Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 30, 2026

Transcript

AGI Dreams Podcast – March 30, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  A
• 0:00
  researcher
• 0:00
  decrypted
• 0:00
  377
• 0:02
  Cloudflare
• 0:02
  Turnstile
• 0:03
  programs
• 0:03
  from
• 0:03
  ChatGPT
• 0:04
  network
• 0:04
  traffic
• 0:05
  and
• 0:05
  found
• 0:05
  something
• 0:05
  that
• 0:05
  goes
• 0:06
  well
• 0:06
  beyond
• 0:06
  standard
• 0:07
  browser
• 0:07
  fingerprinting.
• 0:08
  Each
• 0:08
  program
• 0:08
  collects
• 0:08
  55
• 0:09
  properties
• 0:09
  across
• 0:10
  three
• 0:10
  layers:
• 0:10
  the
• 0:10
  browser
• 0:11
  (GPU,
• 0:11
  screen,
• 0:12
  fonts),
• 0:12
  the
• 0:12
  Cloudflare
• 0:13
  edge
• 0:13
  network
• 0:13
  (city,
• 0:14
  IP,
• 0:14
  region
• 0:15
  from
• 0:15
  edge
• 0:15
  headers),
• 0:15
  and
• 0:15
  --
• 0:16
  the
• 0:16
  part
• 0:16
  that
• 0:16
  matters
• 0:16
  --
• 0:17
  the
• 0:17
  ChatGPT
• 0:17
  React
• 0:18
  application
• 0:18
  itself.
• 0:19
  The
• 0:19
  program
• 0:20
  checks
• 0:20
  internal
• 0:20
  React
• 0:21
  Router
• 0:21
  v6+
• 0:21
  data
• 0:22
  structures,
• 0:23
  route
• 0:23
  loader
• 0:23
  results,
• 0:24
  and
• 0:24
  SSR
• 0:24
  hydration
• 0:25
  state.
• 0:25
  A
• 0:26
  bot
• 0:26
  that
• 0:26
  spoofs
• 0:26
  browser
• 0:27
  fingerprints
• 0:27
  but
• 0:27
  doesn't
• 0:28
  actually
• 0:28
  render
• 0:28
  and
• 0:29
  hydrate
• 0:29
  the
• 0:29
  full
• 0:29
  ChatGPT
• 0:30
  single-page
• 0:30
  application
• 0:31
  will
• 0:31
  fail.
• 0:31
  This
• 0:31
  is
• 0:32
  bot
• 0:32
  detection
• 0:32
  at
• 0:32
  the
• 0:32
  application
• 0:33
  layer,
• 0:33
  not
• 0:34
  the
• 0:34
  browser
• 0:34
  layer.
• 0:35
  The
• 0:35
  encryption
• 0:35
  was
• 0:35
  supposed
• 0:35
  to
• 0:36
  hide
• 0:36
  this.
• 0:36
  The
• 0:36
  outer
• 0:36
  layer
• 0:37
  is
• 0:37
  XOR'd
• 0:37
  with
• 0:38
  a
• 0:38
  token
• 0:38
  from
• 0:38
  the
• 0:38
  prepare
• 0:38
  request
• 0:39
  --
• 0:39
  both
• 0:39
  travel
• 0:40
  in
• 0:40
  the
• 0:40
  same
• 0:40
  HTTP
• 0:41
  exchange.
• 0:42
  Inside
• 0:42
  those
• 0:42
  89
• 0:43
  VM
• 0:43
  instructions
• 0:44
  sits
• 0:44
  a
• 0:44
  19KB
• 0:45
  encrypted
• 0:45
  blob
• 0:46
  containing
• 0:46
  the
• 0:46
  actual
• 0:47
  fingerprinting
• 0:47
  program.
• 0:48
  The
• 0:48
  inner
• 0:48
  key?
• 0:49
  A
• 0:49
  float
• 0:49
  literal
• 0:50
  embedded
• 0:50
  in
• 0:50
  the
• 0:50
  bytecode
• 0:51
  itself,
• 0:51
  verified
• 0:52
  across
• 0:52
  50
• 0:52
  requests,
• 0:53
  50
• 0:53
  out
• 0:54
  of
• 0:54
  50.
• 0:54
  The
• 0:54
  "encryption"
• 0:55
  prevents
• 0:55
  casual
• 0:56
  inspection
• 0:56
  but
• 0:56
  not
• 0:57
  analysis.
• 0:58
  Alongside
• 0:58
  the
• 0:59
  fingerprint,
• 0:59
  Turnstile
• 1:00
  runs
• 1:00
  a
• 1:00
  behavioral
• 1:00
  biometric
• 1:01
  layer
• 1:01
  (keystroke
• 1:01
  timing,
• 1:02
  mouse
• 1:02
  velocity,
• 1:03
  scroll
• 1:03
  patterns,
• 1:04
  idle
• 1:04
  time
• 1:04
  across
• 1:05
  36
• 1:05
  properties)
• 1:05
  and
• 1:06
  a
• 1:06
  SHA-256
• 1:08
  hashcash
• 1:08
  proof-of-work
• 1:09
  challenge.
• 1:10
  The
• 1:10
  privacy
• 1:10
  boundary
• 1:10
  between
• 1:11
  user
• 1:11
  and
• 1:11
  system
• 1:11
  operator
• 1:12
  is
• 1:12
  a
• 1:12
  policy
• 1:12
  decision,
• 1:13
  not
• 1:13
  a
• 1:13
  cryptographic
• 1:14
  one.
• 1:14
  If
• 1:14
  Cloudflare
• 1:15
  has
• 1:15
  to
• 1:15
  fingerprint
• 1:15
  entire
• 1:16
  React
• 1:16
  application
• 1:17
  state
• 1:17
  to
• 1:17
  keep
• 1:17
  bots
• 1:17
  out
• 1:17
  of
• 1:18
  ChatGPT,
• 1:19
  it
• 1:19
  tells
• 1:19
  you
• 1:19
  something
• 1:19
  about
• 1:19
  the
• 1:20
  arms
• 1:20
  race
• 1:20
  on
• 1:20
  the
• 1:20
  other
• 1:20
  side.
• 1:21
  ClawShield
• 1:21
  takes
• 1:22
  the
• 1:22
  inverse
• 1:22
  approach:
• 1:23
  instead
• 1:23
  of
• 1:23
  detecting
• 1:24
  bots
• 1:24
  trying
• 1:24
  to
• 1:24
  reach
• 1:24
  an
• 1:25
  AI,
• 1:25
  it
• 1:25
  secures
• 1:26
  AI
• 1:26
  agents
• 1:26
  trying
• 1:26
  to
• 1:27
  reach
• 1:27
  the
• 1:27
  world.
• 1:28
  This
• 1:28
  open-source
• 1:29
  Go
• 1:29
  proxy
• 1:29
  sits
• 1:29
  between
• 1:30
  users
• 1:30
  and
• 1:30
  any
• 1:30
  OpenAI-compatible
• 1:31
  gateway,
• 1:32
  scanning
• 1:33
  every
• 1:33
  message
• 1:33
  for
• 1:33
  prompt
• 1:34
  injection,
• 1:35
  PII
• 1:35
  leaks,
• 1:35
  secrets,
• 1:36
  and
• 1:36
  vulnerability
• 1:37
  patterns
• 1:37
  before
• 1:38
  they
• 1:38
  reach
• 1:38
  the
• 1:38
  model
• 1:39
  or
• 1:39
  leave
• 1:39
  the
• 1:39
  network.
• 1:40
  The
• 1:40
  architecture
• 1:40
  uses
• 1:41
  defense-in-depth
• 1:42
  across
• 1:42
  three
• 1:42
  layers
• 1:42
  --
• 1:43
  application-level
• 1:44
  content
• 1:44
  scanning
• 1:45
  with
• 1:45
  a
• 1:45
  YAML
• 1:45
  deny-by-default
• 1:46
  policy
• 1:47
  engine,
• 1:47
  iptables-based
• 1:48
  egress
• 1:49
  firewalling
• 1:49
  restricting
• 1:50
  which
• 1:50
  domains
• 1:50
  agents
• 1:51
  can
• 1:51
  reach,
• 1:51
  and
• 1:52
  optional
• 1:52
  eBPF
• 1:53
  kernel-level
• 1:53
  syscall
• 1:54
  monitoring
• 1:54
  for
• 1:54
  detecting
• 1:55
  privilege
• 1:55
  escalation
• 1:56
  and
• 1:56
  fork
• 1:56
  bombs.
• 1:57
  The
• 1:57
  layers
• 1:57
  communicate
• 1:58
  via
• 1:58
  a
• 1:58
  Unix
• 1:58
  socket
• 1:59
  event
• 1:59
  bus,
• 2:00
  enabling
• 2:00
  adaptive
• 2:00
  responses:
• 2:01
  if
• 2:02
  eBPF
• 2:02
  detects
• 2:03
  a
• 2:03
  port
• 2:03
  scan,
• 2:04
  the
• 2:04
  proxy
• 2:04
  elevates
• 2:05
  injection
• 2:05
  sensitivity
• 2:06
  for
• 2:06
  ten
• 2:06
  minutes;
• 2:07
  if
• 2:07
  the
• 2:07
  proxy
• 2:08
  blocks
• 2:08
  three
• 2:08
  injections
• 2:09
  in
• 2:09
  sixty
• 2:09
  seconds,
• 2:10
  the
• 2:10
  default
• 2:10
  action
• 2:11
  forces
• 2:11
  to
• 2:11
  deny.
• 2:12
  Production
• 2:13
  hardening
• 2:13
  includes
• 2:14
  Go-native
• 2:14
  eBPF
• 2:15
  replacing
• 2:15
  a
• 2:15
  Python
• 2:15
  dependency,
• 2:16
  graceful
• 2:17
  degradation
• 2:17
  to
• 2:18
  /proc
• 2:18
  polling
• 2:18
  when
• 2:19
  eBPF
• 2:19
  is
• 2:19
  unavailable,
• 2:20
  and
• 2:21
  streaming
• 2:21
  response
• 2:21
  scanning
• 2:22
  with
• 2:22
  a
• 2:22
  sliding
• 2:22
  200-character
• 2:23
  overlap
• 2:24
  window
• 2:24
  that
• 2:24
  catches
• 2:25
  patterns
• 2:25
  spanning
• 2:25
  chunk
• 2:26
  boundaries.
• 2:27
  NanoClaw
• 2:27
  is
• 2:27
  approaching
• 2:28
  this
• 2:28
  from
• 2:28
  the
• 2:28
  platform
• 2:29
  level.
• 2:29
  Eight
• 2:30
  weeks
• 2:30
  after
• 2:30
  launch,
• 2:31
  the
• 2:31
  project
• 2:31
  has
• 2:31
  crossed
• 2:32
  25,000
• 2:32
  GitHub
• 2:33
  stars
• 2:33
  and
• 2:34
  shipped
• 2:34
  what
• 2:34
  it
• 2:34
  calls
• 2:34
  the
• 2:34
  first
• 2:35
  agent
• 2:35
  vault
• 2:35
  --
• 2:36
  adopting
• 2:36
  OneCLI's
• 2:37
  Agent
• 2:37
  Vault
• 2:37
  as
• 2:37
  the
• 2:37
  default
• 2:38
  credential
• 2:38
  and
• 2:38
  proxying
• 2:39
  layer.
• 2:40
  Every
• 2:40
  agent
• 2:40
  now
• 2:40
  accesses
• 2:41
  external
• 2:41
  services
• 2:42
  through
• 2:42
  a
• 2:42
  vault
• 2:42
  gateway;
• 2:43
  the
• 2:43
  agent
• 2:43
  never
• 2:44
  sees
• 2:44
  raw
• 2:44
  API
• 2:45
  keys.
• 2:46
  Rate
• 2:46
  limiting
• 2:46
  is
• 2:46
  live:
• 2:47
  if
• 2:47
  the
• 2:47
  widely-reported
• 2:48
  Meta
• 2:48
  incident
• 2:49
  (an
• 2:49
  alignment
• 2:49
  director's
• 2:50
  email
• 2:50
  agent
• 2:51
  mass-deleting
• 2:51
  messages
• 2:52
  despite
• 2:53
  explicit
• 2:53
  instructions
• 2:54
  not
• 2:54
  to)
• 2:54
  had
• 2:54
  happened
• 2:55
  with
• 2:55
  a
• 2:55
  rate
• 2:55
  limit
• 2:55
  of
• 2:55
  three
• 2:56
  deletions
• 2:56
  per
• 2:56
  hour,
• 2:57
  the
• 2:57
  damage
• 2:57
  would
• 2:57
  have
• 2:58
  been
• 2:58
  three
• 2:58
  emails,
• 2:59
  not
• 2:59
  an
• 2:59
  entire
• 2:59
  inbox.
• 3:00
  Container
• 3:01
  isolation
• 3:01
  was
• 3:02
  step
• 3:02
  one,
• 3:02
  credential
• 3:03
  isolation
• 3:03
  is
• 3:03
  step
• 3:04
  two.
• 3:04
  The
• 3:05
  harder
• 3:05
  question,
• 3:05
  raised
• 3:06
  in
• 3:06
  the
• 3:06
  thread
• 3:06
  itself:
• 3:07
  the
• 3:07
  vault
• 3:07
  governs
• 3:08
  access,
• 3:09
  but
• 3:09
  the
• 3:09
  agent
• 3:09
  can
• 3:09
  still
• 3:10
  exfiltrate
• 3:10
  data
• 3:11
  it
• 3:11
  retrieves
• 3:11
  through
• 3:11
  those
• 3:11
  keys.
• 3:13
  Credential
• 3:13
  leakage
• 3:13
  and
• 3:14
  data
• 3:14
  exfiltration
• 3:15
  are
• 3:15
  two
• 3:15
  different
• 3:16
  trust
• 3:16
  boundaries,
• 3:17
  and
• 3:17
  NanoClaw
• 3:17
  hasn't
• 3:18
  publicly
• 3:18
  addressed
• 3:18
  the
• 3:19
  second
• 3:19
  one
• 3:19
  yet.
• 3:23
  The
• 3:23
  argument
• 3:24
  for
• 3:24
  a
• 3:24
  single
• 3:24
  omniscient
• 3:24
  agent
• 3:25
  is
• 3:25
  losing
• 3:25
  ground.
• 3:26
  As
• 3:27
  one
• 3:27
  practitioner
• 3:27
  puts
• 3:27
  it:
• 3:28
  everyone
• 3:28
  is
• 3:28
  chasing
• 3:28
  the
• 3:28
  perfect
• 3:29
  agent
• 3:29
  --
• 3:29
  one
• 3:30
  model,
• 3:30
  one
• 3:30
  stack,
• 3:31
  one
• 3:31
  personality
• 3:31
  supposedly
• 3:32
  crushing
• 3:32
  every
• 3:32
  task.
• 3:34
  The
• 3:34
  counter-thesis
• 3:34
  for
• 3:34
  2026
• 3:35
  is
• 3:35
  unbundling
• 3:36
  and
• 3:36
  rebundling.
• 3:37
  Not
• 3:37
  one
• 3:37
  mega-agent,
• 3:37
  but
• 3:38
  specialists:
• 3:39
  context
• 3:39
  retrieval
• 3:39
  that
• 3:40
  finds
• 3:40
  the
• 3:40
  needle
• 3:40
  and
• 3:40
  ignores
• 3:41
  the
• 3:41
  haystack,
• 3:41
  code
• 3:42
  edit
• 3:42
  prediction
• 3:42
  at
• 3:42
  sub-200ms,
• 3:43
  domain
• 3:44
  reasoning
• 3:44
  that
• 3:44
  thinks
• 3:44
  in
• 3:44
  your
• 3:45
  stack,
• 3:45
  verification
• 3:46
  that
• 3:46
  catches
• 3:46
  the
• 3:46
  bug
• 3:47
  you'd
• 3:47
  ship
• 3:47
  at
• 3:47
  2am.
• 3:48
  The
• 3:48
  proof
• 3:49
  is
• 3:49
  already
• 3:49
  shipping:
• 3:50
  Chroma's
• 3:50
  Context-1
• 3:51
  built
• 3:51
  purely
• 3:51
  for
• 3:52
  retrieval,
• 3:52
  Zed's
• 3:53
  Zeta-2
• 3:53
  predicting
• 3:54
  edits
• 3:54
  at
• 3:54
  under
• 3:54
  200
• 3:55
  milliseconds,
• 3:56
  NVIDIA's
• 3:56
  ProRL
• 3:57
  fine-tuning
• 3:57
  a
• 3:57
  single
• 3:57
  step
• 3:58
  of
• 3:58
  the
• 3:58
  reinforcement
• 3:58
  loop.
• 3:59
  You
• 3:59
  optimize
• 4:00
  accuracy,
• 4:00
  latency,
• 4:01
  cost,
• 4:02
  privacy,
• 4:02
  and
• 4:03
  context
• 4:03
  quality
• 4:03
  independently.
• 4:05
  The
• 4:05
  most
• 4:05
  entertaining
• 4:05
  proof
• 4:06
  of
• 4:06
  concept
• 4:06
  is
• 4:06
  an
• 4:06
  AI
• 4:07
  agent
• 4:07
  running
• 4:07
  on
• 4:07
  a
• 4:07
  $7/month
• 4:08
  VPS
• 4:09
  with
• 4:09
  IRC
• 4:09
  as
• 4:09
  its
• 4:10
  transport
• 4:10
  layer.
• 4:11
  George
• 4:11
  Larson's
• 4:12
  "nullclaw"
• 4:13
  is
• 4:13
  a
• 4:13
  678KB
• 4:14
  Zig
• 4:15
  binary
• 4:15
  using
• 4:15
  about
• 4:16
  1MB
• 4:16
  of
• 4:16
  RAM,
• 4:17
  connected
• 4:17
  to
• 4:17
  a
• 4:17
  self-hosted
• 4:18
  Ergo
• 4:18
  IRC
• 4:19
  server.
• 4:20
  Visitors
• 4:20
  hit
• 4:20
  a
• 4:20
  gamja
• 4:21
  web
• 4:21
  client
• 4:21
  embedded
• 4:21
  in
• 4:22
  his
• 4:22
  portfolio
• 4:22
  site;
• 4:23
  the
• 4:23
  agent
• 4:23
  handles
• 4:24
  greetings
• 4:24
  and
• 4:24
  project
• 4:25
  questions
• 4:25
  using
• 4:25
  Haiku
• 4:26
  for
• 4:26
  the
• 4:26
  hot
• 4:26
  path
• 4:26
  and
• 4:26
  Sonnet
• 4:27
  for
• 4:27
  tool
• 4:27
  use
• 4:27
  when
• 4:27
  it
• 4:27
  needs
• 4:28
  to
• 4:28
  clone
• 4:28
  a
• 4:28
  repo
• 4:28
  and
• 4:28
  actually
• 4:29
  read
• 4:29
  test
• 4:29
  files.
• 4:30
  A
• 4:30
  second
• 4:31
  agent,
• 4:31
  "ironclaw,"
• 4:32
  runs
• 4:32
  on
• 4:32
  a
• 4:33
  separate
• 4:33
  box
• 4:33
  via
• 4:33
  Tailscale
• 4:34
  with
• 4:34
  access
• 4:35
  to
• 4:35
  email
• 4:35
  and
• 4:35
  calendar,
• 4:36
  connected
• 4:37
  through
• 4:37
  Google's
• 4:37
  A2A
• 4:37
  protocol.
• 4:39
  The
• 4:39
  security
• 4:39
  model
• 4:39
  is
• 4:40
  deliberate:
• 4:40
  the
• 4:41
  public
• 4:41
  box
• 4:41
  has
• 4:41
  no
• 4:42
  access
• 4:42
  to
• 4:42
  private
• 4:42
  data,
• 4:43
  the
• 4:43
  blast
• 4:44
  radius
• 4:44
  of
• 4:44
  compromise
• 4:45
  is
• 4:45
  an
• 4:45
  IRC
• 4:46
  bot
• 4:46
  with
• 4:46
  a
• 4:46
  $2/day
• 4:48
  inference
• 4:48
  budget,
• 4:49
  and
• 4:49
  prompt
• 4:49
  injection
• 4:50
  is
• 4:50
  mitigated
• 4:51
  by
• 4:51
  strict
• 4:51
  routing
• 4:51
  --
• 4:53
  only
• 4:53
  scheduling,
• 4:54
  availability,
• 4:55
  and
• 4:55
  contact
• 4:55
  requests
• 4:56
  get
• 4:56
  escalated
• 4:56
  to
• 4:56
  ironclaw.
• 4:58
  Total
• 4:58
  footprint:
• 4:59
  under
• 4:59
  10MB
• 5:00
  of
• 5:00
  binaries,
• 5:01
  under
• 5:01
  5MB
• 5:02
  RAM
• 5:02
  at
• 5:02
  idle.
• 5:03
  The
• 5:03
  deeper
• 5:03
  point:
• 5:04
  model
• 5:05
  selection
• 5:05
  is
• 5:05
  a
• 5:05
  design
• 5:06
  decision,
• 5:07
  not
• 5:07
  a
• 5:07
  settings
• 5:07
  toggle.
• 5:08
  Using
• 5:09
  Opus
• 5:09
  for
• 5:09
  a
• 5:09
  concierge
• 5:10
  would
• 5:10
  signal
• 5:10
  the
• 5:10
  opposite
• 5:11
  of
• 5:11
  model
• 5:11
  understanding.
• 5:13
  oh-my-claudecode
• 5:14
  (OMC)
• 5:14
  formalizes
• 5:15
  multi-agent
• 5:16
  orchestration
• 5:16
  as
• 5:16
  a
• 5:16
  first-class
• 5:17
  primitive
• 5:17
  with
• 5:18
  a
• 5:18
  five-stage
• 5:18
  pipeline:
• 5:19
  plan,
• 5:20
  PRD,
• 5:21
  execute,
• 5:22
  verify,
• 5:22
  fix.
• 5:23
  The
• 5:23
  framework
• 5:24
  supports
• 5:24
  three
• 5:25
  worker
• 5:25
  types
• 5:25
  --
• 5:26
  native
• 5:26
  Claude
• 5:26
  teams,
• 5:27
  tmux-spawned
• 5:28
  CLI
• 5:28
  workers
• 5:29
  for
• 5:29
  Codex
• 5:29
  and
• 5:29
  Gemini,
• 5:30
  and
• 5:30
  a
• 5:30
  tri-model
• 5:31
  synthesis
• 5:31
  mode.
• 5:32
  Each
• 5:32
  worker
• 5:33
  is
• 5:33
  a
• 5:33
  real
• 5:33
  CLI
• 5:34
  process
• 5:34
  in
• 5:34
  its
• 5:34
  own
• 5:35
  tmux
• 5:35
  pane,
• 5:36
  not
• 5:36
  simulated
• 5:36
  concurrency.
• 5:38
  The
• 5:38
  deep
• 5:38
  interview
• 5:38
  mode
• 5:39
  uses
• 5:39
  Socratic
• 5:39
  questioning
• 5:40
  to
• 5:40
  expose
• 5:41
  hidden
• 5:41
  assumptions
• 5:41
  before
• 5:42
  any
• 5:42
  code
• 5:42
  is
• 5:42
  written,
• 5:43
  measuring
• 5:43
  requirement
• 5:44
  clarity
• 5:44
  across
• 5:45
  weighted
• 5:45
  dimensions.
• 5:46
  Cost
• 5:46
  optimization
• 5:47
  claims
• 5:47
  30-50%
• 5:50
  token
• 5:50
  savings
• 5:51
  through
• 5:51
  smart
• 5:51
  model
• 5:51
  routing.
• 5:52
  The
• 5:53
  caveats
• 5:53
  are
• 5:53
  real:
• 5:54
  the
• 5:54
  native
• 5:54
  teams
• 5:55
  feature
• 5:55
  depends
• 5:56
  on
• 5:56
  an
• 5:56
  unstable
• 5:56
  Claude
• 5:57
  Code
• 5:57
  flag,
• 5:58
  the
• 5:58
  tmux
• 5:58
  dependency
• 5:59
  kills
• 5:59
  Windows
• 6:00
  compatibility
• 6:00
  outside
• 6:01
  WSL2,
• 6:02
  and
• 6:02
  version
• 6:03
  4.4.0
• 6:05
  removed
• 6:05
  Codex/Gemini
• 6:06
  MCP
• 6:07
  servers
• 6:07
  entirely,
• 6:08
  forcing
• 6:09
  migration
• 6:09
  to
• 6:10
  CLI
• 6:10
  workers.
• 6:11
  Eyes
• 6:11
  for
• 6:12
  Claude
• 6:12
  gives
• 6:12
  coding
• 6:12
  assistants
• 6:13
  a
• 6:13
  webcam.
• 6:14
  The
• 6:15
  tool
• 6:15
  streams
• 6:15
  video
• 6:16
  at
• 6:16
  1
• 6:16
  FPS
• 6:16
  via
• 6:17
  WebSocket
• 6:17
  to
• 6:17
  Google's
• 6:18
  Gemini
• 6:18
  Live
• 6:18
  API,
• 6:19
  which
• 6:20
  builds
• 6:20
  visual
• 6:20
  context
• 6:21
  over
• 6:21
  time
• 6:21
  --
• 6:22
  answering
• 6:23
  questions
• 6:23
  about
• 6:23
  things
• 6:24
  that
• 6:24
  happened
• 6:24
  minutes
• 6:24
  ago,
• 6:25
  not
• 6:25
  just
• 6:25
  what's
• 6:25
  on
• 6:26
  screen
• 6:26
  now.
• 6:27
  The
• 6:27
  motivation:
• 6:28
  iterating
• 6:29
  on
• 6:29
  hardware
• 6:29
  projects
• 6:30
  with
• 6:30
  Claude
• 6:30
  Code
• 6:30
  meant
• 6:31
  a
• 6:31
  screenshot
• 6:31
  loop
• 6:31
  that
• 6:32
  got
• 6:32
  old
• 6:32
  fast.
• 6:33
  Now
• 6:33
  the
• 6:33
  camera
• 6:34
  stays
• 6:34
  pointed
• 6:34
  at
• 6:34
  the
• 6:35
  workspace
• 6:35
  and
• 6:35
  Claude
• 6:35
  queries
• 6:36
  it
• 6:36
  via
• 6:36
  bash
• 6:37
  whenever
• 6:37
  needed.
• 6:38
  Meanwhile,
• 6:39
  notebooklm-py
• 6:40
  ships
• 6:40
  a
• 6:40
  full
• 6:40
  unofficial
• 6:41
  Python
• 6:41
  API
• 6:42
  to
• 6:42
  NotebookLM,
• 6:43
  exposing
• 6:44
  capabilities
• 6:44
  the
• 6:44
  web
• 6:45
  UI
• 6:45
  doesn't
• 6:45
  --
• 6:46
  batch
• 6:46
  downloads,
• 6:47
  quiz/flashcard
• 6:48
  export,
• 6:49
  mind
• 6:49
  map
• 6:49
  JSON
• 6:50
  extraction,
• 6:51
  slide
• 6:51
  deck
• 6:51
  as
• 6:51
  editable
• 6:52
  PPTX,
• 6:53
  and
• 6:53
  programmatic
• 6:54
  sharing.
• 6:55
  It
• 6:55
  integrates
• 6:55
  as
• 6:56
  a
• 6:56
  skill
• 6:56
  for
• 6:56
  Claude
• 6:56
  Code,
• 6:57
  Codex,
• 6:58
  and
• 6:58
  OpenClaw
• 6:59
  agents,
• 7:00
  making
• 7:00
  research
• 7:00
  automation
• 7:01
  a
• 7:01
  composable
• 7:01
  building
• 7:02
  block
• 7:02
  rather
• 7:03
  than
• 7:03
  a
• 7:03
  walled
• 7:03
  garden.
• 7:06
  The
• 7:06
  Mamba-3
• 7:06
  paper
• 7:06
  from
• 7:07
  Carnegie
• 7:07
  Mellon,
• 7:07
  Princeton,
• 7:08
  Together
• 7:08
  AI,
• 7:09
  and
• 7:09
  Cartesia
• 7:09
  introduces
• 7:10
  three
• 7:10
  core
• 7:10
  improvements
• 7:11
  to
• 7:11
  the
• 7:11
  SSM
• 7:11
  lineage.
• 7:12
  First,
• 7:12
  exponential-trapezoidal
• 7:13
  discretization
• 7:14
  provides
• 7:15
  a
• 7:15
  more
• 7:15
  expressive
• 7:15
  recurrence
• 7:16
  that
• 7:16
  empirically
• 7:16
  replaces
• 7:17
  the
• 7:17
  short
• 7:17
  causal
• 7:18
  convolution
• 7:18
  previously
• 7:19
  considered
• 7:19
  essential
• 7:19
  for
• 7:20
  recurrent
• 7:20
  models.
• 7:21
  Second,
• 7:21
  a
• 7:21
  complex-valued
• 7:22
  state
• 7:22
  update
• 7:22
  --
• 7:23
  equivalent
• 7:23
  to
• 7:23
  a
• 7:24
  data-dependent
• 7:24
  rotary
• 7:24
  embedding
• 7:25
  --
• 7:25
  enables
• 7:26
  state
• 7:26
  tracking
• 7:26
  that
• 7:26
  Mamba-2
• 7:27
  simply
• 7:27
  cannot
• 7:27
  do.
• 7:28
  The
• 7:28
  authors
• 7:29
  demonstrate
• 7:29
  this
• 7:29
  concretely:
• 7:30
  Mamba-3
• 7:31
  near-perfectly
• 7:31
  solves
• 7:32
  arithmetic
• 7:32
  tasks
• 7:33
  where
• 7:33
  Mamba-2
• 7:33
  performs
• 7:34
  no
• 7:34
  better
• 7:34
  than
• 7:34
  random
• 7:34
  guessing.
• 7:35
  Third,
• 7:36
  a
• 7:36
  multi-input
• 7:37
  multi-output
• 7:37
  (MIMO)
• 7:37
  formulation
• 7:38
  increases
• 7:39
  decoding
• 7:39
  FLOPs
• 7:39
  by
• 7:40
  up
• 7:40
  to
• 7:40
  4x
• 7:40
  at
• 7:40
  fixed
• 7:41
  state
• 7:41
  size
• 7:41
  while
• 7:41
  maintaining
• 7:42
  similar
• 7:42
  wall-clock
• 7:43
  decode
• 7:43
  latency.
• 7:44
  At
• 7:44
  1.5B
• 7:45
  scale,
• 7:46
  Mamba-3
• 7:46
  MIMO
• 7:47
  improves
• 7:47
  average
• 7:48
  downstream
• 7:48
  accuracy
• 7:48
  by
• 7:49
  2.2
• 7:49
  points
• 7:49
  over
• 7:50
  Transformers,
• 7:51
  1.9
• 7:51
  over
• 7:52
  Mamba-2,
• 7:52
  and
• 7:53
  1.8
• 7:53
  over
• 7:53
  Gated
• 7:54
  DeltaNet.
• 7:55
  Perhaps
• 7:56
  more
• 7:56
  telling:
• 7:57
  Mamba-3
• 7:58
  MIMO
• 7:58
  with
• 7:58
  state
• 7:59
  size
• 7:59
  64
• 7:59
  matches
• 8:00
  Mamba-2's
• 8:01
  perplexity
• 8:01
  at
• 8:01
  state
• 8:01
  size
• 8:02
  128
• 8:02
  --
• 8:03
  same
• 8:03
  performance
• 8:04
  at
• 8:04
  half
• 8:04
  the
• 8:04
  memory.
• 8:05
  At
• 8:05
  the
• 8:05
  other
• 8:05
  end
• 8:05
  of
• 8:06
  the
• 8:06
  parameter
• 8:06
  scale,
• 8:06
  TinyLoRA
• 8:07
  demonstrates
• 8:08
  that
• 8:08
  LoRA
• 8:08
  training
• 8:09
  works
• 8:09
  with
• 8:09
  as
• 8:09
  few
• 8:09
  as
• 8:09
  13
• 8:10
  parameters.
• 8:11
  A
• 8:11
  researcher
• 8:12
  replicated
• 8:12
  the
• 8:13
  paper's
• 8:13
  results
• 8:13
  on
• 8:13
  Qwen
• 8:14
  3.5
• 8:15
  and
• 8:15
  found
• 8:15
  that
• 8:15
  separating
• 8:16
  the
• 8:16
  13
• 8:16
  parameters
• 8:17
  between
• 8:17
  MLP
• 8:18
  and
• 8:18
  attention
• 8:18
  layers
• 8:19
  (26
• 8:19
  total)
• 8:20
  outperformed
• 8:20
  increasing
• 8:21
  the
• 8:21
  global
• 8:21
  parameter
• 8:22
  count
• 8:22
  --
• 8:23
  suggesting
• 8:23
  these
• 8:23
  layers
• 8:24
  benefit
• 8:24
  from
• 8:24
  separate
• 8:25
  optimization
• 8:25
  spaces
• 8:26
  since
• 8:26
  they
• 8:26
  handle
• 8:26
  fundamentally
• 8:27
  different
• 8:27
  operations.
• 8:29
  The
• 8:29
  practical
• 8:29
  hypothesis:
• 8:30
  TinyLoRA
• 8:31
  won't
• 8:31
  memorize
• 8:32
  facts,
• 8:33
  but
• 8:33
  it's
• 8:33
  effective
• 8:33
  at
• 8:33
  altering
• 8:34
  behavior,
• 8:35
  making
• 8:35
  it
• 8:35
  potentially
• 8:36
  useful
• 8:36
  for
• 8:36
  per-user
• 8:37
  behavioral
• 8:37
  adapters.
• 8:38
  Combined
• 8:39
  with
• 8:39
  DeepSeek's
• 8:39
  Engram
• 8:40
  concept,
• 8:41
  this
• 8:41
  could
• 8:41
  enable
• 8:41
  lookup
• 8:42
  tables
• 8:42
  for
• 8:42
  behaviors
• 8:43
  made
• 8:43
  of
• 8:43
  tiny
• 8:43
  LoRA
• 8:44
  adapters
• 8:44
  --
• 8:45
  far
• 8:45
  more
• 8:46
  varied
• 8:46
  than
• 8:46
  mixture-of-experts,
• 8:47
  trainable
• 8:48
  with
• 8:48
  minimal
• 8:48
  memory.
• 8:49
  Related:
• 8:50
  a
• 8:50
  KV
• 8:51
  rotation
• 8:51
  PR
• 8:52
  for
• 8:52
  llama.cpp
• 8:54
  found
• 8:54
  that
• 8:55
  existing
• 8:55
  q8
• 8:56
  KV
• 8:56
  quantization
• 8:57
  tanks
• 8:57
  performance
• 8:57
  on
• 8:57
  AIME25,
• 9:00
  but
• 9:00
  applying
• 9:01
  rotation
• 9:01
  largely
• 9:02
  recovers
• 9:02
  it
• 9:02
  --
• 9:03
  a
• 9:03
  useful
• 9:04
  datapoint
• 9:04
  for
• 9:04
  anyone
• 9:05
  running
• 9:05
  quantized
• 9:06
  inference
• 9:06
  and
• 9:06
  wondering
• 9:07
  why
• 9:07
  their
• 9:07
  math
• 9:07
  benchmarks
• 9:08
  look
• 9:08
  worse
• 9:08
  than
• 9:08
  expected.
• 9:10
  Taalas
• 9:10
  is
• 9:10
  taking
• 9:11
  a
• 9:11
  more
• 9:11
  radical
• 9:11
  approach
• 9:12
  to
• 9:12
  inference
• 9:12
  efficiency:
• 9:13
  burning
• 9:13
  language
• 9:14
  models
• 9:14
  directly
• 9:15
  into
• 9:15
  silicon.
• 9:16
  Their
• 9:17
  ASIC
• 9:17
  card
• 9:17
  reportedly
• 9:18
  processes
• 9:18
  Llama
• 9:19
  3.1
• 9:19
  8B
• 9:20
  at
• 9:20
  17,000
• 9:21
  tokens
• 9:22
  per
• 9:22
  second
• 9:22
  --
• 9:23
  compare
• 9:23
  that
• 9:23
  to
• 9:23
  an
• 9:24
  RTX
• 9:24
  4090
• 9:25
  at
• 9:26
  roughly
• 9:26
  150
• 9:27
  tok/s
• 9:28
  or
• 9:28
  Groq
• 9:28
  at
• 9:28
  around
• 9:28
  1,000.
• 9:30
  A
• 9:30
  new
• 9:30
  medium-sized
• 9:31
  PCIe
• 9:32
  card
• 9:32
  will
• 9:32
  burn
• 9:32
  Qwen
• 9:33
  3.5
• 9:33
  27B
• 9:34
  into
• 9:34
  silicon:
• 9:35
  standard
• 9:36
  PC
• 9:36
  slot,
• 9:37
  10x
• 9:37
  less
• 9:38
  power
• 9:38
  than
• 9:38
  a
• 9:38
  GPU,
• 9:39
  no
• 9:39
  VRAM,
• 9:40
  no
• 9:40
  weights
• 9:40
  loading,
• 9:41
  no
• 9:41
  inference
• 9:41
  framework.
• 9:42
  The
• 9:42
  model
• 9:43
  is
• 9:43
  the
• 9:43
  hardware.
• 9:44
  The
• 9:44
  tradeoffs
• 9:45
  are
• 9:45
  obvious
• 9:45
  and
• 9:45
  acknowledged
• 9:46
  in
• 9:46
  the
• 9:46
  comments:
• 9:47
  you
• 9:47
  can't
• 9:47
  swap
• 9:48
  models
• 9:48
  on
• 9:48
  a
• 9:48
  burned
• 9:49
  chip,
• 9:49
  every
• 9:50
  revision
• 9:50
  requires
• 9:50
  a
• 9:51
  new
• 9:51
  mask
• 9:51
  spin,
• 9:52
  and
• 9:52
  the
• 9:52
  model
• 9:52
  is
• 9:53
  outdated
• 9:53
  the
• 9:53
  moment
• 9:53
  the
• 9:54
  card
• 9:54
  leaves
• 9:54
  the
• 9:54
  factory
• 9:54
  line.
• 9:55
  For
• 9:56
  regulated
• 9:56
  industries
• 9:57
  operating
• 9:57
  under
• 9:57
  HIPAA,
• 9:58
  ITAR,
• 9:58
  or
• 9:59
  CMMC,
• 10:00
  the
• 10:00
  deployment
• 10:00
  model
• 10:01
  --
• 10:02
  air-cooled,
• 10:02
  250W,
• 10:03
  no
• 10:04
  cloud
• 10:04
  dependency,
• 10:05
  no
• 10:05
  data
• 10:05
  leaving
• 10:06
  the
• 10:06
  building
• 10:06
  --
• 10:06
  might
• 10:07
  matter
• 10:07
  more
• 10:07
  than
• 10:07
  flexibility.
• 10:08
  Heretic
• 10:09
  automates
• 10:09
  censorship
• 10:10
  removal
• 10:10
  from
• 10:10
  language
• 10:11
  models
• 10:11
  using
• 10:11
  directional
• 10:12
  ablation
• 10:12
  with
• 10:12
  a
• 10:12
  TPE-based
• 10:13
  parameter
• 10:13
  optimizer.
• 10:15
  The
• 10:15
  tool
• 10:15
  finds
• 10:15
  high-quality
• 10:16
  abliteration
• 10:16
  parameters
• 10:17
  by
• 10:17
  co-minimizing
• 10:18
  refusals
• 10:18
  and
• 10:18
  KL
• 10:19
  divergence
• 10:19
  from
• 10:19
  the
• 10:19
  original
• 10:20
  model,
• 10:21
  producing
• 10:21
  decensored
• 10:21
  models
• 10:22
  that
• 10:22
  retain
• 10:22
  as
• 10:22
  much
• 10:23
  original
• 10:23
  intelligence
• 10:24
  as
• 10:24
  possible.
• 10:25
  On
• 10:25
  Gemma
• 10:25
  3
• 10:26
  12B,
• 10:26
  the
• 10:27
  Heretic
• 10:27
  version
• 10:27
  achieves
• 10:28
  the
• 10:28
  same
• 10:28
  refusal
• 10:28
  suppression
• 10:29
  as
• 10:29
  expert
• 10:29
  manual
• 10:30
  abliterations
• 10:30
  (3/100
• 10:31
  refusals)
• 10:32
  but
• 10:32
  at
• 10:32
  a
• 10:32
  KL
• 10:32
  divergence
• 10:33
  of
• 10:33
  0.16
• 10:34
  versus
• 10:34
  0.45-1.04
• 10:38
  for
• 10:39
  other
• 10:39
  approaches
• 10:39
  --
• 10:40
  less
• 10:40
  damage
• 10:41
  to
• 10:41
  the
• 10:41
  model's
• 10:41
  capabilities.
• 10:42
  The
• 10:43
  community
• 10:43
  has
• 10:43
  published
• 10:43
  over
• 10:44
  1,000
• 10:44
  Heretic
• 10:44
  models
• 10:45
  on
• 10:45
  Hugging
• 10:45
  Face.
• 10:46
  Key
• 10:47
  innovations
• 10:47
  include
• 10:48
  float-valued
• 10:48
  refusal
• 10:49
  direction
• 10:49
  indices
• 10:50
  (interpolating
• 10:50
  between
• 10:50
  layer
• 10:51
  directions
• 10:52
  unlocks
• 10:52
  a
• 10:52
  vast
• 10:52
  space
• 10:53
  beyond
• 10:53
  those
• 10:53
  identified
• 10:54
  by
• 10:54
  difference-of-means)
• 10:55
  and
• 10:55
  separate
• 10:55
  ablation
• 10:56
  weights
• 10:56
  for
• 10:56
  MLP
• 10:57
  versus
• 10:57
  attention
• 10:57
  components,
• 10:58
  since
• 10:59
  MLP
• 10:59
  interventions
• 11:00
  tend
• 11:00
  to
• 11:00
  be
• 11:00
  more
• 11:00
  damaging.
• 11:01
  On
• 11:01
  Apple
• 11:02
  Silicon,
• 11:02
  mlx-snn
• 11:03
  delivers
• 11:03
  a
• 11:03
  spiking
• 11:04
  neural
• 11:04
  network
• 11:04
  library
• 11:05
  for
• 11:05
  the
• 11:05
  MLX
• 11:05
  framework
• 11:06
  that
• 11:06
  benchmarks
• 11:07
  at
• 11:07
  2.6-3.7x
• 11:08
  faster
• 11:09
  training
• 11:09
  than
• 11:09
  Tesla
• 11:09
  V100
• 11:10
  at
• 11:10
  one-seventh
• 11:11
  the
• 11:11
  power
• 11:11
  --
• 11:12
  an
• 11:12
  estimated
• 11:13
  17-25x
• 11:14
  better
• 11:14
  energy
• 11:14
  efficiency.
• 11:16
  The
• 11:16
  library
• 11:16
  ships
• 11:16
  9
• 11:17
  neuron
• 11:17
  models,
• 11:18
  6
• 11:18
  surrogate
• 11:18
  gradients,
• 11:19
  8
• 11:19
  spike
• 11:19
  encodings,
• 11:20
  5
• 11:20
  neuromorphic
• 11:20
  datasets,
• 11:21
  and
• 11:21
  a
• 11:22
  liquid
• 11:22
  state
• 11:22
  machine
• 11:22
  implementation
• 11:23
  with
• 11:23
  configurable
• 11:24
  topology.
• 11:25
  For
• 11:25
  researchers
• 11:25
  working
• 11:26
  on
• 11:26
  neuromorphic
• 11:26
  computing,
• 11:27
  the
• 11:27
  unified
• 11:28
  memory
• 11:28
  architecture
• 11:28
  eliminates
• 11:29
  the
• 11:29
  CPU-GPU
• 11:30
  transfer
• 11:31
  bottleneck
• 11:31
  that
• 11:31
  makes
• 11:31
  recurrent
• 11:32
  spiking
• 11:32
  dynamics
• 11:33
  painful
• 11:33
  on
• 11:33
  discrete
• 11:33
  GPUs.
• 11:36
  An
• 11:36
  essay
• 11:37
  on
• 11:37
  CEO
• 11:37
  accountability
• 11:38
  for
• 11:38
  AI
• 11:38
  makes
• 11:39
  an
• 11:39
  argument
• 11:39
  that
• 11:39
  cuts
• 11:39
  deeper
• 11:40
  than
• 11:40
  the
• 11:40
  usual
• 11:40
  "leaders
• 11:41
  need
• 11:41
  to
• 11:41
  understand
• 11:41
  the
• 11:41
  technology"
• 11:42
  advice.
• 11:43
  The
• 11:43
  real
• 11:43
  divide
• 11:44
  is
• 11:44
  not
• 11:44
  between
• 11:44
  companies
• 11:45
  with
• 11:45
  AI
• 11:45
  pilots
• 11:46
  and
• 11:46
  companies
• 11:46
  without
• 11:47
  them.
• 11:48
  It's
• 11:48
  between
• 11:48
  firms
• 11:48
  generating
• 11:49
  AI
• 11:49
  motion
• 11:50
  and
• 11:50
  firms
• 11:50
  producing
• 11:51
  board-grade
• 11:51
  proof
• 11:51
  that
• 11:51
  the
• 11:52
  economics
• 11:52
  hold
• 11:52
  once
• 11:52
  the
• 11:52
  system
• 11:53
  becomes
• 11:53
  real.
• 11:54
  The
• 11:54
  author
• 11:55
  draws
• 11:55
  a
• 11:55
  sharp
• 11:55
  distinction
• 11:56
  between
• 11:56
  ownership
• 11:56
  (a
• 11:56
  claim
• 11:57
  about
• 11:57
  accountability)
• 11:58
  and
• 11:58
  control
• 11:58
  (an
• 11:58
  operating
• 11:59
  condition):
• 12:00
  a
• 12:00
  CEO
• 12:00
  can
• 12:00
  find
• 12:01
  themselves
• 12:01
  answerable
• 12:02
  for
• 12:02
  a
• 12:02
  system
• 12:02
  whose
• 12:02
  effective
• 12:03
  shape
• 12:03
  is
• 12:03
  still
• 12:03
  being
• 12:04
  determined
• 12:04
  across
• 12:04
  product
• 12:05
  teams,
• 12:05
  IT,
• 12:06
  procurement,
• 12:06
  prompt
• 12:07
  layers,
• 12:07
  data
• 12:08
  pipelines,
• 12:08
  permissions,
• 12:09
  and
• 12:09
  vendor
• 12:09
  configurations.
• 12:11
  The
• 12:11
  first
• 12:11
  serious
• 12:11
  AI
• 12:12
  disappointment
• 12:12
  at
• 12:12
  executive
• 12:13
  level
• 12:13
  is
• 12:13
  rarely
• 12:13
  cinematic
• 12:14
  --
• 12:15
  more
• 12:15
  often
• 12:15
  the
• 12:15
  promised
• 12:15
  savings
• 12:16
  don't
• 12:16
  quite
• 12:16
  land,
• 12:17
  the
• 12:17
  workflow
• 12:17
  needs
• 12:18
  more
• 12:18
  review
• 12:18
  than
• 12:18
  expected,
• 12:19
  exceptions
• 12:20
  pile
• 12:20
  up
• 12:20
  in
• 12:20
  places
• 12:20
  never
• 12:21
  properly
• 12:21
  priced,
• 12:22
  and
• 12:22
  the
• 12:22
  team
• 12:22
  is
• 12:22
  busier
• 12:23
  than
• 12:23
  before
• 12:23
  though
• 12:24
  nominally
• 12:24
  more
• 12:24
  productive.
• 12:25
  The
• 12:25
  prescription:
• 12:26
  pick
• 12:27
  a
• 12:27
  small
• 12:27
  number
• 12:27
  of
• 12:27
  workflows
• 12:28
  that
• 12:28
  genuinely
• 12:28
  matter
• 12:29
  to
• 12:29
  the
• 12:29
  board,
• 12:30
  put
• 12:30
  a
• 12:30
  mixed
• 12:30
  team
• 12:31
  with
• 12:31
  real
• 12:31
  operating
• 12:32
  authority
• 12:32
  around
• 12:32
  them,
• 12:33
  and
• 12:33
  prove
• 12:33
  the
• 12:33
  case
• 12:34
  under
• 12:34
  live
• 12:34
  conditions
• 12:35
  before
• 12:35
  broader
• 12:36
  expansion.
• 12:37
  Price
• 12:38
  supervision,
• 12:38
  reversals,
• 12:39
  rework,
• 12:40
  and
• 12:40
  exception
• 12:41
  handling
• 12:41
  into
• 12:41
  the
• 12:41
  economics
• 12:42
  instead
• 12:42
  of
• 12:43
  treating
• 12:43
  them
• 12:43
  as
• 12:43
  temporary
• 12:44
  noise.
• 12:45
  The
• 12:45
  companion
• 12:45
  piece
• 12:46
  on
• 12:46
  AI
• 12:46
  and
• 12:46
  work
• 12:47
  argues
• 12:47
  that
• 12:47
  smoothness
• 12:48
  is
• 12:48
  not
• 12:48
  the
• 12:48
  same
• 12:48
  as
• 12:48
  relief.
• 12:49
  AI
• 12:50
  lowers
• 12:50
  the
• 12:50
  cost
• 12:50
  of
• 12:50
  starting,
• 12:51
  shortens
• 12:51
  the
• 12:51
  path
• 12:52
  from
• 12:52
  rough
• 12:52
  material
• 12:52
  to
• 12:53
  plausible
• 12:53
  output,
• 12:54
  and
• 12:54
  carries
• 12:54
  context
• 12:55
  further
• 12:55
  across
• 12:55
  boundaries.
• 12:56
  But
• 12:56
  once
• 12:57
  friction
• 12:57
  falls,
• 12:58
  more
• 12:58
  gets
• 12:58
  started
• 12:58
  because
• 12:59
  starting
• 12:59
  is
• 12:59
  easier.
• 13:00
  More
• 13:00
  gets
• 13:01
  sent
• 13:01
  because
• 13:01
  drafting
• 13:02
  is
• 13:02
  easier.
• 13:03
  People
• 13:03
  widen
• 13:04
  scope
• 13:04
  because
• 13:04
  the
• 13:04
  tools
• 13:05
  make
• 13:05
  stretching
• 13:05
  feel
• 13:06
  possible,
• 13:07
  and
• 13:07
  then
• 13:07
  that
• 13:07
  extra
• 13:07
  range
• 13:08
  hardens
• 13:08
  into
• 13:08
  the
• 13:08
  new
• 13:09
  normal.
• 13:10
  The
• 13:10
  work
• 13:10
  may
• 13:10
  ask
• 13:11
  less
• 13:11
  of
• 13:11
  our
• 13:11
  hands
• 13:11
  while
• 13:12
  asking
• 13:12
  more
• 13:12
  of
• 13:12
  our
• 13:12
  name
• 13:13
  --
• 13:14
  less
• 13:14
  time
• 13:14
  producing,
• 13:15
  more
• 13:15
  time
• 13:16
  endorsing,
• 13:16
  approving,
• 13:17
  checking,
• 13:18
  standing
• 13:18
  behind.
• 13:20
  The
• 13:20
  author
• 13:20
  invokes
• 13:21
  Donald
• 13:21
  Schon:
• 13:22
  much
• 13:22
  of
• 13:22
  real
• 13:23
  practice
• 13:23
  is
• 13:23
  discovering
• 13:24
  what
• 13:24
  the
• 13:24
  problem
• 13:24
  actually
• 13:25
  is
• 13:25
  while
• 13:25
  you're
• 13:25
  still
• 13:26
  inside
• 13:26
  it,
• 13:27
  and
• 13:27
  AI
• 13:27
  help
• 13:28
  may
• 13:28
  arrive
• 13:28
  too
• 13:28
  quickly,
• 13:29
  giving
• 13:29
  form
• 13:30
  to
• 13:30
  thought
• 13:30
  before
• 13:30
  thought
• 13:31
  has
• 13:31
  finished
• 13:31
  finding
• 13:31
  its
• 13:32
  own
• 13:32
  shape.
• 13:33
  The
• 13:33
  day
• 13:33
  becomes
• 13:33
  more
• 13:34
  capable
• 13:34
  and
• 13:34
  less
• 13:35
  recoverable.
• 13:36
  A
• 13:36
  data-driven
• 13:37
  post
• 13:37
  on
• 13:37
  Claude
• 13:37
  Code's
• 13:38
  open
• 13:38
  bug
• 13:38
  count
• 13:39
  finds
• 13:39
  5,021
• 13:40
  bugs
• 13:41
  with
• 13:41
  the
• 13:41
  gap
• 13:41
  between
• 13:41
  raised
• 13:42
  and
• 13:42
  closed
• 13:42
  growing
• 13:43
  at
• 13:43
  a
• 13:43
  consistent
• 13:44
  percentage
• 13:44
  --
• 13:45
  meaning
• 13:45
  Anthropic
• 13:46
  is
• 13:46
  permanently
• 13:46
  running
• 13:47
  behind
• 13:47
  at
• 13:47
  a
• 13:47
  steady
• 13:48
  rate
• 13:48
  with
• 13:48
  no
• 13:48
  sign
• 13:49
  of
• 13:49
  catching
• 13:49
  up.
• 13:50
  The
• 13:50
  author's
• 13:51
  point
• 13:51
  isn't
• 13:51
  that
• 13:52
  AI
• 13:52
  tools
• 13:52
  are
• 13:52
  bad
• 13:53
  (his
• 13:53
  team
• 13:53
  uses
• 13:53
  Claude
• 13:54
  Code
• 13:54
  daily
• 13:54
  and
• 13:55
  ships
• 13:55
  aggressively
• 13:55
  with
• 13:55
  it),
• 13:56
  but
• 13:56
  that
• 13:56
  the
• 13:56
  bounded
• 13:57
  nature
• 13:57
  of
• 13:57
  the
• 13:57
  product
• 13:57
  --
• 13:58
  essentially
• 13:59
  a
• 13:59
  while-loop,
• 14:00
  an
• 14:00
  LLM
• 14:01
  connection,
• 14:01
  and
• 14:01
  four
• 14:02
  filesystem
• 14:02
  tools
• 14:02
  --
• 14:03
  makes
• 14:04
  5,000
• 14:04
  open
• 14:05
  bugs
• 14:05
  harder
• 14:05
  to
• 14:05
  excuse.
• 14:07
  The
• 14:07
  quadratic
• 14:07
  projection
• 14:08
  (R²=0.986)
• 14:10
  is
• 14:10
  concerning.
• 14:11
  If
• 14:11
  Anthropic,
• 14:12
  with
• 14:13
  unlimited
• 14:13
  access
• 14:14
  to
• 14:14
  its
• 14:14
  own
• 14:14
  AI
• 14:14
  for
• 14:15
  triage
• 14:15
  and
• 14:15
  fixes,
• 14:16
  can't
• 14:16
  control
• 14:16
  quality
• 14:17
  on
• 14:17
  its
• 14:17
  own
• 14:17
  codebase,
• 14:18
  what
• 14:18
  does
• 14:18
  that
• 14:19
  say
• 14:19
  about
• 14:19
  the
• 14:19
  "AI
• 14:20
  writes
• 14:20
  all
• 14:20
  our
• 14:20
  code"
• 14:20
  thesis?
• 14:22
  The
• 14:22
  commenters
• 14:22
  push
• 14:23
  back:
• 14:24
  many
• 14:24
  bugs
• 14:24
  are
• 14:24
  likely
• 14:25
  duplicates,
• 14:25
  feature
• 14:26
  requests
• 14:26
  mislabeled,
• 14:27
  or
• 14:27
  edge
• 14:27
  cases
• 14:28
  from
• 14:28
  non-technical
• 14:28
  users.
• 14:29
  Fair
• 14:30
  enough.
• 14:31
  But
• 14:31
  one
• 14:31
  reply
• 14:31
  cuts
• 14:32
  to
• 14:32
  it:
• 14:32
  "Either
• 14:33
  the
• 14:33
  tool
• 14:33
  cannot
• 14:33
  do
• 14:34
  what
• 14:34
  it
• 14:34
  says
• 14:34
  and
• 14:35
  creates
• 14:35
  a
• 14:35
  lot
• 14:35
  of
• 14:35
  bugs,
• 14:36
  or
• 14:36
  it
• 14:36
  cannot
• 14:37
  effectively
• 14:37
  triage
• 14:38
  and
• 14:38
  classify
• 14:38
  its
• 14:38
  own
• 14:39
  bugs.
• 14:39
  No
• 14:39
  matter
• 14:40
  which
• 14:40
  way
• 14:40
  you
• 14:40
  slice
• 14:40
  it,
• 14:41
  it
• 14:41
  looks
• 14:41
  bad
• 14:41
  for
• 14:41
  a
• 14:41
  company
• 14:42
  selling
• 14:42
  an
• 14:42
  AI
• 14:43
  dream."
• 14:44
  An
• 14:44
  essay
• 14:44
  on
• 14:44
  leadership
• 14:45
  availability
• 14:45
  --
• 14:46
  the
• 14:46
  practice
• 14:47
  of
• 14:47
  being
• 14:47
  reachable,
• 14:48
  present,
• 14:48
  engaged,
• 14:49
  and
• 14:49
  following
• 14:50
  through
• 14:50
  --
• 14:51
  lands
• 14:51
  differently
• 14:52
  when
• 14:52
  read
• 14:52
  alongside
• 14:52
  the
• 14:52
  AI
• 14:53
  governance
• 14:53
  pieces.
• 14:54
  The
• 14:54
  author
• 14:55
  breaks
• 14:55
  availability
• 14:55
  into
• 14:56
  four
• 14:56
  habits:
• 14:57
  making
• 14:57
  time,
• 14:57
  making
• 14:58
  space,
• 14:58
  making
• 14:59
  decisions,
• 14:59
  and
• 14:59
  closing
• 15:00
  loops.
• 15:00
  Chronic
• 15:01
  unavailability
• 15:01
  teaches
• 15:02
  teams
• 15:02
  to
• 15:02
  think
• 15:02
  smaller,
• 15:03
  route
• 15:03
  around
• 15:03
  their
• 15:04
  leader,
• 15:04
  and
• 15:04
  stop
• 15:05
  surfacing
• 15:05
  problems.
• 15:06
  The
• 15:06
  compounding
• 15:07
  effect
• 15:07
  is
• 15:07
  corrosive:
• 15:08
  a
• 15:08
  "no"
• 15:09
  is
• 15:09
  at
• 15:09
  least
• 15:09
  a
• 15:09
  decision,
• 15:10
  but
• 15:10
  silence
• 15:11
  is
• 15:11
  a
• 15:11
  tax
• 15:11
  on
• 15:11
  everyone
• 15:12
  else's
• 15:12
  time.
• 15:13
  The
• 15:13
  connection
• 15:13
  to
• 15:13
  AI
• 15:14
  governance
• 15:14
  is
• 15:14
  implicit
• 15:15
  but
• 15:15
  real:
• 15:16
  the
• 15:16
  boards
• 15:16
  asking
• 15:17
  "does
• 15:17
  AI
• 15:17
  work?"
• 15:18
  are
• 15:18
  often
• 15:19
  led
• 15:19
  by
• 15:19
  executives
• 15:20
  whose
• 15:20
  follow-through
• 15:21
  problem
• 15:21
  predates
• 15:21
  their
• 15:22
  AI
• 15:22
  strategy.
• 15:23
  Governance
• 15:24
  requires
• 15:24
  the
• 15:24
  exact
• 15:25
  trait
• 15:25
  --
• 15:25
  closing
• 15:26
  loops,
• 15:27
  making
• 15:27
  decisions
• 15:27
  when
• 15:27
  decisions
• 15:28
  are
• 15:28
  needed,
• 15:29
  pricing
• 15:29
  messy
• 15:29
  reality
• 15:30
  into
• 15:30
  the
• 15:30
  story
• 15:30
  --
• 15:31
  that
• 15:31
  too
• 15:31
  many
• 15:31
  leaders
• 15:32
  have
• 15:32
  already
• 15:32
  outsourced
• 15:33
  to
• 15:33
  calendar
• 15:33
  protection
• 15:34
  and
• 15:34
  empowerment
• 15:35
  rhetoric.
• 15:37
  A
• 15:37
  developer
• 15:38
  essay
• 15:38
  adapts
• 15:39
  Liu
• 15:39
  Cixin's
• 15:39
  Dark
• 15:40
  Forest
• 15:40
  hypothesis
• 15:41
  to
• 15:41
  the
• 15:41
  AI
• 15:41
  era.
• 15:42
  The
• 15:42
  original:
• 15:43
  every
• 15:44
  civilization
• 15:44
  that
• 15:44
  reveals
• 15:45
  itself
• 15:45
  gets
• 15:45
  annihilated,
• 15:46
  so
• 15:46
  they
• 15:46
  all
• 15:46
  hide.
• 15:47
  The
• 15:47
  cognitive
• 15:48
  version:
• 15:48
  the
• 15:48
  earlier
• 15:49
  internet
• 15:49
  rewarded
• 15:49
  signaling
• 15:50
  --
• 15:51
  your
• 15:51
  value
• 15:51
  multiplied
• 15:51
  by
• 15:52
  connections,
• 15:52
  and
• 15:52
  execution
• 15:53
  was
• 15:53
  the
• 15:53
  moat.
• 15:54
  Two
• 15:54
  things
• 15:54
  changed.
• 15:56
  The
• 15:56
  web
• 15:56
  got
• 15:56
  consolidated
• 15:57
  by
• 15:57
  corporations
• 15:58
  extracting
• 15:58
  information
• 15:59
  to
• 15:59
  advertise
• 15:59
  and
• 15:59
  governments
• 16:00
  killing
• 16:00
  privacy
• 16:01
  to
• 16:01
  control.
• 16:02
  And
• 16:02
  execution
• 16:03
  got
• 16:03
  cheap.
• 16:04
  Now
• 16:05
  a
• 16:05
  large
• 16:05
  platform
• 16:05
  with
• 16:05
  cash
• 16:06
  can
• 16:06
  incorporate
• 16:06
  your
• 16:07
  innovations
• 16:07
  by
• 16:07
  throwing
• 16:08
  compute
• 16:08
  at
• 16:08
  the
• 16:08
  problem,
• 16:09
  generating
• 16:09
  variations
• 16:10
  every
• 16:10
  few
• 16:10
  days
• 16:11
  until
• 16:11
  your
• 16:11
  uniqueness
• 16:11
  is
• 16:11
  absorbed.
• 16:13
  Every
• 16:13
  prompt
• 16:13
  flows
• 16:14
  through
• 16:14
  centralized
• 16:15
  platforms.
• 16:16
  Every
• 16:16
  prompt
• 16:17
  is
• 16:17
  a
• 16:17
  signal
• 16:17
  revealing
• 16:18
  intent.
• 16:19
  The
• 16:19
  platform
• 16:20
  doesn't
• 16:20
  need
• 16:20
  to
• 16:20
  spy
• 16:20
  on
• 16:20
  you
• 16:21
  specifically
• 16:21
  --
• 16:22
  it
• 16:22
  just
• 16:22
  needs
• 16:23
  to
• 16:23
  see
• 16:23
  where
• 16:23
  the
• 16:23
  questions
• 16:23
  cluster.
• 16:25
  A
• 16:25
  demand
• 16:25
  curve
• 16:25
  made
• 16:26
  of
• 16:26
  human
• 16:26
  interests.
• 16:27
  The
• 16:27
  true
• 16:28
  horror:
• 16:29
  the
• 16:29
  cognitive
• 16:29
  dark
• 16:29
  forest
• 16:30
  doesn't
• 16:30
  kill
• 16:30
  you.
• 16:31
  It
• 16:31
  lets
• 16:31
  you
• 16:32
  live
• 16:32
  and
• 16:32
  feeds
• 16:32
  on
• 16:32
  you.
• 16:33
  Your
• 16:33
  innovation
• 16:34
  becomes
• 16:34
  its
• 16:34
  capabilities.
• 16:36
  Your
• 16:36
  differentiation
• 16:36
  becomes
• 16:37
  its
• 16:37
  median.
• 16:38
  Resistance
• 16:38
  isn't
• 16:39
  suppressed
• 16:39
  --
• 16:40
  it's
• 16:40
  absorbed.
• 16:41
  The
• 16:41
  pharmacovigilance
• 16:42
  weekend
• 16:42
  build
• 16:43
  is
• 16:43
  both
• 16:43
  a
• 16:43
  case
• 16:43
  study
• 16:43
  in
• 16:44
  what
• 16:44
  AI
• 16:44
  enables
• 16:44
  and
• 16:45
  what
• 16:45
  the
• 16:45
  dark
• 16:45
  forest
• 16:45
  predicts.
• 16:46
  A
• 16:47
  developer
• 16:47
  used
• 16:47
  Claude
• 16:47
  Code
• 16:48
  to
• 16:48
  build
• 16:48
  a
• 16:48
  free
• 16:48
  tool
• 16:49
  that
• 16:49
  runs
• 16:49
  signal
• 16:49
  detection
• 16:50
  across
• 16:50
  2.9
• 16:51
  million
• 16:51
  FDA
• 16:51
  adverse
• 16:52
  event
• 16:52
  reports
• 16:52
  --
• 16:53
  disproportionality
• 16:54
  statistics
• 16:54
  (PRR,
• 16:55
  ROR,
• 16:56
  chi-squared)
• 16:56
  that
• 16:56
  commercial
• 16:57
  platforms
• 16:57
  charge
• 16:57
  $50K-500K/year
• 16:59
  for.
• 17:00
  The
• 17:00
  data
• 17:00
  is
• 17:00
  free,
• 17:01
  the
• 17:01
  methods
• 17:02
  are
• 17:02
  published,
• 17:02
  the
• 17:02
  barrier
• 17:03
  was
• 17:03
  always
• 17:03
  the
• 17:03
  engineering.
• 17:04
  The
• 17:05
  community
• 17:05
  response
• 17:05
  was
• 17:05
  brutal:
• 17:06
  the
• 17:06
  first
• 17:07
  outside
• 17:07
  user
• 17:07
  typed
• 17:08
  "ibuprofen"
• 17:09
  and
• 17:09
  got
• 17:09
  a
• 17:09
  Python
• 17:09
  stack
• 17:10
  trace.
• 17:11
  Comments
• 17:11
  ranged
• 17:11
  from
• 17:12
  "vibecoded
• 17:12
  garbage"
• 17:13
  to
• 17:13
  "have
• 17:13
  a
• 17:13
  think
• 17:13
  about
• 17:13
  the
• 17:13
  consequences
• 17:14
  to
• 17:14
  real
• 17:14
  people
• 17:15
  when
• 17:15
  pharmacovigilance
• 17:15
  processes
• 17:16
  fail
• 17:16
  --
• 17:17
  that's
• 17:17
  why
• 17:17
  the
• 17:17
  software
• 17:18
  is
• 17:18
  expensive."
• 17:19
  The
• 17:19
  author
• 17:20
  acknowledged
• 17:20
  it's
• 17:20
  a
• 17:20
  proof
• 17:21
  of
• 17:21
  concept,
• 17:21
  not
• 17:22
  a
• 17:22
  replacement.
• 17:23
  But
• 17:23
  the
• 17:23
  pattern
• 17:24
  is
• 17:24
  instructive:
• 17:25
  AI
• 17:25
  compresses
• 17:26
  the
• 17:26
  distance
• 17:26
  between
• 17:26
  "I
• 17:26
  can
• 17:27
  build
• 17:27
  this"
• 17:27
  and
• 17:27
  "I
• 17:27
  shipped
• 17:28
  this"
• 17:28
  while
• 17:28
  doing
• 17:29
  nothing
• 17:29
  about
• 17:29
  the
• 17:29
  distance
• 17:29
  between
• 17:30
  "I
• 17:30
  shipped
• 17:30
  this"
• 17:31
  and
• 17:31
  "this
• 17:31
  works
• 17:31
  reliably
• 17:32
  under
• 17:32
  production
• 17:32
  conditions."
• 17:34
  Google
• 17:34
  Stitch
• 17:35
  is
• 17:35
  generating
• 17:35
  similar
• 17:36
  reactions
• 17:36
  in
• 17:36
  the
• 17:36
  design
• 17:37
  space
• 17:37
  --
• 17:38
  users
• 17:38
  following
• 17:39
  a
• 17:39
  guide
• 17:39
  and
• 17:39
  producing
• 17:40
  functional
• 17:40
  UI
• 17:41
  outputs
• 17:41
  in
• 17:41
  minutes,
• 17:42
  prompting
• 17:42
  comments
• 17:43
  like
• 17:43
  "my
• 17:44
  job
• 17:44
  is
• 17:44
  officially
• 17:44
  cooked."
• 17:46
  The
• 17:46
  enterprise
• 17:46
  take
• 17:46
  is
• 17:46
  more
• 17:47
  nuanced.
• 17:48
  A
• 17:48
  practitioner
• 17:48
  describes
• 17:49
  arriving
• 17:49
  at
• 17:49
  a
• 17:50
  client
• 17:50
  with
• 17:50
  20-year-old
• 17:51
  code
• 17:51
  nobody
• 17:52
  wants
• 17:52
  to
• 17:52
  touch
• 17:52
  and
• 17:52
  argues
• 17:53
  that
• 17:53
  2026
• 17:54
  is
• 17:54
  when
• 17:55
  legacy
• 17:55
  modernization
• 17:56
  stops
• 17:56
  being
• 17:56
  outsourced
• 17:57
  to
• 17:57
  contractors
• 17:58
  who
• 17:58
  charge
• 17:58
  by
• 17:58
  the
• 17:58
  hour
• 17:59
  and
• 17:59
  starts
• 17:59
  being
• 17:59
  understood
• 18:00
  by
• 18:00
  agents
• 18:00
  who
• 18:01
  charge
• 18:01
  by
• 18:01
  the
• 18:01
  token.
• 18:02
  But
• 18:02
  the
• 18:02
  caveat
• 18:03
  matters:
• 18:04
  bigger
• 18:04
  doesn't
• 18:04
  equal
• 18:05
  better,
• 18:05
  RAG
• 18:06
  is
• 18:06
  for
• 18:06
  documents
• 18:06
  but
• 18:07
  code
• 18:07
  is
• 18:07
  a
• 18:07
  graph,
• 18:08
  structured
• 18:08
  traversal
• 18:09
  beats
• 18:09
  naive
• 18:09
  chunk
• 18:09
  retrieval,
• 18:10
  and
• 18:10
  that
• 18:11
  1998
• 18:12
  codebase
• 18:12
  isn't
• 18:12
  semantically
• 18:13
  similar
• 18:13
  to
• 18:14
  anything
• 18:14
  --
• 18:15
  it's
• 18:15
  connected.
• 18:16
  Same
• 18:17
  harness,
• 18:17
  same
• 18:18
  policies,
• 18:18
  and
• 18:19
  Qwen
• 18:19
  3.5
• 18:20
  beat
• 18:20
  Kimi-K2.5,
• 18:21
  GLM-5,
• 18:23
  and
• 18:23
  DeepSeek-V3.2.
• 18:25
  Vibe
• 18:25
  coding
• 18:26
  doesn't
• 18:26
  have
• 18:26
  a
• 18:26
  place
• 18:27
  in
• 18:27
  an
• 18:27
  enterprise.
• 18:28
  The
• 18:28
  ambient
• 18:28
  intelligence
• 18:29
  editorial
• 18:30
  pushes
• 18:30
  toward
• 18:30
  a
• 18:30
  more
• 18:30
  philosophical
• 18:31
  register,
• 18:32
  arguing
• 18:33
  that
• 18:33
  intelligence
• 18:33
  may
• 18:34
  not
• 18:34
  be
• 18:34
  rare
• 18:34
  or
• 18:34
  centralized
• 18:35
  but
• 18:35
  exists
• 18:35
  as
• 18:36
  ambient
• 18:36
  pattern
• 18:36
  across
• 18:37
  systems,
• 18:38
  waiting
• 18:38
  for
• 18:38
  alignment
• 18:38
  to
• 18:39
  become
• 18:39
  observable.
• 18:40
  The
• 18:40
  WiFi
• 18:40
  DensePose
• 18:41
  work
• 18:41
  --
• 18:42
  extracting
• 18:42
  breath,
• 18:43
  posture,
• 18:44
  and
• 18:44
  presence
• 18:44
  from
• 18:44
  what
• 18:45
  looked
• 18:45
  like
• 18:45
  noise
• 18:45
  --
• 18:46
  made
• 18:46
  this
• 18:46
  tangible.
• 18:47
  What
• 18:48
  seemed
• 18:48
  like
• 18:48
  signal
• 18:49
  noise
• 18:49
  became
• 18:49
  structured
• 18:50
  information
• 18:50
  once
• 18:51
  the
• 18:51
  geometry
• 18:51
  to
• 18:51
  read
• 18:52
  it
• 18:52
  was
• 18:52
  available.
• 18:53
  Whether
• 18:53
  you
• 18:54
  find
• 18:54
  this
• 18:54
  compelling
• 18:54
  depends
• 18:55
  on
• 18:55
  your
• 18:55
  tolerance
• 18:55
  for
• 18:56
  extrapolation
• 18:56
  from
• 18:56
  signal
• 18:57
  processing
• 18:57
  to
• 18:57
  metaphysics,
• 18:59
  but
• 18:59
  the
• 18:59
  core
• 18:59
  observation
• 19:00
  about
• 19:00
  latent
• 19:00
  structure
• 19:01
  in
• 19:01
  environments
• 19:02
  is
• 19:02
  grounded
• 19:02
  in
• 19:02
  real
• 19:03
  engineering.