
• 13:11
  loaded
• 13:11
  carefully
• 13:12
  to
• 13:12
  avoid
• 13:12
  jetsam
• 13:13
  kills.
• 13:14
  The
• 13:14
  rendering
• 13:14
  pipeline
• 13:15
  through
• 13:15
  RealityKit
• 13:15
  ended
• 13:16
  up
• 13:16
  being
• 13:16
  almost
• 13:16
  as
• 13:16
  much
• 13:17
  work
• 13:17
  as
• 13:17
  inference
• 13:17
  itself.
• 13:18
  The
• 13:18
  last
• 13:19
  gap
• 13:19
  in
• 13:19
  the
• 13:19
  local
• 13:19
  stack
• 13:19
  —
• 13:20
  high-quality
• 13:21
  open-source
• 13:21
  speech
• 13:21
  synthesis
• 13:22
  —
• 13:23
  is
• 13:23
  closing
• 13:23
  fast.
• 13:24
  Hume
• 13:24
  AI's
• 13:25
  TADA
• 13:25
  (Text-Acoustic
• 13:26
  Dual
• 13:26
  Alignment)
• 13:27
  resolves
• 13:27
  the
• 13:27
  fundamental
• 13:28
  mismatch
• 13:28
  between
• 13:29
  text
• 13:29
  and
• 13:29
  audio
• 13:30
  representations
• 13:30
  in
• 13:31
  LLM-based
• 13:31
  TTS
• 13:32
  by
• 13:32
  aligning
• 13:33
  audio
• 13:33
  directly
• 13:33
  to
• 13:34
  text
• 13:34
  tokens
• 13:34
  one-to-one,
• 13:35
  creating
• 13:36
  a
• 13:36
  synchronized
• 13:36
  stream
• 13:37
  where
• 13:37
  text
• 13:37
  and
• 13:37
  speech
• 13:38
  move
• 13:38
  in
• 13:38
  lockstep.
• 13:39
  The
• 13:39
  result:
• 13:40
  a
• 13:40
  real-time
• 13:41
  factor
• 13:41
  of
• 13:41
  0.09
• 13:42
  (more
• 13:42
  than
• 13:43
  5x
• 13:43
  faster
• 13:43
  than
• 13:44
  comparable
• 13:44
  systems),
• 13:45
  zero
• 13:45
  hallucinations
• 13:46
  across
• 13:46
  1,000-plus
• 13:48
  test
• 13:48
  samples,
• 13:49
  and
• 13:49
  a
• 13:49
  footprint
• 13:50
  light
• 13:50
  enough
• 13:50
  for
• 13:50
  mobile
• 13:50
  deployment.
• 13:51
  Where
• 13:51
  conventional
• 13:52
  TTS
• 13:52
  exhausts
• 13:53
  a
• 13:53
  2,048-token
• 13:54
  context
• 13:55
  window
• 13:55
  in
• 13:55
  about
• 13:55
  70
• 13:56
  seconds
• 13:56
  of
• 13:56
  audio,
• 13:57
  TADA
• 13:57
  accommodates
• 13:58
  roughly
• 13:58
  700
• 13:58
  seconds
• 13:59
  in
• 13:59
  the
• 13:59
  same
• 13:59
  budget.
• 14:00
  The
• 14:00
  1B
• 14:00
  and
• 14:00
  3B
• 14:01
  parameter
• 14:01
  Llama-based
• 14:02
  models
• 14:02
  are
• 14:02
  available
• 14:03
  now
• 14:03
  under
• 14:03
  an
• 14:03
  open-source
• 14:04
  license.
• 14:08
  The
• 14:08
  open-weight
• 14:08
  community's
• 14:09
  model
• 14:09
  surgery
• 14:10
  techniques
• 14:10
  are
• 14:10
  getting
• 14:10
  increasingly
• 14:11
  sophisticated.
• 14:12
  Heretic's
• 14:13
  new
• 14:13
  Arbitrary-Rank
• 14:14
  Ablation
• 14:14
  (ARA)
• 14:14
  method
• 14:15
  has
• 14:15
  finally
• 14:15
  defeated
• 14:16
  the
• 14:16
  heavy-handed
• 14:16
  alignment
• 14:17
  guardrails
• 14:17
  OpenAI
• 14:18
  applied
• 14:18
  to
• 14:18
  GPT-OSS,
• 14:20
  reducing
• 14:20
  refusals
• 14:21
  from
• 14:21
  74
• 14:21
  (the
• 14:21
  previous
• 14:22
  best
• 14:22
  with
• 14:22
  MPOA+SOMA
• 14:23
  techniques)
• 14:23
  to
• 14:24
  near
• 14:24
  zero
• 14:24
  —
• 14:24
  without
• 14:25
  requiring
• 14:25
  system
• 14:25
  messages.
• 14:26
  ARA
• 14:27
  extends
• 14:27
  the
• 14:27
  rank-1
• 14:27
  ablation
• 14:28
  approach
• 14:28
  that
• 14:28
  was
• 14:28
  already
• 14:29
  effective
• 14:29
  at
• 14:29
  identifying
• 14:30
  and
• 14:30
  removing
• 14:30
  refusal-direction
• 14:31
  vectors
• 14:31
  in
• 14:31
  model
• 14:31
  weights,
• 14:32
  generalizing
• 14:33
  to
• 14:33
  arbitrary
• 14:33
  rank
• 14:33
  for
• 14:33
  finer-grained
• 14:34
  control.
• 14:34
  The
• 14:34
  main
• 14:35
  open
• 14:35
  question
• 14:35
  is
• 14:35
  whether
• 14:36
  the
• 14:36
  extra
• 14:36
  ranks
• 14:36
  capture
• 14:36
  meaningfully
• 14:37
  different
• 14:37
  structure
• 14:37
  or
• 14:38
  risk
• 14:38
  overfitting.
• 14:39
  The
• 14:39
  method
• 14:40
  is
• 14:40
  experimental
• 14:40
  and
• 14:40
  currently
• 14:41
  available
• 14:41
  only
• 14:41
  in
• 14:42
  an
• 14:42
  unreleased
• 14:42
  Heretic
• 14:42
  version,
• 14:43
  but
• 14:43
  when
• 14:43
  it
• 14:43
  ships
• 14:43
  broadly
• 14:44
  it
• 14:44
  represents
• 14:44
  a
• 14:44
  significant
• 14:45
  escalation
• 14:45
  in
• 14:45
  the
• 14:45
  community's
• 14:46
  ability
• 14:46
  to
• 14:46
  post-process
• 14:46
  any
• 14:47
  released
• 14:47
  model
• 14:47
  into
• 14:47
  an
• 14:48
  unconstrained
• 14:48
  variant.
• 14:49
  Fat
• 14:49
  Fish
• 14:49
  takes
• 14:50
  the
• 14:50
  opposite
• 14:50
  approach:
• 14:51
  instead
• 14:51
  of
• 14:51
  removing
• 14:52
  alignment,
• 14:52
  it
• 14:52
  adds
• 14:52
  capacity.
• 14:54
  This
• 14:54
  experimental
• 14:54
  upscale
• 14:55
  of
• 14:55
  Mistral
• 14:55
  Nemo
• 14:56
  goes
• 14:56
  from
• 14:56
  32
• 14:56
  to
• 14:56
  56
• 14:57
  layers,
• 14:57
  32
• 14:58
  to
• 14:58
  48
• 14:58
  attention
• 14:59
  heads,
• 14:59
  and
• 14:59
  8
• 15:00
  to
• 15:00
  12
• 15:00
  key-value
• 15:00
  heads,
• 15:01
  while
• 15:01
  pruning
• 15:01
  intermediate
• 15:02
  size
• 15:02
  from
• 15:02
  14,336
• 15:04
  to
• 15:04
  12,608
• 15:04
  —
• 15:05
  a
• 15:06
  simultaneous
• 15:06
  upscale
• 15:06
  and
• 15:07
  prune
• 15:07
  that
• 15:07
  cost
• 15:07
  roughly
• 15:07
  $1,000
• 15:08
  in
• 15:08
  compute.
• 15:09
  The
• 15:09
  creator's
• 15:10
  rationale
• 15:10
  is
• 15:10
  pragmatic:
• 15:11
  Mistral
• 15:12
  Nemo
• 15:12
  is
• 15:12
  excellent
• 15:12
  but
• 15:13
  dense
• 15:13
  models
• 15:13
  are
• 15:13
  increasingly
• 15:14
  rare
• 15:14
  as
• 15:14
  the
• 15:14
  field
• 15:15
  gravitates
• 15:15
  toward
• 15:15
  mixture-of-experts
• 15:16
  architectures,
• 15:17
  so
• 15:17
  "if
• 15:17
  I
• 15:17
  can't
• 15:17
  get
• 15:18
  a
• 15:18
  new
• 15:18
  base
• 15:18
  model,
• 15:18
  I'll
• 15:19
  make
• 15:19
  one
• 15:19
  myself."
• 15:20
  No
• 15:21
  layers
• 15:21
  were
• 15:21
  zeroed
• 15:21
  out
• 15:21
  (unlike
• 15:22
  typical
• 15:22
  stack
• 15:22
  merges),
• 15:23
  and
• 15:23
  noise
• 15:23
  was
• 15:23
  injected
• 15:24
  to
• 15:24
  encourage
• 15:24
  divergence
• 15:25
  in
• 15:25
  duplicated
• 15:25
  tensors.
• 15:26
  The
• 15:26
  result
• 15:27
  is
• 15:27
  coherent,
• 15:28
  follows
• 15:28
  instructions,
• 15:29
  and
• 15:29
  handles
• 15:29
  new
• 15:30
  languages
• 15:30
  —
• 15:31
  released
• 15:31
  as
• 15:31
  a
• 15:31
  base
• 15:32
  for
• 15:32
  the
• 15:32
  community
• 15:32
  to
• 15:32
  fine-tune
• 15:33
  rather
• 15:34
  than
• 15:34
  a
• 15:34
  finished
• 15:34
  product.
• 15:35
  On
• 15:35
  the
• 15:35
  deployment
• 15:36
  side,
• 15:36
  a
• 15:37
  community
• 15:37
  benchmarking
• 15:38
  effort
• 15:38
  systematically
• 15:39
  optimized
• 15:39
  Qwen3
• 15:39
  Coder
• 15:40
  for
• 15:40
  two
• 15:40
  GPU
• 15:40
  tiers.
• 15:41
  On
• 15:42
  RTX
• 15:42
  5090
• 15:43
  (32
• 15:43
  GB),
• 15:44
  a
• 15:44
  4-bit
• 15:45
  AWQ
• 15:46
  quantized
• 15:46
  Qwen3-Coder-30B-A3B
• 15:49
  hit
• 15:49
  1,157
• 15:51
  tokens
• 15:51
  per
• 15:51
  second
• 15:51
  with
• 15:52
  sub-second
• 15:52
  time-to-first-token
• 15:53
  at
• 15:53
  MCR=16
• 15:54
  via
• 15:55
  vLLM,
• 15:56
  fitting
• 15:56
  114K
• 15:57
  context.
• 15:58
  On
• 15:58
  PRO
• 15:59
  6000
• 15:59
  (96
• 16:00
  GB),
• 16:01
  the
• 16:01
  official
• 16:01
  FP8
• 16:02
  Qwen3-Coder-Next
• 16:03
  reached
• 16:03
  988
• 16:04
  tok/s
• 16:05
  at
• 16:05
  MCR=32
• 16:07
  with
• 16:07
  full
• 16:07
  262K
• 16:08
  context
• 16:08
  —
• 16:10
  though
• 16:10
  vLLM
• 16:10
  peaked
• 16:11
  at
• 16:11
  1,207
• 16:12
  tok/s
• 16:13
  at
• 16:13
  MCR=40
• 16:14
  for
• 16:14
  throughput-optimized
• 16:15
  workloads,
• 16:16
  beating
• 16:16
  SGLang's
• 16:17
  best
• 16:17
  by
• 16:17
  34
• 16:18
  percent.
• 16:19
  The
• 16:19
  entire
• 16:19
  benchmarking
• 16:20
  infrastructure
• 16:20
  is
• 16:20
  open
• 16:21
  via
• 16:21
  DeploDock:
• 16:22
  fork
• 16:22
  the
• 16:22
  repo,
• 16:23
  write
• 16:23
  a
• 16:23
  recipe
• 16:23
  YAML
• 16:24
  specifying
• 16:24
  model,
• 16:25
  engine
• 16:25
  parameters,
• 16:26
  and
• 16:26
  sweep
• 16:26
  matrices,
• 16:27
  open
• 16:28
  a
• 16:28
  PR,
• 16:29
  and
• 16:29
  a
• 16:29
  maintainer
• 16:29
  triggers
• 16:30
  cloud-provisioned
• 16:30
  benchmark
• 16:31
  runs
• 16:31
  with
• 16:31
  results
• 16:32
  posted
• 16:32
  back
• 16:32
  automatically.
• 16:33
  Available
• 16:34
  GPUs
• 16:34
  range
• 16:34
  from
• 16:34
  RTX
• 16:35
  4090
• 16:36
  through
• 16:36
  H200
• 16:37
  and
• 16:37
  B200.
• 16:38
  The
• 16:38
  capital
• 16:38
  flowing
• 16:39
  into
• 16:39
  open-weight
• 16:39
  model
• 16:40
  development
• 16:40
  is
• 16:40
  now
• 16:40
  industrial-scale.
• 16:42
  Yann
• 16:42
  LeCun's
• 16:42
  AMI
• 16:43
  Labs
• 16:43
  raised
• 16:44
  $1.03
• 16:44
  billion
• 16:45
  to
• 16:45
  build
• 16:45
  world
• 16:46
  models
• 16:46
  —
• 16:47
  architectures
• 16:47
  that
• 16:47
  learn
• 16:48
  predictive
• 16:48
  representations
• 16:49
  of
• 16:49
  the
• 16:49
  physical
• 16:49
  world
• 16:50
  rather
• 16:50
  than
• 16:50
  next-token
• 16:50
  probabilities.
• 16:56
  Steve
• 16:56
  Yegge's
• 16:56
  Gas
• 16:57
  Town
• 16:57
  is
• 16:57
  an
• 16:57
  agent
• 16:57
  orchestrator
• 16:58
  in
• 16:58
  the
• 16:58
  most
• 16:58
  literal
• 16:59
  sense
• 16:59
  —
• 17:00
  a
• 17:00
  system
• 17:00
  that
• 17:00
  coordinates
• 17:01
  20
• 17:01
  to
• 17:01
  30
• 17:01
  AI
• 17:01
  coding
• 17:02
  agents
• 17:02
  simultaneously
• 17:03
  with
• 17:03
  defined
• 17:03
  roles,
• 17:04
  workflow
• 17:04
  management,
• 17:05
  merge
• 17:05
  queues,
• 17:06
  patrol
• 17:06
  loops,
• 17:06
  and
• 17:06
  a
• 17:06
  level
• 17:07
  of
• 17:07
  operational
• 17:07
  complexity
• 17:08
  that
• 17:08
  Yegge
• 17:08
  himself
• 17:08
  compares
• 17:09
  to
• 17:09
  Kubernetes
• 17:09
  (and
• 17:10
  to
• 17:10
  a
• 17:10
  late-1800s
• 17:11
  factory
• 17:11
  that
• 17:11
  can
• 17:11
  disembowel
• 17:12
  you).
• 17:12
  People
• 17:12
  are
• 17:12
  using
• 17:12
  it
• 17:13
  to
• 17:13
  churn
• 17:13
  through
• 17:13
  massive
• 17:13
  implementation
• 17:14
  backlogs:
• 17:15
  parallel
• 17:15
  agent
• 17:16
  swarms
• 17:16
  file
• 17:16
  issues,
• 17:17
  write
• 17:17
  code,
• 17:18
  review
• 17:18
  each
• 17:18
  other's
• 17:18
  output,
• 17:19
  merge
• 17:19
  changes,
• 17:20
  and
• 17:20
  land
• 17:20
  features
• 17:20
  while
• 17:20
  the
• 17:21
  developer
• 17:21
  focuses
• 17:21
  on
• 17:21
  design
• 17:22
  and
• 17:22
  direction.
• 17:23
  The
• 17:23
  self-hosted
• 17:24
  reality,
• 17:24
  however,
• 17:24
  is
• 17:24
  punishing
• 17:25
  —
• 17:25
  tmux
• 17:26
  session
• 17:26
  management,
• 17:26
  compute
• 17:27
  provisioning,
• 17:27
  API
• 17:28
  key
• 17:28
  juggling
• 17:28
  across
• 17:28
  providers,
• 17:29
  monitoring,
• 17:30
  and
• 17:30
  recovery
• 17:30
  when
• 17:31
  things
• 17:31
  inevitably
• 17:31
  break.
• 17:33
  Kilo's
• 17:33
  managed
• 17:33
  version
• 17:34
  removes
• 17:34
  that
• 17:34
  operational
• 17:35
  tax.
• 17:36
  The
• 17:36
  full
• 17:36
  Gas
• 17:37
  Town
• 17:37
  environment
• 17:37
  deploys
• 17:38
  in
• 17:38
  seconds,
• 17:39
  agents
• 17:39
  scale
• 17:39
  elastically
• 17:40
  (5
• 17:40
  or
• 17:40
  50
• 17:40
  as
• 17:41
  needed),
• 17:41
  convoys
• 17:42
  run
• 17:42
  around
• 17:42
  the
• 17:42
  clock
• 17:42
  with
• 17:42
  auto-recovery,
• 17:44
  and
• 17:44
  the
• 17:44
  Kilo
• 17:44
  Gateway
• 17:44
  provides
• 17:45
  access
• 17:45
  to
• 17:45
  over
• 17:45
  500
• 17:46
  models
• 17:46
  through
• 17:46
  a
• 17:47
  single
• 17:47
  API
• 17:47
  with
• 17:47
  zero
• 17:48
  markup
• 17:48
  on
• 17:48
  tokens.
• 17:49
  For
• 17:49
  a
• 17:50
  system
• 17:50
  that
• 17:50
  burns
• 17:50
  through
• 17:50
  tokens
• 17:51
  at
• 17:51
  Gas
• 17:51
  Town's
• 17:51
  rate,
• 17:52
  consolidated
• 17:53
  billing
• 17:53
  alone
• 17:53
  justifies
• 17:54
  the
• 17:54
  managed
• 17:54
  layer.
• 17:55
  Yegge
• 17:56
  merged
• 17:56
  over
• 17:56
  100
• 17:57
  PRs
• 17:57
  from
• 17:57
  nearly
• 17:58
  50
• 17:58
  contributors
• 17:59
  in
• 17:59
  the
• 17:59
  first
• 17:59
  12
• 17:59
  days
• 18:00
  after
• 18:00
  launch
• 18:00
  —
• 18:01
  keeping
• 18:01
  up
• 18:02
  with
• 18:02
  that
• 18:02
  pace
• 18:02
  manually
• 18:03
  is
• 18:03
  a
• 18:03
  job
• 18:03
  in
• 18:03
  itself.
• 18:04
  The
• 18:04
  trajectory
• 18:05
  from
• 18:05
  single-agent
• 18:06
  coding
• 18:06
  assistants
• 18:06
  to
• 18:06
  multi-agent
• 18:07
  orchestration
• 18:08
  platforms
• 18:08
  like
• 18:08
  Gas
• 18:08
  Town
• 18:09
  raises
• 18:09
  a
• 18:09
  question
• 18:10
  the
• 18:10
  archive
• 18:10
  feature
• 18:11
  behind
• 18:11
  archive.is
• 18:12
  underscores
• 18:12
  in
• 18:12
  a
• 18:12
  different
• 18:13
  way:
• 18:14
  as
• 18:14
  AI-generated
• 18:15
  content
• 18:15
  proliferates,
• 18:16
  preserving
• 18:17
  and
• 18:17
  accessing
• 18:18
  the
• 18:18
  original
• 18:18
  human-authored
• 18:19
  sources
• 18:19
  becomes
• 18:20
  infrastructure,
• 18:21
  not
• 18:21
  convenience.