Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – July 15, 2026

Transcript

AGI Dreams Podcast – July 15, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Project
• 0:00
  Muteki
• 0:01
  is
• 0:01
  the
• 0:01
  kind
• 0:01
  of
• 0:01
  thing
• 0:01
  that
• 0:01
  keeps
• 0:02
  red-teamers
• 0:02
  up
• 0:02
  at
• 0:02
  night
• 0:03
  —
• 0:03
  not
• 0:03
  because
• 0:03
  it's
• 0:04
  novel
• 0:04
  in
• 0:04
  concept,
• 0:05
  but
• 0:05
  because
• 0:05
  it
• 0:05
  actually
• 0:05
  works
• 0:06
  at
• 0:06
  scale.
• 0:07
  This
• 0:07
  heterogeneous
• 0:07
  multi-model
• 0:08
  AI
• 0:08
  agent
• 0:08
  swarm
• 0:09
  coordinates
• 0:09
  Claude
• 0:09
  Code,
• 0:10
  Codex,
• 0:10
  and
• 0:10
  Cursor
• 0:11
  through
• 0:11
  a
• 0:11
  shared
• 0:11
  blackboard
• 0:11
  architecture
• 0:12
  with
• 0:12
  a
• 0:12
  4-phase
• 0:12
  solve
• 0:13
  pipeline:
• 0:13
  reconnaissance,
• 0:14
  exploitation,
• 0:15
  privilege
• 0:16
  escalation,
• 0:17
  and
• 0:17
  flag
• 0:17
  extraction.
• 0:18
  The
• 0:18
  numbers
• 0:18
  speak
• 0:19
  for
• 0:19
  themselves:
• 0:20
  200
• 0:20
  out
• 0:20
  of
• 0:20
  200
• 0:21
  on
• 0:21
  NYU
• 0:21
  CTF
• 0:22
  Bench,
• 0:22
  8th
• 0:23
  place
• 0:23
  at
• 0:23
  RIFFHACK
• 0:25
  2026
• 0:26
  with
• 0:27
  zero
• 0:27
  human
• 0:27
  intervention
• 0:28
  over
• 0:28
  three
• 0:28
  hours,
• 0:29
  and
• 0:29
  first
• 0:29
  blood
• 0:29
  on
• 0:30
  a
• 0:30
  penetration-testing
• 0:30
  range
• 0:31
  that
• 0:31
  had
• 0:31
  seen
• 0:31
  zero
• 0:31
  solves
• 0:32
  in
• 0:32
  three
• 0:32
  months.
• 0:33
  Total
• 0:33
  cost
• 0:33
  for
• 0:33
  the
• 0:33
  full
• 0:33
  benchmark
• 0:34
  run:
• 0:34
  approximately
• 0:34
  $214.
• 0:36
  What
• 0:36
  makes
• 0:36
  Muteki
• 0:37
  architecturally
• 0:37
  interesting
• 0:38
  is
• 0:38
  the
• 0:38
  heterogeneous
• 0:39
  engine
• 0:39
  design.
• 0:40
  Rather
• 0:40
  than
• 0:40
  running
• 0:41
  multiple
• 0:41
  instances
• 0:41
  of
• 0:42
  the
• 0:42
  same
• 0:42
  model,
• 0:43
  it
• 0:43
  dispatches
• 0:43
  different
• 0:43
  challenge
• 0:44
  types
• 0:44
  to
• 0:44
  different
• 0:44
  AI
• 0:45
  backends
• 0:45
  based
• 0:45
  on
• 0:45
  their
• 0:46
  strengths
• 0:46
  —
• 0:47
  Claude
• 0:47
  Code
• 0:47
  for
• 0:47
  complex
• 0:48
  reasoning
• 0:48
  chains,
• 0:49
  Codex
• 0:49
  for
• 0:49
  rapid
• 0:50
  code
• 0:50
  generation,
• 0:51
  Cursor
• 0:51
  for
• 0:51
  interactive
• 0:52
  debugging
• 0:52
  loops.
• 0:53
  The
• 0:53
  shared
• 0:53
  blackboard
• 0:54
  is
• 0:54
  the
• 0:54
  central
• 0:54
  coordination
• 0:55
  mechanism:
• 0:56
  facts
• 0:56
  discovered
• 0:57
  by
• 0:57
  one
• 0:57
  agent
• 0:57
  are
• 0:57
  immediately
• 0:58
  available
• 0:58
  to
• 0:58
  all
• 0:59
  others,
• 0:59
  dead
• 1:00
  ends
• 1:00
  are
• 1:00
  never
• 1:00
  retried
• 1:01
  by
• 1:01
  a
• 1:01
  different
• 1:01
  worker,
• 1:02
  and
• 1:02
  flags
• 1:02
  are
• 1:02
  accepted
• 1:03
  only
• 1:03
  when
• 1:03
  they
• 1:03
  appear
• 1:04
  verbatim
• 1:04
  in
• 1:04
  real
• 1:04
  execution
• 1:05
  output
• 1:05
  —
• 1:06
  a
• 1:06
  provenance
• 1:06
  gate
• 1:07
  that
• 1:07
  eliminates
• 1:07
  hallucinated
• 1:08
  solutions.
• 1:09
  A
• 1:09
  periodic
• 1:09
  review
• 1:10
  mechanism
• 1:10
  checks
• 1:10
  recorded
• 1:11
  facts
• 1:11
  and
• 1:11
  corrects
• 1:12
  course,
• 1:12
  preventing
• 1:13
  the
• 1:13
  dead-loop
• 1:13
  problem
• 1:13
  that
• 1:14
  plagues
• 1:14
  single-agent
• 1:14
  CTF
• 1:15
  solvers.
• 1:16
  The
• 1:16
  coordination
• 1:17
  loop
• 1:17
  runs
• 1:17
  at
• 1:17
  roughly
• 1:17
  one
• 1:18
  lap
• 1:18
  every
• 1:18
  two
• 1:18
  seconds,
• 1:19
  with
• 1:19
  each
• 1:19
  tick
• 1:19
  following
• 1:20
  a
• 1:20
  read-reason-intent-execute-write
• 1:21
  cycle.
• 1:22
  When
• 1:22
  one
• 1:23
  agent
• 1:23
  gets
• 1:23
  stuck,
• 1:24
  the
• 1:24
  swarm
• 1:24
  scheduler
• 1:24
  reassigns
• 1:25
  the
• 1:25
  problem
• 1:25
  to
• 1:25
  a
• 1:25
  different
• 1:26
  engine
• 1:26
  with
• 1:26
  fresh
• 1:26
  context.
• 1:28
  This
• 1:28
  is
• 1:28
  not
• 1:28
  a
• 1:28
  research
• 1:28
  demo;
• 1:29
  it
• 1:29
  ships
• 1:30
  as
• 1:30
  AGPL-3.0
• 1:32
  with
• 1:32
  full
• 1:33
  benchmark
• 1:33
  reproduction
• 1:34
  scripts
• 1:34
  and
• 1:34
  a
• 1:35
  web
• 1:35
  command
• 1:35
  deck
• 1:35
  for
• 1:35
  real-time
• 1:36
  monitoring.
• 1:37
  The
• 1:37
  broader
• 1:37
  tooling
• 1:37
  ecosystem
• 1:38
  around
• 1:38
  offensive
• 1:39
  AI
• 1:39
  continues
• 1:39
  to
• 1:39
  proliferate
• 1:40
  at
• 1:40
  a
• 1:40
  pace
• 1:41
  that
• 1:41
  should
• 1:41
  concern
• 1:41
  anyone
• 1:42
  responsible
• 1:42
  for
• 1:42
  defending
• 1:43
  networks.
• 1:44
  A
• 1:44
  repository
• 1:44
  called
• 1:45
  recon-skills
• 1:45
  catalogs
• 1:46
  169
• 1:48
  distinct
• 1:48
  offensive
• 1:49
  security
• 1:49
  skills
• 1:49
  spanning
• 1:50
  browser
• 1:50
  fingerprint
• 1:51
  evasion,
• 1:51
  SCADA
• 1:52
  enumeration,
• 1:53
  hardcoded
• 1:53
  credential
• 1:53
  hunting,
• 1:54
  and
• 1:54
  cloud
• 1:54
  metadata
• 1:55
  harvesting
• 1:55
  —
• 1:56
  tested
• 1:56
  across
• 1:57
  600+
• 1:57
  domains
• 1:58
  and
• 1:58
  80+
• 1:59
  vulnerable
• 1:59
  companies
• 2:00
  in
• 2:00
  45+
• 2:01
  sectors.
• 2:02
  This
• 2:02
  is
• 2:02
  essentially
• 2:03
  a
• 2:03
  machine-readable
• 2:04
  playbook
• 2:04
  for
• 2:04
  penetration
• 2:05
  testing
• 2:05
  automation,
• 2:06
  ready
• 2:07
  to
• 2:07
  be
• 2:07
  consumed
• 2:07
  by
• 2:08
  any
• 2:08
  agentic
• 2:08
  framework.
• 2:09
  Meanwhile,
• 2:10
  the
• 2:10
  local
• 2:11
  model
• 2:11
  community
• 2:11
  is
• 2:11
  already
• 2:12
  asking
• 2:12
  whether
• 2:13
  specialized
• 2:13
  cybersecurity
• 2:14
  models
• 2:15
  can
• 2:15
  run
• 2:15
  on
• 2:15
  consumer
• 2:16
  hardware,
• 2:17
  though
• 2:17
  the
• 2:17
  conversation
• 2:17
  remains
• 2:18
  speculative
• 2:18
  rather
• 2:19
  than
• 2:19
  practical
• 2:19
  at
• 2:19
  this
• 2:20
  stage.
• 2:22
  The
• 2:22
  coding
• 2:23
  agent
• 2:23
  landscape
• 2:24
  is
• 2:24
  shifting
• 2:24
  from
• 2:24
  "can
• 2:25
  AI
• 2:25
  write
• 2:25
  code"
• 2:26
  to
• 2:26
  "how
• 2:26
  do
• 2:26
  humans
• 2:27
  actually
• 2:27
  work
• 2:27
  alongside
• 2:28
  AI
• 2:28
  agents."
• 2:29
  Juggler,
• 2:29
  from
• 2:30
  JUCE
• 2:30
  creator
• 2:30
  Julian
• 2:31
  Storer,
• 2:31
  makes
• 2:31
  this
• 2:31
  explicit:
• 2:32
  it
• 2:33
  is
• 2:33
  an
• 2:33
  open-source
• 2:33
  GUI
• 2:33
  coding
• 2:34
  agent
• 2:34
  built
• 2:34
  in
• 2:34
  Go/Wails
• 2:35
  with
• 2:35
  a
• 2:35
  Miller-column
• 2:36
  tree
• 2:36
  UI
• 2:37
  and
• 2:37
  Yjs-based
• 2:37
  document
• 2:38
  model.
• 2:39
  The
• 2:39
  design
• 2:39
  philosophy
• 2:40
  is
• 2:40
  that
• 2:40
  coding
• 2:41
  agents
• 2:41
  need
• 2:41
  a
• 2:41
  proper
• 2:42
  visual
• 2:42
  interface
• 2:42
  —
• 2:44
  one
• 2:44
  that
• 2:44
  shows
• 2:44
  the
• 2:44
  project
• 2:45
  tree,
• 2:45
  the
• 2:45
  agent's
• 2:46
  conversation
• 2:46
  stream,
• 2:46
  and
• 2:46
  the
• 2:47
  edited
• 2:47
  files
• 2:47
  in
• 2:47
  a
• 2:47
  single
• 2:48
  coherent
• 2:48
  layout
• 2:48
  —
• 2:49
  not
• 2:49
  just
• 2:49
  a
• 2:49
  terminal
• 2:50
  scroll.
• 2:50
  The
• 2:51
  plugin
• 2:51
  architecture
• 2:51
  allows
• 2:52
  extending
• 2:52
  the
• 2:52
  agent's
• 2:52
  capabilities,
• 2:53
  and
• 2:54
  the
• 2:54
  Yjs
• 2:54
  foundation
• 2:55
  enables
• 2:55
  real-time
• 2:55
  collaborative
• 2:56
  editing
• 2:56
  between
• 2:56
  human
• 2:56
  and
• 2:57
  agent,
• 2:57
  with
• 2:58
  CRDT-based
• 2:58
  conflict
• 2:59
  resolution
• 2:59
  when
• 2:59
  both
• 3:00
  parties
• 3:00
  modify
• 3:00
  the
• 3:00
  same
• 3:01
  file
• 3:01
  simultaneously.
• 3:02
  The
• 3:02
  AGPL-3.0
• 3:05
  license
• 3:06
  follows
• 3:06
  a
• 3:06
  pattern
• 3:06
  we
• 3:06
  are
• 3:07
  seeing
• 3:07
  more
• 3:07
  frequently
• 3:07
  in
• 3:07
  agent
• 3:08
  tooling:
• 3:09
  open
• 3:09
  source
• 3:09
  with
• 3:09
  a
• 3:09
  copyleft
• 3:10
  boundary
• 3:10
  that
• 3:11
  prevents
• 3:11
  proprietary
• 3:12
  resale
• 3:12
  without
• 3:12
  contribution
• 3:13
  back.
• 3:14
  On
• 3:14
  the
• 3:14
  quality
• 3:15
  control
• 3:15
  front,
• 3:15
  Taste
• 3:16
  Skill
• 3:16
  tackles
• 3:16
  the
• 3:16
  "AI
• 3:17
  slop"
• 3:17
  problem
• 3:17
  head-on.
• 3:18
  It
• 3:19
  is
• 3:19
  a
• 3:19
  portable
• 3:19
  agent
• 3:19
  skill
• 3:20
  framework
• 3:20
  with
• 3:20
  adjustable
• 3:21
  VARIANCE,
• 3:21
  MOTION,
• 3:22
  and
• 3:22
  DENSITY
• 3:23
  dials
• 3:23
  that
• 3:23
  constrain
• 3:24
  AI-generated
• 3:24
  frontend
• 3:25
  code
• 3:25
  toward
• 3:25
  intentional
• 3:26
  design
• 3:26
  rather
• 3:26
  than
• 3:26
  boilerplate
• 3:27
  output.
• 3:28
  The
• 3:28
  v2
• 3:28
  rewrite
• 3:29
  introduces
• 3:29
  brief
• 3:29
  inference,
• 3:30
  design-system
• 3:31
  mapping,
• 3:31
  an
• 3:31
  em-dash
• 3:32
  ban,
• 3:32
  canonical
• 3:33
  GSAP
• 3:33
  code
• 3:34
  skeletons,
• 3:34
  and
• 3:35
  a
• 3:35
  strict
• 3:35
  pre-flight
• 3:35
  check
• 3:35
  —
• 3:36
  essentially
• 3:37
  codifying
• 3:37
  design
• 3:38
  taste
• 3:38
  into
• 3:38
  machine-readable
• 3:39
  rules.
• 3:40
  Vercel
• 3:40
  sponsors
• 3:41
  the
• 3:41
  project,
• 3:41
  and
• 3:42
  the
• 3:42
  skill
• 3:42
  works
• 3:42
  across
• 3:42
  Codex,
• 3:43
  Cursor,
• 3:43
  and
• 3:43
  Claude
• 3:44
  Code.
• 3:45
  Two
• 3:45
  Claude
• 3:45
  Code
• 3:45
  plugins
• 3:46
  address
• 3:46
  different
• 3:46
  knowledge
• 3:47
  gaps.
• 3:48
  Engram
• 3:48
  implements
• 3:49
  spaced
• 3:49
  repetition
• 3:50
  learning
• 3:50
  using
• 3:51
  the
• 3:51
  FSRS-4.5
• 3:54
  scheduling
• 3:54
  algorithm
• 3:55
  with
• 3:55
  a
• 3:55
  separation-of-powers
• 3:56
  architecture:
• 3:57
  the
• 3:57
  tutor
• 3:57
  teaches
• 3:58
  (forcing
• 3:58
  the
• 3:58
  user
• 3:59
  to
• 3:59
  predict
• 3:59
  and
• 3:59
  explain
• 4:00
  back,
• 4:00
  not
• 4:01
  just
• 4:01
  passively
• 4:01
  read),
• 4:02
  the
• 4:02
  assessor
• 4:03
  grades
• 4:03
  blindly
• 4:03
  (0
• 4:04
  out
• 4:04
  of
• 4:04
  198
• 4:05
  grades
• 4:05
  were
• 4:05
  inflated
• 4:06
  in
• 4:06
  testing
• 4:06
  —
• 4:07
  verified
• 4:08
  against
• 4:08
  a
• 4:08
  gold
• 4:09
  standard
• 4:09
  dataset),
• 4:10
  and
• 4:10
  the
• 4:10
  coach
• 4:10
  plans
• 4:11
  review
• 4:11
  schedules
• 4:11
  timed
• 4:12
  to
• 4:12
  the
• 4:12
  forgetting
• 4:12
  curve.
• 4:13
  The
• 4:13
  engine
• 4:13
  runs
• 4:14
  on
• 4:14
  stdlib
• 4:14
  Python
• 4:15
  with
• 4:15
  zero
• 4:15
  network
• 4:16
  code
• 4:16
  (enforced
• 4:16
  by
• 4:16
  a
• 4:17
  permanent
• 4:17
  self-test),
• 4:18
  and
• 4:18
  all
• 4:18
  state
• 4:19
  lives
• 4:19
  as
• 4:19
  plain
• 4:19
  JSON
• 4:20
  files
• 4:20
  on
• 4:20
  your
• 4:20
  machine.
• 4:21
  The
• 4:21
  214
• 4:23
  self-tests
• 4:23
  and
• 4:23
  commons
• 4:24
  data
• 4:24
  contribution
• 4:25
  model
• 4:25
  suggest
• 4:25
  serious
• 4:26
  engineering,
• 4:27
  not
• 4:27
  a
• 4:27
  weekend
• 4:27
  hack.
• 4:28
  RuvNet
• 4:28
  Brain
• 4:29
  takes
• 4:29
  a
• 4:29
  different
• 4:29
  approach,
• 4:30
  bundling
• 4:30
  512MB
• 4:32
  of
• 4:32
  knowledge
• 4:33
  across
• 4:33
  36
• 4:33
  repositories
• 4:34
  with
• 4:34
  dual
• 4:34
  embeddings
• 4:35
  (MiniLM-384
• 4:36
  and
• 4:36
  BGE-768)
• 4:38
  plus
• 4:38
  cross-encoder
• 4:39
  reranking.
• 4:40
  A
• 4:40
  UserPromptSubmit
• 4:41
  grounding
• 4:41
  hook
• 4:42
  intercepts
• 4:42
  every
• 4:42
  prompt
• 4:43
  and
• 4:43
  augments
• 4:43
  it
• 4:43
  with
• 4:43
  relevant
• 4:44
  context
• 4:44
  before
• 4:45
  execution.
• 4:46
  Version
• 4:46
  2.5
• 4:47
  ships
• 4:47
  a
• 4:47
  substitution
• 4:48
  CI
• 4:48
  gate,
• 4:49
  a
• 4:49
  job-proof
• 4:49
  registry,
• 4:50
  and
• 4:50
  a
• 4:50
  subagent
• 4:51
  model
• 4:51
  gate
• 4:51
  —
• 4:52
  features
• 4:52
  that
• 4:53
  reflect
• 4:53
  a
• 4:53
  maturing
• 4:53
  understanding
• 4:54
  of
• 4:54
  how
• 4:54
  plugins
• 4:55
  fail
• 4:55
  in
• 4:55
  production.
• 4:56
  The
• 4:56
  mundane
• 4:57
  reality
• 4:57
  of
• 4:57
  multi-agent
• 4:58
  development
• 4:58
  surfaces
• 4:59
  in
• 4:59
  a
• 4:59
  report
• 4:59
  about
• 5:00
  port
• 5:00
  collisions
• 5:00
  when
• 5:00
  running
• 5:01
  two
• 5:01
  Claude
• 5:01
  Code
• 5:01
  agents
• 5:02
  in
• 5:02
  parallel
• 5:02
  worktrees.
• 5:04
  Both
• 5:04
  agents
• 5:05
  tried
• 5:05
  to
• 5:05
  bind
• 5:05
  the
• 5:05
  same
• 5:06
  port
• 5:06
  for
• 5:06
  their
• 5:06
  dev
• 5:06
  servers,
• 5:07
  crashing
• 5:07
  each
• 5:07
  other.
• 5:09
  The
• 5:09
  solution
• 5:09
  —
• 5:10
  a
• 5:10
  deterministic
• 5:11
  CLI
• 5:11
  tool
• 5:12
  called
• 5:12
  Winter
• 5:12
  that
• 5:12
  assigns
• 5:13
  reserved
• 5:13
  port
• 5:13
  blocks
• 5:14
  per
• 5:14
  feature
• 5:14
  environment
• 5:14
  —
• 5:15
  is
• 5:15
  exactly
• 5:16
  the
• 5:16
  kind
• 5:16
  of
• 5:16
  infrastructure
• 5:17
  problem
• 5:17
  that
• 5:17
  emerges
• 5:18
  when
• 5:18
  "just
• 5:19
  spin
• 5:19
  up
• 5:19
  another
• 5:19
  agent"
• 5:20
  meets
• 5:20
  real
• 5:20
  development
• 5:21
  workflows.
• 5:24
  Voicebox
• 5:24
  consolidates
• 5:25
  seven
• 5:25
  TTS
• 5:26
  engines
• 5:26
  —
• 5:27
  Qwen3-TTS,
• 5:28
  Chatterbox,
• 5:29
  Kokoro,
• 5:29
  TADA,
• 5:30
  and
• 5:30
  others
• 5:30
  —
• 5:31
  into
• 5:31
  a
• 5:31
  single
• 5:32
  open-source
• 5:32
  desktop
• 5:32
  application
• 5:33
  supporting
• 5:33
  23
• 5:34
  languages,
• 5:35
  voice
• 5:35
  cloning,
• 5:36
  an
• 5:36
  MCP
• 5:36
  server
• 5:37
  for
• 5:37
  programmatic
• 5:37
  access,
• 5:38
  and
• 5:38
  system-wide
• 5:39
  global
• 5:39
  dictation.
• 5:40
  Built
• 5:41
  in
• 5:41
  Tauri/Rust
• 5:42
  rather
• 5:42
  than
• 5:42
  Electron
• 5:43
  (a
• 5:43
  meaningful
• 5:43
  performance
• 5:44
  difference
• 5:44
  when
• 5:44
  running
• 5:44
  inference
• 5:45
  locally),
• 5:45
  it
• 5:46
  represents
• 5:46
  the
• 5:46
  strongest
• 5:47
  attempt
• 5:47
  yet
• 5:47
  to
• 5:47
  make
• 5:48
  voice
• 5:48
  synthesis
• 5:48
  a
• 5:48
  first-class
• 5:49
  local
• 5:49
  development
• 5:50
  tool
• 5:50
  rather
• 5:50
  than
• 5:50
  an
• 5:51
  API
• 5:51
  dependency.
• 5:52
  On
• 5:52
  the
• 5:52
  integration
• 5:53
  side,
• 5:53
  Cicero
• 5:54
  offers
• 5:54
  bidirectional
• 5:55
  voice
• 5:55
  for
• 5:55
  any
• 5:56
  AI
• 5:56
  agent
• 5:56
  harness
• 5:56
  —
• 5:57
  Claude
• 5:57
  Code,
• 5:58
  Codex,
• 5:58
  Gemini
• 5:59
  CLI
• 5:59
  —
• 6:00
  via
• 6:00
  ACP
• 6:01
  (Agent
• 6:01
  Communication
• 6:02
  Protocol).
• 6:03
  The
• 6:03
  practical
• 6:03
  insight
• 6:03
  from
• 6:04
  the
• 6:04
  community:
• 6:05
  barge-in
• 6:05
  detection
• 6:05
  requires
• 6:06
  a
• 6:06
  proper
• 6:06
  speech
• 6:07
  model,
• 6:07
  not
• 6:08
  simple
• 6:08
  energy-based
• 6:08
  VAD,
• 6:09
  because
• 6:10
  keyboard
• 6:10
  noise
• 6:10
  and
• 6:10
  ambient
• 6:11
  sound
• 6:11
  trigger
• 6:11
  false
• 6:12
  interruptions.
• 6:13
  Local
• 6:13
  model
• 6:13
  benchmarking
• 6:14
  continues
• 6:14
  to
• 6:14
  produce
• 6:15
  results
• 6:15
  that
• 6:15
  challenge
• 6:15
  the
• 6:16
  assumption
• 6:16
  that
• 6:16
  bigger
• 6:16
  always
• 6:17
  means
• 6:17
  better
• 6:17
  as
• 6:17
  an
• 6:17
  agent.
• 6:18
  Qwen3.6-27B
• 6:21
  at
• 6:21
  INT8
• 6:22
  quantization
• 6:22
  outperformed
• 6:23
  the
• 6:23
  specifically
• 6:23
  agentic-tuned
• 6:24
  Nemotron
• 6:24
  Puzzle-75B-A9B,
• 6:27
  requiring
• 6:27
  only
• 6:27
  6-9
• 6:28
  tool
• 6:28
  calls
• 6:28
  per
• 6:28
  task
• 6:29
  versus
• 6:29
  the
• 6:29
  75B
• 6:30
  model's
• 6:30
  13-23
• 6:31
  —
• 6:32
  suggesting
• 6:32
  that
• 6:32
  the
• 6:32
  smaller
• 6:33
  model
• 6:33
  was
• 6:33
  reasoning
• 6:33
  more
• 6:34
  efficiently
• 6:34
  rather
• 6:34
  than
• 6:34
  compensating
• 6:35
  with
• 6:35
  more
• 6:35
  attempts.
• 6:36
  The
• 6:36
  tester
• 6:37
  also
• 6:37
  discovered
• 6:38
  a
• 6:38
  prefix
• 6:38
  caching
• 6:38
  bug
• 6:39
  with
• 6:39
  fp8
• 6:39
  KV
• 6:40
  cache
• 6:40
  that
• 6:40
  caused
• 6:41
  incorrect
• 6:41
  responses
• 6:41
  —
• 6:42
  a
• 6:42
  reminder
• 6:43
  that
• 6:43
  quantization
• 6:44
  interacts
• 6:44
  with
• 6:44
  serving
• 6:45
  infrastructure
• 6:45
  in
• 6:46
  non-obvious
• 6:46
  ways.
• 6:47
  Separately,
• 6:48
  Qwen3.6's
• 6:50
  35B-A3B
• 6:52
  MoE
• 6:52
  variant
• 6:53
  ran
• 6:53
  a
• 6:53
  single-prompt
• 6:53
  flight
• 6:54
  simulator,
• 6:54
  with
• 6:55
  Q8
• 6:55
  0
• 6:55
  on
• 6:55
  CPU
• 6:56
  outperforming
• 6:56
  Q4
• 6:57
  K
• 6:57
  M
• 6:57
  on
• 6:57
  GPU
• 6:58
  —
• 6:58
  a
• 6:58
  result
• 6:59
  that
• 6:59
  underscores
• 6:59
  how
• 6:59
  quantization
• 7:00
  quality
• 7:00
  can
• 7:01
  matter
• 7:01
  more
• 7:01
  than
• 7:01
  hardware
• 7:02
  acceleration
• 7:02
  for
• 7:02
  MoE
• 7:02
  architectures
• 7:03
  with
• 7:03
  small
• 7:03
  active
• 7:04
  parameter
• 7:04
  counts.
• 7:06
  Goedel-Architect
• 7:07
  achieved
• 7:08
  a
• 7:08
  result
• 7:08
  that
• 7:08
  formal
• 7:08
  methods
• 7:09
  researchers
• 7:09
  have
• 7:09
  been
• 7:09
  chasing
• 7:10
  for
• 7:10
  years:
• 7:10
  100%
• 7:11
  on
• 7:11
  the
• 7:11
  MiniF2F-test
• 7:12
  benchmark,
• 7:13
  the
• 7:13
  first
• 7:13
  system
• 7:14
  ever
• 7:14
  to
• 7:14
  do
• 7:14
  so.
• 7:15
  It
• 7:15
  also
• 7:15
  scored
• 7:16
  57.5%
• 7:17
  on
• 7:17
  PutnamBench
• 7:18
  and
• 7:18
  solved
• 7:18
  4
• 7:18
  of
• 7:18
  6
• 7:19
  problems
• 7:19
  from
• 7:19
  IMO
• 7:20
  2025.
• 7:21
  The
• 7:21
  architecture
• 7:21
  uses
• 7:22
  blueprint
• 7:22
  dependency
• 7:23
  graphs
• 7:23
  in
• 7:23
  Lean
• 7:23
  4
• 7:24
  with
• 7:24
  DeepSeek-V4-Flash
• 7:25
  as
• 7:25
  the
• 7:25
  backbone,
• 7:26
  and
• 7:26
  the
• 7:26
  authors
• 7:27
  claim
• 7:27
  approximately
• 7:28
  100x
• 7:28
  cost
• 7:29
  reduction
• 7:29
  compared
• 7:29
  to
• 7:29
  comparable
• 7:30
  pipelines.
• 7:31
  The
• 7:31
  key
• 7:31
  insight
• 7:32
  is
• 7:32
  not
• 7:32
  the
• 7:32
  language
• 7:32
  model
• 7:33
  itself
• 7:33
  but
• 7:33
  the
• 7:33
  structured
• 7:33
  decomposition:
• 7:35
  breaking
• 7:35
  proofs
• 7:35
  into
• 7:36
  blueprints
• 7:36
  with
• 7:36
  explicit
• 7:37
  dependency
• 7:37
  relationships,
• 7:38
  then
• 7:38
  refining
• 7:39
  each
• 7:39
  component
• 7:40
  through
• 7:40
  targeted
• 7:40
  generation
• 7:41
  and
• 7:41
  verification
• 7:41
  against
• 7:42
  the
• 7:42
  Lean
• 7:42
  type
• 7:42
  checker.
• 7:43
  Where
• 7:43
  brute-force
• 7:44
  approaches
• 7:44
  throw
• 7:45
  entire
• 7:45
  proofs
• 7:45
  at
• 7:45
  a
• 7:45
  model
• 7:46
  and
• 7:46
  hope
• 7:46
  for
• 7:46
  correctness,
• 7:47
  Goedel-Architect
• 7:48
  localizes
• 7:48
  each
• 7:49
  subgoal
• 7:49
  and
• 7:49
  verifies
• 7:49
  it
• 7:50
  independently
• 7:50
  —
• 7:51
  a
• 7:51
  design
• 7:51
  that
• 7:51
  scales
• 7:52
  sublinearly
• 7:52
  with
• 7:52
  proof
• 7:53
  complexity.
• 7:54
  A
• 7:54
  paper
• 7:54
  on
• 7:54
  Looped
• 7:55
  SSMs
• 7:55
  presents
• 7:55
  a
• 7:55
  counterintuitive
• 7:56
  finding:
• 7:57
  state
• 7:57
  space
• 7:58
  models
• 7:58
  with
• 7:58
  shared
• 7:58
  parameters
• 7:59
  across
• 7:59
  depth
• 7:59
  (looped
• 8:00
  architecture)
• 8:00
  match
• 8:00
  or
• 8:01
  outperform
• 8:01
  models
• 8:02
  with
• 8:02
  independent
• 8:02
  parameters
• 8:03
  at
• 8:03
  each
• 8:03
  layer,
• 8:04
  despite
• 8:04
  having
• 8:04
  a
• 8:04
  strictly
• 8:05
  smaller
• 8:05
  hypothesis
• 8:05
  space.
• 8:06
  Tested
• 8:06
  across
• 8:07
  LRU,
• 8:07
  S5,
• 8:08
  LinOSS,
• 8:09
  and
• 8:09
  LrcSSM
• 8:09
  architectures
• 8:10
  on
• 8:10
  six
• 8:10
  benchmarks,
• 8:11
  the
• 8:11
  results
• 8:12
  suggest
• 8:12
  that
• 8:12
  parameter
• 8:13
  sharing
• 8:13
  acts
• 8:13
  as
• 8:14
  a
• 8:14
  beneficial
• 8:14
  inductive
• 8:14
  bias
• 8:15
  rather
• 8:15
  than
• 8:15
  a
• 8:15
  limitation.
• 8:16
  Input
• 8:17
  reshaping
• 8:17
  techniques
• 8:17
  —
• 8:18
  concatenation
• 8:18
  and
• 8:18
  rechunking
• 8:19
  —
• 8:19
  yield
• 8:20
  an
• 8:20
  additional
• 8:20
  1-6%
• 8:21
  improvement.
• 8:22
  This
• 8:22
  matters
• 8:22
  because
• 8:23
  it
• 8:23
  implies
• 8:23
  that
• 8:23
  the
• 8:23
  SSM
• 8:24
  scaling
• 8:24
  narrative
• 8:25
  ("more
• 8:25
  parameters,
• 8:26
  more
• 8:26
  layers")
• 8:27
  may
• 8:27
  be
• 8:27
  leaving
• 8:28
  efficiency
• 8:28
  on
• 8:28
  the
• 8:28
  table.
• 8:29
  From
• 8:29
  Meta
• 8:29
  AI
• 8:30
  and
• 8:30
  ENS
• 8:30
  Paris
• 8:31
  comes
• 8:31
  a
• 8:31
  careful
• 8:31
  investigation
• 8:32
  into
• 8:33
  whether
• 8:33
  the
• 8:33
  brain
• 8:33
  implements
• 8:34
  anything
• 8:34
  resembling
• 8:34
  backpropagation.
• 8:36
  Using
• 8:37
  both
• 8:37
  fMRI
• 8:37
  (spatial
• 8:38
  resolution)
• 8:38
  and
• 8:39
  MEG
• 8:39
  (temporal
• 8:40
  resolution),
• 8:40
  the
• 8:41
  researchers
• 8:41
  found
• 8:41
  that
• 8:41
  backpropagated
• 8:42
  gradients
• 8:43
  from
• 8:43
  deep
• 8:43
  neural
• 8:43
  networks
• 8:44
  do
• 8:44
  predict
• 8:44
  brain
• 8:44
  activity,
• 8:45
  and
• 8:45
  that
• 8:46
  these
• 8:46
  residual
• 8:46
  backprop
• 8:47
  signals
• 8:47
  explain
• 8:48
  variance
• 8:48
  beyond
• 8:48
  what
• 8:49
  forward
• 8:49
  activations
• 8:49
  alone
• 8:50
  capture
• 8:50
  —
• 8:51
  particularly
• 8:51
  in
• 8:51
  higher-level
• 8:52
  visual
• 8:52
  cortex,
• 8:53
  where
• 8:53
  abstract
• 8:53
  representations
• 8:54
  are
• 8:54
  formed.
• 8:55
  However,
• 8:56
  the
• 8:56
  temporal
• 8:56
  and
• 8:57
  spatial
• 8:57
  organization
• 8:58
  of
• 8:58
  these
• 8:58
  gradient-like
• 8:58
  signals
• 8:59
  diverges
• 9:00
  from
• 9:00
  what
• 9:00
  biological
• 9:01
  backpropagation
• 9:01
  would
• 9:02
  predict:
• 9:03
  the
• 9:03
  timing
• 9:03
  is
• 9:03
  wrong,
• 9:04
  and
• 9:04
  the
• 9:04
  anatomical
• 9:05
  distribution
• 9:05
  does
• 9:05
  not
• 9:06
  follow
• 9:06
  the
• 9:06
  layer-by-layer
• 9:07
  backward
• 9:07
  pass
• 9:07
  that
• 9:07
  backprop
• 9:08
  requires.
• 9:09
  The
• 9:09
  conclusion:
• 9:10
  the
• 9:10
  brain
• 9:10
  appears
• 9:11
  to
• 9:11
  use
• 9:11
  a
• 9:11
  different
• 9:11
  credit-assignment
• 9:12
  mechanism
• 9:13
  that
• 9:13
  produces
• 9:13
  computationally
• 9:14
  similar
• 9:15
  outcomes
• 9:15
  without
• 9:15
  the
• 9:15
  same
• 9:16
  algorithmic
• 9:16
  implementation.
• 9:18
  This
• 9:18
  distinction
• 9:18
  matters
• 9:19
  for
• 9:19
  both
• 9:19
  neuroscience
• 9:20
  and
• 9:20
  AI
• 9:20
  architecture
• 9:21
  research
• 9:21
  —
• 9:22
  if
• 9:22
  the
• 9:22
  brain
• 9:22
  achieves
• 9:22
  similar
• 9:23
  results
• 9:23
  through
• 9:23
  a
• 9:23
  fundamentally
• 9:24
  different
• 9:24
  learning
• 9:25
  rule,
• 9:25
  there
• 9:25
  may
• 9:25
  be
• 9:25
  computational
• 9:26
  advantages
• 9:27
  that
• 9:27
  current
• 9:27
  deep
• 9:27
  learning
• 9:28
  is
• 9:28
  missing
• 9:28
  entirely.
• 9:29
  LeMario
• 9:30
  trained
• 9:30
  a
• 9:30
  JEPA
• 9:31
  (Joint
• 9:31
  Embedding
• 9:31
  Predictive
• 9:32
  Architecture)
• 9:32
  world
• 9:32
  model
• 9:33
  on
• 9:33
  Super
• 9:33
  Mario
• 9:33
  Bros
• 9:34
  and
• 9:34
  found
• 9:34
  that
• 9:35
  learned
• 9:35
  short-horizon
• 9:35
  dynamics
• 9:36
  did
• 9:36
  not
• 9:36
  translate
• 9:37
  into
• 9:37
  effective
• 9:37
  long-range
• 9:38
  planning
• 9:38
  —
• 9:39
  even
• 9:39
  when
• 9:39
  next-frame
• 9:40
  prediction
• 9:40
  was
• 9:40
  accurate.
• 9:41
  The
• 9:41
  core
• 9:42
  problem:
• 9:42
  predictive
• 9:43
  state
• 9:43
  is
• 9:43
  not
• 9:43
  the
• 9:43
  same
• 9:44
  as
• 9:44
  control
• 9:44
  state.
• 9:45
  A
• 9:45
  model
• 9:46
  that
• 9:46
  accurately
• 9:46
  predicts
• 9:47
  what
• 9:47
  will
• 9:47
  happen
• 9:47
  next
• 9:47
  can
• 9:48
  still
• 9:48
  fail
• 9:48
  to
• 9:48
  identify
• 9:49
  which
• 9:49
  actions
• 9:50
  lead
• 9:50
  to
• 9:50
  desirable
• 9:50
  outcomes.
• 9:52
  CEM
• 9:52
  (Cross-Entropy
• 9:53
  Method)
• 9:53
  search
• 9:53
  exposed
• 9:54
  the
• 9:54
  model's
• 9:54
  weaknesses,
• 9:55
  and
• 9:55
  scrolling
• 9:56
  camera
• 9:56
  movement
• 9:56
  broke
• 9:57
  embedding
• 9:57
  similarity
• 9:58
  assumptions
• 9:58
  —
• 9:59
  a
• 9:59
  useful
• 10:00
  corrective
• 10:00
  to
• 10:00
  the
• 10:00
  "world
• 10:01
  models
• 10:01
  will
• 10:01
  solve
• 10:01
  planning"
• 10:02
  narrative.
• 10:04
  Ray
• 10:04
  Myers
• 10:05
  published
• 10:05
  a
• 10:05
  detailed
• 10:06
  critique
• 10:06
  of
• 10:06
  Anthropic's
• 10:07
  narrative
• 10:07
  around
• 10:08
  the
• 10:08
  Bun
• 10:08
  team's
• 10:08
  Zig-to-Rust
• 10:09
  rewrite,
• 10:10
  arguing
• 10:10
  that
• 10:10
  the
• 10:10
  story
• 10:11
  was
• 10:11
  a
• 10:11
  marketing
• 10:11
  opportunity
• 10:12
  rather
• 10:12
  than
• 10:12
  a
• 10:13
  genuine
• 10:13
  technical
• 10:13
  case
• 10:14
  study.
• 10:14
  Myers
• 10:15
  identifies
• 10:15
  a
• 10:15
  third
• 10:16
  interpretation
• 10:16
  beyond
• 10:17
  the
• 10:17
  official
• 10:17
  "Zig
• 10:18
  wasn't
• 10:18
  up
• 10:18
  to
• 10:18
  the
• 10:19
  task"
• 10:19
  and
• 10:19
  the
• 10:20
  community's
• 10:20
  "Bun's
• 10:21
  engineering
• 10:21
  decisions
• 10:22
  were
• 10:22
  the
• 10:22
  problem":
• 10:23
  faced
• 10:23
  with
• 10:23
  legitimate
• 10:24
  memory
• 10:24
  bugs,
• 10:24
  management
• 10:25
  eagerly
• 10:25
  approved
• 10:26
  the
• 10:26
  Rust
• 10:26
  rewrite
• 10:27
  because
• 10:27
  it
• 10:27
  showcased
• 10:27
  their
• 10:28
  Fable
• 10:28
  model,
• 10:29
  Anthropic
• 10:29
  already
• 10:29
  had
• 10:30
  Rust
• 10:30
  expertise,
• 10:31
  and
• 10:31
  Zig
• 10:31
  is
• 10:31
  openly
• 10:32
  against
• 10:32
  using
• 10:32
  Anthropic's
• 10:33
  products.
• 10:34
  The
• 10:34
  two-month
• 10:34
  delay
• 10:35
  between
• 10:35
  merging
• 10:35
  the
• 10:35
  migration
• 10:36
  and
• 10:36
  publishing
• 10:36
  the
• 10:37
  rationale
• 10:37
  conveniently
• 10:37
  allowed
• 10:38
  headlines
• 10:38
  like
• 10:38
  "Anthropic's
• 10:39
  Bun
• 10:39
  Rust
• 10:40
  rewrite
• 10:40
  merged
• 10:40
  at
• 10:40
  speed
• 10:41
  of
• 10:41
  AI"
• 10:41
  to
• 10:41
  carry
• 10:41
  the
• 10:41
  story
• 10:42
  first.
• 10:42
  Myers
• 10:43
  points
• 10:43
  to
• 10:43
  TigerBeetle
• 10:43
  —
• 10:44
  a
• 10:44
  production
• 10:45
  database
• 10:45
  written
• 10:45
  in
• 10:45
  Zig
• 10:46
  with
• 10:46
  strong
• 10:46
  memory
• 10:46
  safety
• 10:47
  properties
• 10:47
  —
• 10:47
  as
• 10:48
  a
• 10:48
  counterexample
• 10:48
  that
• 10:48
  undermines
• 10:49
  the
• 10:49
  "Zig
• 10:49
  is
• 10:49
  inherently
• 10:50
  unsafe"
• 10:50
  narrative.
• 10:51
  He
• 10:51
  also
• 10:51
  notes
• 10:52
  that
• 10:52
  the
• 10:52
  AI-driven
• 10:52
  rewrite
• 10:53
  was
• 10:53
  "not
• 10:53
  powerful
• 10:53
  enough
• 10:53
  to
• 10:54
  catch
• 10:54
  a
• 10:54
  use-after-free"
• 10:55
  while
• 10:55
  being
• 10:55
  marketed
• 10:56
  as
• 10:56
  proof
• 10:56
  of
• 10:56
  AI
• 10:56
  superiority.
• 10:58
  When
• 10:58
  an
• 10:58
  AI
• 10:58
  company
• 10:58
  approaching
• 10:59
  a
• 10:59
  trillion-dollar
• 10:59
  IPO
• 11:00
  valuation
• 11:00
  needs
• 11:01
  you
• 11:01
  to
• 11:01
  believe
• 11:01
  coding
• 11:01
  is
• 11:01
  going
• 11:01
  away,
• 11:02
  every
• 11:03
  technical
• 11:03
  blog
• 11:03
  post
• 11:03
  is
• 11:04
  doing
• 11:04
  double
• 11:04
  duty
• 11:04
  as
• 11:04
  investor
• 11:05
  relations.
• 11:06
  Apple
• 11:06
  sued
• 11:07
  OpenAI
• 11:07
  alleging
• 11:08
  trade
• 11:08
  secret
• 11:08
  theft
• 11:09
  at
• 11:09
  every
• 11:09
  organizational
• 11:10
  level,
• 11:10
  from
• 11:10
  individual
• 11:11
  engineers
• 11:11
  to
• 11:11
  institutional
• 11:12
  knowledge.
• 11:13
  The
• 11:13
  community
• 11:14
  discussion
• 11:14
  has
• 11:14
  pivoted
• 11:15
  toward
• 11:15
  what
• 11:15
  this
• 11:15
  means
• 11:15
  for
• 11:16
  on-device
• 11:16
  AI:
• 11:17
  Apple's
• 11:18
  M7/M8
• 11:20
  chip
• 11:20
  trajectory
• 11:21
  suggests
• 11:21
  a
• 11:21
  future
• 11:22
  where
• 11:22
  models
• 11:22
  run
• 11:22
  locally
• 11:23
  rather
• 11:24
  than
• 11:24
  through
• 11:24
  API
• 11:24
  calls,
• 11:25
  which
• 11:25
  would
• 11:25
  fundamentally
• 11:26
  change
• 11:26
  the
• 11:27
  competitive
• 11:27
  landscape
• 11:28
  and
• 11:28
  make
• 11:28
  the
• 11:28
  question
• 11:29
  of
• 11:29
  who
• 11:29
  trained
• 11:29
  on
• 11:29
  whose
• 11:30
  data
• 11:30
  even
• 11:30
  more
• 11:30
  consequential.
• 11:32
  Meanwhile,
• 11:32
  DeepSeek
• 11:33
  is
• 11:33
  developing
• 11:34
  a
• 11:34
  custom
• 11:34
  ASIC
• 11:34
  on
• 11:34
  7nm
• 11:35
  DUV
• 11:36
  (deep
• 11:36
  ultraviolet)
• 11:37
  lithography
• 11:37
  —
• 11:38
  a
• 11:38
  deliberate
• 11:38
  choice
• 11:39
  that
• 11:39
  avoids
• 11:39
  competing
• 11:40
  for
• 11:40
  TSMC's
• 11:40
  oversubscribed
• 11:41
  EUV
• 11:42
  capacity.
• 11:43
  This
• 11:43
  is
• 11:44
  a
• 11:44
  triple-play
• 11:44
  vertical
• 11:45
  integration:
• 11:46
  control
• 11:46
  the
• 11:46
  silicon,
• 11:47
  control
• 11:47
  the
• 11:47
  inference
• 11:48
  cost
• 11:48
  structure,
• 11:49
  and
• 11:49
  sidestep
• 11:50
  US
• 11:50
  export
• 11:50
  controls
• 11:51
  on
• 11:51
  advanced
• 11:51
  lithography
• 11:52
  equipment
• 11:52
  in
• 11:52
  one
• 11:53
  move.
• 11:54
  Given
• 11:54
  DeepSeek's
• 11:55
  track
• 11:55
  record
• 11:55
  of
• 11:55
  achieving
• 11:56
  competitive
• 11:56
  performance
• 11:57
  at
• 11:57
  a
• 11:57
  fraction
• 11:57
  of
• 11:57
  the
• 11:57
  compute
• 11:58
  budget
• 11:58
  (their
• 11:59
  V3
• 11:59
  model
• 11:59
  famously
• 12:00
  trained
• 12:00
  for
• 12:00
  $5-6M),
• 12:02
  custom
• 12:02
  silicon
• 12:02
  purpose-built
• 12:03
  for
• 12:03
  their
• 12:03
  architecture
• 12:04
  could
• 12:04
  further
• 12:05
  widen
• 12:05
  the
• 12:05
  cost
• 12:05
  gap.
• 12:06
  The
• 12:06
  cost
• 12:07
  externalities
• 12:07
  of
• 12:07
  AI
• 12:08
  infrastructure
• 12:08
  are
• 12:09
  becoming
• 12:09
  impossible
• 12:09
  to
• 12:10
  ignore.
• 12:11
  An
• 12:11
  analysis
• 12:11
  of
• 12:12
  the
• 12:12
  PJM
• 12:12
  interconnection
• 12:13
  region
• 12:13
  —
• 12:14
  the
• 12:14
  largest
• 12:14
  wholesale
• 12:15
  electricity
• 12:15
  market
• 12:16
  in
• 12:16
  the
• 12:16
  US,
• 12:16
  spanning
• 12:17
  14
• 12:17
  mid-Atlantic
• 12:18
  and
• 12:18
  Midwest
• 12:18
  states
• 12:19
  —
• 12:19
  projects
• 12:20
  $23
• 12:20
  billion
• 12:21
  in
• 12:21
  customer
• 12:22
  price
• 12:22
  increases
• 12:22
  through
• 12:23
  2028,
• 12:24
  driven
• 12:24
  primarily
• 12:24
  by
• 12:24
  data
• 12:25
  center
• 12:25
  demand.
• 12:26
  The
• 12:26
  exploitation
• 12:27
  mechanism
• 12:28
  is
• 12:28
  specific:
• 12:28
  coincident
• 12:29
  peak
• 12:29
  demand
• 12:30
  billing
• 12:30
  allocates
• 12:30
  grid
• 12:31
  costs
• 12:31
  based
• 12:31
  on
• 12:31
  usage
• 12:32
  at
• 12:32
  the
• 12:32
  moment
• 12:32
  of
• 12:32
  system-wide
• 12:33
  peak
• 12:33
  load.
• 12:34
  Data
• 12:34
  centers
• 12:35
  can
• 12:35
  fine-tune
• 12:36
  their
• 12:36
  consumption
• 12:36
  minute-by-minute,
• 12:38
  algorithmically
• 12:39
  reducing
• 12:39
  draw
• 12:39
  during
• 12:40
  predicted
• 12:40
  peaks
• 12:40
  while
• 12:41
  running
• 12:41
  at
• 12:41
  full
• 12:41
  capacity
• 12:42
  the
• 12:42
  rest
• 12:42
  of
• 12:42
  the
• 12:42
  time
• 12:42
  —
• 12:43
  a
• 12:43
  flexibility
• 12:44
  residential
• 12:45
  customers
• 12:45
  cannot
• 12:46
  replicate.
• 12:47
  When
• 12:47
  utilities
• 12:48
  invest
• 12:48
  in
• 12:48
  grid
• 12:48
  infrastructure
• 12:49
  to
• 12:49
  serve
• 12:49
  data
• 12:50
  center
• 12:50
  loads
• 12:50
  that
• 12:50
  subsequently
• 12:51
  underperform
• 12:52
  projections
• 12:53
  or
• 12:53
  become
• 12:53
  obsolete,
• 12:54
  those
• 12:54
  costs
• 12:55
  get
• 12:55
  spread
• 12:55
  across
• 12:55
  all
• 12:55
  remaining
• 12:56
  customers.
• 12:57
  Every
• 12:58
  state
• 12:58
  except
• 12:58
  Georgia,
• 12:59
  Idaho,
• 12:59
  and
• 13:00
  Louisiana
• 13:00
  has
• 13:00
  a
• 13:01
  consumer
• 13:01
  advocate
• 13:01
  office,
• 13:03
  but
• 13:03
  these
• 13:03
  offices
• 13:03
  are
• 13:04
  often
• 13:04
  legally
• 13:04
  barred
• 13:05
  from
• 13:05
  advocating
• 13:05
  for
• 13:06
  one
• 13:06
  customer
• 13:06
  class
• 13:07
  over
• 13:07
  another
• 13:07
  —
• 13:08
  meaning
• 13:08
  nobody
• 13:08
  effectively
• 13:09
  argues
• 13:09
  against
• 13:10
  data
• 13:10
  centers'
• 13:10
  proposals
• 13:11
  for
• 13:11
  minimal
• 13:11
  cost
• 13:12
  allocation.
• 13:13
  The
• 13:13
  result
• 13:13
  is
• 13:13
  a
• 13:14
  regulatory
• 13:14
  capture
• 13:14
  dynamic
• 13:15
  where
• 13:15
  AI
• 13:16
  compute
• 13:16
  costs
• 13:16
  are
• 13:17
  socialized
• 13:17
  while
• 13:17
  the
• 13:18
  profits
• 13:18
  remain
• 13:18
  private.
• 13:20
  Anthropic
• 13:21
  published
• 13:21
  research
• 13:22
  analyzing
• 13:22
  310,000
• 13:23
  conversations
• 13:24
  across
• 13:24
  languages,
• 13:25
  identifying
• 13:25
  four
• 13:26
  value
• 13:26
  axes
• 13:27
  along
• 13:27
  which
• 13:27
  Claude's
• 13:27
  behavior
• 13:28
  shifts:
• 13:28
  compliance/caution,
• 13:30
  warmth/severity,
• 13:31
  depth/brevity,
• 13:32
  and
• 13:32
  candor/efficiency.
• 13:34
  Russian-language
• 13:34
  conversations
• 13:35
  produce
• 13:35
  the
• 13:35
  most
• 13:35
  strict
• 13:36
  responses;
• 13:36
  Hindi
• 13:37
  the
• 13:37
  warmest.
• 13:38
  The
• 13:38
  researchers
• 13:38
  attribute
• 13:39
  these
• 13:39
  differences
• 13:39
  to
• 13:39
  training
• 13:40
  data
• 13:40
  composition
• 13:41
  rather
• 13:41
  than
• 13:41
  deliberate
• 13:42
  tuning
• 13:42
  —
• 13:42
  which
• 13:43
  is
• 13:43
  arguably
• 13:43
  more
• 13:43
  concerning,
• 13:44
  because
• 13:44
  it
• 13:44
  means
• 13:44
  the
• 13:44
  biases
• 13:45
  are
• 13:45
  structural
• 13:45
  and
• 13:45
  difficult
• 13:46
  to
• 13:46
  correct
• 13:46
  without
• 13:46
  fundamentally
• 13:47
  changing
• 13:47
  the
• 13:47
  data
• 13:48
  pipeline.
• 13:48
  A
• 13:49
  Russian-speaking
• 13:49
  user
• 13:49
  asking
• 13:50
  for
• 13:50
  the
• 13:50
  same
• 13:50
  help
• 13:50
  as
• 13:50
  a
• 13:50
  Hindi-speaking
• 13:51
  user
• 13:51
  gets
• 13:52
  a
• 13:52
  measurably
• 13:52
  different
• 13:52
  personality,
• 13:53
  not
• 13:54
  because
• 13:54
  anyone
• 13:54
  designed
• 13:54
  it
• 13:54
  that
• 13:55
  way,
• 13:55
  but
• 13:55
  because
• 13:55
  the
• 13:55
  internet's
• 13:56
  language
• 13:56
  distribution
• 13:57
  encodes
• 13:57
  cultural
• 13:58
  assumptions
• 13:58
  that
• 13:58
  survive
• 13:58
  through
• 13:59
  training.
• 13:59
  For
• 14:00
  anyone
• 14:00
  deploying
• 14:00
  Claude
• 14:01
  in
• 14:01
  multilingual
• 14:01
  contexts,
• 14:02
  this
• 14:02
  is
• 14:02
  not
• 14:03
  an
• 14:03
  academic
• 14:03
  curiosity;
• 14:04
  it
• 14:05
  is
• 14:05
  a
• 14:05
  deployment
• 14:05
  constraint
• 14:06
  that
• 14:06
  affects
• 14:06
  user
• 14:06
  experience,
• 14:07
  trust,
• 14:08
  and
• 14:08
  retention
• 14:09
  in
• 14:09
  measurable
• 14:09
  ways.
• 14:10
  DoorDash's
• 14:11
  LLM
• 14:11
  jury
• 14:12
  system
• 14:12
  for
• 14:12
  food
• 14:12
  metadata
• 14:13
  generation
• 14:13
  claims
• 14:13
  20%
• 14:14
  higher
• 14:14
  accuracy
• 14:15
  than
• 14:15
  human
• 14:15
  annotators
• 14:15
  —
• 14:16
  a
• 14:16
  result
• 14:17
  worth
• 14:17
  examining
• 14:17
  given
• 14:18
  prior
• 14:18
  concerns
• 14:18
  about
• 14:19
  LLM-as-judge
• 14:19
  circular
• 14:20
  dependencies.
• 14:21
  Their
• 14:21
  approach
• 14:21
  uses
• 14:22
  multiple
• 14:22
  strong
• 14:22
  LLM
• 14:23
  evaluators
• 14:23
  with
• 14:24
  independent
• 14:24
  verdicts
• 14:24
  and
• 14:24
  rationale,
• 14:25
  aggregated
• 14:26
  into
• 14:26
  consensus
• 14:26
  decisions,
• 14:27
  with
• 14:27
  per-tag
• 14:28
  validation
• 14:28
  (protein,
• 14:29
  preparation,
• 14:30
  health
• 14:30
  attributes
• 14:30
  individually)
• 14:31
  rather
• 14:31
  than
• 14:31
  holistic
• 14:32
  item-level
• 14:32
  judgment.
• 14:34
  The
• 14:34
  domain
• 14:34
  is
• 14:34
  genuinely
• 14:35
  hard:
• 14:35
  the
• 14:35
  same
• 14:36
  dish
• 14:36
  can
• 14:36
  be
• 14:36
  described
• 14:37
  in
• 14:37
  countless
• 14:37
  ways
• 14:37
  across
• 14:38
  cuisines,
• 14:39
  and
• 14:39
  only
• 14:39
  a
• 14:39
  handful
• 14:40
  of
• 14:40
  human
• 14:40
  experts
• 14:41
  reliably
• 14:41
  distinguish
• 14:42
  Nepalese
• 14:42
  from
• 14:42
  North
• 14:43
  Indian
• 14:43
  preparations
• 14:44
  or
• 14:44
  interpret
• 14:45
  whether
• 14:45
  "Sichuan-style"
• 14:45
  implies
• 14:46
  a
• 14:46
  specific
• 14:46
  heat
• 14:46
  profile.
• 14:48
  A
• 14:48
  context
• 14:48
  optimization
• 14:49
  loop
• 14:49
  inspired
• 14:50
  by
• 14:50
  reinforcement
• 14:50
  learning
• 14:51
  accelerated
• 14:51
  prompt
• 14:52
  development
• 14:52
  tenfold
• 14:53
  and
• 14:53
  improved
• 14:53
  precision
• 14:54
  by
• 14:54
  over
• 14:54
  20%,
• 14:56
  driven
• 14:56
  by
• 14:56
  failure-signal
• 14:57
  analysis
• 14:58
  rather
• 14:58
  than
• 14:58
  population-based
• 14:59
  evolutionary
• 15:00
  search
• 15:00
  —
• 15:00
  the
• 15:01
  team
• 15:01
  found
• 15:01
  that
• 15:01
  weighting
• 15:02
  failure
• 15:02
  cases
• 15:02
  more
• 15:03
  heavily
• 15:03
  outperformed
• 15:04
  balanced
• 15:04
  sampling.
• 15:05
  Fine-tuned
• 15:06
  small
• 15:06
  language
• 15:06
  models
• 15:07
  match
• 15:07
  frontier
• 15:08
  LLM
• 15:08
  quality
• 15:09
  at
• 15:09
  10%
• 15:09
  of
• 15:09
  inference
• 15:10
  cost
• 15:10
  with
• 15:10
  zero
• 15:11
  human
• 15:11
  annotation
• 15:11
  effort.
• 15:12
  A
• 15:12
  distributed
• 15:13
  pipeline
• 15:13
  using
• 15:13
  Spark-based
• 15:14
  deduplication,
• 15:15
  batch
• 15:15
  LLM
• 15:16
  APIs,
• 15:17
  and
• 15:17
  result
• 15:17
  remapping
• 15:18
  cut
• 15:18
  backfill
• 15:19
  from
• 15:19
  over
• 15:19
  a
• 15:19
  month
• 15:19
  to
• 15:20
  days
• 15:20
  across
• 15:20
  millions
• 15:21
  of
• 15:21
  menu
• 15:21
  items.