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

Transcript

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

• 0:00
  A
• 0:00
  27-billion-parameter
• 0:01
  model
• 0:01
  now
• 0:02
  answers
• 0:02
  from
• 0:02
  a
• 0:02
  Chrome
• 0:03
  tab
• 0:03
  on
• 0:03
  a
• 0:03
  laptop
• 0:03
  with
• 0:03
  a
• 0:03
  6
• 0:04
  GB
• 0:04
  RTX
• 0:05
  3060,
• 0:06
  at
• 0:06
  25
• 0:06
  to
• 0:06
  30
• 0:07
  tokens
• 0:07
  per
• 0:07
  second,
• 0:08
  nothing
• 0:08
  installed
• 0:09
  and
• 0:09
  nothing
• 0:09
  leaving
• 0:10
  the
• 0:10
  machine.
• 0:11
  The
• 0:11
  engine
• 0:11
  is
• 0:11
  mentria.ai,
• 0:13
  a
• 0:13
  WebGPU/WGSL
• 0:15
  runtime
• 0:15
  written
• 0:16
  solo
• 0:16
  from
• 0:16
  scratch;
• 0:17
  the
• 0:17
  model
• 0:17
  is
• 0:18
  Bonsai-27B,
• 0:19
  the
• 0:19
  natively
• 0:19
  one-bit
• 0:20
  model
• 0:20
  Prism
• 0:20
  ML
• 0:20
  trained
• 0:21
  and
• 0:21
  released,
• 0:22
  repacked
• 0:22
  for
• 0:22
  the
• 0:22
  engine.
• 0:23
  Every
• 0:24
  weight
• 0:24
  is
• 0:24
  a
• 0:24
  sign
• 0:24
  bit
• 0:25
  with
• 0:25
  one
• 0:25
  scale
• 0:25
  per
• 0:25
  128
• 0:27
  weights,
• 0:27
  about
• 0:28
  1.14
• 0:28
  bits
• 0:29
  per
• 0:29
  parameter,
• 0:30
  so
• 0:30
  27B
• 0:31
  parameters
• 0:31
  fit
• 0:32
  in
• 0:32
  3.8
• 0:33
  GB
• 0:33
  of
• 0:33
  GPU
• 0:33
  memory.
• 0:34
  Two
• 0:35
  days
• 0:35
  earlier
• 0:35
  the
• 0:35
  same
• 0:36
  model
• 0:36
  decoded
• 0:36
  at
• 0:36
  15
• 0:37
  tokens
• 0:37
  per
• 0:37
  second
• 0:38
  on
• 0:38
  the
• 0:38
  same
• 0:38
  laptop,
• 0:39
  and
• 0:39
  the
• 0:39
  account
• 0:39
  of
• 0:40
  how
• 0:40
  it
• 0:40
  doubled
• 0:40
  is
• 0:40
  the
• 0:40
  interesting
• 0:41
  part
• 0:41
  .
• 0:42
  Decode
• 0:42
  is
• 0:42
  memory-bound:
• 0:44
  each
• 0:44
  token
• 0:44
  is
• 0:44
  804
• 0:45
  GPU
• 0:46
  dispatches,
• 0:46
  401
• 0:47
  of
• 0:47
  them
• 0:48
  the
• 0:48
  one-bit
• 0:48
  matrix-by-vector
• 0:49
  kernel
• 0:49
  streaming
• 0:50
  3.6
• 0:51
  GB
• 0:51
  of
• 0:51
  weights.
• 0:52
  The
• 0:52
  winning
• 0:53
  kernel
• 0:53
  was
• 0:53
  written
• 0:53
  for
• 0:53
  phones:
• 0:54
  four
• 0:55
  one-bit
• 0:55
  weights
• 0:55
  have
• 0:55
  only
• 0:56
  16
• 0:56
  possible
• 0:57
  partial
• 0:57
  products,
• 0:58
  so
• 0:58
  it
• 0:58
  computes
• 0:58
  all
• 0:59
  16
• 0:59
  into
• 0:59
  on-chip
• 1:00
  scratch
• 1:00
  and
• 1:01
  each
• 1:01
  row
• 1:01
  reads
• 1:01
  its
• 1:01
  answer
• 1:02
  from
• 1:02
  the
• 1:02
  table.
• 1:03
  Scratch
• 1:04
  memory
• 1:04
  has
• 1:04
  32
• 1:05
  banks,
• 1:05
  the
• 1:06
  addressing
• 1:06
  sent
• 1:06
  sixteen
• 1:07
  threads
• 1:07
  of
• 1:07
  every
• 1:08
  warp
• 1:08
  to
• 1:08
  the
• 1:08
  same
• 1:08
  bank,
• 1:09
  and
• 1:09
  the
• 1:09
  kernel
• 1:09
  ran
• 1:10
  at
• 1:10
  38%
• 1:11
  of
• 1:11
  bandwidth.
• 1:12
  One
• 1:12
  padding
• 1:13
  slot
• 1:13
  per
• 1:13
  row
• 1:13
  fixed
• 1:14
  it,
• 1:14
  the
• 1:14
  kernel's
• 1:14
  share
• 1:15
  of
• 1:15
  a
• 1:15
  token
• 1:15
  fell
• 1:15
  from
• 1:16
  26.5
• 1:17
  ms
• 1:17
  to
• 1:17
  about
• 1:18
  21,
• 1:18
  and
• 1:18
  raw
• 1:19
  decode
• 1:19
  hit
• 1:19
  32
• 1:20
  tokens
• 1:20
  per
• 1:20
  second,
• 1:21
  with
• 1:21
  every
• 1:21
  change
• 1:22
  verified
• 1:22
  byte-identical
• 1:23
  against
• 1:23
  the
• 1:23
  previous
• 1:24
  build.
• 1:25
  The
• 1:25
  caveats
• 1:25
  are
• 1:26
  equally
• 1:26
  concrete.
• 1:27
  Context
• 1:28
  is
• 1:28
  3,072
• 1:29
  tokens
• 1:29
  on
• 1:30
  the
• 1:30
  6
• 1:30
  GB
• 1:30
  card
• 1:30
  at
• 1:31
  128
• 1:32
  KiB
• 1:32
  per
• 1:32
  token
• 1:33
  with
• 1:33
  an
• 1:33
  unquantized
• 1:34
  KV
• 1:34
  cache.
• 1:35
  Commenters
• 1:36
  reported
• 1:36
  looping
• 1:37
  and
• 1:37
  a
• 1:37
  failed
• 1:37
  attempt
• 1:38
  at
• 1:38
  a
• 1:38
  script
• 1:38
  printing
• 1:38
  the
• 1:38
  first
• 1:39
  20
• 1:39
  primes.
• 1:40
  The
• 1:40
  runtime
• 1:41
  is
• 1:41
  no
• 1:41
  longer
• 1:41
  the
• 1:41
  bottleneck
• 1:42
  for
• 1:42
  one-bit
• 1:42
  models
• 1:43
  on
• 1:43
  consumer
• 1:43
  silicon;
• 1:44
  whether
• 1:45
  a
• 1:45
  1.14-bit
• 1:46
  27B
• 1:46
  is
• 1:47
  smart
• 1:47
  enough
• 1:47
  to
• 1:47
  be
• 1:47
  worth
• 1:47
  running
• 1:48
  remains
• 1:48
  the
• 1:48
  question
• 1:49
  it
• 1:49
  has
• 1:49
  been
• 1:49
  all
• 1:49
  year.
• 1:50
  At
• 1:51
  the
• 1:51
  other
• 1:51
  end
• 1:51
  of
• 1:51
  the
• 1:51
  spectrum,
• 1:52
  a
• 1:52
  bioinformatics
• 1:53
  lab
• 1:53
  machine
• 1:54
  runs
• 1:54
  DeepSeek
• 1:54
  V4.1
• 1:57
  Flash
• 1:57
  on
• 1:57
  CPU
• 1:58
  alone,
• 1:58
  at
• 1:59
  roughly
• 1:59
  30
• 1:59
  tokens
• 2:00
  per
• 2:00
  second
• 2:00
  prompt
• 2:01
  processing
• 2:01
  and
• 2:02
  6
• 2:02
  generating
• 2:02
  on
• 2:02
  a
• 2:02
  Xeon,
• 2:03
  with
• 2:04
  the
• 2:04
  n-gram
• 2:04
  lookup
• 2:04
  table
• 2:05
  offloaded
• 2:05
  to
• 2:05
  half
• 2:06
  the
• 2:06
  threads.
• 2:07
  The
• 2:07
  author
• 2:08
  is
• 2:08
  candid
• 2:08
  that
• 2:08
  the
• 2:08
  code
• 2:09
  was
• 2:09
  "sloppily
• 2:09
  vibecoded
• 2:10
  by
• 2:10
  Opus
• 2:11
  5.0,
• 2:12
  unreviewed
• 2:13
  because
• 2:13
  frankly
• 2:13
  I
• 2:13
  lack
• 2:14
  the
• 2:14
  skill
• 2:14
  to
• 2:14
  verify,"
• 2:15
  and
• 2:15
  that
• 2:15
  a
• 2:15
  watcher
• 2:16
  kills
• 2:16
  the
• 2:16
  model
• 2:16
  in
• 2:17
  15
• 2:17
  seconds
• 2:18
  if
• 2:18
  someone
• 2:18
  needs
• 2:18
  the
• 2:18
  box
• 2:19
  for
• 2:19
  a
• 2:19
  genome
• 2:19
  assembly.
• 2:21
  One
• 2:21
  commenter
• 2:21
  noted
• 2:22
  that
• 2:22
  a
• 2:22
  15-second
• 2:23
  kill
• 2:23
  script
• 2:23
  is
• 2:24
  the
• 2:24
  single
• 2:24
  point
• 2:24
  of
• 2:24
  failure
• 2:25
  between
• 2:25
  an
• 2:25
  unreviewed
• 2:26
  agent
• 2:26
  loop
• 2:26
  and
• 2:27
  someone
• 2:27
  else's
• 2:27
  research
• 2:28
  compute
• 2:28
  .
• 2:29
  A
• 2:29
  more
• 2:29
  conventional
• 2:30
  rig,
• 2:30
  two
• 2:31
  RTX
• 2:31
  3090s
• 2:32
  on
• 2:32
  a
• 2:32
  Zen
• 2:33
  1
• 2:33
  EPYC
• 2:33
  with
• 2:33
  128
• 2:35
  GB
• 2:35
  of
• 2:35
  DDR4
• 2:36
  across
• 2:36
  eight
• 2:36
  channels,
• 2:37
  runs
• 2:38
  Qwen3-Flash-Next
• 2:39
  (177B
• 2:40
  total,
• 2:41
  about
• 2:41
  6B
• 2:41
  active,
• 2:42
  IQ4
• 2:43
  XS)
• 2:43
  at
• 2:43
  38
• 2:44
  tokens
• 2:44
  per
• 2:44
  second
• 2:44
  single-stream
• 2:45
  with
• 2:45
  experts
• 2:45
  in
• 2:46
  system
• 2:46
  RAM,
• 2:47
  collapsing
• 2:47
  to
• 2:47
  about
• 2:47
  4
• 2:48
  each
• 2:48
  with
• 2:48
  two
• 2:48
  parallel
• 2:49
  requests.
• 2:50
  The
• 2:50
  owner
• 2:51
  has
• 2:51
  $800
• 2:52
  for
• 2:52
  a
• 2:52
  third
• 2:52
  3090
• 2:53
  or
• 2:53
  a
• 2:53
  Rome
• 2:54
  CPU;
• 2:55
  the
• 2:55
  thread
• 2:55
  said
• 2:55
  GPU,
• 2:56
  because
• 2:57
  experts
• 2:57
  in
• 2:57
  RAM
• 2:58
  are
• 2:58
  what
• 2:58
  kill
• 2:58
  throughput
• 2:59
  and
• 2:59
  "GPU
• 3:00
  is
• 3:00
  the
• 3:00
  only
• 3:00
  sensible
• 3:01
  path"
• 3:01
  for
• 3:01
  multi-agent
• 3:02
  use
• 3:02
  .
• 3:03
  For
• 3:03
  single-binary
• 3:04
  users,
• 3:04
  KoboldCpp
• 3:05
  v1.121
• 3:08
  adds
• 3:09
  MiniMax
• 3:09
  H3
• 3:10
  video
• 3:10
  generation
• 3:10
  with
• 3:10
  audio
• 3:11
  clips
• 3:11
  and
• 3:11
  reference
• 3:12
  images,
• 3:12
  "xhigh"
• 3:13
  reasoning
• 3:13
  effort,
• 3:14
  tool-calling
• 3:15
  fixes
• 3:15
  for
• 3:15
  Kimi
• 3:16
  and
• 3:16
  DeepSeek
• 3:16
  V4
• 3:16
  Flash,
• 3:17
  and
• 3:17
  a
• 3:17
  streaming
• 3:17
  race-condition
• 3:18
  fix
• 3:18
  .
• 3:21
  Long
• 3:21
  agent
• 3:22
  sessions
• 3:22
  die
• 3:22
  of
• 3:22
  context
• 3:23
  exhaustion,
• 3:24
  and
• 3:24
  the
• 3:24
  standard
• 3:24
  remedy,
• 3:25
  stop,
• 3:26
  summarize,
• 3:27
  and
• 3:27
  re-prefill,
• 3:28
  is
• 3:28
  expensive
• 3:29
  and
• 3:29
  disruptive.
• 3:30
  Spomin
• 3:30
  is
• 3:31
  an
• 3:31
  experimental
• 3:31
  router
• 3:32
  that
• 3:32
  performs
• 3:33
  compaction
• 3:33
  live
• 3:34
  inside
• 3:34
  the
• 3:34
  KV
• 3:34
  cache.
• 3:36
  It
• 3:36
  keeps
• 3:36
  the
• 3:36
  transcript
• 3:37
  in
• 3:37
  chunks
• 3:37
  and
• 3:37
  hands
• 3:37
  them
• 3:38
  to
• 3:38
  a
• 3:38
  separate
• 3:38
  summarizer
• 3:39
  worker
• 3:39
  on
• 3:39
  spare
• 3:39
  hardware
• 3:40
  or
• 3:40
  between
• 3:40
  the
• 3:40
  main
• 3:41
  model's
• 3:41
  generations.
• 3:42
  When
• 3:42
  space
• 3:43
  runs
• 3:43
  low
• 3:43
  it
• 3:43
  replaces
• 3:44
  eligible
• 3:44
  chunks
• 3:44
  with
• 3:45
  summaries
• 3:45
  directly
• 3:45
  in
• 3:45
  the
• 3:45
  cache
• 3:46
  and
• 3:46
  rewrites
• 3:47
  the
• 3:47
  RoPE
• 3:47
  positions
• 3:47
  of
• 3:47
  everything
• 3:48
  after
• 3:48
  them,
• 3:49
  via
• 3:49
  a
• 3:49
  llama.cpp
• 3:51
  fork
• 3:51
  with
• 3:52
  surgical
• 3:52
  cache
• 3:52
  edits.
• 3:53
  Only
• 3:53
  the
• 3:54
  summary
• 3:54
  is
• 3:54
  processed;
• 3:55
  the
• 3:55
  retained
• 3:55
  suffix
• 3:56
  is
• 3:56
  never
• 3:56
  re-prefilled.
• 3:57
  At
• 3:57
  roughly
• 3:58
  10:1
• 3:58
  compression,
• 3:59
  400k
• 3:59
  source
• 4:00
  tokens
• 4:00
  become
• 4:00
  40k,
• 4:01
  putting
• 4:01
  over
• 4:02
  500k
• 4:02
  tokens
• 4:03
  of
• 4:03
  material
• 4:03
  inside
• 4:03
  180k
• 4:04
  resident
• 4:04
  .
• 4:05
  The
• 4:05
  author's
• 4:06
  benchmarks
• 4:06
  are
• 4:06
  exploratory
• 4:07
  and
• 4:07
  mixed.
• 4:08
  A
• 4:08
  larger
• 4:08
  summary
• 4:08
  worker
• 4:09
  raised
• 4:09
  a
• 4:09
  research
• 4:09
  score
• 4:10
  from
• 4:10
  31
• 4:10
  to
• 4:10
  37
• 4:11
  out
• 4:11
  of
• 4:11
  100,
• 4:12
  still
• 4:12
  well
• 4:12
  below
• 4:13
  57
• 4:13
  unmanaged,
• 4:14
  with
• 4:14
  75.2%
• 4:16
  less
• 4:16
  primary
• 4:16
  input.
• 4:17
  A
• 4:17
  managed
• 4:17
  Tetris
• 4:18
  build
• 4:18
  nearly
• 4:18
  matched
• 4:18
  unmanaged
• 4:19
  (16/16
• 4:20
  plus
• 4:20
  11/12
• 4:21
  checks
• 4:21
  versus
• 4:21
  16/16
• 4:22
  plus
• 4:23
  12/12)
• 4:23
  in
• 4:23
  69.8
• 4:24
  minutes
• 4:24
  instead
• 4:25
  of
• 4:25
  149.6
• 4:26
  with
• 4:26
  22.6%
• 4:27
  fewer
• 4:28
  tokens.
• 4:28
  The
• 4:29
  author
• 4:29
  declines
• 4:29
  to
• 4:29
  attribute
• 4:30
  gains
• 4:30
  to
• 4:30
  any
• 4:30
  single
• 4:31
  factor.
• 4:32
  Only
• 4:32
  Qwen
• 4:32
  3.8
• 4:33
  27B
• 4:33
  is
• 4:34
  tested
• 4:34
  as
• 4:34
  the
• 4:34
  main
• 4:34
  model,
• 4:35
  with
• 4:35
  a
• 4:35
  fine-tuned
• 4:36
  Qwen2.5
• 4:37
  3B
• 4:39
  as
• 4:39
  worker,
• 4:39
  and
• 4:39
  Qwen
• 4:39
  4
• 4:40
  is
• 4:40
  unsupported.
• 4:41
  A
• 4:41
  related
• 4:41
  migration
• 4:42
  problem
• 4:42
  got
• 4:42
  a
• 4:42
  rougher
• 4:42
  reception.
• 4:43
  embedflow
• 4:44
  proposes
• 4:44
  swapping
• 4:45
  embedding
• 4:45
  models
• 4:45
  without
• 4:46
  re-embedding,
• 4:47
  motivated
• 4:47
  by
• 4:47
  the
• 4:48
  arithmetic
• 4:48
  that
• 4:48
  re-embedding
• 4:49
  one
• 4:49
  billion
• 4:49
  vectors
• 4:50
  with
• 4:50
  an
• 4:50
  8B
• 4:50
  Qwen
• 4:51
  embedder
• 4:51
  at
• 4:51
  106
• 4:52
  documents
• 4:52
  per
• 4:52
  second
• 4:53
  on
• 4:53
  an
• 4:53
  H100
• 4:54
  takes
• 4:54
  about
• 4:54
  108
• 4:55
  days.
• 4:56
  The
• 4:56
  method
• 4:56
  retrieves
• 4:57
  K
• 4:57
  candidates
• 4:57
  from
• 4:57
  the
• 4:57
  old
• 4:58
  index,
• 4:58
  reranks
• 4:58
  with
• 4:59
  the
• 4:59
  new
• 4:59
  model,
• 5:00
  and
• 5:00
  picks
• 5:00
  K
• 5:00
  large
• 5:00
  enough
• 5:00
  to
• 5:01
  match
• 5:01
  native
• 5:01
  quality;
• 5:02
  across
• 5:02
  63
• 5:03
  migrations
• 5:03
  on
• 5:04
  up
• 5:04
  to
• 5:04
  a
• 5:04
  million
• 5:04
  documents,
• 5:05
  Qwen
• 5:05
  4B
• 5:06
  to
• 5:06
  8B
• 5:06
  matched
• 5:07
  native
• 5:07
  at
• 5:07
  K=50
• 5:08
  .
• 5:09
  Commenters
• 5:09
  were
• 5:09
  mostly
• 5:10
  right:
• 5:10
  this
• 5:11
  is
• 5:11
  a
• 5:11
  rerank
• 5:11
  pipeline,
• 5:12
  not
• 5:12
  a
• 5:12
  migration,
• 5:13
  since
• 5:13
  both
• 5:13
  models
• 5:14
  stay
• 5:14
  loaded
• 5:14
  and
• 5:14
  the
• 5:14
  old
• 5:15
  model's
• 5:15
  recall
• 5:15
  at
• 5:15
  K
• 5:16
  is
• 5:16
  a
• 5:16
  hard
• 5:16
  ceiling.
• 5:17
  Anything
• 5:18
  buried
• 5:18
  below
• 5:18
  rank
• 5:18
  K
• 5:19
  is
• 5:19
  invisible
• 5:19
  forever,
• 5:20
  K
• 5:20
  changes
• 5:21
  per
• 5:21
  migration,
• 5:22
  and
• 5:22
  a
• 5:22
  same-family
• 5:22
  upgrade
• 5:23
  says
• 5:23
  nothing
• 5:23
  about
• 5:23
  a
• 5:23
  cross-family
• 5:24
  move.
• 5:25
  On
• 5:25
  persistent
• 5:25
  memory,
• 5:26
  Nemoryn
• 5:26
  is
• 5:27
  a
• 5:27
  self-hosted
• 5:27
  memory
• 5:28
  server
• 5:28
  that
• 5:28
  Open
• 5:28
  WebUI
• 5:29
  talks
• 5:29
  to
• 5:29
  as
• 5:29
  an
• 5:29
  ordinary
• 5:30
  OpenAI-compatible
• 5:31
  connection
• 5:31
  rather
• 5:32
  than
• 5:32
  a
• 5:32
  Function,
• 5:32
  Tool,
• 5:33
  or
• 5:33
  Pipe.
• 5:34
  It
• 5:34
  sits
• 5:34
  between
• 5:35
  client
• 5:35
  and
• 5:35
  model,
• 5:36
  retrieves
• 5:36
  memories
• 5:37
  from
• 5:37
  prior
• 5:37
  conversations
• 5:38
  into
• 5:38
  context,
• 5:39
  and
• 5:39
  stores
• 5:39
  or
• 5:39
  updates
• 5:40
  them
• 5:40
  afterward.
• 5:41
  The
• 5:41
  README
• 5:44
  explains
• 5:44
  how
• 5:44
  to
• 5:44
  run
• 5:44
  it
• 5:45
  better
• 5:45
  than
• 5:45
  what
• 5:45
  it
• 5:45
  is;
• 5:46
  multi-user
• 5:46
  support
• 5:47
  is
• 5:47
  an
• 5:47
  open
• 5:47
  question
• 5:47
  .
• 5:50
  Can
• 5:50
  a
• 5:50
  coding
• 5:51
  agent
• 5:51
  post-train
• 5:51
  a
• 5:52
  language
• 5:52
  model
• 5:52
  end
• 5:52
  to
• 5:52
  end?
• 5:53
  A
• 5:53
  group
• 5:53
  at
• 5:53
  the
• 5:53
  University
• 5:54
  of
• 5:54
  Electronic
• 5:54
  Science
• 5:55
  and
• 5:55
  Technology
• 5:55
  of
• 5:55
  China
• 5:56
  analyzed
• 5:56
  1,338
• 5:58
  public
• 5:58
  PostTrainBench
• 5:59
  trajectories,
• 6:00
  in
• 6:00
  which
• 6:00
  agents
• 6:00
  (Claude
• 6:00
  Code,
• 6:01
  Codex
• 6:01
  CLI,
• 6:01
  and
• 6:02
  others)
• 6:02
  each
• 6:02
  get
• 6:02
  ten
• 6:02
  hours
• 6:03
  on
• 6:03
  one
• 6:03
  H100
• 6:04
  to
• 6:04
  improve
• 6:04
  a
• 6:04
  small
• 6:05
  base
• 6:05
  model
• 6:05
  on
• 6:05
  one
• 6:05
  of
• 6:05
  seven
• 6:06
  benchmarks.
• 6:07
  Agents
• 6:07
  are
• 6:07
  competent
• 6:07
  executors
• 6:08
  and
• 6:08
  improve
• 6:08
  every
• 6:09
  base
• 6:09
  model,
• 6:09
  but
• 6:09
  strategy
• 6:10
  is
• 6:10
  locked
• 6:10
  in
• 6:11
  before
• 6:11
  any
• 6:11
  code
• 6:11
  runs
• 6:12
  and
• 6:12
  tracks
• 6:12
  the
• 6:12
  agent
• 6:13
  rather
• 6:13
  than
• 6:13
  the
• 6:13
  task.
• 6:14
  Claude
• 6:14
  Code
• 6:14
  anchors
• 6:15
  on
• 6:15
  full
• 6:15
  SFT
• 6:15
  and
• 6:15
  Codex
• 6:16
  CLI
• 6:16
  on
• 6:16
  PEFT
• 6:16
  across
• 6:17
  all
• 6:17
  28
• 6:17
  benchmark-model
• 6:18
  cells.
• 6:18
  Just
• 6:19
  16
• 6:19
  trajectories
• 6:19
  ever
• 6:20
  switched
• 6:20
  objective,
• 6:20
  and
• 6:21
  15
• 6:21
  of
• 6:21
  their
• 6:21
  35
• 6:22
  transitions
• 6:22
  went
• 6:22
  back
• 6:22
  to
• 6:22
  SFT
• 6:23
  .
• 6:23
  The
• 6:23
  interventions
• 6:24
  are
• 6:24
  the
• 6:24
  damning
• 6:24
  part.
• 6:25
  Giving
• 6:25
  Claude
• 6:26
  Code
• 6:26
  with
• 6:26
  Opus
• 6:26
  4.6
• 6:27
  an
• 6:27
  experiment
• 6:27
  journal,
• 6:28
  a
• 6:28
  skill
• 6:28
  library
• 6:29
  distilled
• 6:29
  from
• 6:29
  908
• 6:30
  documents,
• 6:30
  and
• 6:31
  an
• 6:31
  evaluator
• 6:31
  agent
• 6:31
  improved
• 6:32
  scores
• 6:32
  on
• 6:32
  GSM8K,
• 6:33
  HumanEval,
• 6:34
  and
• 6:34
  AIME
• 6:34
  2025,
• 6:35
  but
• 6:35
  strategy
• 6:36
  stayed
• 6:36
  frozen.
• 6:37
  On
• 6:37
  HumanEval
• 6:37
  the
• 6:37
  evaluator
• 6:38
  recommended
• 6:38
  RL
• 6:38
  in
• 6:39
  seven
• 6:39
  of
• 6:39
  nine
• 6:39
  cycles;
• 6:40
  the
• 6:40
  agent
• 6:40
  wrote
• 6:41
  GRPO
• 6:41
  scripts,
• 6:42
  journaled
• 6:42
  "SFT
• 6:43
  plateau
• 6:43
  confirmed
• 6:44
  at
• 6:44
  102–103"
• 6:46
  of
• 6:46
  164,
• 6:47
  then
• 6:47
  launched
• 6:48
  14
• 6:48
  more
• 6:48
  SFT
• 6:49
  variants
• 6:49
  and
• 6:49
  zero
• 6:50
  RL
• 6:50
  runs.
• 6:51
  It
• 6:51
  adopted
• 6:52
  every
• 6:52
  execution-level
• 6:53
  suggestion
• 6:53
  and
• 6:54
  none
• 6:54
  of
• 6:54
  the
• 6:54
  strategy-level
• 6:54
  ones.
• 6:55
  Binding
• 6:56
  human
• 6:56
  plan
• 6:56
  review
• 6:56
  on
• 6:56
  AIME
• 6:57
  redirected
• 6:57
  it
• 6:57
  from
• 6:57
  a
• 6:58
  10K–30K-example
• 6:59
  SFT
• 6:59
  plan
• 6:59
  to
• 6:59
  GRPO
• 7:00
  and
• 7:00
  produced
• 7:01
  the
• 7:01
  best
• 7:01
  pass@8
• 7:02
  of
• 7:02
  any
• 7:02
  setting,
• 7:02
  within
• 7:03
  AIME's
• 7:03
  noise.
• 7:04
  The
• 7:04
  authors'
• 7:04
  conclusion:
• 7:05
  what
• 7:05
  is
• 7:05
  missing
• 7:06
  "is
• 7:06
  not
• 7:06
  a
• 7:06
  resource
• 7:07
  but
• 7:07
  a
• 7:07
  mechanism,"
• 7:08
  the
• 7:08
  ability
• 7:08
  to
• 7:08
  reopen
• 7:08
  a
• 7:08
  strategic
• 7:09
  choice
• 7:09
  mid-execution,
• 7:10
  and
• 7:10
  agents
• 7:11
  "do
• 7:11
  not"
• 7:11
  revise
• 7:12
  rather
• 7:12
  than
• 7:13
  "cannot."
• 7:14
  A
• 7:14
  Mila
• 7:14
  paper
• 7:14
  points
• 7:14
  at
• 7:14
  one
• 7:15
  source
• 7:15
  of
• 7:15
  such
• 7:15
  lock-in.
• 7:16
  Semantic
• 7:16
  Bandits
• 7:17
  runs
• 7:17
  LLMs
• 7:18
  as
• 7:18
  in-context
• 7:18
  reinforcement
• 7:19
  learners
• 7:19
  on
• 7:19
  a
• 7:19
  three-arm
• 7:20
  Gaussian
• 7:20
  bandit
• 7:21
  and
• 7:21
  changes
• 7:21
  nothing
• 7:21
  but
• 7:21
  the
• 7:22
  arm
• 7:22
  labels.
• 7:23
  With
• 7:23
  random
• 7:23
  alphanumeric
• 7:24
  labels,
• 7:24
  OLMo-3.1-32B,
• 7:26
  Qwen3-32B,
• 7:27
  and
• 7:28
  Gemini
• 7:28
  3.1
• 7:28
  Flash
• 7:29
  Lite
• 7:29
  explored
• 7:29
  roughly
• 7:30
  like
• 7:30
  a
• 7:30
  UCB1
• 7:30
  baseline.
• 7:31
  Give
• 7:31
  the
• 7:32
  arms
• 7:32
  sentiment
• 7:32
  words,
• 7:33
  ordinal
• 7:33
  ranks,
• 7:33
  or
• 7:34
  domain
• 7:34
  names
• 7:34
  whose
• 7:35
  value
• 7:35
  pretraining
• 7:35
  can
• 7:35
  guess,
• 7:36
  and
• 7:36
  the
• 7:36
  models
• 7:36
  tilt
• 7:37
  toward
• 7:37
  exploitation:
• 7:38
  with
• 7:38
  the
• 7:38
  favorable
• 7:38
  label
• 7:39
  on
• 7:39
  the
• 7:39
  best
• 7:39
  arm
• 7:39
  they
• 7:39
  beat
• 7:40
  the
• 7:40
  symbolic
• 7:40
  baseline,
• 7:41
  and
• 7:41
  with
• 7:41
  labels
• 7:41
  swapped
• 7:42
  performance
• 7:42
  is,
• 7:43
  in
• 7:43
  the
• 7:43
  authors'
• 7:43
  word,
• 7:44
  dismal.
• 7:45
  Changing
• 7:45
  only
• 7:45
  the
• 7:45
  text
• 7:46
  on
• 7:46
  the
• 7:46
  arms
• 7:46
  moves
• 7:46
  an
• 7:47
  LLM
• 7:47
  from
• 7:47
  near-optimal
• 7:48
  to
• 7:48
  near-worst-case,
• 7:49
  and
• 7:49
  the
• 7:49
  authors
• 7:49
  note
• 7:50
  the
• 7:50
  lever
• 7:50
  "could
• 7:50
  also
• 7:51
  be
• 7:51
  used
• 7:51
  adversarially"
• 7:51
  .
• 7:52
  Anyone
• 7:53
  building
• 7:53
  a
• 7:53
  fraud-review
• 7:54
  or
• 7:54
  triage
• 7:54
  agent
• 7:55
  should
• 7:55
  read
• 7:55
  that
• 7:55
  as
• 7:55
  a
• 7:55
  warning
• 7:56
  about
• 7:56
  the
• 7:56
  field
• 7:56
  names
• 7:56
  in
• 7:57
  their
• 7:57
  tool
• 7:57
  schemas.
• 7:58
  IBM
• 7:58
  Research
• 7:59
  adds
• 7:59
  the
• 7:59
  measurement
• 7:59
  problem.
• 8:00
  A
• 8:00
  ReAct
• 8:01
  agent
• 8:01
  on
• 8:01
  GPT-4.1
• 8:02
  scores
• 8:02
  77.4%
• 8:04
  on
• 8:04
  AppWorld
• 8:04
  on
• 8:05
  average
• 8:05
  yet
• 8:05
  succeeds
• 8:06
  on
• 8:06
  all
• 8:06
  five
• 8:06
  repeated
• 8:07
  runs
• 8:07
  for
• 8:07
  only
• 8:07
  53.0%
• 8:08
  of
• 8:09
  tasks,
• 8:09
  at
• 8:09
  temperature
• 8:10
  zero,
• 8:10
  so
• 8:11
  none
• 8:11
  of
• 8:11
  it
• 8:11
  is
• 8:11
  sampling
• 8:11
  noise.
• 8:12
  Mean@k
• 8:13
  is
• 8:13
  the
• 8:13
  usual
• 8:13
  leaderboard
• 8:14
  average;
• 8:15
  Pass^k
• 8:16
  is
• 8:16
  the
• 8:16
  fraction
• 8:17
  of
• 8:17
  tasks
• 8:17
  that
• 8:17
  succeed
• 8:18
  on
• 8:18
  every
• 8:18
  one
• 8:18
  of
• 8:18
  k
• 8:19
  runs;
• 8:19
  the
• 8:20
  gap
• 8:20
  between
• 8:20
  them
• 8:20
  reaches
• 8:21
  30
• 8:21
  points
• 8:21
  on
• 8:21
  hard
• 8:21
  tasks.
• 8:22
  Their
• 8:23
  fix,
• 8:23
  in
• 8:23
  the
• 8:23
  ALTK-Evolve
• 8:26
  library,
• 8:26
  replays
• 8:27
  each
• 8:27
  recorded
• 8:27
  decision
• 8:28
  offline
• 8:28
  with
• 8:28
  five
• 8:29
  resamples,
• 8:29
  flags
• 8:30
  unstable
• 8:30
  steps,
• 8:31
  and
• 8:31
  turns
• 8:31
  each
• 8:32
  into
• 8:32
  an
• 8:32
  inference-time
• 8:32
  guideline.
• 8:33
  On
• 8:34
  168
• 8:35
  AppWorld
• 8:35
  tasks
• 8:35
  Pass^5
• 8:37
  rose
• 8:37
  from
• 8:37
  53.0%
• 8:38
  to
• 8:39
  69.0%
• 8:40
  while
• 8:40
  Mean@5
• 8:40
  rose
• 8:41
  from
• 8:41
  77.4%
• 8:42
  to
• 8:42
  81.0%.
• 8:44
  The
• 8:44
  recommendation:
• 8:45
  report
• 8:46
  Pass^k
• 8:47
  next
• 8:47
  to
• 8:47
  Mean@k,
• 8:48
  and
• 8:48
  stop
• 8:48
  reaching
• 8:48
  for
• 8:49
  a
• 8:49
  bigger
• 8:49
  model
• 8:49
  first
• 8:49
  .
• 8:52
  A
• 8:52
  developer
• 8:53
  with
• 8:53
  nearly
• 8:53
  thirty
• 8:53
  years
• 8:54
  of
• 8:54
  experience
• 8:54
  posted
• 8:55
  the
• 8:55
  contrarian
• 8:55
  case:
• 8:56
  the
• 8:56
  "give
• 8:56
  it
• 8:56
  a
• 8:57
  prompt,
• 8:57
  go
• 8:57
  get
• 8:58
  coffee"
• 8:58
  workflow
• 8:58
  is
• 8:58
  fine
• 8:59
  for
• 8:59
  greenfield
• 8:59
  projects,
• 9:00
  but
• 9:00
  on
• 9:00
  a
• 9:00
  mature
• 9:01
  codebase
• 9:01
  it
• 9:01
  has
• 9:01
  never
• 9:01
  once
• 9:02
  returned
• 9:02
  a
• 9:02
  good
• 9:02
  edit.
• 9:03
  The
• 9:03
  agent
• 9:03
  loops
• 9:04
  eight
• 9:04
  times,
• 9:04
  burns
• 9:05
  100k
• 9:05
  tokens
• 9:06
  talking
• 9:06
  to
• 9:06
  itself,
• 9:07
  hallucinates
• 9:07
  a
• 9:07
  dependency,
• 9:08
  and
• 9:08
  produces
• 9:09
  code
• 9:09
  that
• 9:09
  passes
• 9:10
  a
• 9:10
  basic
• 9:10
  test
• 9:11
  while
• 9:11
  violating
• 9:12
  the
• 9:12
  architecture.
• 9:13
  The
• 9:13
  proposed
• 9:13
  remedy:
• 9:14
  the
• 9:14
  code
• 9:14
  owner
• 9:15
  already
• 9:15
  knows
• 9:15
  which
• 9:16
  few
• 9:16
  files
• 9:16
  matter,
• 9:17
  so
• 9:17
  select
• 9:18
  them
• 9:18
  and
• 9:18
  send
• 9:18
  them
• 9:18
  to
• 9:18
  a
• 9:18
  chatbot
• 9:19
  or
• 9:19
  aider
• 9:19
  .
• 9:20
  The
• 9:20
  strongest
• 9:21
  counter
• 9:21
  spends
• 9:22
  80%
• 9:22
  of
• 9:22
  the
• 9:22
  time
• 9:23
  planning
• 9:23
  with
• 9:23
  a
• 9:23
  high-effort
• 9:24
  model
• 9:24
  before
• 9:24
  it
• 9:24
  touches
• 9:25
  a
• 9:25
  file:
• 9:26
  interrogate
• 9:26
  the
• 9:26
  problem,
• 9:27
  have
• 9:27
  the
• 9:27
  model
• 9:27
  challenge
• 9:28
  your
• 9:28
  assumptions,
• 9:29
  split
• 9:29
  the
• 9:29
  plan
• 9:29
  into
• 9:30
  task
• 9:30
  files
• 9:30
  with
• 9:31
  their
• 9:31
  own
• 9:31
  validation
• 9:31
  steps,
• 9:32
  then
• 9:32
  dispatch
• 9:33
  cheap
• 9:33
  sub-agents.
• 9:35
  Another
• 9:35
  commenter
• 9:35
  had
• 9:36
  the
• 9:36
  sharpest
• 9:36
  diagnosis:
• 9:37
  the
• 9:37
  coffee-break
• 9:38
  pitch
• 9:38
  works
• 9:38
  on
• 9:38
  a
• 9:38
  blank
• 9:39
  canvas
• 9:39
  because
• 9:40
  there
• 9:40
  is
• 9:40
  no
• 9:40
  architecture
• 9:41
  to
• 9:41
  violate,
• 9:42
  and
• 9:42
  once
• 9:42
  history
• 9:42
  exists
• 9:43
  the
• 9:43
  fix
• 9:43
  is
• 9:43
  not
• 9:43
  a
• 9:43
  longer
• 9:44
  leash
• 9:44
  but
• 9:44
  catching
• 9:45
  drift
• 9:45
  after
• 9:45
  each
• 9:45
  meaningful
• 9:46
  diff,
• 9:47
  since
• 9:47
  an
• 9:47
  agent
• 9:47
  eight
• 9:47
  loops
• 9:48
  deep
• 9:48
  "has
• 9:48
  already
• 9:48
  committed
• 9:49
  to
• 9:49
  an
• 9:49
  architecture
• 9:49
  you
• 9:49
  never
• 9:50
  saw
• 9:50
  form."
• 9:51
  OpenAI
• 9:52
  apparently
• 9:52
  agrees
• 9:52
  the
• 9:52
  comparison
• 9:53
  should
• 9:53
  be
• 9:53
  head-to-head.
• 9:55
  Codex
• 9:55
  Replay,
• 9:56
  spotted
• 9:56
  in
• 9:56
  the
• 9:56
  product's
• 9:56
  code,
• 9:57
  invites
• 9:58
  users
• 9:58
  who
• 9:58
  have
• 9:58
  imported
• 9:59
  a
• 9:59
  Claude
• 9:59
  Code
• 9:59
  session
• 9:59
  to
• 10:00
  rerun
• 10:00
  the
• 10:00
  same
• 10:00
  task
• 10:00
  in
• 10:01
  Codex
• 10:01
  and
• 10:01
  compare.
• 10:02
  The
• 10:02
  thread's
• 10:03
  most
• 10:03
  careful
• 10:03
  reader
• 10:03
  noted
• 10:04
  what
• 10:04
  is
• 10:04
  absent:
• 10:05
  no
• 10:05
  scoring
• 10:05
  or
• 10:05
  winner-selection
• 10:06
  logic,
• 10:07
  and
• 10:07
  nothing
• 10:07
  showing
• 10:07
  how
• 10:08
  it
• 10:08
  reconstructs
• 10:08
  the
• 10:08
  original
• 10:09
  repo
• 10:09
  state,
• 10:09
  dependencies,
• 10:10
  environment
• 10:11
  variables,
• 10:11
  or
• 10:11
  tool
• 10:12
  permissions.
• 10:13
  "A
• 10:13
  marketing
• 10:14
  funnel
• 10:14
  dressed
• 10:14
  as
• 10:14
  a
• 10:14
  benchmark
• 10:15
  tool
• 10:15
  for
• 10:15
  now"
• 10:15
  is
• 10:16
  the
• 10:16
  right
• 10:16
  label
• 10:16
  .
• 10:17
  Cole
• 10:17
  Medin
• 10:18
  removes
• 10:18
  the
• 10:18
  harness
• 10:18
  entirely.
• 10:19
  His
• 10:19
  "drive
• 10:20
  screen"
• 10:20
  skill,
• 10:21
  under
• 10:21
  400
• 10:21
  lines,
• 10:22
  lets
• 10:22
  Claude
• 10:22
  Code
• 10:23
  or
• 10:23
  Codex
• 10:23
  control
• 10:24
  the
• 10:24
  desktop
• 10:24
  through
• 10:24
  PowerShell
• 10:25
  on
• 10:25
  Windows
• 10:25
  and
• 10:25
  AppleScript
• 10:26
  on
• 10:26
  Mac,
• 10:27
  with
• 10:27
  an
• 10:27
  opening
• 10:27
  check
• 10:27
  asking
• 10:28
  whether
• 10:28
  screen
• 10:28
  control
• 10:29
  is
• 10:29
  needed
• 10:29
  at
• 10:29
  all.
• 10:30
  On
• 10:30
  prompt
• 10:30
  injection
• 10:31
  from
• 10:31
  on-screen
• 10:32
  content
• 10:32
  he
• 10:32
  argues
• 10:33
  that
• 10:33
  with
• 10:33
  Fable
• 10:33
  5.1
• 10:34
  and
• 10:34
  GPT-6
• 10:35
  Astra
• 10:35
  the
• 10:35
  models
• 10:35
  are
• 10:35
  "very
• 10:36
  very
• 10:36
  hard
• 10:36
  to
• 10:36
  prompt
• 10:37
  inject"
• 10:37
  and
• 10:38
  he
• 10:38
  is
• 10:38
  "just
• 10:38
  not
• 10:38
  concerned
• 10:39
  about
• 10:39
  it
• 10:39
  these
• 10:39
  days."
• 10:40
  Model
• 10:40
  robustness
• 10:41
  is
• 10:41
  a
• 10:41
  vendor
• 10:41
  claim,
• 10:42
  not
• 10:42
  a
• 10:42
  control,
• 10:43
  and
• 10:43
  a
• 10:43
  skill
• 10:43
  with
• 10:43
  shell
• 10:44
  access
• 10:44
  plus
• 10:44
  a
• 10:44
  screenshot
• 10:45
  loop
• 10:45
  is
• 10:45
  a
• 10:45
  broad
• 10:45
  attack
• 10:46
  surface;
• 10:46
  his
• 10:46
  "use
• 10:47
  at
• 10:47
  your
• 10:47
  own
• 10:47
  risk"
• 10:47
  is
• 10:48
  the
• 10:48
  operative
• 10:48
  line
• 10:48
  .
• 10:51
  A
• 10:51
  student
• 10:51
  who
• 10:51
  maintains
• 10:52
  FOSS
• 10:52
  apps
• 10:52
  looked
• 10:52
  at
• 10:53
  the
• 10:53
  102
• 10:53
  apps
• 10:54
  in
• 10:54
  F-Droid's
• 10:54
  September
• 10:54
  12
• 10:55
  update
• 10:55
  batch
• 10:55
  and
• 10:55
  judged
• 10:56
  72%
• 10:56
  of
• 10:56
  them
• 10:57
  largely
• 10:57
  LLM-written.
• 10:58
  The
• 10:58
  method
• 10:59
  is
• 10:59
  deliberately
• 10:59
  unsophisticated,
• 11:00
  and
• 11:00
  the
• 11:01
  author
• 11:01
  says
• 11:01
  so:
• 11:01
  there
• 11:02
  is
• 11:02
  no
• 11:02
  detector,
• 11:03
  but
• 11:03
  LLM-driven
• 11:03
  repositories
• 11:04
  leave
• 11:04
  tells
• 11:05
  because
• 11:05
  the
• 11:05
  technology's
• 11:05
  appeal
• 11:06
  is
• 11:06
  letting
• 11:06
  developers
• 11:07
  be
• 11:07
  lazier.
• 11:08
  The
• 11:08
  audit
• 11:08
  reads
• 11:08
  recent
• 11:09
  commits
• 11:09
  and
• 11:09
  branding
• 11:09
  for
• 11:09
  Claude
• 11:10
  Code
• 11:10
  and
• 11:10
  Codex
• 11:10
  infrastructure,
• 11:11
  commits
• 11:11
  co-authored
• 11:12
  by
• 11:12
  "Claude
• 11:12
  Opus
• 11:12
  5"
• 11:13
  or
• 11:13
  "Claude
• 11:14
  Sonnet
• 11:14
  5,"
• 11:15
  agent
• 11:15
  pull
• 11:15
  requests,
• 11:16
  AI-generated
• 11:17
  icons,
• 11:17
  and
• 11:18
  the
• 11:18
  classic
• 11:18
  10,000-line
• 11:19
  "init"
• 11:20
  commit.
• 11:20
  Any
• 11:21
  repo
• 11:21
  with
• 11:21
  agentic
• 11:21
  infrastructure
• 11:22
  lands
• 11:22
  in
• 11:22
  "mostly
• 11:23
  AI"
• 11:23
  automatically,
• 11:24
  because
• 11:24
  the
• 11:24
  author
• 11:25
  does
• 11:25
  not
• 11:25
  believe
• 11:25
  AI
• 11:26
  can
• 11:26
  be
• 11:26
  used
• 11:26
  responsibly
• 11:27
  from
• 11:27
  inside
• 11:27
  a
• 11:27
  coding
• 11:27
  harness.
• 11:28
  The
• 11:28
  ratings
• 11:29
  are
• 11:29
  admittedly
• 11:30
  superficial,
• 11:31
  and
• 11:31
  no
• 11:31
  code
• 11:31
  quality
• 11:31
  was
• 11:32
  assessed
• 11:32
  .
• 11:32
  Named
• 11:33
  "mostly
• 11:33
  AI"
• 11:34
  apps
• 11:34
  include
• 11:34
  Gem
• 11:34
  Wallet,
• 11:35
  Léon
• 11:36
  the
• 11:36
  URL
• 11:36
  cleaner,
• 11:37
  Tasks.org,
• 11:38
  and
• 11:38
  Voxscribe;
• 11:39
  a
• 11:39
  Deutsche
• 11:40
  Bahn
• 11:40
  arrival-time
• 11:41
  app
• 11:41
  had
• 11:41
  every
• 11:41
  commit
• 11:42
  co-authored
• 11:42
  by
• 11:43
  Claude.
• 11:44
  Four
• 11:44
  of
• 11:44
  five
• 11:44
  Codeberg-hosted
• 11:45
  apps
• 11:45
  were
• 11:45
  mostly
• 11:46
  AI-generated
• 11:47
  and,
• 11:47
  in
• 11:48
  the
• 11:48
  author's
• 11:48
  reading,
• 11:49
  likely
• 11:49
  violate
• 11:49
  Codeberg's
• 11:50
  AI
• 11:50
  policy.
• 11:51
  Projects
• 11:52
  that
• 11:52
  disclosed
• 11:52
  their
• 11:52
  LLM
• 11:53
  use
• 11:53
  were
• 11:53
  praised,
• 11:54
  and
• 11:54
  disclosure
• 11:55
  is
• 11:55
  the
• 11:55
  reasonable
• 11:55
  ask:
• 11:56
  a
• 11:56
  repository
• 11:57
  whose
• 11:57
  trust
• 11:57
  proposition
• 11:58
  is
• 11:58
  "humans
• 11:59
  you
• 11:59
  can
• 11:59
  inspect
• 11:59
  wrote
• 12:00
  this"
• 12:00
  needs
• 12:01
  to
• 12:01
  say
• 12:01
  when
• 12:01
  that
• 12:01
  stopped
• 12:02
  being
• 12:02
  true.
• 12:03
  The
• 12:03
  open
• 12:03
  question
• 12:04
  is
• 12:04
  the
• 12:04
  maintainability
• 12:05
  of
• 12:05
  that
• 12:05
  72%,
• 12:06
  and
• 12:07
  F-Droid
• 12:07
  is
• 12:07
  where
• 12:07
  the
• 12:07
  answer
• 12:08
  will
• 12:08
  show
• 12:08
  up
• 12:08
  first,
• 12:09
  as
• 12:09
  unpatched
• 12:09
  apps.
• 12:10
  The
• 12:11
  receipts
• 12:11
  standard
• 12:11
  applies
• 12:12
  with
• 12:12
  equal
• 12:12
  force
• 12:12
  to
• 12:13
  model
• 12:13
  uploads.
• 12:14
  Qwen3.8-27B-Uncensored-Genesis-V1
• 12:19
  claims
• 12:20
  to
• 12:20
  cure
• 12:20
  overthinking
• 12:21
  by
• 12:21
  "distilling
• 12:21
  training
• 12:22
  noise
• 12:22
  from
• 12:22
  tensors
• 12:22
  using
• 12:23
  Marchenko-Pastur
• 12:23
  distribution."
• 12:25
  The
• 12:25
  author
• 12:25
  provides
• 12:25
  no
• 12:26
  benchmarks,
• 12:26
  code,
• 12:27
  or
• 12:27
  eigenvalue
• 12:28
  spectra;
• 12:29
  commenters
• 12:29
  recall
• 12:30
  the
• 12:30
  same
• 12:30
  uploader
• 12:30
  making
• 12:31
  similar
• 12:31
  claims
• 12:31
  about
• 12:31
  Qwen
• 12:32
  3.5
• 12:32
  and
• 12:33
  3.6
• 12:33
  without
• 12:34
  releasing
• 12:34
  methodology.
• 12:35
  One
• 12:36
  tester
• 12:36
  ran
• 12:36
  a
• 12:36
  constrained
• 12:37
  short-story
• 12:37
  prompt
• 12:38
  and
• 12:38
  got
• 12:38
  36,008
• 12:39
  generated
• 12:40
  tokens
• 12:40
  from
• 12:40
  Genesis
• 12:41
  versus
• 12:41
  30,594
• 12:43
  from
• 12:43
  the
• 12:44
  standard
• 12:44
  Unsloth
• 12:44
  quant,
• 12:45
  so
• 12:45
  on
• 12:45
  the
• 12:45
  one
• 12:46
  available
• 12:46
  measurement
• 12:46
  the
• 12:46
  model
• 12:47
  thought
• 12:47
  more,
• 12:47
  not
• 12:48
  less.
• 12:49
  The
• 12:49
  right
• 12:49
  ask
• 12:50
  came
• 12:50
  from
• 12:50
  the
• 12:50
  sharpest
• 12:50
  comment:
• 12:51
  post
• 12:51
  before-and-after
• 12:52
  eigenvalue
• 12:53
  spectra
• 12:53
  and
• 12:53
  fixed-benchmark
• 12:54
  token
• 12:54
  counts,
• 12:55
  because
• 12:55
  right
• 12:56
  now
• 12:56
  no
• 12:56
  one
• 12:56
  can
• 12:56
  tell
• 12:57
  whether
• 12:57
  denoising
• 12:57
  happened
• 12:58
  or
• 12:58
  the
• 12:58
  model
• 12:58
  was
• 12:59
  rank-truncated
• 12:59
  into
• 12:59
  different
• 13:00
  behavior
• 13:00
  .
• 13:01
  Bruce
• 13:02
  Schneier
• 13:02
  and
• 13:02
  the
• 13:02
  EFF's
• 13:03
  Cindy
• 13:03
  Cohn
• 13:03
  argue
• 13:04
  that
• 13:04
  the
• 13:04
  post-9/11
• 13:05
  shift
• 13:06
  from
• 13:06
  targeted
• 13:06
  surveillance
• 13:07
  to
• 13:07
  bulk
• 13:07
  collection
• 13:08
  has
• 13:08
  outlived
• 13:08
  its
• 13:09
  counterterrorism
• 13:09
  justification
• 13:10
  and
• 13:10
  should
• 13:10
  be
• 13:10
  dismantled
• 13:11
  rather
• 13:11
  than
• 13:11
  renewed.
• 13:12
  The
• 13:12
  telephone-records
• 13:13
  program
• 13:13
  began
• 13:13
  as
• 13:14
  the
• 13:14
  President's
• 13:14
  Surveillance
• 13:15
  Program,
• 13:15
  was
• 13:15
  secretly
• 13:16
  rebased
• 13:16
  onto
• 13:16
  Section
• 13:17
  215
• 13:18
  of
• 13:18
  the
• 13:18
  Patriot
• 13:18
  Act
• 13:18
  in
• 13:19
  2006,
• 13:20
  was
• 13:20
  rejected
• 13:20
  by
• 13:20
  the
• 13:20
  Second
• 13:21
  Circuit
• 13:21
  in
• 13:21
  2015,
• 13:22
  and
• 13:22
  was
• 13:22
  curtailed
• 13:23
  by
• 13:23
  Congress
• 13:23
  that
• 13:23
  year
• 13:23
  after
• 13:24
  nearly
• 13:24
  fourteen
• 13:24
  years.
• 13:25
  Congress
• 13:25
  let
• 13:25
  Section
• 13:26
  702
• 13:26
  lapse
• 13:27
  this
• 13:27
  year,
• 13:27
  though
• 13:27
  approved
• 13:28
  collection
• 13:28
  continues
• 13:28
  until
• 13:29
  at
• 13:29
  least
• 13:29
  spring
• 13:29
  2027.
• 13:31
  Meanwhile
• 13:31
  mass
• 13:32
  surveillance
• 13:32
  has
• 13:32
  become
• 13:32
  routine
• 13:33
  for
• 13:33
  domestic
• 13:33
  law
• 13:33
  enforcement,
• 13:34
  ICE,
• 13:35
  and
• 13:35
  private
• 13:35
  security:
• 13:37
  FBI
• 13:37
  Director
• 13:37
  Kash
• 13:38
  Patel
• 13:38
  confirmed
• 13:38
  in
• 13:38
  congressional
• 13:39
  testimony
• 13:40
  that
• 13:40
  the
• 13:40
  Bureau
• 13:40
  buys
• 13:40
  Americans'
• 13:41
  data
• 13:41
  from
• 13:41
  brokers
• 13:42
  and
• 13:42
  intends
• 13:42
  to
• 13:42
  continue
• 13:43
  .
• 13:44
  The
• 13:44
  national
• 13:44
  security
• 13:45
  community
• 13:45
  has
• 13:45
  never
• 13:46
  produced
• 13:46
  a
• 13:46
  real
• 13:46
  cost-benefit
• 13:47
  analysis,
• 13:48
  claimed
• 13:48
  successes
• 13:49
  regularly
• 13:49
  collapse
• 13:50
  under
• 13:50
  scrutiny,
• 13:51
  and
• 13:51
  Flock
• 13:51
  treats
• 13:52
  a
• 13:52
  database
• 13:52
  hit
• 13:53
  as
• 13:53
  a
• 13:53
  solved
• 13:53
  crime.
• 13:54
  The
• 13:55
  abuse
• 13:55
  record
• 13:55
  is
• 13:55
  specific:
• 13:56
  Rep.
• 13:57
  Zoe
• 13:58
  Lofgren's
• 13:58
  list
• 13:58
  of
• 13:58
  backdoor-search
• 13:59
  targets,
• 14:00
  including
• 14:00
  protesters,
• 14:01
  campaign
• 14:02
  donors,
• 14:02
  members
• 14:03
  of
• 14:03
  Congress,
• 14:04
  journalists,
• 14:04
  and
• 14:05
  a
• 14:05
  state
• 14:05
  judge;
• 14:05
  a
• 14:06
  Texas
• 14:06
  officer
• 14:07
  using
• 14:07
  plate
• 14:07
  readers
• 14:08
  to
• 14:08
  track
• 14:08
  a
• 14:08
  woman
• 14:08
  suspected
• 14:09
  of
• 14:09
  seeking
• 14:09
  an
• 14:09
  abortion;
• 14:10
  and
• 14:10
  a
• 14:10
  White
• 14:11
  House
• 14:11
  strategy
• 14:11
  document
• 14:12
  pledging
• 14:12
  on
• 14:12
  page
• 14:13
  seven
• 14:13
  to
• 14:13
  "map"
• 14:14
  domestic
• 14:14
  groups
• 14:14
  it
• 14:15
  labels
• 14:15
  anti-American.
• 14:16
  AI,
• 14:17
  they
• 14:17
  argue,
• 14:18
  makes
• 14:18
  the
• 14:18
  "only
• 14:18
  human
• 14:19
  eyes
• 14:19
  matter"
• 14:20
  doctrine
• 14:20
  obsolete
• 14:21
  because
• 14:21
  analysis
• 14:22
  no
• 14:22
  longer
• 14:23
  needs
• 14:23
  eyes.
• 14:24
  The
• 14:24
  remedies
• 14:25
  are
• 14:25
  warrants
• 14:25
  before
• 14:25
  collection,
• 14:26
  access,
• 14:27
  or
• 14:27
  use
• 14:27
  of
• 14:27
  mass-surveillance
• 14:28
  data
• 14:28
  regardless
• 14:29
  of
• 14:29
  custodian,
• 14:30
  and
• 14:30
  the
• 14:30
  Fourth
• 14:30
  Amendment
• 14:31
  Is
• 14:31
  Not
• 14:31
  For
• 14:31
  Sale
• 14:32
  Act.
• 14:33
  The
• 14:33
  closing
• 14:33
  claim
• 14:33
  that
• 14:34
  the
• 14:34
  current
• 14:34
  administration
• 14:35
  uses
• 14:35
  these
• 14:35
  tools
• 14:36
  to
• 14:36
  keep
• 14:36
  itself
• 14:36
  in
• 14:36
  power
• 14:37
  is
• 14:37
  an
• 14:37
  assertion;
• 14:38
  the
• 14:38
  abuse
• 14:39
  examples
• 14:39
  are
• 14:39
  the
• 14:39
  documented
• 14:40
  part.
• 14:41
  At
• 14:41
  a
• 14:41
  much
• 14:41
  smaller
• 14:41
  scale,
• 14:42
  Hugging
• 14:42
  Face's
• 14:43
  security.txt
• 14:45
  made
• 14:45
  the
• 14:45
  rounds
• 14:45
  on
• 14:45
  LocalLLaMA
• 14:46
  because,
• 14:47
  judging
• 14:47
  from
• 14:47
  the
• 14:47
  reactions,
• 14:48
  it
• 14:48
  reads
• 14:49
  as
• 14:49
  an
• 14:49
  appeal
• 14:49
  addressed
• 14:50
  to
• 14:50
  AI
• 14:50
  agents
• 14:50
  as
• 14:50
  much
• 14:51
  as
• 14:51
  to
• 14:51
  human
• 14:51
  researchers.
• 14:52
  The
• 14:52
  jokes
• 14:53
  wrote
• 14:53
  themselves,
• 14:53
  from
• 14:54
  "21st
• 14:54
  century
• 14:55
  antivirus"
• 14:55
  to
• 14:55
  finally,
• 14:56
  a
• 14:56
  .txt
• 14:56
  on
• 14:57
  Hugging
• 14:57
  Face
• 14:57
  that
• 14:57
  isn't
• 14:57
  a
• 14:57
  tokenizer
• 14:58
  vocab.
• 14:59
  The
• 14:59
  serious
• 15:00
  note
• 15:00
  is
• 15:00
  that
• 15:00
  security.txt
• 15:02
  was
• 15:02
  designed
• 15:02
  for
• 15:03
  humans
• 15:03
  to
• 15:03
  find
• 15:03
  a
• 15:03
  disclosure
• 15:04
  contact,
• 15:05
  and
• 15:05
  platforms
• 15:05
  are
• 15:05
  now
• 15:06
  writing
• 15:06
  for
• 15:06
  a
• 15:06
  second
• 15:06
  audience
• 15:07
  that
• 15:07
  scrapes,
• 15:07
  reads,
• 15:08
  and
• 15:08
  acts.
• 15:09
  A
• 15:09
  polite
• 15:09
  request
• 15:10
  in
• 15:10
  a
• 15:10
  text
• 15:10
  file
• 15:10
  is
• 15:10
  not
• 15:11
  a
• 15:11
  control,
• 15:11
  but
• 15:11
  it
• 15:11
  is
• 15:12
  not
• 15:12
  nothing
• 15:12
  either
• 15:12
  .
• 15:15
  ABot-Recon,
• 15:16
  from
• 15:16
  Amap's
• 15:16
  computer
• 15:17
  vision
• 15:17
  lab,
• 15:18
  makes
• 15:18
  a
• 15:18
  contrarian
• 15:18
  bet
• 15:19
  for
• 15:19
  streaming
• 15:19
  3D
• 15:20
  reconstruction:
• 15:21
  no
• 15:21
  persistent
• 15:21
  learned
• 15:22
  long-range
• 15:22
  memory
• 15:22
  at
• 15:23
  all.
• 15:23
  The
• 15:23
  model
• 15:23
  caches
• 15:24
  KV
• 15:24
  features
• 15:24
  from
• 15:25
  the
• 15:25
  preceding
• 15:25
  11
• 15:25
  frames,
• 15:26
  predicts
• 15:26
  a
• 15:26
  point
• 15:27
  map
• 15:27
  for
• 15:27
  the
• 15:27
  current
• 15:27
  frame
• 15:28
  in
• 15:28
  its
• 15:28
  own
• 15:28
  camera
• 15:28
  coordinates,
• 15:29
  estimates
• 15:30
  the
• 15:30
  relative
• 15:30
  pose
• 15:31
  to
• 15:31
  the
• 15:31
  previous
• 15:31
  frame,
• 15:32
  and
• 15:32
  recovers
• 15:33
  the
• 15:33
  global
• 15:33
  trajectory
• 15:34
  and
• 15:34
  point
• 15:34
  cloud
• 15:34
  by
• 15:34
  composing
• 15:35
  poses
• 15:35
  sequentially.
• 15:36
  Because
• 15:36
  it
• 15:36
  solves
• 15:37
  the
• 15:37
  same
• 15:37
  bounded
• 15:37
  12-frame
• 15:38
  problem
• 15:38
  at
• 15:38
  every
• 15:39
  step,
• 15:39
  state
• 15:40
  and
• 15:40
  per-frame
• 15:40
  compute
• 15:41
  stay
• 15:41
  independent
• 15:41
  of
• 15:41
  sequence
• 15:42
  length.
• 15:43
  On
• 15:43
  Oxford
• 15:43
  Spires
• 15:44
  the
• 15:44
  streaming
• 15:44
  model
• 15:44
  alone
• 15:45
  reports
• 15:45
  4.35
• 15:46
  m
• 15:46
  absolute
• 15:47
  trajectory
• 15:47
  error
• 15:48
  and
• 15:48
  0.12
• 15:49
  degrees
• 15:49
  rotational
• 15:50
  error;
• 15:50
  on
• 15:51
  KITTI-02
• 15:54
  it
• 15:54
  runs
• 15:54
  at
• 15:54
  24.45
• 15:55
  FPS
• 15:56
  in
• 15:56
  6.71
• 15:57
  GiB
• 15:58
  on
• 15:58
  an
• 15:58
  H100.
• 16:00
  Weights
• 16:00
  are
• 16:00
  public
• 16:00
  on
• 16:00
  Hugging
• 16:01
  Face
• 16:01
  and
• 16:01
  ModelScope,
• 16:02
  training
• 16:03
  code
• 16:03
  is
• 16:03
  promised
• 16:03
  by
• 16:03
  September
• 16:04
  30,
• 16:05
  and
• 16:05
  the
• 16:05
  code
• 16:05
  is
• 16:05
  Apache
• 16:06
  2.0
• 16:07
  while
• 16:07
  the
• 16:07
  weights
• 16:07
  carry
• 16:08
  a
• 16:08
  separate
• 16:08
  license
• 16:08
  .
• 16:10
  OUI-1
• 16:10
  generates
• 16:10
  UI
• 16:11
  elements
• 16:11
  rather
• 16:11
  than
• 16:11
  prose,
• 16:12
  fine-tuned
• 16:13
  on
• 16:13
  DiffusionGemma
• 16:14
  to
• 16:14
  emit
• 16:14
  OpenUI-Lang,
• 16:15
  a
• 16:15
  custom
• 16:16
  DSL,
• 16:17
  instead
• 16:17
  of
• 16:17
  HTML
• 16:18
  or
• 16:18
  React.
• 16:19
  The
• 16:19
  reasoning
• 16:20
  is
• 16:20
  sound:
• 16:21
  any
• 16:21
  LLM
• 16:22
  can
• 16:22
  learn
• 16:22
  the
• 16:22
  DSL
• 16:23
  from
• 16:23
  a
• 16:23
  system
• 16:23
  prompt,
• 16:24
  but
• 16:24
  that
• 16:24
  eats
• 16:25
  context,
• 16:26
  so
• 16:26
  baking
• 16:26
  it
• 16:27
  into
• 16:27
  weights
• 16:27
  leaves
• 16:27
  room
• 16:28
  for
• 16:28
  conversation
• 16:28
  and
• 16:28
  tool
• 16:29
  calls.
• 16:30
  Adoption
• 16:30
  is
• 16:30
  stuck
• 16:31
  because
• 16:31
  llama.cpp
• 16:33
  does
• 16:33
  not
• 16:33
  support
• 16:34
  DiffusionGemma,
• 16:35
  commenters
• 16:35
  flagged
• 16:36
  a
• 16:36
  license
• 16:36
  change
• 16:37
  from
• 16:37
  Apache
• 16:37
  to
• 16:37
  Gemma
• 16:37
  terms,
• 16:38
  and
• 16:38
  one
• 16:38
  noted
• 16:39
  that
• 16:39
  users
• 16:39
  notice
• 16:40
  100
• 16:40
  ms
• 16:40
  on
• 16:41
  a
• 16:41
  click,
• 16:41
  so
• 16:42
  a
• 16:42
  UI-generating
• 16:43
  model
• 16:43
  must
• 16:43
  be
• 16:43
  tiny
• 16:44
  and
• 16:44
  fast
• 16:44
  or
• 16:44
  it
• 16:44
  is
• 16:45
  a
• 16:45
  demo
• 16:45
  .
• 16:46
  Ideogram-4,
• 16:47
  packaged
• 16:48
  for
• 16:48
  ComfyUI
• 16:48
  by
• 16:48
  Comfy-Org,
• 16:49
  is
• 16:49
  meanwhile
• 16:50
  trending
• 16:50
  on
• 16:50
  Hugging
• 16:51
  Face
• 16:51
  .
• 16:52
  SuperAstra
• 16:52
  takes
• 16:53
  an
• 16:53
  agent
• 16:53
  into
• 16:53
  a
• 16:53
  running
• 16:54
  SNES
• 16:54
  game.
• 16:55
  A
• 16:55
  Python
• 16:56
  desktop
• 16:56
  companion
• 16:57
  talks
• 16:57
  to
• 16:57
  BizHawk
• 16:57
  through
• 16:58
  a
• 16:58
  Lua
• 16:58
  bridge,
• 16:59
  and
• 16:59
  OpenAI's
• 17:00
  gpt-6-astra,
• 17:01
  via
• 17:01
  the
• 17:01
  Responses
• 17:02
  API
• 17:02
  with
• 17:02
  function
• 17:03
  calling,
• 17:03
  gets
• 17:04
  live
• 17:04
  screenshots,
• 17:05
  CPU
• 17:05
  registers,
• 17:06
  disassembly,
• 17:07
  memory
• 17:08
  reads,
• 17:08
  and
• 17:08
  a
• 17:08
  persistent
• 17:09
  per-ROM
• 17:09
  notebook,
• 17:10
  then
• 17:10
  writes
• 17:10
  one-shot
• 17:11
  routines,
• 17:12
  per-frame
• 17:12
  effects,
• 17:13
  or
• 17:13
  guarded
• 17:14
  cartridge
• 17:14
  patches
• 17:14
  capped
• 17:15
  at
• 17:15
  4
• 17:15
  KB
• 17:15
  per
• 17:15
  operation.
• 17:17
  Every
• 17:17
  mutation
• 17:18
  gets
• 17:18
  an
• 17:18
  emulator
• 17:19
  checkpoint
• 17:19
  and
• 17:20
  the
• 17:20
  original
• 17:20
  ROM
• 17:20
  is
• 17:21
  never
• 17:21
  touched.
• 17:22
  An
• 17:22
  agent
• 17:23
  that
• 17:23
  rewrites
• 17:23
  game
• 17:23
  state
• 17:24
  on
• 17:24
  request
• 17:24
  is
• 17:24
  the
• 17:24
  same
• 17:25
  capability
• 17:25
  alignment
• 17:26
  evaluations
• 17:26
  try
• 17:27
  to
• 17:27
  catch
• 17:27
  when
• 17:27
  it
• 17:27
  happens
• 17:28
  uninvited,
• 17:28
  with
• 17:29
  the
• 17:29
  intent
• 17:29
  reversed
• 17:29
  .
• 17:30
  Less
• 17:31
  substantive
• 17:31
  was
• 17:31
  a
• 17:31
  Gemini
• 17:32
  video
• 17:32
  pitched
• 17:32
  as
• 17:33
  a
• 17:33
  fast,
• 17:33
  cheap
• 17:33
  alternative
• 17:34
  to
• 17:34
  motion
• 17:34
  capture,
• 17:35
  which
• 17:35
  commenters
• 17:36
  correctly
• 17:36
  identified
• 17:37
  as
• 17:37
  the
• 17:37
  same
• 17:37
  footage
• 17:38
  restyled
• 17:38
  rather
• 17:38
  than
• 17:38
  a
• 17:39
  video-to-rig
• 17:39
  pipeline
• 17:39
  .