Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – July 14, 2026

Transcript

AGI Dreams Podcast – July 14, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  💰
• 0:00
  Open-Weight
• 0:00
  Cost
• 0:00
  Revolution
• 0:01
  The
• 0:01
  Financial
• 0:01
  Times
• 0:02
  reports
• 0:02
  that
• 0:02
  enterprises
• 0:03
  are
• 0:03
  quietly
• 0:03
  migrating
• 0:04
  to
• 0:04
  Chinese
• 0:04
  open-weight
• 0:05
  models
• 0:05
  to
• 0:05
  cut
• 0:05
  inference
• 0:06
  costs,
• 0:07
  a
• 0:07
  shift
• 0:07
  the
• 0:07
  community
• 0:08
  has
• 0:08
  been
• 0:08
  predicting
• 0:08
  since
• 0:09
  Anthropic's
• 0:09
  Mythos/Fable-era
• 0:11
  licensing
• 0:11
  restrictions.
• 0:12
  The
• 0:12
  corporate
• 0:12
  calculus
• 0:13
  is
• 0:13
  straightforward:
• 0:14
  when
• 0:14
  an
• 0:14
  employee's
• 0:15
  weekly
• 0:15
  AI
• 0:15
  token
• 0:16
  budget
• 0:16
  rivals
• 0:16
  their
• 0:16
  salary,
• 0:17
  and
• 0:17
  the
• 0:17
  CEO
• 0:18
  mandates
• 0:18
  usage
• 0:18
  metrics
• 0:19
  that
• 0:19
  look
• 0:19
  good
• 0:19
  to
• 0:19
  investors,
• 0:20
  something
• 0:21
  has
• 0:21
  to
• 0:21
  give.
• 0:22
  For
• 0:22
  many
• 0:22
  teams,
• 0:23
  that
• 0:23
  something
• 0:23
  is
• 0:23
  the
• 0:23
  provider.
• 0:24
  One
• 0:24
  commenter
• 0:25
  captured
• 0:25
  the
• 0:25
  prevailing
• 0:26
  mood:
• 0:26
  their
• 0:26
  company
• 0:27
  spent
• 0:27
  double
• 0:27
  a
• 0:27
  developer's
• 0:28
  weekly
• 0:28
  pay
• 0:28
  on
• 0:28
  Anthropic
• 0:29
  tokens,
• 0:30
  so
• 0:30
  when
• 0:30
  told
• 0:30
  to
• 0:30
  use
• 0:30
  AI
• 0:31
  harder,
• 0:31
  they
• 0:31
  "looped
• 0:32
  the
• 0:32
  shit
• 0:32
  out
• 0:32
  of
• 0:32
  it."
• 0:33
  The
• 0:33
  geopolitical
• 0:34
  undertone
• 0:34
  is
• 0:34
  loud
• 0:35
  —
• 0:35
  a
• 0:35
  separate
• 0:36
  thread
• 0:36
  asks
• 0:36
  why
• 0:36
  no
• 0:37
  American
• 0:37
  open-source
• 0:38
  lab
• 0:38
  is
• 0:38
  even
• 0:38
  close
• 0:38
  to
• 0:38
  Chinese
• 0:39
  ones
• 0:39
  on
• 0:39
  benchmarks
• 0:39
  —
• 0:40
  and
• 0:40
  the
• 0:40
  community's
• 0:41
  verdict
• 0:41
  is
• 0:41
  blunt:
• 0:42
  this
• 0:42
  is
• 0:42
  not
• 0:42
  an
• 0:42
  "AI
• 0:43
  bubble,"
• 0:43
  it
• 0:44
  is
• 0:44
  a
• 0:44
  "US
• 0:44
  frontier
• 0:45
  models
• 0:45
  bubble"
• 0:45
  .
• 0:46
  The
• 0:46
  cost
• 0:46
  pressure
• 0:46
  is
• 0:47
  compressing
• 0:47
  the
• 0:47
  stack
• 0:47
  all
• 0:48
  the
• 0:48
  way
• 0:48
  down
• 0:48
  to
• 0:48
  the
• 0:48
  device.
• 0:50
  PrismML,
• 0:51
  a
• 0:51
  Caltech
• 0:51
  spinoff
• 0:52
  backed
• 0:52
  by
• 0:52
  Khosla
• 0:52
  Ventures,
• 0:53
  claims
• 0:54
  to
• 0:54
  have
• 0:54
  compressed
• 0:54
  Qwen
• 0:55
  3.6-27B
• 0:58
  from
• 0:58
  54
• 0:59
  GB
• 0:59
  to
• 0:59
  under
• 1:00
  4
• 1:00
  GB
• 1:00
  using
• 1:01
  ternary
• 1:01
  quantization
• 1:01
  —
• 1:02
  small
• 1:03
  enough
• 1:03
  to
• 1:03
  run
• 1:03
  on
• 1:03
  an
• 1:03
  iPhone
• 1:04
  17
• 1:04
  Pro
• 1:04
  with
• 1:05
  all
• 1:05
  27
• 1:05
  billion
• 1:06
  parameters
• 1:06
  active
• 1:06
  simultaneously.
• 1:08
  Apple's
• 1:08
  own
• 1:08
  on-device
• 1:09
  model,
• 1:09
  by
• 1:09
  comparison,
• 1:10
  activates
• 1:11
  only
• 1:11
  1–4
• 1:12
  billion
• 1:12
  of
• 1:13
  its
• 1:13
  20
• 1:13
  billion
• 1:13
  parameters
• 1:14
  at
• 1:14
  a
• 1:14
  time.
• 1:15
  The
• 1:15
  community
• 1:15
  is
• 1:15
  skeptical
• 1:16
  —
• 1:17
  "paper
• 1:17
  or
• 1:17
  it
• 1:17
  didn't
• 1:17
  happen"
• 1:18
  is
• 1:18
  the
• 1:18
  prevailing
• 1:19
  response
• 1:19
  —
• 1:20
  and
• 1:20
  the
• 1:20
  legitimate
• 1:21
  concern
• 1:21
  is
• 1:21
  whether
• 1:22
  QAT
• 1:22
  at
• 1:22
  this
• 1:22
  compression
• 1:23
  ratio
• 1:23
  preserves
• 1:23
  capability
• 1:24
  in
• 1:24
  ways
• 1:24
  that
• 1:25
  matter
• 1:25
  beyond
• 1:25
  benchmarks:
• 1:26
  tool
• 1:26
  use,
• 1:27
  long-context
• 1:28
  reasoning,
• 1:28
  the
• 1:28
  things
• 1:29
  that
• 1:29
  make
• 1:29
  a
• 1:29
  model
• 1:29
  usable
• 1:30
  rather
• 1:30
  than
• 1:30
  merely
• 1:31
  impressive
• 1:31
  .
• 1:32
  While
• 1:32
  PrismML
• 1:33
  squeezes
• 1:33
  parameters,
• 1:34
  BottleCap
• 1:34
  AI
• 1:35
  attacks
• 1:35
  the
• 1:35
  other
• 1:35
  side
• 1:36
  of
• 1:36
  the
• 1:36
  cost
• 1:36
  equation:
• 1:37
  thinking
• 1:37
  tokens.
• 1:38
  Their
• 1:38
  ThinkingCap
• 1:39
  fine-tune
• 1:39
  of
• 1:39
  Qwen
• 1:40
  3.6-27B
• 1:43
  cuts
• 1:43
  reasoning
• 1:44
  tokens
• 1:44
  by
• 1:44
  46%
• 1:45
  on
• 1:45
  average
• 1:45
  —
• 1:46
  over
• 1:46
  60%
• 1:47
  on
• 1:47
  the
• 1:47
  hardest
• 1:48
  benchmarks
• 1:48
  like
• 1:48
  GPQA-Diamond
• 1:49
  —
• 1:51
  while
• 1:51
  tracking
• 1:51
  the
• 1:52
  base
• 1:52
  model's
• 1:52
  accuracy
• 1:53
  within
• 1:53
  a
• 1:53
  percentage
• 1:53
  point
• 1:54
  across
• 1:54
  twelve
• 1:54
  out-of-domain
• 1:55
  benchmarks.
• 1:56
  The
• 1:56
  key
• 1:56
  insight
• 1:57
  is
• 1:57
  that
• 1:57
  long
• 1:58
  reasoning
• 1:58
  traces
• 1:58
  have
• 1:58
  become
• 1:59
  conflated
• 1:59
  with
• 1:59
  intelligence
• 2:00
  when
• 2:00
  they
• 2:01
  often
• 2:01
  represent
• 2:01
  inefficiency:
• 2:03
  the
• 2:03
  model
• 2:03
  narrating
• 2:03
  its
• 2:03
  solution
• 2:04
  rather
• 2:04
  than
• 2:05
  solving.
• 2:06
  An
• 2:06
  accidental
• 2:07
  side
• 2:07
  effect
• 2:07
  —
• 2:08
  shorter
• 2:08
  answers
• 2:08
  after
• 2:09
  the
• 2:09
  think
• 2:09
  block
• 2:09
  —
• 2:10
  turned
• 2:10
  out
• 2:10
  to
• 2:10
  be
• 2:10
  the
• 2:10
  feature
• 2:11
  users
• 2:11
  preferred.
• 2:12
  The
• 2:12
  model
• 2:12
  ships
• 2:13
  under
• 2:13
  Apache
• 2:13
  2.0,
• 2:15
  so
• 2:15
  anyone
• 2:15
  running
• 2:15
  Qwen
• 2:16
  locally
• 2:16
  can
• 2:16
  swap
• 2:16
  it
• 2:16
  in
• 2:17
  and
• 2:17
  halve
• 2:17
  their
• 2:17
  token
• 2:18
  spend
• 2:18
  immediately
• 2:18
  .
• 2:19
  At
• 2:19
  the
• 2:19
  extreme
• 2:20
  end
• 2:20
  of
• 2:20
  the
• 2:20
  democratization
• 2:21
  curve,
• 2:22
  a
• 2:22
  developer
• 2:22
  got
• 2:22
  GLM-5.2
• 2:23
  —
• 2:24
  a
• 2:24
  744-billion-parameter
• 2:25
  MoE
• 2:26
  —
• 2:27
  streaming
• 2:27
  from
• 2:28
  disk
• 2:28
  on
• 2:28
  a
• 2:28
  consumer
• 2:28
  machine
• 2:29
  with
• 2:29
  just
• 2:29
  25
• 2:30
  GB
• 2:30
  of
• 2:30
  RAM.
• 2:31
  The
• 2:31
  speed
• 2:31
  is
• 2:31
  glacial,
• 2:32
  but
• 2:32
  the
• 2:32
  community
• 2:33
  correctly
• 2:33
  identified
• 2:34
  the
• 2:34
  real
• 2:34
  point:
• 2:35
  nobody
• 2:35
  is
• 2:35
  going
• 2:35
  to
• 2:35
  use
• 2:36
  this
• 2:36
  for
• 2:36
  production
• 2:36
  inference.
• 2:38
  The
• 2:38
  interesting
• 2:38
  part
• 2:39
  is
• 2:39
  that
• 2:39
  you
• 2:39
  can
• 2:39
  stream
• 2:39
  744
• 2:41
  billion
• 2:41
  parameters
• 2:41
  from
• 2:42
  an
• 2:42
  NVMe
• 2:42
  at
• 2:42
  all.
• 2:43
  As
• 2:43
  one
• 2:44
  commenter
• 2:44
  noted,
• 2:45
  "we
• 2:45
  went
• 2:45
  from
• 2:45
  overnight
• 2:46
  movie
• 2:46
  downloads
• 2:46
  to
• 2:47
  overnight
• 2:47
  model
• 2:47
  inference
• 2:48
  —
• 2:49
  nature
• 2:49
  is
• 2:49
  healing."
• 2:51
  If
• 2:51
  expert
• 2:51
  routing
• 2:51
  prediction
• 2:52
  improves
• 2:52
  enough
• 2:53
  to
• 2:53
  prefetch
• 2:53
  the
• 2:53
  right
• 2:53
  experts
• 2:54
  before
• 2:54
  they
• 2:54
  are
• 2:54
  needed,
• 2:55
  the
• 2:55
  entire
• 2:56
  economics
• 2:56
  of
• 2:56
  local
• 2:57
  inference
• 2:57
  changes
• 2:57
  .
• 3:00
  🦀
• 3:00
  Inference
• 3:00
  Engines:
• 3:01
  The
• 3:01
  Rust
• 3:02
  Wave
• 3:02
  Three
• 3:02
  projects
• 3:02
  this
• 3:03
  week
• 3:03
  point
• 3:03
  toward
• 3:03
  Rust
• 3:04
  becoming
• 3:04
  a
• 3:04
  serious
• 3:04
  contender
• 3:05
  in
• 3:05
  inference,
• 3:06
  a
• 3:06
  space
• 3:06
  long
• 3:07
  dominated
• 3:07
  by
• 3:07
  C++
• 3:08
  and
• 3:08
  CUDA.
• 3:09
  The
• 3:09
  most
• 3:09
  striking
• 3:10
  benchmark
• 3:10
  comes
• 3:10
  from
• 3:11
  mistral.rs
• 3:12
  v0.9.0,
• 3:14
  which
• 3:14
  claims
• 3:15
  up
• 3:15
  to
• 3:15
  1.8x
• 3:15
  faster
• 3:16
  CPU
• 3:16
  decode
• 3:17
  than
• 3:17
  llama.cpp
• 3:18
  on
• 3:19
  Qwen3
• 3:19
  4B
• 3:19
  Q4
• 3:20
  K
• 3:20
  across
• 3:20
  both
• 3:21
  x86
• 3:21
  (Sapphire
• 3:21
  Rapids)
• 3:22
  and
• 3:22
  ARM
• 3:22
  (GB10).
• 3:23
  The
• 3:23
  optimizations
• 3:24
  are
• 3:24
  granular
• 3:25
  —
• 3:26
  AVX2,
• 3:27
  AVX512,
• 3:28
  ARM
• 3:29
  NEON
• 3:29
  —
• 3:30
  and
• 3:30
  the
• 3:30
  team
• 3:30
  swept
• 3:30
  configurations
• 3:31
  for
• 3:31
  both
• 3:32
  engines
• 3:32
  to
• 3:32
  ensure
• 3:32
  the
• 3:32
  comparison
• 3:33
  favored
• 3:33
  each
• 3:34
  at
• 3:34
  its
• 3:34
  best.
• 3:35
  The
• 3:35
  benchmarks
• 3:35
  were
• 3:35
  run
• 3:36
  on
• 3:36
  Qwen3
• 3:36
  4B
• 3:37
  Q4
• 3:37
  K
• 3:37
  at
• 3:37
  varying
• 3:38
  context
• 3:38
  depths,
• 3:39
  and
• 3:39
  mistral.rs
• 3:40
  came
• 3:40
  out
• 3:40
  ahead
• 3:40
  at
• 3:41
  every
• 3:41
  measured
• 3:41
  point
• 3:41
  on
• 3:42
  both
• 3:42
  architectures.
• 3:43
  Full
• 3:44
  methodology
• 3:44
  and
• 3:44
  repro
• 3:44
  scripts
• 3:45
  are
• 3:45
  published,
• 3:46
  with
• 3:46
  an
• 3:46
  open
• 3:46
  invitation
• 3:47
  for
• 3:47
  independent
• 3:47
  reproductions
• 3:48
  on
• 3:48
  untested
• 3:49
  hardware
• 3:49
  .
• 3:50
  At
• 3:50
  the
• 3:50
  opposite
• 3:50
  end
• 3:51
  of
• 3:51
  the
• 3:51
  maturity
• 3:51
  spectrum,
• 3:52
  a
• 3:52
  developer
• 3:52
  built
• 3:52
  ferrite,
• 3:53
  a
• 3:53
  from-scratch
• 3:54
  inference
• 3:54
  engine
• 3:55
  for
• 3:55
  Qwen3.5's
• 3:57
  hybrid
• 3:58
  Mamba-Transformer
• 3:59
  architecture,
• 3:59
  entirely
• 4:00
  in
• 4:00
  Rust
• 4:00
  with
• 4:00
  no
• 4:00
  ggml,
• 4:01
  Candle,
• 4:02
  or
• 4:02
  PyTorch
• 4:02
  dependency.
• 4:04
  The
• 4:04
  project
• 4:04
  parsed
• 4:05
  raw
• 4:05
  GGUF
• 4:05
  bytes,
• 4:06
  wrote
• 4:06
  its
• 4:06
  own
• 4:07
  dequantization
• 4:07
  kernels
• 4:08
  across
• 4:08
  eight
• 4:08
  mixed
• 4:08
  quantization
• 4:09
  formats,
• 4:10
  and
• 4:10
  implemented
• 4:10
  the
• 4:10
  full
• 4:11
  forward
• 4:11
  pass
• 4:11
  by
• 4:11
  hand.
• 4:13
  The
• 4:13
  architecture
• 4:13
  —
• 4:14
  alternating
• 4:15
  Gated
• 4:15
  DeltaNet
• 4:16
  (a
• 4:16
  Mamba-family
• 4:16
  SSM)
• 4:17
  with
• 4:17
  standard
• 4:18
  GQA
• 4:18
  attention
• 4:19
  every
• 4:19
  fourth
• 4:19
  layer
• 4:19
  —
• 4:21
  exposed
• 4:21
  implementation
• 4:22
  traps
• 4:22
  invisible
• 4:22
  from
• 4:23
  papers
• 4:23
  alone:
• 4:24
  interleaved
• 4:24
  memory
• 4:25
  layouts
• 4:25
  for
• 4:25
  gated
• 4:25
  attention,
• 4:26
  silent
• 4:27
  weight-ordering
• 4:27
  bugs
• 4:28
  that
• 4:28
  produce
• 4:28
  plausible-looking
• 4:29
  but
• 4:29
  wrong
• 4:29
  floats,
• 4:30
  and
• 4:30
  a
• 4:30
  missing
• 4:30
  pre-normalization
• 4:31
  step
• 4:32
  that
• 4:32
  caused
• 4:32
  gate
• 4:32
  values
• 4:33
  to
• 4:33
  grow
• 4:33
  geometrically
• 4:33
  with
• 4:34
  depth.
• 4:35
  The
• 4:35
  generated
• 4:35
  text
• 4:36
  is
• 4:36
  not
• 4:36
  yet
• 4:36
  coherent,
• 4:37
  but
• 4:37
  the
• 4:37
  diagnostic
• 4:38
  methodology
• 4:38
  —
• 4:39
  comparing
• 4:40
  hidden
• 4:40
  states
• 4:40
  layer-by-layer
• 4:41
  against
• 4:41
  the
• 4:41
  HuggingFace
• 4:42
  reference,
• 4:43
  then
• 4:43
  using
• 4:43
  a
• 4:43
  BF16-vs-quantized
• 4:45
  precision
• 4:45
  sweep
• 4:45
  to
• 4:46
  distinguish
• 4:46
  logic
• 4:46
  bugs
• 4:47
  from
• 4:47
  quantization
• 4:47
  noise
• 4:48
  —
• 4:48
  is
• 4:49
  a
• 4:49
  contribution
• 4:49
  in
• 4:49
  its
• 4:49
  own
• 4:50
  right
• 4:50
  .
• 4:51
  Meanwhile,
• 4:51
  llama.cpp
• 4:53
  shipped
• 4:54
  b9978
• 4:55
  to
• 4:55
  fix
• 4:56
  a
• 4:56
  checkpoint
• 4:56
  bug
• 4:57
  that
• 4:57
  disproportionately
• 4:58
  hit
• 4:58
  agentic
• 4:58
  workloads.
• 4:59
  Every
• 5:00
  agent
• 5:00
  turn
• 5:00
  was
• 5:00
  creating
• 5:01
  a
• 5:01
  new
• 5:01
  checkpoint
• 5:02
  regardless
• 5:02
  of
• 5:02
  minimum-step
• 5:03
  spacing,
• 5:04
  collapsing
• 5:05
  the
• 5:05
  coverage
• 5:05
  window
• 5:05
  so
• 5:06
  that
• 5:06
  context
• 5:06
  rewinds
• 5:07
  —
• 5:07
  common
• 5:08
  in
• 5:08
  tool-calling
• 5:08
  loops
• 5:08
  —
• 5:09
  erased
• 5:10
  all
• 5:10
  checkpoints
• 5:10
  and
• 5:11
  forced
• 5:11
  full
• 5:11
  reprocessing.
• 5:13
  For
• 5:13
  multi-turn
• 5:13
  agent
• 5:14
  sessions
• 5:14
  running
• 5:15
  tool-calling
• 5:15
  loops
• 5:15
  with
• 5:16
  frequent
• 5:16
  context
• 5:16
  rewinds,
• 5:17
  this
• 5:18
  was
• 5:18
  the
• 5:18
  difference
• 5:18
  between
• 5:18
  fast
• 5:19
  context
• 5:19
  restoration
• 5:20
  and
• 5:20
  minutes
• 5:20
  of
• 5:20
  wasted
• 5:21
  recomputation
• 5:21
  —
• 5:22
  a
• 5:22
  fix
• 5:23
  that
• 5:23
  matters
• 5:23
  more
• 5:24
  as
• 5:24
  local
• 5:24
  inference
• 5:24
  becomes
• 5:25
  the
• 5:25
  default
• 5:25
  backend
• 5:26
  for
• 5:26
  agentic
• 5:26
  workflows
• 5:27
  .
• 5:30
  🔬
• 5:30
  Model
• 5:30
  Surgery
• 5:31
  and
• 5:31
  Hallucination
• 5:32
  Mapping
• 5:32
  A
• 5:32
  community
• 5:32
  member
• 5:33
  who
• 5:33
  previously
• 5:33
  failed
• 5:34
  at
• 5:34
  expanding
• 5:34
  Gemma
• 5:35
  4's
• 5:35
  depth
• 5:35
  —
• 5:36
  the
• 5:36
  88-layer
• 5:37
  flop,
• 5:37
  documented
• 5:37
  in
• 5:38
  earlier
• 5:38
  posts
• 5:38
  —
• 5:39
  returned
• 5:39
  with
• 5:39
  a
• 5:39
  working
• 5:39
  approach.
• 5:40
  The
• 5:40
  key
• 5:41
  discovery:
• 5:41
  initializing
• 5:42
  new
• 5:42
  layers
• 5:42
  as
• 5:43
  pass-throughs
• 5:43
  starves
• 5:43
  them
• 5:44
  of
• 5:44
  learning
• 5:44
  signal.
• 5:45
  As
• 5:45
  the
• 5:45
  author
• 5:45
  put
• 5:46
  it,
• 5:46
  "a
• 5:46
  worker
• 5:46
  who
• 5:46
  does
• 5:46
  nothing
• 5:47
  gets
• 5:47
  no
• 5:47
  feedback."
• 5:48
  Instead,
• 5:49
  extGemma4-40
• 5:50
  5B
• 5:50
  initializes
• 5:51
  each
• 5:51
  new
• 5:51
  layer
• 5:52
  as
• 5:52
  a
• 5:52
  blend
• 5:52
  of
• 5:52
  its
• 5:52
  two
• 5:52
  neighbors,
• 5:53
  expanding
• 5:54
  from
• 5:54
  60
• 5:54
  to
• 5:54
  77
• 5:55
  layers.
• 5:56
  Two
• 5:56
  rounds
• 5:56
  of
• 5:56
  healing
• 5:57
  —
• 5:58
  first
• 5:58
  training
• 5:58
  only
• 5:58
  the
• 5:58
  new
• 5:59
  layers
• 5:59
  while
• 5:59
  the
• 5:59
  originals
• 6:00
  stayed
• 6:00
  frozen,
• 6:01
  then
• 6:01
  unfreezing
• 6:02
  everything
• 6:02
  —
• 6:03
  produced
• 6:03
  a
• 6:03
  model
• 6:04
  that
• 6:04
  matched
• 6:04
  the
• 6:04
  original
• 6:05
  across
• 6:05
  9
• 6:05
  of
• 6:05
  10
• 6:05
  domains
• 6:06
  and
• 6:06
  beat
• 6:06
  it
• 6:06
  on
• 6:06
  a
• 6:06
  physics
• 6:07
  problem
• 6:07
  the
• 6:07
  parent
• 6:07
  got
• 6:08
  wrong.
• 6:08
  The
• 6:08
  question
• 6:09
  involved
• 6:09
  relativistic
• 6:10
  energy
• 6:10
  at
• 6:10
  identical
• 6:11
  de
• 6:11
  Broglie
• 6:11
  wavelength;
• 6:12
  the
• 6:12
  parent
• 6:12
  forgot
• 6:12
  the
• 6:13
  electron's
• 6:13
  511
• 6:14
  keV
• 6:14
  rest
• 6:14
  mass.
• 6:15
  The
• 6:15
  author
• 6:15
  watched
• 6:16
  the
• 6:16
  model
• 6:16
  heal
• 6:16
  in
• 6:16
  real
• 6:17
  time
• 6:17
  during
• 6:17
  training:
• 6:18
  before
• 6:18
  healing,
• 6:19
  hitting
• 6:19
  a
• 6:19
  rare
• 6:19
  technical
• 6:20
  term
• 6:20
  derailed
• 6:20
  output
• 6:21
  into
• 6:21
  repeated
• 6:21
  non-Latin
• 6:22
  characters;
• 6:22
  13%
• 6:23
  into
• 6:23
  round
• 6:23
  two,
• 6:24
  the
• 6:24
  model
• 6:24
  stumbled
• 6:25
  once
• 6:25
  then
• 6:25
  recovered;
• 6:26
  by
• 6:26
  halfway,
• 6:27
  the
• 6:27
  problem
• 6:27
  vanished
• 6:28
  entirely.
• 6:29
  The
• 6:29
  technique
• 6:29
  challenges
• 6:30
  the
• 6:30
  conventional
• 6:30
  wisdom
• 6:31
  that
• 6:31
  major
• 6:31
  surgery
• 6:32
  on
• 6:32
  a
• 6:32
  fine-tuned
• 6:32
  model
• 6:33
  inevitably
• 6:33
  causes
• 6:34
  catastrophic
• 6:34
  forgetting
• 6:35
  .
• 6:35
  On
• 6:35
  the
• 6:36
  diagnostic
• 6:36
  side,
• 6:37
  an
• 6:37
  independent
• 6:37
  researcher
• 6:38
  stress-tested
• 6:39
  Anthropic's
• 6:39
  J-Space
• 6:40
  hallucination
• 6:40
  signal
• 6:41
  across
• 6:41
  11,400
• 6:42
  examples
• 6:43
  and
• 6:43
  seven
• 6:44
  datasets
• 6:44
  on
• 6:44
  Qwen3-4B,
• 6:46
  drawing
• 6:46
  sharp
• 6:46
  lines
• 6:47
  around
• 6:47
  where
• 6:47
  internal
• 6:47
  workspace
• 6:48
  entropy
• 6:48
  works
• 6:49
  and
• 6:49
  where
• 6:49
  it
• 6:49
  breaks.
• 6:50
  For
• 6:50
  factual
• 6:51
  retrieval
• 6:51
  on
• 6:51
  PopQA,
• 6:52
  J-Space
• 6:53
  caught
• 6:53
  high-confidence
• 6:54
  errors
• 6:54
  with
• 6:54
  100%
• 6:55
  precision
• 6:56
  where
• 6:56
  output
• 6:56
  logprobs
• 6:57
  managed
• 6:57
  worse
• 6:57
  than
• 6:57
  chance.
• 6:58
  On
• 6:59
  TruthfulQA
• 6:59
  —
• 7:00
  adversarial
• 7:01
  human
• 7:01
  myths
• 7:01
  —
• 7:02
  the
• 7:02
  signal
• 7:02
  collapsed
• 7:02
  entirely:
• 7:04
  even
• 7:04
  in
• 7:04
  the
• 7:04
  "safest"
• 7:04
  quadrant
• 7:05
  (high
• 7:05
  confidence,
• 7:06
  clean
• 7:06
  workspace),
• 7:07
  the
• 7:07
  model
• 7:07
  was
• 7:07
  wrong
• 7:08
  84.9%
• 7:09
  of
• 7:09
  the
• 7:09
  time.
• 7:10
  The
• 7:10
  mechanistic
• 7:10
  takeaway:
• 7:11
  workspace
• 7:12
  noise
• 7:12
  detects
• 7:13
  epistemic
• 7:13
  guessing
• 7:14
  but
• 7:14
  cannot
• 7:15
  detect
• 7:15
  ontological
• 7:16
  falsehoods
• 7:16
  burned
• 7:17
  into
• 7:17
  pretraining
• 7:17
  data.
• 7:18
  Math
• 7:19
  posed
• 7:19
  a
• 7:19
  different
• 7:19
  problem
• 7:19
  —
• 7:20
  correct
• 7:20
  GSM8K
• 7:23
  answers
• 7:23
  had
• 7:23
  higher
• 7:24
  mean
• 7:24
  noise
• 7:24
  than
• 7:24
  the
• 7:24
  factual-error
• 7:25
  threshold,
• 7:26
  because
• 7:26
  step-by-step
• 7:27
  reasoning
• 7:27
  is
• 7:27
  structurally
• 7:28
  high-entropy.
• 7:29
  Static
• 7:29
  thresholds
• 7:30
  do
• 7:30
  not
• 7:30
  transfer
• 7:31
  between
• 7:31
  memory
• 7:31
  retrieval
• 7:32
  and
• 7:32
  computation
• 7:32
  .
• 7:35
  🛡️
• 7:35
  Code
• 7:36
  Quality
• 7:36
  in
• 7:36
  the
• 7:36
  Age
• 7:36
  of
• 7:36
  AI
• 7:37
  Onklaud
• 7:37
  5
• 7:38
  takes
• 7:38
  the
• 7:38
  ensemble
• 7:39
  approach
• 7:39
  to
• 7:39
  AI
• 7:39
  code
• 7:40
  generation
• 7:41
  and
• 7:41
  builds
• 7:41
  a
• 7:41
  full
• 7:41
  production
• 7:42
  pipeline
• 7:42
  around
• 7:42
  it.
• 7:43
  Three
• 7:43
  models
• 7:43
  from
• 7:44
  different
• 7:44
  providers
• 7:44
  —
• 7:45
  Kimi
• 7:45
  K2.7,
• 7:46
  GLM
• 7:47
  5.2,
• 7:48
  and
• 7:48
  DeepSeek
• 7:48
  V4
• 7:48
  Pro
• 7:49
  —
• 7:49
  review
• 7:49
  every
• 7:50
  coding
• 7:50
  decision
• 7:50
  through
• 7:51
  a
• 7:51
  structured
• 7:51
  council
• 7:51
  process
• 7:52
  with
• 7:52
  arbitration
• 7:53
  on
• 7:53
  disagreement.
• 7:54
  Four
• 7:54
  zero-cost
• 7:55
  infrastructure
• 7:55
  layers
• 7:55
  handle
• 7:56
  57%
• 7:56
  of
• 7:57
  tasks
• 7:57
  before
• 7:57
  any
• 7:58
  API
• 7:58
  call
• 7:58
  fires:
• 7:59
  a
• 7:59
  stdlib/native
• 8:00
  pattern
• 8:01
  matcher,
• 8:01
  an
• 8:02
  immune
• 8:02
  memory
• 8:02
  system
• 8:03
  storing
• 8:03
  19
• 8:03
  failure
• 8:04
  patterns,
• 8:04
  60–95%
• 8:06
  context
• 8:06
  compression
• 8:07
  for
• 8:07
  long
• 8:07
  sessions,
• 8:08
  and
• 8:08
  a
• 8:08
  quality
• 8:08
  gate
• 8:09
  enforcing
• 8:09
  a
• 8:09
  10/10
• 8:10
  threshold
• 8:10
  across
• 8:10
  seven
• 8:11
  dimensions.
• 8:11
  The
• 8:12
  claimed
• 8:12
  result:
• 8:13
  frontier-matching
• 8:13
  code
• 8:14
  quality
• 8:14
  at
• 8:14
  roughly
• 8:14
  1/100th
• 8:15
  the
• 8:15
  cost
• 8:16
  of
• 8:16
  a
• 8:16
  single-model
• 8:16
  approach,
• 8:17
  shipping
• 8:17
  fully
• 8:18
  open
• 8:18
  source
• 8:18
  under
• 8:18
  BSL
• 8:19
  1.1
• 8:20
  transitioning
• 8:20
  to
• 8:20
  MIT
• 8:21
  in
• 8:21
  2030
• 8:21
  .
• 8:22
  The
• 8:22
  verification
• 8:23
  pipeline
• 8:23
  is
• 8:23
  not
• 8:23
  new,
• 8:24
  but
• 8:24
  the
• 8:24
  parallel
• 8:24
  to
• 8:25
  a
• 8:25
  much
• 8:25
  older
• 8:25
  discipline
• 8:25
  is
• 8:26
  instructive.
• 8:27
  In
• 8:27
  1996,
• 8:28
  Fast
• 8:28
  Company
• 8:29
  profiled
• 8:29
  the
• 8:29
  Space
• 8:29
  Shuttle's
• 8:30
  onboard
• 8:30
  software
• 8:31
  group
• 8:31
  —
• 8:32
  260
• 8:32
  people
• 8:32
  writing
• 8:33
  a
• 8:33
  420,000-line
• 8:35
  program
• 8:35
  that
• 8:35
  had
• 8:35
  accumulated
• 8:36
  exactly
• 8:36
  17
• 8:37
  errors
• 8:37
  across
• 8:37
  its
• 8:37
  last
• 8:38
  eleven
• 8:38
  versions.
• 8:39
  Their
• 8:39
  process
• 8:39
  was
• 8:40
  obsessively
• 8:40
  process-driven:
• 8:41
  every
• 8:42
  change
• 8:42
  logged,
• 8:42
  every
• 8:43
  error
• 8:43
  root-caused,
• 8:44
  every
• 8:44
  fix
• 8:44
  verified
• 8:45
  against
• 8:45
  every
• 8:45
  prior
• 8:45
  fix.
• 8:46
  The
• 8:46
  software
• 8:47
  cost
• 8:47
  $35
• 8:47
  million
• 8:48
  a
• 8:48
  year.
• 8:49
  What
• 8:49
  made
• 8:50
  it
• 8:50
  work
• 8:50
  was
• 8:50
  not
• 8:50
  talent
• 8:51
  or
• 8:51
  tools
• 8:51
  but
• 8:51
  a
• 8:51
  culture
• 8:52
  where
• 8:52
  the
• 8:52
  process
• 8:52
  was
• 8:52
  the
• 8:52
  product
• 8:53
  and
• 8:53
  the
• 8:53
  error
• 8:53
  database
• 8:54
  was
• 8:54
  the
• 8:54
  most
• 8:54
  valuable
• 8:55
  artifact
• 8:55
  the
• 8:55
  team
• 8:56
  owned.
• 8:57
  The
• 8:57
  group
• 8:57
  tracked
• 8:58
  not
• 8:58
  just
• 8:58
  what
• 8:58
  broke
• 8:58
  but
• 8:58
  why
• 8:59
  each
• 8:59
  class
• 8:59
  of
• 8:59
  error
• 8:59
  had
• 9:00
  escaped
• 9:00
  detection,
• 9:01
  then
• 9:01
  fixed
• 9:01
  the
• 9:01
  detection
• 9:02
  rather
• 9:02
  than
• 9:03
  the
• 9:03
  symptom.
• 9:04
  Today's
• 9:04
  multi-model
• 9:05
  verification
• 9:05
  pipelines
• 9:06
  grope
• 9:06
  toward
• 9:07
  the
• 9:07
  same
• 9:07
  principle
• 9:08
  from
• 9:08
  the
• 9:08
  opposite
• 9:08
  direction:
• 9:09
  instead
• 9:10
  of
• 9:10
  perfecting
• 9:10
  a
• 9:10
  single
• 9:10
  codebase
• 9:11
  through
• 9:11
  institutional
• 9:12
  discipline,
• 9:13
  they
• 9:13
  approximate
• 9:13
  zero-defect
• 9:14
  output
• 9:14
  through
• 9:15
  architectural
• 9:15
  redundancy
• 9:16
  .
• 9:16
  Jacquard
• 9:17
  approaches
• 9:17
  the
• 9:17
  trust
• 9:17
  problem
• 9:18
  from
• 9:18
  the
• 9:18
  language
• 9:18
  level.
• 9:19
  Built
• 9:19
  explicitly
• 9:20
  for
• 9:20
  AI-written,
• 9:21
  human-reviewed
• 9:22
  code,
• 9:22
  it
• 9:23
  introduces
• 9:23
  algebraic
• 9:23
  effects
• 9:24
  for
• 9:24
  controlled
• 9:24
  side
• 9:25
  effects,
• 9:25
  capability
• 9:26
  grants
• 9:26
  for
• 9:26
  fine-grained
• 9:27
  permissions,
• 9:28
  and
• 9:28
  content-addressed
• 9:29
  definitions
• 9:29
  so
• 9:29
  that
• 9:29
  every
• 9:30
  function
• 9:30
  version
• 9:30
  is
• 9:31
  immutable
• 9:31
  and
• 9:31
  auditable.
• 9:33
  The
• 9:33
  design
• 9:33
  bet
• 9:33
  is
• 9:33
  that
• 9:34
  if
• 9:34
  the
• 9:34
  language
• 9:34
  itself
• 9:35
  makes
• 9:35
  unsafe
• 9:35
  patterns
• 9:36
  structurally
• 9:36
  unrepresentable
• 9:37
  —
• 9:38
  no
• 9:38
  unchecked
• 9:38
  side
• 9:39
  effects,
• 9:39
  no
• 9:39
  ambient
• 9:40
  authority,
• 9:40
  no
• 9:40
  mutable
• 9:41
  global
• 9:41
  state
• 9:41
  —
• 9:42
  then
• 9:42
  human
• 9:42
  review
• 9:43
  of
• 9:43
  AI-generated
• 9:44
  code
• 9:44
  can
• 9:44
  focus
• 9:45
  on
• 9:45
  whether
• 9:45
  the
• 9:45
  logic
• 9:45
  is
• 9:45
  correct
• 9:46
  rather
• 9:46
  than
• 9:46
  whether
• 9:47
  the
• 9:47
  implementation
• 9:47
  is
• 9:48
  safe
• 9:48
  .
• 9:51
  🏛️
• 9:51
  CMMC
• 9:51
  Reform:
• 9:52
  The
• 9:52
  Floor
• 9:52
  Shifted
• 9:53
  On
• 9:53
  July
• 9:53
  13,
• 9:54
  the
• 9:54
  Department
• 9:54
  of
• 9:55
  Defense
• 9:55
  suspended
• 9:55
  the
• 9:55
  November
• 9:56
  2026
• 9:57
  transition
• 9:57
  to
• 9:57
  mandatory
• 9:58
  third-party
• 9:58
  CMMC
• 9:59
  assessments
• 10:00
  and
• 10:00
  stood
• 10:00
  up
• 10:00
  a
• 10:00
  sixty-day
• 10:01
  reform
• 10:01
  task
• 10:01
  force.
• 10:02
  The
• 10:02
  stated
• 10:02
  reasoning
• 10:03
  —
• 10:03
  that
• 10:03
  the
• 10:03
  certification
• 10:04
  program
• 10:04
  was
• 10:04
  pricing
• 10:05
  small
• 10:05
  firms
• 10:05
  out
• 10:05
  of
• 10:06
  the
• 10:06
  defense
• 10:06
  industrial
• 10:07
  base
• 10:07
  —
• 10:07
  is
• 10:07
  true.
• 10:09
  But
• 10:09
  a
• 10:09
  former
• 10:09
  CISO
• 10:10
  who
• 10:10
  has
• 10:10
  built
• 10:10
  government-community
• 10:11
  enclaves
• 10:11
  from
• 10:11
  both
• 10:12
  sides
• 10:12
  of
• 10:12
  the
• 10:12
  data
• 10:12
  path
• 10:12
  offers
• 10:13
  a
• 10:13
  sharper
• 10:13
  reading:
• 10:14
  the
• 10:14
  security
• 10:15
  baseline
• 10:15
  did
• 10:15
  not
• 10:15
  change.
• 10:17
  NIST
• 10:17
  SP
• 10:17
  800-171
• 10:20
  Rev
• 10:20
  2
• 10:21
  is
• 10:21
  still
• 10:21
  the
• 10:21
  standard.
• 10:22
  DFARS
• 10:23
  is
• 10:23
  still
• 10:23
  in
• 10:23
  force.
• 10:24
  What
• 10:25
  changed
• 10:25
  is
• 10:25
  the
• 10:25
  floor
• 10:26
  underneath
• 10:26
  the
• 10:26
  companies
• 10:27
  that
• 10:27
  were
• 10:27
  never
• 10:27
  going
• 10:27
  to
• 10:27
  take
• 10:27
  the
• 10:28
  baseline
• 10:28
  seriously.
• 10:29
  When
• 10:30
  you
• 10:30
  suspend
• 10:30
  the
• 10:30
  third-party
• 10:31
  check
• 10:31
  but
• 10:31
  leave
• 10:31
  self-assessment
• 10:32
  and
• 10:33
  the
• 10:33
  legal
• 10:33
  attestation
• 10:34
  in
• 10:34
  place,
• 10:34
  you
• 10:35
  raise
• 10:35
  the
• 10:35
  stakes
• 10:35
  for
• 10:35
  honest
• 10:36
  contractors
• 10:36
  —
• 10:37
  who
• 10:37
  now
• 10:37
  sign
• 10:38
  continuous-compliance
• 10:39
  affirmations
• 10:39
  with
• 10:39
  more
• 10:39
  legal
• 10:40
  exposure
• 10:40
  and
• 10:40
  less
• 10:40
  external
• 10:41
  validation
• 10:41
  —
• 10:42
  and
• 10:42
  lower
• 10:43
  them
• 10:43
  for
• 10:43
  everyone
• 10:43
  else.
• 10:44
  The
• 10:44
  author's
• 10:45
  eleven
• 10:45
  proposed
• 10:45
  reforms
• 10:46
  target
• 10:46
  the
• 10:46
  apparatus
• 10:47
  rather
• 10:47
  than
• 10:47
  the
• 10:47
  controls:
• 10:49
  continuous
• 10:50
  evidence
• 10:50
  feeds
• 10:51
  from
• 10:51
  monitored
• 10:51
  commercial
• 10:52
  tooling,
• 10:53
  inter-rater
• 10:53
  consistency
• 10:54
  standards
• 10:54
  for
• 10:55
  assessors,
• 10:55
  independent
• 10:56
  appeals
• 10:56
  that
• 10:56
  do
• 10:56
  not
• 10:57
  route
• 10:57
  through
• 10:57
  the
• 10:57
  assessor
• 10:57
  who
• 10:58
  produced
• 10:58
  the
• 10:58
  finding,
• 10:59
  and
• 10:59
  risk-calibrated
• 11:00
  flowdown
• 11:00
  so
• 11:00
  that
• 11:01
  a
• 11:01
  machine
• 11:01
  shop
• 11:01
  turning
• 11:02
  a
• 11:02
  non-sensitive
• 11:02
  bracket
• 11:03
  stops
• 11:03
  receiving
• 11:04
  the
• 11:04
  same
• 11:04
  obligations
• 11:05
  as
• 11:05
  a
• 11:05
  firm
• 11:05
  holding
• 11:05
  weapons
• 11:06
  design
• 11:06
  data
• 11:06
  .
• 11:10
  🔄
• 11:10
  The
• 11:10
  AI
• 11:10
  Feedback
• 11:11
  Loop
• 11:11
  A
• 11:11
  new
• 11:11
  paper
• 11:11
  from
• 11:11
  METR
• 11:12
  and
• 11:12
  the
• 11:12
  Elasticity
• 11:13
  Institute
• 11:14
  models
• 11:14
  the
• 11:14
  economics
• 11:15
  of
• 11:15
  recursive
• 11:15
  self-improvement
• 11:16
  and
• 11:17
  concludes
• 11:17
  that
• 11:17
  the
• 11:17
  feedback
• 11:18
  loops
• 11:18
  are
• 11:18
  not
• 11:19
  yet
• 11:19
  strong
• 11:19
  enough
• 11:19
  to
• 11:19
  generate
• 11:20
  self-sustaining
• 11:20
  acceleration.
• 11:21
  The
• 11:22
  core
• 11:22
  question:
• 11:23
  for
• 11:23
  a
• 11:23
  one-unit
• 11:23
  increase
• 11:24
  in
• 11:24
  model
• 11:24
  capabilities,
• 11:25
  how
• 11:25
  much
• 11:26
  do
• 11:26
  the
• 11:26
  capabilities
• 11:26
  of
• 11:27
  the
• 11:27
  next
• 11:27
  generation
• 11:27
  increase?
• 11:29
  Their
• 11:29
  calibration
• 11:30
  puts
• 11:30
  the
• 11:30
  current
• 11:30
  AI
• 11:31
  R&D
• 11:31
  boost
• 11:31
  at
• 11:31
  roughly
• 11:32
  9%,
• 11:33
  well
• 11:33
  below
• 11:33
  the
• 11:33
  approximately
• 11:34
  15%
• 11:34
  threshold
• 11:35
  they
• 11:35
  estimate
• 11:35
  would
• 11:35
  be
• 11:36
  needed
• 11:36
  for
• 11:36
  the
• 11:36
  loop
• 11:36
  to
• 11:36
  become
• 11:37
  self-sustaining.
• 11:37
  The
• 11:37
  paper
• 11:38
  models
• 11:38
  the
• 11:38
  relationship
• 11:39
  as
• 11:39
  a
• 11:39
  directed
• 11:39
  graph
• 11:40
  of
• 11:40
  elasticities
• 11:40
  —
• 11:41
  net
• 11:41
  acceleration
• 11:42
  depends
• 11:42
  on
• 11:42
  the
• 11:42
  product
• 11:42
  of
• 11:42
  elasticities
• 11:43
  across
• 11:43
  each
• 11:43
  feedback
• 11:44
  loop
• 11:44
  —
• 11:45
  and
• 11:45
  the
• 11:45
  gap
• 11:45
  between
• 11:45
  current
• 11:46
  and
• 11:46
  critical
• 11:46
  remains
• 11:47
  significant
• 11:47
  enough
• 11:47
  that
• 11:47
  a
• 11:47
  gradual
• 11:48
  strengthening,
• 11:49
  not
• 11:49
  a
• 11:49
  sudden
• 11:49
  ignition,
• 11:50
  is
• 11:50
  the
• 11:50
  expected
• 11:51
  trajectory.
• 11:52
  The
• 11:52
  paper
• 11:52
  distinguishes
• 11:53
  between
• 11:53
  narrow
• 11:54
  capabilities
• 11:54
  —
• 11:55
  improving
• 11:55
  at
• 11:55
  AI
• 11:56
  R&D
• 11:56
  benchmarks
• 11:56
  —
• 11:57
  and
• 11:57
  broad
• 11:57
  capabilities
• 11:58
  —
• 11:59
  improving
• 11:59
  at
• 11:59
  economically
• 12:00
  valuable
• 12:00
  tasks
• 12:00
  —
• 12:01
  a
• 12:01
  distinction
• 12:02
  that
• 12:02
  matters
• 12:02
  because
• 12:02
  a
• 12:02
  system
• 12:03
  that
• 12:03
  optimizes
• 12:03
  itself
• 12:04
  may
• 12:04
  not
• 12:04
  proportionally
• 12:04
  improve
• 12:05
  at
• 12:05
  everything
• 12:05
  else
• 12:05
  .
• 12:06
  Arvind
• 12:06
  Narayanan's
• 12:07
  ICML
• 12:07
  2026
• 12:08
  keynote
• 12:08
  frames
• 12:09
  the
• 12:09
  broader
• 12:09
  context.
• 12:10
  His
• 12:10
  "AI
• 12:11
  as
• 12:11
  Normal
• 12:11
  Technology"
• 12:12
  argument
• 12:12
  holds
• 12:12
  that
• 12:12
  the
• 12:12
  effort
• 12:13
  is
• 12:13
  shifting
• 12:13
  from
• 12:14
  building
• 12:14
  models
• 12:14
  to
• 12:14
  evaluating
• 12:15
  what
• 12:15
  they
• 12:15
  produce
• 12:15
  —
• 12:16
  that
• 12:16
  AI
• 12:17
  is
• 12:17
  transformative
• 12:17
  but
• 12:17
  not
• 12:18
  superintelligent,
• 12:19
  and
• 12:19
  the
• 12:19
  hard
• 12:19
  problems
• 12:20
  are
• 12:20
  now
• 12:20
  measurement,
• 12:20
  integration,
• 12:21
  and
• 12:21
  organizational
• 12:22
  absorption
• 12:22
  rather
• 12:23
  than
• 12:23
  raw
• 12:23
  capability.
• 12:24
  If
• 12:25
  the
• 12:25
  binding
• 12:25
  constraint
• 12:26
  is
• 12:26
  evaluation
• 12:26
  rather
• 12:26
  than
• 12:27
  generation,
• 12:27
  faster
• 12:28
  models
• 12:28
  do
• 12:28
  not
• 12:29
  automatically
• 12:29
  produce
• 12:29
  faster
• 12:30
  progress
• 12:30
  .
• 12:31
  A
• 12:31
  practitioner's
• 12:32
  video
• 12:32
  essay
• 12:32
  on
• 12:33
  organizational
• 12:33
  transformation
• 12:34
  sharpens
• 12:34
  this
• 12:34
  into
• 12:35
  operational
• 12:35
  advice.
• 12:36
  The
• 12:36
  core
• 12:37
  claim:
• 12:38
  the
• 12:38
  distinguishing
• 12:38
  factor
• 12:39
  between
• 12:39
  companies
• 12:39
  like
• 12:39
  Anthropic
• 12:40
  that
• 12:40
  ship
• 12:40
  weekly
• 12:41
  and
• 12:41
  everyone
• 12:41
  else
• 12:41
  is
• 12:41
  not
• 12:42
  AI
• 12:42
  tooling
• 12:42
  but
• 12:42
  organizational
• 12:43
  wiring
• 12:43
  —
• 12:44
  fifteen
• 12:45
  rules
• 12:45
  for
• 12:45
  moving
• 12:45
  a
• 12:45
  company
• 12:46
  toward
• 12:46
  a
• 12:46
  faster
• 12:46
  shipping
• 12:47
  cadence
• 12:47
  that
• 12:47
  boil
• 12:47
  down
• 12:48
  to
• 12:48
  one
• 12:48
  principle:
• 12:49
  find
• 12:49
  the
• 12:49
  point
• 12:49
  of
• 12:49
  greatest
• 12:50
  leverage
• 12:50
  and
• 12:50
  move
• 12:50
  scarce
• 12:51
  human
• 12:51
  judgment
• 12:51
  to
• 12:51
  that
• 12:51
  point.
• 12:53
  The
• 12:53
  photography
• 12:53
  analogy
• 12:53
  lands:
• 12:54
  film's
• 12:55
  cost
• 12:55
  imposed
• 12:55
  discipline
• 12:56
  on
• 12:56
  how
• 12:56
  many
• 12:56
  shots
• 12:57
  you
• 12:57
  took,
• 12:57
  and
• 12:58
  digital
• 12:58
  removed
• 12:58
  that
• 12:59
  discipline
• 12:59
  without
• 12:59
  removing
• 13:00
  the
• 13:00
  need
• 13:00
  for
• 13:00
  curation.
• 13:01
  AI
• 13:02
  has
• 13:02
  done
• 13:02
  the
• 13:02
  same
• 13:02
  to
• 13:03
  all
• 13:03
  of
• 13:03
  work.
• 13:04
  The
• 13:04
  cost
• 13:04
  of
• 13:04
  another
• 13:04
  draft
• 13:05
  is
• 13:05
  collapsing
• 13:05
  to
• 13:05
  zero,
• 13:06
  but
• 13:06
  the
• 13:06
  choice
• 13:07
  of
• 13:07
  what
• 13:07
  deserves
• 13:07
  to
• 13:07
  exist
• 13:08
  has
• 13:08
  become
• 13:08
  harder,
• 13:09
  not
• 13:09
  easier
• 13:09
  .
• 13:12
  🔌
• 13:12
  Creative
• 13:12
  AI
• 13:13
  and
• 13:13
  Agent
• 13:13
  Tooling
• 13:14
  Apple's
• 13:14
  new
• 13:14
  SpeechAnalyzer
• 13:15
  API,
• 13:16
  shipping
• 13:16
  with
• 13:16
  iOS
• 13:16
  26
• 13:17
  and
• 13:17
  macOS
• 13:18
  26,
• 13:19
  replaces
• 13:19
  SFSpeechRecognizer
• 13:20
  with
• 13:20
  no
• 13:21
  published
• 13:21
  accuracy
• 13:21
  figures.
• 13:23
  An
• 13:23
  independent
• 13:23
  benchmark
• 13:24
  on
• 13:24
  M2
• 13:24
  Pro
• 13:24
  fills
• 13:24
  the
• 13:25
  gap:
• 13:25
  SpeechAnalyzer
• 13:26
  achieves
• 13:27
  2.12%
• 13:28
  word
• 13:28
  error
• 13:28
  rate
• 13:28
  on
• 13:28
  LibriSpeech
• 13:29
  test-clean,
• 13:30
  beating
• 13:30
  every
• 13:30
  Whisper
• 13:31
  model
• 13:31
  including
• 13:31
  Whisper
• 13:32
  Small,
• 13:32
  while
• 13:32
  running
• 13:33
  roughly
• 13:33
  three
• 13:33
  times
• 13:33
  faster.
• 13:34
  The
• 13:34
  old
• 13:35
  SFSpeechRecognizer
• 13:36
  came
• 13:36
  last
• 13:36
  —
• 13:37
  behind
• 13:37
  even
• 13:37
  Whisper
• 13:38
  Tiny,
• 13:38
  a
• 13:38
  40
• 13:39
  MB
• 13:39
  model.
• 13:40
  Apple
• 13:40
  published
• 13:41
  no
• 13:41
  accuracy
• 13:41
  figures
• 13:42
  for
• 13:42
  either
• 13:42
  new
• 13:42
  API,
• 13:43
  so
• 13:43
  every
• 13:44
  developer
• 13:44
  deciding
• 13:45
  whether
• 13:45
  to
• 13:45
  migrate
• 13:45
  has
• 13:45
  been
• 13:46
  guessing
• 13:46
  —
• 13:47
  until
• 13:47
  now.
• 13:48
  The
• 13:48
  migration
• 13:48
  case
• 13:49
  is
• 13:49
  closed
• 13:49
  .
• 13:50
  A
• 13:50
  CLI
• 13:50
  tool
• 13:51
  demonstrates
• 13:51
  Claude's
• 13:51
  vision
• 13:52
  capabilities
• 13:52
  in
• 13:52
  an
• 13:53
  unexpectedly
• 13:53
  practical
• 13:54
  niche:
• 13:54
  extracting
• 13:55
  guitar
• 13:55
  tablature
• 13:56
  from
• 13:56
  YouTube
• 13:56
  lesson
• 13:56
  videos.
• 13:58
  The
• 13:58
  pipeline
• 13:58
  downloads
• 13:59
  the
• 13:59
  video,
• 13:59
  samples
• 14:00
  frames,
• 14:00
  uses
• 14:01
  Claude
• 14:01
  vision
• 14:01
  to
• 14:01
  locate
• 14:02
  the
• 14:02
  tab
• 14:02
  region
• 14:02
  via
• 14:03
  labeled
• 14:03
  horizontal-band
• 14:04
  grids,
• 14:04
  deduplicates
• 14:05
  by
• 14:05
  bar
• 14:05
  number,
• 14:06
  and
• 14:06
  stitches
• 14:06
  the
• 14:06
  distinct
• 14:07
  lines
• 14:07
  into
• 14:07
  a
• 14:07
  PDF
• 14:07
  —
• 14:08
  no
• 14:09
  configuration
• 14:09
  needed
• 14:09
  .
• 14:10
  On
• 14:11
  the
• 14:11
  video
• 14:11
  generation
• 14:11
  side,
• 14:12
  a
• 14:12
  curated
• 14:13
  collection
• 14:13
  of
• 14:13
  Blender-to-Seedance
• 14:14
  workflows
• 14:14
  documents
• 14:15
  how
• 14:15
  3D
• 14:16
  scene
• 14:16
  authoring
• 14:17
  in
• 14:17
  Blender
• 14:17
  can
• 14:18
  feed
• 14:18
  into
• 14:18
  AI
• 14:18
  video
• 14:19
  generation
• 14:19
  for
• 14:19
  filmmaking.
• 14:21
  The
• 14:21
  approach
• 14:21
  gives
• 14:21
  creators
• 14:22
  precise
• 14:22
  spatial
• 14:23
  control
• 14:23
  over
• 14:23
  camera,
• 14:24
  lighting,
• 14:24
  and
• 14:25
  composition
• 14:25
  in
• 14:25
  Blender
• 14:26
  while
• 14:26
  letting
• 14:26
  Seedance
• 14:27
  handle
• 14:27
  photorealistic
• 14:28
  rendering
• 14:28
  and
• 14:28
  motion
• 14:29
  synthesis
• 14:29
  —
• 14:30
  bridging
• 14:30
  the
• 14:31
  gap
• 14:31
  between
• 14:31
  deterministic
• 14:32
  3D
• 14:32
  control
• 14:33
  and
• 14:33
  generative
• 14:33
  video
• 14:33
  output
• 14:34
  .
• 14:35
  WorldFoundry
• 14:35
  takes
• 14:35
  a
• 14:35
  more
• 14:36
  systematic
• 14:36
  approach,
• 14:37
  providing
• 14:37
  a
• 14:37
  unified
• 14:38
  framework
• 14:38
  for
• 14:38
  world
• 14:39
  model
• 14:39
  inference
• 14:40
  and
• 14:40
  evaluation
• 14:40
  —
• 14:41
  the
• 14:41
  standardized
• 14:42
  benchmarking
• 14:43
  infrastructure
• 14:43
  that
• 14:43
  video
• 14:44
  generation
• 14:44
  needs
• 14:45
  before
• 14:45
  it
• 14:45
  can
• 14:45
  move
• 14:45
  beyond
• 14:46
  demos
• 14:46
  into
• 14:46
  repeatable
• 14:47
  production
• 14:47
  .
• 14:48
  In
• 14:48
  the
• 14:49
  agent
• 14:49
  infrastructure
• 14:50
  layer,
• 14:50
  OpenConnector
• 14:51
  positions
• 14:51
  itself
• 14:52
  as
• 14:52
  an
• 14:52
  open-source
• 14:53
  alternative
• 14:53
  to
• 14:54
  Composio,
• 14:55
  offering
• 14:55
  a
• 14:55
  unified
• 14:56
  gateway
• 14:56
  to
• 14:56
  over
• 14:56
  1,000
• 14:57
  service
• 14:57
  providers
• 14:58
  with
• 14:58
  MCP-ready
• 14:59
  tool
• 14:59
  exposure
• 15:00
  —
• 15:01
  connect
• 15:01
  once,
• 15:02
  use
• 15:02
  everywhere
• 15:02
  .
• 15:03
  OakLab's
• 15:04
  research
• 15:04
  mission
• 15:05
  aims
• 15:05
  at
• 15:05
  a
• 15:05
  different
• 15:06
  problem
• 15:06
  entirely:
• 15:07
  building
• 15:08
  agent
• 15:08
  architectures
• 15:09
  that
• 15:09
  discover
• 15:09
  temporal
• 15:10
  abstractions
• 15:11
  grounded
• 15:11
  in
• 15:11
  experience,
• 15:12
  learning
• 15:13
  in
• 15:13
  real-time
• 15:14
  without
• 15:14
  storing
• 15:14
  or
• 15:15
  replaying
• 15:15
  data,
• 15:16
  using
• 15:16
  orders
• 15:17
  of
• 15:17
  magnitude
• 15:17
  less
• 15:18
  compute
• 15:18
  than
• 15:18
  existing
• 15:19
  methods.
• 15:20
  Where
• 15:20
  most
• 15:21
  agent
• 15:21
  work
• 15:21
  assumes
• 15:22
  foundation
• 15:22
  models
• 15:23
  as
• 15:23
  the
• 15:23
  reasoning
• 15:23
  substrate,
• 15:24
  OakLab
• 15:25
  is
• 15:25
  working
• 15:25
  on
• 15:25
  the
• 15:25
  substrate
• 15:26
  itself
• 15:26
  —
• 15:27
  algorithms
• 15:28
  that
• 15:28
  assign
• 15:28
  credit
• 15:29
  to
• 15:29
  parameters
• 15:29
  that
• 15:29
  generalize
• 15:30
  well
• 15:31
  rather
• 15:31
  than
• 15:31
  to
• 15:31
  all
• 15:31
  parameters,
• 15:33
  enabling
• 15:33
  learning
• 15:33
  directly
• 15:34
  from
• 15:34
  noisy
• 15:34
  experience
• 15:35
  instead
• 15:35
  of
• 15:35
  curated
• 15:36
  datasets
• 15:36
  .