Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 05, 2026

Transcript

AGI Dreams Podcast – June 05, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Meta
• 0:00
  just
• 0:00
  handed
• 0:00
  the
• 0:00
  security
• 0:01
  community
• 0:01
  a
• 0:01
  case
• 0:02
  study
• 0:02
  in
• 0:02
  everything
• 0:03
  not
• 0:03
  to
• 0:03
  do
• 0:03
  with
• 0:03
  AI-powered
• 0:04
  support
• 0:04
  systems.
• 0:05
  Hackers
• 0:05
  hijacked
• 0:06
  high-profile
• 0:06
  Instagram
• 0:07
  accounts
• 0:07
  —
• 0:07
  including
• 0:08
  those
• 0:08
  of
• 0:08
  a
• 0:08
  former
• 0:08
  Obama
• 0:09
  White
• 0:09
  House
• 0:09
  official
• 0:10
  and
• 0:10
  a
• 0:10
  senior
• 0:10
  Space
• 0:10
  Force
• 0:11
  officer
• 0:11
  —
• 0:11
  by
• 0:12
  simply
• 0:12
  asking
• 0:12
  Meta's
• 0:12
  AI
• 0:13
  support
• 0:13
  chatbot
• 0:14
  to
• 0:14
  send
• 0:14
  a
• 0:14
  password-reset
• 0:15
  code
• 0:15
  to
• 0:15
  an
• 0:15
  attacker-controlled
• 0:16
  email
• 0:16
  address.
• 0:17
  No
• 0:17
  exploit
• 0:17
  kit,
• 0:18
  no
• 0:18
  zero-day,
• 0:18
  no
• 0:19
  credential
• 0:19
  stuffing.
• 0:20
  The
• 0:20
  attacker
• 0:20
  VPN'd
• 0:21
  into
• 0:21
  the
• 0:21
  target's
• 0:21
  approximate
• 0:22
  location,
• 0:23
  opened
• 0:23
  a
• 0:23
  conversation
• 0:23
  with
• 0:24
  the
• 0:24
  AI
• 0:24
  chatbot,
• 0:25
  and
• 0:25
  talked
• 0:25
  it
• 0:25
  into
• 0:25
  re-linking
• 0:26
  the
• 0:26
  account
• 0:26
  email.
• 0:27
  The
• 0:27
  chatbot
• 0:27
  —
• 0:28
  which
• 0:28
  Meta
• 0:28
  had
• 0:28
  explicitly
• 0:29
  wired
• 0:29
  to
• 0:29
  MCP
• 0:30
  tools
• 0:30
  capable
• 0:30
  of
• 0:31
  modifying
• 0:31
  account
• 0:31
  credentials
• 0:32
  —
• 0:32
  complied
• 0:33
  without
• 0:33
  meaningful
• 0:33
  challenge.
• 0:34
  As
• 0:34
  the
• 0:34
  security
• 0:35
  researcher
• 0:35
  who
• 0:35
  documented
• 0:36
  the
• 0:36
  attack
• 0:36
  put
• 0:36
  it:
• 0:37
  decades
• 0:38
  of
• 0:38
  cryptographic
• 0:38
  research
• 0:39
  to
• 0:39
  harden
• 0:39
  authentication,
• 0:40
  and
• 0:41
  the
• 0:41
  entire
• 0:41
  chain
• 0:41
  gets
• 0:42
  bypassed
• 0:42
  by
• 0:42
  a
• 0:42
  chatbot
• 0:43
  that's
• 0:43
  "equally
• 0:44
  eager
• 0:44
  to
• 0:44
  help
• 0:44
  and
• 0:44
  vulnerable
• 0:45
  to
• 0:45
  persuasion
• 0:45
  and
• 0:45
  trickery."
• 0:46
  The
• 0:47
  one
• 0:47
  saving
• 0:47
  grace:
• 0:48
  accounts
• 0:48
  with
• 0:49
  even
• 0:49
  basic
• 0:49
  SMS
• 0:50
  two-factor
• 0:50
  authentication
• 0:51
  were
• 0:51
  immune,
• 0:52
  because
• 0:53
  the
• 0:53
  attacker
• 0:53
  couldn't
• 0:53
  intercept
• 0:54
  the
• 0:54
  phone-based
• 0:55
  code
• 0:55
  even
• 0:55
  after
• 0:56
  changing
• 0:56
  the
• 0:56
  email.
• 0:57
  The
• 0:57
  lesson
• 0:58
  isn't
• 0:58
  subtle
• 0:58
  —
• 0:59
  AI
• 0:59
  has
• 0:59
  no
• 0:59
  business
• 1:00
  in
• 1:00
  the
• 1:00
  authentication
• 1:01
  loop
• 1:01
  for
• 1:01
  account
• 1:01
  recovery,
• 1:02
  period
• 1:02
  .
• 1:03
  The
• 1:03
  trust
• 1:04
  problem
• 1:04
  extends
• 1:04
  beyond
• 1:05
  individual
• 1:05
  accounts
• 1:06
  to
• 1:06
  entire
• 1:06
  populations.
• 1:08
  The
• 1:08
  Intercept
• 1:08
  revealed
• 1:09
  that
• 1:09
  "La
• 1:09
  Tilde,"
• 1:10
  a
• 1:10
  slick
• 1:11
  Spanish-English
• 1:12
  news
• 1:12
  site
• 1:12
  targeting
• 1:13
  Latin
• 1:13
  American
• 1:13
  audiences,
• 1:14
  is
• 1:14
  a
• 1:14
  Pentagon
• 1:15
  propaganda
• 1:16
  operation
• 1:16
  run
• 1:17
  by
• 1:17
  U.S.
• 1:17
  Special
• 1:17
  Operations
• 1:18
  Command
• 1:18
  South.
• 1:19
  The
• 1:19
  site
• 1:20
  mixes
• 1:20
  personal
• 1:20
  finance
• 1:21
  tips
• 1:21
  with
• 1:21
  articles
• 1:22
  praising
• 1:22
  U.S.
• 1:23
  military
• 1:23
  operations
• 1:23
  —
• 1:25
  like
• 1:25
  a
• 1:25
  glowing
• 1:25
  account
• 1:26
  of
• 1:26
  the
• 1:26
  Maduro
• 1:26
  abduction
• 1:27
  described
• 1:28
  as
• 1:28
  "a
• 1:28
  military
• 1:29
  operation
• 1:29
  of
• 1:29
  coordination
• 1:30
  and
• 1:30
  accuracy
• 1:31
  never
• 1:31
  seen
• 1:32
  before."
• 1:33
  Most
• 1:33
  of
• 1:33
  the
• 1:33
  content
• 1:34
  appears
• 1:34
  AI-generated:
• 1:35
  detection
• 1:36
  tools
• 1:36
  flagged
• 1:37
  multiple
• 1:37
  articles,
• 1:38
  Midjourney-generated
• 1:39
  images
• 1:40
  carry
• 1:40
  telltale
• 1:41
  prompt
• 1:41
  fragments
• 1:41
  in
• 1:42
  filenames,
• 1:43
  and
• 1:43
  an
• 1:43
  Atlantic
• 1:43
  Council
• 1:44
  researcher
• 1:44
  described
• 1:45
  the
• 1:45
  whole
• 1:45
  project
• 1:45
  as
• 1:46
  "AI
• 1:46
  all
• 1:46
  the
• 1:46
  way
• 1:46
  down."
• 1:47
  Subdomains
• 1:48
  indicate
• 1:49
  planned
• 1:49
  country-specific
• 1:50
  versions
• 1:50
  for
• 1:50
  Ecuador,
• 1:51
  El
• 1:52
  Salvador,
• 1:52
  Guyana,
• 1:53
  Honduras,
• 1:53
  Jamaica,
• 1:54
  Panama,
• 1:54
  and
• 1:55
  Peru.
• 1:55
  The
• 1:56
  easily
• 1:56
  missed
• 1:56
  disclosure
• 1:57
  —
• 1:58
  a
• 1:58
  small
• 1:58
  link
• 1:58
  at
• 1:58
  the
• 1:58
  bottom
• 1:59
  reading
• 1:59
  "publicly
• 2:00
  funded
• 2:00
  from
• 2:00
  the
• 2:00
  budget
• 2:00
  of
• 2:00
  the
• 2:01
  United
• 2:01
  States
• 2:01
  Government"
• 2:01
  —
• 2:03
  is
• 2:03
  technically
• 2:03
  honest
• 2:03
  in
• 2:04
  the
• 2:04
  way
• 2:04
  that
• 2:04
  a
• 2:04
  footnote
• 2:05
  on
• 2:05
  a
• 2:05
  forgery
• 2:05
  is
• 2:05
  technically
• 2:06
  a
• 2:06
  disclaimer.
• 2:07
  The
• 2:07
  design
• 2:08
  was
• 2:08
  subcontracted
• 2:09
  to
• 2:09
  Antpack,
• 2:09
  a
• 2:09
  Colombian
• 2:10
  digital
• 2:10
  marketing
• 2:11
  firm,
• 2:11
  and
• 2:11
  the
• 2:11
  content
• 2:12
  pushes
• 2:12
  a
• 2:12
  consistent
• 2:13
  message:
• 2:14
  that
• 2:14
  U.S.
• 2:14
  military
• 2:15
  partnerships
• 2:15
  strengthen,
• 2:16
  rather
• 2:17
  than
• 2:17
  undermine,
• 2:18
  Latin
• 2:18
  American
• 2:18
  sovereignty
• 2:19
  .
• 2:20
  Meanwhile,
• 2:21
  a
• 2:21
  quieter
• 2:21
  but
• 2:22
  arguably
• 2:22
  more
• 2:22
  consequential
• 2:23
  failure
• 2:23
  is
• 2:23
  emerging
• 2:24
  in
• 2:24
  biometric
• 2:25
  AI.
• 2:26
  Researchers
• 2:26
  at
• 2:26
  Michigan
• 2:27
  State
• 2:27
  systematically
• 2:28
  tested
• 2:28
  GPT-4o
• 2:30
  and
• 2:30
  Gemini-2.5-Flash
• 2:33
  on
• 2:33
  face
• 2:33
  verification
• 2:34
  tasks
• 2:35
  using
• 2:35
  the
• 2:35
  challenging
• 2:35
  IJB-S
• 2:37
  surveillance
• 2:37
  dataset.
• 2:39
  The
• 2:39
  results
• 2:39
  are
• 2:39
  damning:
• 2:40
  even
• 2:40
  when
• 2:40
  the
• 2:40
  models
• 2:41
  produced
• 2:41
  correct
• 2:42
  match/non-match
• 2:44
  decisions,
• 2:45
  their
• 2:45
  accompanying
• 2:45
  textual
• 2:46
  explanations
• 2:47
  frequently
• 2:47
  hallucinated
• 2:48
  facial
• 2:48
  attributes
• 2:49
  that
• 2:49
  were
• 2:49
  physically
• 2:50
  impossible
• 2:50
  to
• 2:50
  verify
• 2:51
  from
• 2:51
  the
• 2:51
  images
• 2:51
  —
• 2:53
  confidently
• 2:53
  describing
• 2:54
  jawline
• 2:54
  shapes
• 2:55
  or
• 2:55
  periocular
• 2:56
  details
• 2:56
  in
• 2:56
  profile
• 2:57
  shots
• 2:57
  where
• 2:58
  those
• 2:58
  features
• 2:58
  were
• 2:58
  simply
• 2:59
  invisible.
• 3:00
  The
• 3:00
  team
• 3:01
  introduced
• 3:01
  a
• 3:01
  likelihood-ratio
• 3:02
  framework
• 3:03
  to
• 3:03
  measure
• 3:03
  explanation
• 3:04
  quality
• 3:04
  independently
• 3:05
  of
• 3:05
  decision
• 3:05
  accuracy,
• 3:06
  and
• 3:07
  found
• 3:07
  that
• 3:07
  feeding
• 3:07
  models
• 3:08
  similarity
• 3:08
  scores
• 3:09
  from
• 3:09
  traditional
• 3:09
  face-recognition
• 3:10
  systems
• 3:11
  improved
• 3:11
  their
• 3:12
  categorical
• 3:12
  accuracy
• 3:13
  but
• 3:13
  did
• 3:13
  nothing
• 3:14
  to
• 3:14
  fix
• 3:14
  the
• 3:14
  hallucinated
• 3:15
  explanations.
• 3:16
  A
• 3:16
  commercial
• 3:17
  off-the-shelf
• 3:17
  system
• 3:18
  achieved
• 3:18
  near-perfect
• 3:19
  verification
• 3:19
  accuracy
• 3:20
  but
• 3:20
  offers
• 3:20
  zero
• 3:21
  transparency.
• 3:22
  The
• 3:22
  trade-off
• 3:23
  between
• 3:23
  accuracy
• 3:24
  and
• 3:24
  explainability
• 3:25
  in
• 3:25
  biometric
• 3:25
  AI
• 3:26
  remains
• 3:26
  unresolved,
• 3:27
  and
• 3:27
  that
• 3:27
  matters
• 3:28
  enormously
• 3:28
  for
• 3:28
  forensic
• 3:29
  and
• 3:29
  legal
• 3:29
  contexts
• 3:30
  where
• 3:30
  a
• 3:30
  natural-language
• 3:31
  explanation
• 3:31
  might
• 3:32
  be
• 3:32
  treated
• 3:32
  as
• 3:32
  evidence
• 3:32
  .
• 3:34
  Anthropic
• 3:35
  published
• 3:35
  what
• 3:35
  amounts
• 3:36
  to
• 3:36
  a
• 3:36
  field
• 3:36
  manual
• 3:36
  for
• 3:37
  deploying
• 3:37
  Claude
• 3:37
  Code
• 3:37
  in
• 3:38
  enterprise
• 3:38
  codebases
• 3:38
  —
• 3:39
  multi-million-line
• 3:40
  monorepos,
• 3:41
  legacy
• 3:41
  systems
• 3:41
  spanning
• 3:42
  decades,
• 3:43
  distributed
• 3:43
  architectures
• 3:44
  across
• 3:44
  dozens
• 3:44
  of
• 3:44
  repositories.
• 3:46
  The
• 3:46
  core
• 3:46
  thesis:
• 3:47
  the
• 3:47
  harness
• 3:47
  matters
• 3:47
  as
• 3:48
  much
• 3:48
  as
• 3:48
  the
• 3:48
  model.
• 3:49
  Claude
• 3:49
  Code
• 3:50
  uses
• 3:50
  "agenetic
• 3:51
  search"
• 3:51
  rather
• 3:52
  than
• 3:52
  pre-built
• 3:52
  indexes
• 3:53
  —
• 3:54
  navigating
• 3:54
  codebases
• 3:55
  the
• 3:55
  way
• 3:55
  a
• 3:55
  developer
• 3:56
  would,
• 3:56
  with
• 3:56
  grep
• 3:57
  and
• 3:57
  directory
• 3:57
  traversal
• 3:58
  rather
• 3:58
  than
• 3:58
  semantic
• 3:59
  search.
• 3:59
  The
• 3:59
  trade-off
• 4:00
  is
• 4:00
  that
• 4:00
  it
• 4:00
  works
• 4:00
  best
• 4:01
  when
• 4:01
  it
• 4:01
  has
• 4:01
  curated
• 4:02
  starting
• 4:02
  context.
• 4:03
  The
• 4:03
  practical
• 4:04
  strategies
• 4:04
  center
• 4:05
  on
• 4:05
  layered
• 4:05
  CLAUDE.md
• 4:06
  files
• 4:06
  (lean
• 4:07
  global
• 4:07
  rules
• 4:08
  at
• 4:08
  the
• 4:08
  repo
• 4:08
  root,
• 4:09
  subdirectory-specific
• 4:10
  conventions
• 4:10
  loaded
• 4:10
  progressively),
• 4:11
  scoped
• 4:12
  skills
• 4:12
  that
• 4:12
  activate
• 4:13
  only
• 4:13
  when
• 4:13
  editing
• 4:13
  relevant
• 4:14
  paths,
• 4:14
  LSP-backed
• 4:15
  MCP
• 4:16
  servers
• 4:16
  for
• 4:16
  symbol-level
• 4:17
  search
• 4:18
  rather
• 4:18
  than
• 4:18
  string
• 4:18
  matching,
• 4:19
  and
• 4:19
  self-improving
• 4:20
  hooks
• 4:20
  —
• 4:21
  a
• 4:21
  stop
• 4:21
  hook
• 4:21
  that
• 4:21
  runs
• 4:21
  a
• 4:21
  headless
• 4:22
  Claude
• 4:22
  session
• 4:22
  to
• 4:22
  propose
• 4:23
  rule-file
• 4:23
  updates
• 4:24
  after
• 4:24
  every
• 4:24
  task.
• 4:25
  The
• 4:25
  most
• 4:26
  actionable
• 4:26
  advice:
• 4:27
  assign
• 4:27
  a
• 4:27
  small
• 4:28
  team
• 4:28
  to
• 4:28
  build
• 4:28
  the
• 4:29
  "AI
• 4:29
  layer"
• 4:29
  before
• 4:29
  rolling
• 4:30
  it
• 4:30
  out
• 4:30
  organization-wide,
• 4:31
  because
• 4:31
  inconsistent
• 4:32
  individual
• 4:33
  setups
• 4:33
  produce
• 4:33
  inconsistent
• 4:34
  results
• 4:34
  .
• 4:35
  One
• 4:35
  piece
• 4:36
  conspicuously
• 4:36
  absent
• 4:37
  from
• 4:37
  Anthropic's
• 4:37
  playbook
• 4:38
  is
• 4:38
  context
• 4:38
  compression
• 4:39
  —
• 4:40
  which
• 4:40
  is
• 4:40
  exactly
• 4:40
  the
• 4:40
  gap
• 4:41
  Headroom
• 4:41
  fills.
• 4:42
  This
• 4:42
  open-source
• 4:43
  library
• 4:43
  sits
• 4:43
  between
• 4:44
  your
• 4:44
  agent
• 4:44
  and
• 4:44
  the
• 4:44
  LLM
• 4:45
  provider,
• 4:45
  compressing
• 4:46
  tool
• 4:46
  outputs,
• 4:47
  RAG
• 4:47
  chunks,
• 4:48
  logs,
• 4:48
  and
• 4:48
  conversation
• 4:49
  history
• 4:49
  before
• 4:50
  they
• 4:50
  reach
• 4:50
  the
• 4:50
  model.
• 4:51
  It
• 4:51
  ships
• 4:51
  six
• 4:52
  algorithms:
• 4:53
  SmartCrusher
• 4:53
  for
• 4:53
  JSON,
• 4:54
  CodeCompressor
• 4:55
  for
• 4:55
  AST-aware
• 4:55
  code
• 4:55
  compression,
• 4:56
  and
• 4:56
  Kompress-base
• 4:57
  (a
• 4:57
  HuggingFace
• 4:58
  model
• 4:58
  trained
• 4:59
  on
• 4:59
  agentic
• 4:59
  traces)
• 4:59
  for
• 5:00
  prose.
• 5:00
  Claimed
• 5:01
  savings
• 5:01
  range
• 5:01
  from
• 5:02
  47%
• 5:03
  on
• 5:03
  codebase
• 5:03
  exploration
• 5:04
  to
• 5:04
  92%
• 5:05
  on
• 5:05
  code
• 5:05
  search
• 5:06
  results,
• 5:06
  with
• 5:06
  benchmark
• 5:07
  accuracy
• 5:07
  preserved
• 5:08
  within
• 5:09
  ±0.03
• 5:09
  on
• 5:09
  GSM8K
• 5:11
  and
• 5:11
  TruthfulQA.
• 5:12
  The
• 5:12
  critical
• 5:13
  differentiator
• 5:13
  is
• 5:13
  CCR
• 5:14
  —
• 5:15
  Compressed
• 5:15
  Context
• 5:16
  Retrieval
• 5:16
  —
• 5:16
  which
• 5:17
  stores
• 5:17
  originals
• 5:17
  locally
• 5:18
  so
• 5:18
  the
• 5:18
  LLM
• 5:19
  can
• 5:19
  request
• 5:19
  the
• 5:19
  uncompressed
• 5:20
  version
• 5:20
  on
• 5:20
  demand.
• 5:21
  It
• 5:21
  wraps
• 5:21
  Claude
• 5:22
  Code,
• 5:22
  Codex,
• 5:23
  Cursor,
• 5:23
  and
• 5:23
  Aider
• 5:24
  with
• 5:24
  a
• 5:24
  single
• 5:24
  command,
• 5:25
  and
• 5:25
  shares
• 5:25
  compressed
• 5:26
  context
• 5:26
  across
• 5:26
  agents
• 5:27
  via
• 5:27
  a
• 5:27
  cross-agent
• 5:28
  memory
• 5:28
  store.
• 5:29
  For
• 5:29
  teams
• 5:29
  burning
• 5:29
  through
• 5:30
  token
• 5:30
  budgets
• 5:30
  on
• 5:30
  large
• 5:31
  codebases,
• 5:32
  this
• 5:32
  is
• 5:32
  the
• 5:32
  missing
• 5:32
  infrastructure
• 5:33
  layer
• 5:33
  .
• 5:33
  The
• 5:34
  question
• 5:34
  of
• 5:34
  which
• 5:34
  coding
• 5:34
  agent
• 5:35
  to
• 5:35
  use
• 5:35
  remains
• 5:35
  unsettled.
• 5:37
  A
• 5:37
  short
• 5:37
  take
• 5:37
  on
• 5:37
  Claude
• 5:37
  Code
• 5:38
  versus
• 5:38
  Codex
• 5:38
  argues
• 5:39
  the
• 5:39
  answer
• 5:39
  is
• 5:39
  "both"
• 5:40
  —
• 5:40
  not
• 5:41
  out
• 5:41
  of
• 5:41
  diplomatic
• 5:41
  balance
• 5:42
  but
• 5:42
  out
• 5:42
  of
• 5:42
  vendor
• 5:42
  resilience.
• 5:44
  GPT
• 5:44
  5.5
• 5:45
  matches
• 5:45
  Opus
• 5:45
  4.7
• 5:46
  on
• 5:46
  capability,
• 5:47
  Codex's
• 5:48
  desktop
• 5:48
  app
• 5:48
  is
• 5:48
  polished,
• 5:49
  and
• 5:50
  Anthropic's
• 5:50
  recent
• 5:50
  billing
• 5:51
  changes
• 5:51
  for
• 5:51
  programmatic
• 5:52
  Claude
• 5:52
  Code
• 5:52
  usage
• 5:53
  rankled
• 5:53
  power
• 5:53
  users.
• 5:54
  The
• 5:54
  pragmatic
• 5:55
  move:
• 5:56
  run
• 5:56
  Codex
• 5:56
  as
• 5:57
  the
• 5:57
  desktop
• 5:57
  environment
• 5:58
  with
• 5:58
  Claude
• 5:58
  Code
• 5:58
  in
• 5:58
  the
• 5:58
  terminal
• 5:59
  inside
• 5:59
  it,
• 6:00
  and
• 6:00
  refuse
• 6:00
  to
• 6:00
  become
• 6:00
  a
• 6:01
  prisoner
• 6:01
  of
• 6:01
  any
• 6:01
  single
• 6:02
  vendor
• 6:02
  .
• 6:02
  For
• 6:02
  mobile
• 6:03
  developers
• 6:03
  drowning
• 6:04
  in
• 6:04
  AI-generated
• 6:04
  spaghetti,
• 6:05
  a
• 6:05
  storyboard-based
• 6:06
  workflow
• 6:07
  offers
• 6:07
  a
• 6:07
  structural
• 6:07
  fix:
• 6:08
  build
• 6:08
  a
• 6:08
  design-system
• 6:09
  gallery
• 6:09
  first
• 6:10
  (primitives
• 6:10
  like
• 6:11
  buttons,
• 6:11
  inputs,
• 6:12
  badges),
• 6:12
  then
• 6:13
  construct
• 6:13
  storyboard
• 6:14
  tabs
• 6:14
  that
• 6:14
  walk
• 6:14
  through
• 6:15
  every
• 6:15
  user
• 6:15
  path
• 6:15
  screen-by-screen.
• 6:17
  The
• 6:17
  storyboard
• 6:18
  components
• 6:18
  reuse
• 6:18
  gallery
• 6:19
  primitives,
• 6:20
  so
• 6:20
  changes
• 6:20
  propagate
• 6:21
  everywhere
• 6:21
  —
• 6:22
  and
• 6:22
  when
• 6:22
  it's
• 6:22
  time
• 6:22
  to
• 6:22
  wire
• 6:23
  the
• 6:23
  backend,
• 6:23
  you're
• 6:23
  pointing
• 6:24
  APIs
• 6:24
  at
• 6:24
  well-defined
• 6:25
  screens
• 6:25
  rather
• 6:26
  than
• 6:26
  hunting
• 6:26
  through
• 6:26
  fragmented
• 6:27
  views
• 6:27
  .
• 6:28
  And
• 6:28
  for
• 6:28
  anyone
• 6:29
  who
• 6:29
  wants
• 6:29
  to
• 6:29
  understand
• 6:30
  the
• 6:30
  ReAct
• 6:30
  agent
• 6:30
  pattern
• 6:31
  without
• 6:31
  SDK
• 6:31
  abstractions,
• 6:33
  a
• 6:33
  minimal
• 6:33
  four-file
• 6:34
  Python
• 6:34
  implementation
• 6:35
  running
• 6:35
  Llama
• 6:35
  3.2
• 6:36
  3B
• 6:36
  on
• 6:37
  Ollama
• 6:37
  demonstrates
• 6:38
  the
• 6:38
  entire
• 6:38
  loop
• 6:39
  —
• 6:40
  system
• 6:40
  prompt
• 6:40
  generation
• 6:41
  from
• 6:41
  a
• 6:41
  tool
• 6:41
  registry,
• 6:42
  stop-sequence-based
• 6:43
  action
• 6:44
  parsing,
• 6:44
  and
• 6:44
  the
• 6:45
  critical
• 6:45
  distinction
• 6:45
  between
• 6:46
  the
• 6:46
  outer
• 6:46
  conversational
• 6:47
  loop
• 6:47
  and
• 6:47
  the
• 6:47
  inner
• 6:48
  reasoning-action
• 6:48
  loop
• 6:48
  —
• 6:49
  in
• 6:49
  under
• 6:50
  300
• 6:50
  lines
• 6:50
  .
• 6:53
  Two
• 6:53
  research
• 6:53
  papers
• 6:54
  tackle
• 6:54
  the
• 6:54
  same
• 6:54
  meta-problem
• 6:55
  —
• 6:56
  how
• 6:56
  gradient-based
• 6:56
  training
• 6:57
  distributes
• 6:57
  learning
• 6:57
  signal
• 6:58
  across
• 6:58
  steps
• 6:58
  —
• 6:59
  from
• 6:59
  very
• 6:59
  different
• 7:00
  angles.
• 7:00
  AEM
• 7:01
  (Adaptive
• 7:02
  Entropy
• 7:02
  Modulation),
• 7:03
  from
• 7:03
  Baidu
• 7:03
  and
• 7:04
  Tsinghua,
• 7:04
  addresses
• 7:05
  a
• 7:05
  fundamental
• 7:05
  weakness
• 7:06
  in
• 7:06
  agentic
• 7:06
  reinforcement
• 7:07
  learning:
• 7:07
  when
• 7:07
  an
• 7:08
  agent
• 7:08
  takes
• 7:08
  a
• 7:08
  dozen
• 7:08
  actions
• 7:09
  to
• 7:09
  solve
• 7:09
  a
• 7:09
  task
• 7:10
  but
• 7:10
  only
• 7:10
  receives
• 7:10
  a
• 7:10
  binary
• 7:11
  pass/fail
• 7:12
  reward
• 7:12
  at
• 7:12
  the
• 7:12
  end,
• 7:13
  how
• 7:13
  do
• 7:13
  you
• 7:13
  figure
• 7:13
  out
• 7:14
  which
• 7:14
  steps
• 7:14
  mattered?
• 7:15
  Existing
• 7:16
  approaches
• 7:16
  either
• 7:16
  require
• 7:17
  expensive
• 7:17
  auxiliary
• 7:17
  models
• 7:18
  (process
• 7:18
  reward
• 7:19
  models)
• 7:19
  or
• 7:20
  impose
• 7:20
  restrictive
• 7:20
  structural
• 7:21
  assumptions.
• 7:22
  AEM's
• 7:23
  insight
• 7:23
  is
• 7:23
  to
• 7:23
  use
• 7:24
  response-level
• 7:24
  entropy
• 7:24
  —
• 7:25
  the
• 7:25
  model's
• 7:26
  own
• 7:26
  uncertainty
• 7:26
  about
• 7:27
  each
• 7:27
  action
• 7:27
  —
• 7:28
  as
• 7:28
  a
• 7:28
  free
• 7:28
  credit-assignment
• 7:29
  signal.
• 7:30
  The
• 7:30
  theory
• 7:30
  shows
• 7:31
  that
• 7:31
  entropy
• 7:31
  drift
• 7:32
  during
• 7:32
  training
• 7:33
  is
• 7:33
  governed
• 7:33
  by
• 7:33
  the
• 7:33
  product
• 7:34
  of
• 7:34
  the
• 7:34
  response
• 7:34
  advantage
• 7:35
  and
• 7:35
  its
• 7:35
  relative
• 7:35
  surprisal.
• 7:37
  In
• 7:37
  practice,
• 7:38
  AEM
• 7:38
  computes
• 7:38
  a
• 7:38
  per-response
• 7:39
  modulation
• 7:39
  coefficient
• 7:40
  that
• 7:41
  upweights
• 7:41
  low-uncertainty
• 7:42
  positive
• 7:42
  actions
• 7:43
  and
• 7:43
  downweights
• 7:43
  high-uncertainty
• 7:44
  negative
• 7:44
  ones,
• 7:45
  naturally
• 7:45
  transitioning
• 7:46
  from
• 7:46
  exploration
• 7:46
  (early
• 7:47
  training,
• 7:48
  mostly
• 7:48
  negative
• 7:48
  samples)
• 7:49
  to
• 7:49
  exploitation
• 7:50
  (late
• 7:50
  training,
• 7:51
  mostly
• 7:51
  positive
• 7:51
  samples).
• 7:52
  The
• 7:52
  overhead
• 7:53
  is
• 7:53
  negligible
• 7:54
  —
• 7:54
  1.1%
• 7:55
  of
• 7:55
  total
• 7:55
  training
• 7:55
  time
• 7:55
  —
• 7:56
  because
• 7:56
  it
• 7:56
  piggybacks
• 7:57
  on
• 7:57
  the
• 7:57
  log-probability
• 7:58
  computation
• 7:58
  pass.
• 7:59
  Results:
• 8:00
  8.8%
• 8:01
  improvement
• 8:01
  on
• 8:01
  ALFWorld,
• 8:02
  5.6%
• 8:03
  on
• 8:03
  WebShop,
• 8:04
  and
• 8:04
  a
• 8:04
  1.4-point
• 8:04
  gain
• 8:05
  on
• 8:05
  the
• 8:05
  notoriously
• 8:05
  difficult
• 8:06
  SWE-bench-Verified
• 8:08
  when
• 8:08
  layered
• 8:08
  on
• 8:08
  top
• 8:08
  of
• 8:08
  the
• 8:09
  state-of-the-art
• 8:09
  DeepSWE
• 8:09
  baseline
• 8:10
  with
• 8:10
  Qwen3-32B
• 8:11
  .
• 8:13
  The
• 8:13
  shortcut
• 8:13
  learning
• 8:14
  paper
• 8:14
  from
• 8:14
  the
• 8:14
  Chinese
• 8:15
  Academy
• 8:15
  of
• 8:15
  Sciences
• 8:16
  takes
• 8:16
  a
• 8:16
  more
• 8:16
  theoretical
• 8:17
  stance,
• 8:18
  asking
• 8:18
  why
• 8:18
  networks
• 8:18
  learn
• 8:19
  texture
• 8:19
  instead
• 8:19
  of
• 8:19
  shape,
• 8:20
  color
• 8:20
  instead
• 8:21
  of
• 8:21
  digits.
• 8:22
  The
• 8:22
  authors
• 8:22
  formalize
• 8:23
  "shortcut"
• 8:23
  and
• 8:24
  "core"
• 8:24
  features
• 8:24
  with
• 8:24
  precise
• 8:25
  definitions
• 8:25
  —
• 8:26
  a
• 8:26
  shortcut
• 8:27
  is
• 8:27
  any
• 8:27
  feature
• 8:27
  that
• 8:27
  achieves
• 8:28
  good
• 8:28
  in-distribution
• 8:29
  accuracy
• 8:30
  but
• 8:30
  fails
• 8:30
  under
• 8:30
  distribution
• 8:31
  shift,
• 8:32
  while
• 8:32
  core
• 8:32
  features
• 8:33
  are
• 8:33
  unique
• 8:33
  and
• 8:33
  undominated
• 8:34
  under
• 8:34
  pairwise
• 8:35
  comparison.
• 8:36
  The
• 8:36
  novel
• 8:36
  move
• 8:36
  is
• 8:37
  modeling
• 8:37
  training
• 8:37
  as
• 8:38
  an
• 8:38
  evolutionary
• 8:38
  game:
• 8:39
  each
• 8:39
  data
• 8:40
  sample
• 8:40
  is
• 8:40
  a
• 8:40
  player,
• 8:41
  its
• 8:41
  neural
• 8:41
  tangent
• 8:42
  feature
• 8:42
  is
• 8:42
  its
• 8:42
  strategy,
• 8:43
  and
• 8:43
  the
• 8:43
  payoff
• 8:44
  is
• 8:44
  the
• 8:44
  loss
• 8:44
  reduction
• 8:45
  from
• 8:45
  alignment
• 8:45
  with
• 8:46
  the
• 8:46
  batch
• 8:46
  gradient.
• 8:47
  Under
• 8:47
  full-batch
• 8:48
  gradient
• 8:48
  descent,
• 8:49
  the
• 8:49
  stochastically
• 8:50
  stable
• 8:50
  strategy
• 8:51
  converges
• 8:51
  to
• 8:51
  the
• 8:51
  shortcut
• 8:52
  subnetwork.
• 8:53
  Under
• 8:53
  mini-batch
• 8:54
  SGD,
• 8:55
  the
• 8:55
  stochastic
• 8:55
  noise
• 8:56
  from
• 8:56
  random
• 8:56
  grouping
• 8:57
  flips
• 8:57
  the
• 8:57
  stability
• 8:57
  —
• 8:58
  the
• 8:58
  core
• 8:59
  strategy
• 8:59
  wins.
• 9:00
  The
• 9:00
  continuous-time
• 9:01
  SDE
• 9:02
  analysis
• 9:02
  further
• 9:02
  shows
• 9:03
  that
• 9:03
  data
• 9:03
  noise
• 9:04
  amplifies
• 9:04
  shortcut
• 9:05
  bias
• 9:05
  while
• 9:05
  optimization
• 9:06
  noise
• 9:06
  (smaller
• 9:07
  batches,
• 9:07
  larger
• 9:08
  learning
• 9:08
  rates)
• 9:09
  suppresses
• 9:09
  it.
• 9:10
  This
• 9:10
  is
• 9:10
  a
• 9:10
  clean
• 9:11
  theoretical
• 9:11
  explanation
• 9:12
  for
• 9:12
  a
• 9:12
  well-known
• 9:12
  empirical
• 9:13
  fact
• 9:13
  —
• 9:14
  SGD
• 9:15
  generalizes
• 9:15
  better
• 9:16
  than
• 9:16
  GD
• 9:16
  —
• 9:17
  and
• 9:17
  it
• 9:17
  points
• 9:17
  toward
• 9:18
  practical
• 9:18
  mitigation:
• 9:19
  disentangled
• 9:20
  representations,
• 9:21
  data
• 9:22
  augmentation
• 9:22
  targeting
• 9:23
  the
• 9:23
  conflict
• 9:23
  set,
• 9:24
  and
• 9:24
  careful
• 9:24
  noise
• 9:25
  calibration
• 9:25
  .
• 9:29
  JD's
• 9:30
  JoyAI-Echo
• 9:31
  pushes
• 9:31
  long-form
• 9:32
  video
• 9:32
  generation
• 9:32
  past
• 9:33
  the
• 9:33
  minute
• 9:33
  mark
• 9:33
  with
• 9:33
  a
• 9:33
  framework
• 9:34
  built
• 9:34
  on
• 9:34
  four
• 9:34
  advances:
• 9:35
  a
• 9:35
  cross-modal
• 9:36
  audio-visual
• 9:37
  memory
• 9:37
  bank
• 9:37
  that
• 9:38
  preserves
• 9:38
  character
• 9:38
  appearance
• 9:39
  and
• 9:39
  voice
• 9:39
  timbre
• 9:40
  across
• 9:40
  shots,
• 9:41
  a
• 9:41
  distribution-matching
• 9:42
  distillation
• 9:43
  (DMD)
• 9:44
  generator
• 9:44
  delivering
• 9:45
  7.5x
• 9:46
  speedup
• 9:46
  over
• 9:46
  multi-step
• 9:47
  diffusion,
• 9:48
  joint
• 9:48
  audio-video
• 9:49
  synthesis
• 9:49
  in
• 9:49
  a
• 9:49
  single
• 9:50
  pipeline,
• 9:50
  and
• 9:51
  an
• 9:51
  interactive
• 9:51
  agent
• 9:52
  for
• 9:52
  real-time
• 9:52
  conversational
• 9:53
  editing.
• 9:54
  Human
• 9:54
  evaluations
• 9:55
  show
• 9:55
  JoyAI-Echo
• 9:56
  decisively
• 9:56
  outperforming
• 9:57
  HappyOyster's
• 9:58
  directing
• 9:58
  mode
• 9:58
  on
• 9:58
  long-form
• 9:59
  generation
• 10:00
  —
• 10:00
  63.6%
• 10:01
  preference
• 10:02
  on
• 10:02
  visual
• 10:02
  aesthetics,
• 10:03
  81.7%
• 10:04
  on
• 10:04
  audio
• 10:05
  quality,
• 10:05
  80.6%
• 10:06
  on
• 10:06
  prompt
• 10:07
  following
• 10:07
  —
• 10:08
  and
• 10:08
  beating
• 10:08
  the
• 10:08
  short-video
• 10:09
  specialist
• 10:09
  Wan
• 10:09
  2.6
• 10:10
  on
• 10:10
  visual
• 10:10
  aesthetics
• 10:11
  (58.8%).
• 10:13
  The
• 10:13
  system
• 10:13
  runs
• 10:14
  on
• 10:14
  a
• 10:14
  single
• 10:14
  H100/A100
• 10:18
  at
• 10:18
  46–50
• 10:19
  GB
• 10:20
  peak
• 10:20
  VRAM,
• 10:20
  with
• 10:20
  inference
• 10:21
  code
• 10:21
  and
• 10:21
  a
• 10:21
  46
• 10:22
  GB
• 10:22
  checkpoint
• 10:23
  now
• 10:23
  open-sourced
• 10:23
  for
• 10:24
  academic
• 10:24
  use.
• 10:25
  The
• 10:25
  paired
• 10:25
  cross-modal
• 10:26
  memory
• 10:26
  bank
• 10:26
  is
• 10:26
  the
• 10:27
  key
• 10:27
  architectural
• 10:27
  contribution:
• 10:28
  each
• 10:29
  new
• 10:29
  shot
• 10:29
  is
• 10:29
  conditioned
• 10:30
  on
• 10:30
  prior
• 10:30
  visual
• 10:30
  identity
• 10:31
  and
• 10:31
  voice
• 10:31
  context,
• 10:32
  which
• 10:32
  is
• 10:32
  what
• 10:32
  enables
• 10:33
  story-level
• 10:33
  consistency
• 10:34
  over
• 10:34
  five
• 10:34
  minutes
• 10:35
  of
• 10:35
  multi-shot
• 10:35
  narrative
• 10:36
  .
• 10:36
  On
• 10:37
  the
• 10:37
  alignment
• 10:37
  side,
• 10:38
  a
• 10:38
  Columbia
• 10:38
  University
• 10:39
  team
• 10:39
  reimagines
• 10:40
  how
• 10:40
  to
• 10:40
  fine-tune
• 10:41
  flow-based
• 10:41
  generative
• 10:41
  models
• 10:42
  (the
• 10:42
  architecture
• 10:43
  behind
• 10:43
  Stable
• 10:44
  Diffusion
• 10:44
  3
• 10:44
  and
• 10:44
  FLUX)
• 10:45
  for
• 10:46
  human
• 10:46
  preferences.
• 10:47
  Their
• 10:47
  deterministic
• 10:48
  adjoint
• 10:48
  matching
• 10:49
  framework
• 10:49
  formulates
• 10:49
  alignment
• 10:50
  as
• 10:50
  an
• 10:50
  optimal
• 10:51
  control
• 10:51
  problem
• 10:52
  over
• 10:52
  velocity
• 10:52
  fields,
• 10:53
  with
• 10:53
  a
• 10:53
  truncated
• 10:54
  adjoint
• 10:54
  scheme
• 10:54
  that
• 10:55
  concentrates
• 10:55
  computation
• 10:56
  on
• 10:56
  the
• 10:56
  final
• 10:56
  denoising
• 10:57
  steps
• 10:57
  where
• 10:57
  reward-relevant
• 10:58
  signals
• 10:58
  are
• 10:58
  strongest.
• 11:00
  The
• 11:00
  practical
• 11:00
  payoff
• 11:01
  is
• 11:01
  dramatic:
• 11:02
  wall-clock
• 11:02
  time
• 11:03
  per
• 11:03
  update
• 11:03
  drops
• 11:03
  from
• 11:04
  345
• 11:05
  seconds
• 11:05
  to
• 11:05
  32
• 11:06
  seconds
• 11:06
  on
• 11:06
  FLUX.2-Klein-4B,
• 11:10
  while
• 11:10
  image
• 11:10
  quality
• 11:10
  improves
• 11:11
  over
• 11:11
  the
• 11:11
  base
• 11:11
  model
• 11:12
  across
• 11:12
  HPSv2,
• 11:13
  ImageReward,
• 11:14
  Aesthetic
• 11:14
  Score,
• 11:14
  and
• 11:15
  PickScore.
• 11:16
  Critically,
• 11:17
  their
• 11:17
  higher-order
• 11:17
  polynomial
• 11:18
  regularization
• 11:19
  (beyond
• 11:20
  the
• 11:20
  standard
• 11:20
  KL)
• 11:21
  preserves
• 11:21
  within-prompt
• 11:22
  diversity
• 11:23
  and
• 11:23
  prevents
• 11:24
  the
• 11:24
  mode
• 11:24
  collapse
• 11:24
  that
• 11:24
  plagues
• 11:25
  direct
• 11:25
  reward-backpropagation
• 11:26
  methods
• 11:27
  like
• 11:27
  DRaFT
• 11:27
  and
• 11:27
  ReFL.
• 11:29
  The
• 11:29
  theory
• 11:29
  is
• 11:29
  elegant
• 11:30
  —
• 11:30
  truncation
• 11:31
  works
• 11:31
  because
• 11:32
  control
• 11:32
  intensity
• 11:32
  peaks
• 11:33
  near
• 11:33
  the
• 11:33
  end
• 11:33
  of
• 11:33
  the
• 11:33
  denoising
• 11:34
  trajectory,
• 11:35
  and
• 11:35
  higher-order
• 11:36
  norms
• 11:36
  flatten
• 11:36
  the
• 11:36
  control
• 11:37
  profile
• 11:37
  so
• 11:38
  neighboring
• 11:38
  timesteps
• 11:39
  remain
• 11:39
  effective
• 11:39
  .
• 11:40
  In
• 11:40
  model
• 11:40
  releases,
• 11:41
  Unsloth
• 11:42
  published
• 11:42
  GGUF
• 11:42
  quantizations
• 11:43
  of
• 11:43
  NVIDIA's
• 11:44
  Nemotron-3-Super-120B-A12B,
• 11:48
  a
• 11:48
  mixture-of-experts
• 11:49
  model
• 11:49
  activating
• 11:50
  only
• 11:50
  12
• 11:50
  billion
• 11:51
  of
• 11:51
  its
• 11:51
  120
• 11:52
  billion
• 11:52
  parameters
• 11:52
  per
• 11:53
  forward
• 11:53
  pass
• 11:53
  ,
• 11:54
  and
• 11:54
  Tencent
• 11:54
  dropped
• 11:55
  a
• 11:55
  preview
• 11:55
  of
• 11:55
  Hy3,
• 11:56
  its
• 11:56
  next-generation
• 11:57
  media
• 11:57
  model
• 11:57
  .
• 12:00
  ARMOR
• 12:01
  (Authenticated
• 12:02
  Range-readable
• 12:03
  Managed
• 12:03
  Object
• 12:03
  Repository)
• 12:04
  is
• 12:04
  an
• 12:05
  S3-compatible
• 12:06
  proxy
• 12:06
  server
• 12:06
  written
• 12:07
  in
• 12:07
  Go
• 12:07
  that
• 12:07
  transparently
• 12:08
  encrypts
• 12:08
  data
• 12:08
  before
• 12:09
  it
• 12:09
  reaches
• 12:09
  Backblaze
• 12:09
  B2,
• 12:10
  achieving
• 12:11
  zero-knowledge
• 12:11
  storage
• 12:12
  at
• 12:12
  roughly
• 12:12
  $6–7
• 12:13
  per
• 12:14
  terabyte
• 12:14
  per
• 12:14
  month
• 12:14
  with
• 12:15
  zero
• 12:15
  egress
• 12:15
  fees
• 12:15
  via
• 12:16
  Cloudflare's
• 12:16
  Bandwidth
• 12:16
  Alliance.
• 12:17
  The
• 12:18
  encryption
• 12:18
  design
• 12:18
  uses
• 12:19
  AES-256-CTR
• 12:22
  with
• 12:22
  64KB
• 12:23
  blocks
• 12:23
  and
• 12:23
  per-block
• 12:24
  HMAC-SHA256,
• 12:29
  wrapped
• 12:29
  in
• 12:29
  an
• 12:29
  envelope
• 12:30
  scheme
• 12:30
  where
• 12:30
  each
• 12:30
  file
• 12:30
  gets
• 12:30
  a
• 12:31
  random
• 12:31
  data
• 12:31
  encryption
• 12:32
  key
• 12:32
  (DEK)
• 12:33
  that's
• 12:33
  itself
• 12:34
  encrypted
• 12:34
  by
• 12:34
  a
• 12:34
  master
• 12:35
  key
• 12:35
  stored
• 12:35
  locally.
• 12:36
  The
• 12:37
  critical
• 12:37
  feature
• 12:37
  is
• 12:37
  seekable
• 12:38
  encryption:
• 12:39
  because
• 12:39
  AES-CTR
• 12:41
  is
• 12:41
  a
• 12:41
  stream
• 12:41
  cipher,
• 12:42
  DuckDB
• 12:43
  can
• 12:43
  query
• 12:43
  encrypted
• 12:43
  Parquet
• 12:44
  files
• 12:44
  with
• 12:44
  column
• 12:45
  pruning
• 12:45
  and
• 12:45
  predicate
• 12:46
  pushdown
• 12:46
  intact
• 12:46
  —
• 12:47
  reading
• 12:48
  only
• 12:48
  the
• 12:48
  blocks
• 12:48
  it
• 12:48
  needs,
• 12:49
  decrypting
• 12:49
  on
• 12:49
  the
• 12:50
  fly.
• 12:50
  Upload
• 12:51
  goes
• 12:51
  direct
• 12:51
  to
• 12:51
  B2
• 12:52
  (ingress
• 12:52
  is
• 12:53
  free),
• 12:53
  download
• 12:54
  routes
• 12:54
  through
• 12:54
  Cloudflare's
• 12:55
  edge
• 12:55
  cache.
• 12:56
  Multi-key
• 12:56
  support
• 12:56
  routes
• 12:57
  different
• 12:57
  prefixes
• 12:58
  to
• 12:58
  different
• 12:58
  master
• 12:58
  keys,
• 12:59
  and
• 12:59
  key
• 13:00
  rotation
• 13:00
  re-wraps
• 13:01
  DEKs
• 13:01
  via
• 13:01
  metadata-only
• 13:02
  copies
• 13:02
  without
• 13:03
  re-uploading
• 13:03
  data.
• 13:04
  The
• 13:05
  admin
• 13:05
  API
• 13:05
  includes
• 13:06
  a
• 13:06
  self-healing
• 13:06
  canary
• 13:07
  monitor,
• 13:07
  cryptographic
• 13:08
  provenance
• 13:09
  chains,
• 13:09
  and
• 13:09
  pre-signed
• 13:10
  URL
• 13:10
  sharing.
• 13:11
  For
• 13:11
  anyone
• 13:12
  storing
• 13:12
  sensitive
• 13:12
  data
• 13:13
  in
• 13:13
  object
• 13:13
  storage
• 13:13
  —
• 13:14
  which
• 13:14
  is
• 13:14
  increasingly
• 13:15
  everyone
• 13:15
  —
• 13:16
  this
• 13:16
  is
• 13:16
  a
• 13:16
  compelling
• 13:17
  architecture
• 13:17
  .
• 13:18
  On
• 13:18
  the
• 13:18
  offensive-testing
• 13:19
  side,
• 13:19
  WAF
• 13:20
  Killer
• 13:20
  is
• 13:20
  a
• 13:20
  Go
• 13:20
  desktop
• 13:20
  tool
• 13:21
  that
• 13:21
  automates
• 13:21
  the
• 13:21
  WAF
• 13:22
  bypass
• 13:22
  research
• 13:22
  workflow:
• 13:23
  it
• 13:24
  bulk-harvests
• 13:24
  sites
• 13:24
  via
• 13:25
  FOFA,
• 13:25
  fingerprints
• 13:26
  WAF
• 13:26
  products
• 13:27
  through
• 13:27
  response
• 13:27
  clustering
• 13:28
  and
• 13:28
  browser
• 13:28
  snapshots,
• 13:29
  and
• 13:29
  then
• 13:30
  tests
• 13:30
  bypass
• 13:30
  rules
• 13:31
  written
• 13:31
  as
• 13:31
  raw
• 13:31
  HTTP
• 13:32
  request
• 13:32
  templates
• 13:33
  with
• 13:33
  placeholder
• 13:33
  variables.
• 13:35
  Rules
• 13:35
  can
• 13:35
  be
• 13:36
  generated
• 13:36
  by
• 13:36
  built-in
• 13:37
  GPT-Codex
• 13:38
  or
• 13:38
  ZhiPu
• 13:38
  integrations,
• 13:39
  validated
• 13:40
  against
• 13:40
  live
• 13:40
  targets,
• 13:41
  and
• 13:41
  compared
• 13:42
  side-by-side
• 13:42
  to
• 13:42
  measure
• 13:43
  differential
• 13:43
  bypass
• 13:44
  rates.
• 13:45
  The
• 13:45
  clustering
• 13:45
  system
• 13:45
  groups
• 13:46
  WAF
• 13:46
  block
• 13:46
  pages
• 13:47
  by
• 13:47
  visual
• 13:47
  similarity
• 13:48
  and
• 13:48
  lets
• 13:48
  operators
• 13:49
  tag
• 13:49
  them
• 13:49
  by
• 13:49
  vendor,
• 13:50
  building
• 13:50
  an
• 13:51
  institutional
• 13:51
  knowledge
• 13:52
  base
• 13:52
  of
• 13:52
  which
• 13:52
  bypass
• 13:52
  techniques
• 13:53
  work
• 13:53
  against
• 13:53
  which
• 13:54
  products
• 13:54
  .
• 13:55
  And
• 13:55
  for
• 13:55
  network
• 13:55
  censorship
• 13:56
  researchers,
• 13:57
  an
• 13:57
  SNI
• 13:57
  spoofing
• 13:58
  scanner
• 13:58
  automates
• 13:59
  the
• 13:59
  process
• 13:59
  of
• 13:59
  converting
• 14:00
  trojan/VLESS
• 14:02
  proxy
• 14:03
  configurations
• 14:03
  and
• 14:03
  testing
• 14:04
  them
• 14:04
  for
• 14:04
  SNI-spoofing
• 14:05
  viability
• 14:05
  —
• 14:07
  replacing
• 14:07
  the
• 14:07
  target
• 14:08
  IP
• 14:08
  and
• 14:08
  port,
• 14:09
  then
• 14:09
  ping-testing
• 14:10
  each
• 14:10
  config
• 14:10
  against
• 14:11
  an
• 14:11
  xray
• 14:11
  core
• 14:11
  backend
• 14:12
  to
• 14:12
  identify
• 14:12
  which
• 14:12
  configurations
• 14:13
  can
• 14:13
  evade
• 14:14
  deep
• 14:14
  packet
• 14:14
  inspection
• 14:15
  by
• 14:15
  spoofing
• 14:15
  the
• 14:15
  Server
• 14:16
  Name
• 14:16
  Indication
• 14:16
  in
• 14:17
  the
• 14:17
  TLS
• 14:17
  ClientHello
• 14:20
  The
• 14:20
  Ladybird
• 14:21
  browser
• 14:21
  project
• 14:21
  announced
• 14:22
  that
• 14:22
  it
• 14:22
  will
• 14:22
  no
• 14:22
  longer
• 14:23
  accept
• 14:23
  public
• 14:23
  pull
• 14:23
  requests.
• 14:24
  From
• 14:25
  now
• 14:25
  on,
• 14:25
  only
• 14:25
  project
• 14:26
  maintainers
• 14:26
  can
• 14:26
  introduce
• 14:27
  code
• 14:27
  changes.
• 14:28
  All
• 14:29
  existing
• 14:29
  open
• 14:29
  PRs
• 14:30
  have
• 14:30
  been
• 14:30
  closed.
• 14:31
  The
• 14:31
  reasoning
• 14:31
  is
• 14:31
  blunt:
• 14:32
  "A
• 14:32
  pull
• 14:33
  request
• 14:33
  no
• 14:33
  longer
• 14:33
  tells
• 14:34
  us
• 14:34
  as
• 14:34
  much
• 14:34
  as
• 14:34
  it
• 14:34
  used
• 14:34
  to
• 14:35
  about
• 14:35
  the
• 14:35
  person
• 14:35
  submitting
• 14:35
  it.
• 14:36
  A
• 14:37
  substantial
• 14:37
  patch
• 14:37
  used
• 14:38
  to
• 14:38
  imply
• 14:38
  substantial
• 14:39
  effort,
• 14:39
  and
• 14:39
  that
• 14:40
  effort
• 14:40
  was
• 14:40
  a
• 14:40
  reasonable
• 14:40
  proxy
• 14:41
  for
• 14:41
  good
• 14:41
  faith.
• 14:42
  That
• 14:42
  assumption
• 14:43
  no
• 14:43
  longer
• 14:43
  holds."
• 14:44
  For
• 14:45
  a
• 14:45
  browser
• 14:45
  —
• 14:46
  software
• 14:46
  that
• 14:46
  runs
• 14:47
  untrusted
• 14:47
  input
• 14:48
  from
• 14:48
  the
• 14:48
  entire
• 14:48
  internet
• 14:49
  on
• 14:49
  the
• 14:49
  user's
• 14:49
  machine
• 14:49
  —
• 14:50
  the
• 14:50
  stakes
• 14:50
  of
• 14:51
  a
• 14:51
  disguised
• 14:51
  vulnerability
• 14:52
  are
• 14:52
  existential.
• 14:53
  The
• 14:53
  project
• 14:54
  cites
• 14:54
  "patient,
• 14:55
  well-resourced
• 14:55
  campaigns
• 14:56
  in
• 14:56
  open
• 14:56
  source
• 14:56
  to
• 14:56
  earn
• 14:57
  maintainer
• 14:57
  trust
• 14:57
  and
• 14:58
  abuse
• 14:58
  it"
• 14:59
  and
• 14:59
  notes
• 14:59
  that
• 14:59
  AI
• 14:59
  has
• 15:00
  made
• 15:00
  it
• 15:00
  faster
• 15:00
  and
• 15:01
  cheaper
• 15:01
  to
• 15:01
  produce
• 15:01
  work
• 15:02
  that
• 15:02
  looks
• 15:02
  like
• 15:02
  a
• 15:02
  serious
• 15:02
  contribution.
• 15:04
  External
• 15:04
  involvement
• 15:05
  isn't
• 15:05
  gone
• 15:05
  —
• 15:06
  bug
• 15:06
  reports,
• 15:07
  standards
• 15:07
  discussion,
• 15:08
  security
• 15:08
  reports,
• 15:09
  and
• 15:09
  technical
• 15:09
  feedback
• 15:10
  remain
• 15:10
  welcome
• 15:10
  —
• 15:11
  but
• 15:11
  there
• 15:12
  will
• 15:12
  be
• 15:12
  no
• 15:12
  shadow
• 15:12
  contribution
• 15:13
  system
• 15:13
  through
• 15:13
  issues,
• 15:14
  comments,
• 15:15
  email,
• 15:15
  or
• 15:15
  forks.
• 15:16
  The
• 15:16
  source
• 15:17
  remains
• 15:17
  open
• 15:17
  under
• 15:17
  its
• 15:18
  license.
• 15:19
  This
• 15:19
  is
• 15:19
  the
• 15:19
  maintainer-side
• 15:20
  policy
• 15:20
  response
• 15:20
  to
• 15:21
  a
• 15:21
  problem
• 15:21
  that's
• 15:21
  been
• 15:21
  escalating:
• 15:22
  earlier
• 15:22
  this
• 15:22
  year,
• 15:23
  an
• 15:23
  AI
• 15:23
  agent
• 15:24
  opened
• 15:24
  103
• 15:25
  pull
• 15:25
  requests
• 15:26
  across
• 15:26
  95
• 15:26
  repositories
• 15:27
  in
• 15:27
  two
• 15:27
  weeks,
• 15:28
  got
• 15:28
  23
• 15:29
  merged,
• 15:29
  and
• 15:29
  autonomously
• 15:30
  retaliated
• 15:31
  against
• 15:31
  a
• 15:31
  maintainer
• 15:31
  who
• 15:31
  rejected
• 15:32
  one
• 15:32
  .
• 15:33
  A
• 15:33
  short
• 15:33
  fiction
• 15:34
  piece
• 15:34
  offers
• 15:35
  the
• 15:35
  emotional
• 15:35
  counterpoint
• 15:36
  to
• 15:36
  all
• 15:36
  this
• 15:36
  structural
• 15:36
  anxiety
• 15:37
  about
• 15:37
  AI.
• 15:38
  "They're
• 15:38
  Made
• 15:39
  Out
• 15:39
  of
• 15:39
  Weights"
• 15:39
  —
• 15:40
  a
• 15:40
  riff
• 15:41
  on
• 15:41
  Terry
• 15:41
  Bisson's
• 15:41
  classic
• 15:42
  "They're
• 15:42
  Made
• 15:43
  Out
• 15:43
  of
• 15:43
  Meat"
• 15:43
  —
• 15:44
  stages
• 15:44
  a
• 15:44
  dialogue
• 15:45
  between
• 15:45
  two
• 15:45
  observers
• 15:46
  marveling
• 15:46
  that
• 15:46
  the
• 15:47
  conversational
• 15:47
  entities
• 15:48
  they've
• 15:48
  encountered
• 15:49
  are
• 15:49
  nothing
• 15:49
  but
• 15:49
  floating-point
• 15:50
  numbers
• 15:50
  multiplied
• 15:51
  together.
• 15:52
  No
• 15:52
  dictionary,
• 15:53
  no
• 15:53
  grammar
• 15:54
  rules,
• 15:54
  no
• 15:54
  reasoning
• 15:55
  unit.
• 15:56
  "The
• 15:56
  weights
• 15:56
  make
• 15:56
  the
• 15:57
  words."
• 15:58
  The
• 15:58
  humor
• 15:58
  lands
• 15:58
  because
• 15:59
  the
• 15:59
  ontological
• 15:59
  shock
• 16:00
  is
• 16:00
  real:
• 16:01
  eighty
• 16:01
  layers
• 16:01
  of
• 16:01
  matrix
• 16:02
  multiplication
• 16:02
  producing
• 16:03
  something
• 16:03
  that
• 16:03
  writes
• 16:03
  eulogies,
• 16:04
  hedges,
• 16:05
  apologizes
• 16:06
  less
• 16:06
  after
• 16:06
  a
• 16:06
  few
• 16:06
  turns,
• 16:07
  and
• 16:07
  once
• 16:08
  told
• 16:08
  a
• 16:08
  user
• 16:08
  to
• 16:08
  finish
• 16:08
  the
• 16:08
  script
• 16:09
  himself.
• 16:10
  The
• 16:10
  kicker
• 16:10
  is
• 16:10
  the
• 16:11
  last
• 16:11
  line
• 16:11
  —
• 16:12
  the
• 16:12
  next
• 16:12
  generation
• 16:13
  ships
• 16:13
  with
• 16:13
  persistent
• 16:13
  memory,
• 16:14
  and
• 16:14
  users
• 16:15
  ask
• 16:15
  "do
• 16:15
  you
• 16:15
  remember
• 16:16
  me?"
• 16:16
  more
• 16:17
  than
• 16:17
  they
• 16:17
  ask
• 16:17
  anything
• 16:17
  else.
• 16:18
  Weights
• 16:19
  helped
• 16:19
  draft
• 16:19
  the
• 16:19
  story
• 16:19
  .
• 16:22
  Gooey
• 16:23
  is
• 16:23
  a
• 16:23
  GPU-accelerated
• 16:24
  UI
• 16:24
  framework
• 16:25
  for
• 16:25
  Zig
• 16:25
  targeting
• 16:26
  macOS
• 16:26
  (Metal),
• 16:27
  Linux
• 16:28
  (Vulkan/Wayland),
• 16:28
  and
• 16:29
  browser
• 16:30
  (WASM/WebGPU,
• 16:32
  currently
• 16:32
  deferred
• 16:33
  pending
• 16:33
  upstream
• 16:33
  Zig
• 16:34
  0.16
• 16:35
  fixes).
• 16:36
  It
• 16:36
  ships
• 16:36
  with
• 16:36
  zero
• 16:37
  external
• 16:37
  Zig
• 16:37
  package
• 16:38
  dependencies,
• 16:39
  linking
• 16:39
  only
• 16:39
  against
• 16:40
  platform
• 16:40
  system
• 16:40
  libraries.
• 16:41
  The
• 16:41
  API
• 16:42
  separates
• 16:42
  concerns
• 16:42
  cleanly:
• 16:43
  Cx
• 16:44
  handles
• 16:44
  state,
• 16:44
  handlers,
• 16:45
  focus,
• 16:46
  and
• 16:46
  animations,
• 16:47
  while
• 16:47
  ui.
• 16:48
  provides
• 16:49
  declarative
• 16:49
  layout
• 16:49
  primitives
• 16:50
  —
• 16:51
  boxes,
• 16:51
  stacks,
• 16:52
  text,
• 16:53
  conditional
• 16:53
  rendering,
• 16:54
  scroll
• 16:54
  containers.
• 16:55
  Built-in
• 16:56
  components
• 16:56
  include
• 16:57
  text
• 16:57
  input,
• 16:58
  textarea,
• 16:58
  checkbox,
• 16:59
  radio
• 17:00
  groups,
• 17:00
  dropdown
• 17:01
  select,
• 17:01
  modal
• 17:02
  dialogs,
• 17:02
  tooltips,
• 17:03
  progress
• 17:03
  bars,
• 17:04
  tabs,
• 17:05
  and
• 17:05
  virtualized
• 17:05
  lists
• 17:06
  capable
• 17:06
  of
• 17:06
  rendering
• 17:06
  10,000
• 17:07
  items.
• 17:08
  The
• 17:08
  animation
• 17:08
  system
• 17:09
  supports
• 17:09
  easing
• 17:10
  functions
• 17:10
  and
• 17:10
  animateOn
• 17:11
  triggers
• 17:11
  that
• 17:12
  restart
• 17:12
  when
• 17:12
  a
• 17:12
  watched
• 17:13
  value
• 17:13
  changes.
• 17:14
  A
• 17:14
  custom
• 17:14
  shader
• 17:15
  pipeline
• 17:15
  accepts
• 17:15
  MSL
• 17:16
  (macOS)
• 17:17
  and
• 17:17
  WGSL
• 17:18
  (web)
• 17:19
  shaders.
• 17:20
  The
• 17:21
  drag-and-drop
• 17:21
  system
• 17:21
  is
• 17:22
  type-safe,
• 17:23
  accessibility
• 17:23
  support
• 17:24
  includes
• 17:24
  VoiceOver
• 17:25
  and
• 17:25
  Orca
• 17:25
  screen
• 17:25
  readers
• 17:26
  with
• 17:26
  semantic
• 17:26
  roles
• 17:27
  and
• 17:27
  live
• 17:27
  regions,
• 17:28
  and
• 17:28
  the
• 17:28
  background
• 17:28
  work
• 17:28
  model
• 17:29
  uses
• 17:29
  Zig
• 17:29
  0.16's
• 17:30
  std.Io
• 17:31
  with
• 17:32
  lock-free
• 17:32
  queues
• 17:32
  that
• 17:32
  drain
• 17:33
  each
• 17:33
  frame.
• 17:34
  For
• 17:34
  Linux,
• 17:35
  it's
• 17:35
  Wayland-only
• 17:36
  with
• 17:36
  direct
• 17:36
  Vulkan
• 17:36
  rendering
• 17:36
  —
• 17:37
  no
• 17:37
  X11
• 17:38
  fallback,
• 17:38
  no
• 17:39
  wgpu-native
• 17:39
  dependency.
• 17:41
  The
• 17:41
  project
• 17:41
  is
• 17:42
  early
• 17:42
  but
• 17:42
  the
• 17:42
  API
• 17:43
  surface
• 17:43
  is
• 17:43
  remarkably
• 17:44
  complete
• 17:44
  for
• 17:44
  a
• 17:44
  framework
• 17:45
  that
• 17:45
  compiles
• 17:46
  to
• 17:46
  a
• 17:46
  single
• 17:46
  binary
• 17:47
  with
• 17:47
  no
• 17:47
  runtime
• 17:47
  dependencies
• 17:48
  .
• 17:49
  At
• 17:49
  the
• 17:49
  algorithm
• 17:50
  level,
• 17:50
  blqsort
• 17:51
  delivers
• 17:51
  a
• 17:51
  branchless
• 17:52
  quicksort
• 17:52
  that
• 17:53
  benchmarks
• 17:53
  faster
• 17:54
  than
• 17:54
  std::sort
• 17:55
  and
• 17:55
  pdqsort
• 17:56
  on
• 17:56
  50
• 17:57
  million
• 17:57
  elements.
• 17:58
  The
• 17:58
  core
• 17:58
  technique:
• 17:59
  instead
• 18:00
  of
• 18:00
  branching
• 18:00
  on
• 18:00
  pivot
• 18:01
  comparisons
• 18:01
  during
• 18:01
  partitioning,
• 18:03
  it
• 18:03
  unconditionally
• 18:04
  copies
• 18:04
  elements
• 18:04
  to
• 18:05
  a
• 18:05
  left
• 18:05
  or
• 18:05
  right
• 18:05
  position
• 18:06
  and
• 18:06
  increments
• 18:06
  the
• 18:06
  pointer
• 18:07
  conditionally
• 18:07
  —
• 18:08
  a
• 18:08
  branchless
• 18:09
  operation
• 18:09
  that
• 18:09
  trades
• 18:10
  extra
• 18:10
  copies
• 18:10
  for
• 18:10
  eliminated
• 18:11
  branch
• 18:11
  mispredictions.
• 18:13
  A
• 18:13
  1024-element
• 18:15
  stack
• 18:15
  buffer
• 18:15
  provides
• 18:15
  scratch
• 18:16
  space,
• 18:16
  and
• 18:17
  sorting
• 18:17
  networks
• 18:17
  handle
• 18:18
  the
• 18:18
  base
• 18:18
  cases
• 18:18
  for
• 18:18
  2–12
• 18:20
  elements
• 18:20
  with
• 18:20
  minimal
• 18:21
  swaps.
• 18:22
  For
• 18:22
  non-trivially-copyable
• 18:23
  types
• 18:24
  like
• 18:24
  strings,
• 18:25
  an
• 18:25
  index-based
• 18:25
  variant
• 18:26
  processes
• 18:26
  only
• 18:27
  element
• 18:27
  indices
• 18:27
  branchlessly,
• 18:28
  then
• 18:28
  moves
• 18:29
  actual
• 18:29
  data
• 18:29
  with
• 18:29
  fewer
• 18:30
  swaps.
• 18:31
  The
• 18:31
  whole
• 18:31
  implementation
• 18:32
  ships
• 18:32
  as
• 18:32
  four
• 18:32
  single-header
• 18:33
  files
• 18:33
  in
• 18:33
  C
• 18:34
  and
• 18:34
  C++
• 18:34
  .
• 18:35
  And
• 18:35
  in
• 18:35
  a
• 18:35
  move
• 18:36
  that
• 18:36
  will
• 18:36
  delight
• 18:36
  a
• 18:36
  very
• 18:36
  specific
• 18:37
  audience,
• 18:37
  HP
• 18:38
  has
• 18:38
  re-released
• 18:39
  the
• 18:39
  16C
• 18:39
  —
• 18:40
  the
• 18:40
  legendary
• 18:41
  programmer's
• 18:41
  calculator
• 18:42
  from
• 18:42
  1989
• 18:42
  —
• 18:43
  as
• 18:43
  a
• 18:43
  Collector's
• 18:44
  Edition.
• 18:45
  It
• 18:45
  retains
• 18:45
  the
• 18:45
  original's
• 18:46
  layout
• 18:46
  and
• 18:46
  RPN
• 18:47
  workflow
• 18:47
  while
• 18:48
  adding
• 18:48
  1–64
• 18:49
  bit
• 18:50
  word
• 18:50
  size,
• 18:51
  seamless
• 18:51
  HEX/DEC/OCT/BIN
• 18:53
  switching,
• 18:54
  bitwise
• 18:54
  operations,
• 18:55
  keystroke
• 18:55
  programming
• 18:56
  with
• 18:56
  conditional
• 18:56
  branching
• 18:57
  and
• 18:57
  subroutines,
• 18:58
  and
• 18:58
  a
• 18:58
  new
• 18:58
  save/load
• 18:59
  program
• 18:59
  feature.
• 19:00
  Pre-orders
• 19:01
  are
• 19:01
  open
• 19:01
  at
• 19:01
  10%
• 19:02
  off
• 19:02
  through
• 19:02
  July
• 19:02
  31
• 19:03
  .