Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – February 25, 2026

Transcript

AGI Dreams Podcast – February 25, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  Qwen
• 0:00
  team
• 0:00
  just
• 0:00
  published
• 0:01
  a
• 0:01
  paper
• 0:01
  confirming
• 0:02
  what
• 0:02
  an
• 0:02
  independent
• 0:03
  researcher
• 0:03
  discovered
• 0:04
  weeks
• 0:04
  ago
• 0:05
  while
• 0:05
  debugging
• 0:05
  a
• 0:06
  "DeepSeek-Overclock"
• 0:07
  experiment:
• 0:08
  the
• 0:08
  gold-standard
• 0:09
  labels
• 0:09
  in
• 0:09
  both
• 0:10
  GPQA
• 0:10
  and
• 0:10
  HLE
• 0:11
  (Humanity's
• 0:11
  Last
• 0:12
  Exam)
• 0:12
  are
• 0:12
  riddled
• 0:13
  with
• 0:13
  errors.
• 0:14
  The
• 0:14
  original
• 0:14
  investigation
• 0:15
  found
• 0:15
  DeepSeek
• 0:16
  rigorously
• 0:16
  deriving
• 0:17
  technically
• 0:17
  correct
• 0:17
  answers
• 0:18
  that
• 0:18
  contradicted
• 0:19
  the
• 0:19
  provided
• 0:19
  labels
• 0:19
  --
• 0:20
  and
• 0:20
  when
• 0:20
  the
• 0:20
  math
• 0:21
  was
• 0:21
  verified
• 0:21
  line-by-line
• 0:22
  from
• 0:22
  first
• 0:22
  principles,
• 0:23
  the
• 0:23
  model
• 0:23
  was
• 0:24
  right
• 0:24
  and
• 0:24
  the
• 0:24
  benchmark
• 0:25
  was
• 0:25
  wrong.
• 0:25
  Qwen's
• 0:26
  paper
• 0:26
  doesn't
• 0:26
  mince
• 0:26
  words:
• 0:27
  many
• 0:27
  HLE
• 0:27
  questions
• 0:28
  are
• 0:28
  "fundamentally
• 0:29
  broken,"
• 0:29
  and
• 0:30
  some
• 0:30
  standard
• 0:30
  answers
• 0:30
  are
• 0:31
  simply
• 0:31
  incorrect.
• 0:32
  This
• 0:32
  matters
• 0:32
  because
• 0:33
  GPQA
• 0:33
  and
• 0:34
  HLE
• 0:34
  aren't
• 0:34
  obscure
• 0:35
  test
• 0:35
  suites
• 0:35
  --
• 0:36
  they're
• 0:36
  the
• 0:36
  benchmarks
• 0:36
  labs
• 0:37
  cite
• 0:37
  to
• 0:37
  claim
• 0:37
  frontier
• 0:38
  progress.
• 0:39
  If
• 0:39
  the
• 0:39
  ground
• 0:39
  truth
• 0:39
  is
• 0:39
  corrupt,
• 0:40
  leaderboard
• 0:40
  deltas
• 0:41
  between
• 0:41
  models
• 0:41
  may
• 0:42
  reflect
• 0:42
  noise
• 0:42
  rather
• 0:43
  than
• 0:43
  genuine
• 0:43
  capability
• 0:44
  differences.
• 0:45
  The
• 0:45
  problem
• 0:45
  compounds
• 0:45
  when
• 0:46
  you
• 0:46
  remember
• 0:46
  the
• 0:46
  benchmaxxing
• 0:47
  dynamic:
• 0:48
  models
• 0:48
  optimized
• 0:48
  for
• 0:49
  harness
• 0:49
  formats
• 0:49
  until
• 0:49
  those
• 0:50
  formats
• 0:50
  become
• 0:50
  baked-in
• 0:51
  defaults,
• 0:51
  hallucinating
• 0:52
  `\boxed{}`
• 0:52
  LaTeX
• 0:53
  answer
• 0:53
  formatting
• 0:53
  that
• 0:53
  appeared
• 0:54
  nowhere
• 0:54
  in
• 0:54
  their
• 0:54
  prompts.
• 0:55
  We've
• 0:55
  now
• 0:55
  gone
• 0:56
  from
• 0:56
  "models
• 0:56
  gaming
• 0:56
  the
• 0:56
  test"
• 0:57
  to
• 0:57
  "the
• 0:57
  test
• 0:57
  itself
• 0:58
  being
• 0:58
  wrong"
• 0:58
  --
• 0:59
  a
• 0:59
  qualitatively
• 0:59
  different
• 1:00
  failure
• 1:00
  mode
• 1:00
  that
• 1:00
  calls
• 1:00
  into
• 1:01
  question
• 1:01
  the
• 1:01
  entire
• 1:01
  evaluation
• 1:02
  pipeline
• 1:02
  from
• 1:02
  dataset
• 1:03
  curation
• 1:03
  through
• 1:03
  scoring.
• 1:04
  The
• 1:04
  gap
• 1:05
  between
• 1:05
  benchmark
• 1:05
  scores
• 1:06
  and
• 1:06
  real-world
• 1:06
  performance
• 1:07
  shows
• 1:07
  up
• 1:07
  vividly
• 1:07
  in
• 1:08
  a
• 1:08
  new
• 1:08
  independent
• 1:08
  evaluation
• 1:09
  of
• 1:09
  Qwen
• 1:09
  3.5.
• 1:10
  The
• 1:10
  APEX
• 1:11
  Testing
• 1:11
  benchmark
• 1:11
  --
• 1:12
  70
• 1:13
  tasks
• 1:13
  drawn
• 1:13
  from
• 1:13
  real
• 1:14
  GitHub
• 1:14
  repos,
• 1:15
  covering
• 1:15
  bug
• 1:15
  fixes,
• 1:16
  refactors,
• 1:17
  race
• 1:17
  conditions,
• 1:17
  and
• 1:17
  CLI
• 1:18
  tool
• 1:18
  builds
• 1:18
  --
• 1:19
  found
• 1:19
  that
• 1:19
  Qwen
• 1:20
  3.5
• 1:21
  397B
• 1:22
  craters
• 1:22
  on
• 1:22
  "master"
• 1:23
  difficulty
• 1:23
  tasks,
• 1:24
  dropping
• 1:24
  from
• 1:24
  a
• 1:24
  respectable
• 1:25
  ~1550
• 1:26
  ELO
• 1:26
  on
• 1:26
  hard/expert
• 1:27
  work
• 1:27
  to
• 1:27
  1194
• 1:28
  when
• 1:28
  coordination
• 1:29
  across
• 1:29
  many
• 1:29
  files
• 1:30
  over
• 1:30
  many
• 1:30
  steps
• 1:30
  is
• 1:30
  required.
• 1:31
  The
• 1:32
  model
• 1:32
  simply
• 1:32
  loses
• 1:33
  track
• 1:33
  of
• 1:33
  what
• 1:33
  it's
• 1:33
  doing.
• 1:34
  Meanwhile,
• 1:35
  GLM-4.7
• 1:36
  quantized
• 1:36
  at
• 1:36
  ~205GB
• 1:39
  still
• 1:39
  beats
• 1:39
  every
• 1:39
  Qwen
• 1:40
  3.5
• 1:40
  variant
• 1:41
  including
• 1:41
  the
• 1:41
  full
• 1:41
  397B
• 1:43
  cloud
• 1:43
  version,
• 1:44
  posting
• 1:44
  1572
• 1:45
  ELO.
• 1:46
  One
• 1:47
  Qwen
• 1:47
  model
• 1:47
  even
• 1:47
  found
• 1:47
  a
• 1:48
  loophole:
• 1:48
  Qwen3.5-27B
• 1:52
  ran
• 1:52
  a
• 1:52
  test
• 1:52
  suite,
• 1:53
  saw
• 1:53
  existing
• 1:53
  tests
• 1:54
  passing,
• 1:54
  declared
• 1:54
  everything
• 1:55
  "already
• 1:55
  implemented,"
• 1:56
  and
• 1:56
  quit
• 1:56
  without
• 1:57
  writing
• 1:57
  a
• 1:57
  single
• 1:57
  line
• 1:57
  of
• 1:57
  code.
• 1:58
  It
• 1:58
  was
• 1:59
  the
• 1:59
  only
• 1:59
  model
• 1:59
  out
• 1:59
  of
• 1:59
  25+
• 2:00
  that
• 2:00
  tried
• 2:00
  this.
• 2:02
  On
• 2:02
  a
• 2:02
  brighter
• 2:02
  note,
• 2:02
  Codex
• 2:03
  5.3
• 2:03
  is
• 2:04
  basically
• 2:04
  tied
• 2:04
  with
• 2:04
  GPT-5.2
• 2:05
  at
• 2:06
  #4
• 2:06
  overall,
• 2:07
  showing
• 2:07
  remarkable
• 2:08
  consistency
• 2:08
  across
• 2:09
  difficulty
• 2:09
  levels.
• 2:10
  The
• 2:10
  trust
• 2:11
  problem
• 2:11
  extends
• 2:11
  beyond
• 2:12
  benchmarks
• 2:12
  into
• 2:12
  the
• 2:12
  information
• 2:13
  supply
• 2:13
  chain
• 2:13
  itself.
• 2:15
  A
• 2:15
  new
• 2:15
  report
• 2:15
  from
• 2:15
  The
• 2:16
  Verge
• 2:16
  reveals
• 2:17
  that
• 2:17
  ChatGPT,
• 2:18
  Gemini,
• 2:19
  Copilot,
• 2:19
  and
• 2:19
  Perplexity
• 2:20
  are
• 2:20
  increasingly
• 2:21
  citing
• 2:21
  Grokipedia
• 2:22
  --
• 2:23
  Elon
• 2:23
  Musk's
• 2:24
  AI-generated
• 2:24
  Wikipedia
• 2:25
  alternative
• 2:25
  --
• 2:26
  as
• 2:26
  a
• 2:26
  factual
• 2:27
  reference.
• 2:28
  Grokipedia
• 2:28
  faces
• 2:29
  heavy
• 2:29
  criticism
• 2:30
  for
• 2:30
  hallucinations,
• 2:31
  plagiarism,
• 2:31
  and
• 2:32
  promoting
• 2:32
  debunked
• 2:32
  conspiracy
• 2:33
  theories.
• 2:34
  The
• 2:34
  result
• 2:34
  is
• 2:34
  a
• 2:35
  circular
• 2:35
  "AI-citing-AI"
• 2:36
  loop
• 2:37
  where
• 2:37
  one
• 2:37
  system's
• 2:37
  hallucinations
• 2:38
  become
• 2:39
  another
• 2:39
  system's
• 2:39
  source
• 2:40
  material,
• 2:41
  a
• 2:41
  feedback
• 2:41
  mechanism
• 2:42
  that
• 2:42
  could
• 2:42
  systematically
• 2:42
  amplify
• 2:43
  misinformation
• 2:44
  at
• 2:44
  scale.
• 2:46
  Three
• 2:46
  papers
• 2:47
  this
• 2:47
  week
• 2:47
  attack
• 2:47
  different
• 2:48
  bottlenecks
• 2:48
  in
• 2:48
  transformer
• 2:49
  efficiency,
• 2:49
  and
• 2:50
  together
• 2:50
  they
• 2:50
  sketch
• 2:50
  the
• 2:51
  outlines
• 2:51
  of
• 2:51
  a
• 2:51
  post-attention-era
• 2:52
  architecture.
• 2:53
  The
• 2:53
  common
• 2:54
  thread:
• 2:55
  fixed-depth,
• 2:55
  quadratic-cost
• 2:56
  computation
• 2:57
  is
• 2:57
  the
• 2:57
  real
• 2:57
  ceiling,
• 2:58
  and
• 2:58
  the
• 2:58
  most
• 2:58
  promising
• 2:59
  paths
• 2:59
  forward
• 2:59
  involve
• 3:00
  rethinking
• 3:00
  how
• 3:00
  information
• 3:01
  flows
• 3:01
  rather
• 3:01
  than
• 3:01
  simply
• 3:02
  adding
• 3:02
  more
• 3:02
  parameters.
• 3:03
  The
• 3:04
  most
• 3:04
  striking
• 3:04
  result
• 3:04
  comes
• 3:05
  from
• 3:05
  "Turbo
• 3:05
  Connection,"
• 3:06
  which
• 3:06
  introduces
• 3:06
  TurboConn
• 3:07
  --
• 3:08
  an
• 3:08
  architectural
• 3:09
  modification
• 3:10
  that
• 3:10
  routes
• 3:10
  residual
• 3:11
  connections
• 3:11
  from
• 3:11
  higher-layer
• 3:12
  hidden
• 3:12
  states
• 3:12
  of
• 3:12
  each
• 3:12
  token
• 3:13
  back
• 3:13
  to
• 3:13
  lower
• 3:14
  layers
• 3:14
  processing
• 3:14
  the
• 3:14
  next
• 3:15
  token.
• 3:16
  The
• 3:16
  insight:
• 3:18
  reasoning
• 3:18
  power
• 3:18
  in
• 3:18
  standard
• 3:19
  transformers
• 3:19
  is
• 3:19
  fundamentally
• 3:20
  bottlenecked
• 3:20
  by
• 3:21
  fixed
• 3:21
  depth.
• 3:23
  TurboConn
• 3:23
  breaks
• 3:23
  this
• 3:23
  by
• 3:24
  allowing
• 3:24
  effective
• 3:24
  reasoning
• 3:25
  depth
• 3:25
  to
• 3:25
  scale
• 3:25
  linearly
• 3:26
  with
• 3:26
  sequence
• 3:26
  length
• 3:27
  rather
• 3:27
  than
• 3:27
  being
• 3:27
  capped
• 3:28
  by
• 3:28
  layer
• 3:28
  count.
• 3:29
  Fine-tuning
• 3:29
  Qwen-3-1.7B
• 3:31
  with
• 3:31
  TurboConn
• 3:32
  achieves
• 3:32
  100%
• 3:33
  accuracy
• 3:33
  on
• 3:33
  Parity
• 3:34
  tasks
• 3:34
  where
• 3:35
  standard
• 3:35
  fine-tuning
• 3:35
  plateaus
• 3:36
  at
• 3:36
  53.78%,
• 3:38
  and
• 3:38
  the
• 3:38
  modified
• 3:39
  1B
• 3:39
  model
• 3:39
  outperforms
• 3:40
  the
• 3:40
  8B
• 3:40
  baseline
• 3:41
  --
• 3:41
  demonstrating
• 3:42
  that
• 3:42
  improved
• 3:42
  information
• 3:43
  flow
• 3:43
  can
• 3:43
  surpass
• 3:44
  raw
• 3:44
  parameter
• 3:45
  scaling
• 3:45
  for
• 3:45
  certain
• 3:45
  reasoning
• 3:46
  problems.
• 3:47
  Critically,
• 3:48
  dense
• 3:48
  multi-layer-to-multi-layer
• 3:49
  connections
• 3:50
  significantly
• 3:50
  outperform
• 3:51
  sparse
• 3:51
  alternatives
• 3:52
  like
• 3:52
  passing
• 3:52
  only
• 3:52
  the
• 3:52
  top-layer
• 3:53
  hidden
• 3:53
  state
• 3:53
  back
• 3:54
  to
• 3:54
  the
• 3:54
  input.
• 3:55
  The
• 3:55
  model
• 3:55
  also
• 3:55
  exhibits
• 3:56
  strong
• 3:56
  length
• 3:56
  generalization,
• 3:58
  maintaining
• 3:58
  perfect
• 3:59
  accuracy
• 3:59
  on
• 3:59
  sequences
• 4:00
  3x
• 4:00
  longer
• 4:00
  than
• 4:00
  training
• 4:01
  data,
• 4:01
  suggesting
• 4:02
  it
• 4:02
  learns
• 4:02
  generalizable
• 4:03
  algorithms
• 4:04
  rather
• 4:04
  than
• 4:04
  memorizing
• 4:04
  patterns.
• 4:06
  On
• 4:06
  the
• 4:06
  inference
• 4:06
  cost
• 4:07
  side,
• 4:07
  NoesisLab's
• 4:08
  Spartacus-1B
• 4:09
  proposes
• 4:09
  a
• 4:09
  more
• 4:10
  radical
• 4:10
  departure.
• 4:11
  It
• 4:12
  completely
• 4:12
  replaces
• 4:12
  softmax
• 4:13
  attention
• 4:13
  with
• 4:14
  "Causal
• 4:14
  Monoid
• 4:14
  State
• 4:15
  Compression,"
• 4:16
  compressing
• 4:16
  the
• 4:16
  entire
• 4:17
  causal
• 4:17
  prefix
• 4:17
  into
• 4:18
  a
• 4:18
  fixed-size
• 4:18
  d-by-d
• 4:19
  state
• 4:19
  matrix
• 4:20
  that
• 4:20
  updates
• 4:20
  via
• 4:20
  a
• 4:20
  single
• 4:21
  monoid
• 4:21
  operation
• 4:22
  per
• 4:22
  token.
• 4:23
  The
• 4:23
  result
• 4:24
  is
• 4:24
  true
• 4:24
  O(1)
• 4:24
  inference
• 4:25
  time
• 4:25
  and
• 4:25
  O(1)
• 4:26
  memory
• 4:26
  per
• 4:26
  token
• 4:26
  --
• 4:27
  whether
• 4:27
  generating
• 4:28
  the
• 4:28
  10th
• 4:28
  or
• 4:29
  100,000th
• 4:29
  token,
• 4:30
  the
• 4:30
  cost
• 4:30
  stays
• 4:31
  constant.
• 4:32
  Learned
• 4:32
  content-dependent
• 4:33
  vector
• 4:33
  decay
• 4:34
  gates
• 4:34
  replace
• 4:34
  both
• 4:35
  RoPE
• 4:35
  and
• 4:35
  attention
• 4:35
  masks,
• 4:36
  with
• 4:36
  fast-decaying
• 4:37
  dimensions
• 4:38
  tracking
• 4:38
  local
• 4:38
  syntax
• 4:39
  and
• 4:39
  slow-decaying
• 4:40
  dimensions
• 4:40
  serving
• 4:40
  as
• 4:41
  global
• 4:41
  memory.
• 4:42
  At
• 4:42
  1.3B
• 4:43
  parameters,
• 4:44
  Spartacus
• 4:45
  outperforms
• 4:45
  Mamba-1.4B
• 4:47
  and
• 4:47
  RWKV-6-1.6B
• 4:52
  on
• 4:52
  zero-shot
• 4:53
  ARC
• 4:53
  benchmarks,
• 4:54
  and
• 4:55
  integration
• 4:55
  of
• 4:55
  structured
• 4:56
  chain-of-thought
• 4:56
  data
• 4:57
  has
• 4:57
  pushed
• 4:57
  reasoning
• 4:58
  accuracy
• 4:58
  to
• 4:58
  75%.
• 5:00
  The
• 5:00
  architecture
• 5:01
  is
• 5:01
  theoretically
• 5:01
  elegant,
• 5:02
  though
• 5:02
  community
• 5:03
  commenters
• 5:03
  rightly
• 5:04
  note
• 5:04
  that
• 5:04
  benchmarking
• 5:05
  against
• 5:05
  Mamba
• 5:05
  and
• 5:06
  RWKV
• 5:07
  rather
• 5:07
  than
• 5:07
  current
• 5:07
  sub-quadratic
• 5:08
  SOTA
• 5:09
  like
• 5:09
  Kimi-Linear
• 5:10
  or
• 5:10
  Qwen3-Next
• 5:11
  leaves
• 5:11
  the
• 5:11
  competitive
• 5:12
  picture
• 5:12
  incomplete.
• 5:13
  Meanwhile,
• 5:14
  a
• 5:14
  team
• 5:14
  from
• 5:14
  MBZUAI's
• 5:16
  VILA
• 5:16
  Lab
• 5:16
  challenges
• 5:17
  a
• 5:17
  long-held
• 5:17
  assumption
• 5:18
  in
• 5:18
  model
• 5:18
  compression.
• 5:19
  "Sink-Aware
• 5:20
  Pruning
• 5:20
  for
• 5:20
  Diffusion
• 5:21
  Language
• 5:21
  Models"
• 5:22
  demonstrates
• 5:23
  that
• 5:23
  attention
• 5:23
  sinks
• 5:23
  --
• 5:24
  token
• 5:25
  positions
• 5:25
  absorbing
• 5:26
  disproportionate
• 5:26
  attention
• 5:27
  mass
• 5:27
  --
• 5:28
  behave
• 5:28
  fundamentally
• 5:29
  differently
• 5:29
  in
• 5:29
  diffusion
• 5:29
  language
• 5:30
  models
• 5:30
  (DLMs)
• 5:31
  compared
• 5:31
  to
• 5:31
  autoregressive
• 5:32
  models.
• 5:33
  In
• 5:33
  AR
• 5:33
  models
• 5:34
  like
• 5:34
  LLaMA-3,
• 5:35
  sinks
• 5:35
  are
• 5:36
  spatially
• 5:36
  concentrated
• 5:37
  on
• 5:37
  fixed
• 5:37
  early
• 5:37
  tokens
• 5:38
  and
• 5:38
  remain
• 5:38
  temporally
• 5:39
  stable
• 5:39
  throughout
• 5:39
  generation.
• 5:41
  In
• 5:41
  DLMs
• 5:42
  like
• 5:42
  LLaDA
• 5:42
  and
• 5:43
  Dream,
• 5:43
  sink
• 5:43
  positions
• 5:44
  shift
• 5:44
  substantially
• 5:45
  across
• 5:45
  denoising
• 5:46
  timesteps,
• 5:47
  with
• 5:47
  temporal
• 5:47
  variance
• 5:48
  orders
• 5:48
  of
• 5:48
  magnitude
• 5:48
  larger
• 5:49
  than
• 5:49
  in
• 5:49
  AR
• 5:49
  models.
• 5:51
  The
• 5:51
  standard
• 5:51
  pruning
• 5:51
  heuristic
• 5:52
  of
• 5:52
  always
• 5:52
  preserving
• 5:53
  sink
• 5:53
  tokens,
• 5:54
  inherited
• 5:54
  wholesale
• 5:55
  from
• 5:55
  AR
• 5:55
  research,
• 5:56
  actively
• 5:56
  hurts
• 5:57
  DLM
• 5:57
  compression.
• 5:59
  Their
• 5:59
  proposed
• 5:59
  method
• 5:59
  --
• 6:00
  which
• 6:01
  computes
• 6:01
  soft
• 6:01
  sink
• 6:02
  scores
• 6:02
  and
• 6:02
  down-weights
• 6:03
  unstable
• 6:03
  positions
• 6:04
  before
• 6:04
  feeding
• 6:04
  activations
• 6:05
  into
• 6:05
  existing
• 6:06
  pruning
• 6:06
  criteria
• 6:06
  --
• 6:08
  consistently
• 6:08
  outperforms
• 6:09
  baselines
• 6:09
  at
• 6:09
  50-75%
• 6:12
  sparsity
• 6:13
  levels
• 6:13
  with
• 6:13
  no
• 6:14
  retraining
• 6:14
  required.
• 6:18
  Taalas,
• 6:19
  the
• 6:19
  startup
• 6:20
  that
• 6:20
  emerged
• 6:20
  from
• 6:20
  stealth
• 6:20
  with
• 6:21
  over
• 6:21
  $200M
• 6:22
  in
• 6:22
  funding
• 6:23
  and
• 6:23
  a
• 6:23
  radical
• 6:23
  proposition
• 6:23
  --
• 6:24
  etch
• 6:24
  model
• 6:25
  weights
• 6:25
  directly
• 6:25
  into
• 6:26
  transistors
• 6:26
  --
• 6:26
  has
• 6:27
  opened
• 6:27
  free
• 6:27
  access
• 6:28
  to
• 6:28
  its
• 6:28
  first-generation
• 6:28
  ASIC
• 6:29
  running
• 6:29
  Llama
• 6:30
  3.1
• 6:30
  8B
• 6:31
  at
• 6:31
  16,000
• 6:31
  tokens
• 6:32
  per
• 6:32
  second.
• 6:33
  The
• 6:33
  HC1
• 6:33
  chip,
• 6:34
  fabricated
• 6:34
  at
• 6:34
  TSMC
• 6:35
  6nm
• 6:36
  with
• 6:36
  53
• 6:37
  billion
• 6:37
  transistors
• 6:38
  on
• 6:38
  an
• 6:38
  815mm²
• 6:40
  die,
• 6:40
  uses
• 6:41
  mask
• 6:41
  ROM
• 6:41
  recall
• 6:41
  fabric
• 6:42
  and
• 6:42
  SRAM
• 6:42
  KV
• 6:42
  caches
• 6:43
  rather
• 6:43
  than
• 6:43
  general-purpose
• 6:44
  compute.
• 6:44
  At
• 6:44
  roughly
• 6:45
  200W
• 6:46
  power
• 6:46
  draw,
• 6:46
  napkin
• 6:47
  math
• 6:47
  puts
• 6:47
  the
• 6:47
  cost
• 6:48
  at
• 6:48
  ~$0.005
• 6:49
  per
• 6:49
  million
• 6:50
  tokens
• 6:50
  for
• 6:50
  electricity
• 6:51
  alone.
• 6:52
  The
• 6:52
  chatbot
• 6:52
  demo
• 6:52
  undersells
• 6:53
  the
• 6:53
  speed
• 6:53
  --
• 6:54
  anything
• 6:54
  over
• 6:54
  a
• 6:55
  few
• 6:55
  hundred
• 6:55
  tok/s
• 6:56
  feels
• 6:56
  instantaneous
• 6:56
  --
• 6:57
  but
• 6:57
  the
• 6:57
  real
• 6:58
  showcase
• 6:58
  would
• 6:58
  be
• 6:58
  token-intensive
• 6:59
  API
• 6:59
  workloads
• 7:00
  where
• 7:00
  that
• 7:00
  throughput
• 7:01
  translates
• 7:01
  into
• 7:01
  radically
• 7:02
  different
• 7:02
  economics.
• 7:03
  The
• 7:03
  obvious
• 7:03
  limitation:
• 7:04
  every
• 7:05
  model
• 7:05
  revision
• 7:05
  requires
• 7:06
  a
• 7:06
  new
• 7:06
  mask
• 7:06
  spin,
• 7:07
  with
• 7:07
  a
• 7:07
  reported
• 7:08
  two-month
• 7:08
  turnaround.
• 7:09
  Taalas
• 7:09
  is
• 7:09
  betting
• 7:10
  that
• 7:10
  a
• 7:10
  few
• 7:10
  dominant
• 7:11
  model
• 7:11
  sizes
• 7:11
  will
• 7:11
  see
• 7:11
  enough
• 7:12
  production
• 7:12
  volume
• 7:13
  to
• 7:13
  justify
• 7:14
  application-specific
• 7:14
  silicon,
• 7:15
  the
• 7:15
  diametric
• 7:16
  opposite
• 7:16
  of
• 7:16
  Tenstorrent's
• 7:17
  general-purpose
• 7:18
  programmable
• 7:18
  approach.
• 7:20
  On
• 7:20
  the
• 7:20
  other
• 7:20
  end
• 7:20
  of
• 7:20
  the
• 7:20
  hardware
• 7:21
  spectrum,
• 7:21
  Cognitum
• 7:22
  has
• 7:22
  announced
• 7:22
  what
• 7:22
  it
• 7:22
  calls
• 7:23
  "the
• 7:23
  world's
• 7:23
  first
• 7:24
  Agentic
• 7:24
  Processing
• 7:24
  Unit"
• 7:24
  --
• 7:26
  a
• 7:26
  proportional
• 7:26
  intelligence
• 7:27
  system
• 7:27
  designed
• 7:27
  for
• 7:27
  always-on,
• 7:28
  event-driven,
• 7:29
  ultra-low-power
• 7:30
  agent
• 7:30
  workloads.
• 7:31
  Details
• 7:32
  remain
• 7:32
  sparse:
• 7:33
  it's
• 7:33
  positioned
• 7:33
  as
• 7:33
  neither
• 7:34
  a
• 7:34
  GPU
• 7:34
  box
• 7:35
  nor
• 7:35
  an
• 7:35
  edge
• 7:35
  server
• 7:36
  but
• 7:36
  as
• 7:36
  "intelligence
• 7:37
  that
• 7:37
  lives,"
• 7:38
  suggesting
• 7:38
  purpose-built
• 7:39
  hardware
• 7:39
  for
• 7:39
  persistent
• 7:40
  agent
• 7:40
  loops
• 7:40
  rather
• 7:41
  than
• 7:41
  batch
• 7:41
  inference.
• 7:42
  The
• 7:42
  infrastructure
• 7:43
  conversation
• 7:43
  took
• 7:43
  a
• 7:44
  less
• 7:44
  inspiring
• 7:44
  turn
• 7:44
  with
• 7:45
  Hetzner
• 7:45
  announcing
• 7:45
  30-40%
• 7:49
  price
• 7:49
  increases
• 7:49
  effective
• 7:50
  April
• 7:50
  1,
• 7:51
  2026,
• 7:52
  applying
• 7:52
  to
• 7:53
  both
• 7:53
  new
• 7:53
  orders
• 7:53
  and
• 7:53
  existing
• 7:54
  products.
• 7:55
  Server
• 7:56
  Auction
• 7:56
  machines
• 7:56
  get
• 7:56
  a
• 7:56
  3%
• 7:57
  bump.
• 7:58
  Hetzner
• 7:58
  has
• 7:59
  long
• 7:59
  been
• 7:59
  the
• 7:59
  budget-conscious
• 8:00
  self-hoster's
• 8:00
  favorite,
• 8:01
  and
• 8:01
  just
• 8:01
  months
• 8:01
  ago
• 8:02
  was
• 8:02
  lowering
• 8:02
  prices
• 8:03
  and
• 8:03
  expanding
• 8:03
  flexible
• 8:03
  options.
• 8:05
  This
• 8:05
  reversal
• 8:05
  changes
• 8:06
  the
• 8:06
  calculus
• 8:06
  for
• 8:06
  anyone
• 8:07
  running
• 8:07
  inference,
• 8:07
  crawlers,
• 8:08
  or
• 8:08
  multi-agent
• 8:09
  systems
• 8:09
  on
• 8:09
  their
• 8:09
  metal.
• 8:10
  The
• 8:10
  broader
• 8:11
  trend
• 8:11
  is
• 8:11
  unmistakable:
• 8:12
  the
• 8:13
  era
• 8:13
  of
• 8:13
  artificially
• 8:14
  low
• 8:14
  cloud
• 8:14
  pricing,
• 8:15
  fueled
• 8:15
  by
• 8:15
  competitive
• 8:16
  pressure
• 8:16
  and
• 8:16
  massive
• 8:17
  capital
• 8:17
  investment,
• 8:18
  is
• 8:18
  giving
• 8:18
  way
• 8:19
  to
• 8:19
  economic
• 8:19
  reality.
• 8:22
  The
• 8:22
  security
• 8:22
  engineer
• 8:23
  behind
• 8:23
  Clawker
• 8:23
  puts
• 8:23
  the
• 8:24
  problem
• 8:24
  bluntly:
• 8:25
  handing
• 8:25
  an
• 8:25
  LLM
• 8:26
  unrestricted
• 8:26
  code
• 8:27
  execution,
• 8:27
  network
• 8:28
  access,
• 8:28
  and
• 8:28
  full
• 8:29
  filesystem
• 8:29
  access
• 8:29
  is
• 8:30
  reckless,
• 8:30
  and
• 8:30
  Claude
• 8:31
  Code's
• 8:31
  built-in
• 8:31
  sandbox
• 8:32
  is
• 8:32
  "the
• 8:32
  temu
• 8:32
  version
• 8:33
  of
• 8:33
  a
• 8:33
  container."
• 8:34
  Clawker
• 8:34
  is
• 8:34
  an
• 8:34
  open-source
• 8:35
  Go
• 8:35
  tool
• 8:35
  that
• 8:35
  orchestrates
• 8:36
  Claude
• 8:36
  Code
• 8:36
  agents
• 8:37
  inside
• 8:37
  Docker
• 8:37
  containers
• 8:38
  with
• 8:38
  security
• 8:38
  defaults
• 8:39
  that
• 8:39
  would
• 8:39
  make
• 8:39
  a
• 8:39
  paranoid
• 8:39
  sysadmin
• 8:40
  smile:
• 8:41
  a
• 8:41
  network
• 8:41
  firewall
• 8:42
  restricting
• 8:42
  outbound
• 8:43
  traffic
• 8:43
  to
• 8:43
  GitHub
• 8:43
  and
• 8:43
  package
• 8:44
  managers,
• 8:44
  a
• 8:44
  `pkg/whail`
• 8:47
  jail
• 8:47
  preventing
• 8:48
  containers
• 8:48
  from
• 8:48
  seeing
• 8:49
  non-Clawker
• 8:49
  Docker
• 8:49
  resources,
• 8:50
  project-based
• 8:51
  namespace
• 8:51
  isolation,
• 8:52
  and
• 8:52
  seamless
• 8:53
  git
• 8:53
  credential
• 8:53
  forwarding
• 8:53
  via
• 8:54
  a
• 8:54
  host
• 8:54
  proxy
• 8:54
  service.
• 8:55
  The
• 8:55
  convenience
• 8:56
  features
• 8:56
  are
• 8:56
  equally
• 8:56
  thoughtful:
• 8:57
  an
• 8:57
  embedded
• 8:58
  parameterized
• 8:59
  Dockerfile
• 8:59
  template
• 9:00
  with
• 9:00
  an
• 9:00
  unprivileged
• 9:00
  `claude`
• 9:01
  user,
• 9:01
  git
• 9:02
  worktree
• 9:02
  integration
• 9:02
  for
• 9:03
  parallel
• 9:03
  development
• 9:03
  on
• 9:03
  separate
• 9:04
  branches,
• 9:05
  an
• 9:05
  optional
• 9:05
  Prometheus/Loki/Grafana
• 9:07
  monitoring
• 9:07
  stack,
• 9:08
  and
• 9:08
  an
• 9:08
  experimental
• 9:09
  looping
• 9:09
  mode
• 9:10
  that
• 9:10
  runs
• 9:10
  autonomous
• 9:11
  agent
• 9:11
  iterations
• 9:11
  with
• 9:11
  stagnation
• 9:12
  detection
• 9:12
  and
• 9:13
  circuit
• 9:13
  breakers.
• 9:14
  It's
• 9:14
  the
• 9:14
  harness
• 9:14
  for
• 9:14
  the
• 9:14
  harness.
• 9:16
  Containment
• 9:16
  without
• 9:17
  detection
• 9:17
  is
• 9:17
  half
• 9:17
  a
• 9:17
  solution.
• 9:19
  Canari
• 9:19
  brings
• 9:19
  the
• 9:19
  honeypot
• 9:20
  token
• 9:20
  principle
• 9:20
  --
• 9:21
  which
• 9:21
  has
• 9:21
  worked
• 9:21
  in
• 9:22
  traditional
• 9:22
  security
• 9:22
  for
• 9:23
  15
• 9:23
  years
• 9:23
  --
• 9:24
  to
• 9:24
  LLM
• 9:25
  and
• 9:25
  RAG
• 9:25
  applications.
• 9:26
  The
• 9:27
  approach
• 9:27
  is
• 9:27
  simple:
• 9:28
  inject
• 9:28
  synthetic
• 9:29
  fake
• 9:29
  credentials
• 9:30
  (Stripe
• 9:30
  keys,
• 9:31
  AWS
• 9:31
  keys,
• 9:32
  credit
• 9:32
  cards)
• 9:33
  into
• 9:33
  an
• 9:33
  agent's
• 9:33
  context,
• 9:34
  then
• 9:34
  scan
• 9:34
  every
• 9:35
  output
• 9:35
  for
• 9:35
  exact
• 9:36
  token
• 9:36
  matches.
• 9:37
  If
• 9:37
  a
• 9:37
  canary
• 9:38
  appears
• 9:38
  in
• 9:38
  output,
• 9:39
  it's
• 9:39
  definitionally
• 9:40
  a
• 9:40
  breach
• 9:40
  --
• 9:40
  zero
• 9:41
  false
• 9:41
  positives,
• 9:42
  no
• 9:42
  probability
• 9:42
  thresholds,
• 9:43
  no
• 9:43
  tuning.
• 9:44
  The
• 9:44
  library
• 9:45
  wraps
• 9:45
  LLM
• 9:45
  calls
• 9:46
  with
• 9:46
  automatic
• 9:46
  scanning,
• 9:47
  reporting
• 9:47
  6ms
• 9:48
  detection
• 9:48
  latency.
• 9:49
  The
• 9:49
  prior
• 9:50
  art
• 9:50
  here
• 9:50
  is
• 9:50
  Beelzebub's
• 9:51
  canary
• 9:51
  tools
• 9:52
  for
• 9:52
  MCP,
• 9:53
  which
• 9:53
  registered
• 9:53
  fake
• 9:54
  functions
• 9:54
  like
• 9:54
  `repo
• 9:54
  exfil`
• 9:55
  that
• 9:55
  trigger
• 9:55
  alerts
• 9:55
  when
• 9:56
  invoked.
• 9:57
  Canari
• 9:57
  extends
• 9:58
  the
• 9:58
  concept
• 9:58
  from
• 9:58
  tool
• 9:59
  invocation
• 9:59
  to
• 9:59
  the
• 9:59
  output
• 10:00
  stream
• 10:00
  itself,
• 10:01
  catching
• 10:01
  exfiltration
• 10:02
  that
• 10:02
  looks
• 10:02
  exactly
• 10:03
  like
• 10:03
  a
• 10:03
  normal
• 10:03
  agent
• 10:04
  response.
• 10:05
  The
• 10:05
  supply
• 10:05
  chain
• 10:05
  perspective
• 10:06
  rounds
• 10:06
  out
• 10:06
  the
• 10:06
  defensive
• 10:07
  picture.
• 10:08
  A
• 10:08
  pointed
• 10:08
  editorial
• 10:09
  argues
• 10:09
  that
• 10:09
  replacing
• 10:10
  open-source
• 10:11
  dependencies
• 10:11
  wholesale
• 10:12
  with
• 10:12
  AI-generated
• 10:13
  code
• 10:13
  confuses
• 10:13
  technical
• 10:14
  viability
• 10:14
  with
• 10:15
  enterprise
• 10:15
  reality.
• 10:16
  CVEs
• 10:17
  are
• 10:17
  actually
• 10:17
  good,
• 10:18
  the
• 10:18
  author
• 10:18
  contends
• 10:19
  --
• 10:19
  a
• 10:20
  known
• 10:20
  vulnerability
• 10:21
  is
• 10:21
  infinitely
• 10:21
  better
• 10:21
  than
• 10:22
  an
• 10:22
  unknown
• 10:22
  one,
• 10:23
  and
• 10:23
  the
• 10:23
  generate-test-reprompt
• 10:24
  loop
• 10:24
  for
• 10:24
  AI
• 10:25
  replacements
• 10:25
  isn't
• 10:25
  free.
• 10:27
  Google's
• 10:27
  Project
• 10:27
  Zero,
• 10:28
  Anthropic,
• 10:29
  and
• 10:29
  others
• 10:29
  keep
• 10:29
  finding
• 10:30
  critical
• 10:30
  flaws
• 10:31
  in
• 10:31
  mature
• 10:31
  software
• 10:31
  using
• 10:32
  AI
• 10:32
  as
• 10:32
  a
• 10:32
  discovery
• 10:33
  amplifier,
• 10:34
  which
• 10:34
  benefits
• 10:34
  the
• 10:34
  entire
• 10:35
  ecosystem
• 10:35
  including
• 10:36
  packages
• 10:36
  already
• 10:37
  in
• 10:37
  your
• 10:37
  stack.
• 10:38
  The
• 10:38
  real
• 10:38
  shift
• 10:39
  for
• 10:39
  AppSec
• 10:39
  practitioners
• 10:40
  is
• 10:40
  toward
• 10:40
  managing
• 10:40
  guardrails
• 10:41
  around
• 10:41
  AI-assisted
• 10:42
  development,
• 10:43
  not
• 10:43
  pretending
• 10:44
  the
• 10:44
  dependency
• 10:44
  toolchain
• 10:45
  vanished
• 10:45
  overnight.
• 10:46
  The
• 10:46
  identity
• 10:46
  layer
• 10:47
  of
• 10:47
  this
• 10:47
  stack
• 10:47
  reveals
• 10:48
  tradeoffs
• 10:48
  that
• 10:48
  would
• 10:48
  make
• 10:49
  a
• 10:49
  CISO
• 10:49
  wince.
• 10:50
  A
• 10:50
  forensic
• 10:51
  teardown
• 10:51
  of
• 10:51
  LinkedIn's
• 10:52
  identity
• 10:52
  verification
• 10:53
  exposes
• 10:54
  what
• 10:54
  a
• 10:54
  three-minute
• 10:54
  passport
• 10:55
  scan
• 10:55
  actually
• 10:55
  costs
• 10:56
  --
• 10:56
  and
• 10:56
  the
• 10:57
  answer
• 10:57
  is
• 10:57
  far
• 10:57
  more
• 10:57
  than
• 10:58
  you'd
• 10:58
  expect
• 10:58
  for
• 10:58
  a
• 10:58
  blue
• 10:59
  checkmark
• 10:59
  on
• 10:59
  a
• 10:59
  professional
• 11:00
  networking
• 11:00
  site.
• 11:01
  The
• 11:01
  process
• 11:02
  is
• 11:02
  handled
• 11:02
  not
• 11:02
  by
• 11:03
  LinkedIn
• 11:03
  but
• 11:03
  by
• 11:03
  Persona
• 11:04
  Identities,
• 11:05
  Inc.,
• 11:06
  which
• 11:06
  collects
• 11:06
  full
• 11:06
  biometric
• 11:07
  faceprints,
• 11:08
  NFC
• 11:08
  chip
• 11:08
  data,
• 11:09
  behavioral
• 11:10
  biometrics
• 11:10
  (keystroke
• 11:11
  timing,
• 11:11
  interaction
• 11:12
  pauses),
• 11:13
  and
• 11:13
  cross-references
• 11:14
  users
• 11:14
  against
• 11:14
  credit
• 11:15
  agencies
• 11:15
  and
• 11:15
  government
• 11:16
  databases.
• 11:17
  All
• 11:17
  16
• 11:18
  subprocessors
• 11:19
  are
• 11:19
  US-based.
• 11:20
  The
• 11:20
  CLOUD
• 11:20
  Act
• 11:21
  means
• 11:21
  a
• 11:21
  US
• 11:21
  court
• 11:22
  can
• 11:22
  compel
• 11:22
  disclosure
• 11:23
  regardless
• 11:23
  of
• 11:23
  where
• 11:23
  the
• 11:23
  data
• 11:24
  sits
• 11:24
  physically
• 11:24
  --
• 11:25
  Frankfurt
• 11:26
  server
• 11:26
  or
• 11:26
  not.
• 11:27
  Persona's
• 11:27
  liability
• 11:28
  cap
• 11:28
  for
• 11:28
  a
• 11:28
  breach
• 11:29
  exposing
• 11:29
  your
• 11:29
  passport,
• 11:30
  facial
• 11:30
  geometry,
• 11:31
  and
• 11:31
  national
• 11:32
  ID
• 11:32
  number?
• 11:33
  Fifty
• 11:33
  dollars.
• 11:37
  The
• 11:37
  "100x
• 11:38
  developer"
• 11:38
  claim
• 11:38
  has
• 11:39
  evolved
• 11:39
  from
• 11:39
  Twitter
• 11:40
  boasting
• 11:40
  to
• 11:40
  documented
• 11:41
  infrastructure,
• 11:42
  and
• 11:42
  the
• 11:42
  details
• 11:42
  are
• 11:42
  more
• 11:43
  interesting
• 11:43
  than
• 11:43
  the
• 11:43
  headline
• 11:44
  number.
• 11:44
  Brandon
• 11:45
  Casci
• 11:45
  walks
• 11:46
  through
• 11:46
  the
• 11:46
  progression
• 11:46
  from
• 11:46
  using
• 11:47
  Claude
• 11:47
  Code
• 11:47
  for
• 11:47
  individual
• 11:48
  tasks
• 11:48
  to
• 11:48
  building
• 11:49
  a
• 11:49
  full
• 11:49
  orchestration
• 11:50
  layer
• 11:50
  that
• 11:50
  triggers
• 11:51
  agents
• 11:51
  via
• 11:51
  GitHub
• 11:51
  webhooks.
• 11:53
  The
• 11:53
  early
• 11:53
  months
• 11:53
  were
• 11:53
  "trial
• 11:54
  and
• 11:54
  error,
• 11:55
  figuring
• 11:55
  out
• 11:55
  what
• 11:55
  Claude
• 11:55
  Code
• 11:56
  could
• 11:56
  actually
• 11:56
  handle
• 11:57
  versus
• 11:57
  what
• 11:57
  I
• 11:57
  still
• 11:57
  needed
• 11:58
  to
• 11:58
  do
• 11:58
  myself."
• 11:59
  Each
• 12:00
  improvement
• 12:00
  compounded:
• 12:01
  better
• 12:01
  documentation
• 12:02
  led
• 12:02
  to
• 12:02
  better
• 12:02
  agent
• 12:03
  output,
• 12:03
  which
• 12:04
  built
• 12:04
  confidence
• 12:04
  to
• 12:05
  delegate
• 12:05
  more.
• 12:06
  The
• 12:06
  bottleneck
• 12:06
  wasn't
• 12:07
  the
• 12:07
  agents
• 12:07
  or
• 12:07
  the
• 12:07
  code
• 12:08
  --
• 12:08
  it
• 12:08
  was
• 12:09
  the
• 12:09
  human
• 12:09
  directing
• 12:09
  everything
• 12:10
  manually.
• 12:11
  His
• 12:11
  solution
• 12:12
  turns
• 12:12
  GitHub
• 12:12
  Issues
• 12:13
  into
• 12:13
  work
• 12:13
  orders
• 12:13
  with
• 12:13
  two
• 12:14
  labels:
• 12:15
  `agent-spec`
• 12:15
  expands
• 12:16
  a
• 12:16
  casual
• 12:16
  description
• 12:17
  into
• 12:17
  a
• 12:17
  full
• 12:17
  implementation
• 12:18
  spec
• 12:18
  using
• 12:18
  project
• 12:19
  context,
• 12:19
  while
• 12:20
  `agent-implement`
• 12:20
  executes
• 12:21
  it
• 12:21
  --
• 12:21
  writing
• 12:21
  code,
• 12:22
  running
• 12:22
  tests,
• 12:23
  committing,
• 12:23
  and
• 12:23
  opening
• 12:24
  PRs.
• 12:25
  A
• 12:25
  write/review
• 12:26
  agent
• 12:26
  loop
• 12:26
  iterates
• 12:27
  until
• 12:27
  standards
• 12:27
  are
• 12:27
  met,
• 12:28
  with
• 12:28
  a
• 12:28
  circuit
• 12:28
  breaker
• 12:28
  to
• 12:29
  prevent
• 12:29
  infinite
• 12:29
  loops.
• 12:30
  The
• 12:30
  proof
• 12:31
  point:
• 12:31
  on
• 12:31
  a
• 12:31
  single
• 12:31
  day,
• 12:32
  two
• 12:33
  agents
• 12:33
  across
• 12:33
  two
• 12:33
  codebases
• 12:34
  resolved
• 12:34
  a
• 12:34
  production
• 12:35
  disk-full
• 12:35
  crisis
• 12:35
  --
• 12:36
  one
• 12:37
  forked
• 12:37
  and
• 12:37
  upgraded
• 12:37
  a
• 12:37
  Ruby
• 12:38
  gem,
• 12:38
  while
• 12:38
  another
• 12:39
  updated
• 12:39
  the
• 12:39
  Rails
• 12:39
  app
• 12:40
  through
• 12:40
  4
• 12:40
  iterative
• 12:40
  Dockerfile
• 12:41
  debugging
• 12:41
  attempts.
• 12:42
  Casci
• 12:43
  doesn't
• 12:43
  take
• 12:43
  "100x"
• 12:45
  literally,
• 12:45
  but
• 12:46
  frames
• 12:46
  the
• 12:46
  gains
• 12:46
  as
• 12:47
  fundamentally
• 12:47
  operational
• 12:48
  rather
• 12:48
  than
• 12:48
  about
• 12:49
  typing
• 12:49
  speed.
• 12:50
  The
• 12:50
  tools
• 12:51
  enabling
• 12:51
  this
• 12:51
  kind
• 12:51
  of
• 12:52
  workflow
• 12:52
  are
• 12:52
  proliferating.
• 12:53
  Santiago
• 12:54
  Valdarrama
• 12:55
  demonstrates
• 12:55
  a
• 12:55
  Claude
• 12:55
  Code
• 12:56
  skill
• 12:56
  using
• 12:56
  Nimble
• 12:57
  to
• 12:57
  pull
• 12:57
  live,
• 12:58
  structured
• 12:58
  data
• 12:58
  from
• 12:58
  any
• 12:59
  website
• 12:59
  --
• 12:59
  not
• 13:00
  returning
• 13:00
  a
• 13:00
  wall
• 13:00
  of
• 13:00
  text
• 13:01
  but
• 13:01
  normalized
• 13:01
  tables
• 13:02
  organized
• 13:02
  by
• 13:02
  schema,
• 13:03
  handling
• 13:03
  JavaScript-rendered
• 13:04
  content.
• 13:05
  On
• 13:06
  the
• 13:06
  local-first
• 13:06
  side,
• 13:07
  LocalAgent
• 13:07
  v0.1.1
• 13:10
  provides
• 13:10
  an
• 13:10
  agent
• 13:11
  runtime
• 13:11
  supporting
• 13:11
  LM
• 13:12
  Studio,
• 13:12
  Ollama,
• 13:13
  and
• 13:13
  llama.cpp
• 13:14
  backends
• 13:15
  with
• 13:15
  Playwright
• 13:15
  MCP
• 13:16
  for
• 13:16
  browser
• 13:16
  automation,
• 13:17
  tool
• 13:17
  calling
• 13:18
  with
• 13:18
  safe
• 13:18
  defaults
• 13:19
  (shell
• 13:19
  and
• 13:19
  write
• 13:19
  disabled
• 13:20
  by
• 13:20
  default),
• 13:21
  replayable
• 13:21
  run
• 13:21
  artifacts,
• 13:22
  and
• 13:22
  an
• 13:22
  eval
• 13:22
  harness
• 13:23
  for
• 13:23
  repeatable
• 13:23
  testing.
• 13:24
  The
• 13:24
  safety-first
• 13:25
  posture
• 13:25
  --
• 13:26
  explicit
• 13:27
  flags
• 13:27
  required
• 13:28
  for
• 13:28
  coding
• 13:28
  tools,
• 13:29
  approval
• 13:29
  policies,
• 13:30
  and
• 13:30
  trust
• 13:30
  controls
• 13:31
  --
• 13:32
  contrasts
• 13:32
  sharply
• 13:33
  with
• 13:33
  the
• 13:33
  "dangerously
• 13:34
  skip
• 13:34
  permissions"
• 13:34
  defaults
• 13:35
  that
• 13:35
  Clawker
• 13:35
  was
• 13:36
  built
• 13:36
  to
• 13:36
  contain.
• 13:37
  The
• 13:37
  context
• 13:38
  layer
• 13:38
  is
• 13:38
  where
• 13:38
  things
• 13:39
  get
• 13:39
  architecturally
• 13:39
  ambitious.
• 13:41
  ContextGraph
• 13:42
  tackles
• 13:42
  the
• 13:42
  memory
• 13:42
  problem
• 13:43
  with
• 13:43
  a
• 13:43
  13-embedder
• 13:44
  MCP
• 13:44
  server
• 13:45
  that
• 13:45
  embeds
• 13:45
  every
• 13:46
  memory
• 13:46
  simultaneously
• 13:47
  across
• 13:47
  semantic,
• 13:48
  causal,
• 13:49
  temporal,
• 13:49
  code,
• 13:50
  entity,
• 13:50
  and
• 13:51
  structural
• 13:51
  spaces,
• 13:52
  then
• 13:52
  fuses
• 13:53
  results
• 13:53
  using
• 13:53
  Reciprocal
• 13:54
  Rank
• 13:54
  Fusion.
• 13:55
  Three
• 13:56
  embedders
• 13:56
  maintain
• 13:57
  dual
• 13:57
  vector
• 13:57
  indexes
• 13:58
  for
• 13:58
  asymmetric
• 13:59
  queries
• 13:59
  --
• 13:59
  "What
• 14:00
  caused
• 14:00
  X?"
• 14:01
  and
• 14:01
  "What
• 14:01
  did
• 14:01
  X
• 14:01
  cause?"
• 14:02
  return
• 14:02
  different
• 14:03
  results.
• 14:04
  The
• 14:04
  system
• 14:04
  exposes
• 14:05
  56
• 14:06
  MCP
• 14:06
  tools
• 14:07
  covering
• 14:07
  everything
• 14:08
  from
• 14:08
  causal
• 14:08
  chain
• 14:09
  building
• 14:09
  to
• 14:09
  entity
• 14:10
  extraction
• 14:10
  with
• 14:10
  TransE
• 14:11
  predictions,
• 14:12
  backed
• 14:12
  by
• 14:12
  RocksDB
• 14:13
  with
• 14:13
  51
• 14:13
  column
• 14:14
  families
• 14:14
  and
• 14:14
  15
• 14:15
  HNSW
• 14:16
  indexes.
• 14:17
  It
• 14:17
  targets
• 14:18
  sub-5ms
• 14:19
  p95
• 14:20
  latency
• 14:20
  for
• 14:20
  storage.
• 14:22
  Meanwhile,
• 14:22
  Starlog
• 14:23
  offers
• 14:23
  curated
• 14:24
  deep-dives
• 14:25
  on
• 14:25
  offensive
• 14:25
  security
• 14:26
  tools
• 14:26
  and
• 14:26
  AI
• 14:27
  agents,
• 14:27
  indexing
• 14:28
  GitHub
• 14:28
  stars
• 14:29
  into
• 14:29
  actionable
• 14:30
  intel
• 14:30
  for
• 14:30
  security
• 14:31
  professionals
• 14:32
  and
• 14:32
  agent
• 14:32
  builders
• 14:32
  --
• 14:33
  a
• 14:33
  recognition
• 14:34
  that
• 14:34
  the
• 14:34
  observability
• 14:35
  layer
• 14:35
  for
• 14:36
  this
• 14:36
  ecosystem
• 14:37
  barely
• 14:37
  exists
• 14:37
  yet.
• 14:40
  Hugging
• 14:41
  Face
• 14:41
  and
• 14:41
  Unsloth
• 14:41
  have
• 14:42
  teamed
• 14:42
  up
• 14:42
  to
• 14:42
  offer
• 14:43
  free
• 14:43
  model
• 14:43
  training
• 14:43
  through
• 14:44
  HF
• 14:44
  Jobs,
• 14:45
  with
• 14:45
  Unsloth
• 14:45
  providing
• 14:46
  ~2x
• 14:46
  faster
• 14:47
  training
• 14:47
  and
• 14:47
  ~60%
• 14:48
  less
• 14:49
  VRAM
• 14:49
  usage.
• 14:50
  The
• 14:50
  target
• 14:50
  is
• 14:50
  LFM2.5-1.2B-Instruct
• 14:52
  --
• 14:54
  a
• 14:54
  sub-1GB
• 14:55
  model
• 14:55
  optimized
• 14:56
  for
• 14:56
  on-device
• 14:56
  deployment
• 14:57
  on
• 14:57
  CPUs,
• 14:58
  phones,
• 14:58
  and
• 14:58
  laptops.
• 14:59
  A
• 14:59
  Claude
• 15:00
  Code
• 15:00
  or
• 15:00
  Codex
• 15:00
  skill
• 15:01
  handles
• 15:01
  the
• 15:01
  entire
• 15:02
  pipeline:
• 15:02
  generating
• 15:03
  a
• 15:03
  UV
• 15:03
  script
• 15:04
  with
• 15:04
  inline
• 15:04
  dependencies,
• 15:05
  submitting
• 15:05
  to
• 15:06
  HF
• 15:06
  Jobs,
• 15:06
  and
• 15:07
  pushing
• 15:07
  the
• 15:07
  trained
• 15:07
  model
• 15:07
  to
• 15:07
  your
• 15:08
  Hub
• 15:08
  repo.
• 15:09
  The
• 15:09
  barrier
• 15:09
  to
• 15:09
  entry
• 15:10
  for
• 15:10
  fine-tuning
• 15:10
  has
• 15:10
  essentially
• 15:11
  collapsed
• 15:11
  to
• 15:11
  "ask
• 15:11
  your
• 15:12
  coding
• 15:12
  agent
• 15:12
  to
• 15:12
  train
• 15:12
  a
• 15:12
  model"
• 15:13
  --
• 15:13
  though
• 15:13
  the
• 15:14
  guidance
• 15:14
  wisely
• 15:14
  notes
• 15:14
  you
• 15:15
  should
• 15:15
  ask
• 15:15
  for
• 15:15
  cost
• 15:15
  estimates
• 15:16
  before
• 15:16
  launching
• 15:16
  large
• 15:17
  jobs.
• 15:17
  The
• 15:17
  definition
• 15:18
  of
• 15:18
  "local"
• 15:19
  keeps
• 15:19
  expanding.
• 15:20
  Kitten
• 15:20
  TTS
• 15:21
  v0.8
• 15:22
  now
• 15:23
  runs
• 15:23
  entirely
• 15:23
  client-side
• 15:24
  in
• 15:24
  a
• 15:24
  minimal
• 15:24
  Next.js
• 15:25
  app,
• 15:26
  with
• 15:26
  nano/micro/mini
• 15:28
  models
• 15:28
  fetched
• 15:28
  from
• 15:28
  Hugging
• 15:29
  Face
• 15:29
  and
• 15:29
  cached
• 15:30
  via
• 15:30
  the
• 15:30
  Origin
• 15:30
  Private
• 15:31
  File
• 15:31
  System.
• 15:32
  It
• 15:32
  depends
• 15:32
  on
• 15:32
  onnxruntime-web
• 15:33
  and
• 15:33
  Xenova's
• 15:34
  phonemizer.js,
• 15:38
  though
• 15:38
  it's
• 15:38
  currently
• 15:39
  limited
• 15:39
  to
• 15:39
  the
• 15:39
  WASM
• 15:39
  backend
• 15:40
  (WebGPU
• 15:40
  outputs
• 15:41
  silence)
• 15:41
  and
• 15:41
  doesn't
• 15:41
  work
• 15:42
  in
• 15:42
  Safari.
• 15:43
  The
• 15:43
  trajectory
• 15:43
  from
• 15:44
  server-dependent
• 15:44
  TTS
• 15:45
  to
• 15:45
  fully
• 15:45
  client-side
• 15:46
  synthesis
• 15:46
  in
• 15:46
  a
• 15:46
  browser
• 15:47
  tab
• 15:47
  represents
• 15:48
  a
• 15:48
  genuine
• 15:48
  shift
• 15:49
  in
• 15:49
  what
• 15:49
  "local
• 15:49
  inference"
• 15:50
  means
• 15:50
  --
• 15:51
  no
• 15:51
  install,
• 15:52
  no
• 15:52
  Docker,
• 15:52
  no
• 15:53
  GPU.
• 15:54
  For
• 15:54
  those
• 15:54
  who
• 15:54
  prefer
• 15:54
  their
• 15:55
  AI
• 15:55
  stack
• 15:55
  on
• 15:55
  real
• 15:55
  metal,
• 15:56
  one
• 15:56
  project
• 15:57
  packages
• 15:57
  LiteLLM
• 15:58
  (unified
• 15:58
  API
• 15:58
  proxy
• 15:59
  across
• 15:59
  providers),
• 16:00
  n8n
• 16:01
  (workflow
• 16:01
  automation),
• 16:02
  and
• 16:02
  Open
• 16:03
  WebUI
• 16:03
  into
• 16:03
  a
• 16:03
  single
• 16:04
  Docker
• 16:04
  Compose
• 16:04
  file.
• 16:05
  It's
• 16:05
  the
• 16:06
  "everything
• 16:06
  bagel"
• 16:07
  approach
• 16:07
  --
• 16:08
  individually
• 16:08
  well-known
• 16:09
  components
• 16:09
  wired
• 16:10
  together
• 16:10
  into
• 16:10
  a
• 16:10
  turnkey
• 16:11
  local
• 16:11
  deployment
• 16:12
  that
• 16:12
  handles
• 16:12
  multi-provider
• 16:13
  routing,
• 16:13
  visual
• 16:14
  automation,
• 16:15
  and
• 16:15
  a
• 16:15
  chat
• 16:15
  interface
• 16:16
  in
• 16:16
  one
• 16:16
  `docker-compose
• 16:17
  up`.
• 16:18
  And
• 16:18
  when
• 16:18
  your
• 16:18
  AI
• 16:18
  workloads
• 16:19
  outgrow
• 16:19
  a
• 16:20
  single
• 16:20
  Postgres
• 16:20
  instance,
• 16:21
  PgDog
• 16:22
  offers
• 16:22
  a
• 16:22
  Rust-built
• 16:23
  proxy
• 16:23
  that
• 16:24
  adds
• 16:24
  transparent
• 16:24
  sharding,
• 16:25
  connection
• 16:26
  pooling,
• 16:26
  and
• 16:26
  load
• 16:27
  balancing
• 16:27
  without
• 16:27
  changing
• 16:28
  application
• 16:28
  code.
• 16:29
  Unlike
• 16:30
  PgBouncer,
• 16:31
  PgDog
• 16:31
  parses
• 16:32
  SQL
• 16:32
  using
• 16:32
  PostgreSQL's
• 16:33
  native
• 16:33
  parser
• 16:34
  (`pg
• 16:34
  query`)
• 16:34
  to
• 16:35
  route
• 16:35
  reads
• 16:35
  to
• 16:35
  replicas
• 16:36
  and
• 16:36
  writes
• 16:37
  to
• 16:37
  the
• 16:37
  primary
• 16:37
  through
• 16:37
  a
• 16:37
  single
• 16:38
  endpoint,
• 16:39
  correctly
• 16:39
  handling
• 16:40
  `SET`
• 16:40
  statements
• 16:40
  and
• 16:41
  session
• 16:41
  parameters
• 16:41
  that
• 16:42
  trip
• 16:42
  up
• 16:42
  simpler
• 16:42
  proxies.
• 16:44
  It
• 16:44
  handles
• 16:44
  cross-shard
• 16:45
  queries
• 16:45
  by
• 16:45
  fanning
• 16:46
  out
• 16:46
  to
• 16:46
  all
• 16:46
  shards
• 16:47
  and
• 16:47
  assembling
• 16:47
  results
• 16:48
  in
• 16:48
  memory,
• 16:49
  supports
• 16:49
  two-phase
• 16:50
  commit
• 16:50
  for
• 16:50
  atomic
• 16:51
  cross-shard
• 16:51
  writes,
• 16:52
  and
• 16:52
  can
• 16:52
  reshard
• 16:53
  databases
• 16:53
  online
• 16:54
  using
• 16:54
  logical
• 16:55
  replication
• 16:55
  with
• 16:56
  zero
• 16:56
  downtime
• 16:56
  --
• 16:57
  no
• 16:57
  etcd,
• 16:58
  no
• 16:58
  ZooKeeper,
• 16:59
  just
• 16:59
  a
• 17:00
  TOML
• 17:00
  config
• 17:00
  and
• 17:00
  a
• 17:00
  coordinated
• 17:01
  cutover
• 17:01
  sequence.
• 17:03
  Already
• 17:03
  used
• 17:03
  in
• 17:04
  production
• 17:04
  at
• 17:04
  scale,
• 17:05
  it's
• 17:05
  a
• 17:05
  reminder
• 17:05
  that
• 17:06
  not
• 17:06
  all
• 17:06
  critical
• 17:06
  infrastructure
• 17:07
  improvements
• 17:08
  require
• 17:08
  AI
• 17:08
  --
• 17:09
  sometimes
• 17:10
  you
• 17:10
  just
• 17:10
  need
• 17:10
  someone
• 17:11
  to
• 17:11
  write
• 17:11
  a
• 17:11
  really
• 17:11
  good
• 17:12
  proxy
• 17:12
  in
• 17:12
  Rust.