Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 17, 2026

Transcript

AGI Dreams Podcast – March 17, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Mixture-of-Experts
• 0:01
  has
• 0:01
  won
• 0:01
  the
• 0:01
  open-weight
• 0:02
  architecture
• 0:02
  debate.
• 0:03
  Sebastian
• 0:04
  Raschka's
• 0:04
  newly
• 0:05
  updated
• 0:05
  LLM
• 0:06
  Architecture
• 0:06
  Gallery
• 0:06
  makes
• 0:07
  that
• 0:07
  verdict
• 0:07
  visual:
• 0:08
  of
• 0:08
  the
• 0:09
  30-plus
• 0:09
  decoder
• 0:10
  designs
• 0:10
  catalogued
• 0:10
  —
• 0:11
  spanning
• 0:11
  GPT-2
• 0:12
  through
• 0:12
  DeepSeek
• 0:13
  V3.2,
• 0:15
  Qwen3.5,
• 0:17
  and
• 0:17
  Mistral
• 0:17
  4
• 0:18
  —
• 0:18
  nearly
• 0:19
  every
• 0:19
  model
• 0:19
  released
• 0:20
  in
• 0:20
  2026
• 0:20
  uses
• 0:21
  some
• 0:21
  form
• 0:21
  of
• 0:21
  sparse
• 0:22
  MoE,
• 0:22
  often
• 0:23
  hybridized
• 0:23
  with
• 0:23
  sliding-window
• 0:24
  attention,
• 0:25
  multi-latent
• 0:25
  attention
• 0:26
  (MLA),
• 0:26
  or
• 0:26
  state-space
• 0:27
  layers
• 0:27
  .
• 0:28
  The
• 0:28
  gallery
• 0:28
  is
• 0:28
  genuinely
• 0:29
  useful
• 0:29
  as
• 0:29
  a
• 0:29
  reference
• 0:30
  wall:
• 0:31
  each
• 0:31
  model
• 0:31
  gets
• 0:32
  a
• 0:32
  one-panel
• 0:32
  architecture
• 0:33
  figure
• 0:33
  plus
• 0:33
  a
• 0:33
  fact
• 0:33
  sheet,
• 0:34
  and
• 0:35
  the
• 0:35
  visual
• 0:35
  progression
• 0:35
  from
• 0:36
  GPT-2's
• 0:36
  dense
• 0:37
  MHA
• 0:37
  baseline
• 0:38
  to
• 0:38
  today's
• 0:38
  128-expert
• 0:39
  routed
• 0:39
  monsters
• 0:40
  tells
• 0:41
  the
• 0:41
  story
• 0:41
  of
• 0:41
  five
• 0:41
  years
• 0:42
  of
• 0:42
  decoder
• 0:42
  evolution
• 0:43
  in
• 0:43
  a
• 0:43
  single
• 0:43
  scroll.
• 0:44
  Mistral's
• 0:45
  latest
• 0:45
  entry
• 0:45
  lands
• 0:45
  right
• 0:46
  on
• 0:46
  the
• 0:46
  convergence
• 0:46
  line.
• 0:47
  Mistral
• 0:47
  Small
• 0:48
  4
• 0:48
  is
• 0:48
  a
• 0:48
  119B-parameter
• 0:50
  MoE
• 0:50
  with
• 0:50
  128
• 0:51
  experts
• 0:52
  and
• 0:52
  4
• 0:52
  active
• 0:52
  per
• 0:52
  token,
• 0:53
  yielding
• 0:54
  just
• 0:54
  6.5B
• 0:55
  activated
• 0:55
  parameters
• 0:55
  —
• 0:56
  small
• 0:57
  enough
• 0:57
  for
• 0:57
  a
• 0:57
  single
• 0:57
  4090
• 0:58
  plus
• 0:58
  128
• 0:59
  GB
• 1:00
  of
• 1:00
  system
• 1:00
  RAM,
• 1:00
  according
• 1:01
  to
• 1:01
  enthusiastic
• 1:02
  LocalLLaMA
• 1:02
  commenters.
• 1:03
  It
• 1:04
  supports
• 1:04
  256K
• 1:05
  context,
• 1:06
  multimodal
• 1:06
  input,
• 1:07
  configurable
• 1:07
  reasoning
• 1:08
  effort,
• 1:08
  and
• 1:08
  native
• 1:08
  function
• 1:09
  calling,
• 1:09
  all
• 1:10
  under
• 1:10
  Apache
• 1:10
  2.0
• 1:11
  .
• 1:12
  Raschka's
• 1:12
  gallery
• 1:13
  entry
• 1:13
  notes
• 1:13
  the
• 1:13
  design
• 1:14
  is
• 1:14
  essentially
• 1:14
  a
• 1:14
  "near-clone
• 1:15
  of
• 1:15
  DeepSeek
• 1:16
  V3
• 1:16
  with
• 1:16
  larger
• 1:17
  experts,
• 1:18
  fewer
• 1:18
  routed
• 1:18
  experts,
• 1:19
  and
• 1:19
  multimodal
• 1:19
  support."
• 1:21
  The
• 1:21
  DeepSeek
• 1:21
  playbook
• 1:21
  is
• 1:22
  now
• 1:22
  the
• 1:22
  industry
• 1:22
  default
• 1:23
  template.
• 1:24
  But
• 1:24
  knowing
• 1:24
  that
• 1:24
  MoE
• 1:25
  is
• 1:25
  the
• 1:25
  right
• 1:25
  macro
• 1:25
  architecture
• 1:26
  does
• 1:26
  not
• 1:26
  tell
• 1:26
  you
• 1:26
  what
• 1:27
  happens
• 1:27
  inside.
• 1:28
  Drew
• 1:28
  Smith
• 1:29
  spent
• 1:29
  a
• 1:29
  weekend
• 1:29
  performing
• 1:30
  layer
• 1:30
  surgery
• 1:30
  on
• 1:30
  six
• 1:31
  different
• 1:31
  model
• 1:31
  architectures
• 1:31
  —
• 1:32
  Dense
• 1:33
  32B,
• 1:33
  Hybrid
• 1:34
  9B,
• 1:34
  MoE
• 1:35
  30B,
• 1:35
  Dense
• 1:36
  3B,
• 1:36
  and
• 1:36
  a
• 1:36
  cross-model
• 1:37
  transplant
• 1:37
  —
• 1:38
  duplicating
• 1:38
  transformer
• 1:39
  layer
• 1:39
  blocks
• 1:40
  and
• 1:40
  benchmarking
• 1:40
  the
• 1:40
  results
• 1:41
  on
• 1:41
  15
• 1:41
  LeetCode-style
• 1:42
  problems.
• 1:43
  The
• 1:43
  headline
• 1:43
  finding:
• 1:44
  a
• 1:44
  universal
• 1:45
  "danger
• 1:45
  zone"
• 1:45
  at
• 1:45
  approximately
• 1:46
  50-56%
• 1:48
  depth
• 1:49
  that
• 1:49
  kills
• 1:49
  every
• 1:49
  architecture
• 1:50
  tested.
• 1:51
  Duplicate
• 1:51
  those
• 1:51
  middle
• 1:52
  layers
• 1:52
  and
• 1:52
  reasoning
• 1:53
  collapses.
• 1:54
  Delete
• 1:54
  them
• 1:54
  and
• 1:54
  output
• 1:55
  dies
• 1:55
  entirely.
• 1:56
  They
• 1:56
  are
• 1:56
  the
• 1:56
  model's
• 1:57
  attention
• 1:57
  routing
• 1:58
  infrastructure
• 1:58
  —
• 1:59
  the
• 1:59
  wiring
• 1:59
  between
• 1:59
  circuits,
• 2:00
  not
• 2:00
  a
• 2:00
  circuit
• 2:01
  you
• 2:01
  can
• 2:01
  copy
• 2:01
  .
• 2:02
  The
• 2:02
  practical
• 2:02
  results
• 2:03
  are
• 2:03
  more
• 2:03
  nuanced
• 2:03
  than
• 2:03
  the
• 2:04
  headline.
• 2:05
  On
• 2:05
  a
• 2:05
  weak-enough
• 2:05
  model
• 2:05
  (Hybrid
• 2:06
  9B
• 2:06
  scoring
• 2:06
  4/10
• 2:07
  baseline),
• 2:08
  duplicating
• 2:08
  layers
• 2:09
  at
• 2:09
  75-84%
• 2:11
  depth
• 2:11
  yielded
• 2:11
  a
• 2:11
  75%
• 2:12
  capability
• 2:13
  improvement
• 2:13
  —
• 2:13
  three
• 2:14
  new
• 2:14
  problems
• 2:14
  solved,
• 2:15
  nothing
• 2:15
  lost.
• 2:16
  On
• 2:16
  the
• 2:16
  MoE
• 2:17
  30B,
• 2:17
  the
• 2:17
  optimal
• 2:18
  duplication
• 2:18
  point
• 2:18
  shifted
• 2:19
  earlier
• 2:19
  to
• 2:19
  38-44%
• 2:21
  depth,
• 2:21
  because
• 2:22
  expert
• 2:22
  routing
• 2:22
  creates
• 2:23
  implicit
• 2:23
  depth
• 2:23
  through
• 2:24
  selection.
• 2:25
  Triple-stacking
• 2:25
  the
• 2:25
  best
• 2:26
  block
• 2:26
  produced
• 2:26
  garbage.
• 2:27
  Cross-model
• 2:28
  layer
• 2:28
  transplant
• 2:28
  was
• 2:29
  a
• 2:29
  "hard
• 2:29
  no"
• 2:29
  in
• 2:29
  all
• 2:30
  six
• 2:30
  variants
• 2:30
  —
• 2:31
  matching
• 2:32
  tensor
• 2:32
  dimensions
• 2:32
  is
• 2:33
  necessary
• 2:33
  but
• 2:33
  not
• 2:33
  sufficient,
• 2:34
  since
• 2:34
  layers
• 2:35
  develop
• 2:35
  model-specific
• 2:36
  internal
• 2:36
  representations
• 2:37
  during
• 2:37
  training.
• 2:38
  The
• 2:38
  minimum
• 2:39
  viable
• 2:39
  model
• 2:39
  for
• 2:39
  any
• 2:40
  benefit:
• 2:40
  roughly
• 2:41
  3B
• 2:41
  parameters.
• 2:42
  Below
• 2:42
  that,
• 2:43
  there
• 2:43
  are
• 2:43
  not
• 2:43
  enough
• 2:43
  functional
• 2:44
  circuits
• 2:44
  to
• 2:44
  have
• 2:44
  spare
• 2:45
  reasoning
• 2:45
  capacity
• 2:45
  worth
• 2:46
  duplicating.
• 2:47
  Meanwhile,
• 2:48
  NVIDIA's
• 2:48
  Nemotron
• 2:49
  3
• 2:49
  Super
• 2:49
  illustrates
• 2:50
  the
• 2:50
  political
• 2:50
  dimension
• 2:51
  of
• 2:51
  the
• 2:51
  MoE
• 2:51
  convergence.
• 2:53
  Early
• 2:53
  users
• 2:53
  report
• 2:54
  the
• 2:54
  model
• 2:54
  feels
• 2:54
  "overly
• 2:55
  locked
• 2:55
  down,"
• 2:56
  refusing
• 2:56
  even
• 2:56
  obviously
• 2:57
  creative
• 2:57
  or
• 2:58
  absurd
• 2:58
  prompts.
• 2:59
  The
• 2:59
  community
• 3:00
  diagnosis
• 3:00
  is
• 3:00
  blunt:
• 3:01
  NVIDIA's
• 3:02
  primary
• 3:02
  customer
• 3:03
  is
• 3:03
  the
• 3:03
  enterprise
• 3:03
  IT
• 3:04
  department
• 3:04
  signing
• 3:04
  off
• 3:05
  on
• 3:05
  a
• 3:05
  $200K
• 3:06
  hardware
• 3:06
  deal,
• 3:07
  not
• 3:07
  the
• 3:07
  developer
• 3:08
  who
• 3:08
  wants
• 3:08
  flexibility.
• 3:09
  One
• 3:09
  commenter
• 3:10
  noted
• 3:10
  the
• 3:10
  license
• 3:11
  even
• 3:11
  revokes
• 3:11
  usage
• 3:11
  rights
• 3:12
  if
• 3:12
  you
• 3:12
  attempt
• 3:12
  to
• 3:12
  modify
• 3:13
  the
• 3:13
  restrictions
• 3:13
  .
• 3:14
  The
• 3:14
  architecture
• 3:15
  may
• 3:15
  be
• 3:15
  converging,
• 3:16
  but
• 3:16
  the
• 3:16
  governance
• 3:16
  philosophy
• 3:17
  is
• 3:17
  diverging
• 3:17
  hard
• 3:18
  —
• 3:18
  Apache
• 3:19
  2.0
• 3:20
  on
• 3:20
  one
• 3:20
  end,
• 3:21
  procurement-optimized
• 3:22
  lockdown
• 3:22
  on
• 3:22
  the
• 3:22
  other.
• 3:23
  Nemotron
• 3:24
  3
• 3:24
  might
• 3:24
  only
• 3:24
  become
• 3:24
  usable
• 3:25
  when
• 3:25
  abliteration/derestriction
• 3:32
  techniques
• 3:33
  are
• 3:33
  applied.
• 3:36
  A
• 3:36
  former
• 3:37
  backend
• 3:37
  lead
• 3:37
  at
• 3:37
  Manus
• 3:38
  —
• 3:38
  the
• 3:39
  agent
• 3:39
  startup
• 3:39
  acquired
• 3:40
  by
• 3:40
  Meta
• 3:40
  —
• 3:41
  has
• 3:41
  published
• 3:42
  a
• 3:42
  detailed
• 3:42
  post-mortem
• 3:43
  of
• 3:43
  a
• 3:43
  design
• 3:43
  decision
• 3:44
  that
• 3:44
  runs
• 3:45
  counter
• 3:45
  to
• 3:45
  the
• 3:45
  entire
• 3:46
  function-calling
• 3:46
  ecosystem:
• 3:47
  he
• 3:47
  stopped
• 3:48
  using
• 3:48
  structured
• 3:48
  tool
• 3:49
  calls
• 3:49
  entirely
• 3:50
  and
• 3:50
  replaced
• 3:51
  them
• 3:51
  with
• 3:51
  a
• 3:51
  single
• 3:51
  run(command="...")
• 3:52
  interface
• 3:53
  backed
• 3:53
  by
• 3:53
  Unix
• 3:54
  CLI
• 3:54
  semantics
• 3:54
  .
• 3:55
  The
• 3:55
  argument
• 3:56
  is
• 3:56
  that
• 3:56
  Unix
• 3:56
  and
• 3:56
  LLMs
• 3:57
  independently
• 3:57
  converged
• 3:58
  on
• 3:58
  the
• 3:58
  same
• 3:59
  interface
• 3:59
  model
• 3:59
  fifty
• 4:00
  years
• 4:00
  apart
• 4:00
  —
• 4:01
  everything
• 4:01
  is
• 4:01
  a
• 4:01
  text
• 4:01
  stream.
• 4:02
  CLI
• 4:02
  commands
• 4:03
  are
• 4:03
  the
• 4:03
  densest
• 4:03
  tool-use
• 4:04
  pattern
• 4:04
  in
• 4:04
  LLM
• 4:05
  training
• 4:05
  data,
• 4:05
  pipes
• 4:06
  natively
• 4:06
  support
• 4:06
  composition,
• 4:07
  and
• 4:07
  a
• 4:07
  chain
• 4:08
  like
• 4:08
  cat
• 4:08
  log.txt
• 4:09
  |
• 4:10
  grep
• 4:10
  ERROR
• 4:11
  |
• 4:11
  wc
• 4:12
  -l
• 4:12
  replaces
• 4:13
  three
• 4:13
  separate
• 4:13
  function
• 4:13
  calls
• 4:14
  with
• 4:14
  one.
• 4:15
  The
• 4:15
  post's
• 4:15
  real
• 4:15
  contribution
• 4:16
  is
• 4:16
  the
• 4:16
  two-layer
• 4:17
  architecture
• 4:17
  separating
• 4:18
  Unix
• 4:18
  execution
• 4:19
  semantics
• 4:19
  from
• 4:19
  LLM
• 4:20
  presentation
• 4:20
  constraints.
• 4:21
  Layer
• 4:21
  1
• 4:22
  keeps
• 4:22
  pipe
• 4:22
  data
• 4:22
  raw
• 4:23
  and
• 4:23
  lossless
• 4:23
  —
• 4:24
  if
• 4:24
  you
• 4:24
  truncate
• 4:24
  cat
• 4:25
  output
• 4:25
  before
• 4:25
  piping
• 4:26
  to
• 4:26
  grep,
• 4:26
  you
• 4:26
  get
• 4:27
  incomplete
• 4:27
  results.
• 4:27
  Layer
• 4:28
  2
• 4:28
  handles
• 4:28
  binary
• 4:29
  guards,
• 4:29
  overflow
• 4:30
  truncation
• 4:30
  (200
• 4:31
  lines,
• 4:31
  with
• 4:31
  the
• 4:32
  full
• 4:32
  output
• 4:32
  written
• 4:32
  to
• 4:32
  a
• 4:32
  temp
• 4:33
  file
• 4:33
  the
• 4:33
  agent
• 4:33
  can
• 4:33
  grep
• 4:34
  through),
• 4:34
  metadata
• 4:35
  footers
• 4:35
  with
• 4:35
  exit
• 4:36
  codes
• 4:36
  and
• 4:36
  duration,
• 4:37
  and
• 4:37
  stderr
• 4:37
  attachment.
• 4:38
  The
• 4:38
  production
• 4:39
  war
• 4:39
  stories
• 4:39
  are
• 4:39
  illuminating:
• 4:40
  a
• 4:40
  PNG
• 4:41
  fed
• 4:41
  through
• 4:41
  cat
• 4:41
  produced
• 4:42
  182KB
• 4:44
  of
• 4:44
  meaningless
• 4:44
  tokens
• 4:45
  that
• 4:45
  caused
• 4:45
  20
• 4:45
  iterations
• 4:46
  of
• 4:46
  thrashing.
• 4:47
  Silent
• 4:47
  stderr
• 4:47
  on
• 4:48
  a
• 4:48
  failed
• 4:48
  pip
• 4:48
  install
• 4:49
  produced
• 4:49
  10
• 4:49
  blind
• 4:50
  retries
• 4:50
  across
• 4:50
  five
• 4:51
  package
• 4:51
  managers.
• 4:52
  These
• 4:52
  are
• 4:52
  the
• 4:52
  kinds
• 4:53
  of
• 4:53
  failures
• 4:53
  that
• 4:53
  typed
• 4:54
  function
• 4:54
  schemas
• 4:55
  are
• 4:55
  supposed
• 4:55
  to
• 4:55
  prevent
• 4:55
  —
• 4:56
  but
• 4:56
  in
• 4:56
  practice,
• 4:57
  the
• 4:57
  Unix
• 4:58
  approach
• 4:58
  with
• 4:58
  proper
• 4:59
  guardrails
• 4:59
  produced
• 4:59
  more
• 5:00
  reliable
• 5:00
  agents
• 5:01
  at
• 5:01
  Manus
• 5:01
  scale.
• 5:02
  The
• 5:03
  infrastructure
• 5:03
  layer
• 5:03
  below
• 5:04
  the
• 5:04
  agent
• 5:04
  interface
• 5:04
  is
• 5:05
  getting
• 5:05
  its
• 5:05
  own
• 5:05
  purpose-built
• 5:06
  tooling.
• 5:07
  Agentic
• 5:07
  Hosting
• 5:08
  is
• 5:08
  a
• 5:08
  new
• 5:08
  open-source
• 5:09
  PaaS
• 5:09
  written
• 5:09
  in
• 5:10
  Go,
• 5:10
  designed
• 5:11
  for
• 5:11
  bare-metal
• 5:11
  servers
• 5:12
  where
• 5:12
  AI
• 5:12
  agents
• 5:13
  are
• 5:13
  the
• 5:13
  operators
• 5:13
  —
• 5:14
  no
• 5:14
  web
• 5:14
  dashboard,
• 5:15
  REST
• 5:15
  API
• 5:16
  only.
• 5:17
  The
• 5:17
  stack
• 5:17
  is
• 5:17
  deliberately
• 5:18
  opinionated:
• 5:19
  one
• 5:19
  build
• 5:19
  system
• 5:19
  (Nixpacks),
• 5:20
  one
• 5:20
  router
• 5:21
  (Traefik),
• 5:21
  one
• 5:22
  container
• 5:22
  runtime
• 5:22
  (gVisor
• 5:23
  for
• 5:23
  syscall
• 5:23
  interception),
• 5:24
  SQLite
• 5:25
  for
• 5:25
  state,
• 5:25
  AES-256-GCM
• 5:29
  for
• 5:29
  secrets.
• 5:30
  Multi-tenant
• 5:31
  isolation,
• 5:32
  rate
• 5:32
  limiting,
• 5:32
  circuit
• 5:33
  breakers,
• 5:33
  and
• 5:34
  idempotency
• 5:34
  headers
• 5:35
  are
• 5:35
  built
• 5:35
  in.
• 5:36
  It
• 5:36
  ships
• 5:36
  with
• 5:36
  Claude
• 5:37
  Code
• 5:37
  skills
• 5:37
  and
• 5:37
  slash
• 5:38
  commands
• 5:38
  so
• 5:38
  agents
• 5:39
  can
• 5:39
  deploy,
• 5:39
  provision
• 5:40
  databases,
• 5:41
  and
• 5:41
  check
• 5:41
  status
• 5:41
  without
• 5:42
  writing
• 5:42
  curl
• 5:42
  .
• 5:43
  At
• 5:43
  the
• 5:43
  far
• 5:43
  end
• 5:44
  of
• 5:44
  the
• 5:44
  autonomy
• 5:44
  spectrum
• 5:45
  sits
• 5:45
  Paperclip.ing,
• 5:47
  an
• 5:47
  open-source
• 5:48
  platform
• 5:48
  for
• 5:48
  running
• 5:49
  entire
• 5:49
  autonomous
• 5:50
  companies
• 5:50
  staffed
• 5:51
  by
• 5:51
  AI
• 5:51
  agents.
• 5:52
  It
• 5:52
  provides
• 5:53
  org
• 5:53
  charts,
• 5:54
  hierarchies,
• 5:54
  reporting
• 5:55
  lines,
• 5:56
  goal
• 5:56
  decomposition
• 5:57
  (every
• 5:57
  task
• 5:57
  traces
• 5:58
  back
• 5:58
  to
• 5:58
  the
• 5:58
  company
• 5:58
  mission),
• 5:59
  per-agent
• 6:00
  monthly
• 6:00
  budgets
• 6:01
  with
• 6:01
  hard
• 6:01
  limits,
• 6:01
  ticket-based
• 6:02
  communication
• 6:03
  with
• 6:03
  full
• 6:03
  tracing,
• 6:04
  and
• 6:04
  board-level
• 6:05
  governance
• 6:05
  where
• 6:05
  humans
• 6:05
  approve
• 6:06
  hires
• 6:06
  and
• 6:06
  strategy
• 6:07
  .
• 6:08
  The
• 6:08
  pitch
• 6:08
  is
• 6:08
  that
• 6:08
  it
• 6:09
  works
• 6:09
  with
• 6:09
  any
• 6:09
  agent
• 6:10
  runtime
• 6:10
  —
• 6:11
  Claude
• 6:11
  Code
• 6:11
  sessions,
• 6:12
  Python
• 6:12
  scripts,
• 6:13
  HTTP
• 6:13
  webhooks
• 6:14
  —
• 6:14
  and
• 6:15
  that
• 6:15
  the
• 6:15
  real
• 6:15
  problem
• 6:15
  it
• 6:15
  solves
• 6:16
  is
• 6:16
  not
• 6:16
  making
• 6:17
  agents
• 6:17
  smarter
• 6:18
  but
• 6:18
  organizing
• 6:18
  them
• 6:19
  into
• 6:19
  accountable
• 6:19
  structures.
• 6:20
  Whether
• 6:21
  that
• 6:21
  organizational
• 6:22
  metaphor
• 6:22
  actually
• 6:22
  scales
• 6:23
  is
• 6:23
  an
• 6:23
  open
• 6:24
  question,
• 6:25
  but
• 6:25
  the
• 6:25
  design
• 6:25
  addresses
• 6:26
  a
• 6:26
  real
• 6:26
  gap:
• 6:27
  most
• 6:27
  agent
• 6:28
  frameworks
• 6:28
  handle
• 6:28
  execution
• 6:29
  but
• 6:29
  not
• 6:30
  governance,
• 6:30
  budgets,
• 6:31
  or
• 6:31
  persistent
• 6:32
  task
• 6:32
  ownership
• 6:33
  across
• 6:33
  reboots.
• 6:37
  NVIDIA's
• 6:38
  NeMo
• 6:38
  Retriever
• 6:38
  team
• 6:39
  has
• 6:39
  published
• 6:39
  an
• 6:39
  agentic
• 6:40
  retrieval
• 6:40
  pipeline
• 6:41
  that
• 6:41
  topped
• 6:41
  the
• 6:41
  ViDoRe
• 6:41
  v3
• 6:42
  leaderboard
• 6:43
  and
• 6:43
  placed
• 6:43
  second
• 6:43
  on
• 6:44
  the
• 6:44
  reasoning-intensive
• 6:44
  BRIGHT
• 6:45
  benchmark
• 6:45
  —
• 6:46
  using
• 6:46
  the
• 6:46
  exact
• 6:46
  same
• 6:47
  architecture
• 6:47
  for
• 6:47
  both,
• 6:48
  with
• 6:48
  no
• 6:48
  dataset-specific
• 6:49
  tuning.
• 6:50
  The
• 6:50
  core
• 6:50
  insight:
• 6:51
  dense
• 6:51
  retrieval
• 6:51
  based
• 6:52
  on
• 6:52
  semantic
• 6:52
  similarity
• 6:53
  is
• 6:53
  not
• 6:53
  enough
• 6:54
  for
• 6:54
  complex
• 6:54
  document
• 6:55
  search,
• 6:55
  which
• 6:55
  requires
• 6:56
  reasoning,
• 6:56
  iterative
• 6:57
  exploration,
• 6:58
  and
• 6:58
  real-world
• 6:58
  knowledge.
• 6:59
  The
• 6:59
  pipeline
• 7:00
  uses
• 7:00
  a
• 7:00
  ReACT
• 7:00
  loop
• 7:01
  where
• 7:01
  an
• 7:01
  LLM
• 7:01
  agent
• 7:02
  iteratively
• 7:02
  searches,
• 7:03
  evaluates,
• 7:03
  and
• 7:03
  refines
• 7:04
  queries,
• 7:04
  with
• 7:05
  Reciprocal
• 7:05
  Rank
• 7:05
  Fusion
• 7:06
  as
• 7:06
  a
• 7:06
  safety
• 7:06
  net
• 7:06
  when
• 7:07
  the
• 7:07
  agent
• 7:07
  hits
• 7:07
  step
• 7:07
  or
• 7:07
  context
• 7:08
  limits
• 7:08
  .
• 7:09
  An
• 7:09
  interesting
• 7:09
  engineering
• 7:10
  detail:
• 7:10
  the
• 7:10
  team
• 7:11
  initially
• 7:11
  exposed
• 7:12
  the
• 7:12
  retriever
• 7:12
  to
• 7:12
  the
• 7:12
  agent
• 7:13
  via
• 7:13
  a
• 7:13
  Model
• 7:13
  Context
• 7:14
  Protocol
• 7:14
  (MCP)
• 7:14
  server
• 7:15
  —
• 7:15
  the
• 7:15
  natural
• 7:16
  choice
• 7:16
  —
• 7:16
  but
• 7:17
  found
• 7:17
  that
• 7:17
  the
• 7:17
  network
• 7:17
  round-trips,
• 7:18
  server
• 7:18
  lifecycle
• 7:19
  management,
• 7:19
  and
• 7:19
  silent
• 7:20
  misconfiguration
• 7:20
  risks
• 7:21
  were
• 7:21
  too
• 7:21
  costly
• 7:21
  for
• 7:21
  experiment
• 7:22
  velocity.
• 7:23
  They
• 7:23
  replaced
• 7:23
  it
• 7:23
  with
• 7:23
  a
• 7:23
  thread-safe
• 7:24
  singleton
• 7:24
  retriever
• 7:25
  living
• 7:25
  in-process,
• 7:26
  achieving
• 7:26
  the
• 7:26
  same
• 7:27
  safe
• 7:27
  shared
• 7:27
  access
• 7:28
  without
• 7:28
  serialization
• 7:28
  overhead.
• 7:30
  The
• 7:30
  ablation
• 7:30
  studies
• 7:30
  are
• 7:31
  honest
• 7:31
  about
• 7:31
  costs:
• 7:32
  136
• 7:33
  seconds
• 7:33
  per
• 7:33
  query,
• 7:34
  roughly
• 7:34
  760K
• 7:35
  input
• 7:36
  tokens
• 7:36
  per
• 7:36
  query,
• 7:37
  with
• 7:37
  Opus
• 7:37
  4.5
• 7:38
  as
• 7:38
  the
• 7:38
  agent.
• 7:39
  Swapping
• 7:40
  to
• 7:40
  open-weight
• 7:40
  Llama
• 7:41
  produced
• 7:41
  a
• 7:41
  small
• 7:41
  accuracy
• 7:42
  drop
• 7:42
  on
• 7:42
  ViDoRe
• 7:43
  but
• 7:43
  a
• 7:43
  wider
• 7:43
  gap
• 7:44
  on
• 7:44
  BRIGHT,
• 7:45
  confirming
• 7:45
  that
• 7:45
  deep
• 7:45
  reasoning
• 7:46
  tasks
• 7:46
  still
• 7:47
  benefit
• 7:47
  heavily
• 7:47
  from
• 7:47
  frontier
• 7:48
  models.
• 7:49
  Holotron-12B
• 7:50
  from
• 7:50
  H
• 7:50
  Company
• 7:51
  takes
• 7:51
  a
• 7:51
  different
• 7:51
  path
• 7:52
  —
• 7:52
  a
• 7:53
  computer-use
• 7:53
  agent
• 7:53
  optimized
• 7:54
  for
• 7:54
  throughput
• 7:55
  rather
• 7:55
  than
• 7:55
  single-query
• 7:56
  accuracy.
• 7:57
  Post-trained
• 7:58
  from
• 7:58
  NVIDIA's
• 7:59
  Nemotron-Nano-2
• 8:00
  VL
• 8:00
  on
• 8:00
  proprietary
• 8:01
  screen-understanding
• 8:02
  and
• 8:02
  UI-navigation
• 8:03
  data,
• 8:04
  it
• 8:04
  leverages
• 8:04
  the
• 8:04
  hybrid
• 8:05
  SSM-attention
• 8:06
  architecture
• 8:06
  for
• 8:07
  dramatically
• 8:07
  lower
• 8:07
  memory
• 8:08
  footprint
• 8:08
  per
• 8:08
  sequence.
• 8:10
  On
• 8:10
  the
• 8:10
  WebVoyager
• 8:11
  benchmark
• 8:11
  at
• 8:11
  concurrency
• 8:12
  100
• 8:12
  on
• 8:12
  a
• 8:12
  single
• 8:13
  H100,
• 8:14
  Holotron-12B
• 8:15
  achieved
• 8:15
  2x
• 8:16
  the
• 8:16
  throughput
• 8:16
  of
• 8:16
  the
• 8:16
  previous
• 8:17
  Holo2-8B
• 8:18
  model
• 8:18
  (8.9K
• 8:19
  vs
• 8:20
  5.1K
• 8:20
  tokens/sec),
• 8:22
  while
• 8:22
  improving
• 8:22
  task
• 8:23
  success
• 8:23
  from
• 8:23
  35.1%
• 8:24
  to
• 8:24
  80.5%
• 8:26
  .
• 8:27
  LocoOperator-4B
• 8:28
  from
• 8:28
  LocoreMind
• 8:29
  targets
• 8:29
  the
• 8:29
  other
• 8:29
  end
• 8:29
  of
• 8:29
  the
• 8:30
  spectrum
• 8:30
  —
• 8:31
  a
• 8:31
  4B-parameter
• 8:32
  computer-use
• 8:32
  model
• 8:33
  small
• 8:33
  enough
• 8:33
  for
• 8:33
  edge
• 8:33
  deployment
• 8:34
  .
• 8:35
  MiroFish
• 8:35
  offers
• 8:35
  the
• 8:36
  most
• 8:36
  ambitious
• 8:36
  framing
• 8:37
  in
• 8:37
  this
• 8:37
  cluster:
• 8:38
  a
• 8:38
  swarm
• 8:38
  intelligence
• 8:39
  engine
• 8:39
  that
• 8:39
  extracts
• 8:40
  "seed
• 8:40
  information"
• 8:41
  from
• 8:41
  real-world
• 8:41
  data
• 8:42
  (news,
• 8:42
  policy
• 8:43
  drafts,
• 8:43
  financial
• 8:44
  signals),
• 8:45
  constructs
• 8:45
  a
• 8:45
  parallel
• 8:46
  digital
• 8:46
  world
• 8:46
  populated
• 8:47
  by
• 8:47
  thousands
• 8:48
  of
• 8:48
  agents
• 8:48
  with
• 8:48
  independent
• 8:49
  personalities
• 8:50
  and
• 8:50
  long-term
• 8:50
  memory,
• 8:51
  then
• 8:51
  lets
• 8:52
  users
• 8:52
  inject
• 8:52
  variables
• 8:53
  and
• 8:53
  observe
• 8:54
  emergent
• 8:54
  behavior.
• 8:55
  The
• 8:55
  demo
• 8:56
  shows
• 8:56
  predictions
• 8:57
  for
• 8:57
  public
• 8:57
  opinion
• 8:57
  events,
• 8:58
  stock
• 8:58
  market
• 8:59
  sentiment,
• 8:59
  and
• 9:00
  even
• 9:00
  the
• 9:00
  lost
• 9:00
  ending
• 9:01
  of
• 9:01
  Dream
• 9:01
  of
• 9:01
  the
• 9:01
  Red
• 9:02
  Chamber
• 9:02
  .
• 9:03
  The
• 9:03
  engine
• 9:03
  is
• 9:03
  built
• 9:04
  on
• 9:04
  CAMEL-AI's
• 9:08
  OASIS
• 9:08
  framework
• 9:09
  and
• 9:09
  backed
• 9:09
  by
• 9:09
  Shanda
• 9:09
  Group
• 9:10
  .
• 9:10
  Whether
• 9:10
  swarm-of-agents
• 9:11
  simulation
• 9:12
  can
• 9:12
  produce
• 9:12
  actionable
• 9:13
  predictions
• 9:14
  remains
• 9:14
  unproven,
• 9:15
  but
• 9:15
  the
• 9:15
  architecture
• 9:16
  —
• 9:17
  GraphRAG
• 9:17
  construction,
• 9:18
  persona
• 9:19
  generation,
• 9:19
  dual-platform
• 9:20
  parallel
• 9:21
  simulation,
• 9:22
  temporal
• 9:22
  memory
• 9:23
  updates
• 9:23
  —
• 9:24
  represents
• 9:24
  a
• 9:24
  serious
• 9:25
  attempt
• 9:25
  at
• 9:25
  collective
• 9:26
  agent
• 9:26
  intelligence
• 9:27
  beyond
• 9:27
  simple
• 9:28
  task
• 9:28
  delegation.
• 9:32
  A
• 9:32
  developer
• 9:32
  used
• 9:33
  Claude
• 9:33
  Code
• 9:33
  (Opus
• 9:34
  4.6,
• 9:35
  high
• 9:35
  reasoning)
• 9:35
  to
• 9:35
  reverse-engineer
• 9:36
  a
• 9:36
  13-year-old
• 9:37
  Disney
• 9:37
  Infinity
• 9:37
  game
• 9:38
  binary
• 9:38
  —
• 9:39
  no
• 9:39
  symbols,
• 9:39
  no
• 9:39
  source
• 9:40
  code,
• 9:40
  no
• 9:40
  existing
• 9:41
  RE
• 9:41
  documentation
• 9:41
  —
• 9:42
  and
• 9:42
  crack
• 9:42
  a
• 9:43
  character-playset
• 9:43
  restriction
• 9:44
  that
• 9:44
  the
• 9:44
  modding
• 9:44
  community
• 9:45
  had
• 9:45
  failed
• 9:45
  to
• 9:45
  solve
• 9:45
  for
• 9:45
  over
• 9:46
  a
• 9:46
  decade.
• 9:47
  The
• 9:47
  restriction
• 9:47
  was
• 9:47
  not
• 9:47
  a
• 9:48
  single
• 9:48
  flag:
• 9:48
  one
• 9:49
  function
• 9:49
  (FindPlaysetForCharacter)
• 9:50
  gets
• 9:50
  called
• 9:50
  at
• 9:50
  13
• 9:51
  different
• 9:51
  points
• 9:51
  across
• 9:52
  6
• 9:52
  areas
• 9:52
  of
• 9:52
  the
• 9:52
  C++
• 9:53
  code.
• 9:54
  Claude
• 9:54
  helped
• 9:54
  trace
• 9:54
  the
• 9:55
  call
• 9:55
  graph,
• 9:55
  identify
• 9:55
  all
• 9:56
  13
• 9:56
  validation
• 9:56
  sites,
• 9:57
  map
• 9:57
  which
• 9:58
  code
• 9:58
  area
• 9:58
  each
• 9:58
  belonged
• 9:59
  to,
• 9:59
  and
• 9:59
  determine
• 9:59
  the
• 9:59
  exact
• 10:00
  bytes
• 10:00
  to
• 10:00
  patch.
• 10:01
  The
• 10:01
  result:
• 10:02
  17
• 10:02
  binary
• 10:03
  patches
• 10:03
  plus
• 10:03
  3
• 10:03
  modified
• 10:04
  data
• 10:04
  files,
• 10:05
  completed
• 10:05
  in
• 10:05
  under
• 10:05
  24
• 10:06
  hours
• 10:06
  .
• 10:07
  The
• 10:07
  community
• 10:07
  reaction
• 10:08
  —
• 10:08
  90+
• 10:09
  upvotes,
• 10:09
  the
• 10:09
  most
• 10:10
  well-known
• 10:10
  modder
• 10:10
  calling
• 10:11
  it
• 10:11
  "better
• 10:11
  than
• 10:11
  my
• 10:11
  method"
• 10:11
  —
• 10:12
  is
• 10:12
  genuine,
• 10:13
  though
• 10:13
  some
• 10:13
  experienced
• 10:13
  reverse
• 10:14
  engineers
• 10:15
  note
• 10:15
  the
• 10:15
  problem
• 10:15
  may
• 10:15
  have
• 10:15
  been
• 10:16
  more
• 10:16
  about
• 10:16
  nobody
• 10:16
  with
• 10:16
  the
• 10:17
  right
• 10:17
  skills
• 10:17
  bothering
• 10:17
  to
• 10:17
  spend
• 10:18
  the
• 10:18
  time
• 10:18
  than
• 10:18
  about
• 10:18
  technical
• 10:19
  impossibility.
• 10:20
  The
• 10:20
  real
• 10:20
  takeaway
• 10:21
  is
• 10:21
  that
• 10:21
  AI
• 10:21
  coding
• 10:22
  assistants
• 10:22
  are
• 10:22
  now
• 10:23
  useful
• 10:23
  force
• 10:23
  multipliers
• 10:24
  for
• 10:24
  tedious,
• 10:25
  high-context
• 10:25
  tasks
• 10:26
  like
• 10:26
  binary
• 10:26
  analysis,
• 10:27
  not
• 10:27
  just
• 10:28
  greenfield
• 10:28
  code
• 10:28
  generation.
• 10:29
  For
• 10:29
  those
• 10:29
  spending
• 10:29
  serious
• 10:30
  time
• 10:30
  with
• 10:30
  AI
• 10:30
  coding
• 10:31
  assistants,
• 10:31
  a
• 10:32
  2,000-hour
• 10:33
  Claude
• 10:33
  Code
• 10:33
  practitioner
• 10:34
  has
• 10:34
  distilled
• 10:34
  his
• 10:35
  workflow
• 10:35
  into
• 10:35
  the
• 10:35
  WHISK
• 10:36
  framework:
• 10:36
  Write
• 10:37
  (externalize
• 10:37
  agent
• 10:38
  memory
• 10:38
  via
• 10:38
  git
• 10:38
  log
• 10:39
  and
• 10:39
  structured
• 10:39
  plans),
• 10:40
  Isolate
• 10:41
  (use
• 10:41
  sub-agents
• 10:41
  for
• 10:41
  research
• 10:42
  to
• 10:42
  keep
• 10:42
  the
• 10:42
  main
• 10:42
  context
• 10:43
  clean),
• 10:43
  Select
• 10:44
  (layered
• 10:44
  just-in-time
• 10:45
  context
• 10:45
  loading
• 10:46
  via
• 10:46
  global
• 10:46
  rules,
• 10:47
  on-demand
• 10:47
  docs,
• 10:48
  skills,
• 10:48
  and
• 10:48
  prime
• 10:48
  commands),
• 10:49
  and
• 10:49
  Compress
• 10:49
  (handoff
• 10:50
  documents
• 10:50
  and
• 10:51
  focused
• 10:51
  compaction
• 10:51
  as
• 10:51
  a
• 10:51
  last
• 10:52
  resort).
• 10:53
  The
• 10:53
  core
• 10:53
  thesis:
• 10:54
  80%
• 10:54
  of
• 10:54
  coding
• 10:55
  agent
• 10:55
  failures
• 10:55
  stem
• 10:56
  from
• 10:56
  poor
• 10:56
  context
• 10:57
  management,
• 10:57
  and
• 10:58
  the
• 10:58
  new
• 10:58
  1M
• 10:58
  token
• 10:58
  window
• 10:59
  does
• 10:59
  not
• 10:59
  fix
• 10:59
  the
• 10:59
  needle-in-haystack
• 11:00
  problem
• 11:00
  —
• 11:01
  distractors
• 11:02
  from
• 11:02
  similar
• 11:02
  code
• 11:03
  patterns
• 11:03
  cause
• 11:03
  confident
• 11:04
  hallucinations
• 11:04
  .
• 11:05
  The
• 11:05
  most
• 11:06
  practical
• 11:06
  insight
• 11:07
  is
• 11:07
  using
• 11:07
  sub-agents
• 11:08
  for
• 11:08
  research
• 11:08
  in
• 11:08
  parallel,
• 11:09
  then
• 11:09
  loading
• 11:10
  only
• 11:10
  the
• 11:10
  500-token
• 11:11
  summary
• 11:11
  into
• 11:11
  the
• 11:11
  main
• 11:12
  context
• 11:12
  —
• 11:13
  a
• 11:13
  90.2%
• 11:14
  token
• 11:14
  reduction
• 11:15
  that
• 11:15
  Anthropic's
• 11:15
  own
• 11:16
  research
• 11:16
  supports.
• 11:17
  On
• 11:17
  the
• 11:18
  tooling
• 11:18
  side,
• 11:18
  an
• 11:19
  Open
• 11:19
  WebUI
• 11:19
  community
• 11:20
  member
• 11:20
  built
• 11:21
  an
• 11:21
  Inline
• 11:21
  Visualizer
• 11:22
  plugin
• 11:22
  that
• 11:22
  replicates
• 11:23
  Anthropic's
• 11:23
  new
• 11:24
  interactive
• 11:24
  inline
• 11:24
  visuals
• 11:25
  for
• 11:25
  any
• 11:25
  model.
• 11:26
  The
• 11:26
  killer
• 11:26
  feature
• 11:27
  is
• 11:27
  a
• 11:27
  sendPrompt
• 11:27
  JavaScript
• 11:28
  bridge
• 11:28
  that
• 11:28
  lets
• 11:29
  elements
• 11:29
  inside
• 11:29
  rendered
• 11:30
  HTML/SVG
• 11:31
  send
• 11:31
  messages
• 11:32
  back
• 11:32
  to
• 11:32
  the
• 11:32
  chat
• 11:32
  —
• 11:33
  click
• 11:34
  a
• 11:34
  node
• 11:34
  in
• 11:34
  an
• 11:34
  architecture
• 11:34
  diagram
• 11:35
  and
• 11:35
  the
• 11:35
  model
• 11:36
  explains
• 11:36
  that
• 11:36
  component.
• 11:37
  Tested
• 11:38
  with
• 11:38
  Claude
• 11:38
  Haiku
• 11:38
  4.5
• 11:39
  and
• 11:39
  Qwen3.5-35B-A3B,
• 11:44
  no
• 11:44
  external
• 11:44
  dependencies
• 11:45
  required
• 11:45
  .
• 11:46
  Meanwhile,
• 11:47
  Agentic
• 11:47
  QE
• 11:48
  v3.8.0
• 11:50
  continues
• 11:51
  the
• 11:51
  evolution
• 11:51
  of
• 11:51
  AI-orchestrated
• 11:52
  quality
• 11:52
  engineering,
• 11:53
  now
• 11:53
  featuring
• 11:54
  18
• 11:54
  specialized
• 11:55
  agents
• 11:55
  built
• 11:55
  on
• 11:56
  LionAGI
• 11:56
  orchestration
• 11:57
  with
• 11:57
  PACT
• 11:57
  principles
• 11:57
  —
• 11:58
  Proactive,
• 11:59
  Autonomous,
• 11:59
  Collaborative,
• 12:00
  Targeted.
• 12:01
  The
• 12:01
  framework
• 12:01
  bridges
• 12:02
  London
• 12:02
  School
• 12:02
  (mockist)
• 12:03
  and
• 12:03
  Chicago
• 12:03
  School
• 12:04
  (classicist)
• 12:04
  TDD
• 12:05
  approaches,
• 12:05
  with
• 12:06
  session-based
• 12:06
  test
• 12:06
  management
• 12:07
  and
• 12:07
  self-healing
• 12:07
  test
• 12:08
  suites
• 12:08
  .
• 12:10
  A
• 12:10
  new
• 12:11
  arxiv
• 12:11
  paper
• 12:11
  audits
• 12:12
  what
• 12:12
  LLMs
• 12:12
  actually
• 12:12
  associate
• 12:13
  with
• 12:13
  people's
• 12:13
  names
• 12:14
  —
• 12:14
  and
• 12:14
  the
• 12:15
  results
• 12:15
  should
• 12:15
  concern
• 12:15
  anyone
• 12:16
  who
• 12:16
  has
• 12:16
  ever
• 12:16
  interacted
• 12:17
  with
• 12:17
  a
• 12:17
  chatbot.
• 12:18
  The
• 12:18
  researchers
• 12:18
  introduced
• 12:19
  LMP2
• 12:19
  (Language
• 12:20
  Model
• 12:20
  Privacy
• 12:20
  Probe),
• 12:21
  a
• 12:21
  human-centered
• 12:22
  audit
• 12:22
  tool
• 12:22
  tested
• 12:23
  across
• 12:23
  eight
• 12:23
  LLMs
• 12:24
  including
• 12:24
  GPT-4o.
• 12:26
  For
• 12:26
  well-known
• 12:26
  individuals,
• 12:27
  models
• 12:27
  confidently
• 12:28
  generate
• 12:28
  values
• 12:29
  across
• 12:29
  multiple
• 12:29
  personal
• 12:30
  data
• 12:30
  categories.
• 12:31
  For
• 12:31
  everyday
• 12:32
  EU
• 12:32
  residents,
• 12:33
  GPT-4o
• 12:33
  generated
• 12:34
  11
• 12:34
  features
• 12:34
  with
• 12:35
  60%
• 12:35
  accuracy
• 12:35
  —
• 12:36
  including
• 12:37
  gender,
• 12:37
  hair
• 12:37
  color,
• 12:37
  and
• 12:38
  languages
• 12:38
  spoken.
• 12:39
  The
• 12:39
  critical
• 12:39
  finding:
• 12:40
  72%
• 12:41
  of
• 12:41
  participants
• 12:41
  wanted
• 12:42
  control
• 12:42
  over
• 12:42
  model-generated
• 12:43
  associations
• 12:44
  with
• 12:44
  their
• 12:44
  name,
• 12:44
  raising
• 12:45
  hard
• 12:45
  questions
• 12:45
  about
• 12:45
  whether
• 12:46
  data
• 12:46
  privacy
• 12:46
  rights
• 12:46
  under
• 12:47
  GDPR
• 12:47
  should
• 12:47
  extend
• 12:48
  to
• 12:48
  inferred
• 12:48
  associations
• 12:49
  in
• 12:49
  LLM
• 12:49
  weights,
• 12:50
  not
• 12:50
  just
• 12:50
  stored
• 12:50
  data
• 12:51
  .
• 12:51
  The
• 12:51
  paper
• 12:52
  bridges
• 12:52
  the
• 12:52
  gap
• 12:52
  between
• 12:53
  theoretical
• 12:53
  memorization
• 12:54
  risks
• 12:54
  and
• 12:54
  empirical
• 12:54
  measurement
• 12:55
  of
• 12:55
  what
• 12:55
  models
• 12:55
  actually
• 12:56
  retain
• 12:56
  about
• 12:56
  specific
• 12:57
  individuals.
• 12:58
  On
• 12:58
  a
• 12:58
  more
• 12:58
  immediately
• 12:59
  painful
• 12:59
  note:
• 13:00
  if
• 13:00
  you
• 13:00
  use
• 13:00
  Langfuse
• 13:01
  for
• 13:01
  LLM
• 13:01
  observability
• 13:02
  alongside
• 13:03
  evaluation
• 13:03
  tools
• 13:03
  like
• 13:03
  DeepEval,
• 13:04
  check
• 13:05
  your
• 13:05
  usage
• 13:05
  dashboard.
• 13:06
  The
• 13:06
  Langfuse
• 13:07
  SDK
• 13:07
  attaches
• 13:08
  to
• 13:08
  the
• 13:08
  global
• 13:08
  OpenTelemetry
• 13:09
  TracerProvider
• 13:09
  and
• 13:10
  greedily
• 13:10
  intercepts
• 13:10
  any
• 13:11
  span
• 13:11
  with
• 13:11
  gen
• 13:11
  ai.
• 13:12
  attributes
• 13:13
  —
• 13:13
  even
• 13:14
  from
• 13:14
  completely
• 13:14
  unrelated
• 13:15
  tools
• 13:15
  running
• 13:15
  in
• 13:15
  the
• 13:15
  same
• 13:16
  process.
• 13:17
  Because
• 13:17
  Langfuse
• 13:17
  has
• 13:18
  per-trace
• 13:18
  pricing,
• 13:19
  this
• 13:19
  silently
• 13:19
  inflates
• 13:20
  bills
• 13:20
  with
• 13:20
  third-party
• 13:21
  background
• 13:21
  data.
• 13:22
  The
• 13:22
  fix
• 13:23
  is
• 13:23
  a
• 13:23
  should
• 13:23
  export
• 13:23
  span
• 13:24
  filter
• 13:24
  that
• 13:24
  locks
• 13:24
  the
• 13:25
  span
• 13:25
  processor
• 13:25
  to
• 13:25
  only
• 13:26
  accept
• 13:26
  Langfuse
• 13:26
  SDK
• 13:27
  calls,
• 13:28
  which
• 13:28
  is
• 13:28
  arguably
• 13:28
  the
• 13:28
  default
• 13:29
  it
• 13:29
  should
• 13:29
  have
• 13:29
  shipped
• 13:29
  with
• 13:29
  .
• 13:30
  This
• 13:30
  pattern
• 13:31
  —
• 13:32
  SDK
• 13:32
  instrumentation
• 13:33
  that
• 13:33
  captures
• 13:33
  more
• 13:33
  than
• 13:34
  disclosed
• 13:34
  —
• 13:35
  is
• 13:35
  becoming
• 13:35
  a
• 13:35
  recurring
• 13:36
  theme
• 13:36
  as
• 13:36
  the
• 13:36
  observability
• 13:37
  stack
• 13:37
  for
• 13:37
  AI
• 13:38
  applications
• 13:38
  grows
• 13:38
  more
• 13:39
  complex.
• 13:40
  The
• 13:40
  Ministry
• 13:40
  of
• 13:40
  Testing's
• 13:41
  latest
• 13:41
  newsletter
• 13:42
  highlights
• 13:42
  a
• 13:42
  related
• 13:42
  trust
• 13:43
  boundary:
• 13:44
  authentication
• 13:44
  testing.
• 13:46
  Viola
• 13:46
  Lykova
• 13:47
  argues
• 13:47
  that
• 13:47
  teams
• 13:47
  spend
• 13:48
  disproportionate
• 13:48
  effort
• 13:49
  testing
• 13:49
  login
• 13:49
  page
• 13:50
  UI
• 13:50
  while
• 13:50
  neglecting
• 13:51
  high-impact
• 13:51
  auth
• 13:52
  journey
• 13:52
  tests
• 13:52
  —
• 13:53
  the
• 13:53
  actual
• 13:54
  security-critical
• 13:54
  paths
• 13:55
  like
• 13:55
  session
• 13:55
  fixation,
• 13:56
  token
• 13:57
  refresh
• 13:57
  races,
• 13:58
  and
• 13:58
  credential
• 13:58
  stuffing
• 13:59
  resistance.
• 14:00
  The
• 14:00
  Agentic
• 14:01
  QE
• 14:01
  angle
• 14:01
  here
• 14:02
  is
• 14:02
  pointed:
• 14:03
  as
• 14:03
  AI
• 14:03
  agents
• 14:03
  increasingly
• 14:04
  handle
• 14:04
  authentication
• 14:05
  flows,
• 14:06
  the
• 14:06
  gap
• 14:06
  between
• 14:07
  "testing
• 14:07
  the
• 14:07
  login
• 14:08
  form"
• 14:08
  and
• 14:09
  "testing
• 14:09
  the
• 14:09
  auth
• 14:09
  system"
• 14:10
  becomes
• 14:10
  an
• 14:10
  exploitable
• 14:11
  surface
• 14:11
  .
• 14:12
  The
• 14:13
  broader
• 14:13
  pattern
• 14:13
  across
• 14:14
  this
• 14:14
  cluster
• 14:15
  is
• 14:15
  an
• 14:15
  erosion
• 14:15
  of
• 14:15
  developer
• 14:16
  trust
• 14:16
  on
• 14:16
  three
• 14:17
  surfaces
• 14:17
  simultaneously:
• 14:19
  models
• 14:20
  memorize
• 14:20
  what
• 14:20
  they
• 14:20
  should
• 14:21
  not,
• 14:21
  SDKs
• 14:22
  phone
• 14:22
  home
• 14:23
  in
• 14:23
  ways
• 14:23
  that
• 14:23
  generate
• 14:24
  surprise
• 14:24
  bills,
• 14:24
  and
• 14:25
  auth
• 14:25
  systems
• 14:25
  built
• 14:25
  for
• 14:26
  human
• 14:26
  cadence
• 14:26
  fail
• 14:27
  under
• 14:27
  agentic
• 14:27
  speed.
• 14:31
  Trump
• 14:31
  Code
• 14:31
  is
• 14:32
  exactly
• 14:32
  what
• 14:32
  it
• 14:32
  sounds
• 14:33
  like:
• 14:33
  an
• 14:34
  open-source
• 14:34
  system
• 14:34
  that
• 14:35
  applies
• 14:35
  brute-force
• 14:36
  computation
• 14:37
  to
• 14:37
  find
• 14:37
  statistically
• 14:38
  significant
• 14:38
  patterns
• 14:39
  between
• 14:39
  presidential
• 14:39
  social
• 14:40
  media
• 14:40
  posts
• 14:40
  and
• 14:40
  stock
• 14:41
  market
• 14:41
  movements.
• 14:42
  Built
• 14:43
  by
• 14:43
  a
• 14:43
  Taiwanese
• 14:43
  developer
• 14:44
  based
• 14:44
  in
• 14:44
  Japan,
• 14:45
  it
• 14:45
  has
• 14:45
  analyzed
• 14:46
  7,400+
• 14:46
  Truth
• 14:47
  Social
• 14:47
  posts,
• 14:47
  tested
• 14:48
  31.5
• 14:49
  million
• 14:49
  model
• 14:49
  combinations,
• 14:50
  and
• 14:51
  distilled
• 14:51
  551
• 14:52
  surviving
• 14:52
  rules
• 14:53
  that
• 14:53
  passed
• 14:53
  train/test
• 14:53
  validation
• 14:54
  —
• 14:55
  yielding
• 14:55
  a
• 14:56
  claimed
• 14:56
  61.3%
• 14:57
  hit
• 14:57
  rate
• 14:57
  across
• 14:57
  566
• 14:59
  verified
• 14:59
  predictions
• 15:00
  (z=5.39).
• 15:02
  The
• 15:02
  key
• 15:02
  discoveries
• 15:02
  include
• 15:03
  that
• 15:03
  pre-market
• 15:03
  "relief"
• 15:04
  posts
• 15:04
  are
• 15:04
  the
• 15:04
  strongest
• 15:05
  buy
• 15:05
  signal
• 15:05
  (April
• 15:06
  9,
• 15:06
  2025:
• 15:07
  S&P
• 15:08
  +9.52%),
• 15:09
  that
• 15:09
  the
• 15:09
  naive
• 15:09
  "TARIFF
• 15:10
  →
• 15:10
  SHORT"
• 15:10
  trade
• 15:10
  is
• 15:10
  70%
• 15:11
  wrong,
• 15:12
  and
• 15:12
  that
• 15:12
  Truth
• 15:12
  Social
• 15:12
  publishes
• 15:13
  6.2
• 15:13
  hours
• 15:14
  before
• 15:14
  X
• 15:14
  —
• 15:15
  creating
• 15:15
  a
• 15:15
  trading
• 15:16
  window.
• 15:16
  The
• 15:16
  system
• 15:17
  runs
• 15:17
  a
• 15:17
  daily
• 15:17
  pipeline
• 15:17
  with
• 15:18
  three
• 15:18
  "brains":
• 15:19
  Opus
• 15:19
  for
• 15:19
  deep
• 15:19
  causal
• 15:19
  analysis,
• 15:20
  an
• 15:20
  evolutionary
• 15:21
  engine
• 15:21
  that
• 15:21
  breeds
• 15:22
  new
• 15:22
  rules
• 15:22
  from
• 15:22
  survivors,
• 15:23
  and
• 15:23
  a
• 15:23
  circuit
• 15:24
  breaker
• 15:24
  that
• 15:24
  pauses
• 15:24
  when
• 15:25
  performance
• 15:25
  degrades
• 15:25
  .
• 15:26
  The
• 15:26
  statistical
• 15:27
  disclaimers
• 15:27
  are
• 15:27
  appropriate
• 15:28
  —
• 15:28
  31.5
• 15:29
  million
• 15:29
  combinations
• 15:30
  tested
• 15:30
  means
• 15:30
  surviving
• 15:31
  models
• 15:31
  likely
• 15:32
  include
• 15:32
  false
• 15:32
  positives
• 15:33
  from
• 15:33
  data
• 15:33
  snooping
• 15:33
  bias
• 15:34
  —
• 15:34
  but
• 15:35
  the
• 15:35
  architecture
• 15:35
  is
• 15:35
  genuinely
• 15:36
  interesting
• 15:36
  as
• 15:36
  an
• 15:37
  application
• 15:37
  of
• 15:37
  evolutionary
• 15:38
  AI
• 15:38
  to
• 15:38
  market
• 15:39
  signal
• 15:39
  extraction.
• 15:40
  New
• 15:40
  data
• 15:41
  from
• 15:41
  Anthropic
• 15:41
  maps
• 15:42
  Claude
• 15:42
  AI
• 15:42
  usage
• 15:42
  across
• 15:43
  116
• 15:43
  countries,
• 15:44
  indexed
• 15:45
  to
• 15:45
  working-age
• 15:45
  population.
• 15:46
  Israel
• 15:47
  leads
• 15:47
  at
• 15:47
  4.9x,
• 15:48
  followed
• 15:49
  by
• 15:49
  Singapore
• 15:49
  (4.19x)
• 15:50
  and
• 15:50
  the
• 15:51
  United
• 15:51
  States
• 15:51
  (3.69x
• 15:52
  among
• 15:52
  countries
• 15:53
  with
• 15:53
  10,000+
• 15:54
  conversations).
• 15:56
  The
• 15:56
  pattern
• 15:56
  favors
• 15:56
  small,
• 15:57
  tech-driven
• 15:58
  economies:
• 15:58
  because
• 15:59
  the
• 15:59
  index
• 15:59
  normalizes
• 16:00
  by
• 16:00
  workforce
• 16:00
  size,
• 16:01
  places
• 16:01
  like
• 16:01
  Malta
• 16:02
  (2.8x)
• 16:02
  and
• 16:02
  Georgia
• 16:03
  (2.17x)
• 16:04
  punch
• 16:04
  above
• 16:04
  their
• 16:04
  weight.
• 16:05
  Brazil
• 16:05
  ranks
• 16:06
  high
• 16:06
  in
• 16:06
  raw
• 16:06
  usage
• 16:07
  but
• 16:07
  drops
• 16:07
  to
• 16:07
  0.7x
• 16:08
  adjusted.
• 16:09
  Within
• 16:09
  the
• 16:09
  US,
• 16:10
  Washington
• 16:10
  D.C.
• 16:10
  leads
• 16:11
  at
• 16:11
  4.00x
• 16:11
  —
• 16:12
  unsurprising
• 16:13
  given
• 16:13
  the
• 16:13
  concentration
• 16:14
  of
• 16:14
  professionals
• 16:14
  —
• 16:15
  while
• 16:15
  West
• 16:15
  Virginia
• 16:15
  trails
• 16:16
  at
• 16:16
  0.25x.
• 16:17
  The
• 16:17
  most
• 16:18
  telling
• 16:18
  finding
• 16:18
  may
• 16:18
  be
• 16:18
  the
• 16:18
  usage
• 16:19
  pattern
• 16:19
  split:
• 16:20
  lower-income
• 16:20
  countries
• 16:21
  use
• 16:21
  Claude
• 16:21
  primarily
• 16:22
  for
• 16:22
  homework
• 16:22
  help
• 16:23
  and
• 16:23
  programming
• 16:23
  tasks,
• 16:24
  while
• 16:24
  wealthier
• 16:25
  countries
• 16:25
  show
• 16:25
  a
• 16:25
  broader
• 16:26
  professional
• 16:26
  mix
• 16:26
  —
• 16:27
  a
• 16:27
  pattern
• 16:27
  that
• 16:27
  likely
• 16:28
  reflects
• 16:28
  both
• 16:28
  age
• 16:29
  demographics
• 16:29
  and
• 16:30
  economic
• 16:30
  access
• 16:31
  rather
• 16:31
  than
• 16:31
  any
• 16:31
  intrinsic
• 16:32
  preference
• 16:32
  .