Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 21, 2026

Transcript

AGI Dreams Podcast – April 21, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Vercel
• 0:00
  disclosed
• 0:01
  a
• 0:01
  security
• 0:01
  incident
• 0:02
  this
• 0:02
  week
• 0:02
  demonstratng
• 0:03
  how
• 0:03
  third-party
• 0:03
  AI
• 0:04
  tool
• 0:04
  integrations
• 0:05
  have
• 0:05
  become
• 0:05
  a
• 0:05
  soft
• 0:05
  underbelly
• 0:06
  of
• 0:06
  cloud
• 0:06
  infrastructure.
• 0:08
  The
• 0:08
  breach
• 0:08
  originated
• 0:09
  not
• 0:09
  from
• 0:09
  a
• 0:09
  Vercel
• 0:10
  system
• 0:10
  but
• 0:10
  from
• 0:10
  Context.ai,
• 0:11
  an
• 0:11
  AI
• 0:12
  platform
• 0:12
  whose
• 0:12
  compromised
• 0:13
  Google
• 0:13
  Workspace
• 0:13
  OAuth
• 0:14
  application
• 0:14
  gave
• 0:14
  attackers
• 0:15
  a
• 0:15
  foothold
• 0:15
  in
• 0:15
  a
• 0:15
  Vercel
• 0:16
  employee's
• 0:16
  account.
• 0:17
  From
• 0:17
  there,
• 0:17
  the
• 0:18
  attacker
• 0:18
  escalated
• 0:18
  into
• 0:19
  Vercel
• 0:19
  environments
• 0:19
  and
• 0:20
  enumerated
• 0:20
  environment
• 0:20
  variables
• 0:21
  that
• 0:21
  were
• 0:21
  not
• 0:21
  marked
• 0:21
  as
• 0:21
  "sensitive"
• 0:22
  —
• 0:22
  and
• 0:22
  therefore
• 0:23
  not
• 0:23
  encrypted
• 0:23
  at
• 0:23
  rest.
• 0:24
  Vercel
• 0:25
  CEO
• 0:25
  Guillermo
• 0:26
  Rauch
• 0:26
  confirmed
• 0:26
  the
• 0:26
  chain:
• 0:27
  Context.ai
• 0:27
  breach
• 0:28
  →
• 0:28
  employee
• 0:28
  Google
• 0:29
  Workspace
• 0:29
  →
• 0:29
  Vercel
• 0:30
  environments
• 0:30
  →
• 0:30
  unencrypted
• 0:31
  env
• 0:31
  vars
• 0:31
  →
• 0:31
  further
• 0:31
  access.
• 0:33
  A
• 0:33
  threat
• 0:33
  actor
• 0:33
  claiming
• 0:34
  to
• 0:34
  be
• 0:34
  ShinyHunters
• 0:34
  (though
• 0:35
  actual
• 0:35
  ShinyHunters
• 0:35
  members
• 0:36
  deny
• 0:36
  involvement)
• 0:37
  is
• 0:37
  now
• 0:37
  offering
• 0:37
  580
• 0:38
  employee
• 0:39
  records,
• 0:40
  API
• 0:40
  keys,
• 0:41
  NPM
• 0:41
  tokens,
• 0:42
  and
• 0:42
  internal
• 0:42
  deployment
• 0:43
  access
• 0:43
  on
• 0:43
  hacking
• 0:44
  forums,
• 0:45
  with
• 0:45
  an
• 0:45
  alleged
• 0:45
  $2
• 0:45
  million
• 0:46
  ransom
• 0:46
  demand.
• 0:47
  Vercel
• 0:48
  says
• 0:48
  Next.js
• 0:50
  and
• 0:50
  its
• 0:50
  open-source
• 0:50
  projects
• 0:51
  remain
• 0:51
  safe,
• 0:51
  and
• 0:52
  has
• 0:52
  rolled
• 0:52
  out
• 0:52
  dashboard
• 0:53
  improvements
• 0:53
  for
• 0:53
  managing
• 0:53
  sensitive
• 0:54
  variables
• 0:54
  —
• 0:55
  the
• 0:55
  kind
• 0:55
  of
• 0:55
  feature
• 0:56
  you
• 0:56
  implement
• 0:56
  after
• 0:56
  the
• 0:57
  barn
• 0:57
  door
• 0:57
  is
• 0:57
  open.
• 0:58
  The
• 0:58
  European
• 0:58
  Commission's
• 0:59
  week
• 0:59
  was
• 0:59
  arguably
• 1:00
  worse.
• 1:01
  Ursula
• 1:01
  von
• 1:01
  der
• 1:01
  Leyen
• 1:02
  presented
• 1:02
  the
• 1:02
  EU's
• 1:02
  new
• 1:03
  age-verification
• 1:03
  app
• 1:04
  on
• 1:04
  Wednesday,
• 1:05
  declaring
• 1:05
  it
• 1:05
  "technically
• 1:06
  ready"
• 1:06
  and
• 1:06
  "fully
• 1:06
  open
• 1:07
  source."
• 1:08
  Security
• 1:08
  consultant
• 1:09
  Paul
• 1:09
  Moore
• 1:09
  broke
• 1:09
  it
• 1:10
  in
• 1:10
  under
• 1:10
  two
• 1:10
  minutes.
• 1:11
  The
• 1:11
  app
• 1:11
  stored
• 1:12
  sensitive
• 1:12
  data
• 1:12
  unprotected
• 1:13
  on-device,
• 1:14
  and
• 1:14
  French
• 1:14
  white-hat
• 1:15
  Baptiste
• 1:15
  Robert
• 1:16
  confirmed
• 1:16
  that
• 1:16
  biometric
• 1:17
  authentication
• 1:18
  was
• 1:18
  bypassable
• 1:18
  —
• 1:19
  meaning
• 1:19
  your
• 1:19
  nephew
• 1:20
  could
• 1:20
  borrow
• 1:20
  your
• 1:20
  phone
• 1:21
  and
• 1:21
  prove
• 1:21
  he's
• 1:21
  over
• 1:21
  18.
• 1:23
  The
• 1:23
  Commission's
• 1:23
  defense
• 1:24
  evolved
• 1:24
  from
• 1:24
  "it's
• 1:24
  technically
• 1:25
  ready"
• 1:25
  to
• 1:25
  "it's
• 1:26
  a
• 1:26
  demo
• 1:26
  version"
• 1:26
  to
• 1:26
  "the
• 1:27
  code
• 1:27
  will
• 1:27
  be
• 1:27
  constantly
• 1:28
  updated"
• 1:28
  within
• 1:28
  48
• 1:29
  hours,
• 1:30
  while
• 1:30
  both
• 1:30
  researchers
• 1:31
  confirmed
• 1:31
  they
• 1:31
  were
• 1:31
  testing
• 1:32
  the
• 1:32
  latest
• 1:32
  published
• 1:33
  code.
• 1:34
  Over
• 1:34
  400
• 1:35
  security
• 1:35
  experts
• 1:36
  had
• 1:36
  already
• 1:36
  sent
• 1:36
  the
• 1:36
  Commission
• 1:37
  a
• 1:37
  letter
• 1:37
  in
• 1:37
  March
• 1:37
  calling
• 1:38
  for
• 1:38
  a
• 1:38
  moratorium
• 1:39
  on
• 1:39
  age-verification
• 1:40
  deployment
• 1:40
  until
• 1:41
  the
• 1:41
  science
• 1:41
  settled.
• 1:42
  Polish
• 1:42
  MEP
• 1:43
  Piotr
• 1:43
  Müller
• 1:43
  called
• 1:43
  it
• 1:43
  "step-by-step
• 1:44
  creation
• 1:45
  of
• 1:45
  a
• 1:45
  Chinese-style
• 1:45
  internet
• 1:46
  in
• 1:46
  Europe."
• 1:47
  The
• 1:47
  deeper
• 1:47
  issue
• 1:47
  isn't
• 1:48
  the
• 1:48
  bugs
• 1:48
  —
• 1:49
  it's
• 1:49
  the
• 1:49
  political
• 1:49
  pressure
• 1:50
  to
• 1:50
  ship
• 1:50
  identity-linked
• 1:51
  verification
• 1:51
  infrastructure
• 1:52
  before
• 1:52
  anyone
• 1:52
  has
• 1:53
  figured
• 1:53
  out
• 1:53
  how
• 1:53
  to
• 1:53
  do
• 1:53
  it
• 1:54
  safely.
• 1:54
  Trust
• 1:55
  manipulation
• 1:55
  extends
• 1:56
  well
• 1:56
  beyond
• 1:56
  government
• 1:57
  apps.
• 1:57
  A
• 1:58
  peer-reviewed
• 1:58
  Carnegie
• 1:59
  Mellon
• 1:59
  study
• 1:59
  (ICSE
• 2:00
  2026)
• 2:01
  analyzed
• 2:02
  20
• 2:02
  terabytes
• 2:02
  of
• 2:03
  GitHub
• 2:03
  metadata
• 2:04
  and
• 2:04
  found
• 2:04
  approximately
• 2:05
  6
• 2:05
  million
• 2:05
  suspected
• 2:06
  fake
• 2:06
  stars
• 2:06
  distributed
• 2:07
  across
• 2:07
  18,617
• 2:09
  repositories
• 2:10
  by
• 2:10
  roughly
• 2:10
  301,000
• 2:11
  accounts.
• 2:12
  Stars
• 2:13
  sell
• 2:13
  for
• 2:13
  $0.06
• 2:13
  on
• 2:14
  at
• 2:14
  least
• 2:14
  a
• 2:14
  dozen
• 2:14
  websites
• 2:14
  —
• 2:15
  no
• 2:15
  dark
• 2:16
  web
• 2:16
  required.
• 2:17
  The
• 2:17
  investigation
• 2:18
  maps
• 2:18
  the
• 2:18
  full
• 2:18
  pipeline
• 2:19
  from
• 2:19
  star
• 2:19
  farms
• 2:19
  to
• 2:19
  venture
• 2:20
  capital:
• 2:21
  Redpoint
• 2:21
  Ventures'
• 2:22
  Jordan
• 2:22
  Segall
• 2:23
  published
• 2:23
  data
• 2:23
  showing
• 2:24
  the
• 2:24
  median
• 2:24
  GitHub
• 2:24
  star
• 2:25
  count
• 2:25
  at
• 2:25
  seed
• 2:25
  financing
• 2:26
  is
• 2:26
  2,850,
• 2:28
  with
• 2:28
  firms
• 2:28
  running
• 2:29
  automated
• 2:29
  scrapers
• 2:30
  to
• 2:30
  identify
• 2:30
  fast-growing
• 2:31
  repos.
• 2:32
  For
• 2:32
  $85
• 2:33
  to
• 2:33
  $285
• 2:35
  in
• 2:35
  purchased
• 2:35
  stars,
• 2:36
  a
• 2:36
  startup
• 2:37
  can
• 2:37
  manufacture
• 2:37
  that
• 2:37
  seed-stage
• 2:38
  median.
• 2:39
  The
• 2:39
  fork-to-star
• 2:40
  ratio
• 2:40
  emerges
• 2:41
  as
• 2:41
  the
• 2:41
  strongest
• 2:42
  detection
• 2:42
  heuristic
• 2:42
  —
• 2:43
  organic
• 2:44
  projects
• 2:44
  like
• 2:45
  Flask
• 2:45
  show
• 2:45
  235
• 2:46
  forks
• 2:47
  per
• 2:47
  1,000
• 2:47
  stars,
• 2:48
  while
• 2:48
  FreeDomain
• 2:49
  (157,000
• 2:50
  stars)
• 2:51
  manages
• 2:51
  just
• 2:51
  11.
• 2:53
  One
• 2:53
  AI
• 2:53
  project
• 2:54
  topped
• 2:54
  Runa
• 2:54
  Capital's
• 2:55
  ROSS
• 2:55
  Index
• 2:56
  with
• 2:56
  74,300
• 2:57
  stars
• 2:57
  and
• 2:58
  54x
• 2:58
  growth;
• 2:59
  the
• 2:59
  study
• 3:00
  flagged
• 3:00
  nearly
• 3:00
  half
• 3:00
  its
• 3:01
  stars
• 3:01
  as
• 3:01
  artificial.
• 3:02
  The
• 3:02
  FTC's
• 3:03
  2024
• 3:04
  rule
• 3:04
  banning
• 3:04
  fake
• 3:05
  social
• 3:05
  influence
• 3:05
  metrics
• 3:06
  carries
• 3:06
  penalties
• 3:07
  up
• 3:07
  to
• 3:07
  $51,744
• 3:09
  per
• 3:09
  violation,
• 3:10
  and
• 3:10
  the
• 3:10
  SEC
• 3:11
  has
• 3:11
  already
• 3:11
  charged
• 3:12
  founders
• 3:12
  for
• 3:12
  inflating
• 3:13
  traction
• 3:13
  during
• 3:13
  fundraising.
• 3:14
  Nobody
• 3:15
  has
• 3:15
  been
• 3:15
  charged
• 3:15
  specifically
• 3:16
  for
• 3:16
  fake
• 3:16
  GitHub
• 3:17
  stars
• 3:17
  yet.
• 3:18
  Given
• 3:18
  the
• 3:18
  math
• 3:18
  —
• 3:19
  a
• 3:19
  $50
• 3:20
  problem
• 3:20
  with
• 3:20
  a
• 3:20
  $50
• 3:20
  million
• 3:21
  consequence
• 3:21
  —
• 3:22
  it
• 3:22
  may
• 3:22
  only
• 3:22
  be
• 3:23
  a
• 3:23
  matter
• 3:23
  of
• 3:23
  time.
• 3:26
  OpenAI
• 3:27
  released
• 3:27
  what
• 3:27
  amounts
• 3:27
  to
• 3:27
  a
• 3:27
  "Codex
• 3:28
  does
• 3:28
  everything
• 3:28
  now"
• 3:29
  update.
• 3:30
  Background
• 3:31
  computer
• 3:31
  use
• 3:31
  lets
• 3:31
  multiple
• 3:32
  agents
• 3:32
  operate
• 3:33
  your
• 3:33
  Mac
• 3:33
  in
• 3:33
  parallel
• 3:33
  —
• 3:34
  seeing,
• 3:34
  clicking,
• 3:35
  and
• 3:35
  typing
• 3:36
  with
• 3:36
  their
• 3:36
  own
• 3:36
  cursors
• 3:36
  —
• 3:37
  without
• 3:37
  interfering
• 3:37
  with
• 3:38
  your
• 3:38
  work.
• 3:38
  An
• 3:38
  in-app
• 3:39
  browser
• 3:39
  enables
• 3:39
  direct
• 3:40
  annotation
• 3:40
  of
• 3:40
  web
• 3:41
  pages
• 3:41
  for
• 3:41
  frontend
• 3:41
  iteration.
• 3:42
  Memory
• 3:43
  lets
• 3:43
  Codex
• 3:43
  remember
• 3:44
  preferences
• 3:44
  and
• 3:44
  corrections
• 3:45
  across
• 3:45
  sessions.
• 3:46
  Automations
• 3:47
  can
• 3:47
  reuse
• 3:47
  conversation
• 3:48
  threads,
• 3:48
  schedule
• 3:48
  future
• 3:49
  work,
• 3:49
  and
• 3:49
  wake
• 3:50
  up
• 3:50
  autonomously
• 3:50
  days
• 3:51
  later
• 3:51
  to
• 3:51
  continue
• 3:51
  tasks.
• 3:52
  Over
• 3:52
  90
• 3:53
  new
• 3:53
  plugins
• 3:53
  ship,
• 3:54
  spanning
• 3:54
  Atlassian
• 3:55
  Rovo,
• 3:55
  CircleCI,
• 3:55
  GitLab,
• 3:56
  and
• 3:56
  Microsoft
• 3:56
  Suite.
• 3:57
  The
• 3:57
  framing
• 3:58
  is
• 3:58
  explicit:
• 3:59
  Codex
• 3:59
  is
• 3:59
  no
• 3:59
  longer
• 3:59
  a
• 4:00
  code
• 4:00
  assistant
• 4:00
  but
• 4:01
  an
• 4:01
  autonomous
• 4:01
  software
• 4:01
  development
• 4:02
  partner
• 4:02
  that
• 4:02
  operates
• 4:03
  across
• 4:03
  the
• 4:03
  full
• 4:03
  lifecycle.
• 4:04
  The
• 4:04
  logical
• 4:05
  endpoint
• 4:05
  of
• 4:05
  this
• 4:05
  trajectory
• 4:06
  is
• 4:06
  the
• 4:06
  dark
• 4:06
  factory
• 4:06
  —
• 4:08
  a
• 4:08
  codebase
• 4:08
  where
• 4:08
  AI
• 4:09
  handles
• 4:09
  planning,
• 4:10
  implementation,
• 4:11
  review,
• 4:11
  and
• 4:11
  production
• 4:12
  deployment
• 4:12
  with
• 4:12
  zero
• 4:12
  human
• 4:13
  in
• 4:13
  the
• 4:13
  loop.
• 4:14
  Cole
• 4:14
  Medin
• 4:14
  is
• 4:14
  running
• 4:14
  the
• 4:14
  first
• 4:15
  public
• 4:15
  experiment,
• 4:16
  built
• 4:16
  on
• 4:16
  Archon
• 4:16
  (his
• 4:17
  open-source
• 4:17
  harness
• 4:17
  builder)
• 4:17
  driving
• 4:18
  Claude
• 4:18
  Code
• 4:18
  sessions
• 4:19
  routed
• 4:19
  to
• 4:19
  MiniMax
• 4:19
  M2.7
• 4:21
  for
• 4:21
  cost
• 4:21
  reasons.
• 4:22
  The
• 4:22
  architecture
• 4:23
  has
• 4:23
  three
• 4:23
  governance
• 4:24
  documents
• 4:24
  (mission,
• 4:25
  factory
• 4:25
  rules,
• 4:26
  CLAUDE.md)
• 4:26
  loaded
• 4:27
  into
• 4:27
  every
• 4:27
  workflow,
• 4:28
  plus
• 4:28
  a
• 4:28
  triage
• 4:29
  →
• 4:29
  implement
• 4:29
  →
• 4:29
  validate
• 4:30
  loop
• 4:30
  running
• 4:30
  on
• 4:30
  a
• 4:30
  cron
• 4:31
  schedule.
• 4:32
  The
• 4:32
  critical
• 4:32
  design
• 4:32
  decision,
• 4:33
  borrowed
• 4:33
  from
• 4:34
  StrongDM's
• 4:34
  production
• 4:35
  dark
• 4:35
  factory,
• 4:35
  is
• 4:36
  the
• 4:36
  holdout
• 4:36
  pattern:
• 4:37
  the
• 4:37
  validation
• 4:37
  agent
• 4:38
  receives
• 4:38
  the
• 4:38
  user
• 4:38
  journey
• 4:39
  and
• 4:39
  code
• 4:39
  diffs
• 4:39
  but
• 4:40
  zero
• 4:40
  context
• 4:40
  from
• 4:41
  the
• 4:41
  development
• 4:41
  process,
• 4:42
  preventing
• 4:42
  the
• 4:42
  sycophancy
• 4:43
  that
• 4:43
  lets
• 4:43
  AI
• 4:44
  agents
• 4:44
  paper
• 4:44
  over
• 4:44
  their
• 4:44
  own
• 4:45
  bugs.
• 4:45
  Issues
• 4:46
  come
• 4:46
  in
• 4:46
  via
• 4:46
  GitHub,
• 4:47
  get
• 4:47
  triaged
• 4:47
  against
• 4:48
  the
• 4:48
  mission
• 4:48
  scope,
• 4:49
  implemented
• 4:49
  in
• 4:49
  parallel
• 4:50
  worktrees,
• 4:50
  validated
• 4:51
  with
• 4:51
  browser
• 4:51
  automation,
• 4:52
  and
• 4:52
  merged
• 4:53
  without
• 4:53
  human
• 4:53
  approval.
• 4:54
  The
• 4:54
  question
• 4:54
  of
• 4:55
  what
• 4:55
  responsible
• 4:55
  agent-assisted
• 4:56
  open
• 4:56
  source
• 4:57
  looks
• 4:57
  like
• 4:57
  gets
• 4:58
  a
• 4:58
  thoughtful
• 4:58
  answer
• 4:58
  from
• 4:58
  Hugging
• 4:59
  Face.
• 5:00
  Their
• 5:00
  transformers-to-MLX
• 5:01
  Skill
• 5:02
  is
• 5:02
  a
• 5:02
  recipe
• 5:02
  for
• 5:02
  porting
• 5:03
  language
• 5:03
  models
• 5:03
  from
• 5:03
  the
• 5:03
  transformers
• 5:04
  library
• 5:04
  to
• 5:05
  Apple's
• 5:05
  MLX
• 5:05
  framework
• 5:06
  —
• 5:07
  not
• 5:07
  a
• 5:07
  fire-and-forget
• 5:07
  automation
• 5:08
  but
• 5:08
  a
• 5:08
  contributor
• 5:09
  tool
• 5:09
  that
• 5:09
  produces
• 5:09
  PRs
• 5:10
  designed
• 5:10
  to
• 5:10
  look
• 5:10
  like
• 5:11
  careful
• 5:11
  human
• 5:11
  submissions.
• 5:12
  The
• 5:12
  Skill
• 5:13
  runs
• 5:13
  per-layer
• 5:13
  comparisons
• 5:14
  against
• 5:14
  the
• 5:14
  transformers
• 5:15
  baseline,
• 5:16
  detects
• 5:16
  undeclared
• 5:16
  dtypes
• 5:17
  from
• 5:17
  safetensors
• 5:17
  metadata,
• 5:18
  and
• 5:18
  generates
• 5:19
  a
• 5:19
  separate
• 5:19
  non-agentic
• 5:20
  test
• 5:20
  harness
• 5:20
  for
• 5:20
  reproducibility.
• 5:22
  Critically,
• 5:22
  it
• 5:23
  will
• 5:23
  not
• 5:23
  open
• 5:23
  a
• 5:23
  PR
• 5:24
  until
• 5:24
  the
• 5:24
  contributor
• 5:24
  has
• 5:24
  accepted
• 5:25
  the
• 5:25
  results,
• 5:26
  and
• 5:26
  it
• 5:26
  always
• 5:26
  discloses
• 5:26
  agent
• 5:27
  assistance.
• 5:28
  The
• 5:28
  post
• 5:28
  is
• 5:28
  equally
• 5:29
  notable
• 5:29
  for
• 5:29
  its
• 5:29
  diagnosis
• 5:30
  of
• 5:30
  the
• 5:30
  problem:
• 5:31
  PR
• 5:31
  volume
• 5:32
  to
• 5:32
  transformers
• 5:32
  has
• 5:32
  gone
• 5:33
  up
• 5:33
  tenfold
• 5:33
  since
• 5:33
  code
• 5:34
  agents
• 5:34
  started
• 5:34
  working,
• 5:35
  but
• 5:35
  the
• 5:35
  number
• 5:35
  of
• 5:35
  maintainers
• 5:36
  has
• 5:36
  not,
• 5:37
  and
• 5:37
  agents
• 5:37
  "break
• 5:38
  implicit
• 5:38
  contracts
• 5:38
  between
• 5:39
  the
• 5:39
  library
• 5:39
  and
• 5:39
  its
• 5:40
  users"
• 5:40
  by
• 5:40
  following
• 5:40
  "best
• 5:41
  practices"
• 5:42
  that
• 5:42
  violate
• 5:42
  undocumented
• 5:43
  design
• 5:43
  decisions.
• 5:44
  On
• 5:44
  the
• 5:44
  tooling
• 5:45
  front,
• 5:45
  Anthropic
• 5:46
  confirmed
• 5:46
  that
• 5:46
  OpenClaw-style
• 5:47
  Claude
• 5:47
  CLI
• 5:48
  usage
• 5:48
  is
• 5:48
  sanctioned
• 5:49
  again,
• 5:49
  with
• 5:50
  OpenClaw
• 5:50
  supporting
• 5:50
  API
• 5:51
  keys,
• 5:51
  Claude
• 5:52
  CLI
• 5:52
  reuse,
• 5:53
  prompt
• 5:53
  caching,
• 5:54
  and
• 5:54
  a
• 5:54
  1M
• 5:55
  context
• 5:55
  window
• 5:55
  beta
• 5:56
  for
• 5:56
  Opus/Sonnet
• 5:56
  models
• 5:57
  .
• 5:58
  Meanwhile,
• 5:58
  the
• 5:58
  Pi
• 5:58
  coding
• 5:59
  agent
• 5:59
  continues
• 6:00
  gaining
• 6:00
  traction
• 6:00
  as
• 6:00
  a
• 6:01
  minimalist
• 6:01
  Claude
• 6:01
  Code
• 6:02
  alternative
• 6:02
  —
• 6:03
  fighting
• 6:03
  the
• 6:03
  feature
• 6:03
  bloat
• 6:04
  that
• 6:04
  has
• 6:04
  made
• 6:04
  Claude
• 6:04
  Code's
• 6:05
  system
• 6:05
  prompt
• 6:05
  unpredictable
• 6:06
  and
• 6:06
  its
• 6:07
  behavior
• 6:07
  harder
• 6:07
  to
• 6:07
  control.
• 6:08
  Pi's
• 6:09
  philosophy:
• 6:10
  build
• 6:10
  the
• 6:10
  smallest
• 6:10
  extensible
• 6:11
  core,
• 6:12
  then
• 6:12
  let
• 6:12
  the
• 6:12
  agent
• 6:12
  modify
• 6:13
  itself
• 6:13
  to
• 6:13
  add
• 6:14
  any
• 6:14
  capability
• 6:14
  you
• 6:15
  need.
• 6:18
  The
• 6:18
  local
• 6:18
  inference
• 6:19
  community
• 6:19
  is
• 6:19
  working
• 6:19
  through
• 6:20
  the
• 6:20
  Kimi
• 6:20
  K2.6
• 6:22
  deployment
• 6:22
  question,
• 6:23
  and
• 6:23
  the
• 6:23
  numbers
• 6:23
  are
• 6:24
  sobering.
• 6:24
  This
• 6:24
  1.1
• 6:25
  trillion-parameter
• 6:26
  MoE
• 6:26
  model
• 6:27
  needs
• 6:27
  roughly
• 6:27
  768GB
• 6:29
  of
• 6:29
  VRAM
• 6:30
  to
• 6:30
  hit
• 6:30
  25-30
• 6:31
  tokens
• 6:31
  per
• 6:31
  second
• 6:32
  at
• 6:32
  full
• 6:32
  precision
• 6:32
  with
• 6:32
  full
• 6:33
  context.
• 6:34
  One
• 6:34
  user
• 6:34
  got
• 6:34
  it
• 6:34
  running
• 6:35
  on
• 6:35
  a
• 6:35
  512GB
• 6:36
  M3
• 6:37
  Ultra
• 6:37
  at
• 6:37
  a
• 6:37
  "glorious"
• 6:38
  6
• 6:38
  tokens
• 6:38
  per
• 6:39
  second
• 6:39
  using
• 6:39
  an
• 6:39
  IQ2
• 6:40
  KL
• 6:40
  quantization
• 6:41
  —
• 6:41
  functional
• 6:42
  but
• 6:42
  hardly
• 6:42
  production
• 6:43
  speed.
• 6:44
  The
• 6:44
  realistic
• 6:44
  hardware
• 6:44
  estimate
• 6:45
  for
• 6:45
  target
• 6:45
  throughput
• 6:46
  ranges
• 6:46
  from
• 6:46
  $28,000
• 6:47
  (AMD
• 6:47
  Genoa
• 6:48
  +
• 6:49
  2-4
• 6:49
  RTX
• 6:49
  Pro
• 6:49
  6000s)
• 6:50
  to
• 6:51
  $80,000
• 6:51
  (8x
• 6:52
  RTX
• 6:52
  Pro
• 6:52
  6000
• 6:53
  96GB
• 6:53
  each).
• 6:54
  The
• 6:55
  model
• 6:55
  is
• 6:55
  natively
• 6:55
  QAT
• 6:56
  at
• 6:56
  INT4,
• 6:57
  which
• 6:57
  helps
• 6:57
  the
• 6:57
  footprint,
• 6:58
  but
• 6:58
  full
• 6:58
  265K
• 6:59
  context
• 7:00
  at
• 7:00
  F16
• 7:00
  KV
• 7:01
  cache
• 7:01
  still
• 7:01
  demands
• 7:02
  around
• 7:02
  48GB
• 7:03
  on
• 7:03
  top
• 7:03
  of
• 7:03
  the
• 7:03
  ~548GB
• 7:05
  model
• 7:05
  weight.
• 7:05
  At
• 7:06
  the
• 7:06
  opposite
• 7:06
  end
• 7:06
  of
• 7:06
  the
• 7:06
  cost
• 7:07
  spectrum,
• 7:07
  a
• 7:07
  detailed
• 7:08
  build
• 7:08
  guide
• 7:08
  turns
• 7:09
  a
• 7:09
  Xiaomi
• 7:09
  12
• 7:09
  Pro
• 7:09
  into
• 7:10
  a
• 7:10
  24/7
• 7:11
  headless
• 7:11
  AI
• 7:11
  server.
• 7:12
  The
• 7:12
  recipe:
• 7:13
  flash
• 7:13
  LineageOS,
• 7:14
  kill
• 7:14
  the
• 7:14
  Android
• 7:15
  UI
• 7:15
  entirely
• 7:15
  to
• 7:16
  reclaim
• 7:16
  RAM,
• 7:17
  disable
• 7:17
  the
• 7:17
  phantom
• 7:17
  process
• 7:18
  killer
• 7:18
  via
• 7:18
  ADB,
• 7:19
  and
• 7:19
  run
• 7:19
  Ollama
• 7:20
  in
• 7:20
  Termux
• 7:21
  with
• 7:21
  a
• 7:21
  wake
• 7:21
  lock.
• 7:22
  A
• 7:22
  custom
• 7:22
  script
• 7:22
  suite
• 7:23
  ("Zeus")
• 7:23
  handles
• 7:23
  Wi-Fi
• 7:24
  recovery
• 7:24
  after
• 7:24
  UI
• 7:25
  shutdown,
• 7:25
  battery
• 7:26
  cycling
• 7:26
  between
• 7:27
  40-80%
• 7:29
  with
• 7:29
  Telegram
• 7:29
  alerts,
• 7:30
  and
• 7:30
  automated
• 7:30
  fan
• 7:31
  control
• 7:31
  via
• 7:31
  a
• 7:31
  SONOFF
• 7:33
  smart
• 7:33
  plug
• 7:34
  triggered
• 7:34
  at
• 7:34
  45C.
• 7:35
  Benchmarks
• 7:35
  show
• 7:35
  Gemma
• 7:36
  4
• 7:36
  E2B
• 7:36
  at
• 7:36
  Q8
• 7:37
  generating
• 7:37
  2.16
• 7:38
  tokens
• 7:38
  per
• 7:38
  second
• 7:39
  and
• 7:39
  Qwen2.5
• 7:40
  7B
• 7:42
  at
• 7:42
  Q4
• 7:42
  managing
• 7:43
  1.54
• 7:44
  tok/s.
• 7:45
  An
• 7:45
  update
• 7:45
  after
• 7:46
  compiling
• 7:46
  llama.cpp
• 7:48
  with
• 7:48
  ARM
• 7:49
  NEON
• 7:49
  intrinsics
• 7:50
  and
• 7:50
  Q4
• 7:51
  0
• 7:51
  quantization
• 7:52
  reportedly
• 7:52
  improved
• 7:53
  speed
• 7:53
  dramatically.
• 7:54
  Gemma
• 7:55
  4's
• 7:55
  smallest
• 7:55
  variant,
• 7:56
  the
• 7:56
  e4B-IT,
• 7:58
  is
• 7:58
  drawing
• 7:58
  attention
• 7:58
  for
• 7:59
  a
• 7:59
  different
• 7:59
  reason:
• 8:00
  it
• 8:00
  follows
• 8:00
  instructions
• 8:01
  so
• 8:01
  eagerly
• 8:01
  that
• 8:02
  it
• 8:02
  essentially
• 8:02
  has
• 8:02
  no
• 8:03
  refusal
• 8:03
  behavior
• 8:03
  out
• 8:03
  of
• 8:04
  the
• 8:04
  box.
• 8:05
  Users
• 8:05
  report
• 8:05
  the
• 8:05
  vanilla
• 8:06
  model
• 8:06
  handles
• 8:06
  "unhinged
• 8:07
  and
• 8:07
  dark"
• 8:07
  content
• 8:08
  without
• 8:08
  pushback,
• 8:09
  with
• 8:09
  community-produced
• 8:10
  uncensored
• 8:10
  quantizations
• 8:11
  pushing
• 8:12
  the
• 8:12
  envelope
• 8:12
  further.
• 8:16
  A
• 8:16
  benchmark
• 8:16
  of
• 8:16
  four
• 8:16
  agent
• 8:17
  memory
• 8:17
  systems
• 8:17
  delivers
• 8:18
  the
• 8:18
  kind
• 8:18
  of
• 8:18
  results
• 8:19
  that
• 8:19
  should
• 8:19
  make
• 8:19
  anyone
• 8:19
  relying
• 8:20
  on
• 8:20
  these
• 8:20
  tools
• 8:20
  pause.
• 8:22
  Mem0
• 8:22
  scores
• 8:22
  49%
• 8:23
  recall
• 8:23
  —
• 8:24
  worse
• 8:24
  than
• 8:24
  a
• 8:24
  coin
• 8:24
  flip.
• 8:25
  Zep
• 8:26
  burns
• 8:26
  340x
• 8:27
  more
• 8:27
  tokens
• 8:27
  for
• 8:27
  a
• 8:27
  15-point
• 8:28
  improvement.
• 8:29
  The
• 8:29
  community
• 8:29
  response
• 8:29
  was
• 8:30
  blunt:
• 8:30
  "most
• 8:30
  agent
• 8:31
  memory
• 8:31
  systems
• 8:31
  are
• 8:31
  out
• 8:31
  of
• 8:32
  vibe
• 8:32
  coding
• 8:32
  hell
• 8:32
  —
• 8:33
  a
• 8:33
  plague
• 8:33
  like
• 8:33
  Todo-Apps,
• 8:34
  LLM
• 8:34
  Routers
• 8:35
  and
• 8:35
  Flappy-Bird
• 8:35
  Games."
• 8:36
  The
• 8:37
  deeper
• 8:37
  critique
• 8:37
  is
• 8:37
  architectural:
• 8:38
  memory
• 8:39
  is
• 8:39
  token-expensive
• 8:40
  and
• 8:40
  fundamentally
• 8:40
  misaligned
• 8:41
  with
• 8:41
  how
• 8:41
  LLMs
• 8:42
  work.
• 8:42
  These
• 8:43
  aren't
• 8:43
  real
• 8:43
  people
• 8:43
  who
• 8:43
  benefit
• 8:44
  from
• 8:44
  remembering
• 8:44
  —
• 8:45
  they're
• 8:45
  stateless
• 8:46
  inference
• 8:46
  engines
• 8:46
  where
• 8:46
  context
• 8:47
  window
• 8:47
  management
• 8:48
  matters
• 8:48
  more
• 8:48
  than
• 8:48
  persistent
• 8:49
  storage.
• 8:50
  The
• 8:50
  most
• 8:50
  practical
• 8:51
  suggestion
• 8:51
  in
• 8:51
  the
• 8:51
  thread:
• 8:52
  "chronologically
• 8:53
  stacked,
• 8:54
  meticulously
• 8:54
  prompted,
• 8:55
  unpackable
• 8:56
  and
• 8:56
  repackable
• 8:56
  compaction."
• 8:58
  Several
• 8:58
  commenters
• 8:59
  admitted
• 8:59
  to
• 8:59
  manually
• 9:00
  pruning
• 9:00
  JSON
• 9:00
  files
• 9:01
  after
• 9:01
  each
• 9:01
  session
• 9:02
  because
• 9:02
  nothing
• 9:02
  automated
• 9:03
  worked
• 9:03
  without
• 9:03
  destroying
• 9:04
  their
• 9:04
  token
• 9:04
  budget.
• 9:05
  A
• 9:05
  single-file
• 9:06
  OpenWebUI
• 9:07
  filter
• 9:07
  function
• 9:07
  tackles
• 9:08
  these
• 9:08
  problems
• 9:08
  pragmatically:
• 9:09
  semantic
• 9:10
  routing
• 9:10
  classifies
• 9:10
  queries
• 9:11
  into
• 9:11
  five
• 9:11
  categories
• 9:12
  with
• 9:12
  different
• 9:12
  system
• 9:12
  prompts,
• 9:13
  multi-provider
• 9:14
  failover
• 9:14
  switches
• 9:15
  from
• 9:15
  Groq
• 9:15
  to
• 9:15
  Fireworks
• 9:16
  mid-request,
• 9:17
  citation
• 9:17
  verification
• 9:18
  checks
• 9:18
  that
• 9:18
  factual
• 9:19
  responses
• 9:19
  cite
• 9:20
  real
• 9:20
  URLs
• 9:20
  from
• 9:20
  search
• 9:21
  results
• 9:21
  rather
• 9:21
  than
• 9:21
  hallucinated
• 9:22
  ones,
• 9:23
  and
• 9:23
  per-chat
• 9:23
  memory
• 9:24
  stores
• 9:24
  conversation
• 9:25
  turns
• 9:25
  with
• 9:25
  BM25+cosine
• 9:26
  hybrid
• 9:27
  retrieval
• 9:27
  plus
• 9:27
  compression
• 9:28
  of
• 9:28
  oldest
• 9:28
  turns
• 9:28
  .
• 9:29
  Separately,
• 9:30
  the
• 9:30
  Inline
• 9:30
  Visualizer
• 9:31
  v2
• 9:31
  plugin
• 9:32
  enables
• 9:32
  live
• 9:32
  streaming
• 9:33
  visualization
• 9:33
  rendering
• 9:34
  inside
• 9:34
  OpenWebUI
• 9:34
  chats
• 9:35
  —
• 9:36
  SVGs
• 9:36
  assembling
• 9:36
  token-by-token,
• 9:37
  Chart.js
• 9:39
  bars
• 9:39
  populating
• 9:39
  column
• 9:40
  by
• 9:40
  column,
• 9:40
  D3
• 9:41
  force
• 9:41
  graphs
• 9:41
  settling
• 9:42
  in
• 9:42
  real
• 9:42
  time.
• 9:43
  A
• 9:43
  custom
• 9:43
  safe-cut
• 9:44
  HTML
• 9:45
  parser
• 9:45
  plus
• 9:45
  incremental
• 9:46
  DOM
• 9:46
  reconciler
• 9:47
  ensures
• 9:47
  existing
• 9:47
  elements
• 9:48
  never
• 9:48
  re-mount
• 9:48
  and
• 9:48
  animations
• 9:49
  don't
• 9:49
  retrigger.
• 9:50
  Cloudflare
• 9:51
  launched
• 9:51
  a
• 9:51
  scanner
• 9:51
  that
• 9:51
  evaluates
• 9:52
  how
• 9:52
  ready
• 9:52
  websites
• 9:53
  are
• 9:53
  for
• 9:53
  AI
• 9:53
  agent
• 9:53
  interaction,
• 9:54
  checking
• 9:55
  across
• 9:55
  five
• 9:55
  categories:
• 9:56
  discoverability
• 9:57
  (robots.txt,
• 9:58
  sitemaps),
• 9:59
  AI
• 10:00
  bot
• 10:00
  rules,
• 10:00
  agent
• 10:01
  infrastructure
• 10:01
  (server
• 10:02
  cards,
• 10:02
  agent
• 10:02
  skills,
• 10:03
  API
• 10:03
  catalogs,
• 10:04
  OAuth
• 10:04
  discovery),
• 10:05
  and
• 10:05
  Model
• 10:05
  Context
• 10:06
  Protocol
• 10:06
  (MCP)
• 10:07
  support.
• 10:08
  The
• 10:08
  message
• 10:08
  is
• 10:08
  clear:
• 10:09
  the
• 10:09
  web
• 10:09
  is
• 10:09
  being
• 10:10
  restructured
• 10:10
  around
• 10:10
  agent
• 10:11
  access,
• 10:12
  and
• 10:12
  sites
• 10:12
  that
• 10:12
  don't
• 10:12
  publish
• 10:13
  machine-readable
• 10:13
  capability
• 10:14
  descriptions
• 10:15
  will
• 10:15
  be
• 10:15
  invisible
• 10:15
  to
• 10:16
  the
• 10:16
  next
• 10:16
  generation
• 10:17
  of
• 10:17
  automated
• 10:17
  consumers.
• 10:21
  As
• 10:21
  LLMs
• 10:21
  push
• 10:22
  toward
• 10:22
  million-token
• 10:23
  context
• 10:23
  windows,
• 10:24
  the
• 10:24
  KV
• 10:24
  cache
• 10:25
  becomes
• 10:25
  the
• 10:25
  bottleneck
• 10:25
  —
• 10:26
  VRAM
• 10:27
  scales
• 10:27
  linearly
• 10:28
  with
• 10:28
  sequence
• 10:28
  length,
• 10:29
  and
• 10:29
  existing
• 10:30
  Top-K
• 10:30
  eviction
• 10:30
  strategies
• 10:31
  delete
• 10:31
  tokens
• 10:31
  that
• 10:32
  may
• 10:32
  become
• 10:32
  critical
• 10:32
  later.
• 10:33
  A
• 10:33
  new
• 10:34
  approach
• 10:34
  reframes
• 10:34
  the
• 10:34
  problem
• 10:35
  entirely:
• 10:36
  instead
• 10:36
  of
• 10:36
  deciding
• 10:37
  which
• 10:37
  tokens
• 10:37
  to
• 10:37
  remove,
• 10:38
  mathematically
• 10:38
  summarize
• 10:39
  them.
• 10:40
  The
• 10:40
  SRC
• 10:41
  (Selection-Reconstruction-Compression)
• 10:42
  pipeline
• 10:43
  uses
• 10:43
  Shannon
• 10:43
  entropy
• 10:44
  of
• 10:44
  attention
• 10:44
  weights
• 10:44
  to
• 10:45
  identify
• 10:45
  "diffuse"
• 10:45
  tokens
• 10:46
  whose
• 10:46
  information
• 10:46
  is
• 10:47
  less
• 10:47
  specific,
• 10:48
  moves
• 10:48
  them
• 10:48
  to
• 10:48
  a
• 10:48
  recycle
• 10:49
  bin,
• 10:49
  solves
• 10:50
  an
• 10:50
  ordinary
• 10:50
  least
• 10:50
  squares
• 10:51
  problem
• 10:51
  to
• 10:51
  reconstruct
• 10:52
  their
• 10:52
  functional
• 10:52
  contribution
• 10:53
  as
• 10:53
  a
• 10:53
  compact
• 10:53
  weight
• 10:54
  matrix,
• 10:55
  then
• 10:55
  compresses
• 10:55
  via
• 10:55
  SVD
• 10:56
  to
• 10:56
  retain
• 10:56
  only
• 10:57
  the
• 10:57
  dominant
• 10:57
  singular
• 10:57
  values.
• 10:59
  Under
• 10:59
  equal
• 10:59
  memory
• 10:59
  budgets,
• 11:00
  this
• 11:00
  entropy-guided
• 11:01
  approach
• 11:02
  (HAE)
• 11:02
  achieves
• 11:02
  near-zero
• 11:03
  reconstruction
• 11:03
  error
• 11:04
  at
• 11:04
  30%
• 11:04
  keep
• 11:05
  ratio
• 11:05
  while
• 11:05
  Top-K
• 11:06
  degrades
• 11:06
  rapidly.
• 11:07
  It
• 11:07
  actually
• 11:07
  uses
• 11:08
  less
• 11:08
  memory
• 11:08
  than
• 11:08
  Top-K
• 11:09
  across
• 11:09
  all
• 11:09
  ratios
• 11:10
  because
• 11:10
  redundant
• 11:11
  information
• 11:11
  is
• 11:11
  mathematically
• 11:12
  removed
• 11:12
  rather
• 11:13
  than
• 11:13
  explicitly
• 11:13
  stored.
• 11:14
  The
• 11:14
  primary
• 11:15
  trade-off
• 11:15
  is
• 11:15
  computation
• 11:16
  time
• 11:16
  for
• 11:16
  the
• 11:16
  OLS
• 11:17
  and
• 11:17
  SVD
• 11:17
  steps
• 11:18
  —
• 11:19
  the
• 11:19
  authors
• 11:19
  plan
• 11:20
  custom
• 11:20
  Triton
• 11:20
  kernels
• 11:20
  to
• 11:21
  amortize
• 11:21
  that
• 11:21
  cost.
• 11:22
  A
• 11:22
  USC
• 11:23
  paper
• 11:23
  introduces
• 11:24
  Density-Guided
• 11:24
  Response
• 11:25
  Optimization
• 11:25
  (DGRO),
• 11:26
  which
• 11:27
  aligns
• 11:27
  language
• 11:27
  models
• 11:28
  to
• 11:28
  community
• 11:28
  norms
• 11:28
  without
• 11:29
  any
• 11:29
  explicit
• 11:29
  preference
• 11:30
  labels.
• 11:31
  The
• 11:31
  key
• 11:31
  observation:
• 11:32
  responses
• 11:33
  that
• 11:33
  communities
• 11:34
  consistently
• 11:34
  accept
• 11:35
  cluster
• 11:35
  in
• 11:35
  high-density
• 11:36
  regions
• 11:36
  of
• 11:36
  embedding
• 11:37
  space,
• 11:37
  forming
• 11:38
  a
• 11:38
  measurable
• 11:38
  "acceptance
• 11:39
  manifold."
• 11:40
  DGRO
• 11:41
  uses
• 11:41
  local
• 11:42
  kernel
• 11:42
  density
• 11:42
  estimation
• 11:43
  conditioned
• 11:43
  on
• 11:43
  nearby
• 11:44
  conversation
• 11:44
  contexts
• 11:45
  to
• 11:45
  rank
• 11:45
  candidate
• 11:45
  responses,
• 11:47
  then
• 11:47
  feeds
• 11:47
  those
• 11:47
  density-derived
• 11:48
  rankings
• 11:49
  into
• 11:49
  standard
• 11:49
  DPO
• 11:50
  training.
• 11:51
  Tested
• 11:51
  on
• 11:51
  eating
• 11:51
  disorder
• 11:52
  support
• 11:52
  communities
• 11:53
  and
• 11:53
  Russian-language
• 11:53
  conflict
• 11:54
  documentation
• 11:54
  forums
• 11:55
  —
• 11:56
  domains
• 11:56
  where
• 11:56
  explicit
• 11:57
  preference
• 11:57
  annotation
• 11:58
  would
• 11:58
  be
• 11:58
  ethically
• 11:58
  fraught
• 11:59
  or
• 11:59
  culturally
• 11:59
  misaligned
• 12:00
  —
• 12:00
  DGRO-aligned
• 12:01
  models
• 12:02
  consistently
• 12:02
  outperformed
• 12:03
  supervised
• 12:04
  fine-tuning
• 12:04
  baselines
• 12:05
  in
• 12:05
  head-to-head
• 12:05
  evaluation.
• 12:06
  The
• 12:07
  paper
• 12:07
  is
• 12:07
  careful
• 12:08
  about
• 12:08
  limitations:
• 12:09
  DGRO
• 12:10
  reproduces
• 12:10
  whatever
• 12:11
  norms
• 12:11
  exist,
• 12:12
  including
• 12:12
  harmful
• 12:13
  ones,
• 12:13
  and
• 12:14
  is
• 12:14
  vulnerable
• 12:14
  to
• 12:14
  coordinated
• 12:15
  manipulation
• 12:16
  of
• 12:16
  acceptance
• 12:16
  signals.
• 12:17
  It
• 12:17
  is
• 12:18
  positioned
• 12:18
  as
• 12:18
  a
• 12:18
  descriptive
• 12:19
  tool,
• 12:19
  not
• 12:20
  a
• 12:20
  normative
• 12:20
  authority.
• 12:21
  NVIDIA's
• 12:22
  Nemotron
• 12:22
  OCR
• 12:23
  v2
• 12:23
  demonstrates
• 12:24
  the
• 12:24
  power
• 12:24
  of
• 12:24
  synthetic
• 12:25
  data
• 12:25
  at
• 12:25
  scale.
• 12:27
  A
• 12:27
  modified
• 12:28
  SynthDoG
• 12:28
  pipeline
• 12:29
  generates
• 12:29
  12
• 12:30
  million
• 12:30
  training
• 12:30
  images
• 12:31
  across
• 12:31
  six
• 12:31
  languages
• 12:32
  with
• 12:32
  pixel-perfect
• 12:33
  ground
• 12:33
  truth
• 12:33
  at
• 12:33
  word,
• 12:34
  line,
• 12:34
  and
• 12:35
  paragraph
• 12:35
  levels
• 12:35
  —
• 12:36
  including
• 12:37
  reading-order
• 12:37
  relation
• 12:38
  graphs
• 12:38
  that
• 12:38
  most
• 12:39
  real-world
• 12:39
  OCR
• 12:40
  datasets
• 12:40
  lack.
• 12:41
  The
• 12:41
  result:
• 12:42
  NED
• 12:42
  scores
• 12:43
  drop
• 12:43
  from
• 12:43
  0.56-0.92
• 12:46
  (unusable)
• 12:49
  to
• 12:50
  0.035-0.069
• 12:55
  on
• 12:55
  non-English
• 12:55
  languages.
• 12:57
  A
• 12:57
  single
• 12:57
  unified
• 12:58
  model
• 12:58
  handles
• 12:58
  all
• 12:59
  languages
• 12:59
  simultaneously
• 13:00
  at
• 13:00
  34.7
• 13:01
  pages
• 13:02
  per
• 13:02
  second
• 13:02
  on
• 13:02
  one
• 13:03
  A100,
• 13:04
  outperforming
• 13:05
  even
• 13:05
  language-specialized
• 13:06
  PaddleOCR
• 13:07
  variants
• 13:07
  while
• 13:07
  running
• 13:08
  28x
• 13:08
  faster
• 13:08
  .
• 13:09
  In
• 13:10
  video
• 13:10
  generation,
• 13:11
  a
• 13:11
  Max
• 13:11
  Planck
• 13:11
  team
• 13:12
  introduces
• 13:12
  Physical
• 13:13
  Simulator
• 13:13
  In-the-Loop
• 13:14
  Video
• 13:14
  Generation
• 13:15
  (PSIVG),
• 13:16
  which
• 13:16
  integrates
• 13:17
  an
• 13:17
  MPM-based
• 13:17
  physics
• 13:18
  simulator
• 13:18
  directly
• 13:19
  into
• 13:19
  the
• 13:19
  diffusion
• 13:19
  process.
• 13:21
  The
• 13:21
  pipeline
• 13:21
  reconstructs
• 13:22
  4D
• 13:22
  scenes
• 13:23
  from
• 13:23
  template
• 13:23
  videos,
• 13:24
  initializes
• 13:24
  them
• 13:25
  in
• 13:25
  a
• 13:25
  physics
• 13:25
  engine,
• 13:26
  and
• 13:26
  uses
• 13:26
  simulated
• 13:27
  trajectories
• 13:27
  to
• 13:27
  guide
• 13:28
  generation
• 13:28
  toward
• 13:29
  physically
• 13:29
  coherent
• 13:29
  motion.
• 13:31
  A
• 13:31
  test-time
• 13:31
  texture
• 13:32
  consistency
• 13:32
  optimization
• 13:33
  (TTCO)
• 13:33
  technique
• 13:34
  adapts
• 13:34
  text
• 13:35
  embeddings
• 13:35
  based
• 13:35
  on
• 13:35
  pixel
• 13:36
  correspondences
• 13:37
  from
• 13:37
  the
• 13:37
  simulator.
• 13:38
  In
• 13:38
  user
• 13:38
  studies,
• 13:39
  PSIVG
• 13:40
  was
• 13:40
  preferred
• 13:41
  in
• 13:41
  82.3%
• 13:42
  of
• 13:42
  comparisons
• 13:43
  against
• 13:43
  baseline
• 13:43
  methods.
• 13:46
  AutoProber
• 13:46
  is
• 13:47
  the
• 13:47
  kind
• 13:47
  of
• 13:47
  project
• 13:47
  that
• 13:47
  makes
• 13:48
  you
• 13:48
  wonder
• 13:48
  what
• 13:48
  garage
• 13:49
  security
• 13:49
  research
• 13:49
  will
• 13:49
  look
• 13:50
  like
• 13:50
  in
• 13:50
  five
• 13:50
  years.
• 13:51
  Built
• 13:51
  from
• 13:52
  a
• 13:52
  $300
• 13:52
  CNC
• 13:53
  machine,
• 13:53
  a
• 13:53
  USB
• 13:54
  microscope,
• 13:54
  a
• 13:54
  Siglent
• 13:55
  oscilloscope,
• 13:56
  and
• 13:56
  3D-printed
• 13:56
  parts,
• 13:57
  it
• 13:57
  gives
• 13:58
  an
• 13:58
  AI
• 13:58
  agent
• 13:58
  everything
• 13:59
  it
• 13:59
  needs
• 13:59
  to
• 13:59
  autonomously
• 14:00
  probe
• 14:00
  hardware
• 14:00
  targets.
• 14:01
  The
• 14:01
  workflow:
• 14:02
  the
• 14:02
  agent
• 14:02
  identifies
• 14:03
  the
• 14:03
  target
• 14:03
  on
• 14:03
  the
• 14:04
  CNC
• 14:04
  bed,
• 14:05
  takes
• 14:05
  microscope
• 14:05
  frames
• 14:06
  while
• 14:06
  tracking
• 14:06
  XYZ
• 14:06
  coordinates,
• 14:07
  stitches
• 14:08
  them
• 14:08
  into
• 14:08
  an
• 14:08
  annotated
• 14:08
  map
• 14:09
  identifying
• 14:09
  pads,
• 14:10
  pins,
• 14:10
  and
• 14:10
  chips,
• 14:11
  proposes
• 14:11
  probe
• 14:12
  targets
• 14:12
  for
• 14:12
  human
• 14:12
  approval,
• 14:13
  then
• 14:13
  physically
• 14:14
  probes
• 14:14
  the
• 14:14
  approved
• 14:14
  points.
• 14:15
  Safety
• 14:16
  is
• 14:16
  taken
• 14:16
  seriously
• 14:16
  —
• 14:17
  an
• 14:17
  independent
• 14:18
  optical
• 14:18
  endstop
• 14:19
  on
• 14:19
  oscilloscope
• 14:19
  Channel
• 14:20
  4
• 14:20
  provides
• 14:21
  continuous
• 14:21
  monitoring
• 14:22
  during
• 14:22
  any
• 14:22
  motion,
• 14:23
  with
• 14:23
  any
• 14:23
  anomaly
• 14:24
  triggering
• 14:24
  an
• 14:24
  immediate
• 14:24
  feed
• 14:25
  hold.
• 14:25
  The
• 14:26
  entire
• 14:26
  stack
• 14:26
  is
• 14:26
  Python-controlled
• 14:27
  with
• 14:27
  a
• 14:27
  Flask
• 14:28
  dashboard,
• 14:29
  and
• 14:29
  the
• 14:29
  agent
• 14:29
  handles
• 14:29
  everything
• 14:30
  from
• 14:30
  homing
• 14:30
  and
• 14:31
  calibration
• 14:31
  to
• 14:31
  target
• 14:32
  identification.
• 14:33
  Meanwhile,
• 14:34
  a
• 14:34
  Chinese-language
• 14:34
  GitHub
• 14:35
  repository
• 14:35
  called
• 14:36
  "awesome-human-distillation"
• 14:37
  catalogs
• 14:38
  over
• 14:38
  204
• 14:39
  projects
• 14:39
  that
• 14:39
  turn
• 14:40
  real
• 14:40
  people
• 14:40
  —
• 14:41
  colleagues,
• 14:42
  mentors,
• 14:42
  ex-partners,
• 14:43
  parents,
• 14:44
  deceased
• 14:44
  family
• 14:45
  members
• 14:45
  —
• 14:46
  into
• 14:46
  callable
• 14:47
  Claude
• 14:47
  Code
• 14:47
  Agent
• 14:48
  Skills.
• 14:49
  Categories
• 14:50
  span
• 14:50
  workplace
• 14:51
  (distill
• 14:51
  your
• 14:51
  boss),
• 14:52
  academia
• 14:52
  (distill
• 14:53
  your
• 14:53
  thesis
• 14:53
  advisor),
• 14:54
  intimate
• 14:54
  relationships
• 14:55
  (distill
• 14:55
  your
• 14:56
  ex,
• 14:56
  your
• 14:57
  crush,
• 14:57
  your
• 14:58
  first
• 14:58
  love),
• 14:59
  and
• 14:59
  public
• 14:59
  figures
• 14:59
  (from
• 15:00
  Confucius
• 15:00
  to
• 15:00
  Trump
• 15:00
  to
• 15:01
  Jack
• 15:01
  Ma).
• 15:02
  The
• 15:02
  repository's
• 15:03
  disclaimer
• 15:03
  is
• 15:03
  characteristically
• 15:04
  self-aware:
• 15:05
  "We
• 15:05
  have
• 15:06
  no
• 15:06
  intention
• 15:06
  of
• 15:06
  becoming
• 15:07
  traitors
• 15:07
  to
• 15:07
  humanity
• 15:08
  —
• 15:09
  down
• 15:09
  with
• 15:09
  AI-centrism!"
• 15:11
  Anti-distillation
• 15:12
  skills
• 15:13
  also
• 15:13
  exist,
• 15:14
  designed
• 15:14
  to
• 15:14
  make
• 15:15
  your
• 15:15
  distilled
• 15:15
  skill
• 15:15
  file
• 15:16
  look
• 15:16
  complete
• 15:16
  while
• 15:17
  keeping
• 15:17
  the
• 15:17
  real
• 15:17
  knowledge
• 15:18
  to
• 15:18
  yourself.
• 15:19
  On
• 15:19
  the
• 15:19
  evaluation
• 15:20
  side,
• 15:20
  Charles
• 15:21
  Martin's
• 15:21
  analysis
• 15:22
  of
• 15:22
  LLM-as-a-Verifier
• 15:23
  techniques
• 15:23
  highlights
• 15:24
  a
• 15:24
  shift
• 15:24
  in
• 15:24
  how
• 15:25
  we
• 15:25
  extract
• 15:25
  judgment
• 15:26
  from
• 15:26
  models.
• 15:27
  Rather
• 15:27
  than
• 15:27
  taking
• 15:28
  the
• 15:28
  single
• 15:28
  score
• 15:28
  token
• 15:29
  a
• 15:29
  model
• 15:29
  outputs,
• 15:30
  the
• 15:30
  system
• 15:30
  examines
• 15:31
  logprobs
• 15:32
  across
• 15:32
  all
• 15:32
  possible
• 15:32
  score
• 15:33
  tokens,
• 15:33
  computes
• 15:34
  expected
• 15:34
  scores
• 15:35
  from
• 15:35
  the
• 15:35
  full
• 15:35
  distribution,
• 15:36
  repeats
• 15:37
  verification
• 15:37
  multiple
• 15:38
  times,
• 15:39
  and
• 15:39
  decomposes
• 15:39
  evaluation
• 15:40
  across
• 15:40
  multiple
• 15:41
  criteria.
• 15:42
  Results
• 15:43
  include
• 15:43
  77.8%
• 15:45
  on
• 15:45
  SWE-Bench
• 15:45
  Verified
• 15:46
  and
• 15:46
  78.9%
• 15:48
  pairwise
• 15:48
  verification
• 15:49
  accuracy
• 15:49
  on
• 15:49
  Terminal-Bench.
• 15:51
  The
• 15:51
  implication:
• 15:52
  some
• 15:53
  older
• 15:53
  best
• 15:53
  practices
• 15:54
  for
• 15:54
  LLM
• 15:54
  judges
• 15:55
  need
• 15:55
  revision
• 15:56
  as
• 15:56
  frontier
• 15:56
  models
• 15:57
  improve
• 15:57
  at
• 15:57
  nuanced
• 15:58
  evaluation
• 15:58
  —
• 15:59
  but
• 15:59
  the
• 16:00
  gains
• 16:00
  appear
• 16:00
  tied
• 16:01
  to
• 16:01
  the
• 16:01
  specific
• 16:01
  verifier
• 16:02
  setup,
• 16:03
  not
• 16:03
  a
• 16:03
  blanket
• 16:03
  "switch
• 16:04
  from
• 16:04
  1-5
• 16:04
  to
• 16:04
  1-100"
• 16:06
  recommendation.