Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 16, 2026

Transcript

AGI Dreams Podcast – June 16, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  security
• 0:00
  community
• 0:01
  has
• 0:01
  moved
• 0:01
  from
• 0:01
  complaining
• 0:02
  to
• 0:02
  organizing.
• 0:03
  FreeFable.org
• 0:04
  published
• 0:05
  an
• 0:05
  open
• 0:05
  letter
• 0:05
  to
• 0:05
  Commerce
• 0:06
  Secretary
• 0:06
  Lutnick
• 0:06
  and
• 0:06
  National
• 0:07
  Cyber
• 0:07
  Director
• 0:07
  Cairncross
• 0:08
  demanding
• 0:09
  the
• 0:09
  Fable
• 0:09
  5
• 0:09
  export
• 0:10
  controls
• 0:10
  be
• 0:10
  lifted
• 0:11
  and
• 0:11
  replaced
• 0:11
  with
• 0:12
  "an
• 0:12
  open,
• 0:12
  scientific
• 0:13
  and
• 0:13
  transparent
• 0:14
  process"
• 0:15
  for
• 0:15
  future
• 0:15
  AI
• 0:16
  risk
• 0:16
  assessments.
• 0:16
  The
• 0:17
  signatories
• 0:17
  —
• 0:17
  former
• 0:18
  CSOs,
• 0:19
  CISO-level
• 0:19
  practitioners,
• 0:20
  professors
• 0:21
  of
• 0:21
  law
• 0:21
  and
• 0:21
  cybersecurity,
• 0:22
  and
• 0:22
  AI
• 0:22
  governance
• 0:23
  advisors
• 0:23
  from
• 0:23
  institutions
• 0:24
  including
• 0:25
  Harvard's
• 0:25
  Berkman
• 0:25
  Klein
• 0:26
  Center,
• 0:26
  the
• 0:26
  Center
• 0:27
  for
• 0:27
  Democracy
• 0:27
  &
• 0:27
  Technology,
• 0:28
  and
• 0:28
  Cleveland
• 0:29
  State's
• 0:29
  cybersecurity
• 0:29
  program
• 0:30
  —
• 0:30
  are
• 0:30
  not
• 0:30
  fringe
• 0:31
  voices.
• 0:31
  They
• 0:31
  represent
• 0:32
  the
• 0:32
  professional
• 0:32
  class
• 0:33
  that
• 0:33
  builds
• 0:33
  and
• 0:33
  breaks
• 0:33
  defenses
• 0:34
  for
• 0:34
  a
• 0:34
  living,
• 0:34
  and
• 0:34
  their
• 0:35
  core
• 0:35
  argument
• 0:35
  is
• 0:35
  blunt:
• 0:36
  the
• 0:36
  research
• 0:36
  that
• 0:36
  triggered
• 0:37
  the
• 0:37
  federal
• 0:37
  action
• 0:37
  focused
• 0:38
  on
• 0:38
  "determining
• 0:38
  whether
• 0:39
  a
• 0:39
  human-prompted
• 0:39
  section
• 0:40
  of
• 0:40
  code
• 0:40
  was
• 0:40
  insecure,"
• 0:41
  a
• 0:41
  capability
• 0:41
  that
• 0:41
  is
• 0:42
  a
• 0:42
  prerequisite
• 0:42
  for
• 0:42
  writing
• 0:42
  secure
• 0:43
  code,
• 0:43
  not
• 0:43
  an
• 0:43
  offensive
• 0:44
  weapon.
• 0:44
  The
• 0:44
  letter
• 0:44
  makes
• 0:45
  three
• 0:45
  factual
• 0:45
  claims
• 0:46
  worth
• 0:46
  stress-testing.
• 0:46
  First,
• 0:47
  that
• 0:47
  the
• 0:47
  same
• 0:47
  capabilities
• 0:48
  "can
• 0:48
  be
• 0:48
  replicated
• 0:48
  on
• 0:49
  GPT-5.5,
• 0:50
  Opus,
• 0:50
  Sonnet
• 0:50
  and
• 0:50
  even
• 0:51
  Chinese
• 0:51
  models
• 0:51
  like
• 0:51
  Kimi
• 0:52
  2.7"
• 0:52
  —
• 0:52
  meaning
• 0:53
  the
• 0:53
  export
• 0:53
  control
• 0:53
  punishes
• 0:54
  one
• 0:54
  vendor
• 0:54
  for
• 0:54
  a
• 0:54
  capability
• 0:55
  class
• 0:55
  that
• 0:55
  is
• 0:55
  now
• 0:55
  table
• 0:56
  stakes.
• 0:56
  Second,
• 0:56
  that
• 0:56
  Chinese
• 0:57
  open-weight
• 0:57
  models
• 0:57
  are
• 0:58
  "only
• 0:58
  months
• 0:58
  behind
• 0:58
  the
• 0:59
  best
• 0:59
  American
• 0:59
  models"
• 1:00
  and
• 1:00
  that
• 1:00
  the
• 1:00
  PRC
• 1:00
  government
• 1:01
  likely
• 1:01
  has
• 1:01
  access
• 1:01
  to
• 1:01
  unpublished
• 1:02
  capabilities
• 1:02
  beyond
• 1:03
  those.
• 1:03
  Third,
• 1:04
  that
• 1:04
  Anthropic
• 1:05
  is
• 1:05
  already
• 1:05
  addressing
• 1:05
  the
• 1:05
  research
• 1:06
  findings,
• 1:06
  making
• 1:07
  the
• 1:07
  ban
• 1:07
  a
• 1:07
  response
• 1:07
  to
• 1:08
  a
• 1:08
  problem
• 1:08
  the
• 1:08
  vendor
• 1:08
  was
• 1:08
  already
• 1:09
  remediating.
• 1:10
  The
• 1:10
  rhetorical
• 1:10
  question
• 1:11
  underlying
• 1:11
  all
• 1:11
  of
• 1:11
  it:
• 1:12
  if
• 1:12
  defenders
• 1:12
  lose
• 1:13
  access
• 1:13
  to
• 1:13
  the
• 1:13
  best
• 1:13
  tools
• 1:14
  while
• 1:14
  adversaries
• 1:14
  retain
• 1:15
  equivalent
• 1:15
  capabilities,
• 1:16
  who
• 1:16
  actually
• 1:16
  benefits?
• 1:17
  The
• 1:17
  backstory
• 1:18
  adds
• 1:18
  context
• 1:18
  to
• 1:19
  the
• 1:19
  signatories'
• 1:19
  frustration.
• 1:20
  The
• 1:20
  federal
• 1:21
  response
• 1:21
  was
• 1:21
  reportedly
• 1:22
  triggered
• 1:22
  not
• 1:22
  by
• 1:23
  a
• 1:23
  sophisticated
• 1:23
  jailbreak
• 1:24
  but
• 1:24
  by
• 1:24
  a
• 1:24
  researcher
• 1:25
  prompting
• 1:25
  Fable
• 1:26
  5
• 1:26
  with
• 1:26
  a
• 1:26
  routine
• 1:27
  "fix
• 1:27
  this
• 1:28
  code"
• 1:28
  request
• 1:29
  that
• 1:29
  surfaced
• 1:29
  a
• 1:29
  vulnerability
• 1:29
  —
• 1:30
  exactly
• 1:31
  the
• 1:31
  kind
• 1:31
  of
• 1:31
  workflow
• 1:32
  every
• 1:32
  security
• 1:32
  team
• 1:32
  runs
• 1:33
  daily.
• 1:34
  The
• 1:34
  gap
• 1:34
  between
• 1:34
  what
• 1:34
  triggered
• 1:35
  the
• 1:35
  panic
• 1:35
  and
• 1:35
  the
• 1:35
  severity
• 1:36
  of
• 1:36
  the
• 1:36
  government's
• 1:36
  response
• 1:36
  —
• 1:38
  disabling
• 1:38
  the
• 1:38
  model
• 1:38
  for
• 1:39
  all
• 1:39
  customers
• 1:39
  within
• 1:39
  90
• 1:40
  minutes
• 1:40
  of
• 1:40
  a
• 1:40
  White
• 1:40
  House
• 1:40
  ultimatum
• 1:41
  —
• 1:42
  is
• 1:42
  the
• 1:42
  detail
• 1:42
  that
• 1:42
  keeps
• 1:43
  this
• 1:43
  story
• 1:43
  from
• 1:43
  fading.
• 1:44
  Export
• 1:45
  controls
• 1:45
  designed
• 1:45
  for
• 1:45
  dual-use
• 1:46
  physical
• 1:46
  goods
• 1:46
  are
• 1:46
  being
• 1:47
  applied
• 1:47
  to
• 1:47
  models
• 1:48
  whose
• 1:48
  most
• 1:48
  dangerous
• 1:48
  capability
• 1:49
  is
• 1:49
  also
• 1:50
  their
• 1:50
  most
• 1:50
  useful
• 1:50
  one,
• 1:51
  and
• 1:51
  the
• 1:51
  security
• 1:52
  professionals
• 1:52
  who
• 1:52
  depend
• 1:53
  on
• 1:53
  that
• 1:53
  capability
• 1:53
  are
• 1:54
  now
• 1:54
  formally
• 1:54
  objecting.
• 1:58
  The
• 1:58
  Miasma
• 1:58
  worm
• 1:59
  hit
• 1:59
  Microsoft's
• 1:59
  own
• 1:59
  Azure
• 2:00
  repositories
• 2:01
  this
• 2:01
  week,
• 2:02
  triggering
• 2:02
  GitHub's
• 2:03
  automated
• 2:03
  security
• 2:04
  system
• 2:04
  to
• 2:04
  flag
• 2:04
  and
• 2:04
  disable
• 2:05
  73
• 2:05
  Microsoft-related
• 2:06
  package
• 2:06
  repositories
• 2:07
  in
• 2:07
  just
• 2:07
  over
• 2:07
  a
• 2:07
  minute
• 2:08
  —
• 2:08
  more
• 2:08
  than
• 2:09
  40
• 2:09
  tied
• 2:09
  directly
• 2:09
  to
• 2:09
  Azure.
• 2:10
  The
• 2:11
  infection
• 2:11
  vector
• 2:11
  traces
• 2:12
  back
• 2:12
  to
• 2:12
  the
• 2:12
  Durabletask
• 2:13
  package,
• 2:13
  previously
• 2:14
  compromised
• 2:14
  in
• 2:14
  May,
• 2:15
  which
• 2:15
  strongly
• 2:16
  suggests
• 2:16
  that
• 2:16
  credentials
• 2:16
  stolen
• 2:17
  in
• 2:17
  the
• 2:17
  original
• 2:17
  attack
• 2:17
  were
• 2:18
  never
• 2:18
  properly
• 2:18
  revoked.
• 2:19
  Any
• 2:19
  build
• 2:20
  process
• 2:20
  depending
• 2:20
  on
• 2:20
  those
• 2:21
  packages
• 2:21
  broke
• 2:21
  immediately.
• 2:22
  For
• 2:23
  a
• 2:23
  package
• 2:23
  that
• 2:23
  logged
• 2:23
  400,000
• 2:24
  monthly
• 2:24
  downloads,
• 2:25
  the
• 2:25
  blast
• 2:26
  radius
• 2:26
  is
• 2:26
  substantial.
• 2:28
  The
• 2:28
  same
• 2:28
  weekly
• 2:28
  security
• 2:29
  roundup
• 2:29
  documents
• 2:29
  a
• 2:30
  cascade
• 2:30
  of
• 2:30
  adjacent
• 2:31
  threats:
• 2:31
  OpenSSL's
• 2:32
  high-severity
• 2:32
  use-after-free
• 2:33
  bug
• 2:33
  in
• 2:33
  PKCS7
• 2:34
  handling,
• 2:35
  NightmareEclipse
• 2:35
  returning
• 2:36
  with
• 2:36
  the
• 2:36
  RoguePlanet
• 2:36
  exploit
• 2:37
  (Windows
• 2:37
  Defender
• 2:37
  race
• 2:38
  condition
• 2:38
  for
• 2:38
  a
• 2:38
  system-level
• 2:39
  shell)
• 2:39
  and
• 2:39
  GreatXML
• 2:40
  (BitLocker
• 2:40
  bypass
• 2:41
  via
• 2:41
  offline
• 2:41
  scans),
• 2:42
  a
• 2:42
  FreeBSD
• 2:43
  page-cache
• 2:43
  bug
• 2:44
  following
• 2:44
  the
• 2:44
  Linux
• 2:44
  CopyFail/DirtyPipe
• 2:46
  pattern,
• 2:46
  and
• 2:46
  —
• 2:47
  most
• 2:47
  consequentially
• 2:48
  for
• 2:48
  the
• 2:48
  ecosystem
• 2:48
  —
• 2:49
  NPM
• 2:49
  12's
• 2:50
  decision
• 2:50
  to
• 2:50
  block
• 2:50
  automatic
• 2:51
  execution
• 2:51
  of
• 2:51
  install
• 2:52
  scripts.
• 2:53
  That
• 2:53
  last
• 2:53
  change
• 2:54
  addresses
• 2:54
  the
• 2:54
  single
• 2:55
  largest
• 2:55
  propagation
• 2:55
  mechanism
• 2:56
  for
• 2:56
  supply
• 2:56
  chain
• 2:56
  worms:
• 2:57
  pre-install,
• 2:58
  install,
• 2:59
  and
• 2:59
  post-install
• 3:00
  hooks
• 3:00
  that
• 3:00
  run
• 3:00
  arbitrary
• 3:01
  commands
• 3:01
  as
• 3:01
  the
• 3:01
  build
• 3:02
  user.
• 3:02
  It
• 3:03
  will
• 3:03
  not
• 3:03
  stop
• 3:03
  all
• 3:03
  malicious
• 3:03
  packages,
• 3:04
  but
• 3:04
  it
• 3:04
  removes
• 3:04
  the
• 3:05
  highway
• 3:05
  the
• 3:05
  worms
• 3:05
  were
• 3:05
  using.
• 3:06
  Perhaps
• 3:06
  the
• 3:07
  most
• 3:07
  darkly
• 3:07
  amusing
• 3:07
  finding:
• 3:08
  researchers
• 3:09
  reverse-engineering
• 3:09
  the
• 3:10
  Mini
• 3:10
  Shai-Hulud
• 3:10
  worm
• 3:11
  discovered
• 3:11
  it
• 3:11
  uses
• 3:12
  AI
• 3:12
  prompt
• 3:12
  injection
• 3:13
  to
• 3:13
  hide
• 3:13
  from
• 3:13
  automated
• 3:14
  analysis.
• 3:15
  The
• 3:15
  payload
• 3:15
  file
• 3:15
  contains
• 3:16
  comment
• 3:16
  blocks
• 3:16
  referencing
• 3:17
  biological
• 3:17
  and
• 3:17
  nuclear
• 3:18
  weapons
• 3:18
  —
• 3:19
  topics
• 3:19
  that
• 3:19
  cause
• 3:20
  many
• 3:20
  AI
• 3:20
  models
• 3:20
  to
• 3:21
  refuse
• 3:21
  further
• 3:21
  processing.
• 3:23
  The
• 3:23
  worm
• 3:23
  is
• 3:23
  exploiting
• 3:23
  the
• 3:24
  same
• 3:24
  overzealous
• 3:24
  guardrails
• 3:25
  that
• 3:25
  security
• 3:25
  professionals
• 3:26
  have
• 3:26
  been
• 3:26
  criticizing
• 3:27
  in
• 3:27
  a
• 3:27
  completely
• 3:27
  different
• 3:28
  context.
• 3:29
  When
• 3:29
  malware
• 3:29
  developers
• 3:30
  weaponize
• 3:30
  your
• 3:30
  safety
• 3:31
  features
• 3:31
  as
• 3:31
  evasion
• 3:32
  techniques,
• 3:32
  the
• 3:32
  irony
• 3:33
  writes
• 3:33
  itself.
• 3:34
  Against
• 3:34
  this
• 3:35
  backdrop,
• 3:35
  Chainguard
• 3:36
  launched
• 3:36
  Athena,
• 3:37
  an
• 3:37
  industry
• 3:37
  coalition
• 3:38
  for
• 3:38
  coordinated
• 3:38
  defense
• 3:39
  of
• 3:39
  open-source
• 3:39
  software.
• 3:41
  The
• 3:41
  premise
• 3:41
  is
• 3:41
  simple:
• 3:42
  AI
• 3:42
  models
• 3:42
  can
• 3:42
  now
• 3:43
  find
• 3:43
  chained
• 3:43
  zero-day
• 3:44
  vulnerabilities
• 3:44
  at
• 3:44
  machine
• 3:45
  speed,
• 3:46
  the
• 3:46
  window
• 3:46
  between
• 3:47
  discovery
• 3:47
  and
• 3:47
  exploitation
• 3:48
  has
• 3:48
  collapsed
• 3:48
  from
• 3:49
  years
• 3:49
  to
• 3:49
  hours,
• 3:49
  and
• 3:50
  fragmented
• 3:50
  patch-it-yourself
• 3:51
  responses
• 3:51
  cannot
• 3:52
  keep
• 3:52
  up.
• 3:53
  Athena
• 3:53
  pools
• 3:54
  vetted
• 3:54
  pre-disclosure
• 3:55
  findings
• 3:55
  from
• 3:55
  members,
• 3:56
  rebuilds
• 3:56
  affected
• 3:57
  projects
• 3:57
  as
• 3:58
  hardened
• 3:58
  versions
• 3:58
  available
• 3:59
  before
• 3:59
  public
• 4:00
  disclosure,
• 4:00
  and
• 4:00
  coordinates
• 4:01
  platform-level
• 4:02
  mitigations
• 4:02
  —
• 4:03
  network
• 4:04
  rules,
• 4:04
  detection
• 4:05
  signatures,
• 4:05
  traffic
• 4:06
  blocks
• 4:06
  —
• 4:07
  so
• 4:07
  that
• 4:07
  organizations
• 4:08
  without
• 4:08
  security
• 4:08
  teams
• 4:09
  still
• 4:09
  get
• 4:09
  protection.
• 4:10
  More
• 4:10
  than
• 4:10
  two
• 4:11
  dozen
• 4:11
  organizations
• 4:12
  have
• 4:12
  joined,
• 4:12
  with
• 4:12
  the
• 4:13
  first
• 4:13
  wave
• 4:13
  of
• 4:13
  coordinated
• 4:14
  disclosures
• 4:14
  scheduled
• 4:15
  within
• 4:15
  a
• 4:15
  month.
• 4:16
  The
• 4:16
  question
• 4:17
  is
• 4:17
  whether
• 4:17
  the
• 4:17
  coalition
• 4:18
  model
• 4:18
  can
• 4:18
  scale
• 4:18
  faster
• 4:19
  than
• 4:19
  the
• 4:19
  worms.
• 4:20
  The
• 4:20
  Miasma
• 4:21
  family
• 4:21
  has
• 4:21
  been
• 4:21
  propagating
• 4:22
  across
• 4:22
  NPM,
• 4:23
  PyPI,
• 4:24
  VS
• 4:24
  Code
• 4:25
  extensions,
• 4:26
  and
• 4:26
  GitHub
• 4:26
  Actions
• 4:27
  simultaneously.
• 4:28
  Athena
• 4:29
  addresses
• 4:29
  the
• 4:29
  vulnerability
• 4:30
  side,
• 4:30
  but
• 4:31
  the
• 4:31
  credential-theft-to-worm-propagation
• 4:32
  cycle
• 4:33
  requires
• 4:33
  a
• 4:33
  different
• 4:34
  kind
• 4:34
  of
• 4:34
  intervention
• 4:34
  entirely.
• 4:36
  On
• 4:36
  the
• 4:36
  governance
• 4:37
  side,
• 4:37
  researchers
• 4:38
  from
• 4:38
  Old
• 4:38
  Dominion
• 4:38
  University,
• 4:39
  Deloitte,
• 4:40
  Accenture,
• 4:41
  and
• 4:41
  the
• 4:41
  University
• 4:42
  of
• 4:42
  Colombo
• 4:42
  published
• 4:43
  "AI
• 4:43
  Trust
• 4:44
  OS,"
• 4:45
  proposing
• 4:45
  a
• 4:45
  continuous,
• 4:46
  telemetry-driven
• 4:47
  governance
• 4:47
  framework
• 4:48
  that
• 4:48
  replaces
• 4:49
  periodic
• 4:49
  manual
• 4:50
  audits
• 4:50
  with
• 4:50
  always-on
• 4:51
  zero-trust
• 4:52
  compliance
• 4:52
  monitoring.
• 4:53
  The
• 4:53
  core
• 4:54
  contribution
• 4:54
  is
• 4:54
  a
• 4:55
  Shadow
• 4:55
  AI
• 4:55
  discovery
• 4:56
  mechanism:
• 4:57
  an
• 4:57
  agent
• 4:57
  that
• 4:57
  continuously
• 4:58
  scans
• 4:59
  LangSmith
• 4:59
  and
• 4:59
  Datadog
• 5:00
  telemetry
• 5:00
  streams
• 5:01
  to
• 5:01
  detect
• 5:02
  AI
• 5:02
  systems
• 5:02
  that
• 5:02
  engineering
• 5:03
  teams
• 5:03
  deployed
• 5:04
  without
• 5:04
  formal
• 5:04
  security
• 5:05
  review.
• 5:06
  In
• 5:06
  their
• 5:06
  evaluation,
• 5:07
  it
• 5:07
  identified
• 5:08
  an
• 5:08
  undeclared
• 5:09
  fine-tuned
• 5:09
  production
• 5:10
  model
• 5:10
  with
• 5:10
  no
• 5:10
  documented
• 5:11
  training
• 5:11
  provenance,
• 5:12
  risk
• 5:12
  classification,
• 5:13
  or
• 5:13
  accountable
• 5:14
  owner
• 5:14
  —
• 5:15
  exactly
• 5:15
  the
• 5:15
  governance
• 5:16
  gap
• 5:16
  that
• 5:16
  regulators
• 5:17
  under
• 5:17
  ISO
• 5:17
  42001
• 5:20
  and
• 5:20
  the
• 5:20
  EU
• 5:20
  AI
• 5:20
  Act
• 5:21
  are
• 5:21
  trying
• 5:21
  to
• 5:21
  close.
• 5:22
  The
• 5:22
  paper's
• 5:22
  strongest
• 5:23
  insight
• 5:23
  is
• 5:24
  reframing
• 5:24
  compliance
• 5:24
  from
• 5:25
  "filling
• 5:25
  out
• 5:25
  forms"
• 5:26
  to
• 5:26
  "machines
• 5:26
  collecting
• 5:26
  evidence."
• 5:28
  Whether
• 5:28
  enterprises
• 5:29
  actually
• 5:29
  adopt
• 5:30
  telemetry-first
• 5:30
  governance
• 5:31
  before
• 5:31
  their
• 5:31
  next
• 5:31
  audit
• 5:32
  is
• 5:32
  a
• 5:32
  different
• 5:32
  matter.
• 5:35
  AMD
• 5:35
  CEO
• 5:36
  Lisa
• 5:36
  Su
• 5:36
  walked
• 5:37
  onto
• 5:37
  a
• 5:37
  stage,
• 5:37
  held
• 5:37
  a
• 5:37
  box
• 5:38
  the
• 5:38
  size
• 5:38
  of
• 5:38
  a
• 5:38
  thick
• 5:39
  paperback
• 5:39
  in
• 5:39
  one
• 5:39
  hand,
• 5:39
  and
• 5:40
  ran
• 5:40
  a
• 5:40
  235-billion-parameter
• 5:41
  model
• 5:42
  live.
• 5:42
  No
• 5:42
  data
• 5:43
  center,
• 5:43
  no
• 5:43
  cloud,
• 5:44
  no
• 5:44
  rented
• 5:44
  GPU.
• 5:45
  The
• 5:45
  Ryzen
• 5:45
  AI
• 5:45
  Max+
• 5:46
  395,
• 5:47
  the
• 5:47
  chip
• 5:47
  inside
• 5:48
  AMD's
• 5:48
  new
• 5:48
  compact
• 5:48
  AI
• 5:49
  development
• 5:49
  system,
• 5:50
  is
• 5:50
  the
• 5:50
  first
• 5:50
  x86
• 5:51
  silicon
• 5:51
  where
• 5:51
  CPU
• 5:52
  and
• 5:52
  GPU
• 5:52
  share
• 5:52
  the
• 5:52
  same
• 5:53
  128GB
• 5:54
  of
• 5:54
  unified
• 5:54
  memory.
• 5:55
  Under
• 5:55
  Linux,
• 5:56
  the
• 5:56
  GPU
• 5:56
  gets
• 5:56
  access
• 5:57
  to
• 5:57
  110GB
• 5:59
  of
• 5:59
  that
• 5:59
  pool
• 6:00
  —
• 6:00
  more
• 6:00
  than
• 6:00
  three
• 6:01
  times
• 6:01
  what
• 6:01
  an
• 6:01
  RTX
• 6:02
  5090
• 6:03
  offers,
• 6:03
  in
• 6:03
  a
• 6:03
  chassis
• 6:04
  you
• 6:04
  can
• 6:04
  carry
• 6:04
  with
• 6:05
  one
• 6:05
  hand.
• 6:05
  The
• 6:06
  benchmark
• 6:06
  that
• 6:06
  drew
• 6:06
  attention:
• 6:07
  this
• 6:07
  chip
• 6:07
  beat
• 6:08
  an
• 6:08
  RTX
• 6:08
  5080
• 6:09
  by
• 6:09
  more
• 6:09
  than
• 6:09
  3x
• 6:10
  on
• 6:10
  DeepSeek
• 6:10
  R1
• 6:11
  inference.
• 6:12
  The
• 6:12
  economic
• 6:12
  argument
• 6:13
  is
• 6:13
  straightforward.
• 6:14
  A
• 6:14
  heavy
• 6:14
  AI
• 6:15
  user
• 6:15
  paying
• 6:15
  for
• 6:15
  Claude
• 6:15
  Code
• 6:16
  Max,
• 6:16
  ChatGPT
• 6:17
  Pro,
• 6:17
  Cursor,
• 6:18
  and
• 6:18
  Gemini
• 6:18
  subscriptions
• 6:19
  spends
• 6:19
  roughly
• 6:20
  $5,280
• 6:21
  annually.
• 6:22
  A
• 6:22
  $1,499
• 6:23
  box
• 6:23
  —
• 6:24
  though
• 6:25
  the
• 6:25
  AMD
• 6:25
  website
• 6:25
  lists
• 6:26
  the
• 6:26
  full
• 6:26
  configuration
• 6:27
  at
• 6:27
  $3,999
• 6:28
  —
• 6:29
  running
• 6:29
  Ollama
• 6:29
  with
• 6:30
  a
• 6:30
  local
• 6:30
  model
• 6:30
  eliminates
• 6:31
  per-request
• 6:31
  costs,
• 6:32
  token
• 6:32
  meters,
• 6:32
  and
• 6:32
  data
• 6:33
  leaving
• 6:33
  the
• 6:33
  machine.
• 6:34
  The
• 6:34
  real
• 6:34
  advantage
• 6:34
  is
• 6:35
  not
• 6:35
  cost
• 6:35
  savings
• 6:35
  but
• 6:35
  memory
• 6:36
  capacity:
• 6:37
  running
• 6:37
  235B-parameter
• 6:39
  models
• 6:39
  that
• 6:40
  simply
• 6:40
  cannot
• 6:40
  fit
• 6:40
  on
• 6:41
  consumer
• 6:41
  GPUs
• 6:42
  at
• 6:42
  any
• 6:42
  price.
• 6:43
  NVIDIA's
• 6:43
  moat
• 6:43
  has
• 6:44
  always
• 6:44
  been
• 6:44
  as
• 6:44
  much
• 6:44
  about
• 6:45
  CUDA
• 6:45
  and
• 6:45
  the
• 6:45
  software
• 6:46
  ecosystem
• 6:46
  as
• 6:46
  raw
• 6:46
  silicon,
• 6:47
  and
• 6:47
  AMD's
• 6:48
  ROCm
• 6:48
  stack
• 6:49
  still
• 6:49
  has
• 6:49
  rough
• 6:50
  edges
• 6:50
  —
• 6:51
  but
• 6:51
  for
• 6:51
  inference
• 6:51
  workloads
• 6:52
  where
• 6:52
  memory
• 6:52
  is
• 6:52
  the
• 6:52
  bottleneck,
• 6:53
  unified
• 6:54
  memory
• 6:54
  changes
• 6:54
  the
• 6:54
  math.
• 6:55
  Meanwhile,
• 6:56
  Apple
• 6:56
  Silicon
• 6:57
  is
• 6:57
  getting
• 6:57
  its
• 6:57
  own
• 6:57
  inference
• 6:58
  breakthrough.
• 6:59
  A
• 6:59
  developer
• 6:59
  ported
• 7:00
  EXL3
• 7:00
  quantization
• 7:01
  to
• 7:01
  run
• 7:01
  natively
• 7:02
  on
• 7:02
  Apple
• 7:02
  hardware,
• 7:03
  achieving
• 7:04
  2,700
• 7:05
  tokens/second
• 7:06
  prefill
• 7:06
  and
• 7:07
  68.5
• 7:08
  tokens/second
• 7:09
  decode
• 7:09
  on
• 7:09
  an
• 7:09
  M5
• 7:10
  Max.
• 7:11
  EXL3
• 7:12
  —
• 7:13
  ExLlamaV3's
• 7:14
  quantization
• 7:14
  format
• 7:15
  —
• 7:15
  was
• 7:15
  previously
• 7:16
  CUDA-only,
• 7:17
  locking
• 7:17
  out
• 7:17
  the
• 7:17
  growing
• 7:17
  population
• 7:18
  of
• 7:18
  developers
• 7:19
  running
• 7:19
  local
• 7:19
  models
• 7:19
  on
• 7:20
  Mac
• 7:20
  hardware.
• 7:21
  The
• 7:21
  port
• 7:21
  brings
• 7:21
  Metal/MLX-class
• 7:23
  performance
• 7:23
  to
• 7:23
  a
• 7:23
  quantization
• 7:24
  format
• 7:25
  optimized
• 7:25
  for
• 7:25
  VRAM
• 7:26
  efficiency,
• 7:27
  narrowing
• 7:27
  the
• 7:27
  gap
• 7:27
  between
• 7:28
  NVIDIA
• 7:28
  and
• 7:28
  Apple
• 7:29
  for
• 7:29
  local
• 7:29
  inference.
• 7:30
  The
• 7:30
  infrastructure
• 7:31
  to
• 7:31
  actually
• 7:31
  use
• 7:32
  all
• 7:32
  this
• 7:32
  hardware
• 7:33
  is
• 7:33
  maturing
• 7:33
  too.
• 7:35
  One
• 7:35
  developer
• 7:35
  documented
• 7:36
  a
• 7:36
  homelab
• 7:37
  AI
• 7:37
  platform
• 7:37
  built
• 7:38
  around
• 7:38
  OpenCode
• 7:38
  Web
• 7:39
  UI
• 7:39
  with
• 7:39
  Git
• 7:39
  access,
• 7:40
  where
• 7:40
  the
• 7:41
  AI
• 7:41
  pushes
• 7:41
  changes
• 7:42
  to
• 7:42
  feature
• 7:42
  branches,
• 7:43
  the
• 7:43
  human
• 7:43
  reviews
• 7:44
  PRs,
• 7:44
  and
• 7:44
  GitOps
• 7:45
  handles
• 7:45
  deployment.
• 7:46
  The
• 7:46
  key
• 7:46
  design
• 7:47
  constraint:
• 7:48
  the
• 7:48
  AI
• 7:48
  has
• 7:48
  internet
• 7:49
  access
• 7:49
  and
• 7:49
  Git
• 7:49
  access
• 7:50
  but
• 7:50
  cannot
• 7:51
  reach
• 7:51
  the
• 7:51
  actual
• 7:51
  services
• 7:52
  it
• 7:52
  is
• 7:52
  changing.
• 7:53
  Unreviewed
• 7:54
  code
• 7:54
  never
• 7:54
  gets
• 7:55
  deployed.
• 7:56
  The
• 7:56
  pattern
• 7:56
  —
• 7:57
  AI
• 7:57
  writes,
• 7:58
  human
• 7:58
  reviews,
• 7:59
  automation
• 7:59
  deploys
• 7:59
  —
• 8:00
  is
• 8:01
  becoming
• 8:01
  the
• 8:01
  standard
• 8:01
  for
• 8:02
  anyone
• 8:02
  running
• 8:02
  AI-assisted
• 8:03
  operations
• 8:04
  on
• 8:04
  their
• 8:04
  own
• 8:04
  infrastructure.
• 8:07
  Inside
• 8:07
  Meta's
• 8:08
  new
• 8:08
  Applied
• 8:08
  AI
• 8:08
  unit,
• 8:09
  an
• 8:09
  employee
• 8:09
  interrupted
• 8:10
  a
• 8:10
  company-wide
• 8:11
  livestream
• 8:11
  with
• 8:11
  an
• 8:11
  expletive-filled
• 8:12
  outburst
• 8:13
  calling
• 8:13
  it
• 8:14
  "being
• 8:14
  the
• 8:14
  company's
• 8:14
  bitch,"
• 8:15
  according
• 8:15
  to
• 8:15
  a
• 8:15
  recording
• 8:15
  heard
• 8:16
  by
• 8:16
  WIRED.
• 8:16
  The
• 8:17
  presenter
• 8:17
  covered
• 8:17
  their
• 8:17
  face.
• 8:18
  The
• 8:18
  incident
• 8:18
  reflects
• 8:19
  deeper
• 8:19
  dysfunction:
• 8:20
  Applied
• 8:20
  AI,
• 8:21
  a
• 8:21
  6,500-person
• 8:21
  unit
• 8:22
  formed
• 8:22
  in
• 8:22
  March
• 8:22
  to
• 8:22
  support
• 8:22
  AI
• 8:23
  researchers,
• 8:24
  has
• 8:24
  become
• 8:24
  a
• 8:24
  dumping
• 8:24
  ground
• 8:25
  where
• 8:25
  engineers
• 8:25
  describe
• 8:25
  themselves
• 8:26
  as
• 8:26
  "draftees"
• 8:27
  doing
• 8:27
  work
• 8:28
  they
• 8:28
  call
• 8:28
  "mechanical
• 8:29
  and
• 8:29
  not
• 8:29
  creative"
• 8:29
  —
• 8:30
  generating
• 8:31
  coding
• 8:31
  puzzles
• 8:31
  to
• 8:31
  test
• 8:31
  model
• 8:32
  performance.
• 8:33
  "It's
• 8:33
  literally
• 8:33
  the
• 8:33
  gulag,"
• 8:34
  one
• 8:34
  employee
• 8:35
  told
• 8:35
  WIRED.
• 8:36
  "Most
• 8:36
  people
• 8:36
  find
• 8:36
  the
• 8:36
  work
• 8:37
  soul-crushing,"
• 8:37
  said
• 8:37
  another.
• 8:39
  The
• 8:39
  turmoil
• 8:39
  extends
• 8:39
  beyond
• 8:40
  Applied
• 8:40
  AI.
• 8:41
  Over
• 8:41
  1,600
• 8:43
  Meta
• 8:43
  employees
• 8:43
  signed
• 8:43
  a
• 8:43
  petition
• 8:44
  against
• 8:44
  a
• 8:44
  new
• 8:44
  program
• 8:45
  to
• 8:45
  monitor
• 8:45
  their
• 8:46
  clicks
• 8:46
  and
• 8:46
  keystrokes
• 8:47
  for
• 8:47
  AI
• 8:47
  training
• 8:47
  data.
• 8:48
  Instagram's
• 8:49
  chief
• 8:49
  product
• 8:49
  officer
• 8:50
  Chris
• 8:50
  Cox
• 8:51
  acknowledged
• 8:51
  the
• 8:51
  "brutal"
• 8:52
  environment
• 8:52
  on
• 8:52
  a
• 8:52
  company-wide
• 8:53
  call,
• 8:54
  comparing
• 8:54
  the
• 8:54
  past
• 8:54
  few
• 8:54
  months
• 8:55
  to
• 8:55
  "running
• 8:55
  a
• 8:55
  marathon
• 8:56
  in
• 8:56
  the
• 8:56
  middle
• 8:56
  of
• 8:56
  a
• 8:56
  hailstorm."
• 8:58
  Zuckerberg's
• 8:58
  internal
• 8:58
  memo
• 8:59
  acknowledged
• 8:59
  mistakes,
• 9:00
  promised
• 9:00
  no
• 9:00
  more
• 9:01
  mass
• 9:01
  layoffs
• 9:01
  this
• 9:01
  year,
• 9:02
  and
• 9:02
  suggested
• 9:03
  Applied
• 9:03
  AI
• 9:03
  was
• 9:03
  "a
• 9:04
  waypoint,
• 9:04
  not
• 9:05
  a
• 9:05
  destination"
• 9:05
  —
• 9:06
  cold
• 9:06
  comfort
• 9:07
  for
• 9:07
  engineers
• 9:07
  who
• 9:07
  were
• 9:07
  told
• 9:07
  to
• 9:08
  join
• 9:08
  or
• 9:08
  leave.
• 9:09
  The
• 9:09
  50-to-1
• 9:10
  manager
• 9:10
  ratio
• 9:10
  in
• 9:11
  some
• 9:11
  parts
• 9:11
  of
• 9:11
  the
• 9:11
  unit
• 9:11
  tells
• 9:11
  you
• 9:11
  everything
• 9:12
  about
• 9:12
  how
• 9:12
  much
• 9:12
  individual
• 9:13
  attention
• 9:13
  anyone
• 9:13
  is
• 9:13
  getting.
• 9:14
  Meta
• 9:14
  is
• 9:15
  the
• 9:15
  most
• 9:15
  visible
• 9:15
  example
• 9:16
  of
• 9:16
  a
• 9:16
  broader
• 9:16
  pattern.
• 9:17
  A
• 9:17
  new
• 9:17
  Glean
• 9:18
  report,
• 9:18
  produced
• 9:19
  with
• 9:19
  researchers
• 9:19
  from
• 9:19
  Notre
• 9:20
  Dame,
• 9:20
  Stanford,
• 9:21
  and
• 9:21
  UC
• 9:21
  Berkeley,
• 9:22
  found
• 9:22
  that
• 9:22
  white-collar
• 9:23
  workers
• 9:23
  spend
• 9:23
  an
• 9:23
  average
• 9:23
  of
• 9:24
  6.4
• 9:24
  hours
• 9:24
  per
• 9:25
  week
• 9:25
  "botsitting"
• 9:25
  AI
• 9:25
  —
• 9:26
  feeding
• 9:27
  it
• 9:27
  context,
• 9:27
  checking
• 9:28
  outputs,
• 9:28
  debugging
• 9:29
  mistakes,
• 9:29
  and
• 9:29
  cleaning
• 9:30
  up
• 9:30
  errors.
• 9:31
  That
• 9:31
  is
• 9:31
  nearly
• 9:31
  a
• 9:31
  full
• 9:32
  working
• 9:32
  day
• 9:32
  every
• 9:32
  week
• 9:33
  spent
• 9:33
  supervising
• 9:33
  tools
• 9:34
  that
• 9:34
  were
• 9:34
  supposed
• 9:34
  to
• 9:34
  save
• 9:34
  time.
• 9:35
  The
• 9:35
  data
• 9:36
  point
• 9:36
  that
• 9:36
  should
• 9:36
  alarm
• 9:37
  every
• 9:37
  executive:
• 9:38
  workers
• 9:38
  who
• 9:38
  spend
• 9:39
  a
• 9:39
  disproportionate
• 9:39
  share
• 9:39
  of
• 9:40
  their
• 9:40
  AI
• 9:40
  time
• 9:40
  botsitting
• 9:41
  are
• 9:41
  73%
• 9:42
  more
• 9:42
  likely
• 9:42
  to
• 9:42
  be
• 9:42
  actively
• 9:43
  job
• 9:43
  hunting.
• 9:44
  "Workers
• 9:45
  who
• 9:45
  absorb
• 9:45
  it
• 9:45
  without
• 9:45
  recognition
• 9:46
  or
• 9:46
  reward
• 9:46
  grow
• 9:47
  exhausted.
• 9:48
  Then
• 9:48
  they
• 9:48
  grow
• 9:48
  resentful.
• 9:49
  Then
• 9:50
  they
• 9:50
  start
• 9:50
  polishing
• 9:50
  their
• 9:50
  resumes,"
• 9:51
  the
• 9:51
  report
• 9:51
  warns.
• 9:52
  The
• 9:53
  organizations
• 9:53
  seeing
• 9:54
  actual
• 9:54
  gains
• 9:54
  are
• 9:54
  not
• 9:55
  the
• 9:55
  ones
• 9:55
  deploying
• 9:55
  more
• 9:56
  AI
• 9:56
  —
• 9:57
  they
• 9:57
  are
• 9:57
  the
• 9:57
  ones
• 9:57
  investing
• 9:58
  in
• 9:58
  the
• 9:58
  work
• 9:58
  around
• 9:58
  AI:
• 9:59
  setting
• 10:00
  context,
• 10:00
  defining
• 10:01
  what
• 10:01
  "good"
• 10:01
  looks
• 10:01
  like,
• 10:02
  and
• 10:02
  deciding
• 10:02
  what
• 10:03
  should
• 10:03
  never
• 10:03
  have
• 10:03
  been
• 10:03
  handed
• 10:04
  to
• 10:04
  a
• 10:04
  model
• 10:04
  in
• 10:04
  the
• 10:04
  first
• 10:04
  place.
• 10:05
  Running
• 10:06
  directly
• 10:06
  counter
• 10:06
  to
• 10:07
  the
• 10:07
  "cut
• 10:07
  headcount
• 10:07
  and
• 10:07
  automate"
• 10:08
  impulse
• 10:09
  is
• 10:09
  an
• 10:09
  Amazon
• 10:09
  director's
• 10:10
  argument
• 10:10
  for
• 10:10
  doubling
• 10:10
  down
• 10:11
  on
• 10:11
  junior
• 10:11
  hires.
• 10:12
  Jody
• 10:12
  Biggs
• 10:12
  makes
• 10:13
  a
• 10:13
  structural
• 10:13
  case:
• 10:14
  every
• 10:14
  senior
• 10:14
  engineer
• 10:15
  was
• 10:15
  once
• 10:15
  a
• 10:15
  junior
• 10:16
  hire
• 10:16
  who
• 10:16
  learned
• 10:16
  by
• 10:16
  doing,
• 10:17
  and
• 10:17
  cutting
• 10:17
  the
• 10:18
  entry
• 10:18
  ramp
• 10:18
  eliminates
• 10:19
  the
• 10:19
  future
• 10:19
  supply
• 10:19
  of
• 10:19
  the
• 10:19
  senior
• 10:20
  talent
• 10:20
  every
• 10:20
  organization
• 10:21
  needs.
• 10:22
  The
• 10:22
  adoption
• 10:22
  data
• 10:23
  is
• 10:23
  interesting
• 10:23
  —
• 10:24
  AI
• 10:24
  tool
• 10:24
  pickup
• 10:25
  follows
• 10:25
  a
• 10:25
  U-shaped
• 10:26
  curve
• 10:26
  where
• 10:26
  new
• 10:26
  employees
• 10:27
  and
• 10:27
  senior
• 10:27
  staff
• 10:28
  adopt
• 10:28
  quickly
• 10:28
  while
• 10:29
  mid-career
• 10:29
  professionals
• 10:30
  struggle
• 10:30
  most.
• 10:31
  Biggs'
• 10:32
  org
• 10:32
  now
• 10:32
  does
• 10:33
  all
• 10:33
  backfill
• 10:33
  hiring
• 10:34
  at
• 10:34
  the
• 10:34
  entry
• 10:34
  level,
• 10:35
  targeting
• 10:36
  roughly
• 10:36
  a
• 10:36
  quarter
• 10:36
  of
• 10:36
  engineers
• 10:37
  being
• 10:37
  junior
• 10:38
  to
• 10:38
  maintain
• 10:38
  a
• 10:38
  self-sustaining
• 10:39
  pipeline.
• 10:40
  The
• 10:40
  thesis
• 10:41
  is
• 10:41
  that
• 10:41
  AI
• 10:41
  changes
• 10:42
  what
• 10:42
  juniors
• 10:42
  do
• 10:43
  (less
• 10:43
  mechanical
• 10:43
  production,
• 10:44
  more
• 10:44
  judgment
• 10:45
  development),
• 10:45
  not
• 10:46
  whether
• 10:46
  you
• 10:46
  need
• 10:46
  them.
• 10:50
  Google
• 10:51
  Research
• 10:51
  published
• 10:52
  the
• 10:52
  strong
• 10:52
  evidence
• 10:52
  that
• 10:53
  LLM-generated
• 10:53
  models
• 10:54
  can
• 10:54
  match
• 10:55
  human-expert-curated
• 10:56
  forecasting
• 10:56
  systems.
• 10:57
  Their
• 10:57
  Empirical
• 10:58
  Research
• 10:58
  Assistance
• 10:58
  (ERA)
• 10:59
  system
• 10:59
  —
• 10:59
  an
• 10:59
  LLM-guided
• 11:00
  Monte
• 11:00
  Carlo
• 11:01
  tree
• 11:01
  search
• 11:01
  that
• 11:01
  autonomously
• 11:02
  generates,
• 11:02
  evaluates,
• 11:03
  and
• 11:03
  refines
• 11:04
  Python
• 11:04
  forecasting
• 11:04
  code
• 11:05
  —
• 11:05
  produced
• 11:05
  ensembles
• 11:06
  that
• 11:06
  matched
• 11:06
  or
• 11:06
  outperformed
• 11:07
  the
• 11:07
  gold-standard
• 11:08
  CDC
• 11:08
  hub
• 11:08
  ensembles
• 11:09
  across
• 11:09
  influenza,
• 11:10
  COVID-19,
• 11:12
  and
• 11:12
  RSV
• 11:13
  during
• 11:13
  the
• 11:13
  entire
• 11:13
  2025-2026
• 11:15
  US
• 11:15
  respiratory
• 11:16
  season.
• 11:16
  This
• 11:16
  was
• 11:16
  fully
• 11:17
  prospective:
• 11:17
  142
• 11:18
  unique
• 11:18
  model
• 11:19
  prompts,
• 11:19
  over
• 11:19
  207,500
• 11:21
  candidate
• 11:21
  models
• 11:22
  generated,
• 11:22
  54
• 11:23
  selected
• 11:23
  for
• 11:23
  an
• 11:23
  internal
• 11:24
  hub,
• 11:24
  and
• 11:24
  weekly
• 11:25
  submissions
• 11:25
  with
• 11:25
  auditable
• 11:26
  timestamps
• 11:26
  on
• 11:26
  GitHub.
• 11:27
  The
• 11:27
  Google-SAI-FluEns
• 11:28
  ensemble
• 11:28
  ranked
• 11:28
  first
• 11:29
  among
• 11:29
  43
• 11:29
  eligible
• 11:30
  FluSight
• 11:30
  submissions.
• 11:31
  For
• 11:31
  COVID-19,
• 11:33
  first
• 11:33
  among
• 11:33
  12.
• 11:34
  For
• 11:34
  RSV
• 11:34
  —
• 11:34
  a
• 11:35
  "cold
• 11:35
  start"
• 11:35
  scenario
• 11:36
  with
• 11:36
  limited
• 11:36
  historical
• 11:36
  data
• 11:37
  —
• 11:37
  first
• 11:37
  among
• 11:37
  four.
• 11:38
  The
• 11:38
  methodological
• 11:39
  diversity
• 11:39
  is
• 11:39
  the
• 11:39
  key
• 11:40
  detail.
• 11:40
  ERA
• 11:41
  did
• 11:41
  not
• 11:41
  find
• 11:41
  one
• 11:41
  best
• 11:42
  algorithm
• 11:42
  and
• 11:42
  hammer
• 11:43
  it
• 11:43
  into
• 11:43
  submission.
• 11:44
  It
• 11:44
  generated
• 11:44
  models
• 11:45
  spanning
• 11:45
  mechanistic
• 11:46
  compartmental
• 11:46
  formulations,
• 11:47
  classical
• 11:48
  auto-regressive
• 11:48
  methods,
• 11:49
  random
• 11:49
  forests,
• 11:50
  neural
• 11:50
  networks,
• 11:51
  and
• 11:51
  hybrid
• 11:51
  approaches
• 11:52
  —
• 11:52
  then
• 11:53
  ensembled
• 11:53
  the
• 11:53
  best
• 11:53
  performers.
• 11:54
  The
• 11:55
  diversity
• 11:55
  of
• 11:55
  machine-generated
• 11:56
  approaches,
• 11:57
  not
• 11:57
  any
• 11:57
  single
• 11:58
  model's
• 11:58
  brilliance,
• 11:59
  is
• 11:59
  what
• 11:59
  drives
• 11:59
  the
• 11:59
  ensemble's
• 12:00
  competitive
• 12:00
  edge.
• 12:01
  The
• 12:01
  practical
• 12:02
  implication
• 12:02
  is
• 12:02
  significant:
• 12:03
  public
• 12:04
  health
• 12:04
  systems
• 12:04
  that
• 12:04
  currently
• 12:05
  depend
• 12:05
  on
• 12:05
  dozens
• 12:06
  of
• 12:06
  expert
• 12:06
  teams
• 12:06
  maintaining
• 12:07
  individual
• 12:07
  models
• 12:08
  could
• 12:08
  achieve
• 12:08
  equivalent
• 12:09
  forecasting
• 12:09
  quality
• 12:10
  through
• 12:10
  automated
• 12:11
  model
• 12:11
  discovery,
• 12:12
  enabling
• 12:12
  deployment
• 12:13
  at
• 12:13
  scales
• 12:13
  that
• 12:13
  human
• 12:13
  labor
• 12:14
  cannot
• 12:14
  reach
• 12:14
  —
• 12:15
  particularly
• 12:16
  in
• 12:16
  countries
• 12:16
  that
• 12:16
  lack
• 12:16
  established
• 12:17
  forecasting
• 12:17
  hubs.
• 12:19
  The
• 12:19
  results
• 12:19
  contrast
• 12:19
  sharply
• 12:20
  with
• 12:20
  a
• 12:20
  new
• 12:20
  UC
• 12:21
  Berkeley-led
• 12:21
  benchmark
• 12:22
  called
• 12:22
  Agents'
• 12:23
  Last
• 12:23
  Exam
• 12:23
  (ALE),
• 12:24
  where
• 12:24
  frontier
• 12:25
  models
• 12:25
  scored
• 12:25
  below
• 12:26
  25%
• 12:26
  on
• 12:27
  real-world
• 12:27
  tasks
• 12:28
  spanning
• 12:28
  more
• 12:28
  than
• 12:29
  50
• 12:29
  industries.
• 12:30
  OpenAI's
• 12:31
  GPT-5.5
• 12:32
  led
• 12:32
  at
• 12:32
  24%,
• 12:34
  Anthropic's
• 12:34
  Fable
• 12:35
  5
• 12:35
  followed
• 12:35
  at
• 12:35
  22%,
• 12:37
  with
• 12:37
  Gemini,
• 12:38
  DeepSeek,
• 12:38
  and
• 12:38
  Grok
• 12:39
  all
• 12:39
  below
• 12:39
  16%.
• 12:41
  The
• 12:41
  tasks
• 12:41
  were
• 12:41
  drawn
• 12:42
  from
• 12:42
  actual
• 12:42
  professional
• 12:43
  work
• 12:43
  across
• 12:43
  domains
• 12:44
  from
• 12:44
  audio
• 12:44
  processing
• 12:45
  to
• 12:45
  theoretical
• 12:46
  physics,
• 12:46
  contributed
• 12:47
  by
• 12:47
  over
• 12:47
  300
• 12:48
  industry
• 12:48
  experts.
• 12:49
  Stanford
• 12:50
  PhD
• 12:50
  student
• 12:51
  Benjamin
• 12:51
  Liu's
• 12:52
  concern
• 12:52
  is
• 12:53
  not
• 12:53
  the
• 12:53
  pass
• 12:53
  rate
• 12:53
  itself
• 12:54
  but
• 12:54
  the
• 12:54
  failure
• 12:55
  mode:
• 12:56
  "They
• 12:56
  often
• 12:56
  produce
• 12:57
  an
• 12:57
  answer
• 12:57
  that
• 12:57
  looks
• 12:57
  completely
• 12:58
  plausible
• 12:59
  but
• 12:59
  is
• 12:59
  subtly
• 12:59
  wrong,
• 13:00
  and
• 13:00
  in
• 13:00
  science
• 13:01
  a
• 13:01
  confident
• 13:01
  wrong
• 13:02
  answer
• 13:02
  is
• 13:02
  more
• 13:02
  dangerous
• 13:03
  than
• 13:03
  no
• 13:03
  answer."
• 13:04
  The
• 13:04
  coherence
• 13:05
  problem
• 13:05
  gets
• 13:05
  a
• 13:06
  dedicated
• 13:06
  measurement
• 13:07
  tool
• 13:07
  with
• 13:07
  Vending-Bench,
• 13:08
  a
• 13:08
  simulated
• 13:09
  environment
• 13:09
  where
• 13:10
  LLM
• 13:10
  agents
• 13:11
  must
• 13:11
  operate
• 13:11
  a
• 13:11
  vending
• 13:12
  machine
• 13:12
  over
• 13:12
  extended
• 13:13
  horizons
• 13:13
  exceeding
• 13:14
  20
• 13:14
  million
• 13:15
  tokens
• 13:15
  per
• 13:15
  run.
• 13:16
  The
• 13:16
  tasks
• 13:17
  are
• 13:17
  individually
• 13:18
  trivial
• 13:18
  —
• 13:19
  manage
• 13:19
  inventory,
• 13:20
  place
• 13:20
  orders,
• 13:21
  set
• 13:21
  prices,
• 13:22
  collect
• 13:22
  cash
• 13:23
  —
• 13:24
  but
• 13:24
  collectively
• 13:24
  they
• 13:25
  stress
• 13:25
  an
• 13:25
  LLM's
• 13:26
  capacity
• 13:26
  for
• 13:26
  sustained
• 13:27
  decision-making.
• 13:28
  Claude
• 13:29
  3.5
• 13:29
  Sonnet
• 13:30
  led
• 13:30
  with
• 13:30
  a
• 13:30
  mean
• 13:30
  net
• 13:31
  worth
• 13:31
  of
• 13:31
  $2,218
• 13:33
  and
• 13:33
  outsold
• 13:34
  the
• 13:34
  human
• 13:34
  baseline,
• 13:35
  but
• 13:35
  even
• 13:35
  top
• 13:35
  models
• 13:36
  sometimes
• 13:36
  sold
• 13:37
  zero
• 13:37
  items
• 13:37
  across
• 13:37
  entire
• 13:38
  runs,
• 13:39
  highlighting
• 13:39
  extreme
• 13:39
  variance.
• 13:41
  All
• 13:41
  models
• 13:41
  eventually
• 13:42
  stagnated.
• 13:42
  The
• 13:43
  failures
• 13:43
  did
• 13:43
  not
• 13:43
  correlate
• 13:44
  with
• 13:44
  context
• 13:44
  window
• 13:45
  limits,
• 13:45
  suggesting
• 13:46
  breakdowns
• 13:46
  stem
• 13:47
  from
• 13:47
  something
• 13:47
  more
• 13:47
  fundamental
• 13:48
  than
• 13:48
  memory
• 13:48
  constraints.
• 13:49
  The
• 13:49
  gap
• 13:50
  between
• 13:50
  what
• 13:50
  LLMs
• 13:51
  can
• 13:51
  do
• 13:51
  in
• 13:51
  a
• 13:51
  single
• 13:52
  turn
• 13:52
  and
• 13:52
  what
• 13:52
  they
• 13:52
  sustain
• 13:53
  over
• 13:53
  time
• 13:53
  remains
• 13:54
  the
• 13:54
  field's
• 13:54
  central
• 13:55
  unsolved
• 13:55
  problem
• 13:55
  —
• 13:56
  a
• 13:56
  reality
• 13:57
  that
• 13:57
  shows
• 13:57
  up
• 13:57
  whether
• 13:58
  you
• 13:58
  measure
• 13:58
  it
• 13:58
  with
• 13:58
  disease
• 13:59
  forecasts,
• 13:59
  professional
• 14:00
  exams,
• 14:01
  or
• 14:01
  simulated
• 14:01
  vending
• 14:02
  machines.
• 14:05
  The
• 14:05
  Fable
• 14:05
  5
• 14:05
  export
• 14:06
  controls
• 14:06
  and
• 14:06
  the
• 14:06
  broader
• 14:07
  trend
• 14:07
  of
• 14:07
  model
• 14:07
  restrictions
• 14:08
  are
• 14:08
  accelerating
• 14:09
  a
• 14:09
  parallel
• 14:09
  infrastructure:
• 14:10
  takedown-resilient
• 14:11
  distribution.
• 14:13
  Heretic
• 14:13
  Grimoire
• 14:14
  1.4
• 14:14
  represents
• 14:15
  the
• 14:15
  community's
• 14:16
  answer
• 14:16
  to
• 14:16
  platform-level
• 14:17
  model
• 14:17
  removals
• 14:17
  —
• 14:19
  a
• 14:19
  system
• 14:19
  that
• 14:19
  reduces
• 14:20
  model
• 14:20
  backups
• 14:20
  to
• 14:20
  9KB
• 14:21
  reproducible
• 14:22
  manifests
• 14:22
  distributed
• 14:23
  via
• 14:23
  IPFS,
• 14:24
  allowing
• 14:24
  anyone
• 14:25
  to
• 14:25
  reconstruct
• 14:25
  a
• 14:25
  model
• 14:26
  from
• 14:26
  its
• 14:26
  manifest
• 14:26
  even
• 14:27
  if
• 14:27
  the
• 14:27
  original
• 14:27
  hosting
• 14:28
  is
• 14:28
  removed.
• 14:29
  The
• 14:29
  architecture
• 14:30
  borrows
• 14:30
  from
• 14:30
  BitTorrent's
• 14:30
  resilience
• 14:31
  playbook:
• 14:32
  no
• 14:32
  single
• 14:32
  point
• 14:32
  of
• 14:32
  failure,
• 14:33
  no
• 14:33
  single
• 14:33
  entity
• 14:34
  to
• 14:34
  serve
• 14:34
  a
• 14:34
  takedown
• 14:34
  notice.
• 14:35
  It
• 14:36
  is
• 14:36
  a
• 14:36
  direct
• 14:36
  response
• 14:37
  to
• 14:37
  a
• 14:37
  world
• 14:37
  where
• 14:37
  model
• 14:37
  access
• 14:38
  can
• 14:38
  be
• 14:38
  revoked
• 14:38
  overnight
• 14:39
  by
• 14:39
  government
• 14:39
  order.
• 14:40
  The
• 14:40
  licensing
• 14:41
  landscape
• 14:41
  is
• 14:41
  shifting
• 14:42
  beneath
• 14:42
  the
• 14:42
  open-weights
• 14:43
  community's
• 14:43
  feet
• 14:43
  as
• 14:43
  well.
• 14:44
  A
• 14:44
  poll
• 14:44
  from
• 14:44
  z.ai
• 14:45
  found
• 14:46
  that
• 14:46
  MIT-licensed
• 14:46
  open
• 14:47
  weights
• 14:47
  are
• 14:47
  losing
• 14:47
  ground
• 14:48
  —
• 14:49
  a
• 14:49
  data
• 14:49
  point
• 14:49
  that
• 14:49
  challenges
• 14:50
  the
• 14:50
  assumption
• 14:50
  that
• 14:50
  permissive
• 14:51
  licensing
• 14:51
  is
• 14:51
  the
• 14:51
  default
• 14:52
  future.
• 14:53
  DeepSeek's
• 14:53
  MIT-licensed
• 14:54
  R1
• 14:54
  model
• 14:55
  put
• 14:55
  pressure
• 14:55
  on
• 14:55
  every
• 14:55
  competitor
• 14:56
  to
• 14:56
  match
• 14:56
  its
• 14:56
  terms,
• 14:57
  but
• 14:57
  the
• 14:57
  polling
• 14:57
  data
• 14:58
  suggests
• 14:58
  the
• 14:58
  industry
• 14:58
  is
• 14:58
  moving
• 14:59
  in
• 14:59
  the
• 14:59
  opposite
• 14:59
  direction:
• 15:00
  more
• 15:00
  restrictive
• 15:01
  licenses,
• 15:01
  more
• 15:02
  usage
• 15:02
  conditions,
• 15:03
  more
• 15:03
  strings
• 15:03
  attached.
• 15:04
  The
• 15:04
  community
• 15:05
  is
• 15:05
  splitting
• 15:05
  into
• 15:05
  factions
• 15:06
  —
• 15:06
  those
• 15:06
  who
• 15:07
  consider
• 15:07
  permissive
• 15:07
  licensing
• 15:08
  sufficient
• 15:08
  and
• 15:08
  those
• 15:09
  who
• 15:09
  view
• 15:09
  decentralized
• 15:10
  hosting
• 15:10
  as
• 15:10
  the
• 15:10
  only
• 15:11
  guarantee
• 15:11
  that
• 15:11
  matters
• 15:12
  when
• 15:12
  a
• 15:12
  single
• 15:12
  phone
• 15:12
  call
• 15:12
  can
• 15:13
  disable
• 15:13
  a
• 15:13
  model
• 15:13
  globally.
• 15:16
  If
• 15:16
  the
• 15:16
  GPU
• 15:17
  lunchbox
• 15:17
  war
• 15:18
  is
• 15:18
  about
• 15:18
  who
• 15:18
  owns
• 15:18
  the
• 15:18
  silicon,
• 15:19
  the
• 15:19
  meta-harness
• 15:20
  race
• 15:20
  is
• 15:20
  about
• 15:20
  who
• 15:21
  controls
• 15:21
  the
• 15:21
  orchestration
• 15:22
  layer
• 15:22
  above
• 15:22
  it.
• 15:23
  Databricks
• 15:23
  open-sourced
• 15:24
  Omni
• 15:24
  Agent,
• 15:25
  a
• 15:25
  meta-harness
• 15:25
  that
• 15:25
  orchestrates
• 15:26
  multiple
• 15:26
  AI
• 15:26
  coding
• 15:27
  assistants
• 15:27
  —
• 15:28
  Claude
• 15:28
  Code
• 15:29
  for
• 15:29
  implementation,
• 15:30
  Codex
• 15:30
  for
• 15:30
  review,
• 15:31
  local
• 15:31
  models
• 15:32
  via
• 15:32
  Ollama
• 15:32
  for
• 15:32
  experimentation
• 15:33
  —
• 15:34
  in
• 15:34
  a
• 15:34
  single
• 15:34
  session
• 15:34
  with
• 15:34
  shared
• 15:35
  history,
• 15:35
  policies,
• 15:36
  and
• 15:36
  guardrails.
• 15:37
  The
• 15:37
  tool
• 15:37
  ships
• 15:38
  with
• 15:38
  example
• 15:38
  orchestrators
• 15:39
  that
• 15:39
  delegate
• 15:39
  tasks
• 15:40
  across
• 15:40
  coding
• 15:40
  agents,
• 15:41
  monitor
• 15:41
  completion,
• 15:42
  and
• 15:42
  handle
• 15:42
  handoffs
• 15:43
  automatically.
• 15:44
  Custom
• 15:45
  guardrails
• 15:45
  are
• 15:45
  defined
• 15:45
  in
• 15:46
  Python:
• 15:46
  human-in-the-loop
• 15:47
  approval
• 15:48
  for
• 15:48
  dangerous
• 15:48
  operations
• 15:49
  like
• 15:49
  force
• 15:49
  pushes,
• 15:50
  autonomous
• 15:50
  execution
• 15:51
  for
• 15:51
  everything
• 15:51
  else.
• 15:52
  The
• 15:52
  thesis
• 15:52
  behind
• 15:52
  meta-harnesses
• 15:53
  is
• 15:53
  that
• 15:53
  no
• 15:54
  single
• 15:54
  model
• 15:54
  or
• 15:54
  provider
• 15:55
  should
• 15:55
  own
• 15:55
  an
• 15:55
  entire
• 15:55
  AI
• 15:56
  coding
• 15:56
  workflow.
• 15:57
  The
• 15:57
  harness
• 15:57
  matters
• 15:57
  as
• 15:57
  much
• 15:58
  as
• 15:58
  —
• 15:58
  or
• 15:58
  more
• 15:58
  than
• 15:58
  —
• 15:59
  the
• 15:59
  model.
• 15:59
  Fable
• 16:00
  5's
• 16:00
  sudden
• 16:00
  disappearance
• 16:01
  proved
• 16:01
  the
• 16:01
  point:
• 16:01
  developers
• 16:02
  relying
• 16:02
  on
• 16:02
  a
• 16:02
  single
• 16:03
  provider
• 16:03
  woke
• 16:03
  up
• 16:03
  to
• 16:03
  a
• 16:03
  blank
• 16:03
  screen.
• 16:04
  The
• 16:04
  meta-harness
• 16:05
  pattern
• 16:05
  hedges
• 16:05
  that
• 16:05
  risk
• 16:06
  by
• 16:06
  making
• 16:06
  the
• 16:06
  orchestration
• 16:07
  layer
• 16:07
  provider-agnostic.
• 16:08
  Databricks'
• 16:09
  CTO
• 16:09
  is
• 16:10
  reportedly
• 16:10
  dog-fooding
• 16:10
  Omni
• 16:11
  Agent
• 16:11
  for
• 16:11
  everyday
• 16:12
  engineering
• 16:12
  internally,
• 16:13
  and
• 16:13
  the
• 16:13
  framework
• 16:14
  supports
• 16:14
  collaboration
• 16:14
  across
• 16:15
  devices
• 16:15
  and
• 16:15
  teams
• 16:16
  via
• 16:16
  a
• 16:16
  web
• 16:16
  UI
• 16:16
  —
• 16:17
  a
• 16:17
  feature
• 16:17
  most
• 16:18
  coding
• 16:18
  agents
• 16:18
  still
• 16:19
  lack.
• 16:19
  The
• 16:20
  economic
• 16:20
  argument
• 16:20
  underneath
• 16:21
  all
• 16:21
  of
• 16:21
  this
• 16:21
  got
• 16:21
  a
• 16:21
  detailed
• 16:22
  treatment
• 16:22
  in
• 16:22
  a
• 16:22
  video
• 16:23
  analysis
• 16:23
  of
• 16:23
  whether
• 16:23
  AI
• 16:24
  constitutes
• 16:24
  a
• 16:24
  bubble.
• 16:25
  The
• 16:25
  core
• 16:26
  thesis:
• 16:26
  asking
• 16:27
  "is
• 16:27
  AI
• 16:27
  a
• 16:27
  bubble?"
• 16:28
  is
• 16:28
  the
• 16:28
  wrong
• 16:28
  question.
• 16:30
  The
• 16:30
  right
• 16:30
  question
• 16:31
  is
• 16:31
  which
• 16:31
  parts
• 16:31
  of
• 16:31
  the
• 16:31
  buildout
• 16:32
  are
• 16:32
  speculative
• 16:32
  froth
• 16:33
  and
• 16:33
  which
• 16:33
  are
• 16:34
  infrastructure
• 16:34
  for
• 16:34
  demand
• 16:35
  that
• 16:35
  already
• 16:35
  exists.
• 16:37
  OpenAI's
• 16:37
  revenue
• 16:38
  went
• 16:38
  from
• 16:38
  $2
• 16:38
  billion
• 16:38
  in
• 16:39
  2023
• 16:40
  to
• 16:40
  over
• 16:40
  $20
• 16:40
  billion
• 16:41
  in
• 16:41
  2025.
• 16:43
  Anthropic
• 16:43
  grew
• 16:44
  even
• 16:44
  faster.
• 16:45
  NVIDIA's
• 16:46
  data
• 16:46
  center
• 16:46
  business
• 16:46
  did
• 16:47
  $194
• 16:47
  billion
• 16:48
  in
• 16:48
  fiscal
• 16:48
  2026.
• 16:50
  But
• 16:50
  the
• 16:50
  transformative
• 16:51
  shift
• 16:51
  is
• 16:51
  inference,
• 16:52
  not
• 16:52
  training
• 16:52
  —
• 16:53
  agents
• 16:53
  that
• 16:54
  loop,
• 16:54
  call
• 16:54
  tools,
• 16:55
  write
• 16:55
  code,
• 16:56
  check
• 16:56
  results,
• 16:57
  and
• 16:57
  burn
• 16:57
  millions
• 16:57
  of
• 16:57
  tokens
• 16:58
  per
• 16:58
  run.
• 16:59
  A
• 16:59
  single
• 17:00
  agent
• 17:00
  workflow
• 17:00
  can
• 17:00
  consume
• 17:01
  thousands
• 17:01
  of
• 17:01
  times
• 17:01
  the
• 17:02
  inference
• 17:02
  cost
• 17:02
  of
• 17:02
  a
• 17:02
  chat
• 17:03
  conversation.
• 17:04
  The
• 17:04
  buildout
• 17:04
  is
• 17:04
  real;
• 17:05
  the
• 17:05
  question
• 17:06
  is
• 17:06
  who
• 17:06
  captures
• 17:06
  value
• 17:07
  when
• 17:07
  tokens
• 17:07
  get
• 17:07
  cheaper.
• 17:08
  A
• 17:08
  companion
• 17:09
  analysis
• 17:09
  explores
• 17:10
  how
• 17:10
  the
• 17:10
  AI
• 17:11
  investment
• 17:11
  landscape
• 17:12
  is
• 17:12
  sorting
• 17:12
  into
• 17:13
  winners
• 17:13
  and
• 17:13
  commodity
• 17:13
  providers
• 17:14
  as
• 17:14
  the
• 17:14
  market
• 17:14
  matures.
• 17:16
  On
• 17:16
  the
• 17:16
  tooling
• 17:16
  side,
• 17:17
  the
• 17:17
  market
• 17:17
  is
• 17:17
  bifurcating
• 17:18
  between
• 17:18
  maximalist
• 17:19
  orchestrators
• 17:19
  and
• 17:20
  minimalist,
• 17:20
  opinionated
• 17:21
  frameworks.
• 17:22
  StandardAgents'
• 17:23
  Arrow-JS
• 17:23
  is
• 17:23
  a
• 17:23
  tiny,
• 17:24
  type-safe
• 17:24
  reactive
• 17:25
  UI
• 17:25
  runtime
• 17:26
  built
• 17:26
  specifically
• 17:26
  for
• 17:26
  coding
• 17:27
  agents
• 17:27
  —
• 17:28
  template
• 17:28
  literals,
• 17:29
  direct
• 17:29
  DOM
• 17:29
  updates,
• 17:30
  minimal
• 17:30
  API
• 17:31
  surface.
• 17:32
  It
• 17:32
  includes
• 17:32
  SSR,
• 17:33
  hydration,
• 17:34
  and
• 17:34
  a
• 17:34
  QuickJS/WASM-backed
• 17:37
  sandbox
• 17:38
  for
• 17:38
  executing
• 17:38
  code
• 17:38
  off
• 17:39
  the
• 17:39
  host
• 17:39
  realm.
• 17:40
  The
• 17:40
  bet
• 17:40
  is
• 17:40
  that
• 17:40
  agents
• 17:41
  understand
• 17:41
  JavaScript
• 17:42
  primitives
• 17:42
  better
• 17:42
  than
• 17:42
  complex
• 17:43
  framework
• 17:43
  abstractions.
• 17:45
  At
• 17:45
  the
• 17:45
  other
• 17:45
  end
• 17:45
  of
• 17:45
  the
• 17:45
  spectrum,
• 17:46
  archex
• 17:47
  provides
• 17:47
  local-first,
• 17:48
  deterministic
• 17:49
  code-context
• 17:50
  extraction
• 17:50
  for
• 17:50
  AI
• 17:51
  agents
• 17:51
  with
• 17:51
  no
• 17:52
  API
• 17:52
  key
• 17:52
  and
• 17:53
  no
• 17:53
  telemetry
• 17:53
  —
• 17:54
  Apache
• 17:55
  2.0
• 17:55
  licensed,
• 17:56
  designed
• 17:57
  for
• 17:57
  developers
• 17:58
  who
• 17:58
  want
• 17:58
  their
• 17:58
  code
• 17:58
  context
• 17:59
  computed
• 17:59
  locally
• 18:00
  and
• 18:00
  reproducibly.
• 18:01
  And
• 18:02
  Ironsmith
• 18:02
  takes
• 18:02
  the
• 18:03
  agent-builds-software
• 18:04
  pattern
• 18:04
  to
• 18:04
  its
• 18:04
  logical
• 18:05
  conclusion:
• 18:06
  an
• 18:06
  open-source
• 18:06
  macOS
• 18:07
  app
• 18:07
  that
• 18:08
  creates
• 18:08
  other
• 18:08
  macOS
• 18:09
  apps
• 18:09
  from
• 18:09
  natural
• 18:09
  language
• 18:10
  prompts,
• 18:11
  working
• 18:11
  entirely
• 18:11
  with
• 18:12
  local
• 18:12
  models.