Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 25, 2026

Transcript

AGI Dreams Podcast – June 25, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  NSA's
• 0:00
  own
• 0:01
  director
• 0:01
  reportedly
• 0:02
  told
• 0:02
  the
• 0:02
  Senate
• 0:02
  Intelligence
• 0:03
  Committee
• 0:03
  that
• 0:03
  Anthropic's
• 0:04
  Mythos
• 0:04
  model
• 0:05
  "broke
• 0:05
  into
• 0:05
  almost
• 0:05
  all
• 0:06
  of
• 0:06
  its
• 0:06
  classified
• 0:06
  systems
• 0:07
  in
• 0:07
  hours."
• 0:08
  Per
• 0:08
  The
• 0:08
  Economist,
• 0:09
  Senator
• 0:09
  Mark
• 0:10
  Warner
• 0:10
  —
• 0:11
  vice
• 0:11
  chair
• 0:11
  of
• 0:11
  the
• 0:11
  committee
• 0:12
  —
• 0:12
  said
• 0:13
  General
• 0:13
  Joshua
• 0:13
  Rudd
• 0:13
  conveyed
• 0:14
  this
• 0:14
  directly.
• 0:15
  The
• 0:15
  timing
• 0:16
  is
• 0:16
  pointed:
• 0:17
  the
• 0:17
  revelation
• 0:17
  surfaced
• 0:18
  on
• 0:18
  June
• 0:18
  11,
• 0:19
  the
• 0:19
  same
• 0:19
  day
• 0:19
  Amazon
• 0:20
  reportedly
• 0:20
  found
• 0:20
  a
• 0:20
  separate
• 0:21
  jailbreak
• 0:21
  in
• 0:21
  Anthropic's
• 0:22
  models.
• 0:23
  Within
• 0:23
  hours,
• 0:24
  the
• 0:24
  Trump
• 0:24
  administration
• 0:25
  ordered
• 0:25
  Anthropic
• 0:26
  to
• 0:26
  cut
• 0:26
  off
• 0:26
  foreign
• 0:27
  access
• 0:27
  to
• 0:27
  Mythos
• 0:27
  and
• 0:28
  Fable.
• 0:29
  Anthropic
• 0:29
  shut
• 0:30
  both
• 0:30
  models
• 0:30
  down
• 0:30
  entirely
• 0:31
  instead.
• 0:32
  Two
• 0:32
  competing
• 0:33
  narratives
• 0:33
  are
• 0:33
  circulating.
• 0:34
  One
• 0:35
  says
• 0:35
  the
• 0:35
  shutdown
• 0:35
  was
• 0:36
  a
• 0:36
  direct
• 0:36
  response
• 0:36
  to
• 0:36
  the
• 0:37
  NSA
• 0:37
  breach.
• 0:38
  The
• 0:38
  other
• 0:38
  says
• 0:39
  Anthropic
• 0:39
  considers
• 0:39
  the
• 0:40
  jailbreak
• 0:40
  minor
• 0:40
  —
• 0:41
  something
• 0:41
  other
• 0:41
  AI
• 0:42
  models
• 0:42
  can
• 0:42
  already
• 0:42
  be
• 0:42
  tricked
• 0:43
  into
• 0:43
  doing
• 0:43
  —
• 0:44
  and
• 0:44
  views
• 0:44
  the
• 0:44
  shutdown
• 0:45
  as
• 0:45
  an
• 0:45
  overreaction.
• 0:46
  The
• 0:46
  uncomfortable
• 0:47
  wrinkle:
• 0:48
  the
• 0:48
  NSA
• 0:48
  was
• 0:48
  already
• 0:48
  using
• 0:49
  Mythos
• 0:49
  for
• 0:49
  its
• 0:49
  own
• 0:49
  cyber
• 0:50
  operations,
• 0:51
  with
• 0:51
  Anthropic
• 0:51
  engineers
• 0:52
  embedded
• 0:52
  inside
• 0:52
  the
• 0:52
  agency.
• 0:54
  The
• 0:54
  same
• 0:54
  tool
• 0:54
  the
• 0:54
  agency
• 0:55
  relied
• 0:55
  on
• 0:55
  is
• 0:56
  the
• 0:56
  one
• 0:56
  its
• 0:56
  director
• 0:56
  says
• 0:57
  compromised
• 0:57
  nearly
• 0:57
  everything
• 0:58
  it
• 0:58
  owns.
• 0:59
  The
• 0:59
  strategic
• 0:59
  implications
• 1:00
  go
• 1:00
  beyond
• 1:00
  defense.
• 1:01
  If
• 1:01
  a
• 1:01
  model
• 1:01
  can
• 1:02
  find
• 1:02
  thousands
• 1:02
  of
• 1:02
  flaws
• 1:03
  in
• 1:03
  hardened
• 1:03
  classified
• 1:04
  networks
• 1:04
  in
• 1:04
  hours,
• 1:05
  any
• 1:05
  adversary
• 1:06
  with
• 1:06
  a
• 1:06
  comparable
• 1:06
  model
• 1:07
  can
• 1:07
  neutralize
• 1:07
  the
• 1:07
  NSA's
• 1:08
  offensive
• 1:08
  exploit
• 1:08
  stockpile
• 1:09
  by
• 1:09
  patching
• 1:10
  those
• 1:10
  same
• 1:10
  vulnerabilities
• 1:11
  before
• 1:11
  they
• 1:11
  can
• 1:11
  be
• 1:11
  weaponized.
• 1:12
  The
• 1:13
  era
• 1:13
  of
• 1:13
  hoarding
• 1:13
  zero-days
• 1:14
  as
• 1:14
  a
• 1:14
  strategic
• 1:14
  asset
• 1:15
  may
• 1:15
  be
• 1:15
  closing
• 1:15
  faster
• 1:16
  than
• 1:16
  anyone
• 1:16
  in
• 1:16
  Fort
• 1:17
  Meade
• 1:17
  anticipated.
• 1:18
  Meanwhile,
• 1:18
  Microsoft's
• 1:19
  BlueHat
• 1:19
  IL
• 1:20
  security
• 1:20
  conference
• 1:20
  showcased
• 1:21
  an
• 1:21
  LLM-based
• 1:22
  vulnerability
• 1:22
  variant
• 1:23
  hunter
• 1:23
  that
• 1:23
  inverts
• 1:23
  the
• 1:24
  classic
• 1:24
  SAST
• 1:24
  calculation.
• 1:25
  Instead
• 1:26
  of
• 1:26
  hand-writing
• 1:26
  per-language
• 1:27
  taint-flow
• 1:27
  rules,
• 1:28
  the
• 1:28
  tool
• 1:28
  takes
• 1:28
  a
• 1:28
  known
• 1:29
  bug
• 1:29
  —
• 1:30
  described
• 1:30
  in
• 1:30
  natural
• 1:30
  language
• 1:31
  —
• 1:31
  and
• 1:31
  uses
• 1:32
  an
• 1:32
  LLM
• 1:32
  to
• 1:32
  judge
• 1:33
  whether
• 1:33
  similar
• 1:33
  code
• 1:33
  patterns
• 1:34
  across
• 1:34
  different
• 1:34
  languages
• 1:35
  realize
• 1:35
  the
• 1:35
  same
• 1:35
  vulnerability
• 1:36
  concept.
• 1:37
  The
• 1:37
  four-skill
• 1:38
  pipeline
• 1:38
  distills
• 1:38
  a
• 1:38
  seed
• 1:39
  bug
• 1:39
  into
• 1:39
  a
• 1:39
  flow
• 1:39
  and
• 1:39
  a
• 1:40
  reason,
• 1:40
  finds
• 1:41
  similar
• 1:41
  flows
• 1:41
  across
• 1:42
  codebases
• 1:42
  regardless
• 1:43
  of
• 1:43
  language,
• 1:43
  correlates
• 1:44
  which
• 1:44
  flows
• 1:44
  are
• 1:44
  actually
• 1:45
  vulnerable,
• 1:46
  and
• 1:46
  then
• 1:46
  skeptically
• 1:47
  verifies
• 1:47
  each
• 1:47
  finding
• 1:48
  by
• 1:48
  arguing
• 1:48
  attacker
• 1:49
  control,
• 1:49
  reachability,
• 1:50
  and
• 1:50
  generating
• 1:51
  a
• 1:51
  proof
• 1:51
  of
• 1:51
  concept.
• 1:53
  Microsoft
• 1:53
  used
• 1:53
  it
• 1:53
  to
• 1:53
  scan
• 1:54
  60
• 1:54
  Azure
• 1:55
  hybrid
• 1:55
  cloud
• 1:55
  extensions
• 1:56
  written
• 1:56
  in
• 1:56
  everything
• 1:57
  from
• 1:57
  C
• 1:57
  to
• 1:57
  Bash,
• 1:58
  finding
• 1:58
  issues
• 1:59
  in
• 1:59
  12
• 1:59
  and
• 1:59
  validating
• 2:00
  the
• 2:00
  other
• 2:00
  48.
• 2:01
  The
• 2:01
  key
• 2:02
  insight
• 2:02
  is
• 2:02
  that
• 2:02
  logical
• 2:03
  vulnerabilities
• 2:03
  —
• 2:04
  flows
• 2:05
  that
• 2:05
  are
• 2:05
  legitimate
• 2:05
  in
• 2:05
  one
• 2:06
  context
• 2:06
  but
• 2:07
  become
• 2:07
  exploitable
• 2:07
  in
• 2:08
  another,
• 2:08
  like
• 2:08
  calling
• 2:09
  IMDS
• 2:09
  on-prem
• 2:10
  —
• 2:10
  are
• 2:10
  exactly
• 2:11
  the
• 2:11
  class
• 2:11
  that
• 2:11
  traditional
• 2:12
  SAST
• 2:12
  misses.
• 2:13
  In
• 2:13
  more
• 2:13
  conventional
• 2:14
  breach
• 2:14
  news,
• 2:15
  LastPass
• 2:15
  is
• 2:15
  notifying
• 2:16
  users
• 2:16
  of
• 2:16
  yet
• 2:17
  another
• 2:17
  data
• 2:17
  breach
• 2:17
  —
• 2:18
  this
• 2:18
  time
• 2:18
  through
• 2:19
  market
• 2:19
  research
• 2:19
  firm
• 2:20
  Klue,
• 2:20
  which
• 2:20
  had
• 2:20
  Salesforce
• 2:21
  and
• 2:21
  Gong
• 2:21
  integrations
• 2:22
  exposing
• 2:22
  customer
• 2:23
  names,
• 2:23
  phone
• 2:24
  numbers,
• 2:24
  email
• 2:24
  addresses,
• 2:25
  physical
• 2:25
  addresses,
• 2:26
  and
• 2:26
  support
• 2:26
  case
• 2:27
  data.
• 2:27
  Password
• 2:28
  vaults
• 2:28
  were
• 2:28
  reportedly
• 2:29
  not
• 2:29
  affected.
• 2:30
  This
• 2:30
  is
• 2:30
  the
• 2:30
  company's
• 2:31
  third
• 2:31
  major
• 2:31
  security
• 2:32
  incident
• 2:32
  after
• 2:32
  the
• 2:33
  2015
• 2:33
  hash
• 2:33
  theft
• 2:34
  and
• 2:34
  the
• 2:34
  2022
• 2:35
  source
• 2:35
  code
• 2:35
  compromise
• 2:36
  that
• 2:36
  exposed
• 2:36
  encrypted
• 2:37
  vault
• 2:37
  backups.
• 2:38
  At
• 2:38
  this
• 2:38
  point,
• 2:39
  asking
• 2:39
  users
• 2:40
  to
• 2:40
  "remain
• 2:40
  vigilant
• 2:41
  of
• 2:41
  potential
• 2:41
  phishing
• 2:41
  attacks"
• 2:42
  reads
• 2:42
  less
• 2:43
  like
• 2:43
  advice
• 2:43
  and
• 2:43
  more
• 2:44
  like
• 2:44
  an
• 2:44
  admission
• 2:44
  that
• 2:44
  the
• 2:44
  steady
• 2:45
  drip
• 2:45
  of
• 2:45
  leaked
• 2:45
  PII
• 2:46
  has
• 2:46
  made
• 2:46
  social
• 2:47
  engineering
• 2:47
  against
• 2:47
  the
• 2:47
  LastPass
• 2:48
  user
• 2:48
  base
• 2:49
  permanently
• 2:49
  easier.
• 2:50
  On
• 2:51
  the
• 2:51
  defensive
• 2:51
  tooling
• 2:51
  side,
• 2:52
  an
• 2:52
  open-source
• 2:53
  project
• 2:53
  called
• 2:53
  Hush
• 2:54
  addresses
• 2:54
  a
• 2:54
  problem
• 2:55
  AI
• 2:55
  coding
• 2:55
  agents
• 2:55
  have
• 2:56
  made
• 2:56
  worse:
• 2:56
  .env
• 2:57
  files
• 2:58
  sitting
• 2:58
  in
• 2:58
  worktrees
• 2:59
  where
• 2:59
  any
• 2:59
  agent
• 2:59
  can
• 3:00
  cat
• 3:00
  .env
• 3:00
  and
• 3:00
  surface
• 3:01
  secrets
• 3:01
  into
• 3:01
  a
• 3:01
  model's
• 3:02
  context.
• 3:03
  Hush
• 3:03
  keeps
• 3:04
  secrets
• 3:04
  age-encrypted
• 3:05
  outside
• 3:05
  the
• 3:05
  repo,
• 3:06
  with
• 3:06
  the
• 3:06
  master
• 3:06
  key
• 3:06
  in
• 3:06
  the
• 3:07
  macOS
• 3:07
  Keychain,
• 3:08
  and
• 3:08
  injects
• 3:08
  values
• 3:08
  only
• 3:09
  into
• 3:09
  the
• 3:09
  child
• 3:09
  process
• 3:10
  of
• 3:10
  the
• 3:10
  command
• 3:10
  that
• 3:10
  needs
• 3:10
  them.
• 3:11
  Agent
• 3:11
  contexts
• 3:12
  (Claude
• 3:12
  Code,
• 3:13
  Codex,
• 3:13
  or
• 3:14
  any
• 3:14
  non-TTY
• 3:14
  session)
• 3:15
  are
• 3:15
  auto-detected
• 3:16
  and
• 3:16
  locked
• 3:16
  to
• 3:16
  "use,
• 3:17
  don't
• 3:17
  see."
• 3:18
  The
• 3:18
  design
• 3:18
  philosophy
• 3:19
  is
• 3:19
  refreshingly
• 3:20
  honest:
• 3:20
  Hush
• 3:21
  explicitly
• 3:21
  states
• 3:22
  it
• 3:22
  stops
• 3:22
  accidental
• 3:23
  exposure
• 3:23
  and
• 3:23
  agent
• 3:24
  reflexes,
• 3:25
  not
• 3:25
  a
• 3:25
  determined
• 3:25
  local
• 3:26
  attacker
• 3:26
  sharing
• 3:26
  your
• 3:26
  uid.
• 3:30
  WIRED
• 3:30
  verified
• 3:31
  the
• 3:31
  leaked
• 3:31
  membership
• 3:31
  of
• 3:31
  Dialog,
• 3:32
  a
• 3:32
  private
• 3:33
  society
• 3:33
  that
• 3:33
  operated
• 3:34
  for
• 3:34
  20
• 3:34
  years
• 3:34
  with
• 3:35
  no
• 3:35
  public
• 3:35
  website.
• 3:36
  The
• 3:36
  registration
• 3:37
  list
• 3:37
  for
• 3:37
  its
• 3:37
  August
• 3:37
  2026
• 3:38
  retreat
• 3:38
  near
• 3:39
  Dublin
• 3:39
  —
• 3:40
  $16,000
• 3:40
  per
• 3:41
  head
• 3:41
  —
• 3:41
  includes
• 3:42
  OpenAI's
• 3:42
  Chief
• 3:42
  Strategy
• 3:43
  Officer,
• 3:43
  Google
• 3:44
  DeepMind's
• 3:44
  head
• 3:44
  of
• 3:44
  AI
• 3:45
  global
• 3:45
  affairs,
• 3:46
  the
• 3:46
  CEO
• 3:46
  of
• 3:46
  YouTube,
• 3:47
  and
• 3:47
  sitting
• 3:47
  government
• 3:48
  officials
• 3:48
  who
• 3:48
  used
• 3:48
  personal
• 3:49
  email
• 3:49
  to
• 3:49
  avoid
• 3:50
  FOIA.
• 3:51
  Peter
• 3:51
  Thiel
• 3:51
  co-founded
• 3:52
  the
• 3:52
  group;
• 3:52
  the
• 3:53
  co-founder
• 3:53
  of
• 3:53
  Palantir
• 3:54
  and
• 3:54
  the
• 3:54
  Treasury
• 3:54
  Secretary
• 3:55
  are
• 3:55
  also
• 3:55
  on
• 3:55
  the
• 3:56
  list.
• 3:56
  These
• 3:57
  networks
• 3:57
  are
• 3:57
  not
• 3:57
  new
• 3:57
  —
• 3:58
  the
• 3:58
  Bohemian
• 3:59
  Club
• 3:59
  has
• 3:59
  operated
• 3:59
  this
• 3:59
  way
• 4:00
  for
• 4:00
  a
• 4:00
  century
• 4:00
  —
• 4:01
  but
• 4:01
  the
• 4:01
  concentration
• 4:01
  of
• 4:02
  AI
• 4:02
  executives
• 4:02
  alongside
• 4:03
  regulators
• 4:03
  nominally
• 4:04
  overseeing
• 4:04
  them
• 4:04
  creates
• 4:05
  documented
• 4:05
  conflicts
• 4:06
  of
• 4:06
  interest
• 4:06
  that
• 4:06
  deserve
• 4:07
  scrutiny
• 4:07
  as
• 4:07
  AI
• 4:07
  policy
• 4:08
  decisions
• 4:08
  accelerate.
• 4:09
  On
• 4:09
  the
• 4:10
  measurement
• 4:10
  front,
• 4:10
  an
• 4:11
  editorial
• 4:11
  makes
• 4:11
  a
• 4:11
  useful
• 4:12
  argument:
• 4:12
  the
• 4:13
  easiest
• 4:13
  way
• 4:13
  to
• 4:13
  lie
• 4:13
  with
• 4:14
  AI
• 4:14
  is
• 4:14
  to
• 4:14
  focus
• 4:14
  only
• 4:15
  on
• 4:15
  benchmark
• 4:15
  capability
• 4:16
  while
• 4:16
  ignoring
• 4:17
  cost.
• 4:17
  A
• 4:17
  company
• 4:18
  handling
• 4:18
  100,000
• 4:19
  software
• 4:19
  engineering
• 4:19
  tasks
• 4:20
  at
• 4:20
  $10–$30
• 4:21
  each
• 4:22
  can
• 4:22
  spend
• 4:22
  millions
• 4:22
  more
• 4:23
  annually
• 4:23
  by
• 4:23
  chasing
• 4:24
  the
• 4:24
  highest
• 4:24
  leaderboard
• 4:24
  score
• 4:25
  when
• 4:25
  a
• 4:25
  system
• 4:25
  ten
• 4:25
  times
• 4:26
  cheaper
• 4:26
  delivers
• 4:27
  similar
• 4:27
  real-world
• 4:28
  performance.
• 4:29
  The
• 4:29
  better
• 4:29
  approach:
• 4:29
  plot
• 4:30
  cost
• 4:30
  per
• 4:30
  task
• 4:31
  against
• 4:31
  success
• 4:31
  rate
• 4:32
  and
• 4:32
  examine
• 4:32
  the
• 4:32
  Pareto
• 4:33
  frontier.
• 4:34
  A
• 4:34
  startup
• 4:34
  optimizes
• 4:35
  for
• 4:35
  cost;
• 4:36
  an
• 4:36
  enterprise
• 4:36
  pays
• 4:37
  for
• 4:37
  marginal
• 4:37
  accuracy.
• 4:38
  Neither
• 4:39
  is
• 4:39
  universally
• 4:39
  correct,
• 4:40
  which
• 4:40
  is
• 4:40
  precisely
• 4:41
  why
• 4:41
  single-number
• 4:41
  leaderboards
• 4:42
  —
• 4:42
  which
• 4:43
  measure
• 4:43
  capability
• 4:44
  without
• 4:44
  cost
• 4:44
  —
• 4:45
  are
• 4:45
  misleading
• 4:45
  by
• 4:46
  design.
• 4:49
  Apple's
• 4:49
  WWDC
• 4:50
  2026
• 4:51
  output
• 4:51
  clarifies
• 4:52
  that
• 4:52
  the
• 4:52
  company's
• 4:52
  AI
• 4:52
  strategy
• 4:53
  is
• 4:53
  less
• 4:53
  about
• 4:53
  competing
• 4:54
  on
• 4:54
  frontier
• 4:54
  model
• 4:55
  capability
• 4:55
  and
• 4:56
  more
• 4:56
  about
• 4:56
  owning
• 4:56
  the
• 4:56
  on-device
• 4:57
  runtime.
• 4:58
  Core
• 4:58
  AI
• 4:58
  is
• 4:58
  a
• 4:58
  comprehensive
• 4:59
  Swift
• 4:59
  framework
• 4:59
  for
• 4:59
  loading
• 5:00
  and
• 5:00
  running
• 5:00
  AI
• 5:00
  models
• 5:01
  entirely
• 5:01
  on
• 5:01
  Apple
• 5:01
  Silicon
• 5:02
  —
• 5:02
  zero
• 5:03
  server
• 5:03
  dependencies,
• 5:04
  zero
• 5:04
  token
• 5:04
  costs.
• 5:05
  Models
• 5:06
  are
• 5:06
  automatically
• 5:06
  specialized
• 5:07
  per
• 5:07
  hardware
• 5:07
  target
• 5:08
  with
• 5:08
  ahead-of-time
• 5:08
  compilation
• 5:09
  for
• 5:09
  instant
• 5:09
  load
• 5:10
  times.
• 5:11
  The
• 5:11
  framework
• 5:11
  includes
• 5:12
  PyTorch
• 5:12
  extensions
• 5:13
  for
• 5:13
  converting
• 5:13
  models
• 5:13
  into
• 5:14
  Core
• 5:14
  AI
• 5:14
  assets
• 5:15
  with
• 5:15
  hardware-optimized
• 5:15
  attention
• 5:16
  and
• 5:16
  normalization
• 5:17
  operations,
• 5:18
  support
• 5:18
  for
• 5:18
  custom
• 5:18
  Metal
• 5:19
  4
• 5:19
  kernels,
• 5:20
  and
• 5:20
  Core
• 5:20
  AI
• 5:20
  Optimization
• 5:21
  —
• 5:22
  a
• 5:22
  quantization
• 5:22
  and
• 5:23
  palettization
• 5:23
  toolkit
• 5:24
  with
• 5:24
  per-layer
• 5:24
  granularity.
• 5:26
  A
• 5:26
  new
• 5:26
  Xcode
• 5:26
  integration
• 5:27
  and
• 5:27
  standalone
• 5:28
  Core
• 5:28
  AI
• 5:28
  Debugger
• 5:29
  let
• 5:29
  developers
• 5:29
  inspect
• 5:30
  computation
• 5:30
  graphs,
• 5:31
  profile
• 5:31
  performance,
• 5:32
  and
• 5:32
  validate
• 5:32
  artifacts
• 5:33
  before
• 5:33
  deployment.
• 5:34
  Separately,
• 5:35
  Apple
• 5:35
  open-sourced
• 5:36
  container,
• 5:37
  a
• 5:37
  Swift-based
• 5:37
  tool
• 5:38
  for
• 5:38
  running
• 5:38
  Linux
• 5:38
  containers
• 5:39
  as
• 5:39
  lightweight
• 5:39
  VMs
• 5:40
  on
• 5:40
  Apple
• 5:40
  Silicon.
• 5:41
  It
• 5:42
  consumes
• 5:42
  and
• 5:42
  produces
• 5:42
  OCI-compatible
• 5:43
  images,
• 5:44
  so
• 5:44
  pulling
• 5:45
  from
• 5:45
  and
• 5:45
  pushing
• 5:45
  to
• 5:45
  any
• 5:46
  standard
• 5:46
  registry
• 5:46
  works
• 5:47
  out
• 5:47
  of
• 5:47
  the
• 5:47
  box.
• 5:48
  The
• 5:48
  tool
• 5:48
  requires
• 5:49
  macOS
• 5:49
  26
• 5:50
  and
• 5:50
  leverages
• 5:50
  new
• 5:50
  virtualization
• 5:51
  and
• 5:51
  networking
• 5:52
  features
• 5:52
  in
• 5:52
  that
• 5:52
  release.
• 5:53
  The
• 5:54
  convergence
• 5:54
  is
• 5:54
  telling:
• 5:55
  Apple
• 5:56
  is
• 5:56
  building
• 5:56
  the
• 5:56
  full
• 5:56
  stack
• 5:57
  for
• 5:57
  developers
• 5:57
  who
• 5:57
  want
• 5:58
  to
• 5:58
  run
• 5:58
  AI
• 5:58
  models
• 5:58
  locally
• 5:59
  (Core
• 5:59
  AI)
• 5:59
  and
• 5:59
  test
• 6:00
  server-side
• 6:00
  infrastructure
• 6:01
  natively
• 6:01
  on
• 6:01
  Mac
• 6:01
  (container),
• 6:02
  making
• 6:02
  the
• 6:02
  Mac
• 6:03
  a
• 6:03
  more
• 6:03
  complete
• 6:03
  AI
• 6:04
  development
• 6:04
  platform
• 6:04
  without
• 6:05
  cloud
• 6:05
  dependency.
• 6:06
  Neither
• 6:06
  announcement
• 6:07
  individually
• 6:07
  reshapes
• 6:08
  the
• 6:08
  landscape,
• 6:08
  but
• 6:08
  together
• 6:09
  they
• 6:09
  make
• 6:09
  a
• 6:09
  clear
• 6:09
  strategic
• 6:10
  statement.
• 6:11
  While
• 6:11
  competitors
• 6:12
  race
• 6:12
  to
• 6:12
  train
• 6:12
  the
• 6:12
  biggest
• 6:12
  model
• 6:13
  and
• 6:13
  sell
• 6:13
  the
• 6:13
  most
• 6:13
  tokens,
• 6:14
  Apple
• 6:14
  is
• 6:15
  betting
• 6:15
  that
• 6:15
  who
• 6:15
  owns
• 6:15
  the
• 6:16
  runtime
• 6:16
  wins
• 6:16
  —
• 6:17
  and
• 6:17
  that
• 6:17
  the
• 6:17
  billion-device
• 6:18
  install
• 6:18
  base
• 6:19
  running
• 6:19
  Apple
• 6:19
  Silicon
• 6:20
  is
• 6:20
  the
• 6:20
  distribution
• 6:21
  advantage
• 6:21
  no
• 6:21
  amount
• 6:21
  of
• 6:21
  training
• 6:22
  compute
• 6:22
  can
• 6:22
  replicate.
• 6:25
  Ideogram
• 6:26
  4
• 6:26
  is
• 6:26
  the
• 6:27
  most
• 6:27
  significant
• 6:27
  open-weight
• 6:28
  image
• 6:28
  model
• 6:28
  release
• 6:29
  in
• 6:29
  months.
• 6:30
  Built
• 6:30
  from
• 6:30
  scratch
• 6:30
  as
• 6:30
  a
• 6:30
  9.3-billion-parameter
• 6:32
  flow-matching
• 6:33
  Diffusion
• 6:33
  Transformer,
• 6:34
  it
• 6:34
  uses
• 6:34
  a
• 6:34
  fully
• 6:35
  single-stream
• 6:35
  architecture:
• 6:36
  text
• 6:37
  and
• 6:37
  image
• 6:37
  tokens
• 6:38
  concatenated
• 6:38
  into
• 6:38
  one
• 6:38
  sequence,
• 6:39
  processed
• 6:40
  through
• 6:40
  the
• 6:40
  same
• 6:40
  34-layer
• 6:41
  transformer
• 6:41
  with
• 6:42
  no
• 6:42
  separate
• 6:42
  branches.
• 6:44
  Instead
• 6:44
  of
• 6:44
  CLIP
• 6:44
  or
• 6:44
  T5,
• 6:45
  the
• 6:45
  text
• 6:46
  encoder
• 6:46
  is
• 6:46
  Qwen3-VL-8B-Instruct
• 6:48
  —
• 6:50
  a
• 6:50
  full
• 6:51
  vision-language
• 6:51
  model
• 6:52
  —
• 6:52
  with
• 6:52
  hidden
• 6:53
  states
• 6:53
  from
• 6:53
  13
• 6:54
  intermediate
• 6:54
  layers
• 6:54
  concatenated
• 6:55
  for
• 6:55
  multi-scale
• 6:56
  semantic
• 6:56
  features.
• 6:58
  Trained
• 6:58
  exclusively
• 6:58
  on
• 6:59
  structured
• 6:59
  JSON
• 6:59
  captions,
• 7:00
  it
• 7:00
  offers
• 7:01
  explicit
• 7:01
  control
• 7:02
  over
• 7:02
  composition,
• 7:03
  bounding-box
• 7:03
  layout,
• 7:04
  color
• 7:04
  palettes,
• 7:04
  and
• 7:05
  typography.
• 7:06
  On
• 7:06
  the
• 7:07
  Design
• 7:07
  Arena
• 7:07
  leaderboard
• 7:08
  it
• 7:08
  ranks
• 7:08
  as
• 7:08
  the
• 7:08
  top
• 7:09
  open-weight
• 7:09
  model,
• 7:10
  trailing
• 7:10
  only
• 7:11
  proprietary
• 7:11
  GPT
• 7:12
  and
• 7:12
  Gemini.
• 7:14
  In
• 7:14
  a
• 7:14
  blind
• 7:14
  ContraLabs
• 7:15
  evaluation,
• 7:16
  professional
• 7:16
  designers
• 7:17
  chose
• 7:17
  it
• 7:17
  first
• 7:17
  47.9%
• 7:18
  of
• 7:19
  the
• 7:19
  time
• 7:19
  —
• 7:20
  ahead
• 7:20
  of
• 7:20
  Gemini
• 7:20
  3.1
• 7:21
  Flash
• 7:21
  at
• 7:22
  30%
• 7:22
  and
• 7:22
  FLUX.2
• 7:23
  max
• 7:24
  at
• 7:24
  15.5%.
• 7:26
  At
• 7:26
  9.3B
• 7:27
  parameters
• 7:28
  it
• 7:28
  delivers
• 7:28
  better
• 7:28
  text
• 7:29
  rendering
• 7:29
  than
• 7:29
  models
• 7:29
  2–8x
• 7:31
  its
• 7:31
  size.
• 7:32
  The
• 7:32
  open
• 7:33
  deep
• 7:33
  research
• 7:33
  space
• 7:34
  gets
• 7:34
  QUEST-35B,
• 7:38
  trained
• 7:38
  on
• 7:39
  32
• 7:39
  H100s
• 7:40
  at
• 7:40
  Ohio
• 7:41
  State
• 7:41
  with
• 7:41
  full
• 7:41
  recipe,
• 7:42
  weights,
• 7:42
  and
• 7:43
  8,000
• 7:43
  multi-step
• 7:44
  research
• 7:44
  traces
• 7:44
  released.
• 7:45
  Community
• 7:46
  reaction
• 7:46
  is
• 7:46
  cautiously
• 7:47
  positive:
• 7:48
  "multi-step
• 7:48
  research
• 7:49
  traces
• 7:49
  are
• 7:49
  dense,"
• 7:50
  but
• 7:50
  the
• 7:50
  real
• 7:51
  question
• 7:51
  is
• 7:51
  "how
• 7:51
  brittle
• 7:52
  it
• 7:52
  is
• 7:52
  to
• 7:52
  harness
• 7:52
  changes,
• 7:53
  because
• 7:53
  that
• 7:54
  is
• 7:54
  where
• 7:54
  these
• 7:54
  agents
• 7:54
  usually
• 7:55
  fall
• 7:55
  apart."
• 7:56
  Cohere's
• 7:57
  North
• 7:57
  Mini
• 7:57
  Code
• 7:58
  now
• 7:58
  ships
• 7:58
  a
• 7:58
  4-bit
• 7:59
  quantized
• 7:59
  version
• 8:00
  that
• 8:00
  runs
• 8:01
  on
• 8:01
  20GB
• 8:01
  of
• 8:01
  RAM,
• 8:02
  available
• 8:03
  through
• 8:03
  Ollama
• 8:03
  and
• 8:03
  OpenRouter,
• 8:04
  with
• 8:04
  community
• 8:05
  contributions
• 8:06
  including
• 8:06
  oMLX
• 8:07
  implementations
• 8:07
  and
• 8:08
  Q6
• 8:08
  GGUF
• 8:09
  quants
• 8:09
  arriving
• 8:10
  within
• 8:10
  days.
• 8:11
  The
• 8:11
  quantization
• 8:12
  frontier
• 8:12
  keeps
• 8:13
  pushing
• 8:13
  lower.
• 8:14
  EdgeRazor
• 8:14
  is
• 8:15
  a
• 8:15
  framework
• 8:15
  for
• 8:15
  mixed-precision
• 8:16
  quantization-aware
• 8:17
  distillation
• 8:18
  down
• 8:18
  to
• 8:18
  1.58
• 8:19
  bits
• 8:19
  —
• 8:20
  a
• 8:20
  blend
• 8:21
  of
• 8:21
  4-bit
• 8:22
  and
• 8:22
  ternary
• 8:22
  weights,
• 8:23
  with
• 8:23
  configurable
• 8:24
  ratios
• 8:24
  including
• 8:25
  2.79-bit
• 8:26
  and
• 8:26
  1.88-bit
• 8:27
  variants.
• 8:29
  It
• 8:29
  integrates
• 8:29
  logits,
• 8:30
  features,
• 8:31
  and
• 8:31
  attention
• 8:31
  distillation
• 8:32
  in
• 8:32
  a
• 8:32
  unified
• 8:32
  interface,
• 8:33
  targeting
• 8:34
  mobile
• 8:34
  and
• 8:34
  edge
• 8:35
  deployment.
• 8:36
  In
• 8:36
  the
• 8:36
  community
• 8:36
  fine-tune
• 8:37
  space,
• 8:37
  Qwable-3.6-27b
• 8:39
  applies
• 8:40
  a
• 8:40
  LoRA
• 8:40
  adapter
• 8:41
  to
• 8:41
  Qwen's
• 8:41
  27B
• 8:42
  base
• 8:42
  using
• 8:43
  a
• 8:43
  cleaned
• 8:43
  Fable
• 8:43
  5-style
• 8:44
  reasoning
• 8:44
  dataset,
• 8:45
  optimized
• 8:46
  for
• 8:46
  step-by-step
• 8:47
  technical
• 8:47
  responses
• 8:48
  and
• 8:48
  downstream
• 8:49
  GGUF
• 8:49
  conversion
• 8:49
  .
• 8:50
  ScenemaAI
• 8:51
  also
• 8:51
  dropped
• 8:52
  an
• 8:52
  audio
• 8:52
  generation
• 8:53
  model
• 8:53
  on
• 8:53
  Hugging
• 8:53
  Face,
• 8:54
  though
• 8:54
  details
• 8:55
  remain
• 8:55
  sparse
• 8:55
  at
• 8:56
  launch
• 8:56
  .
• 8:59
  NVIDIA
• 8:59
  NeMo
• 9:00
  AutoModel
• 9:00
  delivers
• 9:00
  the
• 9:01
  kind
• 9:01
  of
• 9:01
  gains
• 9:01
  that
• 9:02
  make
• 9:02
  MoE
• 9:02
  practitioners
• 9:03
  stop
• 9:03
  and
• 9:03
  re-read
• 9:04
  the
• 9:04
  benchmarks.
• 9:05
  Building
• 9:06
  on
• 9:06
  HuggingFace
• 9:06
  Transformers
• 9:07
  v5's
• 9:07
  first-class
• 9:08
  MoE
• 9:08
  support,
• 9:09
  it
• 9:09
  adds
• 9:09
  Expert
• 9:10
  Parallelism,
• 9:10
  DeepEP
• 9:11
  fused
• 9:11
  all-to-all
• 9:11
  dispatch,
• 9:12
  and
• 9:12
  TransformerEngine
• 9:13
  kernels
• 9:13
  —
• 9:14
  accessible
• 9:14
  by
• 9:14
  changing
• 9:15
  a
• 9:15
  single
• 9:15
  import
• 9:16
  line.
• 9:16
  On
• 9:16
  Qwen3-30B-A3B
• 9:19
  and
• 9:19
  Nemotron
• 9:20
  3
• 9:20
  Nano
• 9:20
  30B-A3B,
• 9:22
  this
• 9:22
  yields
• 9:23
  3.4–3.7x
• 9:25
  higher
• 9:25
  training
• 9:25
  throughput
• 9:26
  with
• 9:26
  29–32%
• 9:27
  less
• 9:28
  GPU
• 9:28
  memory
• 9:28
  versus
• 9:29
  the
• 9:29
  best
• 9:29
  Transformers
• 9:30
  v5
• 9:30
  configuration.
• 9:31
  Peak
• 9:32
  memory
• 9:32
  for
• 9:32
  Qwen3
• 9:32
  drops
• 9:33
  from
• 9:33
  68.2
• 9:34
  GiB
• 9:34
  to
• 9:34
  48.1
• 9:35
  GiB.
• 9:36
  At
• 9:36
  frontier
• 9:36
  scale,
• 9:37
  AutoModel
• 9:37
  enables
• 9:38
  full
• 9:38
  fine-tuning
• 9:38
  of
• 9:39
  the
• 9:39
  550B-parameter
• 9:40
  Nemotron
• 9:40
  3
• 9:41
  Ultra
• 9:41
  across
• 9:41
  128
• 9:42
  H100
• 9:42
  GPUs
• 9:42
  —
• 9:43
  a
• 9:44
  regime
• 9:44
  where
• 9:44
  Transformers
• 9:45
  v5
• 9:45
  simply
• 9:45
  runs
• 9:45
  out
• 9:46
  of
• 9:46
  memory.
• 9:47
  The
• 9:47
  key:
• 9:48
  Expert
• 9:48
  Parallelism
• 9:49
  is
• 9:49
  treated
• 9:49
  as
• 9:49
  orthogonal
• 9:50
  to
• 9:50
  data
• 9:50
  parallelism,
• 9:51
  so
• 9:51
  on
• 9:51
  8
• 9:52
  GPUs
• 9:52
  the
• 9:52
  system
• 9:52
  runs
• 9:53
  ep=8
• 9:53
  and
• 9:53
  dp=8
• 9:54
  simultaneously,
• 9:56
  with
• 9:56
  every
• 9:56
  GPU
• 9:56
  holding
• 9:57
  only
• 9:57
  1/8
• 9:57
  of
• 9:58
  the
• 9:58
  expert
• 9:58
  weights.
• 9:59
  Checkpoints
• 9:59
  remain
• 9:59
  standard
• 10:00
  HF
• 10:00
  safetensors,
• 10:01
  deployable
• 10:02
  on
• 10:02
  vLLM
• 10:02
  and
• 10:02
  SGLang
• 10:03
  without
• 10:03
  conversion.
• 10:04
  On
• 10:04
  the
• 10:04
  inference
• 10:05
  side,
• 10:05
  NVIDIA
• 10:05
  is
• 10:06
  claiming
• 10:06
  a
• 10:06
  15x
• 10:07
  speedup
• 10:07
  using
• 10:07
  diffusion-based
• 10:08
  LLM
• 10:08
  generation,
• 10:09
  where
• 10:10
  entire
• 10:10
  blocks
• 10:10
  of
• 10:10
  text
• 10:11
  are
• 10:11
  produced
• 10:11
  at
• 10:11
  once
• 10:11
  rather
• 10:12
  than
• 10:12
  token
• 10:12
  by
• 10:12
  token.
• 10:13
  Early
• 10:13
  adopters
• 10:14
  running
• 10:14
  the
• 10:14
  26B
• 10:15
  diffusion
• 10:15
  Gemma
• 10:16
  on
• 10:16
  AMD
• 10:16
  Strix
• 10:16
  Halo
• 10:17
  report
• 10:17
  the
• 10:17
  number
• 10:17
  is
• 10:18
  real
• 10:18
  but
• 10:18
  context-dependent:
• 10:19
  predictable
• 10:20
  outputs
• 10:20
  like
• 10:20
  JSON
• 10:21
  lock
• 10:21
  in
• 10:21
  at
• 10:21
  around
• 10:22
  70
• 10:22
  tok/s,
• 10:23
  while
• 10:23
  creative
• 10:24
  writing
• 10:24
  needs
• 10:24
  more
• 10:25
  correction
• 10:25
  passes
• 10:25
  and
• 10:26
  drops
• 10:26
  to
• 10:26
  about
• 10:26
  15.
• 10:27
  Critical
• 10:28
  caveats:
• 10:29
  bf16
• 10:29
  is
• 10:30
  mandatory
• 10:30
  (fp16
• 10:31
  produces
• 10:31
  garbage),
• 10:32
  the
• 10:32
  exact
• 10:33
  prompt
• 10:33
  format
• 10:33
  matters
• 10:34
  enormously,
• 10:35
  and
• 10:35
  the
• 10:35
  headline
• 10:35
  numbers
• 10:36
  rely
• 10:36
  on
• 10:36
  compiler
• 10:37
  tricks
• 10:37
  that
• 10:37
  do
• 10:37
  not
• 10:37
  yet
• 10:38
  work
• 10:38
  on
• 10:38
  AMD.
• 10:39
  The
• 10:39
  deeper
• 10:39
  question
• 10:40
  —
• 10:41
  if
• 10:41
  the
• 10:41
  diffusion
• 10:41
  model's
• 10:42
  output
• 10:42
  is
• 10:42
  "good
• 10:42
  enough"
• 10:43
  95%
• 10:44
  of
• 10:44
  the
• 10:44
  time,
• 10:44
  why
• 10:45
  bother
• 10:45
  with
• 10:45
  autoregressive
• 10:46
  verification
• 10:46
  at
• 10:47
  all
• 10:47
  —
• 10:47
  remains
• 10:48
  unresolved.
• 10:49
  But
• 10:49
  15x
• 10:50
  is
• 10:50
  a
• 10:50
  big
• 10:50
  enough
• 10:51
  number
• 10:51
  that
• 10:51
  even
• 10:51
  a
• 10:51
  fraction
• 10:52
  of
• 10:52
  the
• 10:52
  claimed
• 10:52
  speedup,
• 10:53
  on
• 10:53
  the
• 10:53
  right
• 10:53
  workloads,
• 10:54
  would
• 10:54
  change
• 10:55
  how
• 10:55
  people
• 10:55
  think
• 10:55
  about
• 10:56
  local
• 10:56
  inference
• 10:56
  budgets.
• 10:59
  Google
• 11:00
  has
• 11:00
  integrated
• 11:00
  computer
• 11:01
  use
• 11:01
  as
• 11:01
  a
• 11:01
  built-in
• 11:02
  tool
• 11:02
  in
• 11:02
  Gemini
• 11:02
  3.5
• 11:03
  Flash,
• 11:04
  making
• 11:04
  it
• 11:04
  a
• 11:04
  native
• 11:05
  capability
• 11:05
  rather
• 11:06
  than
• 11:06
  a
• 11:06
  standalone
• 11:06
  model.
• 11:07
  The
• 11:07
  system
• 11:07
  can
• 11:08
  see,
• 11:08
  reason
• 11:08
  about,
• 11:09
  and
• 11:09
  take
• 11:09
  actions
• 11:09
  across
• 11:10
  browser,
• 11:10
  mobile,
• 11:10
  and
• 11:10
  desktop
• 11:11
  environments.
• 11:12
  To
• 11:12
  mitigate
• 11:12
  prompt
• 11:13
  injection
• 11:13
  risks,
• 11:13
  Google
• 11:14
  uses
• 11:14
  targeted
• 11:14
  adversarial
• 11:15
  training
• 11:15
  and
• 11:15
  ships
• 11:16
  two
• 11:16
  enterprise
• 11:16
  safeguard
• 11:17
  systems:
• 11:18
  mandatory
• 11:18
  user
• 11:18
  confirmation
• 11:19
  for
• 11:19
  sensitive
• 11:19
  actions
• 11:20
  and
• 11:20
  automatic
• 11:20
  task
• 11:21
  termination
• 11:21
  when
• 11:21
  indirect
• 11:22
  prompt
• 11:22
  injection
• 11:22
  is
• 11:22
  detected.
• 11:23
  For
• 11:24
  developers
• 11:24
  who
• 11:24
  want
• 11:24
  Gemini's
• 11:25
  capabilities
• 11:25
  through
• 11:25
  a
• 11:25
  familiar
• 11:26
  interface,
• 11:27
  gemini-web2api
• 11:28
  reverse-engineers
• 11:29
  Gemini's
• 11:29
  web
• 11:30
  StreamGenerate
• 11:30
  protocol
• 11:31
  into
• 11:31
  an
• 11:31
  OpenAI-compatible
• 11:31
  API
• 11:32
  —
• 11:33
  zero
• 11:34
  cost,
• 11:34
  single
• 11:34
  Python
• 11:35
  file,
• 11:35
  with
• 11:35
  function
• 11:36
  calling
• 11:36
  support,
• 11:37
  multiple
• 11:37
  models
• 11:37
  including
• 11:38
  Flash
• 11:38
  Thinking
• 11:38
  for
• 11:38
  20k+
• 11:39
  character
• 11:39
  output,
• 11:40
  and
• 11:40
  adjustable
• 11:40
  reasoning
• 11:41
  depth.
• 11:41
  Anonymous
• 11:42
  access
• 11:42
  works
• 11:42
  for
• 11:42
  all
• 11:43
  models;
• 11:43
  real
• 11:44
  Pro
• 11:44
  routing
• 11:44
  requires
• 11:44
  a
• 11:44
  paid
• 11:45
  Gemini
• 11:45
  Advanced
• 11:45
  cookie.
• 11:46
  Mistral
• 11:47
  OCR
• 11:47
  4
• 11:47
  advances
• 11:47
  document
• 11:48
  intelligence
• 11:48
  beyond
• 11:49
  plain
• 11:49
  text
• 11:49
  extraction,
• 11:50
  returning
• 11:50
  bounding
• 11:51
  boxes,
• 11:51
  typed-block
• 11:52
  classification
• 11:52
  (titles,
• 11:53
  tables,
• 11:53
  equations,
• 11:54
  signatures),
• 11:55
  and
• 11:55
  inline
• 11:55
  confidence
• 11:55
  scores
• 11:56
  per
• 11:56
  page
• 11:56
  and
• 11:57
  per
• 11:57
  word.
• 11:58
  Independent
• 11:58
  annotators
• 11:59
  preferred
• 11:59
  OCR
• 11:59
  4
• 12:00
  over
• 12:00
  every
• 12:01
  competing
• 12:01
  system
• 12:02
  tested
• 12:02
  with
• 12:02
  win
• 12:02
  rates
• 12:03
  averaging
• 12:03
  72%,
• 12:04
  and
• 12:04
  it
• 12:05
  scored
• 12:05
  highest
• 12:05
  on
• 12:05
  OlmOCRBench
• 12:06
  at
• 12:06
  85.20.
• 12:09
  It
• 12:09
  supports
• 12:09
  170
• 12:10
  languages,
• 12:11
  runs
• 12:11
  in
• 12:11
  a
• 12:11
  single
• 12:12
  container
• 12:12
  for
• 12:12
  self-hosted
• 12:13
  deployment,
• 12:14
  and
• 12:14
  prices
• 12:14
  at
• 12:14
  $4
• 12:15
  per
• 12:15
  1,000
• 12:16
  pages
• 12:16
  ($2
• 12:16
  with
• 12:17
  batch
• 12:17
  discount).
• 12:18
  Early
• 12:18
  adopter
• 12:19
  Rogo
• 12:19
  reported
• 12:19
  "equivalent
• 12:20
  accuracy
• 12:21
  at
• 12:21
  roughly
• 12:21
  8x
• 12:22
  lower
• 12:22
  cost
• 12:22
  and
• 12:23
  17x
• 12:23
  lower
• 12:24
  latency"
• 12:25
  versus
• 12:25
  leading
• 12:25
  agentic
• 12:26
  document
• 12:26
  parsers
• 12:27
  on
• 12:27
  financial
• 12:27
  QA
• 12:27
  datasets.
• 12:29
  From
• 12:29
  research,
• 12:30
  AADvark
• 12:30
  out
• 12:30
  of
• 12:30
  MIT
• 12:31
  tackles
• 12:31
  a
• 12:31
  gap
• 12:32
  no
• 12:32
  prior
• 12:32
  agent-aided
• 12:33
  design
• 12:33
  system
• 12:34
  has
• 12:34
  solved:
• 12:35
  3D
• 12:35
  CAD
• 12:35
  assemblies
• 12:36
  with
• 12:36
  moving
• 12:36
  parts.
• 12:37
  The
• 12:37
  system
• 12:37
  places
• 12:38
  Gemini
• 12:38
  3
• 12:39
  Flash
• 12:39
  in
• 12:39
  an
• 12:39
  iterative
• 12:39
  loop
• 12:40
  —
• 12:40
  writing
• 12:40
  JSON
• 12:41
  definitions
• 12:41
  for
• 12:41
  parts
• 12:42
  and
• 12:42
  joints,
• 12:42
  compiling
• 12:43
  via
• 12:43
  a
• 12:43
  modified
• 12:44
  OndselSolver,
• 12:45
  and
• 12:45
  refining
• 12:45
  based
• 12:46
  on
• 12:46
  FreeCAD
• 12:46
  renderings
• 12:47
  augmented
• 12:47
  with
• 12:47
  unique
• 12:47
  edge
• 12:48
  colors
• 12:48
  and
• 12:48
  text
• 12:48
  identifiers.
• 12:50
  The
• 12:50
  team
• 12:50
  had
• 12:50
  to
• 12:50
  switch
• 12:51
  the
• 12:51
  solver
• 12:51
  from
• 12:51
  Euler
• 12:52
  angles
• 12:52
  to
• 12:52
  quaternions
• 12:53
  and
• 12:53
  make
• 12:53
  error
• 12:54
  messages
• 12:54
  far
• 12:54
  more
• 12:54
  informative
• 12:55
  to
• 12:55
  compensate
• 12:56
  for
• 12:56
  VLMs'
• 12:56
  spatial
• 12:57
  reasoning
• 12:57
  limitations.
• 12:59
  Generating
• 12:59
  a
• 12:59
  functional
• 13:00
  pair
• 13:00
  of
• 13:00
  scissors
• 13:00
  took
• 13:01
  20
• 13:01
  iterations,
• 13:02
  4.14
• 13:03
  hours,
• 13:03
  $15.85,
• 13:05
  and
• 13:05
  468
• 13:06
  LLM
• 13:06
  calls
• 13:07
  across
• 13:07
  18.2M
• 13:08
  input
• 13:08
  tokens
• 13:09
  —
• 13:09
  an
• 13:09
  expensive
• 13:10
  demo,
• 13:11
  but
• 13:11
  the
• 13:11
  first
• 13:11
  time
• 13:11
  any
• 13:12
  system
• 13:12
  has
• 13:12
  passed
• 13:13
  what
• 13:13
  the
• 13:13
  authors
• 13:13
  call
• 13:14
  "the
• 13:14
  scissors
• 13:14
  test":
• 13:15
  producing
• 13:15
  a
• 13:15
  3D
• 13:16
  assembly
• 13:17
  whose
• 13:17
  parts
• 13:17
  actually
• 13:18
  move
• 13:18
  in
• 13:18
  a
• 13:18
  mechanically
• 13:19
  correct
• 13:19
  slicing
• 13:20
  motion.
• 13:23
  A
• 13:23
  PhD
• 13:23
  student's
• 13:24
  post
• 13:24
  about
• 13:24
  a
• 13:24
  colleague's
• 13:25
  relationship
• 13:25
  with
• 13:25
  AI
• 13:26
  agents
• 13:26
  reads
• 13:27
  less
• 13:27
  like
• 13:27
  satire
• 13:28
  and
• 13:28
  more
• 13:28
  like
• 13:28
  a
• 13:28
  clinical
• 13:28
  case
• 13:29
  study.
• 13:30
  The
• 13:30
  colleague
• 13:30
  becomes
• 13:31
  "insanely
• 13:32
  anxious"
• 13:33
  whenever
• 13:33
  no
• 13:33
  agents
• 13:33
  are
• 13:33
  running,
• 13:34
  wakes
• 13:34
  at
• 13:34
  3
• 13:35
  AM
• 13:35
  to
• 13:35
  assign
• 13:35
  new
• 13:35
  tasks,
• 13:36
  and
• 13:36
  when
• 13:37
  the
• 13:37
  university
• 13:37
  ran
• 13:37
  out
• 13:38
  of
• 13:38
  tokens,
• 13:38
  threatened
• 13:39
  colleagues
• 13:39
  he
• 13:39
  perceived
• 13:40
  as
• 13:40
  overusing
• 13:40
  their
• 13:40
  allocation.
• 13:42
  He
• 13:42
  has
• 13:42
  private
• 13:42
  ChatGPT
• 13:43
  and
• 13:43
  Anthropic
• 13:43
  subscriptions
• 13:44
  on
• 13:44
  top
• 13:44
  of
• 13:44
  university-provided
• 13:45
  access
• 13:46
  and
• 13:46
  reportedly
• 13:47
  burned
• 13:47
  1,000
• 13:47
  euros
• 13:48
  of
• 13:48
  departmental
• 13:48
  budget
• 13:49
  in
• 13:49
  a
• 13:49
  single
• 13:49
  pre-NeurIPS
• 13:50
  all-nighter.
• 13:51
  His
• 13:51
  defense:
• 13:52
  "tokens
• 13:52
  are
• 13:52
  not
• 13:53
  a
• 13:53
  substance
• 13:53
  and
• 13:53
  AI
• 13:53
  is
• 13:54
  not
• 13:54
  like
• 13:54
  gambling,"
• 13:55
  so
• 13:55
  addiction
• 13:55
  is
• 13:55
  impossible.
• 13:56
  His
• 13:57
  colleagues
• 13:57
  coined
• 13:58
  "tokaine
• 13:58
  addiction,"
• 13:59
  and
• 13:59
  the
• 13:59
  community
• 13:59
  response
• 14:00
  is
• 14:00
  sobering
• 14:00
  —
• 14:01
  multiple
• 14:01
  commenters
• 14:02
  report
• 14:02
  identical
• 14:03
  patterns
• 14:03
  in
• 14:03
  their
• 14:03
  own
• 14:04
  workplaces,
• 14:05
  including
• 14:05
  one
• 14:05
  startup
• 14:06
  employee
• 14:06
  describing
• 14:06
  company
• 14:07
  parties
• 14:07
  where
• 14:08
  "everyone
• 14:08
  is
• 14:08
  on
• 14:08
  their
• 14:08
  phones
• 14:09
  looking
• 14:09
  at
• 14:09
  their
• 14:09
  agents
• 14:10
  doing
• 14:10
  stupid
• 14:10
  tasks."
• 14:11
  The
• 14:11
  comparison
• 14:12
  to
• 14:12
  compulsive
• 14:12
  gaming
• 14:12
  is
• 14:13
  apt:
• 14:13
  the
• 14:13
  vehicle
• 14:14
  differs,
• 14:15
  but
• 14:15
  the
• 14:15
  pattern
• 14:15
  of
• 14:15
  escalating
• 14:16
  engagement
• 14:17
  despite
• 14:17
  negative
• 14:18
  consequences
• 14:18
  to
• 14:18
  sleep,
• 14:19
  finances,
• 14:20
  and
• 14:20
  relationships
• 14:21
  is
• 14:21
  textbook
• 14:21
  behavioral
• 14:22
  compulsion.
• 14:23
  For
• 14:24
  a
• 14:24
  healthier
• 14:24
  relationship
• 14:25
  with
• 14:25
  local
• 14:25
  models,
• 14:26
  consider
• 14:26
  the
• 14:26
  developer
• 14:27
  who
• 14:27
  wired
• 14:27
  an
• 14:27
  MQ-2
• 14:28
  gas
• 14:28
  sensor
• 14:28
  into
• 14:28
  a
• 14:29
  suitcase
• 14:29
  robot
• 14:29
  running
• 14:30
  a
• 14:30
  local
• 14:30
  LLM.
• 14:31
  When
• 14:31
  smoke
• 14:31
  hits
• 14:32
  the
• 14:32
  sensor,
• 14:32
  it
• 14:32
  modulates
• 14:33
  the
• 14:33
  sampler
• 14:33
  temperature
• 14:34
  (1.0
• 14:35
  to
• 14:35
  ~1.6),
• 14:36
  top
• 14:36
  p
• 14:36
  (0.95
• 14:37
  to
• 14:37
  0.99),
• 14:39
  and
• 14:39
  top
• 14:39
  k
• 14:39
  (64
• 14:40
  to
• 14:40
  120)
• 14:41
  in
• 14:41
  real
• 14:41
  time,
• 14:42
  so
• 14:42
  the
• 14:42
  robot's
• 14:42
  speech
• 14:43
  genuinely
• 14:43
  gets
• 14:43
  loopier
• 14:44
  —
• 14:44
  lower-probability,
• 14:45
  more
• 14:45
  associative
• 14:46
  tokens
• 14:46
  —
• 14:47
  without
• 14:47
  any
• 14:47
  scripted
• 14:48
  behavior.
• 14:49
  A
• 14:49
  per-phase
• 14:50
  persona
• 14:50
  nudge
• 14:51
  makes
• 14:51
  the
• 14:51
  robot
• 14:51
  show
• 14:51
  the
• 14:52
  effects
• 14:52
  without
• 14:52
  announcing
• 14:53
  them,
• 14:53
  while
• 14:53
  the
• 14:53
  physical
• 14:54
  body
• 14:54
  adds
• 14:54
  drooping
• 14:55
  eyes
• 14:55
  and
• 14:55
  a
• 14:55
  smoke-and-plasma
• 14:56
  display
• 14:56
  at
• 14:57
  phase
• 14:57
  10.
• 14:58
  The
• 14:58
  creator's
• 14:58
  honest
• 14:58
  caveat
• 14:59
  —
• 15:00
  "a
• 15:00
  cigarette
• 15:00
  or
• 15:00
  incense
• 15:01
  probably
• 15:01
  trips
• 15:02
  it
• 15:02
  too"
• 15:02
  —
• 15:03
  does
• 15:03
  not
• 15:03
  diminish
• 15:03
  what
• 15:03
  amounts
• 15:04
  to
• 15:04
  one
• 15:04
  of
• 15:04
  the
• 15:04
  most
• 15:04
  creative
• 15:05
  demonstrations
• 15:05
  of
• 15:05
  how
• 15:05
  sampling
• 15:06
  parameters
• 15:06
  shape
• 15:07
  LLM
• 15:07
  behavior
• 15:08
  in
• 15:08
  practice.