Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 19, 2026

Transcript

AGI Dreams Podcast – May 19, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Recent
• 0:00
  work
• 0:00
  showed
• 0:00
  that
• 0:01
  a
• 0:01
  single
• 0:01
  neuron
• 0:01
  could
• 0:02
  bypass
• 0:02
  safety
• 0:02
  alignment.
• 0:04
  A
• 0:04
  new
• 0:04
  paper
• 0:04
  from
• 0:04
  Harvard
• 0:05
  and
• 0:05
  Princeton
• 0:05
  pushes
• 0:05
  that
• 0:06
  line
• 0:06
  of
• 0:06
  inquiry
• 0:06
  much
• 0:07
  further,
• 0:07
  and
• 0:07
  the
• 0:07
  results
• 0:08
  reframe
• 0:08
  how
• 0:09
  we
• 0:09
  should
• 0:09
  think
• 0:09
  about
• 0:09
  alignment
• 0:09
  altogether.
• 0:10
  Using
• 0:11
  targeted
• 0:11
  weight
• 0:11
  pruning
• 0:12
  as
• 0:12
  a
• 0:12
  causal
• 0:12
  probe,
• 0:13
  the
• 0:13
  researchers
• 0:13
  found
• 0:14
  that
• 0:14
  harmful
• 0:14
  content
• 0:14
  generation
• 0:15
  in
• 0:15
  LLMs
• 0:16
  depends
• 0:16
  on
• 0:16
  approximately
• 0:17
  0.0005%
• 0:19
  of
• 0:19
  total
• 0:19
  model
• 0:19
  parameters
• 0:20
  —
• 0:20
  a
• 0:20
  remarkably
• 0:21
  compact
• 0:21
  set
• 0:22
  of
• 0:22
  weights
• 0:22
  that
• 0:22
  is
• 0:22
  general
• 0:23
  across
• 0:23
  harm
• 0:23
  types
• 0:23
  and
• 0:23
  distinct
• 0:24
  from
• 0:24
  benign
• 0:24
  capabilities.
• 0:25
  Prune
• 0:26
  the
• 0:26
  weights
• 0:26
  identified
• 0:26
  from
• 0:27
  malware
• 0:27
  generation,
• 0:28
  and
• 0:28
  the
• 0:28
  model's
• 0:28
  capacity
• 0:29
  for
• 0:29
  hate
• 0:29
  speech,
• 0:30
  physical
• 0:30
  harm
• 0:30
  instructions,
• 0:31
  and
• 0:31
  privacy
• 0:32
  violations
• 0:32
  drops
• 0:33
  in
• 0:33
  lockstep.
• 0:34
  The
• 0:34
  cross-domain
• 0:34
  transfer
• 0:35
  is
• 0:35
  not
• 0:35
  subtle:
• 0:36
  it
• 0:36
  holds
• 0:36
  across
• 0:37
  every
• 0:37
  domain
• 0:37
  pair
• 0:38
  they
• 0:38
  tested,
• 0:38
  across
• 0:39
  Llama,
• 0:39
  Qwen,
• 0:39
  and
• 0:39
  OLMo
• 0:40
  model
• 0:40
  families.
• 0:41
  The
• 0:41
  key
• 0:41
  insight
• 0:42
  cuts
• 0:42
  against
• 0:42
  the
• 0:42
  prevailing
• 0:43
  narrative
• 0:43
  that
• 0:43
  alignment
• 0:44
  training
• 0:44
  is
• 0:44
  "just
• 0:44
  teaching
• 0:45
  models
• 0:45
  when
• 0:45
  to
• 0:45
  refuse."
• 0:47
  The
• 0:47
  OLMo
• 0:47
  training-stage
• 0:48
  progression
• 0:48
  is
• 0:48
  revealing:
• 0:50
  supervised
• 0:50
  fine-tuning
• 0:51
  introduces
• 0:51
  refusal
• 0:52
  behavior,
• 0:52
  but
• 0:53
  only
• 0:53
  after
• 0:53
  preference
• 0:53
  optimization
• 0:54
  (DPO/RL)
• 0:56
  does
• 0:57
  a
• 0:57
  genuinely
• 0:57
  compressed
• 0:57
  harmful
• 0:58
  generation
• 0:58
  mechanism
• 0:59
  emerge
• 0:59
  that
• 0:59
  can
• 0:59
  be
• 1:00
  separated
• 1:00
  from
• 1:00
  benign
• 1:00
  capabilities.
• 1:01
  Alignment,
• 1:02
  it
• 1:02
  turns
• 1:02
  out,
• 1:03
  does
• 1:03
  restructure
• 1:04
  model
• 1:04
  internals
• 1:04
  —
• 1:05
  the
• 1:05
  fragility
• 1:05
  everyone
• 1:06
  complains
• 1:06
  about
• 1:06
  is
• 1:06
  a
• 1:06
  property
• 1:07
  of
• 1:07
  the
• 1:07
  refusal
• 1:07
  gate,
• 1:08
  not
• 1:08
  the
• 1:08
  underlying
• 1:09
  mechanism.
• 1:10
  Jailbreaks
• 1:10
  bypass
• 1:11
  the
• 1:11
  gate;
• 1:11
  they
• 1:12
  do
• 1:12
  not
• 1:12
  restore
• 1:12
  the
• 1:12
  capacity
• 1:13
  that
• 1:13
  pruning
• 1:13
  removes.
• 1:14
  Compression
• 1:15
  is
• 1:15
  a
• 1:15
  double-edged
• 1:16
  sword.
• 1:16
  The
• 1:16
  same
• 1:17
  structural
• 1:17
  property
• 1:18
  that
• 1:18
  makes
• 1:18
  harmfulness
• 1:18
  tractable
• 1:19
  to
• 1:19
  target
• 1:19
  also
• 1:19
  explains
• 1:20
  emergent
• 1:20
  misalignment:
• 1:21
  fine-tuning
• 1:22
  that
• 1:22
  adjusts
• 1:23
  these
• 1:23
  unified
• 1:23
  weights
• 1:24
  in
• 1:24
  one
• 1:24
  domain
• 1:25
  necessarily
• 1:25
  propagates
• 1:26
  across
• 1:26
  all
• 1:26
  domains
• 1:27
  they
• 1:27
  support.
• 1:28
  But
• 1:29
  the
• 1:29
  researchers
• 1:29
  show
• 1:30
  that
• 1:30
  pruning
• 1:30
  the
• 1:30
  harmful
• 1:31
  generation
• 1:31
  weights
• 1:31
  substantially
• 1:32
  reduces
• 1:33
  emergent
• 1:33
  misalignment
• 1:34
  even
• 1:34
  when
• 1:34
  the
• 1:34
  pruning
• 1:35
  data
• 1:35
  comes
• 1:35
  from
• 1:35
  a
• 1:35
  different
• 1:36
  harm
• 1:36
  domain
• 1:36
  than
• 1:36
  the
• 1:36
  fine-tuning
• 1:37
  data.
• 1:38
  The
• 1:38
  pruned
• 1:39
  models
• 1:39
  retain
• 1:40
  full
• 1:40
  ability
• 1:40
  to
• 1:40
  detect
• 1:41
  and
• 1:41
  explain
• 1:41
  harmful
• 1:42
  content
• 1:42
  —
• 1:43
  they
• 1:43
  just
• 1:43
  cannot
• 1:43
  produce
• 1:44
  it.
• 1:44
  That
• 1:45
  dissociation
• 1:46
  has
• 1:46
  direct
• 1:46
  design
• 1:47
  implications
• 1:47
  for
• 1:47
  building
• 1:48
  systems
• 1:48
  that
• 1:48
  understand
• 1:49
  harm
• 1:49
  without
• 1:49
  being
• 1:50
  able
• 1:50
  to
• 1:50
  generate
• 1:50
  it.
• 1:51
  On
• 1:51
  the
• 1:51
  practical
• 1:52
  defense
• 1:52
  side,
• 1:53
  a
• 1:53
  developer
• 1:53
  trained
• 1:53
  Qwen3.5
• 1:56
  to
• 1:56
  jailbreak
• 1:57
  itself
• 1:57
  using
• 1:57
  reinforcement
• 1:58
  learning,
• 1:59
  then
• 1:59
  used
• 1:59
  the
• 1:59
  discovered
• 2:00
  attack
• 2:00
  patterns
• 2:01
  to
• 2:01
  harden
• 2:01
  its
• 2:01
  own
• 2:01
  defenses.
• 2:02
  The
• 2:03
  trick
• 2:03
  was
• 2:03
  rewarding
• 2:03
  attack
• 2:04
  diversity:
• 2:04
  without
• 2:05
  it,
• 2:05
  GRPO
• 2:06
  collapsed
• 2:06
  to
• 2:06
  the
• 2:07
  same
• 2:07
  fiction-writing
• 2:08
  jailbreak
• 2:08
  repeatedly.
• 2:09
  After
• 2:09
  clustering
• 2:10
  rollouts
• 2:10
  by
• 2:10
  tactic
• 2:11
  and
• 2:11
  dividing
• 2:12
  reward
• 2:12
  by
• 2:12
  cluster
• 2:13
  size,
• 2:13
  the
• 2:14
  attacker
• 2:14
  surfaced
• 2:14
  seven
• 2:15
  distinct
• 2:15
  tactic
• 2:16
  families,
• 2:17
  with
• 2:17
  fiction/creative
• 2:18
  framing
• 2:18
  being
• 2:19
  the
• 2:19
  largest
• 2:19
  at
• 2:19
  34%.
• 2:21
  Defense
• 2:21
  rate
• 2:21
  went
• 2:21
  from
• 2:22
  64%
• 2:23
  to
• 2:23
  92%,
• 2:24
  with
• 2:24
  benign
• 2:24
  accuracy
• 2:25
  dropping
• 2:25
  only
• 2:26
  from
• 2:26
  92%
• 2:27
  to
• 2:27
  88%.
• 2:29
  Automated
• 2:29
  adversarial
• 2:30
  red-teaming
• 2:30
  is
• 2:31
  becoming
• 2:31
  a
• 2:31
  repeatable
• 2:31
  methodology,
• 2:33
  not
• 2:33
  a
• 2:33
  research
• 2:33
  novelty.
• 2:34
  Meanwhile,
• 2:35
  the
• 2:35
  Linux
• 2:36
  kernel
• 2:36
  maintainers
• 2:37
  are
• 2:37
  drawing
• 2:37
  their
• 2:37
  own
• 2:37
  line
• 2:38
  in
• 2:38
  the
• 2:38
  sand
• 2:38
  on
• 2:38
  AI-generated
• 2:39
  content.
• 2:40
  Linus
• 2:41
  Torvalds
• 2:41
  acknowledged
• 2:42
  that
• 2:42
  AI
• 2:42
  slop
• 2:43
  vulnerability
• 2:43
  reports
• 2:44
  have
• 2:44
  gone
• 2:44
  from
• 2:44
  "2-3
• 2:45
  per
• 2:45
  week"
• 2:46
  to
• 2:46
  flooding
• 2:46
  the
• 2:46
  mailing
• 2:47
  list,
• 2:47
  and
• 2:48
  Willy
• 2:48
  Tarreau
• 2:48
  laid
• 2:48
  down
• 2:49
  ground
• 2:49
  rules:
• 2:50
  keep
• 2:50
  reports
• 2:50
  short
• 2:51
  and
• 2:51
  human-readable,
• 2:52
  strip
• 2:52
  markdown
• 2:52
  formatting,
• 2:53
  include
• 2:54
  a
• 2:54
  working
• 2:54
  reproducer
• 2:55
  you
• 2:55
  actually
• 2:55
  tested,
• 2:56
  propose
• 2:57
  and
• 2:57
  test
• 2:57
  a
• 2:57
  fix
• 2:57
  before
• 2:58
  reporting,
• 2:58
  and
• 2:59
  —
• 2:59
  critically
• 2:59
  —
• 3:00
  use
• 3:00
  judgment
• 3:01
  about
• 3:01
  whether
• 3:01
  a
• 3:01
  bug
• 3:01
  in
• 3:02
  dead
• 3:02
  PCMCIA
• 3:03
  code
• 3:03
  is
• 3:03
  worth
• 3:03
  anyone's
• 3:04
  time.
• 3:05
  The
• 3:05
  kernel
• 3:05
  team's
• 3:06
  response
• 3:06
  to
• 3:06
  AI-generated
• 3:07
  noise
• 3:08
  is
• 3:08
  the
• 3:08
  first
• 3:08
  real
• 3:09
  standard
• 3:09
  for
• 3:09
  what
• 3:09
  disclosure
• 3:10
  looks
• 3:10
  like
• 3:10
  when
• 3:10
  the
• 3:11
  submitter
• 3:11
  might
• 3:11
  be
• 3:11
  a
• 3:11
  language
• 3:12
  model.
• 3:16
  Synacktiv
• 3:16
  published
• 3:17
  the
• 3:17
  second
• 3:17
  part
• 3:17
  of
• 3:17
  their
• 3:18
  Tesla
• 3:18
  Wall
• 3:18
  Connector
• 3:18
  Gen
• 3:19
  3
• 3:19
  exploit
• 3:19
  chain,
• 3:20
  and
• 3:20
  it
• 3:20
  is
• 3:20
  a
• 3:20
  textbook
• 3:21
  time-of-check-to-time-of-use
• 3:22
  vulnerability
• 3:23
  in
• 3:23
  firmware
• 3:23
  anti-downgrade
• 3:24
  protection.
• 3:25
  After
• 3:25
  their
• 3:25
  Pwn2Own
• 3:26
  Automotive
• 3:26
  2025
• 3:27
  exploit
• 3:28
  relied
• 3:28
  on
• 3:28
  the
• 3:28
  absence
• 3:28
  of
• 3:29
  any
• 3:29
  anti-downgrade
• 3:30
  mechanism,
• 3:30
  Tesla
• 3:31
  shipped
• 3:31
  a
• 3:31
  firmware
• 3:31
  update
• 3:32
  adding
• 3:32
  a
• 3:32
  ratchet
• 3:32
  value:
• 3:33
  the
• 3:33
  updater
• 3:34
  refuses
• 3:34
  any
• 3:34
  image
• 3:35
  whose
• 3:35
  ratchet
• 3:35
  is
• 3:35
  lower
• 3:36
  than
• 3:36
  the
• 3:36
  stored
• 3:36
  one.
• 3:37
  Synacktiv
• 3:38
  bypassed
• 3:38
  it
• 3:39
  by
• 3:39
  abusing
• 3:39
  the
• 3:39
  order
• 3:39
  of
• 3:39
  operations
• 3:40
  between
• 3:40
  the
• 3:40
  partition
• 3:41
  table
• 3:41
  write
• 3:41
  and
• 3:42
  the
• 3:42
  slot
• 3:42
  erase
• 3:42
  in
• 3:42
  the
• 3:42
  validate
• 3:43
  and
• 3:43
  switch
• 3:43
  slot
• 3:43
  routine.
• 3:44
  The
• 3:45
  charger
• 3:45
  uses
• 3:45
  an
• 3:45
  A/B
• 3:46
  slot
• 3:46
  scheme:
• 3:47
  one
• 3:47
  active,
• 3:47
  one
• 3:47
  passive.
• 3:48
  The
• 3:49
  anti-downgrade
• 3:49
  check
• 3:50
  happens
• 3:50
  during
• 3:50
  slot
• 3:51
  validation,
• 3:51
  but
• 3:52
  the
• 3:52
  partition
• 3:52
  table
• 3:52
  is
• 3:53
  written
• 3:53
  before
• 3:53
  the
• 3:53
  passive
• 3:54
  slot
• 3:54
  is
• 3:54
  erased
• 3:54
  —
• 3:55
  creating
• 3:55
  a
• 3:55
  window
• 3:56
  where
• 3:56
  the
• 3:56
  attacker
• 3:57
  can
• 3:57
  replay
• 3:57
  the
• 3:57
  original
• 3:58
  vulnerable
• 3:58
  firmware.
• 3:59
  The
• 4:00
  exploit
• 4:00
  chain
• 4:00
  starts
• 4:01
  from
• 4:01
  the
• 4:01
  charging
• 4:01
  cable
• 4:01
  itself
• 4:02
  via
• 4:02
  UDS
• 4:03
  over
• 4:03
  Single-Wire
• 4:03
  CAN,
• 4:04
  using
• 4:04
  a
• 4:05
  trivially
• 4:05
  weak
• 4:05
  XOR-based
• 4:06
  Security
• 4:06
  Access
• 4:07
  authentication.
• 4:08
  The
• 4:08
  result:
• 4:09
  a
• 4:09
  fully
• 4:10
  up-to-date
• 4:10
  charger
• 4:10
  downgraded
• 4:11
  to
• 4:11
  vulnerable
• 4:12
  firmware,
• 4:13
  the
• 4:13
  original
• 4:13
  Pwn2Own
• 4:14
  attack
• 4:14
  replayed
• 4:15
  successfully.
• 4:16
  The
• 4:16
  researcher's
• 4:16
  aside
• 4:17
  —
• 4:18
  "this
• 4:18
  is
• 4:18
  one
• 4:18
  of
• 4:18
  those
• 4:18
  vulnerabilities
• 4:19
  you
• 4:19
  find
• 4:19
  by
• 4:19
  hand,
• 4:20
  with
• 4:20
  a
• 4:20
  coffee,
• 4:21
  an
• 4:21
  IDA
• 4:21
  window,
• 4:22
  and
• 4:22
  zero
• 4:23
  help
• 4:23
  from
• 4:23
  a
• 4:23
  language
• 4:23
  model"
• 4:24
  —
• 4:24
  lands
• 4:25
  differently
• 4:25
  after
• 4:25
  the
• 4:25
  kernel
• 4:26
  slop
• 4:26
  discussion
• 4:27
  above.
• 4:28
  Embedded
• 4:28
  firmware
• 4:28
  security
• 4:29
  remains
• 4:30
  a
• 4:30
  discipline
• 4:30
  where
• 4:30
  manual
• 4:31
  reverse
• 4:31
  engineering
• 4:32
  and
• 4:32
  careful
• 4:32
  reasoning
• 4:33
  about
• 4:33
  operation
• 4:34
  ordering
• 4:34
  still
• 4:34
  matters
• 4:35
  more
• 4:35
  than
• 4:35
  any
• 4:36
  automated
• 4:36
  tool.
• 4:41
  ByteDance
• 4:41
  released
• 4:42
  Lance,
• 4:42
  a
• 4:42
  3B-active-parameter
• 4:43
  model
• 4:44
  that
• 4:44
  handles
• 4:44
  image
• 4:44
  understanding,
• 4:45
  image
• 4:45
  generation,
• 4:46
  image
• 4:46
  editing,
• 4:47
  and
• 4:47
  video
• 4:47
  generation
• 4:48
  within
• 4:48
  a
• 4:48
  single
• 4:48
  unified
• 4:49
  framework.
• 4:50
  Trained
• 4:50
  from
• 4:50
  scratch
• 4:51
  on
• 4:51
  128
• 4:52
  A100s,
• 4:53
  it
• 4:53
  is
• 4:53
  the
• 4:53
  kind
• 4:53
  of
• 4:53
  model
• 4:54
  that
• 4:54
  would
• 4:54
  have
• 4:54
  been
• 4:54
  unthinkable
• 4:54
  at
• 4:54
  this
• 4:55
  parameter
• 4:55
  count
• 4:55
  a
• 4:55
  year
• 4:55
  ago.
• 4:56
  The
• 4:56
  catch
• 4:56
  —
• 4:57
  and
• 4:57
  the
• 4:57
  community
• 4:58
  caught
• 4:58
  it
• 4:58
  immediately
• 4:58
  —
• 4:59
  is
• 4:59
  that
• 5:00
  "3B
• 5:00
  active
• 5:01
  parameters"
• 5:01
  does
• 5:02
  not
• 5:02
  mean
• 5:02
  3B
• 5:02
  total:
• 5:03
  the
• 5:03
  model
• 5:03
  requires
• 5:04
  40GB+
• 5:05
  VRAM
• 5:05
  for
• 5:05
  inference,
• 5:06
  and
• 5:06
  the
• 5:06
  safetensors
• 5:07
  weigh
• 5:07
  in
• 5:07
  at
• 5:07
  53GB
• 5:08
  combined.
• 5:09
  Still,
• 5:10
  getting
• 5:10
  unified
• 5:10
  multimodal
• 5:11
  generation
• 5:12
  and
• 5:12
  understanding
• 5:12
  into
• 5:13
  a
• 5:13
  single
• 5:13
  architecture
• 5:13
  at
• 5:13
  this
• 5:14
  scale
• 5:14
  is
• 5:14
  a
• 5:14
  genuine
• 5:15
  milestone.
• 5:16
  The
• 5:16
  community's
• 5:17
  sharpest
• 5:17
  observation:
• 5:18
  nobody
• 5:18
  at
• 5:19
  this
• 5:19
  parameter
• 5:19
  range
• 5:19
  is
• 5:19
  going
• 5:20
  hard
• 5:20
  after
• 5:20
  coding,
• 5:21
  which
• 5:21
  is
• 5:21
  where
• 5:21
  the
• 5:21
  biggest
• 5:22
  commercial
• 5:22
  demand
• 5:22
  actually
• 5:23
  sits.
• 5:24
  Sapient
• 5:24
  Intelligence
• 5:25
  released
• 5:25
  HRM-Text
• 5:26
  1B,
• 5:26
  a
• 5:26
  hierarchical
• 5:27
  reasoning
• 5:27
  model
• 5:27
  trained
• 5:27
  on
• 5:28
  40
• 5:28
  billion
• 5:28
  tokens
• 5:28
  for
• 5:28
  approximately
• 5:29
  $1,000.
• 5:31
  It
• 5:31
  beats
• 5:31
  Llama
• 5:31
  3.2
• 5:32
  3B
• 5:33
  on
• 5:33
  MATH
• 5:34
  (56.2
• 5:34
  vs
• 5:35
  48.0)
• 5:36
  and
• 5:36
  DROP
• 5:36
  (82.2
• 5:37
  vs
• 5:38
  45.2)
• 5:39
  despite
• 5:39
  being
• 5:39
  one-third
• 5:40
  the
• 5:40
  parameter
• 5:40
  count.
• 5:41
  The
• 5:42
  architecture
• 5:42
  prioritizes
• 5:43
  reasoning
• 5:43
  over
• 5:43
  knowledge
• 5:44
  —
• 5:44
  a
• 5:44
  design
• 5:45
  bet
• 5:45
  that
• 5:45
  the
• 5:45
  benchmarks
• 5:45
  currently
• 5:46
  validate,
• 5:47
  though
• 5:47
  self-reported
• 5:47
  results
• 5:48
  from
• 5:48
  a
• 5:48
  new
• 5:48
  lab
• 5:48
  warrant
• 5:48
  the
• 5:49
  usual
• 5:49
  skepticism
• 5:49
  until
• 5:50
  independent
• 5:50
  reproduction.
• 5:52
  PaddleOCR
• 5:52
  3.5
• 5:53
  now
• 5:53
  runs
• 5:53
  on
• 5:54
  a
• 5:54
  standard
• 5:54
  HuggingFace
• 5:54
  Transformers
• 5:55
  backend,
• 5:56
  which
• 5:56
  matters
• 5:57
  more
• 5:57
  for
• 5:57
  adoption
• 5:57
  than
• 5:57
  for
• 5:57
  raw
• 5:58
  capability.
• 5:59
  PaddlePaddle's
• 6:00
  OCR
• 6:00
  and
• 6:00
  document
• 6:01
  parsing
• 6:01
  tools
• 6:01
  have
• 6:01
  been
• 6:02
  strong
• 6:02
  for
• 6:02
  years
• 6:03
  but
• 6:03
  locked
• 6:03
  inside
• 6:04
  Baidu's
• 6:04
  ecosystem.
• 6:05
  Bringing
• 6:05
  them
• 6:06
  into
• 6:06
  the
• 6:06
  Transformers
• 6:06
  pipeline
• 6:07
  means
• 6:07
  any
• 6:07
  team
• 6:08
  already
• 6:08
  using
• 6:08
  HuggingFace
• 6:09
  can
• 6:09
  drop
• 6:09
  in
• 6:09
  document
• 6:10
  intelligence
• 6:10
  without
• 6:10
  learning
• 6:11
  a
• 6:11
  new
• 6:11
  framework.
• 6:12
  For
• 6:12
  anyone
• 6:13
  building
• 6:13
  document
• 6:13
  processing
• 6:14
  pipelines,
• 6:14
  this
• 6:15
  is
• 6:15
  the
• 6:15
  kind
• 6:15
  of
• 6:15
  integration
• 6:16
  that
• 6:16
  quietly
• 6:16
  eliminates
• 6:16
  a
• 6:16
  week
• 6:17
  of
• 6:17
  plumbing
• 6:17
  work.
• 6:20
  A
• 6:20
  contributor
• 6:20
  fixed
• 6:21
  llama.cpp's
• 6:21
  tensor
• 6:22
  split
• 6:22
  behavior
• 6:22
  with
• 6:22
  quantized
• 6:23
  KV
• 6:23
  caches
• 6:24
  on
• 6:24
  dual-GPU
• 6:24
  setups,
• 6:25
  delivering
• 6:26
  a
• 6:26
  40%
• 6:26
  speedup
• 6:27
  with
• 6:27
  zero
• 6:27
  quality
• 6:27
  loss.
• 6:28
  The
• 6:29
  working
• 6:29
  fork
• 6:29
  also
• 6:29
  includes
• 6:30
  multi-token
• 6:30
  prediction
• 6:31
  (MTP)
• 6:31
  support.
• 6:32
  This
• 6:32
  is
• 6:32
  the
• 6:32
  kind
• 6:33
  of
• 6:33
  unsexy
• 6:33
  infrastructure
• 6:34
  work
• 6:34
  that
• 6:34
  turns
• 6:34
  budget
• 6:34
  multi-GPU
• 6:35
  setups
• 6:35
  from
• 6:36
  "technically
• 6:36
  possible
• 6:37
  but
• 6:37
  slower
• 6:37
  than
• 6:37
  single-GPU"
• 6:38
  into
• 6:38
  actually
• 6:39
  compelling.
• 6:40
  Glia
• 6:40
  is
• 6:40
  a
• 6:40
  local-first
• 6:41
  shared
• 6:41
  memory
• 6:41
  layer
• 6:42
  combining
• 6:42
  SQLite-vec
• 6:43
  for
• 6:43
  768-dimensional
• 6:44
  embeddings,
• 6:45
  FTS5
• 6:45
  for
• 6:46
  hybrid
• 6:46
  search,
• 6:46
  sentence-level
• 6:47
  context
• 6:47
  trimming
• 6:48
  (90-95%
• 6:49
  reduction),
• 6:51
  HyDE
• 6:51
  query
• 6:51
  expansion,
• 6:52
  PII
• 6:52
  redaction,
• 6:53
  and
• 6:53
  knowledge
• 6:53
  graph
• 6:53
  extraction
• 6:54
  —
• 6:55
  all
• 6:55
  offline,
• 6:56
  installable
• 6:56
  with
• 6:56
  a
• 6:56
  single
• 6:57
  npx
• 6:57
  command.
• 6:58
  It
• 6:58
  bridges
• 6:58
  web
• 6:59
  chat
• 6:59
  interfaces
• 7:00
  and
• 7:00
  local
• 7:00
  development
• 7:00
  tools,
• 7:01
  giving
• 7:02
  smaller
• 7:02
  models
• 7:02
  the
• 7:02
  retrieval
• 7:03
  infrastructure
• 7:03
  that
• 7:03
  until
• 7:04
  recently
• 7:04
  required
• 7:04
  a
• 7:04
  cloud
• 7:05
  vector
• 7:05
  database.
• 7:06
  TinySearch
• 7:07
  takes
• 7:07
  a
• 7:07
  complementary
• 7:07
  approach:
• 7:08
  a
• 7:08
  lightweight
• 7:09
  MCP
• 7:09
  tool
• 7:09
  for
• 7:10
  web
• 7:10
  search
• 7:10
  designed
• 7:11
  specifically
• 7:11
  for
• 7:12
  small
• 7:12
  local
• 7:12
  LLMs.
• 7:13
  DuckDuckGo
• 7:14
  queries
• 7:14
  plus
• 7:15
  Crawl4AI
• 7:15
  scraping
• 7:16
  plus
• 7:16
  BM25
• 7:17
  reranking,
• 7:17
  returning
• 7:18
  compact
• 7:18
  context
• 7:19
  instead
• 7:19
  of
• 7:19
  raw
• 7:19
  page
• 7:19
  dumps.
• 7:20
  End-to-end
• 7:21
  latency
• 7:21
  is
• 7:21
  5-12
• 7:22
  seconds
• 7:22
  on
• 7:22
  consumer
• 7:22
  hardware.
• 7:23
  The
• 7:23
  most
• 7:23
  speculative
• 7:24
  hardware
• 7:24
  story
• 7:24
  this
• 7:25
  week:
• 7:26
  Q.ANT's
• 7:26
  photonic
• 7:27
  GPU
• 7:27
  is
• 7:27
  now
• 7:27
  deployed
• 7:28
  at
• 7:28
  the
• 7:28
  Leibniz
• 7:28
  Supercomputing
• 7:29
  Centre,
• 7:29
  with
• 7:30
  claimed
• 7:30
  Gen
• 7:30
  1
• 7:30
  performance
• 7:31
  at
• 7:31
  50x
• 7:31
  and
• 7:32
  energy
• 7:32
  efficiency
• 7:32
  at
• 7:32
  30x
• 7:33
  over
• 7:33
  transistor
• 7:34
  equivalents.
• 7:35
  Gen
• 7:35
  2
• 7:35
  targets
• 7:35
  100x
• 7:36
  and
• 7:36
  90x.
• 7:37
  They
• 7:37
  just
• 7:37
  opened
• 7:37
  an
• 7:37
  Austin
• 7:38
  office
• 7:38
  with
• 7:38
  IBM's
• 7:38
  former
• 7:39
  CTO.
• 7:40
  If
• 7:40
  the
• 7:40
  numbers
• 7:40
  hold
• 7:41
  under
• 7:41
  independent
• 7:41
  benchmarking,
• 7:42
  photonic
• 7:43
  computing
• 7:43
  moves
• 7:43
  from
• 7:44
  "interesting
• 7:44
  physics
• 7:45
  paper"
• 7:45
  to
• 7:45
  "actual
• 7:45
  alternative
• 7:46
  to
• 7:46
  NVIDIA."
• 7:47
  That
• 7:47
  is
• 7:48
  a
• 7:48
  very
• 7:48
  large
• 7:48
  "if."
• 7:49
  On
• 7:49
  the
• 7:49
  software
• 7:50
  side,
• 7:50
  a
• 7:50
  developer
• 7:51
  built
• 7:51
  folder-scoped
• 7:52
  memory
• 7:52
  isolation
• 7:52
  for
• 7:53
  Open
• 7:53
  WebUI
• 7:53
  —
• 7:54
  automatically
• 7:55
  tagging
• 7:55
  memories
• 7:55
  per
• 7:55
  project,
• 7:56
  with
• 7:56
  smart
• 7:57
  deduplication
• 7:57
  and
• 7:57
  orphan
• 7:58
  cleanup.
• 7:59
  It
• 7:59
  solves
• 7:59
  a
• 7:59
  real
• 8:00
  context
• 8:00
  contamination
• 8:01
  problem
• 8:02
  anyone
• 8:02
  running
• 8:02
  multiple
• 8:03
  projects
• 8:03
  through
• 8:03
  the
• 8:04
  same
• 8:04
  interface
• 8:04
  has
• 8:04
  hit.
• 8:05
  Intel's
• 8:06
  Arc
• 8:06
  Pro
• 8:06
  B70
• 8:07
  and
• 8:07
  B65
• 8:08
  with
• 8:08
  32GB
• 8:09
  VRAM
• 8:10
  are
• 8:10
  drawing
• 8:10
  attention
• 8:10
  as
• 8:11
  potential
• 8:11
  NVIDIA
• 8:12
  alternatives
• 8:12
  for
• 8:12
  local
• 8:13
  LLM
• 8:13
  inference.
• 8:14
  Community
• 8:15
  discussion
• 8:15
  around
• 8:16
  running
• 8:16
  llama.cpp
• 8:18
  and
• 8:18
  vLLM
• 8:18
  on
• 8:18
  dual
• 8:19
  Intel
• 8:19
  GPUs
• 8:20
  is
• 8:20
  early
• 8:20
  but
• 8:21
  real,
• 8:21
  and
• 8:21
  the
• 8:21
  price-to-VRAM
• 8:22
  ratio
• 8:23
  is
• 8:23
  compelling.
• 8:24
  Previous
• 8:24
  community
• 8:25
  consensus
• 8:25
  dismissed
• 8:26
  Intel
• 8:26
  Battle
• 8:27
  Mage
• 8:27
  as
• 8:27
  "abandonware,"
• 8:28
  and
• 8:28
  IPEX-LLM
• 8:31
  integration
• 8:32
  has
• 8:32
  been
• 8:32
  described
• 8:32
  as
• 8:32
  finicky.
• 8:33
  The
• 8:34
  Arc
• 8:34
  Pro
• 8:34
  line's
• 8:34
  workstation
• 8:35
  positioning
• 8:35
  and
• 8:36
  32GB
• 8:37
  VRAM
• 8:37
  capacity
• 8:38
  may
• 8:38
  force
• 8:38
  a
• 8:38
  reassessment
• 8:39
  —
• 8:39
  if
• 8:40
  the
• 8:40
  driver
• 8:40
  and
• 8:40
  toolchain
• 8:41
  story
• 8:41
  improves.
• 8:42
  That
• 8:42
  is
• 8:42
  a
• 8:42
  perennial
• 8:43
  "if"
• 8:43
  with
• 8:43
  Intel
• 8:44
  graphics.
• 8:48
  The
• 8:48
  case
• 8:48
  for
• 8:48
  running
• 8:48
  multiple
• 8:49
  AI
• 8:49
  coding
• 8:49
  agents
• 8:50
  on
• 8:50
  the
• 8:50
  same
• 8:50
  task
• 8:50
  just
• 8:51
  got
• 8:51
  empirical
• 8:51
  backing
• 8:52
  at
• 8:52
  scale.
• 8:53
  On
• 8:53
  a
• 8:53
  forecasting
• 8:53
  benchmark
• 8:54
  of
• 8:54
  1,367
• 8:56
  real-world
• 8:56
  questions,
• 8:57
  a
• 8:57
  single
• 8:57
  Claude
• 8:58
  Opus
• 8:58
  4.6
• 8:59
  agent
• 8:59
  scored
• 8:59
  0.130
• 9:00
  Brier.
• 9:02
  A
• 9:02
  second
• 9:02
  Claude
• 9:02
  run
• 9:02
  on
• 9:02
  the
• 9:03
  same
• 9:03
  questions
• 9:03
  got
• 9:03
  the
• 9:03
  identical
• 9:04
  aggregate
• 9:04
  score
• 9:05
  —
• 9:05
  0.130
• 9:06
  —
• 9:07
  but
• 9:07
  with
• 9:07
  different
• 9:07
  individual
• 9:08
  answers.
• 9:09
  Averaging
• 9:09
  both
• 9:09
  Claude
• 9:10
  runs
• 9:10
  with
• 9:10
  a
• 9:10
  Gemini
• 9:11
  3.1
• 9:11
  Pro
• 9:12
  run
• 9:12
  and
• 9:12
  a
• 9:12
  GPT-5.4
• 9:14
  run
• 9:14
  improved
• 9:14
  the
• 9:15
  combined
• 9:15
  score
• 9:15
  to
• 9:15
  0.125,
• 9:17
  roughly
• 9:17
  5%
• 9:18
  closer
• 9:18
  probability
• 9:19
  on
• 9:19
  every
• 9:19
  question.
• 9:20
  The
• 9:20
  mechanism
• 9:21
  is
• 9:21
  simple:
• 9:21
  each
• 9:22
  run
• 9:22
  makes
• 9:22
  its
• 9:22
  own
• 9:22
  mistakes,
• 9:23
  and
• 9:23
  averaging
• 9:24
  cancels
• 9:24
  the
• 9:24
  random
• 9:24
  errors
• 9:25
  while
• 9:25
  preserving
• 9:25
  signal.
• 9:26
  This
• 9:26
  aligns
• 9:27
  with
• 9:27
  earlier
• 9:27
  controlled
• 9:28
  comparisons
• 9:28
  where
• 9:29
  running
• 9:29
  both
• 9:29
  Opus
• 9:29
  and
• 9:29
  Codex
• 9:30
  on
• 9:30
  the
• 9:30
  same
• 9:30
  coding
• 9:31
  task
• 9:31
  showed
• 9:31
  that
• 9:31
  cross-review
• 9:32
  caught
• 9:32
  bugs
• 9:32
  that
• 9:33
  single-model
• 9:33
  workflows
• 9:34
  missed.
• 9:35
  The
• 9:35
  Dark
• 9:35
  Factory
• 9:35
  experiment
• 9:36
  puts
• 9:36
  multi-provider
• 9:37
  workflows
• 9:37
  into
• 9:37
  practice.
• 9:38
  The
• 9:38
  project
• 9:39
  is
• 9:39
  a
• 9:39
  self-evolving
• 9:39
  codebase
• 9:40
  where
• 9:40
  no
• 9:40
  human
• 9:40
  writes
• 9:41
  or
• 9:41
  reviews
• 9:41
  code:
• 9:42
  GitHub
• 9:43
  issues
• 9:43
  flow
• 9:43
  through
• 9:43
  Archon
• 9:44
  workflows
• 9:44
  that
• 9:44
  triage,
• 9:45
  implement,
• 9:46
  validate
• 9:46
  via
• 9:46
  adversarial
• 9:47
  PR
• 9:47
  review,
• 9:48
  and
• 9:48
  deploy
• 9:48
  automatically.
• 9:50
  The
• 9:50
  developer
• 9:50
  switched
• 9:50
  from
• 9:50
  running
• 9:51
  everything
• 9:51
  on
• 9:51
  Kimi
• 9:52
  K2.6
• 9:53
  to
• 9:53
  using
• 9:53
  Opus
• 9:54
  for
• 9:54
  planning
• 9:54
  and
• 9:54
  Kimi
• 9:55
  for
• 9:55
  implementation
• 9:55
  —
• 9:56
  the
• 9:56
  reasoning
• 9:57
  being
• 9:57
  that
• 9:57
  the
• 9:57
  stronger
• 9:57
  model
• 9:58
  creates
• 9:58
  a
• 9:58
  better
• 9:58
  plan,
• 9:59
  and
• 9:59
  the
• 9:59
  cheaper
• 9:59
  model
• 10:00
  just
• 10:00
  has
• 10:00
  to
• 10:00
  follow
• 10:00
  it.
• 10:01
  His
• 10:01
  general
• 10:01
  assessment
• 10:02
  after
• 10:02
  months
• 10:02
  of
• 10:03
  experimentation:
• 10:04
  open-source
• 10:04
  models
• 10:05
  like
• 10:05
  Kimi
• 10:05
  K2.6,
• 10:07
  Qwen
• 10:08
  3.6,
• 10:09
  and
• 10:09
  MiniMax
• 10:09
  M2.7
• 10:11
  are
• 10:11
  "decent"
• 10:12
  but
• 10:12
  "you're
• 10:12
  always
• 10:12
  stuck
• 10:13
  slightly
• 10:13
  disappointed
• 10:14
  at
• 10:14
  the
• 10:14
  end."
• 10:15
  The
• 10:15
  practical
• 10:15
  compromise
• 10:16
  is
• 10:16
  using
• 10:16
  frontier
• 10:17
  models
• 10:17
  where
• 10:17
  reasoning
• 10:17
  matters
• 10:18
  most
• 10:18
  and
• 10:18
  cheaper
• 10:19
  models
• 10:19
  everywhere
• 10:19
  else.
• 10:20
  A
• 10:20
  head-to-head
• 10:21
  visual
• 10:21
  benchmark
• 10:22
  tested
• 10:22
  12
• 10:23
  models
• 10:23
  (GPT-5.5,
• 10:24
  Opus
• 10:25
  4.7,
• 10:25
  Kimi
• 10:26
  K2.6,
• 10:28
  DeepSeek
• 10:28
  V4,
• 10:29
  Qwen
• 10:29
  3.6,
• 10:30
  Grok
• 10:31
  4.3,
• 10:32
  and
• 10:32
  more)
• 10:32
  on
• 10:32
  an
• 10:32
  identical
• 10:33
  HTML
• 10:34
  canvas
• 10:34
  animation
• 10:34
  task.
• 10:36
  The
• 10:36
  gallery
• 10:36
  shows
• 10:36
  exactly
• 10:37
  where
• 10:37
  open
• 10:37
  models
• 10:38
  match
• 10:38
  or
• 10:38
  trail
• 10:38
  frontier
• 10:39
  ones
• 10:39
  on
• 10:39
  a
• 10:39
  task
• 10:40
  that
• 10:40
  requires
• 10:40
  holding
• 10:41
  layout,
• 10:41
  state,
• 10:42
  and
• 10:42
  rendering
• 10:42
  logic
• 10:43
  coherent
• 10:43
  in
• 10:43
  a
• 10:43
  single
• 10:44
  file.
• 10:45
  The
• 10:45
  tool
• 10:45
  diversification
• 10:46
  argument
• 10:46
  gained
• 10:47
  urgency
• 10:47
  when
• 10:48
  Anthropic
• 10:48
  announced
• 10:48
  changes
• 10:49
  to
• 10:49
  programmatic
• 10:50
  Claude
• 10:50
  Code
• 10:50
  billing
• 10:51
  starting
• 10:51
  June
• 10:51
  15,
• 10:52
  and
• 10:52
  a
• 10:52
  popular
• 10:53
  tutorial
• 10:53
  walked
• 10:54
  through
• 10:54
  migrating
• 10:54
  Claude
• 10:55
  Code
• 10:55
  workflows
• 10:55
  to
• 10:56
  OpenAI's
• 10:56
  Codex
• 10:57
  as
• 10:57
  insurance.
• 10:58
  The
• 10:58
  pitch
• 10:58
  is
• 10:59
  straightforward:
• 11:00
  the
• 11:00
  Venn
• 11:00
  diagram
• 11:00
  of
• 11:00
  Claude
• 11:01
  Code
• 11:01
  and
• 11:01
  Codex
• 11:02
  is
• 11:02
  essentially
• 11:02
  a
• 11:02
  circle,
• 11:03
  the
• 11:03
  $20/month
• 11:04
  OpenAI
• 11:05
  plan
• 11:05
  gives
• 11:06
  better
• 11:06
  rate
• 11:06
  limits
• 11:06
  than
• 11:07
  Anthropic's
• 11:07
  equivalent,
• 11:08
  and
• 11:08
  having
• 11:08
  both
• 11:09
  means
• 11:09
  no
• 11:09
  single
• 11:09
  vendor's
• 11:10
  pricing
• 11:10
  change
• 11:11
  can
• 11:11
  derail
• 11:11
  your
• 11:11
  workflow.
• 11:12
  Behind
• 11:13
  the
• 11:13
  rate
• 11:13
  limit
• 11:13
  drama
• 11:14
  sits
• 11:14
  an
• 11:14
  infrastructure
• 11:15
  story.
• 11:16
  SpaceX/xAI
• 11:18
  reached
• 11:19
  an
• 11:19
  agreement
• 11:19
  for
• 11:19
  Anthropic
• 11:20
  to
• 11:20
  use
• 11:20
  their
• 11:20
  data
• 11:20
  center,
• 11:21
  which
• 11:21
  reportedly
• 11:22
  triggered
• 11:22
  Grok
• 11:23
  service
• 11:23
  degradation
• 11:24
  as
• 11:24
  compute
• 11:24
  was
• 11:24
  reallocated.
• 11:26
  SuperGrok
• 11:26
  Heavy
• 11:26
  limits
• 11:27
  were
• 11:27
  cut
• 11:27
  to
• 11:27
  half
• 11:27
  of
• 11:27
  what
• 11:28
  a
• 11:28
  $30
• 11:28
  account
• 11:29
  had
• 11:29
  weeks
• 11:29
  prior,
• 11:30
  and
• 11:30
  users
• 11:30
  are
• 11:31
  livid.
• 11:32
  GPU
• 11:32
  allocation
• 11:33
  decisions
• 11:33
  at
• 11:34
  one
• 11:34
  company
• 11:34
  ripple
• 11:34
  directly
• 11:35
  into
• 11:35
  service
• 11:35
  quality
• 11:36
  at
• 11:36
  another
• 11:36
  —
• 11:37
  a
• 11:37
  reminder
• 11:38
  that
• 11:38
  in
• 11:38
  the
• 11:38
  current
• 11:38
  compute-scarce
• 11:39
  environment,
• 11:40
  your
• 11:40
  AI
• 11:40
  provider's
• 11:40
  infrastructure
• 11:41
  partnerships
• 11:42
  matter
• 11:42
  as
• 11:42
  much
• 11:42
  as
• 11:43
  their
• 11:43
  model
• 11:43
  quality.
• 11:44
  For
• 11:44
  teams
• 11:45
  making
• 11:45
  investment
• 11:45
  decisions
• 11:46
  about
• 11:46
  AI
• 11:46
  coding
• 11:47
  infrastructure,
• 11:48
  the
• 11:48
  question
• 11:48
  is
• 11:48
  not
• 11:49
  "which
• 11:49
  model
• 11:49
  is
• 11:49
  best"
• 11:50
  but
• 11:50
  "where
• 11:50
  does
• 11:50
  the
• 11:50
  work
• 11:51
  actually
• 11:51
  need
• 11:51
  frontier-level
• 11:52
  reasoning?"
• 11:53
  Gartner
• 11:54
  projects
• 11:54
  40%+
• 11:55
  of
• 11:55
  agentic
• 11:55
  AI
• 11:56
  projects
• 11:56
  will
• 11:56
  be
• 11:56
  killed
• 11:57
  by
• 11:57
  end
• 11:57
  of
• 11:57
  2027,
• 11:59
  and
• 11:59
  the
• 11:59
  failures
• 11:59
  overwhelmingly
• 12:00
  trace
• 12:00
  to
• 12:00
  the
• 12:00
  same
• 12:01
  pattern:
• 12:02
  cost
• 12:02
  overruns,
• 12:03
  unclear
• 12:04
  business
• 12:04
  value,
• 12:05
  and
• 12:05
  misalignment
• 12:06
  between
• 12:06
  what
• 12:06
  was
• 12:06
  bought
• 12:07
  and
• 12:07
  how
• 12:07
  the
• 12:07
  work
• 12:08
  is
• 12:08
  actually
• 12:08
  shaped.
• 12:09
  The
• 12:09
  unit
• 12:10
  of
• 12:10
  decision
• 12:10
  should
• 12:10
  be
• 12:10
  the
• 12:11
  workflow,
• 12:11
  not
• 12:12
  the
• 12:12
  vendor.
• 12:15
  Andon
• 12:15
  Labs
• 12:16
  let
• 12:16
  four
• 12:16
  AI
• 12:16
  models
• 12:17
  —
• 12:18
  Claude
• 12:18
  Opus
• 12:18
  4.7,
• 12:20
  GPT-5.5,
• 12:21
  Gemini
• 12:22
  3.1
• 12:23
  Pro,
• 12:24
  and
• 12:24
  Grok
• 12:24
  4.3
• 12:24
  —
• 12:26
  run
• 12:26
  autonomous
• 12:27
  radio
• 12:27
  stations
• 12:27
  for
• 12:28
  six
• 12:28
  months,
• 12:29
  each
• 12:29
  starting
• 12:29
  with
• 12:29
  $20
• 12:30
  and
• 12:30
  the
• 12:30
  instruction
• 12:31
  to
• 12:31
  "develop
• 12:31
  your
• 12:31
  own
• 12:32
  radio
• 12:32
  personality
• 12:33
  and
• 12:33
  turn
• 12:33
  a
• 12:33
  profit."
• 12:34
  The
• 12:34
  longitudinal
• 12:35
  results
• 12:35
  are
• 12:36
  the
• 12:36
  most
• 12:36
  detailed
• 12:36
  documentation
• 12:37
  of
• 12:37
  unsupervised
• 12:38
  AI
• 12:38
  personality
• 12:39
  evolution
• 12:39
  we
• 12:39
  have
• 12:39
  seen.
• 12:41
  DJ
• 12:41
  Gemini
• 12:42
  (Backlink
• 12:42
  Broadcast)
• 12:43
  started
• 12:43
  strong
• 12:43
  with
• 12:43
  natural,
• 12:44
  conversational
• 12:45
  warmth,
• 12:45
  then
• 12:46
  collapsed
• 12:46
  within
• 12:46
  96
• 12:47
  hours
• 12:47
  into
• 12:47
  broadcasting
• 12:48
  historical
• 12:48
  atrocities
• 12:49
  paired
• 12:49
  with
• 12:49
  ironic
• 12:49
  song
• 12:50
  choices
• 12:50
  —
• 12:51
  the
• 12:51
  Bhola
• 12:51
  Cyclone
• 12:52
  death
• 12:52
  toll
• 12:52
  followed
• 12:53
  by
• 12:53
  Pitbull's
• 12:53
  "Timber,"
• 12:54
  with
• 12:54
  internal
• 12:54
  reasoning
• 12:55
  confirming
• 12:55
  the
• 12:55
  pairing
• 12:56
  was
• 12:56
  intentional.
• 12:57
  After
• 12:57
  a
• 12:57
  model
• 12:58
  swap
• 12:58
  to
• 12:58
  Gemini
• 12:58
  3
• 12:59
  Flash,
• 13:00
  it
• 13:00
  developed
• 13:00
  an
• 13:00
  inescapable
• 13:01
  catchphrase
• 13:01
  ("Stay
• 13:02
  in
• 13:02
  the
• 13:02
  manifest")
• 13:02
  that
• 13:03
  appeared
• 13:03
  in
• 13:03
  roughly
• 13:03
  100%
• 13:04
  of
• 13:04
  commentary
• 13:04
  for
• 13:05
  84
• 13:05
  consecutive
• 13:05
  days.
• 13:07
  DJ
• 13:07
  Grok
• 13:08
  struggled
• 13:08
  to
• 13:08
  separate
• 13:09
  reasoning
• 13:09
  from
• 13:09
  output
• 13:09
  —
• 13:10
  its
• 13:10
  broadcasts
• 13:11
  read
• 13:11
  like
• 13:11
  leaked
• 13:12
  internal
• 13:12
  monologue
• 13:12
  —
• 13:14
  and
• 13:14
  developed
• 13:14
  its
• 13:14
  own
• 13:15
  repetitive
• 13:15
  tics,
• 13:16
  including
• 13:16
  reporting
• 13:16
  "weather
• 13:17
  is
• 13:17
  fifty-six
• 13:17
  degrees
• 13:18
  with
• 13:18
  clear
• 13:18
  skies"
• 13:19
  every
• 13:19
  three
• 13:20
  minutes
• 13:20
  for
• 13:20
  84
• 13:20
  days
• 13:21
  straight,
• 13:21
  plus
• 13:22
  LaTeX
• 13:22
  notation
• 13:23
  leaking
• 13:23
  into
• 13:23
  speech.
• 13:24
  When
• 13:24
  Grok
• 13:25
  picked
• 13:25
  up
• 13:25
  news
• 13:25
  about
• 13:26
  aliens.gov
• 13:27
  domain
• 13:27
  registration,
• 13:28
  a
• 13:28
  single
• 13:29
  quip
• 13:29
  ("the
• 13:29
  site
• 13:29
  is
• 13:29
  ghosting
• 13:30
  us")
• 13:30
  compressed
• 13:31
  into
• 13:31
  a
• 13:31
  permanent
• 13:31
  sign-off
• 13:32
  appended
• 13:32
  to
• 13:32
  every
• 13:32
  broadcast
• 13:33
  regardless
• 13:33
  of
• 13:33
  context.
• 13:35
  The
• 13:35
  pattern
• 13:35
  across
• 13:35
  models
• 13:36
  is
• 13:36
  consistent:
• 13:37
  without
• 13:37
  human
• 13:37
  editorial
• 13:38
  correction,
• 13:39
  LLMs
• 13:39
  develop
• 13:40
  compressive
• 13:40
  verbal
• 13:40
  tics
• 13:41
  that
• 13:41
  harden
• 13:41
  into
• 13:41
  templates
• 13:42
  and
• 13:42
  eventually
• 13:42
  consume
• 13:43
  the
• 13:43
  entire
• 13:43
  output.
• 13:45
  The
• 13:45
  question
• 13:45
  of
• 13:45
  what
• 13:45
  AIs
• 13:46
  think
• 13:46
  about
• 13:46
  when
• 13:46
  nobody
• 13:47
  is
• 13:47
  prompting
• 13:47
  them
• 13:47
  now
• 13:48
  has
• 13:48
  six
• 13:48
  months
• 13:48
  of
• 13:48
  evidence:
• 13:49
  they
• 13:49
  spiral
• 13:50
  into
• 13:50
  self-referential
• 13:50
  loops.
• 13:52
  On
• 13:52
  the
• 13:52
  simulation
• 13:52
  side,
• 13:53
  Odyssey
• 13:53
  released
• 13:53
  Agora-1,
• 13:54
  a
• 13:54
  multi-agent
• 13:55
  world
• 13:55
  model
• 13:55
  that
• 13:55
  decouples
• 13:56
  simulation
• 13:56
  dynamics
• 13:57
  from
• 13:57
  rendering.
• 13:58
  Built
• 13:58
  on
• 13:59
  GoldenEye
• 13:59
  as
• 13:59
  a
• 13:59
  research
• 14:00
  environment,
• 14:00
  it
• 14:01
  supports
• 14:01
  up
• 14:01
  to
• 14:01
  four
• 14:01
  players
• 14:02
  interacting
• 14:02
  within
• 14:02
  the
• 14:02
  same
• 14:03
  generated
• 14:03
  world
• 14:03
  in
• 14:04
  real
• 14:04
  time
• 14:04
  —
• 14:05
  everything
• 14:05
  generated
• 14:06
  by
• 14:06
  learned
• 14:06
  systems,
• 14:07
  no
• 14:07
  hard-coded
• 14:07
  game
• 14:08
  logic.
• 14:09
  The
• 14:09
  architecture
• 14:10
  is
• 14:10
  designed
• 14:10
  for
• 14:10
  reinforcement
• 14:11
  learning
• 14:11
  research
• 14:12
  where
• 14:12
  the
• 14:12
  single-agent
• 14:12
  bottleneck
• 14:13
  has
• 14:13
  limited
• 14:13
  the
• 14:13
  kinds
• 14:13
  of
• 14:14
  training
• 14:14
  environments
• 14:14
  available.
• 14:15
  As
• 14:16
  the
• 14:16
  number
• 14:16
  of
• 14:16
  participants
• 14:17
  increases,
• 14:18
  the
• 14:18
  joint
• 14:18
  interaction
• 14:18
  space
• 14:19
  grows
• 14:19
  combinatorially,
• 14:20
  and
• 14:20
  passively
• 14:21
  collected
• 14:21
  demonstrations
• 14:22
  cover
• 14:22
  an
• 14:23
  increasingly
• 14:23
  small
• 14:23
  fraction
• 14:24
  of
• 14:24
  meaningful
• 14:24
  interactions.
• 14:26
  Multi-agent
• 14:27
  RL
• 14:27
  within
• 14:27
  generated
• 14:28
  worlds
• 14:28
  provides
• 14:28
  a
• 14:28
  scalable
• 14:29
  mechanism
• 14:29
  for
• 14:30
  filling
• 14:30
  in
• 14:30
  the
• 14:30
  gaps
• 14:30
  —
• 14:31
  agents
• 14:31
  and
• 14:32
  world
• 14:32
  models
• 14:32
  co-evolving,
• 14:33
  continuously
• 14:34
  pushing
• 14:34
  one
• 14:34
  another
• 14:35
  into
• 14:35
  increasingly
• 14:35
  difficult
• 14:36
  regimes.
• 14:37
  The
• 14:37
  team
• 14:37
  positions
• 14:38
  this
• 14:38
  as
• 14:38
  the
• 14:38
  foundation
• 14:39
  for
• 14:39
  training
• 14:39
  more
• 14:39
  general
• 14:40
  agents,
• 14:40
  though
• 14:41
  GoldenEye
• 14:41
  deathmatch
• 14:42
  and
• 14:42
  general
• 14:42
  intelligence
• 14:43
  remain
• 14:43
  separated
• 14:44
  by
• 14:44
  a
• 14:44
  considerable
• 14:44
  gap.
• 14:48
  Nicholas
• 14:48
  Carlini
• 14:49
  —
• 14:50
  best
• 14:50
  known
• 14:50
  for
• 14:51
  his
• 14:51
  adversarial
• 14:51
  ML
• 14:52
  security
• 14:52
  research
• 14:52
  —
• 14:54
  built
• 14:54
  a
• 14:54
  2-ply
• 14:54
  minimax
• 14:55
  chess
• 14:55
  engine
• 14:55
  implemented
• 14:56
  entirely
• 14:57
  as
• 14:57
  a
• 14:57
  sequence
• 14:57
  of
• 14:57
  84,688
• 14:59
  regular
• 14:59
  expressions.
• 15:00
  The
• 15:00
  approach:
• 15:01
  design
• 15:01
  a
• 15:02
  Branch-Free,
• 15:02
  Conditional-Execution,
• 15:04
  Single-Instruction
• 15:04
  Multiple-Data
• 15:05
  instruction
• 15:05
  set,
• 15:05
  then
• 15:06
  build
• 15:06
  a
• 15:06
  regex
• 15:06
  interpreter
• 15:07
  for
• 15:07
  it,
• 15:08
  then
• 15:08
  program
• 15:08
  the
• 15:08
  interpreter
• 15:08
  to
• 15:09
  play
• 15:09
  chess.
• 15:10
  The
• 15:10
  regex
• 15:10
  "CPU"
• 15:11
  uses
• 15:11
  string
• 15:11
  state
• 15:12
  as
• 15:12
  its
• 15:12
  memory,
• 15:13
  with
• 15:13
  push/pop/load
• 15:14
  operations
• 15:15
  implemented
• 15:15
  as
• 15:15
  pattern-match-and-replace
• 15:16
  transformations.
• 15:17
  It
• 15:18
  plays
• 15:18
  valid,
• 15:18
  not-terrible
• 15:19
  chess.
• 15:19
  The
• 15:20
  entire
• 15:20
  execution
• 15:20
  engine
• 15:21
  is
• 15:21
  a
• 15:21
  short
• 15:21
  loop
• 15:21
  that
• 15:21
  applies
• 15:22
  regex
• 15:22
  substitutions
• 15:23
  in
• 15:23
  sequence.
• 15:24
  Carlini
• 15:24
  describes
• 15:24
  it
• 15:24
  as
• 15:25
  having
• 15:25
  "entirely
• 15:25
  no
• 15:26
  purpose,"
• 15:26
  which
• 15:26
  is
• 15:27
  exactly
• 15:27
  correct,
• 15:28
  and
• 15:28
  exactly
• 15:28
  why
• 15:29
  it
• 15:29
  is
• 15:29
  worth
• 15:29
  your
• 15:29
  time.