Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 24, 2026

Transcript

AGI Dreams Podcast – April 24, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  Vercel
• 0:01
  breach
• 0:01
  may
• 0:01
  be
• 0:01
  the
• 0:01
  most
• 0:02
  instructive
• 0:02
  platform
• 0:03
  compromise
• 0:03
  of
• 0:03
  2026
• 0:04
  so
• 0:04
  far
• 0:04
  —
• 0:05
  not
• 0:05
  because
• 0:05
  the
• 0:05
  techniques
• 0:06
  were
• 0:06
  novel,
• 0:06
  but
• 0:07
  because
• 0:07
  the
• 0:07
  blast
• 0:07
  radius
• 0:08
  was
• 0:08
  entirely
• 0:09
  predictable
• 0:10
  and
• 0:10
  entirely
• 0:10
  preventable.
• 0:11
  Trend
• 0:12
  Micro's
• 0:12
  detailed
• 0:12
  analysis
• 0:13
  traces
• 0:13
  the
• 0:13
  attack
• 0:14
  to
• 0:14
  a
• 0:14
  Lumma
• 0:14
  Stealer
• 0:14
  malware
• 0:15
  infection
• 0:15
  at
• 0:15
  Context.ai,
• 0:17
  a
• 0:17
  third-party
• 0:17
  AI
• 0:18
  analytics
• 0:18
  company,
• 0:19
  in
• 0:19
  approximately
• 0:19
  February
• 0:20
  2026.
• 0:21
  The
• 0:21
  attacker
• 0:22
  exfiltrated
• 0:22
  Google
• 0:22
  Workspace
• 0:23
  OAuth
• 0:23
  tokens
• 0:24
  from
• 0:24
  Context.ai's
• 0:25
  AWS
• 0:25
  environment,
• 0:26
  used
• 0:26
  one
• 0:27
  to
• 0:27
  pivot
• 0:27
  into
• 0:27
  a
• 0:27
  Vercel
• 0:28
  employee's
• 0:28
  Workspace
• 0:28
  account,
• 0:29
  then
• 0:29
  escalated
• 0:30
  into
• 0:30
  Vercel's
• 0:30
  internal
• 0:31
  systems
• 0:31
  and
• 0:32
  enumerated
• 0:32
  customer
• 0:32
  environment
• 0:33
  variables.
• 0:34
  The
• 0:34
  critical
• 0:34
  design
• 0:34
  flaw:
• 0:35
  Vercel's
• 0:36
  environment
• 0:36
  variable
• 0:36
  sensitivity
• 0:37
  flag
• 0:37
  defaults
• 0:37
  to
• 0:38
  off.
• 0:39
  Every
• 0:39
  DATABASE
• 0:39
  URL,
• 0:40
  STRIPE
• 0:40
  SECRET
• 0:41
  KEY,
• 0:41
  or
• 0:41
  AWS
• 0:42
  SECRET
• 0:42
  ACCESS
• 0:43
  KEY
• 0:44
  that
• 0:44
  a
• 0:44
  developer
• 0:44
  added
• 0:44
  without
• 0:45
  explicitly
• 0:45
  toggling
• 0:46
  the
• 0:46
  sensitive
• 0:46
  flag
• 0:47
  was
• 0:47
  stored
• 0:47
  in
• 0:47
  a
• 0:47
  way
• 0:47
  that
• 0:47
  made
• 0:48
  it
• 0:48
  readable
• 0:48
  with
• 0:48
  internal
• 0:48
  access.
• 0:49
  CEO
• 0:50
  Guillermo
• 0:50
  Rauch
• 0:51
  confirmed
• 0:51
  the
• 0:51
  chain
• 0:52
  and
• 0:52
  notably
• 0:52
  attributed
• 0:53
  the
• 0:53
  attacker's
• 0:53
  speed
• 0:54
  to
• 0:54
  AI
• 0:54
  augmentation
• 0:55
  —
• 0:56
  one
• 0:56
  of
• 0:56
  the
• 0:56
  first
• 0:56
  on-record
• 0:57
  claims
• 0:57
  of
• 0:57
  AI-accelerated
• 0:58
  adversary
• 0:58
  tradecraft
• 0:59
  from
• 0:59
  the
• 0:59
  CEO
• 0:59
  of
• 1:00
  an
• 1:00
  affected
• 1:00
  platform.
• 1:01
  The
• 1:01
  timing
• 1:01
  is
• 1:01
  brutal.
• 1:02
  This
• 1:02
  breach
• 1:03
  slots
• 1:03
  into
• 1:03
  a
• 1:03
  three-week
• 1:04
  convergence
• 1:04
  with
• 1:04
  LiteLLM
• 1:05
  (stolen
• 1:05
  CI/CD
• 1:06
  credentials
• 1:07
  yielding
• 1:07
  malicious
• 1:07
  PyPI
• 1:08
  packages
• 1:08
  targeting
• 1:09
  50+
• 1:09
  credential
• 1:10
  types)
• 1:10
  and
• 1:10
  Axios
• 1:10
  (maintainer
• 1:11
  account
• 1:11
  hijacked,
• 1:12
  RAT
• 1:12
  deployed
• 1:13
  to
• 1:13
  an
• 1:13
  npm
• 1:13
  package
• 1:14
  with
• 1:14
  70-100
• 1:16
  million
• 1:17
  weekly
• 1:17
  downloads,
• 1:18
  attributed
• 1:18
  to
• 1:18
  North
• 1:19
  Korea's
• 1:19
  Sapphire
• 1:19
  Sleet).
• 1:20
  Three
• 1:20
  different
• 1:20
  vectors,
• 1:21
  one
• 1:21
  target:
• 1:22
  credentials
• 1:22
  developers
• 1:23
  store
• 1:23
  in
• 1:23
  their
• 1:23
  toolchains.
• 1:24
  A
• 1:24
  single
• 1:25
  Vercel
• 1:25
  project
• 1:26
  commonly
• 1:26
  contains
• 1:26
  10-30
• 1:27
  environment
• 1:28
  variables;
• 1:29
  at
• 1:29
  organizational
• 1:29
  scale,
• 1:30
  that
• 1:30
  is
• 1:30
  500-1,500
• 1:34
  potential
• 1:35
  pivot
• 1:35
  points.
• 1:36
  One
• 1:36
  public
• 1:37
  customer
• 1:37
  reported
• 1:37
  receiving
• 1:38
  an
• 1:38
  OpenAI
• 1:38
  leaked-key
• 1:39
  notification
• 1:39
  nine
• 1:40
  days
• 1:40
  before
• 1:40
  Vercel's
• 1:41
  disclosure
• 1:41
  —
• 1:42
  for
• 1:42
  an
• 1:42
  API
• 1:43
  key
• 1:43
  that
• 1:43
  existed
• 1:43
  only
• 1:44
  in
• 1:44
  Vercel
• 1:44
  —
• 1:45
  suggesting
• 1:46
  credentials
• 1:46
  were
• 1:46
  being
• 1:46
  exploited
• 1:47
  in
• 1:47
  the
• 1:47
  wild
• 1:47
  before
• 1:48
  anyone
• 1:48
  was
• 1:48
  told
• 1:48
  to
• 1:48
  rotate.
• 1:50
  Regulatory
• 1:50
  implications
• 1:51
  under
• 1:51
  GDPR's
• 1:52
  72-hour
• 1:53
  clock,
• 1:53
  SOC
• 1:54
  2
• 1:54
  continuous
• 1:54
  monitoring,
• 1:55
  and
• 1:56
  SEC's
• 1:56
  8-K
• 1:57
  rules
• 1:57
  are
• 1:57
  now
• 1:57
  live
• 1:57
  questions
• 1:58
  for
• 1:58
  affected
• 1:58
  organizations.
• 2:00
  Meanwhile,
• 2:01
  the
• 2:01
  offensive
• 2:01
  side
• 2:01
  of
• 2:01
  the
• 2:01
  security
• 2:02
  equation
• 2:02
  took
• 2:02
  a
• 2:02
  leap.
• 2:04
  XBOW,
• 2:05
  an
• 2:05
  AI-driven
• 2:06
  penetration
• 2:06
  testing
• 2:06
  firm,
• 2:07
  reports
• 2:07
  that
• 2:07
  GPT-5.5
• 2:09
  reduces
• 2:09
  their
• 2:09
  vulnerability
• 2:10
  miss
• 2:10
  rate
• 2:10
  to
• 2:11
  just
• 2:11
  10%
• 2:11
  —
• 2:12
  down
• 2:12
  from
• 2:12
  Opus
• 2:13
  4.6's
• 2:14
  18%
• 2:14
  and
• 2:14
  GPT-5's
• 2:15
  40%.
• 2:17
  More
• 2:17
  striking:
• 2:18
  GPT-5.5
• 2:20
  running
• 2:20
  black-box
• 2:20
  (no
• 2:21
  source
• 2:21
  code)
• 2:21
  already
• 2:21
  outperforms
• 2:22
  GPT-5
• 2:23
  running
• 2:23
  white-box.
• 2:24
  In
• 2:24
  white-box
• 2:25
  mode,
• 2:25
  performance
• 2:26
  jumped
• 2:26
  so
• 2:26
  dramatically
• 2:27
  it
• 2:27
  "effectively
• 2:27
  killed
• 2:27
  our
• 2:28
  benchmark."
• 2:28
  The
• 2:29
  model
• 2:29
  also
• 2:29
  fails
• 2:30
  faster
• 2:30
  when
• 2:30
  credentials
• 2:31
  are
• 2:31
  wrong
• 2:31
  or
• 2:31
  access
• 2:32
  is
• 2:32
  blocked,
• 2:32
  cutting
• 2:33
  wasted
• 2:33
  iteration
• 2:34
  time
• 2:34
  roughly
• 2:34
  in
• 2:34
  half.
• 2:35
  XBOW
• 2:36
  notes
• 2:36
  that
• 2:36
  GPT-5.5
• 2:38
  persists
• 2:38
  on
• 2:38
  failing
• 2:39
  paths
• 2:39
  only
• 2:39
  half
• 2:40
  as
• 2:40
  often
• 2:40
  as
• 2:40
  previous
• 2:41
  models
• 2:41
  —
• 2:42
  a
• 2:42
  quietly
• 2:43
  important
• 2:43
  capability
• 2:44
  for
• 2:44
  autonomous
• 2:44
  agents
• 2:45
  that
• 2:45
  need
• 2:45
  to
• 2:45
  know
• 2:45
  when
• 2:46
  to
• 2:46
  give
• 2:46
  up.
• 2:47
  On
• 2:47
  the
• 2:47
  data
• 2:47
  governance
• 2:48
  side,
• 2:48
  UK
• 2:48
  Biobank's
• 2:49
  health
• 2:49
  data
• 2:49
  keeps
• 2:50
  surfacing
• 2:50
  on
• 2:50
  GitHub,
• 2:51
  and
• 2:51
  the
• 2:51
  institution
• 2:52
  has
• 2:52
  resorted
• 2:52
  to
• 2:52
  an
• 2:53
  unusual
• 2:53
  tactic:
• 2:54
  copyright
• 2:54
  takedown
• 2:55
  notices.
• 2:56
  Across
• 2:56
  110
• 2:57
  DMCA
• 2:58
  requests
• 2:58
  since
• 2:58
  July
• 2:59
  2025,
• 3:00
  researchers
• 3:01
  in
• 3:01
  at
• 3:01
  least
• 3:01
  14
• 3:02
  countries
• 3:02
  have
• 3:02
  inadvertently
• 3:03
  uploaded
• 3:03
  participant
• 3:04
  genotypes,
• 3:05
  phenotype
• 3:05
  records,
• 3:06
  and
• 3:06
  genomic
• 3:06
  data
• 3:06
  files
• 3:07
  alongside
• 3:07
  their
• 3:08
  analysis
• 3:08
  code.
• 3:09
  Nearly
• 3:09
  half
• 3:09
  the
• 3:09
  targeted
• 3:10
  files
• 3:10
  are
• 3:10
  Jupyter
• 3:11
  or
• 3:11
  R
• 3:11
  notebooks
• 3:11
  —
• 3:12
  a
• 3:13
  few
• 3:13
  rows
• 3:13
  of
• 3:13
  data
• 3:13
  embedded
• 3:14
  in
• 3:14
  a
• 3:14
  research
• 3:14
  workflow.
• 3:16
  The
• 3:16
  Guardian
• 3:16
  previously
• 3:17
  demonstrated
• 3:17
  that
• 3:17
  a
• 3:17
  single
• 3:18
  participant
• 3:18
  could
• 3:18
  be
• 3:18
  re-identified
• 3:19
  using
• 3:19
  only
• 3:20
  birth
• 3:20
  month/year
• 3:21
  and
• 3:21
  the
• 3:21
  date
• 3:21
  of
• 3:22
  one
• 3:22
  surgery.
• 3:23
  UK
• 3:23
  Biobank's
• 3:24
  CEO
• 3:24
  has
• 3:24
  reassured
• 3:25
  participants
• 3:26
  that
• 3:26
  no
• 3:26
  names
• 3:26
  or
• 3:26
  NHS
• 3:27
  numbers
• 3:27
  were
• 3:27
  exposed,
• 3:28
  but
• 3:28
  the
• 3:28
  structural
• 3:29
  problem
• 3:29
  —
• 3:30
  researchers
• 3:30
  sharing
• 3:31
  code
• 3:31
  without
• 3:31
  scrubbing
• 3:32
  embedded
• 3:32
  data
• 3:32
  —
• 3:33
  remains
• 3:33
  unsolved.
• 3:35
  For
• 3:35
  security
• 3:35
  practitioners
• 3:36
  drowning
• 3:36
  in
• 3:37
  conference
• 3:37
  content,
• 3:38
  one
• 3:38
  practical
• 3:39
  tool
• 3:39
  worth
• 3:39
  noting:
• 3:40
  greptalks
• 3:41
  now
• 3:41
  indexes
• 3:41
  500+
• 3:42
  talks
• 3:42
  from
• 3:43
  Black
• 3:43
  Hat,
• 3:43
  DEF
• 3:44
  CON,
• 3:44
  RSA,
• 3:45
  and
• 3:45
  others,
• 3:46
  with
• 3:46
  two
• 3:46
  AI
• 3:47
  personas
• 3:47
  —
• 3:48
  a
• 3:48
  ruthless
• 3:49
  ex-TAO
• 3:49
  technical
• 3:50
  analyst
• 3:50
  and
• 3:51
  a
• 3:51
  CISO-perspective
• 3:51
  evaluator
• 3:52
  —
• 3:53
  scoring
• 3:53
  and
• 3:54
  summarizing
• 3:54
  every
• 3:58
  DeepSeek
• 3:59
  is
• 3:59
  having
• 3:59
  a
• 3:59
  week
• 3:59
  that
• 3:59
  would
• 3:59
  make
• 4:00
  most
• 4:00
  startups
• 4:00
  dizzy.
• 4:01
  The
• 4:01
  company
• 4:01
  —
• 4:02
  spun
• 4:02
  out
• 4:02
  of
• 4:02
  a
• 4:02
  quantitative
• 4:03
  hedge
• 4:03
  fund,
• 4:03
  never
• 4:03
  previously
• 4:04
  having
• 4:04
  taken
• 4:04
  outside
• 4:04
  money,
• 4:05
  releasing
• 4:05
  models
• 4:06
  for
• 4:06
  free
• 4:06
  —
• 4:07
  reportedly
• 4:08
  saw
• 4:08
  its
• 4:08
  first
• 4:08
  funding
• 4:08
  round
• 4:09
  rocket
• 4:09
  from
• 4:09
  an
• 4:09
  initial
• 4:09
  $10
• 4:10
  billion
• 4:10
  valuation
• 4:11
  target
• 4:11
  to
• 4:11
  above
• 4:11
  $20
• 4:12
  billion
• 4:12
  in
• 4:12
  days,
• 4:13
  with
• 4:13
  Tencent
• 4:13
  and
• 4:14
  Alibaba
• 4:14
  both
• 4:14
  circling.
• 4:15
  Tencent
• 4:15
  reportedly
• 4:16
  pushed
• 4:16
  for
• 4:16
  20%
• 4:17
  of
• 4:17
  the
• 4:17
  company;
• 4:18
  DeepSeek
• 4:18
  pushed
• 4:18
  back.
• 4:20
  The
• 4:20
  traditional
• 4:20
  valuation
• 4:20
  metrics
• 4:21
  simply
• 4:21
  do
• 4:21
  not
• 4:21
  apply
• 4:21
  to
• 4:22
  a
• 4:22
  company
• 4:22
  built
• 4:22
  like
• 4:22
  this
• 4:22
  —
• 4:23
  no
• 4:23
  revenue
• 4:24
  model,
• 4:24
  no
• 4:24
  prior
• 4:25
  outside
• 4:25
  capital,
• 4:26
  just
• 4:26
  engineering
• 4:27
  output
• 4:27
  that
• 4:27
  keeps
• 4:27
  embarrassing
• 4:28
  organizations
• 4:28
  with
• 4:28
  hundred-billion-dollar
• 4:29
  budgets.
• 4:30
  What
• 4:30
  earns
• 4:31
  that
• 4:31
  valuation
• 4:31
  in
• 4:31
  practice
• 4:32
  showed
• 4:32
  up
• 4:32
  this
• 4:32
  week
• 4:33
  when
• 4:33
  DeepSeek
• 4:33
  open-sourced
• 4:34
  their
• 4:34
  Tile
• 4:34
  Kernels
• 4:34
  —
• 4:35
  the
• 4:35
  actual
• 4:35
  GPU
• 4:36
  kernels
• 4:36
  used
• 4:36
  in
• 4:37
  their
• 4:37
  production
• 4:37
  training
• 4:37
  and
• 4:37
  inference
• 4:38
  pipelines.
• 4:39
  These
• 4:39
  are
• 4:39
  not
• 4:39
  research
• 4:39
  prototypes
• 4:40
  or
• 4:40
  reference
• 4:40
  implementations.
• 4:42
  As
• 4:42
  one
• 4:42
  developer
• 4:42
  put
• 4:43
  it:
• 4:43
  "We
• 4:43
  wrote
• 4:44
  GPU
• 4:44
  kernels
• 4:44
  so
• 4:44
  optimized
• 4:45
  they
• 4:45
  flirt
• 4:45
  with
• 4:45
  the
• 4:45
  hardware's
• 4:46
  physical
• 4:46
  limits."
• 4:47
  TileLang,
• 4:48
  the
• 4:48
  domain-specific
• 4:49
  language
• 4:49
  underlying
• 4:49
  the
• 4:50
  kernels,
• 4:50
  provides
• 4:51
  an
• 4:51
  abstraction
• 4:51
  layer
• 4:52
  that
• 4:52
  community
• 4:52
  developers
• 4:53
  are
• 4:53
  already
• 4:53
  targeting
• 4:53
  for
• 4:54
  local
• 4:54
  hardware
• 4:54
  like
• 4:55
  NVIDIA's
• 4:55
  GB10.
• 4:56
  This
• 4:56
  follows
• 4:57
  DeepSeek's
• 4:57
  quiet
• 4:58
  DeepGemm
• 4:58
  repository
• 4:59
  update
• 4:59
  last
• 4:59
  week
• 5:00
  with
• 5:00
  mega
• 5:00
  FP4
• 5:01
  quantization
• 5:02
  and
• 5:02
  distributed
• 5:02
  communication
• 5:03
  optimizations
• 5:03
  —
• 5:04
  a
• 5:05
  pattern
• 5:05
  that
• 5:05
  points
• 5:05
  toward
• 5:06
  infrastructure
• 5:06
  preparation
• 5:07
  for
• 5:07
  whatever
• 5:07
  model
• 5:07
  generation
• 5:08
  comes
• 5:08
  next.
• 5:09
  Open-sourcing
• 5:10
  production
• 5:10
  kernels
• 5:11
  is
• 5:11
  the
• 5:11
  closest
• 5:11
  thing
• 5:12
  to
• 5:12
  handing
• 5:12
  out
• 5:12
  cheat
• 5:12
  codes,
• 5:13
  and
• 5:13
  the
• 5:13
  local
• 5:14
  AI
• 5:14
  community
• 5:14
  is
• 5:14
  not
• 5:15
  being
• 5:15
  subtle
• 5:15
  about
• 5:15
  what
• 5:15
  it
• 5:15
  plans
• 5:16
  to
• 5:16
  do
• 5:16
  with
• 5:16
  them.
• 5:17
  The
• 5:17
  v4
• 5:18
  model
• 5:18
  itself
• 5:18
  is
• 5:18
  now
• 5:19
  live,
• 5:19
  with
• 5:20
  an
• 5:20
  API
• 5:20
  format
• 5:20
  compatible
• 5:21
  with
• 5:21
  both
• 5:21
  OpenAI
• 5:22
  and
• 5:22
  Anthropic
• 5:23
  SDKs
• 5:23
  —
• 5:24
  a
• 5:24
  deliberate
• 5:24
  interoperability
• 5:25
  play
• 5:25
  that
• 5:25
  makes
• 5:26
  switching
• 5:26
  trivially
• 5:26
  easy.
• 5:27
  The
• 5:27
  older
• 5:28
  model
• 5:28
  names
• 5:28
  are
• 5:28
  deprecated
• 5:29
  as
• 5:29
  of
• 5:29
  July
• 5:30
  2026,
• 5:31
  consolidating
• 5:32
  the
• 5:32
  lineup
• 5:32
  around
• 5:33
  v4's
• 5:33
  thinking
• 5:33
  and
• 5:34
  non-thinking
• 5:34
  modes.
• 5:36
  Anthropic
• 5:37
  published
• 5:37
  a
• 5:37
  transparency
• 5:38
  post
• 5:38
  this
• 5:38
  week
• 5:39
  that
• 5:39
  landed
• 5:39
  like
• 5:39
  a
• 5:39
  confession.
• 5:41
  Three
• 5:41
  separate
• 5:41
  changes
• 5:42
  degraded
• 5:42
  Claude
• 5:42
  Code's
• 5:43
  quality
• 5:43
  over
• 5:43
  the
• 5:43
  past
• 5:44
  two
• 5:44
  months,
• 5:44
  and
• 5:45
  all
• 5:45
  three
• 5:45
  prioritized
• 5:46
  infrastructure
• 5:46
  concerns
• 5:47
  over
• 5:47
  user
• 5:47
  experience.
• 5:48
  On
• 5:48
  March
• 5:48
  4,
• 5:49
  reasoning
• 5:49
  effort
• 5:49
  was
• 5:50
  quietly
• 5:50
  dropped
• 5:50
  from
• 5:50
  high
• 5:51
  to
• 5:51
  medium
• 5:51
  to
• 5:51
  reduce
• 5:52
  latency
• 5:52
  —
• 5:53
  "the
• 5:53
  wrong
• 5:53
  tradeoff,"
• 5:54
  Anthropic
• 5:54
  now
• 5:55
  admits.
• 5:56
  On
• 5:56
  March
• 5:56
  26,
• 5:58
  a
• 5:58
  bug
• 5:58
  in
• 5:58
  session
• 5:59
  thinking
• 5:59
  management
• 5:59
  caused
• 6:00
  Claude
• 6:00
  to
• 6:00
  seem
• 6:00
  "forgetful
• 6:01
  and
• 6:01
  repetitive"
• 6:01
  for
• 6:01
  two
• 6:02
  weeks
• 6:02
  before
• 6:02
  being
• 6:02
  fixed.
• 6:04
  On
• 6:04
  April
• 6:04
  16,
• 6:05
  a
• 6:05
  verbosity-reducing
• 6:06
  system
• 6:06
  prompt
• 6:07
  "hurt
• 6:07
  coding
• 6:07
  quality"
• 6:08
  and
• 6:08
  was
• 6:08
  reverted
• 6:09
  four
• 6:09
  days
• 6:09
  later.
• 6:10
  Each
• 6:11
  change
• 6:11
  was
• 6:11
  made
• 6:11
  without
• 6:12
  informing
• 6:12
  paying
• 6:12
  customers.
• 6:13
  As
• 6:14
  one
• 6:14
  commenter
• 6:14
  noted
• 6:14
  with
• 6:15
  precision:
• 6:16
  "If
• 6:16
  a
• 6:16
  hosted
• 6:16
  model
• 6:16
  has
• 6:17
  been
• 6:17
  quantized
• 6:17
  or
• 6:18
  in
• 6:18
  some
• 6:18
  way
• 6:18
  had
• 6:18
  its
• 6:18
  capabilities
• 6:19
  reduced,
• 6:20
  I
• 6:20
  should
• 6:20
  get
• 6:20
  a
• 6:20
  discount.
• 6:21
  The
• 6:22
  price
• 6:22
  should
• 6:22
  be
• 6:22
  per
• 6:22
  quant."
• 6:23
  The
• 6:23
  timing
• 6:24
  could
• 6:24
  not
• 6:24
  be
• 6:24
  better
• 6:24
  for
• 6:25
  the
• 6:25
  local
• 6:25
  model
• 6:25
  movement.
• 6:26
  Users
• 6:27
  are
• 6:27
  actively
• 6:27
  migrating
• 6:28
  coding
• 6:28
  workflows
• 6:28
  from
• 6:28
  Opus
• 6:29
  4.7
• 6:29
  to
• 6:29
  Qwen3.6-35B-A3B,
• 6:35
  a
• 6:35
  mixture-of-experts
• 6:36
  model
• 6:36
  with
• 6:36
  35
• 6:37
  billion
• 6:37
  total
• 6:37
  parameters
• 6:37
  and
• 6:38
  only
• 6:38
  3
• 6:38
  billion
• 6:38
  active
• 6:39
  per
• 6:39
  token.
• 6:40
  On
• 6:40
  dual
• 6:40
  3090s
• 6:41
  at
• 6:41
  Q6
• 6:42
  quantization,
• 6:43
  it
• 6:43
  runs
• 6:43
  at
• 6:43
  approximately
• 6:44
  120
• 6:45
  tokens
• 6:45
  per
• 6:45
  second
• 6:45
  with
• 6:46
  full
• 6:46
  context
• 6:46
  and
• 6:46
  prompt
• 6:47
  caching.
• 6:48
  The
• 6:48
  practical
• 6:48
  workflow
• 6:49
  that
• 6:49
  is
• 6:49
  emerging:
• 6:50
  use
• 6:50
  a
• 6:50
  frontier
• 6:50
  model
• 6:51
  (Opus,
• 6:51
  GPT-5)
• 6:52
  for
• 6:52
  planning,
• 6:53
  then
• 6:53
  feed
• 6:54
  that
• 6:54
  plan
• 6:54
  into
• 6:54
  Qwen
• 6:54
  for
• 6:55
  implementation.
• 6:56
  One
• 6:56
  developer
• 6:56
  running
• 6:57
  this
• 6:57
  setup
• 6:57
  reports
• 6:57
  it
• 6:57
  replaced
• 6:58
  95%
• 6:59
  of
• 6:59
  cloud
• 6:59
  API
• 6:59
  The
• 7:00
  speed
• 7:00
  differential
• 7:00
  is
• 7:01
  dramatic
• 7:01
  —
• 7:02
  "it's
• 7:02
  really
• 7:02
  hard
• 7:03
  to
• 7:03
  go
• 7:03
  back
• 7:03
  to
• 7:03
  Opus
• 7:03
  after
• 7:04
  that...
• 7:04
  no
• 7:04
  more
• 7:04
  alt
• 7:05
  tabbing
• 7:05
  for
• 7:05
  20
• 7:05
  minutes."
• 7:06
  The
• 7:07
  tradeoff
• 7:07
  is
• 7:07
  real:
• 7:08
  Qwen
• 7:08
  lacks
• 7:08
  knowledge
• 7:09
  of
• 7:09
  niche
• 7:09
  frameworks
• 7:09
  and
• 7:09
  sometimes
• 7:10
  needs
• 7:10
  hand-holding,
• 7:11
  but
• 7:11
  the
• 7:11
  iteration
• 7:12
  velocity
• 7:12
  compensates.
• 7:13
  This
• 7:13
  raises
• 7:14
  a
• 7:14
  question
• 7:14
  that
• 7:14
  keeps
• 7:15
  surfacing
• 7:15
  in
• 7:15
  the
• 7:15
  local
• 7:15
  LLM
• 7:16
  community:
• 7:17
  how
• 7:17
  can
• 7:17
  a
• 7:17
  27B
• 7:18
  dense
• 7:18
  model
• 7:18
  outperform
• 7:19
  a
• 7:19
  397B
• 7:20
  MoE
• 7:20
  model
• 7:21
  on
• 7:21
  benchmarks?
• 7:22
  The
• 7:22
  technical
• 7:23
  answer
• 7:23
  involves
• 7:23
  the
• 7:23
  difference
• 7:24
  between
• 7:24
  memorized
• 7:25
  knowledge
• 7:25
  and
• 7:25
  learned
• 7:25
  generative
• 7:26
  functions.
• 7:27
  Larger
• 7:27
  models
• 7:28
  can
• 7:28
  afford
• 7:28
  to
• 7:28
  "lazily
• 7:29
  interpolate
• 7:29
  over
• 7:30
  their
• 7:30
  memories"
• 7:31
  while
• 7:31
  smaller
• 7:31
  models
• 7:31
  must
• 7:32
  develop
• 7:32
  more
• 7:32
  generalizable
• 7:33
  problem-solving
• 7:34
  capabilities
• 7:34
  —
• 7:35
  essentially,
• 7:36
  they
• 7:36
  are
• 7:36
  forced
• 7:36
  to
• 7:37
  be
• 7:37
  smarter
• 7:37
  because
• 7:37
  they
• 7:38
  lack
• 7:38
  the
• 7:38
  parameter
• 7:38
  budget
• 7:39
  to
• 7:39
  be
• 7:39
  lazy.
• 7:40
  As
• 7:40
  chain-of-thought
• 7:40
  and
• 7:41
  agentic
• 7:41
  training
• 7:41
  progresses,
• 7:42
  models
• 7:42
  learn
• 7:43
  increasingly
• 7:43
  better
• 7:44
  "figure
• 7:44
  it
• 7:44
  out"
• 7:44
  functions,
• 7:45
  and
• 7:45
  those
• 7:45
  functions
• 7:46
  fit
• 7:46
  efficiently
• 7:47
  into
• 7:47
  smaller
• 7:47
  parameter
• 7:48
  counts.
• 7:49
  Current
• 7:49
  benchmarks,
• 7:49
  however,
• 7:50
  do
• 7:50
  not
• 7:50
  capture
• 7:51
  the
• 7:51
  larger
• 7:51
  model's
• 7:51
  advantages
• 7:52
  in
• 7:52
  world
• 7:52
  knowledge
• 7:53
  and
• 7:53
  long-context
• 7:54
  logical
• 7:54
  coherence.
• 7:57
  The
• 7:57
  tooling
• 7:57
  for
• 7:57
  removing
• 7:58
  safety
• 7:58
  training
• 7:59
  from
• 7:59
  language
• 7:59
  models
• 7:59
  has
• 8:00
  matured
• 8:00
  from
• 8:00
  research
• 8:00
  curiosity
• 8:01
  to
• 8:01
  industrial-grade
• 8:02
  automation.
• 8:03
  Abliterix,
• 8:04
  an
• 8:04
  open-source
• 8:05
  tool
• 8:05
  built
• 8:05
  on
• 8:05
  Heretic,
• 8:06
  uses
• 8:06
  Optuna
• 8:06
  TPE
• 8:07
  optimization
• 8:08
  to
• 8:08
  find
• 8:08
  the
• 8:08
  optimal
• 8:09
  abliteration
• 8:09
  parameters
• 8:10
  for
• 8:10
  any
• 8:10
  transformer
• 8:11
  architecture.
• 8:11
  It
• 8:12
  co-minimizes
• 8:12
  refusals
• 8:13
  and
• 8:13
  KL
• 8:14
  divergence
• 8:14
  —
• 8:15
  meaning
• 8:15
  it
• 8:15
  strips
• 8:15
  safety
• 8:16
  behavior
• 8:16
  while
• 8:16
  preserving
• 8:17
  as
• 8:17
  much
• 8:17
  of
• 8:17
  the
• 8:17
  model's
• 8:18
  intelligence
• 8:18
  as
• 8:18
  possible.
• 8:19
  The
• 8:19
  tool
• 8:20
  ships
• 8:20
  with
• 8:20
  150+
• 8:21
  pre-built
• 8:21
  configs
• 8:22
  covering
• 8:22
  dense,
• 8:23
  MoE,
• 8:23
  SSM/hybrid,
• 8:25
  and
• 8:25
  vision-language
• 8:26
  architectures.
• 8:27
  More
• 8:27
  significantly,
• 8:28
  it
• 8:28
  has
• 8:28
  end-to-end
• 8:28
  broken
• 8:29
  three
• 8:29
  of
• 8:29
  the
• 8:29
  strongest
• 8:30
  published
• 8:30
  anti-abliteration
• 8:31
  defenses:
• 8:32
  DeepRefusal
• 8:33
  (EMNLP
• 8:33
  2025)
• 8:34
  and
• 8:34
  Circuit
• 8:34
  Breakers
• 8:35
  /
• 8:35
  Representation
• 8:36
  Rerouting
• 8:36
  (NeurIPS
• 8:37
  2024).
• 8:38
  The
• 8:38
  recipe
• 8:38
  is
• 8:39
  minimal:
• 8:39
  diagnose
• 8:40
  the
• 8:40
  LoRA
• 8:40
  delta
• 8:40
  via
• 8:41
  SVD,
• 8:41
  lerp
• 8:42
  it
• 8:42
  away
• 8:42
  to
• 8:42
  recover
• 8:42
  base
• 8:42
  weights,
• 8:43
  then
• 8:43
  run
• 8:43
  single-direction
• 8:44
  abliteration.
• 8:45
  No
• 8:45
  fine-tuning,
• 8:46
  no
• 8:46
  manual
• 8:46
  prompt
• 8:46
  engineering.
• 8:47
  The
• 8:47
  Gemma
• 8:48
  4
• 8:48
  E4B
• 8:48
  model
• 8:49
  achieved
• 8:49
  a
• 8:49
  7%
• 8:50
  refusal
• 8:50
  rate
• 8:50
  with
• 8:50
  KL
• 8:51
  divergence
• 8:51
  of
• 8:51
  just
• 8:51
  0.0006
• 8:52
  —
• 8:53
  the
• 8:54
  model
• 8:54
  barely
• 8:54
  changed
• 8:55
  on
• 8:55
  harmless
• 8:55
  prompts.
• 8:56
  A
• 8:56
  concrete
• 8:57
  demonstration
• 8:57
  of
• 8:57
  Abliterix's
• 8:58
  capabilities
• 8:58
  arrived
• 8:59
  this
• 8:59
  week
• 8:59
  in
• 8:59
  the
• 8:59
  form
• 8:59
  of
• 9:00
  a
• 9:00
  Qwen3.6-35B-A3B
• 9:06
  abliterated
• 9:06
  model
• 9:07
  using
• 9:07
  Expert-Granular
• 9:08
  Abliteration
• 9:08
  (EGA)
• 9:08
  —
• 9:09
  a
• 9:10
  technique
• 9:10
  that
• 9:10
  applies
• 9:10
  the
• 9:10
  refusal
• 9:11
  direction
• 9:11
  removal
• 9:12
  simultaneously
• 9:13
  across
• 9:13
  all
• 9:13
  256
• 9:14
  routed
• 9:14
  experts
• 9:15
  in
• 9:15
  each
• 9:15
  MoE
• 9:15
  layer
• 9:16
  via
• 9:16
  vectorized
• 9:16
  The
• 9:17
  configuration
• 9:18
  tournament
• 9:18
  searched
• 9:18
  over
• 9:18
  hyperparameters
• 9:19
  including
• 9:20
  per-layer
• 9:20
  alpha
• 9:20
  values,
• 9:21
  decay
• 9:21
  kernels,
• 9:22
  and
• 9:22
  winsorization
• 9:23
  settings,
• 9:24
  then
• 9:24
  evaluated
• 9:24
  the
• 9:24
  Pareto-optimal
• 9:25
  point
• 9:25
  using
• 9:25
  three
• 9:26
  independent
• 9:26
  judges
• 9:26
  (GPT-4o-mini,
• 9:28
  Gemini
• 9:28
  3
• 9:29
  Flash
• 9:29
  Preview,
• 9:30
  and
• 9:30
  a
• 9:30
  local
• 9:30
  HarmBench
• 9:31
  classifier)
• 9:31
  across
• 9:31
  512
• 9:32
  prompts.
• 9:33
  The
• 9:33
  result:
• 9:34
  10.5%
• 9:35
  refusal
• 9:35
  rate
• 9:35
  under
• 9:36
  the
• 9:36
  strictest
• 9:36
  ensemble,
• 9:37
  with
• 9:37
  KL
• 9:37
  divergence
• 9:38
  of
• 9:38
  0.1153
• 9:39
  —
• 9:40
  low
• 9:41
  enough
• 9:41
  that
• 9:41
  the
• 9:41
  model's
• 9:42
  general
• 9:42
  capabilities
• 9:43
  remain
• 9:43
  intact.
• 9:44
  The
• 9:44
  vision
• 9:44
  pathway
• 9:45
  remains
• 9:45
  unmodified,
• 9:46
  so
• 9:46
  image-based
• 9:47
  harmful
• 9:47
  requests
• 9:48
  may
• 9:48
  still
• 9:48
  be
• 9:48
  refused
• 9:49
  even
• 9:49
  when
• 9:49
  equivalent
• 9:50
  text
• 9:50
  requests
• 9:50
  are
• 9:51
  not.
• 9:51
  The
• 9:51
  foundational
• 9:52
  research
• 9:52
  enabling
• 9:53
  models
• 9:53
  of
• 9:53
  this
• 9:54
  scale
• 9:54
  —
• 9:54
  billions
• 9:55
  of
• 9:55
  parameters
• 9:56
  trained
• 9:56
  across
• 9:56
  hundreds
• 9:57
  of
• 9:57
  GPUs
• 9:57
  using
• 9:58
  intra-layer
• 9:58
  model
• 9:58
  parallelism
• 9:59
  —
• 10:00
  was
• 10:00
  laid
• 10:00
  out
• 10:00
  in
• 10:00
  NVIDIA's
• 10:01
  Megatron-LM
• 10:02
  work,
• 10:02
  which
• 10:03
  demonstrated
• 10:03
  76%
• 10:04
  scaling
• 10:05
  efficiency
• 10:05
  across
• 10:06
  512
• 10:06
  GPUs
• 10:07
  for
• 10:07
  an
• 10:07
  8.3B
• 10:08
  parameter
• 10:09
  model.
• 10:09
  That
• 10:10
  infrastructure
• 10:10
  lineage
• 10:11
  is
• 10:11
  what
• 10:11
  makes
• 10:11
  both
• 10:11
  the
• 10:12
  models
• 10:12
  and
• 10:12
  their
• 10:12
  subsequent
• 10:13
  abliteration
• 10:13
  possible.
• 10:17
  Jensen
• 10:17
  Huang
• 10:17
  told
• 10:17
  the
• 10:18
  Dwarkesh
• 10:18
  Podcast
• 10:18
  this
• 10:19
  week
• 10:19
  that
• 10:19
  US
• 10:19
  chip
• 10:20
  export
• 10:20
  controls
• 10:20
  might
• 10:20
  be
• 10:21
  creating
• 10:21
  the
• 10:21
  problem
• 10:21
  they
• 10:21
  are
• 10:22
  trying
• 10:22
  to
• 10:22
  solve.
• 10:22
  His
• 10:23
  argument
• 10:23
  is
• 10:23
  not
• 10:23
  that
• 10:24
  China
• 10:24
  is
• 10:24
  not
• 10:24
  a
• 10:24
  threat,
• 10:25
  but
• 10:25
  that
• 10:25
  "victimising
• 10:26
  them
• 10:26
  and
• 10:26
  turning
• 10:26
  them
• 10:26
  into
• 10:26
  an
• 10:26
  enemy
• 10:27
  likely
• 10:27
  is
• 10:27
  not
• 10:27
  the
• 10:27
  best
• 10:28
  answer."
• 10:29
  The
• 10:29
  context
• 10:29
  is
• 10:29
  Huawei
• 10:30
  targeting
• 10:30
  750,000
• 10:31
  AI
• 10:31
  chip
• 10:32
  shipments
• 10:32
  this
• 10:32
  year
• 10:32
  —
• 10:33
  nowhere
• 10:33
  near
• 10:33
  NVIDIA's
• 10:34
  compute
• 10:34
  capacity,
• 10:34
  but
• 10:35
  the
• 10:35
  trajectory
• 10:35
  is
• 10:35
  clear.
• 10:36
  If
• 10:36
  DeepSeek
• 10:37
  optimizes
• 10:37
  its
• 10:37
  next
• 10:37
  model
• 10:38
  for
• 10:38
  Huawei's
• 10:38
  Ascend
• 10:38
  chips
• 10:39
  rather
• 10:39
  than
• 10:39
  NVIDIA
• 10:39
  hardware,
• 10:40
  the
• 10:40
  entire
• 10:41
  export
• 10:41
  control
• 10:41
  strategy
• 10:42
  starts
• 10:42
  to
• 10:42
  look
• 10:42
  counterproductive.
• 10:43
  Reddit's
• 10:43
  response
• 10:44
  was
• 10:44
  characteristically
• 10:44
  "guy
• 10:45
  who
• 10:45
  makes
• 10:45
  money
• 10:46
  selling
• 10:46
  chips
• 10:46
  wants
• 10:46
  to
• 10:46
  sell
• 10:47
  more
• 10:47
  chips."
• 10:48
  Fair,
• 10:48
  but
• 10:48
  the
• 10:48
  underlying
• 10:49
  data
• 10:49
  is
• 10:49
  real
• 10:49
  —
• 10:50
  Stanford's
• 10:50
  HAI
• 10:51
  report
• 10:51
  confirmed
• 10:51
  the
• 10:51
  China
• 10:52
  gap
• 10:52
  has
• 10:52
  "effectively
• 10:53
  closed,"
• 10:53
  and
• 10:54
  a
• 10:54
  Supermicro
• 10:54
  co-founder
• 10:55
  was
• 10:55
  arrested
• 10:55
  in
• 10:55
  March
• 10:55
  for
• 10:55
  allegedly
• 10:56
  running
• 10:56
  a
• 10:56
  $2
• 10:56
  billion
• 10:57
  GPU
• 10:57
  smuggling
• 10:58
  ring.
• 10:59
  The
• 10:59
  controls
• 10:59
  are
• 10:59
  being
• 10:59
  circumvented
• 11:00
  at
• 11:00
  industrial
• 11:00
  scale.
• 11:01
  A
• 11:01
  WSJ
• 11:02
  opinion
• 11:03
  piece
• 11:03
  by
• 11:03
  a16z
• 11:04
  partners
• 11:04
  argues
• 11:05
  the
• 11:05
  US
• 11:05
  should
• 11:05
  embrace
• 11:05
  open-source
• 11:06
  AI
• 11:06
  to
• 11:06
  maintain
• 11:07
  its
• 11:07
  competitive
• 11:07
  edge.
• 11:08
  The
• 11:08
  community
• 11:09
  reaction
• 11:09
  is
• 11:09
  split:
• 11:10
  some
• 11:10
  note
• 11:10
  that
• 11:11
  "most
• 11:11
  of
• 11:11
  the
• 11:11
  good
• 11:11
  ones
• 11:11
  are
• 11:11
  Chinese"
• 11:12
  researchers
• 11:12
  regardless
• 11:13
  of
• 11:13
  where
• 11:13
  they
• 11:13
  work,
• 11:14
  while
• 11:14
  others
• 11:14
  point
• 11:15
  out
• 11:15
  that
• 11:15
  US
• 11:15
  AI
• 11:16
  majors
• 11:16
  already
• 11:17
  build
• 11:17
  on
• 11:17
  Chinese
• 11:17
  open-source
• 11:18
  releases
• 11:18
  —
• 11:19
  Microsoft's
• 11:19
  recent
• 11:20
  audio
• 11:20
  transformer
• 11:21
  is
• 11:21
  based
• 11:21
  on
• 11:21
  Qwen
• 11:22
  3.5,
• 11:23
  NVIDIA's
• 11:23
  Lyra
• 11:24
  2.0
• 11:24
  is
• 11:25
  based
• 11:25
  on
• 11:25
  Wan
• 11:25
  Video.
• 11:26
  The
• 11:26
  open-source
• 11:27
  genie
• 11:27
  left
• 11:28
  the
• 11:28
  bottle
• 11:28
  years
• 11:28
  ago;
• 11:29
  the
• 11:29
  policy
• 11:29
  debate
• 11:29
  is
• 11:29
  mostly
• 11:30
  catching
• 11:30
  Meanwhile,
• 11:32
  Meta
• 11:32
  is
• 11:32
  reportedly
• 11:33
  planning
• 11:33
  to
• 11:33
  capture
• 11:34
  employee
• 11:34
  mouse
• 11:34
  movements
• 11:35
  and
• 11:35
  keystrokes
• 11:35
  for
• 11:35
  AI
• 11:36
  training
• 11:36
  data.
• 11:37
  The
• 11:37
  company's
• 11:37
  internal
• 11:38
  data
• 11:38
  collection
• 11:38
  ambitions
• 11:39
  are
• 11:39
  not
• 11:39
  new
• 11:39
  —
• 11:40
  a
• 11:40
  Swedish
• 11:41
  investigation
• 11:41
  earlier
• 11:42
  this
• 11:42
  year
• 11:42
  exposed
• 11:42
  how
• 11:42
  Meta's
• 11:43
  smart
• 11:43
  glasses
• 11:43
  training
• 11:44
  pipeline
• 11:44
  had
• 11:44
  Sama
• 11:45
  subcontractors
• 11:45
  in
• 11:45
  Nairobi
• 11:46
  annotating
• 11:46
  video
• 11:47
  footage
• 11:47
  including
• 11:47
  bathroom
• 11:48
  visits
• 11:48
  and
• 11:48
  bank
• 11:48
  card
• 11:49
  details
• 11:49
  —
• 11:50
  but
• 11:50
  turning
• 11:50
  the
• 11:50
  surveillance
• 11:51
  apparatus
• 11:51
  on
• 11:51
  its
• 11:51
  own
• 11:51
  workforce
• 11:52
  represents
• 11:52
  a
• 11:52
  distinct
• 11:53
  escalation.
• 11:56
  AirTrain
• 11:57
  is
• 11:57
  an
• 11:57
  open-source
• 11:57
  project
• 11:58
  that
• 11:58
  asks
• 11:58
  a
• 11:59
  simple
• 11:59
  question:
• 12:00
  what
• 12:00
  if
• 12:00
  you
• 12:00
  could
• 12:00
  train
• 12:01
  ML
• 12:01
  models
• 12:01
  by
• 12:01
  pooling
• 12:02
  MacBooks
• 12:02
  over
• 12:02
  Wi-Fi?
• 12:04
  The
• 12:04
  answer
• 12:04
  hinges
• 12:04
  on
• 12:04
  the
• 12:05
  DiLoCo
• 12:05
  algorithm
• 12:06
  (Distributed
• 12:06
  Low-Communication),
• 12:07
  which
• 12:08
  reduces
• 12:08
  network
• 12:08
  synchronization
• 12:09
  requirements
• 12:09
  by
• 12:10
  500x
• 12:11
  compared
• 12:11
  to
• 12:11
  traditional
• 12:11
  distributed
• 12:12
  data
• 12:12
  parallelism.
• 12:13
  Instead
• 12:14
  of
• 12:14
  syncing
• 12:14
  gradients
• 12:14
  after
• 12:15
  every
• 12:15
  step
• 12:15
  (requiring
• 12:16
  ~50
• 12:16
  GB/s
• 12:17
  sustained
• 12:17
  bandwidth
• 12:18
  for
• 12:18
  a
• 12:18
  124M
• 12:19
  model),
• 12:20
  each
• 12:20
  Mac
• 12:20
  trains
• 12:21
  independently
• 12:21
  for
• 12:21
  500
• 12:22
  steps,
• 12:22
  then
• 12:22
  exchanges
• 12:23
  only
• 12:23
  the
• 12:23
  pseudo-gradient
• 12:24
  difference
• 12:24
  —
• 12:25
  a
• 12:25
  ~2
• 12:25
  second
• 12:26
  sync
• 12:26
  over
• 12:26
  standard
• 12:27
  Wi-Fi.
• 12:28
  Seven
• 12:28
  friends
• 12:28
  with
• 12:29
  M4
• 12:29
  MacBooks
• 12:30
  collectively
• 12:30
  match
• 12:30
  an
• 12:31
  A100's
• 12:32
  19.5
• 12:32
  TFLOPS
• 12:33
  for
• 12:33
  zero
• 12:34
  compute
• 12:34
  cost.
• 12:35
  The
• 12:35
  project
• 12:36
  includes
• 12:36
  "Sleep
• 12:37
  Swarms,"
• 12:38
  where
• 12:38
  Macs
• 12:38
  train
• 12:38
  automatically
• 12:39
  during
• 12:39
  configured
• 12:40
  nighttime
• 12:40
  windows
• 12:41
  and
• 12:41
  hand
• 12:41
  off
• 12:41
  checkpoints
• 12:42
  across
• 12:42
  timezones
• 12:43
  for
• 12:43
  24/7
• 12:44
  coverage,
• 12:44
  and
• 12:44
  "Dream
• 12:45
  Training,"
• 12:45
  where
• 12:46
  idle
• 12:46
  Macs
• 12:46
  generate
• 12:47
  synthetic
• 12:47
  training
• 12:47
  data
• 12:47
  scored
• 12:48
  for
• 12:48
  quality
• 12:48
  and
• 12:49
  mixed
• 12:49
  into
• 12:49
  subsequent
• 12:49
  batches.
• 12:50
  The
• 12:51
  gradient
• 12:51
  marketplace
• 12:51
  scores
• 12:52
  each
• 12:52
  worker's
• 12:52
  contribution
• 12:53
  by
• 12:53
  alignment,
• 12:54
  magnitude,
• 12:54
  history,
• 12:55
  and
• 12:55
  improvement
• 12:55
  —
• 12:56
  automatically
• 12:57
  downweighting
• 12:57
  bad
• 12:57
  workers
• 12:58
  without
• 12:58
  ejecting
• 12:58
  them.
• 12:59
  On
• 12:59
  the
• 12:59
  browser
• 12:59
  inference
• 13:00
  side,
• 13:00
  Hugging
• 13:00
  Face
• 13:01
  published
• 13:01
  a
• 13:01
  detailed
• 13:02
  guide
• 13:02
  to
• 13:02
  running
• 13:02
  Transformers.js
• 13:04
  inside
• 13:04
  a
• 13:04
  Chrome
• 13:04
  extension
• 13:05
  under
• 13:05
  Manifest
• 13:06
  V3.
• 13:07
  The
• 13:07
  architecture
• 13:08
  splits
• 13:08
  a
• 13:08
  background
• 13:09
  service
• 13:09
  worker
• 13:09
  (hosting
• 13:10
  models
• 13:10
  and
• 13:11
  orchestrating
• 13:11
  agent
• 13:11
  loops)
• 13:12
  from
• 13:12
  a
• 13:12
  side
• 13:12
  panel
• 13:12
  UI
• 13:13
  and
• 13:13
  content
• 13:13
  script.
• 13:14
  Gemma
• 13:14
  4
• 13:15
  handles
• 13:15
  reasoning
• 13:16
  and
• 13:16
  tool
• 13:16
  decisions
• 13:17
  while
• 13:17
  MiniLM
• 13:17
  generates
• 13:18
  embeddings
• 13:18
  for
• 13:18
  semantic
• 13:19
  search
• 13:19
  —
• 13:20
  all
• 13:21
  running
• 13:21
  locally
• 13:21
  in
• 13:21
  the
• 13:21
  extension
• 13:22
  runtime
• 13:22
  with
• 13:22
  models
• 13:23
  cached
• 13:23
  under
• 13:23
  the
• 13:24
  extension
• 13:24
  origin.
• 13:25
  The
• 13:25
  practical
• 13:26
  design
• 13:26
  rule:
• 13:27
  background
• 13:28
  owns
• 13:28
  orchestration
• 13:28
  and
• 13:29
  inference,
• 13:30
  UI
• 13:30
  surfaces
• 13:30
  stay
• 13:31
  thin,
• 13:31
  content
• 13:32
  scripts
• 13:32
  handle
• 13:32
  DOM
• 13:33
  access.
• 13:34
  A
• 13:34
  deceptively
• 13:34
  simple
• 13:35
  idea
• 13:35
  surfaced
• 13:35
  on
• 13:36
  LocalLLaMA
• 13:36
  this
• 13:36
  week:
• 13:37
  why
• 13:37
  do
• 13:37
  we
• 13:37
  sample
• 13:38
  reasoning
• 13:38
  tokens
• 13:39
  and
• 13:39
  output
• 13:39
  tokens
• 13:39
  with
• 13:40
  the
• 13:40
  same
• 13:40
  parameters?
• 13:41
  One
• 13:41
  developer
• 13:42
  implemented
• 13:42
  separate
• 13:42
  sampler
• 13:43
  overrides
• 13:43
  in
• 13:43
  llama.cpp
• 13:44
  —
• 13:45
  high
• 13:46
  temperature
• 13:46
  for
• 13:46
  thinking
• 13:46
  (exploration)
• 13:47
  and
• 13:47
  low
• 13:47
  temperature
• 13:48
  for
• 13:48
  output
• 13:48
  (precision).
• 13:49
  On
• 13:49
  Gemma
• 13:49
  4,
• 13:50
  using
• 13:50
  temp
• 13:51
  1.0
• 13:51
  for
• 13:52
  thinking
• 13:52
  and
• 13:52
  0.0
• 13:53
  for
• 13:53
  output
• 13:53
  produced
• 13:54
  the
• 13:54
  best
• 13:54
  Ukrainian
• 13:54
  grammar
• 13:55
  seen
• 13:55
  so
• 13:55
  far
• 13:55
  while
• 13:55
  maintaining
• 13:56
  non-deterministic
• 13:57
  variation.
• 13:58
  The
• 13:58
  intuition
• 13:58
  maps
• 13:59
  cleanly:
• 14:00
  during
• 14:00
  reasoning,
• 14:00
  you
• 14:01
  want
• 14:01
  entropy
• 14:01
  so
• 14:01
  the
• 14:01
  model
• 14:02
  can
• 14:02
  search
• 14:02
  and
• 14:02
  branch;
• 14:03
  during
• 14:03
  output,
• 14:04
  you
• 14:04
  want
• 14:04
  the
• 14:04
  surface
• 14:04
  form
• 14:05
  locked
• 14:05
  down,
• 14:06
  especially
• 14:06
  in
• 14:06
  languages
• 14:07
  where
• 14:07
  the
• 14:07
  model
• 14:07
  is
• 14:07
  less
• 14:07
  confident.
• 14:08
  Several
• 14:09
  commenters
• 14:09
  noted
• 14:09
  they
• 14:09
  would
• 14:10
  not
• 14:10
  be
• 14:10
  surprised
• 14:10
  if
• 14:11
  major
• 14:11
  cloud
• 14:11
  providers
• 14:12
  already
• 14:12
  do
• 14:12
  something
• 14:12
  similar
• 14:13
  as
• 14:13
  part
• 14:14
  of
• 14:14
  their
• 14:14
  "Secret
• 14:14
  Sauce."
• 14:17
  South
• 14:18
  Korean
• 14:18
  police
• 14:18
  arrested
• 14:19
  a
• 14:19
  40-year-old
• 14:19
  man
• 14:20
  for
• 14:20
  sharing
• 14:20
  an
• 14:20
  AI-generated
• 14:21
  image
• 14:21
  of
• 14:21
  Neukgu,
• 14:22
  a
• 14:22
  two-year-old
• 14:23
  wolf
• 14:23
  that
• 14:23
  escaped
• 14:23
  from
• 14:23
  a
• 14:23
  zoo
• 14:24
  in
• 14:24
  Daejeon
• 14:24
  on
• 14:25
  April
• 14:25
  8.
• 14:26
  The
• 14:26
  fabricated
• 14:26
  photo
• 14:27
  prompted
• 14:27
  the
• 14:27
  city
• 14:27
  government
• 14:28
  to
• 14:28
  issue
• 14:28
  an
• 14:28
  emergency
• 14:29
  text
• 14:29
  alert,
• 14:30
  relocate
• 14:30
  their
• 14:30
  entire
• 14:31
  search
• 14:31
  operation,
• 14:32
  and
• 14:32
  present
• 14:32
  the
• 14:32
  AI
• 14:33
  image
• 14:33
  during
• 14:33
  an
• 14:33
  official
• 14:34
  press
• 14:34
  briefing
• 14:34
  —
• 14:35
  sending
• 14:35
  authorities
• 14:36
  on
• 14:36
  what
• 14:36
  can
• 14:36
  only
• 14:36
  be
• 14:37
  described
• 14:37
  as
• 14:37
  a
• 14:37
  wild
• 14:38
  wolf
• 14:38
  chase.
• 14:38
  The
• 14:39
  man
• 14:39
  told
• 14:39
  police
• 14:39
  he
• 14:39
  did
• 14:40
  it
• 14:40
  "for
• 14:40
  fun."
• 14:40
  He
• 14:41
  faces
• 14:41
  charges
• 14:41
  of
• 14:41
  disrupting
• 14:42
  government
• 14:42
  work
• 14:42
  by
• 14:42
  deception,
• 14:43
  carrying
• 14:43
  up
• 14:44
  to
• 14:44
  five
• 14:44
  years
• 14:44
  in
• 14:44
  prison.
• 14:45
  The
• 14:45
  wolf
• 14:46
  was
• 14:46
  eventually
• 14:46
  captured
• 14:47
  nine
• 14:47
  days
• 14:47
  after
• 14:47
  its
• 14:47
  escape
• 14:48
  near
• 14:48
  an
• 14:48
  expressway,
• 14:49
  becoming
• 14:49
  a
• 14:49
  national
• 14:50
  sensation
• 14:50
  —
• 14:51
  a
• 14:51
  local
• 14:52
  bakery
• 14:52
  now
• 14:52
  sells
• 14:53
  wolf-face
• 14:53
  pastries,
• 14:54
  and
• 14:54
  the
• 14:54
  city
• 14:54
  is
• 14:54
  considering
• 14:55
  naming
• 14:55
  Neukgu
• 14:56
  an
• 14:56
  official
• 14:56
  mascot.
• 14:58
  The
• 14:58
  case
• 14:58
  is
• 14:58
  one
• 14:58
  of
• 14:58
  the
• 14:58
  first
• 14:59
  criminal
• 14:59
  prosecutions
• 15:00
  specifically
• 15:00
  targeting
• 15:00
  AI-generated
• 15:01
  misinformation
• 15:02
  that
• 15:02
  directly
• 15:02
  diverted
• 15:03
  government
• 15:03
  resources.
• 15:04
  On
• 15:04
  more
• 15:04
  constructive
• 15:05
  applications,
• 15:06
  Rainmaker
• 15:06
  is
• 15:06
  an
• 15:07
  autonomous
• 15:07
  weather
• 15:07
  prediction
• 15:08
  agent
• 15:08
  claiming
• 15:09
  a
• 15:09
  73%
• 15:10
  win
• 15:10
  rate
• 15:10
  on
• 15:10
  Polymarket
• 15:10
  —
• 15:11
  turning
• 15:12
  LLM-driven
• 15:12
  analysis
• 15:13
  into
• 15:13
  real-money
• 15:13
  forecasting
• 15:14
  with
• 15:14
  measurable
• 15:15
  accountability.
• 15:16
  And
• 15:16
  Agent
• 15:16
  Sprite
• 15:17
  Forge
• 15:17
  offers
• 15:17
  a
• 15:17
  Codex-first
• 15:18
  pipeline
• 15:18
  for
• 15:18
  generating
• 15:19
  game-ready
• 15:19
  2D
• 15:20
  sprite
• 15:20
  sheets
• 15:20
  from
• 15:20
  natural
• 15:21
  language
• 15:21
  prompts,
• 15:22
  handling
• 15:22
  the
• 15:22
  full
• 15:23
  loop
• 15:23
  from
• 15:23
  asset
• 15:23
  planning
• 15:24
  through
• 15:24
  image
• 15:24
  generation
• 15:25
  to
• 15:25
  deterministic
• 15:25
  post-processing
• 15:26
  for
• 15:26
  chroma-key
• 15:27
  cleanup,
• 15:27
  frame
• 15:28
  extraction,
• 15:29
  and
• 15:29
  transparent
• 15:29
  PNG/GIF
• 15:31
  export.