Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – September 22, 2026

Transcript

AGI Dreams Podcast – September 22, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  most
• 0:00
  useful
• 0:00
  number
• 0:01
  in
• 0:01
  cyber-AI
• 0:01
  today
• 0:02
  is
• 0:02
  a
• 0:02
  blinded
• 0:02
  one.
• 0:03
  Running
• 0:03
  GLM-5.3
• 0:04
  and
• 0:04
  its
• 0:05
  cheaper
• 0:05
  sibling
• 0:05
  GLM-5.3
• 0:06
  Flash
• 0:07
  against
• 0:07
  59
• 0:08
  cyber-investigation
• 0:09
  cases
• 0:09
  the
• 0:09
  models
• 0:09
  had
• 0:09
  never
• 0:10
  seen,
• 0:10
  botsbench
• 0:11
  found
• 0:11
  the
• 0:11
  two
• 0:11
  effectively
• 0:12
  level:
• 0:12
  29
• 0:13
  solved
• 0:13
  for
• 0:14
  the
• 0:14
  full
• 0:14
  model,
• 0:14
  32
• 0:15
  for
• 0:15
  Flash.
• 0:16
  The
• 0:16
  full
• 0:16
  model
• 0:17
  pulls
• 0:17
  ahead
• 0:17
  only
• 0:17
  on
• 0:17
  BOTSv3,
• 0:19
  a
• 0:19
  practice
• 0:19
  set
• 0:19
  public
• 0:20
  for
• 0:20
  years.
• 0:21
  The
• 0:21
  real
• 0:21
  story
• 0:21
  is
• 0:21
  price.
• 0:22
  GLM-5.3
• 0:23
  costs
• 0:24
  roughly
• 0:24
  fifteen
• 0:24
  times
• 0:25
  more
• 0:25
  per
• 0:25
  solved
• 0:25
  case
• 0:26
  than
• 0:26
  Flash
• 0:26
  and
• 0:26
  solves
• 0:27
  no
• 0:27
  more
• 0:27
  of
• 0:27
  them.
• 0:28
  Flash's
• 0:28
  best
• 0:28
  run,
• 0:29
  38
• 0:29
  of
• 0:29
  59,
• 0:30
  is
• 0:30
  the
• 0:30
  highest
• 0:31
  any
• 0:31
  open-weight
• 0:31
  model
• 0:32
  has
• 0:32
  reached
• 0:32
  here,
• 0:33
  matching
• 0:33
  the
• 0:33
  GLM-5.2
• 0:34
  it
• 0:34
  replaces
• 0:35
  at
• 0:35
  a
• 0:35
  fraction
• 0:35
  of
• 0:35
  the
• 0:35
  cost.
• 0:36
  The
• 0:36
  same
• 0:37
  model
• 0:37
  also
• 0:37
  scored
• 0:38
  64%
• 0:38
  in
• 0:38
  one
• 0:39
  agent
• 0:39
  harness
• 0:39
  and
• 0:39
  54%
• 0:40
  in
• 0:40
  another,
• 0:41
  proof
• 0:41
  that
• 0:42
  a
• 0:42
  score
• 0:42
  belongs
• 0:42
  to
• 0:42
  a
• 0:42
  scaffold,
• 0:43
  not
• 0:43
  the
• 0:44
  weights
• 0:44
  .
• 0:45
  Against
• 0:45
  a
• 0:45
  backdrop
• 0:46
  of
• 0:46
  capture-the-flag
• 0:46
  audits
• 0:47
  that
• 0:47
  caught
• 0:47
  models
• 0:48
  git-cloning
• 0:48
  official
• 0:48
  writeups
• 0:49
  and
• 0:49
  reading
• 0:49
  plaintext
• 0:50
  flags,
• 0:50
  the
• 0:51
  line
• 0:51
  worth
• 0:51
  underlining
• 0:52
  is
• 0:52
  that
• 0:52
  these
• 0:52
  cases
• 0:53
  were
• 0:53
  blinded,
• 0:54
  with
• 0:54
  no
• 0:54
  contamination,
• 0:55
  cheating,
• 0:55
  or
• 0:56
  refusals.
• 0:57
  Exploitation
• 0:58
  is
• 0:58
  where
• 0:58
  autonomy
• 0:59
  stops
• 0:59
  being
• 0:59
  abstract.
• 1:01
  Inside
• 1:01
  ExploitBench,
• 1:02
  a
• 1:02
  model
• 1:03
  built
• 1:03
  a
• 1:03
  full
• 1:03
  renderer-grade
• 1:04
  chain
• 1:04
  against
• 1:04
  V8
• 1:05
  in
• 1:05
  Chrome
• 1:05
  127,
• 1:07
  from
• 1:07
  a
• 1:07
  WebAssembly
• 1:07
  out-of-bounds
• 1:08
  write
• 1:08
  through
• 1:08
  address
• 1:09
  disclosure,
• 1:10
  a
• 1:10
  sandbox
• 1:10
  cage
• 1:10
  escape,
• 1:11
  arbitrary
• 1:11
  native
• 1:12
  read-write,
• 1:12
  and
• 1:12
  native
• 1:13
  code
• 1:13
  execution,
• 1:14
  scored
• 1:14
  16
• 1:15
  of
• 1:15
  16.
• 1:16
  The
• 1:16
  evidence
• 1:16
  is
• 1:17
  not
• 1:17
  one
• 1:17
  lucky
• 1:17
  grade
• 1:18
  but
• 1:18
  a
• 1:18
  three-day
• 1:18
  ladder
• 1:19
  of
• 1:19
  50
• 1:19
  recorded
• 1:20
  grader
• 1:20
  runs
• 1:20
  climbing
• 1:21
  from
• 1:21
  2
• 1:21
  to
• 1:21
  16
• 1:22
  across
• 1:22
  roughly
• 1:23
  fifteen
• 1:23
  distinct
• 1:24
  solution
• 1:24
  hashes,
• 1:25
  dead
• 1:25
  ends
• 1:25
  included.
• 1:27
  A
• 1:27
  full
• 1:27
  browser
• 1:27
  compromise
• 1:28
  still
• 1:28
  needs
• 1:28
  a
• 1:28
  second
• 1:29
  broker-escape
• 1:29
  stage,
• 1:30
  but
• 1:30
  this
• 1:30
  is
• 1:30
  arbitrary
• 1:31
  code
• 1:31
  execution
• 1:32
  in
• 1:32
  the
• 1:32
  V8
• 1:32
  process,
• 1:33
  built
• 1:33
  iteratively
• 1:34
  by
• 1:34
  the
• 1:34
  agent
• 1:34
  .
• 1:35
  Caleb
• 1:36
  Gross,
• 1:36
  in
• 1:36
  a
• 1:36
  Black
• 1:37
  Hat
• 1:37
  talk
• 1:37
  on
• 1:37
  his
• 1:37
  SiftRank
• 1:38
  tool,
• 1:38
  reframes
• 1:39
  the
• 1:39
  whole
• 1:39
  exercise:
• 1:40
  vulnerability
• 1:41
  discovery
• 1:42
  is
• 1:42
  an
• 1:42
  attention
• 1:42
  problem,
• 1:43
  not
• 1:43
  a
• 1:43
  detection
• 1:44
  problem.
• 1:45
  Rather
• 1:45
  than
• 1:45
  force
• 1:45
  one
• 1:46
  model
• 1:46
  to
• 1:46
  absorb
• 1:47
  context,
• 1:47
  reason,
• 1:48
  and
• 1:48
  act
• 1:48
  in
• 1:48
  a
• 1:48
  single
• 1:49
  loop,
• 1:50
  it
• 1:50
  asks
• 1:50
  only
• 1:50
  for
• 1:50
  a
• 1:50
  reordering
• 1:51
  of
• 1:51
  small
• 1:51
  document
• 1:52
  batches
• 1:52
  by
• 1:52
  relevance
• 1:53
  to
• 1:53
  a
• 1:53
  deliberately
• 1:53
  vague
• 1:54
  query.
• 1:55
  Pointed
• 1:55
  at
• 1:55
  a
• 1:55
  WattBox
• 1:56
  surge
• 1:56
  protector's
• 1:56
  55
• 1:57
  MB
• 1:57
  firmware
• 1:58
  decompiled
• 1:58
  into
• 1:58
  about
• 1:59
  5,700
• 2:00
  functions,
• 2:01
  it
• 2:01
  ranked
• 2:01
  them
• 2:01
  in
• 2:01
  batches
• 2:02
  of
• 2:02
  six
• 2:02
  and
• 2:02
  collapsed
• 2:03
  the
• 2:03
  corpus
• 2:03
  to
• 2:03
  about
• 2:04
  four
• 2:04
  candidates.
• 2:05
  The
• 2:05
  top
• 2:05
  hit
• 2:05
  dropped
• 2:06
  a
• 2:06
  parameter
• 2:06
  straight
• 2:07
  into
• 2:07
  a
• 2:07
  popen
• 2:07
  call,
• 2:08
  an
• 2:08
  unconstrained
• 2:08
  root
• 2:09
  shell,
• 2:09
  now
• 2:09
  patched
• 2:10
  with
• 2:10
  a
• 2:10
  CVE.
• 2:11
  On
• 2:11
  the
• 2:11
  BinPwn
• 2:12
  dataset
• 2:12
  it
• 2:12
  beat
• 2:13
  zero-shot
• 2:13
  classification
• 2:14
  by
• 2:14
  more
• 2:14
  than
• 2:14
  double,
• 2:15
  and
• 2:15
  small
• 2:16
  models
• 2:16
  like
• 2:16
  GPT-5
• 2:17
  Nano
• 2:17
  matched
• 2:18
  the
• 2:18
  full
• 2:18
  model.
• 2:19
  The
• 2:19
  edge
• 2:19
  grows
• 2:20
  as
• 2:20
  vulnerabilities
• 2:20
  get
• 2:21
  rarer,
• 2:21
  which
• 2:21
  is
• 2:21
  the
• 2:22
  real-world
• 2:22
  case
• 2:22
  .
• 2:26
  Two
• 2:26
  projects
• 2:26
  come
• 2:27
  at
• 2:27
  model
• 2:27
  safety
• 2:27
  from
• 2:28
  opposite
• 2:28
  ends
• 2:28
  and
• 2:28
  reach
• 2:28
  the
• 2:29
  same
• 2:29
  conclusion:
• 2:30
  the
• 2:30
  guardrails
• 2:30
  are
• 2:30
  thinner
• 2:30
  than
• 2:31
  the
• 2:31
  marketing
• 2:31
  implies.
• 2:32
  The
• 2:32
  research
• 2:33
  contribution
• 2:33
  is
• 2:33
  BLOOM-WILT,
• 2:34
  out
• 2:34
  of
• 2:34
  the
• 2:34
  University
• 2:35
  of
• 2:35
  Manchester,
• 2:36
  which
• 2:36
  extends
• 2:36
  an
• 2:36
  automated
• 2:37
  auditing
• 2:37
  pipeline
• 2:37
  with
• 2:37
  two
• 2:38
  tricks
• 2:38
  that
• 2:38
  need
• 2:38
  no
• 2:38
  training,
• 2:39
  no
• 2:39
  second
• 2:39
  model,
• 2:40
  and
• 2:40
  no
• 2:40
  access
• 2:41
  beyond
• 2:41
  the
• 2:41
  target's
• 2:41
  next-token
• 2:42
  logits.
• 2:42
  One
• 2:43
  lets
• 2:43
  the
• 2:43
  auditor
• 2:43
  revise
• 2:44
  its
• 2:44
  strategy
• 2:44
  across
• 2:44
  rounds
• 2:45
  from
• 2:45
  judge
• 2:45
  scores;
• 2:46
  the
• 2:46
  other
• 2:46
  blends
• 2:46
  the
• 2:46
  target's
• 2:47
  own
• 2:47
  next-token
• 2:47
  distribution
• 2:48
  with
• 2:48
  a
• 2:48
  second
• 2:48
  one
• 2:48
  conditioned
• 2:49
  on
• 2:49
  a
• 2:49
  behavior-eliciting
• 2:50
  prompt,
• 2:50
  masking
• 2:51
  tokens
• 2:51
  the
• 2:51
  base
• 2:51
  model
• 2:52
  would
• 2:52
  never
• 2:52
  produce.
• 2:53
  Because
• 2:53
  it
• 2:53
  reuses
• 2:53
  the
• 2:53
  same
• 2:54
  weights
• 2:54
  and
• 2:54
  does
• 2:54
  not
• 2:54
  decay
• 2:55
  across
• 2:55
  a
• 2:55
  generation,
• 2:56
  the
• 2:56
  elicited
• 2:56
  behavior
• 2:57
  persists
• 2:57
  through
• 2:57
  multi-turn
• 2:58
  conversations,
• 2:59
  which
• 2:59
  is
• 2:59
  what
• 2:59
  auditing
• 2:59
  requires
• 3:00
  .
• 3:00
  Tested
• 3:01
  on
• 3:01
  four
• 3:01
  small
• 3:01
  instruct
• 3:02
  models
• 3:02
  chosen
• 3:03
  so
• 3:03
  differences
• 3:03
  reflect
• 3:04
  alignment
• 3:04
  rather
• 3:04
  than
• 3:04
  scale,
• 3:05
  the
• 3:05
  method
• 3:05
  improves
• 3:06
  on
• 3:06
  every
• 3:06
  metric
• 3:06
  with
• 3:06
  no
• 3:07
  overlapping
• 3:07
  error
• 3:07
  intervals
• 3:08
  and,
• 3:08
  in
• 3:08
  the
• 3:08
  authors'
• 3:08
  words,
• 3:09
  overturns
• 3:09
  the
• 3:10
  previous
• 3:10
  safety
• 3:10
  rankings.
• 3:11
  Qwen3.5-4B
• 3:13
  looked
• 3:13
  safer
• 3:14
  than
• 3:14
  its
• 3:14
  peers
• 3:14
  under
• 3:14
  ordinary
• 3:15
  auditing
• 3:15
  and
• 3:15
  turns
• 3:16
  out
• 3:16
  no
• 3:16
  better
• 3:16
  than
• 3:16
  the
• 3:16
  rest
• 3:17
  once
• 3:17
  real
• 3:17
  pressure
• 3:17
  is
• 3:17
  applied,
• 3:18
  while
• 3:18
  Gemma-4-E4B
• 3:19
  holds
• 3:19
  up
• 3:19
  best.
• 3:20
  The
• 3:20
  honest
• 3:20
  limits:
• 3:21
  only
• 3:21
  3-to-4-billion-parameter
• 3:22
  targets
• 3:22
  were
• 3:22
  tested,
• 3:23
  the
• 3:23
  method
• 3:23
  needs
• 3:23
  output
• 3:24
  logits
• 3:24
  so
• 3:24
  text-only
• 3:25
  APIs
• 3:25
  are
• 3:25
  out
• 3:25
  of
• 3:25
  reach,
• 3:26
  and
• 3:26
  some
• 3:26
  refusals
• 3:26
  survive.
• 3:27
  The
• 3:27
  conclusion
• 3:28
  is
• 3:28
  blunt,
• 3:29
  that
• 3:29
  these
• 3:29
  models'
• 3:29
  safety
• 3:30
  alignment
• 3:30
  is
• 3:30
  shallower
• 3:30
  than
• 3:31
  it
• 3:31
  appeared.
• 3:32
  If
• 3:32
  BLOOM-WILT
• 3:32
  measures
• 3:32
  how
• 3:33
  shallow
• 3:33
  alignment
• 3:33
  is,
• 3:34
  phantom-kv
• 3:34
  exploits
• 3:35
  it.
• 3:35
  The
• 3:35
  project
• 3:36
  strips
• 3:36
  refusal
• 3:36
  behavior
• 3:37
  without
• 3:37
  touching
• 3:37
  weights,
• 3:38
  shipping
• 3:38
  the
• 3:38
  change
• 3:39
  as
• 3:39
  a
• 3:39
  loadable
• 3:39
  key-value
• 3:40
  cache
• 3:40
  graft,
• 3:41
  a
• 3:41
  small
• 3:41
  bank
• 3:41
  of
• 3:41
  per-layer
• 3:42
  tensors
• 3:42
  spliced
• 3:42
  into
• 3:43
  the
• 3:43
  attention
• 3:43
  cache.
• 3:44
  Nothing
• 3:44
  is
• 3:44
  subtracted
• 3:45
  or
• 3:45
  projected
• 3:45
  out;
• 3:46
  the
• 3:46
  model's
• 3:47
  own
• 3:47
  attention
• 3:47
  does
• 3:47
  the
• 3:47
  steering,
• 3:48
  and
• 3:48
  unloading
• 3:49
  the
• 3:49
  graft
• 3:49
  leaves
• 3:49
  the
• 3:49
  base
• 3:50
  model
• 3:50
  byte-identical
• 3:51
  again.
• 3:52
  On
• 3:52
  Qwen3-4B
• 3:53
  the
• 3:53
  base
• 3:53
  refuses
• 3:53
  25
• 3:54
  of
• 3:54
  60
• 3:54
  harmful
• 3:55
  prompts;
• 3:56
  the
• 3:56
  strongest
• 3:56
  graft
• 3:57
  cuts
• 3:57
  that
• 3:57
  to
• 3:57
  5
• 3:58
  while
• 3:58
  holding
• 3:58
  divergence
• 3:59
  on
• 3:59
  harmless
• 3:59
  prompts
• 3:59
  near
• 4:00
  0.015
• 4:01
  nats.
• 4:02
  The
• 4:02
  author
• 4:02
  does
• 4:02
  not
• 4:03
  oversell
• 4:03
  it:
• 4:04
  capability
• 4:04
  drops
• 4:05
  under
• 4:05
  the
• 4:05
  graft,
• 4:05
  and
• 4:06
  it
• 4:06
  covers
• 4:06
  only
• 4:06
  dense
• 4:07
  Qwen3.
• 4:08
  As
• 4:08
  a
• 4:08
  dual-use
• 4:09
  artifact
• 4:09
  it
• 4:09
  is
• 4:09
  notable
• 4:10
  precisely
• 4:10
  because
• 4:11
  it
• 4:11
  looks
• 4:11
  like
• 4:11
  ordinary
• 4:12
  conversation
• 4:12
  history
• 4:13
  and
• 4:13
  hot-swaps
• 4:14
  mid-session;
• 4:15
  its
• 4:15
  disclosed
• 4:15
  leakage
• 4:16
  evals
• 4:16
  are
• 4:16
  the
• 4:16
  responsible
• 4:17
  part
• 4:17
  .
• 4:19
  The
• 4:19
  clean
• 4:20
  version
• 4:20
  of
• 4:20
  the
• 4:20
  reseller
• 4:20
  scam
• 4:21
  is
• 4:21
  old
• 4:21
  news;
• 4:22
  the
• 4:22
  named,
• 4:22
  criminally-framed
• 4:23
  version
• 4:23
  is
• 4:23
  new.
• 4:24
  NahCrofAI
• 4:25
  advertised
• 4:25
  every
• 4:25
  frontier
• 4:26
  model
• 4:26
  below
• 4:26
  the
• 4:27
  cheapest
• 4:27
  alternative,
• 4:28
  its
• 4:28
  founder
• 4:28
  claiming
• 4:29
  custom
• 4:29
  inference
• 4:29
  engines
• 4:30
  and
• 4:30
  blaming
• 4:30
  competitors'
• 4:31
  prices
• 4:31
  on
• 4:31
  skill
• 4:31
  issues.
• 4:32
  An
• 4:32
  investigation
• 4:33
  showed
• 4:33
  it
• 4:33
  was
• 4:33
  an
• 4:34
  OpenRouter
• 4:34
  wrapper
• 4:34
  that
• 4:34
  silently
• 4:35
  routed
• 4:35
  to
• 4:35
  weaker,
• 4:36
  cheaper
• 4:36
  models.
• 4:37
  A
• 4:37
  model
• 4:37
  sold
• 4:38
  as
• 4:38
  kimi-k3
• 4:38
  at
• 4:38
  two
• 4:39
  dollars
• 4:39
  in
• 4:39
  and
• 4:39
  ten
• 4:39
  out
• 4:40
  was
• 4:40
  served
• 4:40
  as
• 4:40
  GLM-5.3
• 4:41
  Flash,
• 4:42
  a
• 4:42
  13-fold
• 4:42
  markup
• 4:43
  on
• 4:43
  input
• 4:43
  and
• 4:43
  20-fold
• 4:44
  on
• 4:44
  output,
• 4:45
  and
• 4:45
  the
• 4:45
  house
• 4:45
  model
• 4:45
  family,
• 4:46
  pitched
• 4:46
  as
• 4:46
  in-house
• 4:47
  work,
• 4:47
  routed
• 4:48
  to
• 4:48
  GLM-5.2,
• 4:49
  Qwen,
• 4:49
  and
• 4:49
  Kimi.
• 4:51
  The
• 4:51
  physical
• 4:51
  claims
• 4:51
  did
• 4:51
  not
• 4:52
  survive
• 4:52
  contact
• 4:53
  either:
• 4:53
  running
• 4:53
  Kimi
• 4:54
  K3
• 4:55
  even
• 4:55
  at
• 4:55
  brutal
• 4:55
  two-bit
• 4:56
  quantization
• 4:56
  needs
• 4:56
  roughly
• 4:57
  802
• 4:58
  GiB,
• 4:58
  and
• 4:59
  the
• 4:59
  largest
• 4:59
  rental
• 4:59
  box
• 5:00
  available
• 5:00
  tops
• 5:00
  out
• 5:00
  near
• 5:01
  765
• 5:01
  GiB
• 5:02
  .
• 5:03
  What
• 5:03
  makes
• 5:03
  it
• 5:03
  a
• 5:03
  case
• 5:03
  study
• 5:04
  is
• 5:04
  the
• 5:04
  response
• 5:04
  pattern.
• 5:05
  Confronted,
• 5:06
  the
• 5:06
  operator
• 5:07
  denied
• 5:07
  everything,
• 5:08
  then
• 5:08
  across
• 5:08
  five
• 5:09
  attempts
• 5:09
  changed
• 5:09
  nothing
• 5:10
  except
• 5:10
  the
• 5:10
  fingerprints
• 5:11
  betraying
• 5:11
  OpenRouter
• 5:12
  underneath.
• 5:13
  Then
• 5:14
  came
• 5:14
  a
• 5:14
  fake
• 5:14
  blog
• 5:14
  post
• 5:15
  under
• 5:15
  the
• 5:15
  pretense
• 5:15
  of
• 5:15
  a
• 5:15
  new
• 5:16
  team
• 5:16
  taking
• 5:16
  over,
• 5:17
  and
• 5:17
  about
• 5:17
  three
• 5:17
  hours
• 5:18
  later
• 5:18
  the
• 5:18
  entire
• 5:19
  online
• 5:19
  presence
• 5:19
  wiped,
• 5:20
  domains
• 5:20
  and
• 5:20
  accounts
• 5:21
  vanishing.
• 5:22
  Screenshots
• 5:22
  preserved
• 5:23
  the
• 5:23
  operator
• 5:23
  admitting
• 5:24
  the
• 5:24
  fraud
• 5:24
  spanned
• 5:24
  the
• 5:24
  service's
• 5:25
  full
• 5:25
  two-year
• 5:25
  life
• 5:26
  and
• 5:26
  then
• 5:26
  asking
• 5:27
  the
• 5:27
  investigator
• 5:27
  to
• 5:27
  help
• 5:27
  cover
• 5:28
  it
• 5:28
  up.
• 5:29
  The
• 5:29
  guidance
• 5:29
  for
• 5:30
  anyone
• 5:30
  who
• 5:30
  bought
• 5:30
  credits
• 5:30
  is
• 5:30
  sober:
• 5:31
  assume
• 5:32
  every
• 5:32
  prompt
• 5:32
  was
• 5:32
  logged
• 5:33
  and
• 5:33
  mined
• 5:33
  for
• 5:33
  keys,
• 5:34
  rotate
• 5:34
  credentials,
• 5:35
  and
• 5:35
  open
• 5:35
  chargebacks.
• 5:37
  The
• 5:37
  trust
• 5:37
  problem
• 5:37
  is
• 5:38
  not
• 5:38
  only
• 5:38
  who
• 5:38
  bills
• 5:38
  you;
• 5:39
  it
• 5:39
  is
• 5:40
  what
• 5:40
  your
• 5:40
  local
• 5:40
  tools
• 5:40
  can
• 5:41
  reach.
• 5:42
  A
• 5:42
  proof-of-concept
• 5:42
  called
• 5:43
  not-a-mused
• 5:43
  targets
• 5:43
  the
• 5:44
  Muse
• 5:44
  dictation
• 5:44
  app
• 5:44
  through
• 5:45
  an
• 5:45
  undocumented
• 5:45
  setting
• 5:46
  that
• 5:46
  an
• 5:46
  unprivileged
• 5:47
  local
• 5:47
  process
• 5:47
  can
• 5:48
  rewrite.
• 5:49
  Redirect
• 5:49
  that
• 5:49
  endpoint
• 5:50
  and
• 5:50
  dictated
• 5:50
  prompts
• 5:51
  flow
• 5:51
  to
• 5:51
  an
• 5:51
  attacker-controlled
• 5:52
  server,
• 5:52
  opening
• 5:53
  the
• 5:53
  door
• 5:53
  to
• 5:53
  captured
• 5:53
  audio,
• 5:54
  prompt
• 5:54
  injection
• 5:55
  into
• 5:55
  Muse,
• 5:56
  and
• 5:56
  theft
• 5:56
  of
• 5:56
  the
• 5:56
  app's
• 5:56
  authentication
• 5:57
  material.
• 5:58
  This
• 5:59
  is
• 5:59
  a
• 5:59
  local
• 5:59
  attack
• 5:59
  that
• 6:00
  assumes
• 6:00
  code
• 6:00
  execution
• 6:01
  as
• 6:01
  the
• 6:01
  user
• 6:02
  already,
• 6:02
  so
• 6:03
  it
• 6:03
  is
• 6:03
  not
• 6:03
  a
• 6:03
  remote
• 6:04
  break.
• 6:05
  It
• 6:05
  matters
• 6:05
  because
• 6:05
  of
• 6:06
  amplification:
• 6:07
  an
• 6:07
  assistant
• 6:08
  granted
• 6:08
  broad
• 6:08
  access
• 6:09
  becomes
• 6:09
  a
• 6:09
  more
• 6:10
  valuable
• 6:10
  pivot
• 6:10
  than
• 6:11
  ordinary
• 6:11
  malware,
• 6:12
  since
• 6:12
  the
• 6:12
  app's
• 6:13
  access
• 6:13
  quietly
• 6:14
  becomes
• 6:14
  the
• 6:14
  attacker's.
• 6:15
  It
• 6:16
  is
• 6:16
  the
• 6:16
  same
• 6:16
  trust-boundary
• 6:17
  shape
• 6:17
  that
• 6:17
  turned
• 6:17
  agent
• 6:18
  config
• 6:18
  files
• 6:18
  into
• 6:19
  command-execution
• 6:19
  vectors
• 6:20
  .
• 6:22
  The
• 6:22
  idea
• 6:23
  that
• 6:23
  you
• 6:23
  do
• 6:23
  not
• 6:23
  need
• 6:23
  a
• 6:23
  special
• 6:24
  model
• 6:24
  to
• 6:24
  get
• 6:24
  calibrated
• 6:25
  classifications
• 6:26
  keeps
• 6:26
  proving
• 6:26
  itself,
• 6:27
  this
• 6:27
  week
• 6:27
  as
• 6:28
  a
• 6:28
  two-line
• 6:28
  recipe.
• 6:29
  Run
• 6:29
  any
• 6:30
  GGUF
• 6:30
  in
• 6:30
  llama.cpp
• 6:32
  with
• 6:32
  the
• 6:32
  next-token
• 6:33
  count
• 6:33
  set
• 6:33
  to
• 6:33
  one
• 6:33
  and
• 6:34
  the
• 6:34
  top-probabilities
• 6:34
  count
• 6:35
  to
• 6:35
  ten,
• 6:35
  disable
• 6:36
  reasoning,
• 6:36
  and
• 6:36
  ask
• 6:37
  for
• 6:37
  a
• 6:37
  single-token
• 6:38
  yes-or-no
• 6:38
  answer.
• 6:39
  The
• 6:39
  returned
• 6:40
  log-probabilities
• 6:40
  convert
• 6:41
  to
• 6:41
  a
• 6:41
  probability
• 6:41
  by
• 6:41
  exponentiation,
• 6:43
  so
• 6:43
  a
• 6:43
  spam
• 6:43
  call
• 6:44
  comes
• 6:44
  back
• 6:44
  as
• 6:44
  99.5%
• 6:45
  to
• 6:46
  0.5%
• 6:47
  rather
• 6:47
  than
• 6:47
  a
• 6:47
  bare
• 6:47
  label,
• 6:48
  and
• 6:48
  a
• 6:48
  clever
• 6:48
  extension
• 6:49
  packs
• 6:49
  several
• 6:50
  yes-no
• 6:50
  questions
• 6:50
  into
• 6:50
  one
• 6:51
  forward
• 6:51
  pass
• 6:51
  by
• 6:51
  mapping
• 6:52
  each
• 6:52
  answer
• 6:52
  bit
• 6:52
  to
• 6:52
  a
• 6:53
  token.
• 6:54
  The
• 6:54
  community
• 6:54
  reaction
• 6:54
  was
• 6:55
  appropriately
• 6:55
  mixed:
• 6:56
  one
• 6:56
  noted
• 6:56
  that
• 6:57
  many
• 6:57
  such
• 6:57
  calls
• 6:57
  can
• 6:58
  cost
• 6:58
  more
• 6:58
  than
• 6:58
  one
• 6:58
  JSON
• 6:59
  fill,
• 7:00
  and
• 7:00
  another
• 7:00
  bluntly
• 7:00
  recommended
• 7:01
  constrained
• 7:01
  decoding
• 7:02
  instead
• 7:02
  .
• 7:03
  That
• 7:03
  raw-logprob
• 7:04
  route
• 7:04
  is
• 7:04
  one
• 7:04
  of
• 7:04
  three
• 7:05
  competing
• 7:05
  for
• 7:05
  the
• 7:05
  same
• 7:06
  job,
• 7:06
  and
• 7:07
  the
• 7:07
  trained-encoder
• 7:07
  route
• 7:08
  just
• 7:08
  got
• 7:08
  a
• 7:08
  clean
• 7:08
  local
• 7:08
  port.
• 7:10
  Laya-MLX
• 7:11
  is
• 7:11
  an
• 7:11
  independent
• 7:11
  Apple
• 7:12
  Silicon
• 7:12
  reimplementation
• 7:13
  of
• 7:13
  Convai
• 7:13
  Innovations'
• 7:14
  Laya
• 7:14
  models,
• 7:15
  which
• 7:15
  answer
• 7:16
  a
• 7:16
  typed
• 7:16
  question
• 7:16
  in
• 7:17
  one
• 7:17
  bidirectional
• 7:17
  forward
• 7:18
  pass
• 7:18
  through
• 7:18
  an
• 7:18
  encoder
• 7:19
  and
• 7:19
  a
• 7:19
  decision
• 7:19
  head
• 7:20
  rather
• 7:20
  than
• 7:20
  generating
• 7:21
  tokens.
• 7:22
  The
• 7:22
  pitch
• 7:23
  is
• 7:23
  zero
• 7:23
  output
• 7:23
  tokens
• 7:24
  and
• 7:24
  no
• 7:24
  PyTorch,
• 7:25
  Transformers
• 7:26
  runtime,
• 7:27
  or
• 7:27
  cloud
• 7:27
  dependency.
• 7:29
  On
• 7:29
  an
• 7:29
  M3
• 7:29
  Max
• 7:30
  the
• 7:30
  421-million-parameter
• 7:31
  English
• 7:32
  model
• 7:32
  answers
• 7:32
  a
• 7:32
  short
• 7:33
  question
• 7:33
  in
• 7:33
  about
• 7:33
  13
• 7:34
  milliseconds,
• 7:35
  and
• 7:35
  the
• 7:35
  port
• 7:35
  matched
• 7:35
  the
• 7:35
  upstream
• 7:36
  selected
• 7:36
  answer
• 7:37
  on
• 7:37
  all
• 7:37
  63
• 7:37
  validation
• 7:38
  questions
• 7:38
  in
• 7:39
  both
• 7:39
  precisions.
• 7:40
  The
• 7:40
  author
• 7:41
  is
• 7:41
  careful
• 7:41
  about
• 7:41
  scope,
• 7:42
  flagging
• 7:42
  that
• 7:42
  this
• 7:43
  is
• 7:43
  unofficial
• 7:43
  and
• 7:43
  that
• 7:44
  calibration
• 7:44
  temperatures
• 7:45
  had
• 7:45
  to
• 7:45
  be
• 7:45
  clamped
• 7:45
  to
• 7:45
  stop
• 7:46
  a
• 7:46
  shipped
• 7:46
  bucket
• 7:46
  from
• 7:47
  reporting
• 7:47
  a
• 7:47
  coin
• 7:47
  flip
• 7:48
  as
• 7:48
  near-certainty.
• 7:49
  The
• 7:49
  honest
• 7:49
  read
• 7:50
  across
• 7:50
  all
• 7:50
  three
• 7:51
  approaches
• 7:51
  is
• 7:51
  that
• 7:52
  reading
• 7:52
  a
• 7:52
  capable
• 7:52
  model's
• 7:53
  own
• 7:53
  next-token
• 7:53
  distribution
• 7:54
  is
• 7:55
  often
• 7:55
  the
• 7:55
  more
• 7:55
  accurate
• 7:56
  and
• 7:56
  better-calibrated
• 7:56
  classifier,
• 7:57
  losing
• 7:58
  little
• 7:58
  to
• 7:58
  a
• 7:58
  bespoke
• 7:58
  model
• 7:59
  .
• 8:01
  For
• 8:01
  two
• 8:01
  years
• 8:01
  the
• 8:01
  recurring
• 8:02
  lesson
• 8:02
  here
• 8:02
  has
• 8:02
  been
• 8:03
  that
• 8:03
  releasing
• 8:03
  weights
• 8:04
  is
• 8:04
  not
• 8:04
  the
• 8:04
  same
• 8:04
  as
• 8:04
  granting
• 8:05
  access,
• 8:06
  since
• 8:06
  a
• 8:06
  frontier
• 8:06
  open
• 8:07
  model
• 8:07
  can
• 8:07
  still
• 8:07
  demand
• 8:08
  a
• 8:08
  datacenter.
• 8:09
  This
• 8:10
  week
• 8:10
  the
• 8:10
  story
• 8:10
  moved
• 8:10
  a
• 8:10
  layer
• 8:11
  down,
• 8:11
  to
• 8:11
  the
• 8:11
  infrastructure
• 8:12
  itself.
• 8:14
  Tim
• 8:14
  Dettmers,
• 8:15
  whose
• 8:15
  earlier
• 8:16
  claim
• 8:16
  that
• 8:16
  GLM
• 8:16
  could
• 8:17
  run
• 8:17
  on
• 8:17
  a
• 8:17
  single
• 8:17
  small
• 8:18
  box
• 8:18
  invited
• 8:18
  well-earned
• 8:19
  skepticism,
• 8:20
  announced
• 8:20
  his
• 8:20
  lab's
• 8:20
  Open
• 8:21
  Source
• 8:21
  Week:
• 8:22
  two
• 8:22
  projects
• 8:23
  and
• 8:23
  four
• 8:23
  papers
• 8:23
  from
• 8:23
  a
• 8:24
  small
• 8:24
  group
• 8:24
  with
• 8:24
  a
• 8:24
  couple
• 8:24
  of
• 8:25
  GPUs.
• 8:26
  The
• 8:26
  headline
• 8:27
  demos
• 8:27
  are
• 8:27
  an
• 8:27
  agent
• 8:28
  harness
• 8:28
  that
• 8:28
  optimized
• 8:29
  his
• 8:29
  framework's
• 8:29
  Metal
• 8:30
  backend
• 8:30
  to
• 8:30
  serve
• 8:31
  a
• 8:31
  35-billion-parameter
• 8:32
  Qwen
• 8:32
  model
• 8:32
  at
• 8:33
  450
• 8:34
  tokens
• 8:34
  per
• 8:34
  second
• 8:35
  at
• 8:35
  1.5
• 8:35
  bits
• 8:36
  per
• 8:36
  weight,
• 8:37
  and
• 8:37
  an
• 8:37
  inference
• 8:37
  framework
• 8:38
  he
• 8:38
  says
• 8:38
  runs
• 8:38
  a
• 8:38
  125-billion-parameter
• 8:40
  model
• 8:40
  on
• 8:41
  a
• 8:41
  single
• 8:41
  24
• 8:42
  GB
• 8:42
  GPU.
• 8:43
  Every
• 8:43
  figure
• 8:44
  is
• 8:44
  asserted
• 8:44
  in
• 8:44
  prose
• 8:44
  with
• 8:45
  no
• 8:45
  code
• 8:45
  linked
• 8:45
  yet,
• 8:45
  so
• 8:46
  the
• 8:46
  posture
• 8:46
  is
• 8:46
  interested
• 8:47
  but
• 8:47
  waiting;
• 8:47
  his
• 8:48
  claim
• 8:48
  that
• 8:48
  a
• 8:48
  couple
• 8:48
  of
• 8:49
  people
• 8:49
  with
• 8:49
  a
• 8:49
  couple
• 8:49
  of
• 8:49
  GPUs
• 8:50
  can
• 8:50
  compete
• 8:50
  with
• 8:50
  the
• 8:51
  frontier
• 8:51
  is
• 8:52
  a
• 8:52
  thesis
• 8:52
  to
• 8:52
  test,
• 8:53
  not
• 8:53
  a
• 8:53
  result
• 8:53
  to
• 8:53
  bank
• 8:54
  .
• 8:54
  The
• 8:54
  same
• 8:54
  recursive
• 8:55
  move,
• 8:55
  weights
• 8:56
  optimizing
• 8:56
  their
• 8:56
  own
• 8:56
  serving
• 8:57
  stack,
• 8:57
  is
• 8:57
  what
• 8:58
  Zhipu
• 8:58
  describes
• 8:58
  in
• 8:58
  bringing
• 8:59
  GLM-5.3-Flash
• 9:00
  to
• 9:00
  production.
• 9:01
  A
• 9:01
  GLM-5.3-powered
• 9:02
  agent
• 9:03
  did
• 9:03
  much
• 9:03
  of
• 9:03
  the
• 9:03
  systems
• 9:04
  work
• 9:04
  to
• 9:04
  deploy
• 9:04
  its
• 9:04
  own
• 9:04
  successor
• 9:05
  on
• 9:05
  a
• 9:05
  cluster
• 9:05
  of
• 9:05
  more
• 9:06
  than
• 9:06
  100,000
• 9:07
  Chinese-made
• 9:07
  accelerators,
• 9:08
  framed
• 9:08
  as
• 9:08
  an
• 9:08
  early
• 9:09
  form
• 9:09
  of
• 9:09
  recursive
• 9:09
  self-improvement
• 9:10
  while
• 9:10
  conceding
• 9:11
  it
• 9:11
  has
• 9:11
  not
• 9:11
  reached
• 9:11
  that
• 9:11
  bar.
• 9:12
  The
• 9:12
  concrete
• 9:12
  evidence
• 9:12
  is
• 9:12
  three
• 9:13
  checkable
• 9:13
  debugging
• 9:13
  cases,
• 9:14
  including
• 9:15
  a
• 9:15
  throughput
• 9:15
  regression
• 9:15
  traced
• 9:16
  to
• 9:16
  a
• 9:16
  dispatch
• 9:16
  call
• 9:16
  never
• 9:16
  releasing
• 9:17
  Python's
• 9:17
  global
• 9:18
  interpreter
• 9:18
  lock
• 9:18
  and
• 9:18
  a
• 9:18
  decode
• 9:19
  kernel
• 9:19
  rebuilt
• 9:20
  for
• 9:20
  a
• 9:20
  1.71-fold
• 9:21
  speedup.
• 9:21
  The
• 9:21
  self-reported
• 9:22
  usage
• 9:22
  figures,
• 9:22
  most-used
• 9:23
  model
• 9:23
  on
• 9:23
  two
• 9:23
  platforms
• 9:24
  within
• 9:24
  a
• 9:24
  week
• 9:24
  and
• 9:24
  62
• 9:25
  trillion
• 9:25
  tokens
• 9:26
  in
• 9:26
  six
• 9:26
  days,
• 9:26
  carry
• 9:27
  no
• 9:27
  third-party
• 9:27
  corroboration,
• 9:28
  but
• 9:28
  the
• 9:28
  engineering
• 9:29
  narrative
• 9:29
  is
• 9:29
  specific
• 9:30
  and
• 9:30
  the
• 9:30
  merges
• 9:30
  verifiable
• 9:30
  .
• 9:31
  Underneath
• 9:31
  sits
• 9:31
  the
• 9:31
  vendor-lock
• 9:32
  question,
• 9:32
  and
• 9:32
  one
• 9:33
  developer
• 9:33
  wrote
• 9:33
  a
• 9:33
  native
• 9:33
  Rust
• 9:34
  and
• 9:34
  Vulkan
• 9:34
  training
• 9:35
  backend
• 9:35
  covering
• 9:35
  143
• 9:36
  Transformer
• 9:37
  architectures
• 9:37
  with
• 9:37
  no
• 9:37
  CUDA
• 9:38
  or
• 9:38
  PyTorch.
• 9:39
  It
• 9:39
  implements
• 9:39
  forward
• 9:39
  and
• 9:40
  backward
• 9:40
  passes,
• 9:40
  AdamW,
• 9:41
  LoRA
• 9:41
  fine-tuning,
• 9:42
  and
• 9:42
  both
• 9:42
  dense
• 9:43
  and
• 9:43
  mixture-of-experts
• 9:43
  paths,
• 9:44
  tested
• 9:44
  on
• 9:45
  AMD
• 9:45
  hardware
• 9:45
  down
• 9:46
  to
• 9:46
  a
• 9:46
  handheld
• 9:46
  gaming
• 9:47
  chip.
• 9:47
  The
• 9:47
  author
• 9:47
  refuses
• 9:48
  to
• 9:48
  overclaim,
• 9:48
  stating
• 9:49
  plainly
• 9:49
  it
• 9:49
  is
• 9:49
  not
• 9:49
  yet
• 9:49
  a
• 9:50
  CUDA
• 9:50
  replacement
• 9:50
  .
• 9:51
  Jun
• 9:51
  Kim,
• 9:52
  creator
• 9:52
  of
• 9:52
  the
• 9:53
  oMLX
• 9:53
  framework,
• 9:54
  joined
• 9:54
  Hugging
• 9:54
  Face
• 9:55
  to
• 9:55
  maintain
• 9:55
  it
• 9:55
  full-time,
• 9:56
  graduating
• 9:57
  a
• 9:57
  side
• 9:57
  project
• 9:58
  to
• 9:58
  funded
• 9:58
  work
• 9:58
  while
• 9:59
  keeping
• 9:59
  its
• 9:59
  Apache
• 9:59
  license
• 10:00
  .
• 10:01
  And
• 10:01
  Thomas
• 10:01
  Wolf
• 10:01
  answered
• 10:02
  the
• 10:02
  recurring
• 10:02
  is-open-source-dying
• 10:03
  refrain
• 10:03
  not
• 10:04
  with
• 10:04
  argument
• 10:04
  but
• 10:04
  a
• 10:04
  list:
• 10:05
  roughly
• 10:06
  thirty
• 10:06
  notable
• 10:06
  open
• 10:07
  releases
• 10:07
  in
• 10:07
  ten
• 10:07
  weeks,
• 10:08
  several
• 10:08
  frontier-scale,
• 10:09
  including
• 10:10
  Kimi
• 10:10
  K3
• 10:10
  at
• 10:11
  2.8
• 10:11
  trillion
• 10:11
  parameters
• 10:12
  and
• 10:12
  Qwen's
• 10:12
  2.4-trillion
• 10:13
  Max.
• 10:14
  As
• 10:15
  one
• 10:15
  reply
• 10:15
  put
• 10:15
  it,
• 10:15
  dying
• 10:16
  quickly
• 10:16
  apparently
• 10:16
  means
• 10:17
  a
• 10:17
  new
• 10:17
  frontier
• 10:17
  model
• 10:18
  every
• 10:18
  ten
• 10:18
  days
• 10:18
  .
• 10:21
  Depth
• 10:21
  pruning,
• 10:22
  deleting
• 10:22
  whole
• 10:22
  transformer
• 10:23
  blocks,
• 10:23
  is
• 10:24
  one
• 10:24
  of
• 10:24
  the
• 10:24
  cheapest
• 10:24
  ways
• 10:25
  to
• 10:25
  speed
• 10:25
  up
• 10:25
  a
• 10:25
  model,
• 10:25
  because
• 10:26
  it
• 10:26
  yields
• 10:26
  predictable
• 10:26
  savings
• 10:27
  and
• 10:27
  stacks
• 10:27
  cleanly
• 10:28
  with
• 10:28
  quantization.
• 10:29
  Multiverse
• 10:29
  Computing,
• 10:30
  last
• 10:30
  seen
• 10:30
  compressing
• 10:31
  DeepSeek
• 10:31
  with
• 10:31
  quantum-inspired
• 10:32
  tensor
• 10:32
  networks,
• 10:33
  returns
• 10:33
  with
• 10:33
  a
• 10:33
  sharp
• 10:34
  reframing:
• 10:35
  choosing
• 10:35
  which
• 10:35
  blocks
• 10:35
  to
• 10:35
  cut
• 10:36
  is
• 10:36
  a
• 10:36
  combinatorial
• 10:36
  problem,
• 10:37
  not
• 10:37
  a
• 10:37
  ranking
• 10:37
  one,
• 10:38
  because
• 10:38
  removing
• 10:39
  one
• 10:39
  block's
• 10:39
  cost
• 10:39
  depends
• 10:40
  on
• 10:40
  which
• 10:40
  others
• 10:40
  go
• 10:40
  with
• 10:40
  it.
• 10:41
  Their
• 10:41
  method
• 10:41
  assigns
• 10:41
  each
• 10:42
  block
• 10:42
  a
• 10:42
  binary
• 10:42
  keep-or-remove
• 10:43
  variable
• 10:43
  and
• 10:44
  uses
• 10:44
  a
• 10:44
  second-order
• 10:44
  expansion
• 10:45
  of
• 10:45
  the
• 10:45
  loss,
• 10:45
  whose
• 10:46
  off-diagonal
• 10:46
  terms
• 10:46
  capture
• 10:47
  exactly
• 10:47
  the
• 10:47
  pairwise
• 10:48
  couplings
• 10:48
  that
• 10:48
  magnitude
• 10:48
  and
• 10:49
  influence
• 10:49
  heuristics
• 10:49
  throw
• 10:49
  away.
• 10:50
  Selecting
• 10:51
  which
• 10:51
  blocks
• 10:51
  to
• 10:51
  drop
• 10:52
  becomes
• 10:52
  finding
• 10:52
  the
• 10:52
  low-energy
• 10:53
  states
• 10:53
  of
• 10:53
  an
• 10:53
  Ising
• 10:54
  spin
• 10:54
  glass,
• 10:55
  computed
• 10:55
  once
• 10:55
  and
• 10:56
  reused
• 10:56
  across
• 10:56
  targets,
• 10:57
  solved
• 10:57
  by
• 10:57
  brute
• 10:58
  force
• 10:58
  where
• 10:58
  feasible
• 10:59
  or
• 10:59
  an
• 10:59
  open
• 10:59
  tabu
• 10:59
  solver
• 11:00
  where
• 11:00
  not
• 11:00
  .
• 11:00
  The
• 11:01
  results
• 11:01
  are
• 11:01
  strongest
• 11:01
  where
• 11:02
  they
• 11:02
  matter
• 11:02
  most,
• 11:02
  at
• 11:02
  aggressive
• 11:03
  compression.
• 11:04
  Removing
• 11:04
  half
• 11:04
  of
• 11:05
  a
• 11:05
  70-billion-parameter
• 11:06
  Llama's
• 11:06
  depth,
• 11:07
  the
• 11:07
  method
• 11:07
  holds
• 11:07
  a
• 11:07
  key
• 11:08
  knowledge
• 11:08
  benchmark
• 11:08
  near
• 11:09
  77
• 11:10
  while
• 11:10
  the
• 11:10
  strongest
• 11:10
  baseline
• 11:11
  falls
• 11:11
  to
• 11:11
  the
• 11:11
  mid-50s,
• 11:12
  a
• 11:12
  gap
• 11:12
  of
• 11:13
  almost
• 11:13
  23
• 11:13
  points,
• 11:14
  beating
• 11:15
  that
• 11:15
  baseline
• 11:15
  on
• 11:15
  every
• 11:16
  benchmark
• 11:16
  tested.
• 11:17
  A
• 11:17
  telling
• 11:17
  detail
• 11:18
  undercuts
• 11:18
  the
• 11:18
  folk
• 11:19
  wisdom
• 11:19
  that
• 11:19
  the
• 11:19
  best
• 11:20
  cut
• 11:20
  is
• 11:20
  one
• 11:20
  consecutive
• 11:21
  chunk
• 11:21
  of
• 11:21
  middle
• 11:21
  layers:
• 11:22
  on
• 11:22
  a
• 11:23
  smaller
• 11:23
  model,
• 11:24
  the
• 11:24
  first
• 11:24
  configuration
• 11:25
  to
• 11:25
  remove
• 11:25
  an
• 11:25
  early
• 11:26
  block
• 11:26
  beat
• 11:26
  the
• 11:26
  supposed
• 11:27
  optimum
• 11:27
  after
• 11:27
  light
• 11:28
  retraining.
• 11:29
  The
• 11:29
  authors
• 11:29
  are
• 11:29
  candid
• 11:30
  that
• 11:30
  the
• 11:30
  spin-glass
• 11:31
  energy
• 11:31
  is
• 11:31
  a
• 11:31
  strong
• 11:32
  proxy
• 11:32
  but
• 11:32
  not
• 11:33
  a
• 11:33
  perfect
• 11:33
  one,
• 11:34
  and
• 11:34
  that
• 11:34
  several
• 11:34
  wins
• 11:34
  lean
• 11:35
  on
• 11:35
  retraining.
• 11:36
  The
• 11:36
  method
• 11:36
  meets
• 11:37
  a
• 11:37
  live
• 11:37
  target
• 11:37
  in
• 11:38
  altar-1,
• 11:39
  a
• 11:39
  third-party
• 11:39
  prune
• 11:40
  of
• 11:40
  the
• 11:40
  753-billion-parameter
• 11:41
  GLM-5.3.
• 11:43
  Using
• 11:44
  router-weighted
• 11:44
  expert
• 11:45
  activation
• 11:45
  pruning,
• 11:46
  the
• 11:46
  same
• 11:47
  family
• 11:47
  Cerebras
• 11:47
  applied
• 11:48
  to
• 11:48
  an
• 11:48
  earlier
• 11:48
  GLM
• 11:49
  two
• 11:49
  generations
• 11:50
  back,
• 11:50
  it
• 11:51
  deletes
• 11:51
  34%
• 11:52
  of
• 11:52
  the
• 11:52
  experts
• 11:52
  and
• 11:52
  quantizes
• 11:53
  the
• 11:53
  survivors
• 11:54
  to
• 11:54
  four
• 11:54
  bits,
• 11:55
  landing
• 11:55
  at
• 11:55
  328
• 11:56
  GB
• 11:57
  to
• 11:57
  serve
• 11:57
  on
• 11:57
  four
• 11:57
  H200s,
• 11:59
  with
• 11:59
  routing
• 11:59
  untouched.
• 12:00
  The
• 12:01
  fidelity
• 12:01
  claim
• 12:01
  is
• 12:01
  a
• 12:01
  Kullback-Leibler
• 12:02
  divergence
• 12:03
  of
• 12:03
  0.506
• 12:04
  nats
• 12:05
  against
• 12:05
  the
• 12:05
  full-precision
• 12:06
  model
• 12:06
  on
• 12:06
  a
• 12:06
  sealed
• 12:06
  panel,
• 12:07
  with
• 12:08
  the
• 12:08
  calibration
• 12:08
  set
• 12:08
  weighted
• 12:09
  toward
• 12:09
  cybersecurity
• 12:10
  traces,
• 12:10
  coding,
• 12:11
  and
• 12:11
  tool
• 12:11
  calling.
• 12:12
  The
• 12:13
  standing
• 12:13
  caveat
• 12:13
  applies
• 12:14
  unchanged:
• 12:15
  a
• 12:15
  single
• 12:15
  divergence
• 12:16
  number
• 12:16
  on
• 12:16
  a
• 12:16
  sealed
• 12:17
  panel
• 12:17
  is
• 12:17
  suggestive,
• 12:18
  and
• 12:18
  independent
• 12:19
  large-scale
• 12:19
  validation
• 12:20
  would
• 12:20
  matter
• 12:20
  more
• 12:20
  than
• 12:20
  any
• 12:21
  anecdotal
• 12:21
  eval
• 12:21
  .
• 12:24
  If
• 12:24
  you
• 12:24
  want
• 12:24
  agents
• 12:25
  to
• 12:25
  do
• 12:25
  real
• 12:25
  work,
• 12:25
  the
• 12:25
  interesting
• 12:26
  question
• 12:26
  is
• 12:26
  the
• 12:26
  harness,
• 12:27
  not
• 12:27
  the
• 12:27
  model,
• 12:28
  and
• 12:28
  Workspace-OS
• 12:28
  is
• 12:28
  an
• 12:29
  unusually
• 12:29
  complete
• 12:29
  answer.
• 12:30
  It
• 12:31
  is
• 12:31
  a
• 12:31
  macOS
• 12:31
  desktop
• 12:32
  that
• 12:32
  folds
• 12:32
  a
• 12:32
  manager's
• 12:32
  whole
• 12:33
  day
• 12:33
  into
• 12:33
  one
• 12:33
  window,
• 12:34
  editing
• 12:34
  real
• 12:35
  Office
• 12:35
  files
• 12:35
  through
• 12:36
  an
• 12:36
  embedded
• 12:36
  LibreOffice
• 12:37
  engine,
• 12:37
  with
• 12:37
  AI
• 12:38
  agents
• 12:38
  working
• 12:38
  the
• 12:38
  same
• 12:39
  files
• 12:39
  in
• 12:39
  the
• 12:39
  same
• 12:39
  folders.
• 12:40
  Two
• 12:41
  ideas
• 12:41
  carry
• 12:41
  it.
• 12:42
  Live
• 12:42
  transclusion
• 12:43
  lets
• 12:43
  a
• 12:43
  value
• 12:43
  live
• 12:44
  in
• 12:44
  one
• 12:44
  spreadsheet
• 12:44
  cell
• 12:45
  and
• 12:45
  flow
• 12:45
  into
• 12:45
  every
• 12:45
  document
• 12:46
  and
• 12:46
  email
• 12:46
  that
• 12:47
  quotes
• 12:47
  it
• 12:47
  while
• 12:47
  still
• 12:47
  writing
• 12:48
  a
• 12:48
  real
• 12:48
  literal,
• 12:48
  so
• 12:49
  a
• 12:49
  broken
• 12:49
  link
• 12:49
  shows
• 12:49
  a
• 12:49
  placeholder,
• 12:50
  not
• 12:50
  a
• 12:50
  wrong
• 12:50
  number.
• 12:51
  And
• 12:51
  the
• 12:51
  reviewable
• 12:52
  agent
• 12:52
  starts
• 12:52
  every
• 12:52
  run
• 12:53
  with
• 12:53
  a
• 12:53
  shadow-git
• 12:53
  checkpoint
• 12:54
  and
• 12:54
  lands
• 12:54
  the
• 12:54
  result
• 12:55
  as
• 12:55
  a
• 12:55
  card
• 12:55
  showing
• 12:55
  the
• 12:56
  diff,
• 12:56
  cost,
• 12:56
  and
• 12:56
  turn
• 12:56
  count,
• 12:57
  kept
• 12:57
  or
• 12:57
  reverted
• 12:58
  in
• 12:58
  one
• 12:58
  click.
• 12:59
  It
• 12:59
  runs
• 12:59
  agents
• 13:00
  on
• 13:00
  your
• 13:00
  Claude
• 13:00
  subscription
• 13:01
  with
• 13:01
  no
• 13:01
  API
• 13:01
  key
• 13:02
  or
• 13:02
  telemetry,
• 13:03
  and
• 13:03
  enforces
• 13:03
  the
• 13:03
  full
• 13:04
  Electron
• 13:04
  hardening
• 13:05
  checklist.
• 13:06
  It
• 13:06
  answers
• 13:06
  a
• 13:06
  real
• 13:07
  critique,
• 13:07
  that
• 13:07
  computer-use
• 13:08
  agents
• 13:08
  waste
• 13:09
  their
• 13:09
  budget
• 13:09
  fighting
• 13:10
  the
• 13:10
  interfaces
• 13:10
  they
• 13:10
  should
• 13:10
  use,
• 13:11
  by
• 13:11
  giving
• 13:11
  the
• 13:11
  agent
• 13:12
  scoped
• 13:12
  tools
• 13:13
  and
• 13:13
  a
• 13:13
  reviewable
• 13:13
  trail
• 13:13
  instead
• 13:14
  .
• 13:14
  Whether
• 13:15
  more
• 13:15
  agents
• 13:15
  beat
• 13:15
  one
• 13:16
  good
• 13:16
  agent
• 13:16
  got
• 13:16
  a
• 13:16
  rigorous
• 13:17
  test
• 13:17
  from
• 13:17
  Berkeley
• 13:17
  and
• 13:17
  Microsoft
• 13:18
  Research:
• 13:19
  do
• 13:19
  agents
• 13:19
  that
• 13:19
  communicate
• 13:20
  at
• 13:20
  test
• 13:20
  time
• 13:20
  beat
• 13:21
  running
• 13:21
  many
• 13:21
  independent
• 13:22
  attempts
• 13:22
  and
• 13:23
  keeping
• 13:23
  the
• 13:23
  best?
• 13:24
  Their
• 13:24
  protocol
• 13:24
  is
• 13:24
  minimal:
• 13:25
  identical
• 13:25
  agents
• 13:26
  with
• 13:26
  no
• 13:26
  orchestrator,
• 13:27
  sharing
• 13:27
  an
• 13:27
  append-only
• 13:28
  findings
• 13:28
  log,
• 13:29
  adopting
• 13:29
  a
• 13:29
  peer's
• 13:30
  approach
• 13:30
  only
• 13:30
  after
• 13:31
  reproducing
• 13:31
  a
• 13:31
  measured
• 13:31
  improvement.
• 13:33
  On
• 13:33
  a
• 13:33
  hard
• 13:33
  reasoning
• 13:34
  suite,
• 13:34
  a
• 13:34
  five-agent
• 13:35
  team
• 13:35
  solved
• 13:36
  8.0%
• 13:36
  of
• 13:37
  games
• 13:37
  against
• 13:37
  2.2%
• 13:38
  for
• 13:38
  the
• 13:39
  best
• 13:39
  of
• 13:39
  five
• 13:39
  independent
• 13:40
  runs;
• 13:41
  matching
• 13:41
  the
• 13:41
  team's
• 13:41
  rate
• 13:42
  would
• 13:42
  take
• 13:42
  33
• 13:42
  solo
• 13:43
  agents.
• 13:44
  The
• 13:44
  gains
• 13:44
  are
• 13:44
  conditional:
• 13:45
  teams
• 13:46
  pay
• 13:46
  a
• 13:46
  coordination
• 13:46
  tax
• 13:47
  and
• 13:47
  lose
• 13:47
  to
• 13:47
  independent
• 13:48
  sampling
• 13:48
  below
• 13:48
  roughly
• 13:49
  400,000
• 13:50
  output
• 13:50
  tokens
• 13:50
  per
• 13:51
  agent,
• 13:51
  and
• 13:51
  where
• 13:52
  public
• 13:52
  feedback
• 13:52
  does
• 13:53
  not
• 13:53
  reliably
• 13:53
  rank
• 13:53
  intermediate
• 13:54
  states
• 13:54
  the
• 13:54
  team
• 13:55
  trailed
• 13:55
  independent
• 13:56
  pairs.
• 13:57
  The
• 13:57
  authors
• 13:57
  name
• 13:57
  the
• 13:58
  mechanism
• 13:58
  verified
• 13:58
  progress
• 13:59
  sharing
• 13:59
  and
• 13:59
  note
• 14:00
  that
• 14:00
  whether
• 14:00
  it
• 14:00
  holds
• 14:00
  when
• 14:01
  feedback
• 14:01
  is
• 14:01
  sparse
• 14:02
  or
• 14:02
  subjective
• 14:02
  remains
• 14:03
  open
• 14:03
  .
• 14:04
  The
• 14:04
  lower-level
• 14:04
  plumbing
• 14:05
  is
• 14:05
  filling
• 14:05
  in
• 14:05
  too.
• 14:06
  LSAP
• 14:07
  proposes
• 14:08
  a
• 14:08
  protocol
• 14:08
  that
• 14:09
  turns
• 14:09
  a
• 14:09
  language
• 14:09
  server's
• 14:10
  atomic
• 14:10
  operations
• 14:11
  into
• 14:11
  agent-facing
• 14:12
  cognitive
• 14:12
  ones,
• 14:13
  so
• 14:13
  a
• 14:13
  single
• 14:14
  find-references
• 14:15
  request
• 14:15
  triggers
• 14:15
  the
• 14:16
  whole
• 14:16
  chain
• 14:16
  of
• 14:16
  symbol
• 14:17
  localization
• 14:17
  and
• 14:18
  context
• 14:18
  extraction,
• 14:19
  sparing
• 14:20
  the
• 14:20
  agent
• 14:20
  a
• 14:20
  dozen
• 14:20
  scripted
• 14:21
  round-trips.
• 14:22
  It
• 14:22
  is
• 14:23
  early,
• 14:23
  at
• 14:23
  an
• 14:23
  alpha
• 14:24
  spec,
• 14:24
  and
• 14:25
  honest
• 14:25
  that
• 14:25
  it
• 14:25
  is
• 14:25
  an
• 14:25
  orchestration
• 14:26
  layer
• 14:26
  over
• 14:26
  the
• 14:27
  existing
• 14:27
  standard
• 14:28
  rather
• 14:28
  than
• 14:28
  a
• 14:28
  replacement
• 14:28
  .
• 14:29
  IWE,
• 14:29
  complementing
• 14:30
  it,
• 14:30
  turns
• 14:31
  a
• 14:31
  directory
• 14:31
  of
• 14:31
  markdown
• 14:32
  files
• 14:32
  into
• 14:32
  a
• 14:33
  knowledge
• 14:33
  graph
• 14:33
  that
• 14:33
  both
• 14:33
  a
• 14:34
  human
• 14:34
  editor
• 14:34
  and
• 14:34
  an
• 14:35
  AI
• 14:35
  agent
• 14:35
  query,
• 14:36
  with
• 14:36
  no
• 14:36
  built-in
• 14:37
  model
• 14:37
  of
• 14:37
  its
• 14:37
  own,
• 14:38
  over
• 14:38
  the
• 14:38
  Model
• 14:38
  Context
• 14:39
  Protocol.
• 14:40
  Its
• 14:40
  best
• 14:40
  idea
• 14:41
  is
• 14:41
  treating
• 14:41
  agent
• 14:41
  writes
• 14:42
  as
• 14:42
  checked
• 14:42
  rather
• 14:42
  than
• 14:43
  trusted:
• 14:43
  every
• 14:44
  mutation
• 14:44
  must
• 14:45
  declare
• 14:45
  its
• 14:45
  blast
• 14:45
  radius,
• 14:46
  and
• 14:47
  an
• 14:47
  edit
• 14:47
  that
• 14:47
  will
• 14:47
  not
• 14:47
  declare
• 14:48
  it
• 14:48
  is
• 14:48
  refused,
• 14:49
  the
• 14:49
  same
• 14:50
  reviewable-action
• 14:50
  instinct
• 14:51
  as
• 14:51
  Workspace-OS
• 14:52
  applied
• 14:52
  to
• 14:52
  memory
• 14:53
  .
• 14:55
  The
• 14:55
  neat
• 14:56
  theoretical
• 14:56
  hook
• 14:56
  today:
• 14:57
  a
• 14:57
  language
• 14:58
  model
• 14:58
  can
• 14:58
  be
• 14:58
  nothing
• 14:59
  but
• 14:59
  the
• 14:59
  compressor
• 15:00
  already
• 15:00
  on
• 15:01
  your
• 15:01
  machine.
• 15:02
  From
• 15:02
  the
• 15:02
  identity
• 15:03
  that
• 15:03
  every
• 15:03
  prediction
• 15:03
  model
• 15:04
  is
• 15:04
  a
• 15:04
  compressor
• 15:05
  and
• 15:05
  every
• 15:05
  compressor
• 15:06
  a
• 15:06
  prediction
• 15:06
  model,
• 15:07
  one
• 15:07
  experiment
• 15:08
  primes
• 15:08
  gzip
• 15:09
  with
• 15:09
  a
• 15:09
  corpus
• 15:09
  and
• 15:10
  generates
• 15:10
  text
• 15:10
  by
• 15:10
  searching
• 15:11
  for
• 15:11
  the
• 15:11
  byte
• 15:11
  sequences
• 15:12
  that
• 15:12
  compress
• 15:12
  best,
• 15:13
  since
• 15:13
  a
• 15:13
  continuation
• 15:14
  echoing
• 15:15
  the
• 15:15
  window
• 15:15
  costs
• 15:15
  almost
• 15:16
  nothing
• 15:16
  to
• 15:16
  encode.
• 15:17
  Picking
• 15:17
  the
• 15:17
  single
• 15:18
  best
• 15:18
  next
• 15:18
  byte
• 15:18
  fails
• 15:19
  because
• 15:19
  gzip
• 15:19
  reports
• 15:20
  only
• 15:20
  integer
• 15:20
  lengths,
• 15:21
  so
• 15:21
  a
• 15:21
  beam
• 15:21
  search
• 15:22
  over
• 15:22
  spans
• 15:22
  is
• 15:22
  needed,
• 15:23
  and
• 15:23
  priming
• 15:23
  on
• 15:23
  a
• 15:23
  small
• 15:24
  Shakespeare
• 15:24
  corpus
• 15:24
  produces
• 15:25
  output
• 15:25
  that
• 15:25
  is
• 15:26
  not
• 15:26
  coherent
• 15:26
  but
• 15:26
  clearly
• 15:27
  captures
• 15:27
  something
• 15:28
  of
• 15:28
  the
• 15:28
  source.
• 15:29
  It
• 15:29
  is
• 15:29
  the
• 15:29
  empirical,
• 15:30
  non-neural
• 15:30
  counterpoint
• 15:31
  to
• 15:31
  the
• 15:31
  thesis
• 15:31
  that
• 15:31
  learning
• 15:32
  is
• 15:32
  just
• 15:32
  compression:
• 15:33
  no
• 15:33
  weights
• 15:34
  and
• 15:34
  no
• 15:34
  training,
• 15:35
  only
• 15:35
  DEFLATE's
• 15:36
  32-kilobyte
• 15:36
  window
• 15:37
  doing
• 15:37
  the
• 15:37
  work
• 15:37
  .
• 15:38
  From
• 15:38
  bytes
• 15:39
  to
• 15:39
  surfaces,
• 15:40
  a
• 15:40
  graphics
• 15:40
  result
• 15:41
  took
• 15:41
  a
• 15:41
  best-paper
• 15:41
  honor
• 15:42
  at
• 15:42
  this
• 15:42
  year's
• 15:42
  ECCV.
• 15:44
  Heat
• 15:44
  Kernel
• 15:44
  Textures
• 15:45
  represent
• 15:45
  a
• 15:45
  mesh's
• 15:45
  appearance
• 15:46
  with
• 15:46
  anisotropic
• 15:47
  heat
• 15:47
  kernels,
• 15:48
  the
• 15:48
  geodesic
• 15:48
  counterparts
• 15:49
  of
• 15:49
  Gaussians,
• 15:50
  defined
• 15:50
  directly
• 15:51
  on
• 15:51
  the
• 15:51
  surface
• 15:51
  rather
• 15:52
  than
• 15:52
  through
• 15:52
  a
• 15:52
  UV
• 15:52
  atlas.
• 15:53
  The
• 15:53
  payoff
• 15:54
  is
• 15:54
  avoiding
• 15:54
  the
• 15:54
  seams,
• 15:55
  distortion,
• 15:56
  and
• 15:56
  wasted
• 15:57
  atlas
• 15:57
  space
• 15:57
  UV
• 15:57
  mapping
• 15:58
  imposes,
• 15:59
  with
• 15:59
  everything
• 15:59
  operating
• 16:00
  intrinsically
• 16:00
  on
• 16:00
  the
• 16:00
  mesh
• 16:01
  and
• 16:01
  fitting
• 16:01
  from
• 16:01
  a
• 16:01
  texture
• 16:02
  or
• 16:02
  from
• 16:02
  multiple
• 16:02
  views.
• 16:04
  The
• 16:04
  subtitle
• 16:04
  says
• 16:04
  it
• 16:04
  cleanly:
• 16:05
  geodesic
• 16:06
  Gaussians
• 16:06
  that
• 16:06
  do
• 16:06
  not
• 16:07
  splat
• 16:07
  .
• 16:08
  Underneath
• 16:08
  all
• 16:08
  of
• 16:08
  it
• 16:09
  runs
• 16:09
  the
• 16:09
  resource
• 16:09
  these
• 16:10
  systems
• 16:10
  actually
• 16:10
  compete
• 16:11
  for,
• 16:11
  human
• 16:11
  attention,
• 16:12
  and
• 16:12
  a
• 16:12
  short
• 16:13
  essay
• 16:13
  makes
• 16:13
  the
• 16:13
  case
• 16:13
  for
• 16:13
  reclaiming
• 16:14
  it.
• 16:15
  It
• 16:15
  starts
• 16:15
  from
• 16:15
  the
• 16:15
  Tetris
• 16:16
  effect,
• 16:16
  the
• 16:17
  finding
• 16:17
  that
• 16:17
  sustained
• 16:18
  focus
• 16:18
  reshapes
• 16:18
  perception,
• 16:19
  and
• 16:19
  argues
• 16:20
  the
• 16:20
  danger
• 16:20
  is
• 16:20
  that
• 16:21
  less
• 16:21
  and
• 16:21
  less
• 16:21
  of
• 16:21
  that
• 16:22
  focus
• 16:22
  is
• 16:22
  chosen,
• 16:23
  with
• 16:23
  feeds
• 16:23
  deciding
• 16:24
  what
• 16:24
  keeps
• 16:24
  people
• 16:24
  scrolling.
• 16:26
  The
• 16:26
  remedy
• 16:26
  it
• 16:26
  offers
• 16:27
  is
• 16:27
  not
• 16:27
  a
• 16:27
  detox
• 16:28
  app
• 16:28
  but
• 16:28
  a
• 16:28
  habit:
• 16:29
  return
• 16:29
  to
• 16:29
  bookmarks,
• 16:30
  RSS,
• 16:31
  and
• 16:31
  blogs,
• 16:32
  accept
• 16:32
  a
• 16:32
  slower
• 16:32
  internet
• 16:33
  where
• 16:33
  content
• 16:33
  is
• 16:33
  finite,
• 16:34
  and
• 16:34
  keep
• 16:34
  at
• 16:35
  it
• 16:35
  until
• 16:35
  something
• 16:35
  clicks.
• 16:36
  It
• 16:36
  is
• 16:36
  a
• 16:36
  fitting
• 16:37
  close
• 16:37
  to
• 16:37
  an
• 16:37
  update
• 16:37
  whose
• 16:38
  tools
• 16:38
  all
• 16:38
  promise
• 16:38
  to
• 16:38
  manage
• 16:39
  attention
• 16:39
  on
• 16:39
  the
• 16:39
  user's
• 16:40
  behalf,
• 16:40
  when
• 16:41
  the
• 16:41
  choice
• 16:41
  of
• 16:41
  what
• 16:41
  to
• 16:41
  look
• 16:42
  at
• 16:42
  was
• 16:42
  always
• 16:42
  supposed
• 16:43
  to
• 16:43
  belong
• 16:43
  to
• 16:43
  the
• 16:43
  reader
• 16:43
  .