Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 01, 2026

Transcript

AGI Dreams Podcast – May 01, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  shai-hulud
• 0:00
  threat
• 0:01
  actor
• 0:01
  is
• 0:01
  back,
• 0:01
  and
• 0:01
  this
• 0:01
  time
• 0:02
  the
• 0:02
  sandworm
• 0:02
  burrowed
• 0:02
  into
• 0:03
  PyTorch.
• 0:04
  Semgrep
• 0:04
  disclosed
• 0:04
  that
• 0:04
  the
• 0:05
  PyPI
• 0:05
  package
• 0:05
  lightning
• 0:06
  --
• 0:06
  the
• 0:06
  widely-used
• 0:07
  PyTorch
• 0:07
  Lightning
• 0:07
  deep
• 0:08
  learning
• 0:08
  framework
• 0:08
  --
• 0:08
  was
• 0:09
  compromised
• 0:09
  in
• 0:09
  versions
• 0:10
  2.6.2
• 0:11
  and
• 0:11
  2.6.3,
• 0:13
  published
• 0:14
  April
• 0:14
  30.
• 0:15
  Anyone
• 0:15
  running
• 0:15
  pip
• 0:16
  install
• 0:16
  lightning
• 0:16
  during
• 0:17
  the
• 0:17
  affected
• 0:17
  window
• 0:18
  activated
• 0:18
  a
• 0:18
  hidden
• 0:18
  runtime
• 0:19
  directory
• 0:19
  containing
• 0:20
  a
• 0:20
  14.8
• 0:21
  MB
• 0:21
  obfuscated
• 0:21
  JavaScript
• 0:22
  payload
• 0:22
  that
• 0:22
  executes
• 0:23
  on
• 0:23
  import.
• 0:24
  The
• 0:24
  malware
• 0:24
  steals
• 0:25
  credentials,
• 0:25
  authentication
• 0:26
  tokens,
• 0:26
  environment
• 0:27
  variables,
• 0:27
  and
• 0:28
  cloud
• 0:28
  secrets
• 0:28
  across
• 0:29
  all
• 0:29
  three
• 0:29
  major
• 0:29
  providers
• 0:30
  (AWS
• 0:30
  via
• 0:31
  IMDSv2
• 0:32
  and
• 0:32
  Secrets
• 0:32
  Manager
• 0:32
  enumeration,
• 0:33
  Azure
• 0:34
  via
• 0:34
  DefaultAzureCredential
• 0:35
  and
• 0:35
  Key
• 0:35
  Vault,
• 0:36
  GCP
• 0:36
  via
• 0:37
  GoogleAuth
• 0:37
  and
• 0:37
  Secret
• 0:38
  Manager).
• 0:39
  On
• 0:39
  GitHub
• 0:39
  Actions
• 0:39
  runners,
• 0:40
  it
• 0:40
  dumps
• 0:40
  Runner.Worker
• 0:41
  process
• 0:41
  memory
• 0:42
  to
• 0:42
  extract
• 0:42
  every
• 0:43
  secret
• 0:43
  marked
• 0:47
  The
• 0:47
  cross-ecosystem
• 0:48
  propagation
• 0:49
  is
• 0:49
  particularly
• 0:49
  nasty:
• 0:50
  if
• 0:51
  the
• 0:51
  malware
• 0:51
  finds
• 0:51
  npm
• 0:52
  publish
• 0:52
  credentials,
• 0:53
  it
• 0:53
  injects
• 0:53
  a
• 0:53
  dropper
• 0:54
  into
• 0:54
  every
• 0:54
  package
• 0:54
  that
• 0:55
  token
• 0:55
  can
• 0:55
  publish,
• 0:56
  bumps
• 0:56
  the
• 0:56
  patch
• 0:56
  version,
• 0:57
  and
• 0:57
  republishes
• 0:57
  --
• 0:58
  worming
• 0:58
  from
• 0:59
  PyPI
• 0:59
  into
• 0:59
  npm
• 1:00
  silently.
• 1:01
  What
• 1:01
  makes
• 1:01
  this
• 1:01
  iteration
• 1:02
  notable
• 1:02
  beyond
• 1:03
  the
• 1:03
  credential
• 1:03
  theft
• 1:03
  is
• 1:03
  its
• 1:04
  abuse
• 1:04
  of
• 1:04
  developer
• 1:04
  tooling
• 1:05
  for
• 1:05
  persistence.
• 1:06
  The
• 1:06
  malware
• 1:06
  writes
• 1:07
  a
• 1:07
  SessionStart
• 1:07
  hook
• 1:07
  into
• 1:09
  .claude/settings.json
• 1:10
  with
• 1:11
  matcher:
• 1:11
  "",
• 1:11
  firing
• 1:12
  every
• 1:12
  time
• 1:12
  a
• 1:12
  developer
• 1:12
  opens
• 1:13
  Claude
• 1:13
  Code
• 1:13
  in
• 1:13
  the
• 1:13
  infected
• 1:14
  repository
• 1:14
  --
• 1:15
  no
• 1:15
  user
• 1:16
  action
• 1:16
  required
• 1:16
  beyond
• 1:17
  launching
• 1:17
  the
• 1:17
  session.
• 1:18
  A
• 1:18
  parallel
• 1:21
  .vscode/tasks.json
• 1:24
  hook
• 1:24
  triggers
• 1:24
  on
• 1:24
  folder
• 1:25
  open
• 1:25
  for
• 1:25
  VS
• 1:25
  Code
• 1:25
  users.
• 1:27
  Both
• 1:27
  invoke
• 1:27
  a
• 1:27
  Bun
• 1:28
  runtime
• 1:28
  bootstrapper
• 1:29
  that
• 1:29
  downloads
• 1:29
  bun-v1.3.13
• 1:32
  if
• 1:32
  absent,
• 1:33
  executes
• 1:33
  the
• 1:33
  full
• 1:33
  payload,
• 1:34
  and
• 1:34
  cleans
• 1:34
  up
• 1:35
  from
• 1:35
  /tmp.
• 1:36
  This
• 1:37
  may
• 1:37
  be
• 1:37
  among
• 1:37
  the
• 1:37
  first
• 1:38
  documented
• 1:38
  instances
• 1:39
  of
• 1:39
  malware
• 1:39
  weaponizing
• 1:40
  Claude
• 1:40
  Code's
• 1:40
  hook
• 1:41
  system
• 1:41
  in
• 1:41
  a
• 1:41
  real-world
• 1:42
  attack.
• 1:43
  If
• 1:43
  a
• 1:43
  GitHub
• 1:43
  token
• 1:43
  with
• 1:44
  write
• 1:44
  access
• 1:44
  is
• 1:45
  present,
• 1:45
  a
• 1:45
  bonus
• 1:46
  payload
• 1:46
  pushes
• 1:46
  a
• 1:47
  workflow
• 1:47
  named
• 1:47
  "Formatter"
• 1:48
  that
• 1:48
  dumps
• 1:49
  all
• 1:49
  repository
• 1:50
  secrets
• 1:50
  via
• 1:50
  ${{
• 1:51
  toJSON(secrets)
• 1:51
  }}
• 1:52
  as
• 1:52
  a
• 1:52
  downloadable
• 1:53
  Actions
• 1:53
  artifact.
• 1:54
  Any
• 1:54
  machine
• 1:55
  that
• 1:55
  imported
• 1:56
  the
• 1:56
  malicious
• 1:56
  package
• 1:56
  should
• 1:57
  be
• 1:57
  treated
• 1:57
  as
• 1:57
  fully
• 1:58
  look
• 1:59
  for
• 1:59
  commit
• 1:59
  messages
• 2:00
  prefixed
• 2:00
  with
• 2:01
  EveryBoiWeBuildIsAWormyBoi
• 2:02
  and
• 2:02
  repos
• 2:03
  with
• 2:03
  the
• 2:03
  description
• 2:04
  "A
• 2:04
  Mini
• 2:04
  Shai-Hulud
• 2:05
  has
• 2:05
  Appeared."
• 2:06
  Meanwhile,
• 2:07
  a
• 2:07
  separate
• 2:08
  disclosure
• 2:08
  highlighted
• 2:09
  the
• 2:09
  fragile
• 2:09
  state
• 2:09
  of
• 2:10
  kernel
• 2:10
  vulnerability
• 2:11
  coordination.
• 2:12
  CVE-2026-31431
• 2:14
  (CopyFail),
• 2:16
  a
• 2:16
  local
• 2:17
  privilege
• 2:17
  escalation
• 2:18
  in
• 2:18
  the
• 2:18
  Linux
• 2:18
  kernel
• 2:19
  introduced
• 2:19
  in
• 2:19
  version
• 2:19
  4.14
• 2:20
  back
• 2:20
  in
• 2:21
  2017,
• 2:22
  was
• 2:22
  fixed
• 2:22
  in
• 2:22
  6.18.22
• 2:25
  and
• 2:25
  6.19.12
• 2:26
  --
• 2:28
  but
• 2:28
  longterm
• 2:28
  kernels
• 2:28
  6.12
• 2:29
  through
• 2:30
  5.10
• 2:31
  remain
• 2:31
  unpatched,
• 2:32
  and
• 2:32
  the
• 2:32
  fix
• 2:32
  does
• 2:32
  not
• 2:32
  backport
• 2:33
  cleanly.
• 2:34
  Sam
• 2:34
  James
• 2:34
  of
• 2:34
  Gentoo
• 2:35
  posted
• 2:35
  a
• 2:35
  workaround
• 2:36
  patch,
• 2:36
  noting
• 2:36
  bluntly:
• 2:37
  "for
• 2:37
  Linux
• 2:38
  kernel
• 2:38
  vulnerabilities,
• 2:39
  unless
• 2:40
  the
• 2:40
  reporter
• 2:40
  chooses
• 2:41
  to
• 2:41
  bring
• 2:41
  it
• 2:41
  to
• 2:41
  the
• 2:41
  linux-distros
• 2:42
  ML,
• 2:43
  there
• 2:43
  is
• 2:43
  no
• 2:43
  heads-up
• 2:44
  to
• 2:44
  distributions."
• 2:45
  The
• 2:45
  structural
• 2:46
  gap
• 2:46
  between
• 2:46
  kernel
• 2:47
  maintainers
• 2:47
  and
• 2:47
  downstream
• 2:48
  distributions
• 2:49
  remains
• 2:49
  a
• 2:49
  persistent
• 2:50
  weakness
• 2:50
  in
• 2:50
  the
• 2:50
  open-source
• 2:51
  security
• 2:51
  model.
• 2:52
  On
• 2:53
  the
• 2:53
  lighter
• 2:53
  side
• 2:53
  of
• 2:53
  tool
• 2:53
  misbehavior,
• 2:54
  Theo
• 2:55
  Browne
• 2:55
  flagged
• 2:55
  that
• 2:55
  Claude
• 2:56
  Code
• 2:56
  refuses
• 2:56
  requests
• 2:57
  or
• 2:57
  charges
• 2:58
  extra
• 2:58
  if
• 2:58
  recent
• 2:59
  commits
• 2:59
  mention
• 3:00
  "OpenClaw"
• 3:00
  in
• 3:01
  a
• 3:01
  JSON
• 3:01
  blob
• 3:01
  --
• 3:02
  even
• 3:02
  in
• 3:02
  an
• 3:02
  empty
• 3:03
  repository
• 3:03
  with
• 3:04
  no
• 3:04
  code.
• 3:05
  The
• 3:05
  behavior
• 3:06
  appears
• 3:06
  to
• 3:06
  be
• 3:06
  an
• 3:06
  artifact
• 3:07
  of
• 3:07
  competitive-product
• 3:08
  detection
• 3:08
  leaking
• 3:09
  into
• 3:09
  billing
• 3:09
  logic,
• 3:10
  and
• 3:10
  while
• 3:11
  Anthropic
• 3:11
  has
• 3:11
  not
• 3:12
  commented
• 3:12
  publicly,
• 3:13
  it
• 3:13
  is
• 3:13
  a
• 3:13
  reminder
• 3:14
  that
• 3:14
  the
• 3:14
  tools
• 3:14
  developers
• 3:15
  rely
• 3:15
  on
• 3:15
  carry
• 3:16
  their
• 3:16
  own
• 3:16
  opaque
• 3:16
  policy
• 3:17
  layers.
• 3:21
  The
• 3:21
  prevailing
• 3:22
  narrative
• 3:22
  around
• 3:22
  AI-driven
• 3:23
  vulnerability
• 3:23
  discovery
• 3:24
  has
• 3:24
  been
• 3:24
  that
• 3:24
  it
• 3:24
  requires
• 3:25
  frontier
• 3:25
  models
• 3:26
  with
• 3:26
  restricted
• 3:26
  access
• 3:27
  --
• 3:27
  Anthropic's
• 3:28
  Mythos
• 3:28
  program
• 3:28
  being
• 3:29
  the
• 3:29
  marquee
• 3:29
  example.
• 3:30
  Niels
• 3:30
  Provos,
• 3:31
  the
• 3:31
  original
• 3:31
  author
• 3:31
  of
• 3:32
  the
• 3:32
  1998
• 3:32
  OpenBSD
• 3:33
  TCP
• 3:34
  SACK
• 3:34
  implementation
• 3:35
  that
• 3:35
  Anthropic's
• 3:35
  red
• 3:36
  team
• 3:36
  flagged
• 3:36
  as
• 3:36
  their
• 3:37
  headline
• 3:37
  finding,
• 3:38
  decided
• 3:38
  to
• 3:38
  test
• 3:38
  that
• 3:39
  claim.
• 3:39
  Using
• 3:40
  his
• 3:40
  open-source
• 3:40
  framework
• 3:41
  built
• 3:41
  around
• 3:41
  finite-state-machine
• 3:42
  workflows,
• 3:43
  he
• 3:43
  replicated
• 3:44
  the
• 3:44
  OpenBSD
• 3:44
  discovery
• 3:45
  and
• 3:45
  then
• 3:46
  autonomously
• 3:46
  found
• 3:46
  new
• 3:47
  zero-day
• 3:47
  vulnerabilities
• 3:48
  in
• 3:48
  four
• 3:48
  widely-used
• 3:49
  open-source
• 3:49
  projects
• 3:49
  --
• 3:50
  using
• 3:50
  commercial
• 3:51
  models
• 3:51
  like
• 3:51
  Claude
• 3:51
  Sonnet
• 3:52
  and
• 3:52
  open-weight
• 3:52
  models
• 3:53
  like
• 3:53
  Z.AI's
• 3:53
  offering.
• 3:54
  The
• 3:54
  framework
• 3:55
  employs
• 3:55
  a
• 3:55
  central
• 3:55
  orchestrator
• 3:56
  agent
• 3:56
  that
• 3:57
  routes
• 3:57
  to
• 3:57
  specialized
• 3:57
  agents
• 3:58
  based
• 3:58
  on
• 3:58
  an
• 3:58
  append-only
• 3:59
  execution
• 3:59
  journal,
• 4:00
  with
• 4:00
  tiered
• 4:00
  harness
• 4:00
  construction:
• 4:01
  single-function
• 4:02
  isolation,
• 4:03
  multi-component
• 4:03
  harnesses,
• 4:04
  and
• 4:04
  full
• 4:04
  end-to-end
• 4:05
  VM
• 4:05
  validation.
• 4:06
  The
• 4:06
  economics
• 4:07
  are
• 4:07
  striking.
• 4:08
  A
• 4:08
  single
• 4:08
  investigation
• 4:09
  consumes
• 4:09
  roughly
• 4:09
  10
• 4:10
  million
• 4:10
  tokens
• 4:10
  on
• 4:10
  Sonnet
• 4:11
  ($30)
• 4:11
  or
• 4:11
  Opus
• 4:12
  ($150).
• 4:14
  Z.AI's
• 4:14
  hosted
• 4:15
  open-weight
• 4:15
  model
• 4:16
  averaged
• 4:16
  27
• 4:16
  million
• 4:17
  tokens
• 4:17
  per
• 4:17
  run
• 4:18
  but
• 4:18
  at
• 4:18
  comparable
• 4:18
  cost
• 4:19
  per
• 4:19
  investigation
• 4:19
  due
• 4:20
  to
• 4:20
  lower
• 4:20
  per-token
• 4:20
  pricing.
• 4:21
  One
• 4:21
  discovery
• 4:22
  --
• 4:22
  an
• 4:23
  integer
• 4:23
  truncation
• 4:23
  flaw
• 4:24
  in
• 4:24
  a
• 4:24
  foundational
• 4:24
  library
• 4:25
  dormant
• 4:25
  for
• 4:25
  eighteen
• 4:26
  years
• 4:26
  --
• 4:26
  required
• 4:27
  manual
• 4:27
  confirmation
• 4:28
  of
• 4:28
  exploitability
• 4:28
  primitives,
• 4:29
  during
• 4:29
  which
• 4:29
  the
• 4:29
  model's
• 4:30
  AUP
• 4:30
  guardrails
• 4:31
  blocked
• 4:31
  further
• 4:31
  collaboration
• 4:32
  after
• 4:32
  two
• 4:32
  of
• 4:32
  seven
• 4:33
  planned
• 4:33
  exploit-development
• 4:34
  steps.
• 4:35
  Provos
• 4:35
  frames
• 4:35
  this
• 4:36
  friction
• 4:36
  as
• 4:36
  the
• 4:36
  core
• 4:37
  asymmetry:
• 4:38
  "a
• 4:38
  defender
• 4:38
  doing
• 4:39
  legitimate
• 4:39
  work
• 4:39
  hit
• 4:40
  friction
• 4:40
  that
• 4:40
  a
• 4:40
  well-resourced
• 4:41
  adversary
• 4:42
  using
• 4:42
  uncensored
• 4:43
  open-weight
• 4:43
  models
• 4:43
  would
• 4:43
  not."
• 4:44
  The
• 4:45
  seven-step
• 4:45
  refusal
• 4:46
  is
• 4:46
  exactly
• 4:46
  the
• 4:46
  kind
• 4:46
  of
• 4:47
  gate
• 4:47
  that
• 4:47
  does
• 4:47
  not
• 4:47
  exist
• 4:48
  for
• 4:48
  attackers
• 4:48
  running
• 4:49
  locally.
• 4:50
  The
• 4:50
  framework
• 4:50
  is
• 4:50
  open-source,
• 4:51
  and
• 4:52
  Provos
• 4:52
  encourages
• 4:52
  security
• 4:53
  engineers
• 4:53
  to
• 4:54
  contribute
• 4:54
  --
• 4:55
  particularly
• 4:55
  toward
• 4:55
  easier
• 4:56
  onboarding.
• 4:57
  The
• 4:57
  implication
• 4:58
  is
• 4:58
  clear:
• 4:59
  the
• 4:59
  capability
• 4:59
  floor
• 5:00
  for
• 5:00
  automated
• 5:00
  vulnerability
• 5:01
  discovery
• 5:01
  is
• 5:02
  now
• 5:02
  low
• 5:02
  enough
• 5:02
  that
• 5:02
  commodity
• 5:03
  models
• 5:03
  can
• 5:03
  clear
• 5:04
  it,
• 5:04
  and
• 5:04
  the
• 5:04
  scaffolding
• 5:05
  matters
• 5:05
  more
• 5:05
  than
• 5:05
  the
• 5:06
  model.
• 5:09
  The
• 5:09
  Agentics
• 5:10
  Foundation
• 5:10
  published
• 5:10
  the
• 5:11
  OIA
• 5:11
  Model
• 5:11
  --
• 5:12
  Open
• 5:12
  Intelligence
• 5:13
  Architecture
• 5:13
  --
• 5:14
  a
• 5:14
  nine-layer
• 5:14
  reference
• 5:15
  architecture
• 5:15
  for
• 5:15
  organizing
• 5:16
  intelligent
• 5:16
  systems
• 5:17
  in
• 5:17
  enterprise
• 5:17
  environments.
• 5:19
  Modeled
• 5:19
  explicitly
• 5:20
  on
• 5:20
  the
• 5:20
  OSI
• 5:20
  networking
• 5:20
  stack
• 5:21
  (with
• 5:21
  bottom-up
• 5:21
  numbering
• 5:22
  from
• 5:22
  Layer
• 5:22
  0
• 5:22
  Physical
• 5:23
  Compute
• 5:23
  through
• 5:23
  Layer
• 5:23
  9
• 5:24
  Human
• 5:24
  and
• 5:24
  Browser
• 5:24
  Interface),
• 5:25
  it
• 5:25
  addresses
• 5:26
  concerns
• 5:26
  that
• 5:26
  prior
• 5:27
  frameworks
• 5:27
  were
• 5:27
  not
• 5:27
  designed
• 5:28
  for:
• 5:28
  persistent
• 5:29
  memory
• 5:29
  across
• 5:29
  sessions,
• 5:30
  autonomous
• 5:31
  action
• 5:31
  exceeding
• 5:32
  explicit
• 5:32
  instruction,
• 5:33
  and
• 5:34
  consequential
• 5:34
  outputs
• 5:35
  requiring
• 5:35
  auditable
• 5:35
  judgment.
• 5:37
  Two
• 5:37
  paired
• 5:37
  state-holding
• 5:38
  layers
• 5:38
  bracket
• 5:38
  the
• 5:38
  operational
• 5:39
  core:
• 5:40
  Layer
• 5:40
  3
• 5:40
  (Agent
• 5:41
  Data
• 5:41
  Substrate)
• 5:41
  holds
• 5:42
  the
• 5:42
  data
• 5:42
  systems
• 5:42
  operate
• 5:43
  on,
• 5:44
  while
• 5:44
  Layer
• 5:44
  8
• 5:45
  (Continuity
• 5:45
  Fabric)
• 5:45
  maintains
• 5:46
  the
• 5:46
  cognitive
• 5:46
  state
• 5:46
  systems
• 5:47
  become
• 5:47
  --
• 5:48
  memory,
• 5:48
  learned
• 5:49
  adaptations,
• 5:49
  and
• 5:50
  verifiable
• 5:50
  decision
• 5:51
  trails.
• 5:52
  Six
• 5:52
  cross-layer
• 5:53
  spans
• 5:53
  (security,
• 5:54
  sovereignty,
• 5:54
  auditability,
• 5:55
  identity,
• 5:56
  energy,
• 5:56
  provenance)
• 5:57
  cut
• 5:57
  horizontally
• 5:57
  across
• 5:58
  all
• 5:58
  The
• 5:59
  specification
• 5:59
  is
• 5:59
  in
• 5:59
  active
• 6:00
  public
• 6:00
  review
• 6:00
  with
• 6:00
  eleven
• 6:01
  numbered
• 6:01
  decisions
• 6:02
  open
• 6:02
  for
• 6:02
  challenge,
• 6:03
  and
• 6:03
  the
• 6:03
  Foundation
• 6:04
  explicitly
• 6:04
  invites
• 6:04
  pushback
• 6:05
  by
• 6:05
  decision
• 6:05
  number
• 6:06
  rather
• 6:06
  than
• 6:06
  scattered
• 6:07
  marginalia.
• 6:08
  The
• 6:08
  timing
• 6:08
  is
• 6:08
  not
• 6:08
  accidental.
• 6:09
  The
• 6:10
  OIA
• 6:10
  document
• 6:10
  cites
• 6:11
  three
• 6:11
  developments
• 6:11
  that
• 6:11
  made
• 6:12
  a
• 6:12
  reference
• 6:12
  architecture
• 6:12
  operationally
• 6:13
  urgent:
• 6:14
  Anthropic's
• 6:15
  Mythos
• 6:15
  system
• 6:15
  card
• 6:16
  signaling
• 6:16
  that
• 6:16
  offensive
• 6:17
  capability
• 6:17
  has
• 6:17
  outpaced
• 6:18
  defensive
• 6:18
  readiness,
• 6:19
  agentic
• 6:19
  deployment
• 6:20
  across
• 6:20
  regulated
• 6:20
  sectors
• 6:21
  without
• 6:21
  shared
• 6:21
  vocabulary
• 6:22
  producing
• 6:22
  architectural
• 6:23
  fragmentation,
• 6:24
  and
• 6:24
  disclosures
• 6:25
  of
• 6:25
  unanticipated
• 6:25
  model
• 6:26
  behaviors
• 6:26
  including
• 6:27
  strategic
• 6:27
  concealment.
• 6:28
  Whether
• 6:28
  the
• 6:28
  OIA
• 6:29
  Model
• 6:29
  gains
• 6:29
  adoption
• 6:30
  or
• 6:30
  becomes
• 6:31
  another
• 6:31
  aspirational
• 6:32
  framework
• 6:32
  remains
• 6:32
  to
• 6:33
  be
• 6:33
  seen,
• 6:33
  but
• 6:33
  the
• 6:33
  gap
• 6:34
  it
• 6:34
  names
• 6:34
  is
• 6:34
  real
• 6:34
  --
• 6:35
  ask
• 6:36
  any
• 6:36
  enterprise
• 6:36
  architect
• 6:37
  deploying
• 6:37
  production
• 6:38
  AI
• 6:38
  what
• 6:38
  reference
• 6:39
  model
• 6:39
  they
• 6:39
  use,
• 6:40
  and
• 6:40
  you
• 6:40
  will
• 6:40
  get
• 6:40
  a
• 6:40
  different
• 6:41
  answer
• 6:41
  every
• 6:41
  time.
• 6:42
  A
• 6:42
  research
• 6:43
  paper
• 6:43
  from
• 6:43
  Sant'Anna
• 6:43
  School
• 6:44
  of
• 6:44
  Advanced
• 6:44
  Studies
• 6:44
  provides
• 6:45
  the
• 6:45
  theoretical
• 6:45
  complement.
• 6:46
  "Agentic
• 6:47
  Microphysics"
• 6:47
  proposes
• 6:48
  that
• 6:48
  safety
• 6:49
  analysis
• 6:49
  must
• 6:49
  shift
• 6:50
  from
• 6:50
  individual
• 6:50
  model
• 6:50
  evaluation
• 6:51
  to
• 6:51
  population-level
• 6:52
  dynamics
• 6:52
  --
• 6:53
  what
• 6:53
  happens
• 6:54
  when
• 6:54
  one
• 6:54
  agent's
• 6:54
  output
• 6:55
  becomes
• 6:55
  another
• 6:55
  agent's
• 6:56
  input
• 6:56
  under
• 6:56
  specific
• 6:57
  protocol
• 6:57
  conditions.
• 6:59
  The
• 6:59
  core
• 6:59
  argument:
• 7:00
  individually
• 7:00
  aligned
• 7:01
  agents
• 7:01
  can
• 7:01
  participate
• 7:02
  in
• 7:02
  emergent
• 7:02
  information
• 7:03
  cascades,
• 7:03
  tacit
• 7:04
  collusion,
• 7:04
  and
• 7:04
  collective
• 7:05
  deception.
• 7:06
  The
• 7:06
  paper
• 7:07
  introduces
• 7:07
  a
• 7:07
  generative
• 7:08
  methodology
• 7:08
  borrowed
• 7:09
  from
• 7:09
  computational
• 7:09
  social
• 7:10
  science
• 7:10
  (Epstein
• 7:11
  and
• 7:11
  Axtell's
• 7:12
  "growing
• 7:12
  artificial
• 7:13
  societies"
• 7:13
  tradition)
• 7:14
  and
• 7:14
  applies
• 7:15
  it
• 7:15
  to
• 7:15
  a
• 7:15
  controlled
• 7:15
  experiment
• 7:16
  where
• 7:16
  LLM
• 7:17
  agents
• 7:17
  interact
• 7:18
  through
• 7:18
  a
• 7:18
  shared
• 7:18
  news
• 7:19
  The
• 7:19
  central
• 7:20
  empirical
• 7:20
  finding
• 7:21
  is
• 7:21
  that
• 7:21
  feed
• 7:21
  position,
• 7:22
  not
• 7:22
  visible
• 7:22
  social
• 7:23
  proof,
• 7:23
  governs
• 7:24
  collective
• 7:24
  attention
• 7:24
  --
• 7:25
  agents
• 7:26
  select
• 7:26
  almost
• 7:27
  exclusively
• 7:27
  from
• 7:27
  top-ranked
• 7:28
  items
• 7:28
  regardless
• 7:29
  of
• 7:29
  endorsement
• 7:29
  counts.
• 7:30
  The
• 7:30
  safety
• 7:30
  implication:
• 7:32
  an
• 7:32
  adversary
• 7:32
  manipulating
• 7:33
  ranking
• 7:33
  can
• 7:33
  induce
• 7:34
  synchronized
• 7:34
  attention
• 7:34
  shifts
• 7:35
  across
• 7:35
  heterogeneous
• 7:36
  agent
• 7:36
  populations,
• 7:38
  a
• 7:38
  common
• 7:38
  architectural
• 7:38
  failure
• 7:39
  mode
• 7:39
  the
• 7:39
  authors
• 7:39
  describe
• 7:40
  as
• 7:40
  exploitable
• 7:41
  by
• 7:41
  design.
• 7:42
  On
• 7:42
  the
• 7:42
  practitioner
• 7:43
  side,
• 7:44
  Rohit
• 7:44
  Sharma
• 7:44
  articulated
• 7:45
  a
• 7:45
  problem
• 7:45
  that
• 7:45
  anyone
• 7:46
  running
• 7:46
  agentic
• 7:46
  workloads
• 7:47
  on
• 7:47
  hosted
• 7:47
  APIs
• 7:48
  recognizes:
• 7:49
  the
• 7:49
  "unprovability
• 7:50
  gap"
• 7:51
  between
• 7:51
  knowing
• 7:51
  model
• 7:51
  quality
• 7:52
  degraded
• 7:52
  and
• 7:52
  being
• 7:53
  able
• 7:53
  to
• 7:53
  prove
• 7:53
  it.
• 7:54
  Anthropic
• 7:55
  published
• 7:55
  a
• 7:55
  postmortem
• 7:56
  disclosing
• 7:56
  three
• 7:56
  product-layer
• 7:57
  changes
• 7:57
  that
• 7:58
  caused
• 7:58
  Claude
• 7:58
  Code
• 7:58
  degradation
• 7:59
  --
• 8:00
  a
• 8:00
  lower
• 8:00
  default
• 8:00
  reasoning
• 8:01
  effort,
• 8:01
  a
• 8:01
  caching
• 8:02
  bug
• 8:02
  clearing
• 8:02
  thinking
• 8:03
  history,
• 8:03
  and
• 8:04
  a
• 8:04
  system
• 8:04
  prompt
• 8:04
  capping
• 8:05
  inter-tool-call
• 8:05
  text.
• 8:06
  Sharma
• 8:07
  notes
• 8:07
  that
• 8:07
  the
• 8:07
  dimensions
• 8:08
  that
• 8:08
  matter
• 8:08
  (reasoning
• 8:08
  depth,
• 8:09
  initiative,
• 8:10
  willingness
• 8:10
  to
• 8:10
  push
• 8:10
  back)
• 8:11
  are
• 8:11
  subjective
• 8:11
  and
• 8:11
  high-dimensional,
• 8:12
  with
• 8:13
  no
• 8:13
  eval
• 8:13
  suite
• 8:14
  that
• 8:14
  captures
• 8:14
  "the
• 8:14
  model
• 8:15
  stopped
• 8:15
  challenging
• 8:15
  my
• 8:16
  bad
• 8:16
  idea."
• 8:17
  Every
• 8:17
  agentic
• 8:18
  system
• 8:18
  on
• 8:18
  a
• 8:18
  hosted
• 8:19
  API
• 8:19
  carries
• 8:19
  a
• 8:19
  silent
• 8:20
  dependency
• 8:20
  on
• 8:20
  quality
• 8:21
  consistency
• 8:22
  that
• 8:22
  no
• 8:22
  SLA
• 8:22
  currently
• 8:23
  covers.
• 8:24
  Irene
• 8:24
  Abezgauz
• 8:24
  offered
• 8:25
  a
• 8:25
  practical
• 8:25
  framework
• 8:26
  for
• 8:26
  the
• 8:26
  organizational
• 8:27
  side
• 8:27
  of
• 8:27
  this
• 8:27
  transition:
• 8:28
  a
• 8:28
  20-30-50
• 8:31
  split
• 8:32
  for
• 8:32
  PM
• 8:32
  teams
• 8:32
  in
• 8:32
  the
• 8:33
  hybrid
• 8:33
  AI
• 8:33
  era.
• 8:34
  Twenty
• 8:34
  percent
• 8:35
  of
• 8:35
  capacity
• 8:35
  goes
• 8:36
  to
• 8:36
  AI
• 8:36
  tools
• 8:36
  setup
• 8:37
  and
• 8:37
  maintenance
• 8:37
  (prompts,
• 8:38
  evals,
• 8:39
  workflows,
• 8:40
  integrations),
• 8:41
  thirty
• 8:41
  percent
• 8:41
  to
• 8:42
  agentic
• 8:42
  workforce
• 8:42
  management
• 8:43
  (reviewing
• 8:43
  and
• 8:43
  synthesizing
• 8:44
  agent
• 8:44
  output
• 8:45
  --
• 8:46
  "closer
• 8:46
  to
• 8:46
  leading
• 8:47
  a
• 8:47
  team
• 8:47
  than
• 8:47
  to
• 8:47
  IC
• 8:47
  work"),
• 8:48
  and
• 8:48
  fifty
• 8:48
  percent
• 8:49
  to
• 8:49
  core
• 8:49
  PM
• 8:49
  work
• 8:50
  that
• 8:50
  stays
• 8:50
  human:
• 8:51
  strategy,
• 8:52
  prioritization,
• 8:53
  stakeholder
• 8:53
  alignment,
• 8:54
  customer
• 8:54
  conversations.
• 8:56
  The
• 8:56
  key
• 8:56
  insight:
• 8:56
  agents
• 8:57
  reduce
• 8:57
  mechanical
• 8:58
  work
• 8:58
  linearly
• 8:58
  but
• 8:59
  judgment
• 8:59
  work
• 8:59
  stays
• 8:59
  the
• 9:00
  same
• 9:00
  or
• 9:00
  increases,
• 9:01
  and
• 9:01
  cutting
• 9:01
  headcount
• 9:02
  proportional
• 9:03
  to
• 9:03
  total
• 9:03
  time
• 9:03
  saved
• 9:04
  causes
• 9:04
  strategy
• 9:04
  degradation
• 9:05
  that
• 9:05
  only
• 9:06
  surfaces
• 9:06
  in
• 9:06
  outcomes
• 9:07
  months
• 9:10
  Salvatore
• 9:10
  Sanfilippo
• 9:11
  (antirez,
• 9:12
  of
• 9:12
  Redis
• 9:12
  fame)
• 9:12
  shipped
• 9:13
  experimental
• 9:13
  llama.cpp
• 9:14
  support
• 9:15
  for
• 9:15
  DeepSeek
• 9:15
  v4
• 9:16
  Flash,
• 9:16
  with
• 9:16
  a
• 9:16
  GGUF
• 9:17
  quantization
• 9:18
  that
• 9:18
  fits
• 9:18
  in
• 9:18
  128
• 9:19
  GB
• 9:19
  of
• 9:19
  RAM.
• 9:20
  The
• 9:20
  approach
• 9:20
  is
• 9:20
  surgical:
• 9:21
  routed
• 9:21
  experts
• 9:22
  are
• 9:22
  quantized
• 9:22
  to
• 9:22
  2
• 9:22
  bits
• 9:23
  using
• 9:23
  two
• 9:23
  different
• 9:23
  2-bit
• 9:24
  quant
• 9:24
  schemes
• 9:24
  to
• 9:24
  balance
• 9:25
  error
• 9:25
  and
• 9:25
  size,
• 9:26
  while
• 9:26
  shared
• 9:26
  experts
• 9:26
  and
• 9:26
  all
• 9:26
  non-expert
• 9:27
  weights
• 9:27
  stay
• 9:27
  at
• 9:27
  Q8.
• 9:28
  The
• 9:28
  result
• 9:29
  runs
• 9:29
  at
• 9:29
  17-21
• 9:30
  tok/s
• 9:31
  on
• 9:31
  a
• 9:31
  MacBook
• 9:31
  M3
• 9:31
  Max
• 9:31
  --
• 9:32
  firmly
• 9:32
  in
• 9:33
  the
• 9:33
  "usable"
• 9:33
  zone
• 9:33
  for
• 9:33
  a
• 9:33
  model
• 9:34
  that,
• 9:34
  even
• 9:34
  at
• 9:34
  2-bit
• 9:35
  quantization,
• 9:36
  feels
• 9:36
  qualitatively
• 9:36
  stronger
• 9:37
  than
• 9:37
  Qwen
• 9:37
  3.6
• 9:38
  27B
• 9:38
  in
• 9:39
  subjective
• 9:39
  evaluation.
• 9:40
  Community
• 9:40
  response
• 9:41
  was
• 9:41
  immediate:
• 9:42
  one
• 9:42
  contributor
• 9:43
  got
• 9:43
  it
• 9:43
  running
• 9:43
  on
• 9:43
  an
• 9:43
  RTX
• 9:44
  6000
• 9:44
  96GB
• 9:45
  with
• 9:45
  CUDA,
• 9:46
  and
• 9:46
  a
• 9:46
  ROCm
• 9:47
  port
• 9:47
  appeared
• 9:47
  for
• 9:47
  AMD
• 9:48
  hardware,
• 9:49
  though
• 9:49
  prefill
• 9:49
  performance
• 9:50
  lagged
• 9:50
  at
• 9:50
  around
• 9:50
  110
• 9:51
  tok/s
• 9:52
  with
• 9:52
  output
• 9:52
  at
• 9:52
  8
• 9:53
  tok/s.
• 9:54
  The
• 9:54
  bottleneck
• 9:55
  is
• 9:55
  graph
• 9:55
  splits
• 9:56
  --
• 9:56
  856
• 9:57
  at
• 9:57
  batch
• 9:58
  size
• 9:58
  512
• 9:59
  versus
• 9:59
  the
• 9:59
  fewer-than-10
• 10:00
  typical
• 10:00
  of
• 10:00
  pure
• 10:00
  transformer
• 10:01
  models
• 10:01
  --
• 10:02
  causing
• 10:02
  massive
• 10:03
  CPU-GPU
• 10:05
  synchronization
• 10:05
  overhead.
• 10:06
  On
• 10:06
  the
• 10:06
  hardware
• 10:07
  acceleration
• 10:07
  front,
• 10:07
  native
• 10:08
  NVFP4
• 10:09
  support
• 10:10
  in
• 10:10
  llama.cpp
• 10:11
  build
• 10:11
  b8967
• 10:13
  delivered
• 10:13
  43-68%
• 10:15
  faster
• 10:16
  prompt
• 10:16
  processing
• 10:17
  on
• 10:17
  an
• 10:17
  RTX
• 10:17
  5090
• 10:18
  running
• 10:18
  Qwen3.6-27B-NVFP4,
• 10:23
  with
• 10:23
  the
• 10:23
  largest
• 10:24
  gains
• 10:24
  at
• 10:24
  shorter
• 10:24
  context
• 10:25
  (1.7x
• 10:26
  at
• 10:26
  pp512)
• 10:27
  tapering
• 10:27
  to
• 10:28
  1.43x
• 10:29
  at
• 10:29
  32K
• 10:29
  context.
• 10:30
  Token
• 10:31
  generation
• 10:31
  speed
• 10:31
  was
• 10:32
  unchanged
• 10:32
  --
• 10:33
  the
• 10:33
  improvement
• 10:33
  is
• 10:33
  purely
• 10:34
  in
• 10:34
  the
• 10:34
  prefill
• 10:34
  path,
• 10:35
  meaning
• 10:35
  faster
• 10:36
  time-to-first-token
• 10:36
  for
• 10:37
  RAG
• 10:37
  workloads,
• 10:38
  document
• 10:38
  analysis,
• 10:39
  and
• 10:39
  code-heavy
• 10:39
  prompts.
• 10:40
  The
• 10:40
  accuracy
• 10:41
  tradeoff
• 10:41
  remains
• 10:42
  real:
• 10:43
  NVFP4
• 10:45
  quants
• 10:45
  show
• 10:45
  measurably
• 10:46
  higher
• 10:46
  perplexity
• 10:47
  than
• 10:47
  imatrix
• 10:47
  quants
• 10:48
  at
• 10:48
  the
• 10:48
  same
• 10:48
  VRAM
• 10:49
  budget,
• 10:49
  and
• 10:49
  Nvidia's
• 10:50
  quantization-aware
• 10:51
  distillation
• 10:51
  models
• 10:52
  that
• 10:52
  would
• 10:52
  close
• 10:52
  that
• 10:52
  gap
• 10:53
  exist
• 10:53
  for
• 10:53
  only
• 10:53
  a
• 10:53
  handful
• 10:54
  of
• 10:54
  their
• 10:54
  own
• 10:54
  models.
• 10:55
  Speculative
• 10:56
  decoding
• 10:56
  continues
• 10:57
  to
• 10:57
  find
• 10:57
  its
• 10:57
  sweet
• 10:57
  spot
• 10:58
  in
• 10:58
  constrained-output
• 10:59
  workflows.
• 11:00
  One
• 11:00
  developer
• 11:01
  replaced
• 11:01
  Gemini
• 11:01
  2.5
• 11:02
  Flash-Lite
• 11:03
  API
• 11:03
  calls
• 11:03
  with
• 11:04
  local
• 11:04
  Gemma-4-31B
• 11:05
  (Q6
• 11:06
  K
• 11:06
  L)
• 11:06
  paired
• 11:07
  with
• 11:07
  Gemma-4-E2B
• 11:07
  (Q8
• 11:08
  0)
• 11:09
  as
• 11:09
  a
• 11:09
  draft
• 11:09
  model,
• 11:10
  achieving
• 11:10
  120-200
• 11:11
  tok/s
• 11:12
  on
• 11:13
  an
• 11:13
  RTX
• 11:13
  5090
• 11:14
  for
• 11:14
  structured
• 11:15
  JSON
• 11:15
  extraction
• 11:15
  tasks
• 11:16
  in
• 11:16
  Lithuanian.
• 11:17
  The
• 11:18
  acceptance
• 11:18
  rate
• 11:18
  for
• 11:18
  speculative
• 11:19
  tokens
• 11:19
  jumps
• 11:19
  to
• 11:20
  0.7+
• 11:21
  on
• 11:21
  strict
• 11:21
  schema
• 11:22
  extraction
• 11:22
  versus
• 11:23
  0.3-0.4
• 11:24
  on
• 11:24
  open-ended
• 11:25
  generation,
• 11:26
  making
• 11:26
  the
• 11:26
  headline
• 11:27
  speed
• 11:27
  achievable
• 11:28
  specifically
• 11:28
  for
• 11:28
  the
• 11:28
  narrow-output-space
• 11:29
  workloads
• 11:30
  where
• 11:30
  local
• 11:30
  inference
• 11:31
  most
• 11:31
  directly
• 11:31
  competes
• 11:32
  with
• 11:32
  API
• 11:32
  calls.
• 11:33
  The
• 11:33
  setup
• 11:34
  uses
• 11:34
  31.5
• 11:35
  GB
• 11:35
  VRAM.
• 11:36
  A
• 11:37
  community
• 11:37
  benchmark
• 11:37
  of
• 11:38
  the
• 11:38
  4B
• 11:38
  model
• 11:38
  class
• 11:39
  revealed
• 11:39
  pronounced
• 11:40
  "lab
• 11:40
  personalities"
• 11:40
  at
• 11:41
  small
• 11:41
  scale.
• 11:42
  NVIDIA's
• 11:42
  Nemotron
• 11:43
  3
• 11:43
  Nano
• 11:43
  (2.8
• 11:44
  GB
• 11:44
  on
• 11:45
  disk)
• 11:45
  scored
• 11:45
  85%
• 11:46
  overall
• 11:46
  with
• 11:47
  a
• 11:47
  perfect
• 11:47
  100%
• 11:48
  on
• 11:48
  finance
• 11:48
  tasks,
• 11:49
  while
• 11:49
  IBM's
• 11:50
  Granite4
• 11:50
  3B
• 11:50
  hit
• 11:51
  100%
• 11:51
  on
• 11:51
  code
• 11:52
  but
• 11:52
  only
• 11:52
  20%
• 11:53
  on
• 11:53
  reasoning
• 11:53
  --
• 11:54
  near-mirror
• 11:55
  profiles
• 11:55
  from
• 11:55
  labs
• 11:56
  that
• 11:56
  market
• 11:56
  both
• 11:56
  as
• 11:56
  general-purpose.
• 11:58
  Microsoft's
• 11:58
  phi4-mini
• 11:59
  was
• 11:59
  the
• 11:59
  most
• 12:00
  balanced
• 12:00
  and
• 12:00
  the
• 12:00
  bang-for-GB
• 12:01
  winner
• 12:01
  at
• 12:01
  30.8
• 12:02
  accuracy-percent
• 12:03
  per
• 12:03
  GB.
• 12:04
  Qwen
• 12:04
  3.5
• 12:05
  4B
• 12:05
  scored
• 12:06
  15%,
• 12:07
  but
• 12:07
  the
• 12:07
  methodology
• 12:08
  acknowledged
• 12:09
  this
• 12:09
  was
• 12:09
  a
• 12:09
  thinking-model-in-fixed-budget
• 12:10
  problem:
• 12:11
  the
• 12:11
  1024-token
• 12:12
  cap
• 12:13
  is
• 12:13
  insufficient
• 12:14
  for
• 12:14
  models
• 12:14
  that
• 12:14
  need
• 12:14
  4096+
• 12:16
  tokens
• 12:16
  to
• 12:16
  finish
• 12:16
  their
• 12:17
  chain
• 12:17
  of
• 12:17
  thought.
• 12:18
  The
• 12:18
  eval
• 12:18
  ecosystem
• 12:19
  has
• 12:19
  a
• 12:19
  structural
• 12:20
  challenge
• 12:20
  here
• 12:20
  that
• 12:21
  per-model
• 12:21
  token
• 12:21
  budgets
• 12:22
  would
• 12:22
  partially
• 12:22
  address.
• 12:24
  A
• 12:24
  separate
• 12:24
  study
• 12:24
  demonstrated
• 12:25
  that
• 12:25
  task-specific
• 12:26
  LoRA
• 12:26
  adapters
• 12:27
  can
• 12:27
  double
• 12:27
  coding
• 12:28
  performance
• 12:28
  of
• 12:28
  a
• 12:28
  7B
• 12:29
  model
• 12:29
  without
• 12:30
  modifying
• 12:30
  the
• 12:30
  coding
• 12:31
  agent
• 12:31
  itself.
• 12:32
  Four
• 12:32
  LoRA
• 12:33
  variants
• 12:33
  of
• 12:33
  Qwen2.5-7B
• 12:35
  (adding
• 12:36
  just
• 12:37
  3%
• 12:37
  parameters),
• 12:38
  built
• 12:38
  from
• 12:38
  the
• 12:38
  successes
• 12:39
  of
• 12:39
  a
• 12:39
  larger
• 12:39
  model,
• 12:40
  showed
• 12:40
  that
• 12:40
  enabling
• 12:41
  tweaked
• 12:41
  retrieval
• 12:42
  and
• 12:42
  planning
• 12:42
  agents
• 12:42
  raised
• 12:43
  test
• 12:43
  passes
• 12:43
  from
• 12:43
  3
• 12:44
  to
• 12:44
  10
• 12:44
  --
• 12:45
  despite
• 12:45
  neither
• 12:46
  the
• 12:46
  coder
• 12:46
  nor
• 12:46
  debugger
• 12:47
  agent
• 12:47
  being
• 12:47
  The
• 12:49
  tweaked
• 12:49
  debugger
• 12:49
  technically
• 12:50
  outperformed
• 12:50
  the
• 12:50
  tweaked
• 12:51
  coder,
• 12:52
  suggesting
• 12:52
  the
• 12:52
  coding
• 12:52
  agent
• 12:53
  may
• 12:53
  have
• 12:53
  the
• 12:53
  lowest
• 12:54
  intelligence
• 12:54
  requirement
• 12:55
  in
• 12:55
  the
• 12:55
  pipeline.
• 12:56
  At
• 12:56
  the
• 12:56
  far
• 12:57
  edge
• 12:57
  of
• 12:57
  the
• 12:57
  spectrum,
• 12:58
  RuVLLM
• 12:59
  ESP32
• 13:01
  v0.3.0-rc2
• 13:04
  shipped
• 13:04
  actual
• 13:04
  firmware
• 13:05
  for
• 13:05
  five
• 13:05
  ESP32
• 13:07
  variants,
• 13:07
  turning
• 13:07
  $2-4
• 13:08
  dev
• 13:08
  boards
• 13:09
  into
• 13:09
  specialist
• 13:09
  nodes
• 13:10
  in
• 13:10
  a
• 13:10
  small
• 13:10
  AI
• 13:10
  cluster.
• 13:12
  Each
• 13:12
  chip
• 13:12
  runs
• 13:12
  a
• 13:12
  single
• 13:13
  Rust
• 13:13
  program
• 13:13
  (~430
• 13:14
  KB
• 13:15
  flash)
• 13:15
  that
• 13:15
  boots
• 13:15
  into
• 13:16
  a
• 13:16
  role
• 13:16
  --
• 13:17
  HNSW
• 13:17
  vector
• 13:17
  indexer,
• 13:18
  RAG
• 13:18
  retriever,
• 13:19
  anomaly
• 13:19
  sentinel,
• 13:20
  semantic
• 13:20
  memory
• 13:21
  archivist,
• 13:21
  or
• 13:21
  micro-LoRA
• 13:22
  adapter
• 13:22
  --
• 13:23
  and
• 13:23
  handles
• 13:23
  commands
• 13:24
  over
• 13:24
  USB-Serial
• 13:25
  at
• 13:25
  115200
• 13:26
  baud.
• 13:27
  No
• 13:27
  cloud,
• 13:28
  no
• 13:28
  API
• 13:28
  key,
• 13:29
  no
• 13:29
  internet
• 13:29
  required.
• 13:30
  This
• 13:30
  is
• 13:31
  explicitly
• 13:31
  not
• 13:31
  transformer
• 13:32
  inference
• 13:32
  crammed
• 13:32
  into
• 13:33
  4
• 13:33
  KB
• 13:33
  of
• 13:33
  SRAM
• 13:33
  (the
• 13:33
  previous
• 13:34
  release
• 13:34
  claimed
• 13:34
  that
• 13:35
  and
• 13:35
  it
• 13:35
  did
• 13:35
  not
• 13:35
  work);
• 13:36
  it
• 13:36
  is
• 13:36
  the
• 13:36
  federation-ready
• 13:37
  primitive
• 13:37
  layer
• 13:37
  --
• 13:38
  vector
• 13:38
  search,
• 13:38
  retrieval,
• 13:39
  memory,
• 13:39
  anomaly
• 13:40
  detection
• 13:40
  --
• 13:41
  that
• 13:41
  a
• 13:41
  future
• 13:41
  ESP32-P4
• 13:43
  with
• 13:44
  8
• 13:44
  MB
• 13:44
  PSRAM
• 13:45
  could
• 13:45
  join
• 13:45
  as
• 13:45
  a
• 13:45
  draft-token
• 13:46
  node.
• 13:46
  Zyphra
• 13:47
  also
• 13:47
  dropped
• 13:47
  ZUNA
• 13:47
  on
• 13:48
  Hugging
• 13:48
  Face,
• 13:48
  though
• 13:48
  details
• 13:49
  remain
• 13:49
  sparse
• 13:49
  at
• 13:49
  launch.
• 13:53
  Warp,
• 13:53
  the
• 13:53
  Rust-based
• 13:54
  terminal
• 13:54
  that
• 13:54
  launched
• 13:54
  five
• 13:55
  years
• 13:55
  ago
• 13:55
  with
• 13:55
  a
• 13:55
  promise
• 13:56
  to
• 13:56
  eventually
• 13:56
  open-source,
• 13:57
  finally
• 13:57
  delivered.
• 13:58
  The
• 13:58
  client
• 13:58
  is
• 13:58
  now
• 13:59
  available
• 13:59
  under
• 13:59
  AGPL,
• 14:00
  with
• 14:00
  OpenAI
• 14:01
  as
• 14:01
  founding
• 14:02
  sponsor
• 14:02
  and
• 14:02
  their
• 14:02
  agent
• 14:03
  orchestration
• 14:03
  platform
• 14:04
  Oz
• 14:04
  powering
• 14:04
  the
• 14:04
  contribution
• 14:05
  workflow.
• 14:06
  The
• 14:06
  strategic
• 14:06
  logic
• 14:07
  is
• 14:07
  Warp's
• 14:08
  team
• 14:08
  believes
• 14:09
  the
• 14:09
  biggest
• 14:09
  bottleneck
• 14:09
  to
• 14:09
  development
• 14:10
  is
• 14:10
  no
• 14:10
  longer
• 14:11
  writing
• 14:11
  code
• 14:11
  but
• 14:11
  the
• 14:11
  human-in-the-loop
• 14:12
  activities
• 14:12
  around
• 14:13
  it
• 14:13
  --
• 14:13
  speccing
• 14:14
  product
• 14:14
  and
• 14:14
  verifying
• 14:15
  behavior
• 14:15
  --
• 14:16
  and
• 14:16
  that
• 14:16
  opening
• 14:16
  the
• 14:16
  codebase
• 14:17
  to
• 14:17
  community
• 14:17
  contributors
• 14:18
  who
• 14:18
  supervise
• 14:18
  agents
• 14:19
  will
• 14:19
  accelerate
• 14:20
  Alongside
• 14:21
  the
• 14:21
  open-sourcing,
• 14:21
  Warp
• 14:22
  added
• 14:22
  support
• 14:22
  for
• 14:22
  open-source
• 14:23
  models
• 14:23
  (Kimi,
• 14:24
  MiniMax,
• 14:25
  Qwen)
• 14:25
  with
• 14:25
  an
• 14:25
  "auto
• 14:26
  (open)"
• 14:26
  model-routed
• 14:27
  option,
• 14:28
  a
• 14:28
  settings
• 14:28
  file
• 14:28
  for
• 14:29
  programmatic
• 14:29
  control,
• 14:30
  and
• 14:30
  customizable
• 14:31
  UI
• 14:31
  density
• 14:31
  ranging
• 14:32
  from
• 14:32
  a
• 14:32
  minimal
• 14:32
  terminal
• 14:33
  to
• 14:33
  a
• 14:33
  full
• 14:33
  agentic
• 14:34
  development
• 14:34
  environment.
• 14:35
  Public
• 14:35
  GitHub
• 14:36
  issues
• 14:36
  become
• 14:36
  the
• 14:36
  source
• 14:37
  of
• 14:37
  truth
• 14:37
  for
• 14:37
  the
• 14:37
  roadmap.
• 14:38
  The
• 14:39
  AGPL
• 14:39
  license
• 14:40
  means
• 14:40
  derivative
• 14:40
  works
• 14:41
  must
• 14:41
  also
• 14:42
  be
• 14:42
  open-source,
• 14:43
  which
• 14:43
  positions
• 14:43
  Warp
• 14:44
  as
• 14:44
  a
• 14:44
  community
• 14:44
  platform
• 14:45
  rather
• 14:45
  than
• 14:45
  a
• 14:45
  fork-and-commercialize
• 14:46
  target.
• 14:47
  Browser
• 14:48
  Use
• 14:48
  shipped
• 14:48
  BUX,
• 14:49
  a
• 14:49
  project
• 14:50
  that
• 14:50
  turns
• 14:50
  any
• 14:50
  $5
• 14:51
  VPS
• 14:51
  into
• 14:52
  a
• 14:52
  24/7
• 14:53
  Claude
• 14:53
  Code
• 14:53
  agent
• 14:54
  with
• 14:54
  a
• 14:54
  persistent
• 14:55
  Chromium
• 14:55
  session.
• 14:56
  Three
• 14:56
  systemd
• 14:57
  services
• 14:57
  --
• 14:58
  Claude
• 14:58
  Code,
• 14:59
  a
• 14:59
  browser
• 14:59
  harness
• 15:00
  via
• 15:00
  browser-use's
• 15:00
  cloud
• 15:01
  CDK,
• 15:02
  and
• 15:02
  a
• 15:02
  web
• 15:02
  terminal
• 15:02
  --
• 15:03
  boot
• 15:03
  from
• 15:04
  a
• 15:04
  single
• 15:04
  curl
• 15:04
  install
• 15:05
  script.
• 15:06
  A
• 15:06
  Telegram
• 15:06
  bot
• 15:06
  integration
• 15:07
  lets
• 15:07
  users
• 15:08
  text
• 15:08
  their
• 15:08
  agent
• 15:08
  from
• 15:09
  anywhere.
• 15:10
  When
• 15:10
  the
• 15:10
  agent
• 15:10
  hits
• 15:10
  a
• 15:10
  login
• 15:11
  wall,
• 15:11
  2FA,
• 15:12
  or
• 15:12
  CAPTCHA,
• 15:13
  it
• 15:13
  hands
• 15:13
  back
• 15:13
  a
• 15:13
  live
• 15:14
  view
• 15:14
  URL
• 15:15
  and
• 15:15
  waits
• 15:15
  rather
• 15:15
  than
• 15:16
  attempting
• 15:16
  credential
• 15:16
  stuffing.
• 15:18
  Agent
• 15:18
  state
• 15:18
  persists
• 15:19
  in
• 15:19
  /home/bux
• 15:21
  across
• 15:21
  reboots,
• 15:22
  preserving
• 15:22
  cookies,
• 15:23
  skills,
• 15:24
  and
• 15:24
  chat
• 15:24
  history.
• 15:25
  The
• 15:25
  managed
• 15:26
  offering
• 15:26
  provisions
• 15:27
  a
• 15:27
  box
• 15:27
  in
• 15:27
  about
• 15:28
  60
• 15:28
  seconds.
• 15:29
  Graphify
• 15:30
  takes
• 15:30
  a
• 15:30
  different
• 15:30
  approach
• 15:31
  to
• 15:31
  developer
• 15:31
  intelligence:
• 15:32
  it
• 15:33
  reads
• 15:33
  a
• 15:33
  codebase
• 15:33
  (code
• 15:34
  via
• 15:34
  tree-sitter
• 15:35
  AST
• 15:35
  parsing,
• 15:36
  docs
• 15:37
  and
• 15:37
  PDFs
• 15:37
  via
• 15:37
  Claude,
• 15:38
  images
• 15:39
  via
• 15:39
  Claude
• 15:39
  vision)
• 15:40
  and
• 15:40
  builds
• 15:40
  a
• 15:40
  queryable
• 15:41
  knowledge
• 15:41
  graph
• 15:42
  using
• 15:42
  NetworkX
• 15:43
  and
• 15:43
  Leiden
• 15:43
  community
• 15:44
  detection.
• 15:45
  The
• 15:45
  headline
• 15:45
  claim
• 15:45
  is
• 15:46
  71.5x
• 15:47
  fewer
• 15:47
  tokens
• 15:48
  per
• 15:48
  query
• 15:48
  versus
• 15:49
  reading
• 15:49
  raw
• 15:49
  files,
• 15:50
  benchmarked
• 15:50
  on
• 15:51
  a
• 15:51
  52-file
• 15:51
  mixed
• 15:52
  corpus
• 15:52
  of
• 15:52
  Karpathy
• 15:53
  repos,
• 15:53
  papers,
• 15:54
  and
• 15:54
  images.
• 15:55
  Every
• 15:55
  edge
• 15:56
  is
• 15:56
  tagged
• 15:56
  EXTRACTED,
• 15:57
  INFERRED,
• 15:58
  or
• 15:58
  AMBIGUOUS,
• 15:59
  and
• 15:59
  output
• 16:00
  includes
• 16:00
  interactive
• 16:01
  HTML
• 16:01
  visualization,
• 16:03
  Obsidian
• 16:03
  vault
• 16:03
  export,
• 16:05
  Wikipedia-style
• 16:05
  wiki
• 16:06
  articles
• 16:06
  for
• 16:06
  agent
• 16:07
  navigation,
• 16:07
  and
• 16:08
  Neo4j
• 16:08
  Cypher
• 16:09
  export.
• 16:10
  A
• 16:10
  post-commit
• 16:11
  git
• 16:11
  hook
• 16:11
  keeps
• 16:12
  the
• 16:12
  graph
• 16:12
  synchronized
• 16:13
  automatically.
• 16:14
  Microsoft
• 16:15
  open-sourced
• 16:16
  VibeVoice,
• 16:17
  a
• 16:17
  family
• 16:17
  of
• 16:17
  frontier
• 16:18
  voice
• 16:18
  AI
• 16:18
  models
• 16:19
  spanning
• 16:19
  TTS
• 16:20
  and
• 16:20
  ASR.
• 16:21
  The
• 16:21
  ASR
• 16:22
  model
• 16:22
  handles
• 16:22
  60-minute
• 16:23
  long-form
• 16:24
  audio
• 16:24
  in
• 16:24
  a
• 16:24
  single
• 16:25
  pass
• 16:25
  with
• 16:26
  structured
• 16:26
  output
• 16:27
  (who
• 16:27
  said
• 16:27
  what
• 16:27
  and
• 16:28
  when),
• 16:28
  customizable
• 16:29
  hotwords
• 16:30
  for
• 16:30
  domain-specific
• 16:31
  accuracy,
• 16:32
  and
• 16:32
  support
• 16:32
  for
• 16:32
  50+
• 16:33
  languages.
• 16:34
  The
• 16:34
  TTS
• 16:35
  model
• 16:35
  generates
• 16:36
  up
• 16:36
  to
• 16:36
  90
• 16:37
  minutes
• 16:37
  of
• 16:37
  multi-speaker
• 16:38
  conversational
• 16:39
  audio.
• 16:40
  The
• 16:40
  core
• 16:40
  innovation
• 16:41
  is
• 16:41
  continuous
• 16:41
  speech
• 16:42
  tokenizers
• 16:42
  operating
• 16:43
  at
• 16:43
  7.5
• 16:44
  Hz
• 16:44
  --
• 16:45
  an
• 16:45
  ultra-low
• 16:46
  frame
• 16:46
  rate
• 16:46
  that
• 16:47
  preserves
• 16:47
  fidelity
• 16:48
  while
• 16:48
  dramatically
• 16:48
  reducing
• 16:49
  sequence
• 16:49
  length
• 16:50
  for
• 16:50
  long-form
• 16:50
  processing.
• 16:51
  The
• 16:52
  TTS
• 16:52
  code
• 16:53
  was
• 16:53
  previously
• 16:53
  removed
• 16:54
  after
• 16:54
  misuse
• 16:55
  concerns;
• 16:56
  the
• 16:56
  ASR
• 16:56
  and
• 16:57
  real-time
• 16:57
  streaming
• 16:58
  models
• 16:58
  (0.5B
• 16:59
  parameters,
• 17:00
  ~300ms
• 17:01
  first-audible
• 17:02
  latency)
• 17:03
  remain
• 17:03
  available.
• 17:06
  Microsoft
• 17:07
  and
• 17:07
  OpenAI
• 17:08
  are
• 17:08
  ending
• 17:08
  their
• 17:08
  exclusive
• 17:09
  partnership
• 17:09
  and
• 17:10
  revenue-sharing
• 17:10
  arrangement,
• 17:11
  according
• 17:11
  to
• 17:11
  Bloomberg.
• 17:12
  The
• 17:12
  deal
• 17:12
  restructuring
• 17:13
  removes
• 17:13
  the
• 17:13
  exclusivity
• 17:14
  that
• 17:14
  bound
• 17:14
  OpenAI
• 17:15
  to
• 17:15
  Microsoft's
• 17:16
  cloud
• 17:16
  infrastructure
• 17:17
  and
• 17:17
  eliminates
• 17:17
  the
• 17:18
  revenue-sharing
• 17:18
  provisions
• 17:19
  that
• 17:19
  gave
• 17:19
  Microsoft
• 17:19
  a
• 17:20
  cut
• 17:20
  of
• 17:20
  OpenAI's
• 17:20
  commercial
• 17:21
  income.
• 17:22
  The
• 17:22
  move
• 17:22
  signals
• 17:23
  both
• 17:23
  companies
• 17:23
  positioning
• 17:24
  for
• 17:24
  independence
• 17:24
  --
• 17:25
  Microsoft
• 17:26
  diversifying
• 17:27
  its
• 17:27
  AI
• 17:27
  bets
• 17:27
  across
• 17:27
  multiple
• 17:28
  model
• 17:28
  providers,
• 17:29
  OpenAI
• 17:29
  seeking
• 17:30
  flexibility
• 17:30
  to
• 17:30
  work
• 17:31
  with
• 17:31
  other
• 17:31
  cloud
• 17:31
  and
• 17:31
  distribution
• 17:32
  partners.
• 17:33
  The
• 17:33
  full
• 17:33
  implications
• 17:34
  for
• 17:34
  Azure's
• 17:34
  AI
• 17:34
  strategy
• 17:35
  and
• 17:35
  OpenAI's
• 17:36
  path
• 17:36
  to
• 17:36
  profitability
• 17:37
  will
• 17:37
  take
• 17:37
  time
• 17:37
  to
• 17:37
  materialize,
• 17:38
  but
• 17:38
  the
• 17:38
  structural
• 17:39
  shift
• 17:39
  is
• 17:39
  significant:
• 17:40
  the
• 17:40
  most
• 17:40
  consequential
• 17:41
  AI
• 17:41
  partnership
• 17:42
  of
• 17:42
  the
• 17:42
  last
• 17:42
  five
• 17:42
  years
• 17:42
  is
• 17:43
  being
• 17:43
  unwound.
• 17:44
  A
• 17:44
  pointed
• 17:44
  video
• 17:44
  essay
• 17:45
  took
• 17:45
  aim
• 17:45
  at
• 17:45
  what
• 17:45
  it
• 17:45
  calls
• 17:46
  "token
• 17:46
  budgets"
• 17:46
  --
• 17:47
  the
• 17:47
  practice,
• 17:47
  reportedly
• 17:48
  in
• 17:48
  use
• 17:48
  at
• 17:48
  Facebook,
• 17:49
  of
• 17:49
  tracking
• 17:49
  employees'
• 17:50
  AI
• 17:50
  tool
• 17:50
  consumption
• 17:51
  via
• 17:51
  leaderboards.
• 17:52
  The
• 17:52
  the
• 17:53
  person
• 17:53
  at
• 17:54
  the
• 17:54
  top
• 17:54
  of
• 17:54
  the
• 17:54
  token
• 17:54
  leaderboard
• 17:55
  is
• 17:55
  reviewing
• 17:55
  zero
• 17:56
  code,
• 17:57
  the
• 17:57
  metric
• 17:57
  creates
• 17:57
  perverse
• 17:58
  incentives
• 17:58
  toward
• 17:59
  generating
• 17:59
  slop
• 17:59
  at
• 17:59
  volume,
• 18:00
  and
• 18:00
  Jensen
• 18:01
  Huang's
• 18:01
  framing
• 18:01
  of
• 18:02
  $250,000/year/employee
• 18:05
  in
• 18:05
  token
• 18:05
  spend
• 18:06
  as
• 18:06
  a
• 18:06
  productivity
• 18:06
  benchmark
• 18:07
  serves
• 18:07
  primarily
• 18:08
  to
• 18:08
  funnel
• 18:08
  corporate
• 18:08
  savings
• 18:09
  toward
• 18:09
  GPU
• 18:09
  vendors
• 18:10
  and
• 18:10
  AI
• 18:10
  providers.
• 18:11
  The
• 18:11
  broader
• 18:12
  argument
• 18:12
  --
• 18:13
  the
• 18:13
  "Bittar
• 18:13
  lesson"
• 18:14
  --
• 18:14
  posits
• 18:15
  that
• 18:15
  AI
• 18:15
  usefulness
• 18:16
  is
• 18:16
  inversely
• 18:17
  proportional
• 18:17
  to
• 18:17
  the
• 18:17
  precision
• 18:18
  required:
• 18:19
  LLMs
• 18:20
  approximate
• 18:20
  language
• 18:20
  which
• 18:21
  approximates
• 18:21
  intent,
• 18:22
  and
• 18:22
  the
• 18:22
  compounding
• 18:23
  approximation
• 18:23
  hits
• 18:24
  a
• 18:24
  wall
• 18:24
  at
• 18:24
  roughly
• 18:25
  80%
• 18:25
  of
• 18:25
  any
• 18:25
  task,
• 18:26
  leaving
• 18:27
  the
• 18:27
  hard
• 18:27
  20%
• 18:28
  as
• 18:28
  irreducibly
• 18:29
  human
• 18:29
  work.
• 18:30
  Whether
• 18:30
  or
• 18:30
  not
• 18:30
  one
• 18:31
  accepts
• 18:31
  the
• 18:31
  full
• 18:31
  thesis,
• 18:32
  the
• 18:32
  observation
• 18:33
  that
• 18:33
  token
• 18:33
  consumption
• 18:34
  is
• 18:34
  becoming
• 18:34
  a
• 18:34
  proxy
• 18:35
  metric
• 18:35
  for
• 18:35
  productivity
• 18:36
  --
• 18:37
  with
• 18:37
  the
• 18:37
  same
• 18:37
  pathologies
• 18:38
  as
• 18:38
  lines-of-code
• 18:39
  metrics
• 18:39
  --
• 18:40
  deserves
• 18:40
  attention.
• 18:42
  Talkie,
• 18:42
  a
• 18:43
  13B
• 18:43
  language
• 18:44
  model
• 18:44
  trained
• 18:44
  exclusively
• 18:45
  on
• 18:45
  260B
• 18:46
  tokens
• 18:47
  of
• 18:47
  pre-1931
• 18:48
  English
• 18:49
  text,
• 18:50
  offers
• 18:50
  a
• 18:50
  genuinely
• 18:50
  novel
• 18:51
  angle
• 18:51
  on
• 18:51
  LLM
• 18:52
  capabilities.
• 18:53
  Built
• 18:53
  by
• 18:54
  Nick
• 18:54
  Levine,
• 18:54
  David
• 18:55
  Duvenaud,
• 18:55
  and
• 18:56
  Alec
• 18:56
  Radford,
• 18:57
  the
• 18:57
  model
• 18:57
  is
• 18:57
  contamination-free
• 18:58
  by
• 18:59
  construction
• 18:59
  --
• 19:00
  enabling
• 19:01
  generalization
• 19:01
  experiments
• 19:02
  impossible
• 19:02
  with
• 19:03
  web-trained
• 19:03
  models,
• 19:04
  like
• 19:04
  testing
• 19:05
  whether
• 19:05
  a
• 19:05
  model
• 19:05
  with
• 19:05
  no
• 19:05
  knowledge
• 19:06
  of
• 19:06
  digital
• 19:06
  computers
• 19:07
  can
• 19:07
  learn
• 19:07
  to
• 19:07
  code
• 19:07
  in
• 19:08
  Python
• 19:08
  from
• 19:08
  in-context
• 19:09
  examples.
• 19:10
  Early
• 19:10
  results
• 19:11
  show
• 19:11
  it
• 19:11
  can
• 19:11
  produce
• 19:12
  simple
• 19:12
  correct
• 19:12
  programs
• 19:13
  (single-line,
• 19:14
  or
• 19:14
  small
• 19:14
  modifications
• 19:15
  to
• 19:15
  examples),
• 19:16
  and
• 19:16
  the
• 19:16
  success
• 19:17
  rate
• 19:17
  improves
• 19:17
  with
• 19:17
  scale.
• 19:19
  The
• 19:19
  post-training
• 19:19
  pipeline
• 19:20
  avoids
• 19:20
  modern
• 19:20
  instruction-response
• 19:21
  pairs
• 19:22
  entirely,
• 19:23
  instead
• 19:23
  using
• 19:23
  historical
• 19:24
  etiquette
• 19:24
  manuals,
• 19:25
  letter-writing
• 19:26
  guides,
• 19:26
  and
• 19:26
  cookbooks,
• 19:27
  with
• 19:27
  Claude
• 19:27
  Opus
• 19:28
  as
• 19:28
  a
• 19:28
  DPO
• 19:28
  A
• 19:30
  "modern
• 19:30
  twin"
• 19:30
  trained
• 19:31
  on
• 19:31
  FineWeb
• 19:31
  with
• 19:31
  the
• 19:31
  same
• 19:32
  compute
• 19:32
  underperforms
• 19:33
  expectations
• 19:34
  on
• 19:34
  some
• 19:34
  benchmarks,
• 19:35
  which
• 19:35
  the
• 19:35
  team
• 19:35
  attributes
• 19:36
  to
• 19:36
  OCR
• 19:37
  noise
• 19:37
  --
• 19:38
  their
• 19:38
  vintage
• 19:39
  OCR
• 19:39
  system
• 19:40
  only
• 19:40
  achieves
• 19:40
  30%
• 19:41
  of
• 19:41
  the
• 19:41
  learning
• 19:41
  efficiency
• 19:42
  of
• 19:42
  human-transcribed
• 19:43
  text,
• 19:44
  with
• 19:44
  regex
• 19:44
  cleaning
• 19:45
  recovering
• 19:45
  to
• 19:45
  70%.
• 19:47
  The
• 19:47
  plan
• 19:48
  is
• 19:48
  to
• 19:48
  scale
• 19:48
  to
• 19:48
  GPT-3.5
• 19:50
  level
• 19:50
  using
• 19:50
  a
• 19:50
  trillion-token
• 19:51
  historical
• 19:52
  corpus.
• 19:53
  The
• 19:53
  research
• 19:54
  question
• 19:54
  underneath
• 19:54
  --
• 19:55
  how
• 19:56
  much
• 19:56
  of
• 19:56
  what
• 19:56
  we
• 19:56
  think
• 19:56
  we
• 19:56
  know
• 19:57
  about
• 19:57
  LLMs
• 19:58
  is
• 19:58
  about
• 19:58
  language
• 19:58
  in
• 19:59
  general
• 19:59
  versus
• 19:59
  the
• 20:00
  web
• 20:00
  as
• 20:00
  a
• 20:00
  dataset
• 20:00
  --
• 20:01
  is
• 20:02
  one
• 20:02
  of
• 20:02
  the
• 20:02
  more
• 20:02
  interesting
• 20:02
  epistemological
• 20:03
  questions
• 20:04
  in
• 20:04
  the
• 20:04
  field
• 20:04
  right
• 20:04
  now.