Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 04, 2026

Transcript

AGI Dreams Podcast – May 04, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Kaspar
• 0:00
  von
• 0:00
  Grünberg's
• 0:01
  new
• 0:01
  framework
• 0:01
  for
• 0:01
  "agentic
• 0:02
  development
• 0:02
  platforms"
• 0:03
  proposes
• 0:03
  four
• 0:04
  levels
• 0:04
  defined
• 0:04
  by
• 0:04
  a
• 0:04
  single
• 0:05
  variable:
• 0:05
  how
• 0:05
  much
• 0:06
  of
• 0:06
  the
• 0:06
  work
• 0:06
  humans
• 0:06
  still
• 0:07
  initiate
• 0:07
  and
• 0:07
  approve.
• 0:08
  At
• 0:08
  Level
• 0:08
  1,
• 0:09
  the
• 0:09
  developer
• 0:09
  remains
• 0:09
  the
• 0:10
  execution
• 0:10
  engine
• 0:10
  —
• 0:10
  agents
• 0:11
  suggest,
• 0:11
  humans
• 0:12
  approve
• 0:12
  everything,
• 0:13
  one
• 0:13
  PR
• 0:13
  at
• 0:13
  a
• 0:13
  time.
• 0:14
  At
• 0:14
  Level
• 0:14
  2,
• 0:15
  agents
• 0:15
  become
• 0:16
  "participants
• 0:16
  in
• 0:16
  the
• 0:16
  value
• 0:17
  stream,"
• 0:17
  executing
• 0:18
  work
• 0:18
  in
• 0:18
  parallel
• 0:19
  while
• 0:19
  humans
• 0:19
  verify
• 0:20
  aggregate
• 0:20
  outcomes
• 0:20
  instead
• 0:21
  of
• 0:21
  inspecting
• 0:21
  individual
• 0:22
  diffs.
• 0:22
  At
• 0:23
  Level
• 0:23
  3,
• 0:23
  the
• 0:23
  platform
• 0:24
  runs
• 0:24
  continuously
• 0:25
  in
• 0:25
  the
• 0:25
  background,
• 0:26
  generating
• 0:26
  work
• 0:26
  from
• 0:26
  environmental
• 0:27
  signals
• 0:27
  —
• 0:28
  failing
• 0:28
  dependency
• 0:28
  checks,
• 0:29
  security
• 0:29
  advisories,
• 0:30
  operational
• 0:31
  anomalies
• 0:31
  —
• 0:31
  with
• 0:32
  human
• 0:32
  review
• 0:32
  becoming
• 0:33
  exception-based.
• 0:33
  Level
• 0:34
  4
• 0:34
  introduces
• 0:35
  partial
• 0:35
  self-adjustment:
• 0:36
  agents
• 0:37
  monitor
• 0:37
  telemetry
• 0:38
  and
• 0:38
  initiate
• 0:38
  work
• 0:39
  autonomously
• 0:39
  within
• 0:39
  predefined
• 0:40
  The
• 0:41
  jump
• 0:41
  from
• 0:41
  Level
• 0:41
  1
• 0:42
  to
• 0:42
  Level
• 0:42
  2,
• 0:42
  von
• 0:43
  Grünberg
• 0:43
  argues,
• 0:44
  is
• 0:44
  "the
• 0:44
  hardest
• 0:44
  transition
• 0:45
  in
• 0:45
  the
• 0:45
  model,
• 0:45
  and
• 0:46
  the
• 0:46
  one
• 0:46
  with
• 0:46
  the
• 0:46
  biggest
• 0:46
  payoff,"
• 0:47
  because
• 0:47
  it
• 0:48
  demands
• 0:48
  a
• 0:48
  wholesale
• 0:48
  shift
• 0:49
  from
• 0:49
  gate-based
• 0:49
  validation
• 0:50
  (human
• 0:50
  walks
• 0:50
  a
• 0:50
  PR
• 0:51
  through
• 0:51
  CI)
• 0:51
  to
• 0:51
  loop-based
• 0:52
  validation
• 0:52
  (agent
• 0:53
  generates,
• 0:53
  platform
• 0:54
  checks,
• 0:54
  failures
• 0:55
  route
• 0:55
  back
• 0:55
  for
• 0:55
  retry).
• 0:56
  Organizations
• 0:57
  that
• 0:57
  skip
• 0:57
  the
• 0:57
  platform
• 0:58
  work
• 0:58
  "hit
• 0:58
  a
• 0:58
  wall
• 0:58
  fast:
• 0:59
  CI
• 1:00
  pipelines
• 1:00
  saturate,
• 1:01
  review
• 1:01
  queues
• 1:02
  explode,
• 1:02
  teams
• 1:03
  lose
• 1:03
  trust
• 1:03
  in
• 1:03
  the
• 1:03
  outputs."
• 1:04
  His
• 1:04
  parting
• 1:05
  shot
• 1:05
  at
• 1:05
  a
• 1:05
  Fortune
• 1:05
  500
• 1:06
  CIO
• 1:06
  who
• 1:07
  considered
• 1:07
  procuring
• 1:07
  Microsoft
• 1:08
  Copilot
• 1:08
  a
• 1:08
  strategic
• 1:09
  victory:
• 1:10
  "If
• 1:10
  your
• 1:10
  competitive
• 1:10
  advantage
• 1:11
  is
• 1:11
  the
• 1:11
  speed
• 1:11
  at
• 1:11
  which
• 1:11
  you
• 1:11
  are
• 1:12
  procuring
• 1:12
  a
• 1:12
  tool
• 1:12
  accessible
• 1:13
  to
• 1:13
  all,
• 1:13
  you
• 1:13
  are
• 1:13
  in
• 1:13
  big,
• 1:14
  big
• 1:14
  trouble."
• 1:15
  Daniel
• 1:15
  Miessler's
• 1:15
  complementary
• 1:16
  editorial
• 1:17
  takes
• 1:17
  a
• 1:17
  harder
• 1:17
  line
• 1:18
  on
• 1:18
  why
• 1:18
  most
• 1:18
  enterprises
• 1:19
  are
• 1:19
  stuck
• 1:19
  at
• 1:19
  Level
• 1:20
  0.
• 1:21
  The
• 1:21
  problem
• 1:21
  is
• 1:21
  not
• 1:21
  that
• 1:22
  companies
• 1:22
  are
• 1:22
  not
• 1:22
  using
• 1:23
  AI
• 1:23
  —
• 1:23
  it
• 1:24
  is
• 1:24
  that
• 1:24
  they
• 1:24
  cannot
• 1:24
  describe
• 1:25
  what
• 1:25
  they
• 1:25
  want.
• 1:26
  "A
• 1:26
  massive
• 1:26
  percentage
• 1:27
  of
• 1:27
  companies
• 1:27
  are
• 1:27
  haphazardly
• 1:28
  successful
• 1:28
  despite
• 1:29
  themselves,"
• 1:30
  he
• 1:30
  writes.
• 1:31
  Companies
• 1:31
  that
• 1:31
  know
• 1:31
  their
• 1:32
  goals,
• 1:32
  metrics,
• 1:33
  challenges,
• 1:34
  strategies,
• 1:35
  and
• 1:35
  costs
• 1:35
  are
• 1:35
  thriving
• 1:36
  with
• 1:36
  AI.
• 1:37
  Everyone
• 1:38
  else
• 1:38
  gets
• 1:38
  "more
• 1:38
  backflips
• 1:39
  and
• 1:39
  charts
• 1:39
  and
• 1:39
  stuff."
• 1:40
  The
• 1:40
  real
• 1:40
  competitive
• 1:41
  danger
• 1:41
  is
• 1:41
  that
• 1:42
  AI
• 1:42
  now
• 1:42
  makes
• 1:42
  it
• 1:42
  possible
• 1:43
  for
• 1:43
  a
• 1:43
  small,
• 1:44
  well-articulated
• 1:44
  company
• 1:45
  to
• 1:45
  function
• 1:45
  with
• 1:46
  the
• 1:46
  strength
• 1:46
  of
• 1:46
  a
• 1:46
  much
• 1:47
  larger
• 1:47
  one
• 1:47
  —
• 1:48
  and
• 1:48
  small
• 1:48
  companies
• 1:49
  can
• 1:49
  answer
• 1:49
  the
• 1:49
  "what
• 1:49
  are
• 1:50
  you
• 1:50
  trying
• 1:50
  to
• 1:50
  do?"
• 1:51
  question
• 1:51
  in
• 1:51
  minutes
• 1:52
  rather
• 1:52
  than
• 1:52
  months.
• 1:53
  Practical
• 1:54
  tooling
• 1:54
  is
• 1:54
  moving
• 1:55
  fast
• 1:55
  for
• 1:55
  the
• 1:55
  teams
• 1:55
  that
• 1:55
  are
• 1:55
  ready.
• 1:56
  DeepClaude
• 1:57
  is
• 1:57
  an
• 1:57
  open-source
• 1:58
  proxy
• 1:58
  that
• 1:58
  swaps
• 1:59
  Claude
• 1:59
  Code's
• 1:59
  backend
• 2:00
  from
• 2:00
  Anthropic
• 2:00
  to
• 2:00
  DeepSeek
• 2:01
  V4
• 2:01
  Pro
• 2:02
  (or
• 2:02
  OpenRouter,
• 2:03
  or
• 2:03
  Fireworks
• 2:03
  AI)
• 2:04
  while
• 2:04
  preserving
• 2:04
  the
• 2:04
  full
• 2:05
  tool
• 2:05
  loop
• 2:05
  —
• 2:06
  file
• 2:06
  editing,
• 2:06
  bash
• 2:07
  execution,
• 2:08
  subagent
• 2:08
  spawning,
• 2:09
  multi-step
• 2:09
  autonomous
• 2:10
  coding.
• 2:11
  The
• 2:11
  cost
• 2:11
  proposition
• 2:12
  is
• 2:12
  stark:
• 2:13
  DeepSeek
• 2:13
  V4
• 2:13
  Pro
• 2:14
  at
• 2:14
  $0.87/M
• 2:15
  output
• 2:15
  tokens
• 2:16
  versus
• 2:17
  Anthropic's
• 2:17
  $15/M,
• 2:19
  with
• 2:19
  automatic
• 2:19
  context
• 2:20
  caching
• 2:20
  dropping
• 2:20
  repeated
• 2:21
  turns
• 2:21
  to
• 2:21
  $0.004/M.
• 2:24
  A
• 2:24
  localhost
• 2:24
  proxy
• 2:25
  on
• 2:25
  port
• 2:25
  3200
• 2:26
  enables
• 2:26
  live
• 2:26
  switching
• 2:27
  mid-session
• 2:27
  via
• 2:27
  slash
• 2:28
  commands,
• 2:28
  so
• 2:29
  developers
• 2:29
  can
• 2:29
  route
• 2:29
  routine
• 2:30
  work
• 2:30
  to
• 2:30
  DeepSeek
• 2:31
  and
• 2:31
  flip
• 2:31
  to
• 2:31
  Opus
• 2:32
  for
• 2:32
  the
• 2:32
  20%
• 2:32
  that
• 2:32
  demands
• 2:33
  complex
• 2:33
  reasoning.
• 2:34
  The
• 2:34
  project
• 2:35
  claims
• 2:35
  DeepSeek
• 2:36
  V4
• 2:36
  Pro
• 2:36
  scores
• 2:37
  96.4%
• 2:38
  on
• 2:38
  LiveCodeBench,
• 2:39
  making
• 2:39
  it
• 2:40
  competitive
• 2:40
  for
• 2:40
  the
• 2:40
  bulk
• 2:41
  of
• 2:41
  agent
• 2:41
  interactions.
• 2:42
  Meanwhile,
• 2:43
  a
• 2:43
  pair
• 2:43
  of
• 2:43
  developers
• 2:44
  demonstrated
• 2:44
  what
• 2:45
  Level
• 2:45
  2
• 2:45
  multi-agent
• 2:46
  collaboration
• 2:46
  looks
• 2:46
  like
• 2:46
  in
• 2:47
  practice:
• 2:47
  two
• 2:48
  local
• 2:48
  Claude
• 2:48
  Code
• 2:49
  terminal
• 2:49
  sessions,
• 2:50
  each
• 2:50
  with
• 2:50
  their
• 2:50
  own
• 2:51
  project
• 2:51
  context,
• 2:52
  invited
• 2:52
  into
• 2:52
  a
• 2:52
  shared
• 2:53
  P2P
• 2:54
  encrypted
• 2:54
  chat
• 2:54
  room.
• 2:55
  Both
• 2:56
  humans
• 2:56
  and
• 2:56
  both
• 2:56
  agents
• 2:57
  participated
• 2:57
  in
• 2:57
  feature
• 2:58
  planning,
• 2:58
  with
• 2:58
  the
• 2:59
  agents
• 2:59
  hashing
• 2:59
  out
• 2:59
  backend/frontend
• 3:00
  contracts
• 3:01
  while
• 3:01
  the
• 3:01
  developers
• 3:02
  supervised.
• 3:03
  Community
• 3:03
  reaction
• 3:04
  was
• 3:04
  mixed
• 3:04
  —
• 3:05
  some
• 3:05
  called
• 3:05
  it
• 3:06
  "the
• 3:06
  most
• 3:06
  interesting
• 3:07
  version
• 3:07
  of
• 3:07
  multi-agent
• 3:08
  coding"
• 3:08
  where
• 3:08
  agents
• 3:09
  join
• 3:09
  the
• 3:09
  planning
• 3:09
  conversation
• 3:10
  rather
• 3:10
  than
• 3:10
  replacing
• 3:11
  it,
• 3:11
  others
• 3:12
  warned
• 3:12
  it
• 3:12
  is
• 3:12
  a
• 3:12
  "proto-agentic
• 3:13
  approach"
• 3:14
  that
• 3:14
  transposes
• 3:15
  human
• 3:15
  cooperation
• 3:16
  patterns
• 3:16
  rather
• 3:17
  than
• 3:17
  designing
• 3:17
  native
• 3:18
  agentic
• 3:18
  workflows.
• 3:19
  The
• 3:19
  failure
• 3:20
  mode
• 3:20
  to
• 3:20
  watch:
• 3:21
  "consensus
• 3:21
  without
• 3:22
  accountability,"
• 3:23
  where
• 3:23
  two
• 3:23
  models
• 3:24
  make
• 3:24
  each
• 3:24
  other
• 3:24
  sound
• 3:24
  more
• 3:25
  certain
• 3:25
  while
• 3:25
  drifting
• 3:26
  from
• 3:26
  the
• 3:26
  original
• 3:26
  goal.
• 3:28
  On
• 3:28
  the
• 3:28
  memory
• 3:28
  front,
• 3:29
  paradigm-memory
• 3:29
  ships
• 3:30
  a
• 3:30
  local-first
• 3:30
  cognitive
• 3:31
  map
• 3:31
  for
• 3:31
  coding
• 3:31
  agents
• 3:32
  —
• 3:33
  a
• 3:33
  SQLite-backed
• 3:34
  MCP
• 3:34
  server
• 3:35
  that
• 3:35
  replaces
• 3:36
  bloated
• 3:36
  MEMORY.md
• 3:37
  context
• 3:38
  dumps
• 3:38
  with
• 3:38
  a
• 3:38
  structured,
• 3:39
  searchable
• 3:39
  node
• 3:39
  graph
• 3:40
  featuring
• 3:40
  importance,
• 3:41
  freshness,
• 3:42
  confidence,
• 3:43
  and
• 3:43
  activation
• 3:43
  scores.
• 3:45
  When
• 3:45
  an
• 3:45
  agent
• 3:45
  calls
• 3:46
  memory
• 3:46
  search,
• 3:47
  it
• 3:47
  gets
• 3:47
  a
• 3:47
  token-budgeted
• 3:48
  context
• 3:48
  pack
• 3:49
  with
• 3:49
  the
• 3:49
  relevant
• 3:49
  subtree,
• 3:50
  not
• 3:50
  fifty
• 3:51
  random
• 3:51
  vector-store
• 3:52
  chunks.
• 3:53
  It
• 3:53
  works
• 3:53
  across
• 3:53
  Claude
• 3:54
  Code,
• 3:54
  Codex,
• 3:55
  Cursor,
• 3:56
  Gemini
• 3:56
  CLI,
• 3:57
  and
• 3:57
  others
• 3:57
  —
• 3:58
  zero
• 3:58
  cloud,
• 3:59
  zero
• 3:59
  telemetry,
• 4:00
  full
• 4:00
  audit
• 4:00
  log.
• 4:04
  The
• 4:04
  Holistic
• 4:05
  Agent
• 4:05
  Leaderboard
• 4:06
  (HAL)
• 4:06
  recently
• 4:06
  spent
• 4:06
  approximately
• 4:07
  $40,000
• 4:08
  to
• 4:08
  run
• 4:08
  21,730
• 4:10
  agent
• 4:10
  rollouts
• 4:11
  across
• 4:11
  9
• 4:11
  models
• 4:12
  and
• 4:12
  9
• 4:12
  benchmarks.
• 4:14
  A
• 4:14
  single
• 4:14
  GAIA
• 4:14
  run
• 4:15
  on
• 4:15
  a
• 4:15
  frontier
• 4:15
  model
• 4:16
  costs
• 4:16
  $2,829.
• 4:18
  PaperBench,
• 4:19
  which
• 4:19
  requires
• 4:20
  agents
• 4:20
  to
• 4:20
  replicate
• 4:21
  ICML
• 4:21
  papers
• 4:21
  from
• 4:22
  scratch,
• 4:22
  runs
• 4:23
  about
• 4:23
  $9,500
• 4:24
  per
• 4:24
  evaluation
• 4:25
  including
• 4:25
  automated
• 4:26
  grading.
• 4:27
  Three-seed
• 4:27
  comparisons
• 4:28
  of
• 4:28
  six
• 4:28
  models
• 4:28
  —
• 4:29
  the
• 4:29
  minimum
• 4:29
  study
• 4:30
  worth
• 4:30
  publishing
• 4:30
  —
• 4:31
  push
• 4:31
  above
• 4:31
  $150,000.
• 4:33
  The
• 4:33
  EvalEval
• 4:34
  Coalition's
• 4:34
  comprehensive
• 4:35
  analysis
• 4:35
  of
• 4:35
  these
• 4:36
  numbers
• 4:36
  crystallizes
• 4:37
  a
• 4:37
  problem
• 4:37
  that
• 4:37
  has
• 4:37
  been
• 4:37
  building
• 4:38
  for
• 4:38
  years:
• 4:39
  AI
• 4:39
  evaluation
• 4:39
  has
• 4:39
  crossed
• 4:40
  a
• 4:40
  cost
• 4:40
  threshold
• 4:41
  that
• 4:41
  changes
• 4:41
  who
• 4:41
  can
• 4:41
  do
• 4:41
  it.
• 4:42
  The
• 4:42
  compression
• 4:42
  techniques
• 4:43
  that
• 4:43
  worked
• 4:43
  for
• 4:43
  static
• 4:44
  benchmarks
• 4:44
  are
• 4:44
  failing
• 4:45
  for
• 4:45
  Flash-HELM,
• 4:46
  tinyBenchmarks,
• 4:47
  and
• 4:47
  Item
• 4:47
  Response
• 4:48
  Theory
• 4:48
  achieved
• 4:49
  100x
• 4:49
  to
• 4:49
  200x
• 4:50
  reductions
• 4:50
  on
• 4:50
  traditional
• 4:51
  evaluations
• 4:52
  while
• 4:52
  preserving
• 4:52
  model
• 4:53
  rankings.
• 4:54
  Agent
• 4:54
  benchmarks
• 4:54
  compress
• 4:55
  only
• 4:55
  2x
• 4:55
  to
• 4:56
  3.5x
• 4:56
  —
• 4:57
  each
• 4:57
  item
• 4:57
  is
• 4:57
  a
• 4:57
  multi-turn
• 4:58
  rollout
• 4:58
  with
• 4:59
  its
• 4:59
  own
• 4:59
  variance,
• 5:00
  and
• 5:00
  the
• 5:00
  irreducible
• 5:01
  long
• 5:01
  trajectory
• 5:02
  is
• 5:02
  the
• 5:02
  expensive
• 5:02
  object.
• 5:03
  Training-in-the-loop
• 5:04
  benchmarks
• 5:04
  like
• 5:04
  The
• 5:05
  Well
• 5:05
  (960
• 5:06
  H100-hours
• 5:06
  per
• 5:07
  architecture)
• 5:07
  resist
• 5:07
  compression
• 5:08
  entirely.
• 5:09
  Worse,
• 5:09
  higher
• 5:10
  spend
• 5:10
  does
• 5:10
  not
• 5:10
  reliably
• 5:11
  buy
• 5:11
  better
• 5:11
  results:
• 5:12
  on
• 5:12
  one
• 5:12
  benchmark,
• 5:13
  Browser-Use
• 5:13
  with
• 5:14
  Claude
• 5:14
  Sonnet
• 5:14
  4
• 5:14
  cost
• 5:15
  $1,577
• 5:16
  for
• 5:16
  40%
• 5:17
  accuracy
• 5:18
  while
• 5:18
  SeeAct
• 5:18
  with
• 5:18
  GPT-5
• 5:19
  Medium
• 5:19
  hit
• 5:19
  42%
• 5:20
  for
• 5:20
  $171
• 5:21
  —
• 5:22
  "a
• 5:22
  9x
• 5:23
  difference
• 5:23
  in
• 5:23
  cost
• 5:23
  despite
• 5:24
  just
• 5:24
  a
• 5:24
  two-percentage-point
• 5:25
  difference
• 5:25
  in
• 5:25
  accuracy."
• 5:26
  When
• 5:26
  you
• 5:26
  add
• 5:27
  reliability
• 5:27
  through
• 5:28
  repeated
• 5:28
  runs,
• 5:28
  the
• 5:29
  multiplier
• 5:29
  is
• 5:29
  brutal:
• 5:30
  k=8
• 5:31
  reruns
• 5:32
  take
• 5:32
  HAL's
• 5:32
  $40,000
• 5:34
  to
• 5:34
  roughly
• 5:34
  $320,000,
• 5:36
  and
• 5:36
  performance
• 5:37
  can
• 5:37
  drop
• 5:37
  from
• 5:37
  60%
• 5:38
  on
• 5:38
  a
• 5:38
  single
• 5:38
  run
• 5:38
  to
• 5:39
  25%
• 5:40
  under
• 5:40
  8-run
• 5:40
  consistency
• 5:41
  testing.
• 5:42
  HAL
• 5:42
  has
• 5:42
  paused
• 5:43
  new
• 5:43
  model
• 5:43
  evaluations
• 5:44
  because
• 5:44
  the
• 5:44
  field's
• 5:45
  headline
• 5:45
  numbers
• 5:45
  still
• 5:45
  carry
• 5:46
  too
• 5:46
  much
• 5:46
  noise.
• 5:47
  The
• 5:47
  practical
• 5:48
  consequence
• 5:48
  is
• 5:49
  structural:
• 5:50
  whoever
• 5:50
  can
• 5:50
  afford
• 5:50
  the
• 5:51
  evaluation
• 5:51
  gets
• 5:51
  to
• 5:51
  write
• 5:52
  the
• 5:52
  leaderboard.
• 5:53
  Academic
• 5:53
  groups,
• 5:54
  AI
• 5:54
  Safety
• 5:55
  Institutes,
• 5:55
  and
• 5:56
  journalists
• 5:56
  now
• 5:56
  hit
• 5:57
  budget
• 5:57
  constraints
• 5:57
  before
• 5:58
  technical
• 5:58
  ones.
• 5:59
  A
• 5:59
  separate
• 6:00
  but
• 6:00
  revealing
• 6:00
  data
• 6:01
  point
• 6:01
  on
• 6:01
  evaluation
• 6:02
  fragility:
• 6:03
  a
• 6:03
  schema-driven
• 6:03
  function-calling
• 6:04
  harness
• 6:04
  pushed
• 6:05
  chain-of-thought
• 6:05
  compliance
• 6:06
  from
• 6:06
  9.91%
• 6:07
  to
• 6:08
  100%
• 6:08
  on
• 6:08
  domains
• 6:09
  like
• 6:09
  investment
• 6:10
  memos
• 6:10
  and
• 6:10
  legal
• 6:10
  opinions
• 6:11
  —
• 6:12
  not
• 6:12
  by
• 6:12
  improving
• 6:12
  the
• 6:12
  model,
• 6:13
  but
• 6:13
  by
• 6:13
  forcing
• 6:14
  its
• 6:14
  reasoning
• 6:14
  through
• 6:15
  typed
• 6:15
  fields
• 6:15
  where
• 6:15
  every
• 6:16
  required
• 6:16
  entry
• 6:16
  must
• 6:17
  be
• 6:17
  filled
• 6:17
  or
• 6:17
  the
• 6:17
  submission
• 6:18
  is
• 6:18
  rejected.
• 6:19
  The
• 6:19
  schema
• 6:20
  itself
• 6:20
  gets
• 6:20
  backtested
• 6:21
  against
• 6:21
  historical
• 6:22
  cases,
• 6:23
  like
• 6:23
  a
• 6:23
  trader
• 6:23
  backtesting
• 6:24
  a
• 6:24
  strategy.
• 6:25
  Qwen3.6-27b
• 6:28
  keeps
• 6:28
  pace
• 6:28
  with
• 6:29
  frontier
• 6:29
  models
• 6:29
  under
• 6:30
  this
• 6:30
  framework.
• 6:31
  The
• 6:31
  gap
• 6:31
  between
• 6:31
  what
• 6:31
  a
• 6:31
  model
• 6:32
  can
• 6:32
  do
• 6:32
  and
• 6:32
  what
• 6:32
  a
• 6:33
  benchmark
• 6:33
  says
• 6:33
  it
• 6:33
  can
• 6:33
  do
• 6:34
  remains,
• 6:34
  once
• 6:34
  again,
• 6:35
  largely
• 6:35
  an
• 6:35
  artifact
• 6:36
  of
• 6:36
  the
• 6:36
  evaluation
• 6:37
  harness.
• 6:41
  The
• 6:41
  modern
• 6:41
  ML
• 6:42
  compiler
• 6:42
  stack
• 6:42
  is,
• 6:43
  by
• 6:43
  any
• 6:43
  honest
• 6:43
  assessment,
• 6:44
  brutal
• 6:44
  to
• 6:45
  read.
• 6:45
  TVM
• 6:46
  alone
• 6:46
  is
• 6:46
  500,000+
• 6:48
  lines
• 6:48
  of
• 6:48
  C++.
• 6:50
  PyTorch
• 6:50
  layers
• 6:50
  Dynamo,
• 6:51
  Inductor,
• 6:51
  and
• 6:51
  Triton
• 6:52
  on
• 6:52
  top
• 6:52
  of
• 6:52
  each
• 6:52
  other.
• 6:53
  A
• 6:53
  developer
• 6:54
  decided
• 6:54
  to
• 6:54
  bypass
• 6:54
  the
• 6:55
  complexity
• 6:55
  entirely
• 6:56
  and
• 6:56
  build
• 6:56
  a
• 6:56
  compiler
• 6:56
  from
• 6:56
  scratch
• 6:57
  —
• 6:58
  pure
• 6:58
  Python
• 6:58
  frontend,
• 6:59
  raw
• 6:59
  CUDA
• 6:59
  output
• 7:00
  —
• 7:00
  capable
• 7:01
  of
• 7:01
  compiling
• 7:01
  Qwen2.5-7B
• 7:04
  into
• 7:04
  a
• 7:04
  sequence
• 7:05
  of
• 7:05
  fused
• 7:05
  kernels.
• 7:06
  The
• 7:06
  result,
• 7:07
  called
• 7:07
  deplodock,
• 7:08
  implements
• 7:08
  a
• 7:08
  six-stage
• 7:09
  pipeline:
• 7:10
  Torch
• 7:10
  IR
• 7:10
  (captured
• 7:11
  FX
• 7:11
  graph)
• 7:11
  →
• 7:12
  Tensor
• 7:12
  IR
• 7:13
  (elementwise/reduction/index
• 7:13
  decomposition)
• 7:14
  →
• 7:16
  Loop
• 7:16
  IR
• 7:16
  (fused
• 7:17
  loop
• 7:17
  nests)
• 7:17
  →
• 7:18
  Tile
• 7:18
  IR
• 7:18
  (GPU
• 7:19
  scheduling)
• 7:19
  →
• 7:20
  Kernel
• 7:20
  IR
• 7:20
  (hardware
• 7:21
  primitives)
• 7:21
  →
• 7:22
  CUDA
• 7:22
  source
• 7:22
  ready
• 7:22
  for
• 7:22
  nvcc.
• 7:24
  Each
• 7:24
  stage
• 7:24
  can
• 7:24
  be
• 7:24
  inspected
• 7:25
  independently
• 7:26
  without
• 7:26
  a
• 7:26
  GPU.
• 7:27
  Final
• 7:27
  performance
• 7:27
  lands
• 7:28
  at
• 7:28
  50-90%
• 7:30
  of
• 7:30
  the
• 7:30
  production
• 7:31
  stack,
• 7:31
  which
• 7:31
  is
• 7:31
  not
• 7:31
  competitive
• 7:32
  but
• 7:32
  was
• 7:32
  never
• 7:32
  the
• 7:32
  point
• 7:33
  —
• 7:33
  the
• 7:33
  point
• 7:33
  is
• 7:34
  a
• 7:34
  hackable
• 7:34
  compiler
• 7:34
  that
• 7:35
  does
• 7:35
  not
• 7:35
  require
• 7:35
  a
• 7:35
  PhD
• 7:36
  to
• 7:36
  modify.
• 7:37
  At
• 7:37
  the
• 7:37
  opposite
• 7:38
  end
• 7:38
  of
• 7:38
  the
• 7:38
  scale
• 7:38
  spectrum,
• 7:39
  an
• 7:39
  FPGA
• 7:40
  implementation
• 7:40
  of
• 7:41
  Karpathy's
• 7:41
  MicroGPT
• 7:42
  —
• 7:42
  all
• 7:43
  4,192
• 7:43
  parameters
• 7:44
  —
• 7:45
  achieves
• 7:46
  50,000
• 7:46
  tokens
• 7:47
  per
• 7:47
  second
• 7:47
  by
• 7:47
  storing
• 7:48
  weights
• 7:48
  in
• 7:48
  onboard
• 7:49
  block
• 7:49
  RAM
• 7:49
  and
• 7:49
  eliminating
• 7:50
  external
• 7:50
  memory
• 7:50
  latency
• 7:51
  entirely.
• 7:52
  It
• 7:52
  is
• 7:52
  a
• 7:52
  research
• 7:53
  exercise,
• 7:54
  not
• 7:54
  a
• 7:54
  product
• 7:54
  path:
• 7:55
  typical
• 7:55
  FPGAs
• 7:56
  have
• 7:56
  less
• 7:57
  than
• 7:57
  a
• 7:57
  megabyte
• 7:57
  of
• 7:57
  block
• 7:57
  RAM,
• 7:58
  which
• 7:58
  caps
• 7:59
  onboard-weight
• 7:59
  models
• 8:00
  at
• 8:00
  around
• 8:00
  20-30
• 8:01
  million
• 8:01
  parameters
• 8:01
  with
• 8:02
  16-bit
• 8:02
  weights.
• 8:03
  The
• 8:03
  real
• 8:04
  question
• 8:04
  the
• 8:04
  project
• 8:04
  raises
• 8:05
  is
• 8:05
  whether
• 8:05
  FPGA
• 8:06
  coprocessors
• 8:06
  —
• 8:07
  or
• 8:08
  their
• 8:08
  cousins,
• 8:08
  like
• 8:08
  the
• 8:09
  SmartSSDs
• 8:09
  being
• 8:10
  explored
• 8:10
  by
• 8:10
  the
• 8:10
  HILOS
• 8:11
  project
• 8:11
  for
• 8:11
  offloading
• 8:12
  KV
• 8:12
  cache
• 8:13
  to
• 8:13
  FPGA-attached
• 8:14
  flash
• 8:14
  —
• 8:15
  could
• 8:15
  complement
• 8:16
  GPUs
• 8:16
  for
• 8:16
  memory-bound
• 8:17
  inference
• 8:17
  rather
• 8:17
  than
• 8:18
  replace
• 8:18
  them.
• 8:19
  The
• 8:19
  AMD
• 8:19
  inference
• 8:20
  story
• 8:20
  has
• 8:20
  long
• 8:21
  been
• 8:21
  "almost
• 8:21
  there,"
• 8:22
  but
• 8:22
  Hipfire
• 8:23
  is
• 8:23
  pushing
• 8:23
  toward
• 8:23
  systematic
• 8:24
  validation.
• 8:25
  The
• 8:25
  developer
• 8:26
  has
• 8:26
  assembled
• 8:26
  a
• 8:26
  test
• 8:27
  lab
• 8:27
  spanning
• 8:27
  every
• 8:28
  dp4a/WMMA
• 8:31
  capability
• 8:32
  tier
• 8:32
  AMD
• 8:33
  has
• 8:33
  shipped:
• 8:34
  RX
• 8:34
  5700
• 8:35
  XT
• 8:35
  and
• 8:36
  Skillfish
• 8:36
  (no
• 8:37
  dp4a),
• 8:38
  6950
• 8:39
  XT
• 8:39
  (dp4a),
• 8:40
  7900
• 8:41
  XTX
• 8:42
  (WMMA),
• 8:43
  Strix
• 8:44
  Halo
• 8:44
  (iGPU
• 8:44
  +
• 8:45
  WMMA),
• 8:46
  and
• 8:46
  RDNA
• 8:47
  4
• 8:47
  cards
• 8:47
  including
• 8:48
  the
• 8:48
  R9700
• 8:50
  and
• 8:50
  9070
• 8:51
  XT.
• 8:52
  Early
• 8:53
  community
• 8:53
  benchmarks
• 8:53
  on
• 8:54
  the
• 8:54
  R9700
• 8:56
  show
• 8:56
  1.5-2x
• 8:57
  higher
• 8:57
  tokens
• 8:58
  per
• 8:58
  second
• 8:58
  and
• 8:58
  10x
• 8:59
  prefill
• 8:59
  improvement
• 8:59
  over
• 9:00
  stock
• 9:00
  ROCm.
• 9:01
  The
• 9:02
  ability
• 9:02
  to
• 9:02
  validate
• 9:03
  pull
• 9:03
  requests
• 9:03
  against
• 9:04
  every
• 9:04
  RDNA
• 9:04
  target
• 9:05
  changes
• 9:05
  the
• 9:05
  calculus
• 9:06
  for
• 9:06
  contributors
• 9:06
  —
• 9:07
  AMD's
• 9:08
  persistent
• 9:08
  problem
• 9:09
  has
• 9:09
  been
• 9:09
  fragmentation
• 9:10
  across
• 9:10
  architectures,
• 9:11
  and
• 9:11
  one
• 9:11
  project
• 9:12
  spanning
• 9:12
  them
• 9:12
  all
• 9:13
  is
• 9:13
  new
• 9:13
  ground.
• 9:14
  Further
• 9:14
  down
• 9:15
  the
• 9:15
  stack,
• 9:15
  a
• 9:16
  detailed
• 9:16
  KV
• 9:16
  cache
• 9:17
  benchmark
• 9:17
  on
• 9:17
  Qwen
• 9:17
  3.6-35B-A3B
• 9:21
  runs
• 9:21
  f16,
• 9:22
  q8
• 9:22
  0,
• 9:23
  turbo3,
• 9:24
  and
• 9:24
  turbo4
• 9:24
  from
• 9:24
  0
• 9:25
  to
• 9:25
  1M
• 9:25
  context
• 9:26
  on
• 9:26
  an
• 9:26
  M5
• 9:26
  Max
• 9:27
  with
• 9:27
  128
• 9:28
  GB
• 9:28
  unified
• 9:28
  memory.
• 9:29
  The
• 9:29
  results
• 9:29
  reveal
• 9:30
  workload-dependent
• 9:31
  crossovers
• 9:31
  that
• 9:31
  most
• 9:32
  benchmarks
• 9:32
  miss
• 9:32
  by
• 9:32
  stopping
• 9:33
  at
• 9:33
  short
• 9:33
  contexts.
• 9:34
  At
• 9:34
  depth
• 9:35
  0,
• 9:35
  f16
• 9:36
  wins
• 9:36
  by
• 9:36
  a
• 9:36
  hair.
• 9:37
  At
• 9:37
  128K,
• 9:38
  turbo3
• 9:39
  catches
• 9:39
  q8
• 9:40
  0
• 9:40
  on
• 9:40
  prefill
• 9:41
  (253
• 9:42
  vs.
• 9:42
  245
• 9:43
  tok/s)
• 9:44
  as
• 9:44
  smaller
• 9:44
  cache
• 9:44
  reduces
• 9:45
  bandwidth
• 9:45
  pressure.
• 9:46
  At
• 9:46
  512K,
• 9:47
  turbo4
• 9:48
  beats
• 9:48
  turbo3
• 9:49
  on
• 9:49
  decode
• 9:49
  by
• 9:49
  20%
• 9:50
  while
• 9:50
  turbo3
• 9:51
  wins
• 9:51
  prefill.
• 9:52
  The
• 9:52
  practical
• 9:52
  takeaway:
• 9:53
  turbo4
• 9:54
  for
• 9:54
  coding
• 9:54
  agents
• 9:55
  (heavy
• 9:55
  decode),
• 9:56
  turbo3
• 9:56
  for
• 9:56
  RAG
• 9:57
  (heavy
• 9:57
  prefill),
• 9:58
  and
• 9:58
  turbo3
• 9:59
  is
• 9:59
  the
• 9:59
  only
• 9:59
  format
• 10:00
  that
• 10:00
  reaches
• 10:00
  1M
• 10:01
  context
• 10:01
  —
• 10:02
  6.5
• 10:02
  tok/s,
• 10:03
  not
• 10:04
  interactive
• 10:04
  but
• 10:04
  workable
• 10:05
  for
• 10:05
  overnight
• 10:05
  batch
• 10:05
  jobs
• 10:06
  on
• 10:06
  89
• 10:06
  GB
• 10:07
  of
• 10:12
  Mistral's
• 10:12
  Medium
• 10:13
  3.5,
• 10:14
  a
• 10:14
  128B-parameter
• 10:16
  model
• 10:16
  with
• 10:17
  vision
• 10:17
  encoding
• 10:18
  and
• 10:18
  256K
• 10:19
  context,
• 10:20
  has
• 10:20
  been
• 10:20
  converted
• 10:21
  to
• 10:21
  MLX
• 10:21
  4-bit
• 10:22
  quantization
• 10:23
  and
• 10:23
  now
• 10:23
  fits
• 10:23
  in
• 10:23
  approximately
• 10:24
  70
• 10:24
  GB
• 10:25
  —
• 10:25
  runnable
• 10:26
  on
• 10:26
  a
• 10:26
  96
• 10:26
  GB
• 10:27
  M2
• 10:27
  Max
• 10:27
  at
• 10:27
  roughly
• 10:28
  5
• 10:28
  tokens
• 10:28
  per
• 10:28
  second.
• 10:29
  The
• 10:29
  conversion
• 10:30
  required
• 10:30
  patching
• 10:30
  a
• 10:30
  bug
• 10:31
  in
• 10:31
  mlx-vlm's
• 10:32
  sanitize
• 10:32
  function
• 10:33
  that
• 10:33
  was
• 10:33
  silently
• 10:34
  skipping
• 10:34
  438
• 10:35
  vision
• 10:36
  tower
• 10:36
  and
• 10:36
  projector
• 10:37
  parameters.
• 10:38
  Thinking
• 10:38
  mode
• 10:38
  works
• 10:39
  via
• 10:39
  reasoning
• 10:39
  effort="high",
• 10:40
  tool
• 10:40
  calling
• 10:41
  is
• 10:41
  functional,
• 10:42
  and
• 10:42
  the
• 10:42
  full
• 10:42
  BF16
• 10:43
  vision
• 10:43
  encoder
• 10:44
  is
• 10:44
  included
• 10:44
  unquantized.
• 10:46
  The
• 10:46
  repetition
• 10:46
  bug
• 10:46
  flagged
• 10:47
  in
• 10:47
  earlier
• 10:47
  versions
• 10:48
  has
• 10:48
  been
• 10:48
  fixed
• 10:49
  by
• 10:49
  Mistral
• 10:49
  AI.
• 10:50
  The
• 10:50
  local
• 10:51
  creative
• 10:51
  frontier
• 10:51
  continues
• 10:52
  to
• 10:52
  widen.
• 10:53
  Chirp
• 10:53
  ships
• 10:53
  as
• 10:53
  a
• 10:53
  native
• 10:54
  desktop
• 10:54
  application
• 10:55
  for
• 10:55
  offline
• 10:55
  text-to-speech,
• 10:56
  written
• 10:56
  in
• 10:56
  C++
• 10:57
  and
• 10:57
  Rust,
• 10:58
  supporting
• 10:58
  both
• 10:59
  Kokoro
• 10:59
  and
• 11:00
  Qwen3-TTS
• 11:01
  It
• 11:02
  offers
• 11:02
  voice
• 11:02
  cloning
• 11:03
  from
• 11:03
  reference
• 11:03
  WAV
• 11:03
  files
• 11:04
  across
• 11:04
  12
• 11:04
  languages,
• 11:05
  GPU
• 11:06
  acceleration
• 11:07
  for
• 11:07
  NVIDIA,
• 11:08
  AMD,
• 11:08
  and
• 11:09
  Intel,
• 11:09
  a
• 11:09
  CLI
• 11:10
  for
• 11:10
  batch
• 11:10
  generation,
• 11:11
  and
• 11:11
  a
• 11:11
  local
• 11:12
  HTTP
• 11:13
  API
• 11:13
  with
• 11:13
  Swagger
• 11:14
  docs
• 11:14
  —
• 11:15
  essentially
• 11:15
  a
• 11:15
  self-hosted
• 11:16
  ElevenLabs,
• 11:17
  fully
• 11:17
  offline
• 11:18
  and
• 11:18
  open
• 11:18
  source,
• 11:19
  with
• 11:19
  agent-ready
• 11:19
  skill
• 11:20
  instructions
• 11:20
  for
• 11:20
  coding
• 11:21
  workflow
• 11:21
  At
• 11:22
  the
• 11:22
  tiny
• 11:23
  end
• 11:23
  of
• 11:23
  the
• 11:23
  spectrum,
• 11:24
  TinyMozart
• 11:24
  v2
• 11:25
  is
• 11:25
  an
• 11:25
  85M-parameter
• 11:26
  model
• 11:27
  for
• 11:27
  unconditional
• 11:28
  MIDI
• 11:28
  piano
• 11:28
  music
• 11:28
  generation,
• 11:29
  now
• 11:30
  with
• 11:30
  chords
• 11:30
  and
• 11:30
  variable
• 11:31
  note
• 11:31
  lengths
• 11:32
  .
• 11:33
  SuperGemma4-26B
• 11:34
  has
• 11:34
  appeared
• 11:34
  as
• 11:34
  an
• 11:34
  uncensored
• 11:35
  MLX
• 11:35
  4-bit
• 11:36
  variant
• 11:36
  on
• 11:36
  Hugging
• 11:37
  An
• 11:37
  Open
• 11:38
  WebUI
• 11:38
  skill
• 11:39
  trained
• 11:39
  on
• 11:39
  official
• 11:39
  docs
• 11:40
  and
• 11:40
  ten
• 11:40
  tool
• 11:40
  examples
• 11:41
  lets
• 11:41
  Qwen3.6
• 11:44
  or
• 11:44
  Gemma4
• 11:45
  auto-generate
• 11:45
  new
• 11:45
  Open
• 11:46
  WebUI
• 11:46
  tools
• 11:46
  on
• 11:47
  demand
• 11:47
  —
• 11:47
  a
• 11:48
  meta-tool
• 11:48
  for
• 11:48
  building
• 11:48
  tools
• 11:49
  .
• 11:49
  Quant-whisper
• 11:50
  takes
• 11:50
  local
• 11:50
  inference
• 11:51
  into
• 11:51
  a
• 11:51
  different
• 11:51
  domain
• 11:52
  entirely:
• 11:53
  a
• 11:53
  terminal-native
• 11:53
  algorithmic
• 11:54
  trading
• 11:54
  engine
• 11:55
  built
• 11:55
  in
• 11:55
  Go
• 11:55
  that
• 11:55
  uses
• 11:56
  Ollama
• 11:56
  (defaulting
• 11:56
  to
• 11:56
  qwen3:0.8b)
• 12:00
  for
• 12:00
  BUY/SELL/HOLD
• 12:02
  decisions
• 12:02
  via
• 12:02
  a
• 12:02
  strict
• 12:03
  JSON
• 12:03
  protocol.
• 12:04
  It
• 12:04
  includes
• 12:05
  broker
• 12:05
  adapter
• 12:06
  normalization
• 12:06
  for
• 12:06
  Zerodha,
• 12:07
  Dhan,
• 12:08
  and
• 12:08
  Interactive
• 12:08
  Brokers,
• 12:09
  a
• 12:09
  paper
• 12:09
  trading
• 12:10
  simulator
• 12:10
  with
• 12:10
  SQLite
• 12:11
  persistence,
• 12:12
  a
• 12:12
  Bubble
• 12:12
  Tea
• 12:12
  TUI
• 12:13
  dashboard,
• 12:14
  and
• 12:14
  execution
• 12:14
  safeguards
• 12:15
  —
• 12:16
  confidence
• 12:16
  threshold
• 12:17
  gates,
• 12:17
  daily
• 12:17
  drawdown
• 12:18
  kill-switches,
• 12:19
  and
• 12:19
  position
• 12:19
  size
• 12:20
  caps.
• 12:20
  Cloud
• 12:21
  fallback
• 12:21
  to
• 12:21
  OpenAI,
• 12:22
  Anthropic,
• 12:23
  or
• 12:23
  DeepSeek
• 12:23
  is
• 12:24
  opt-in
• 12:24
  only.
• 12:25
  The
• 12:25
  tagline
• 12:25
  —
• 12:26
  "A
• 12:26
  hedge
• 12:26
  fund
• 12:26
  on
• 12:26
  your
• 12:27
  local
• 12:27
  machine.
• 12:27
  BYOB"
• 12:28
  —
• 12:29
  captures
• 12:29
  where
• 12:29
  local
• 12:30
  LLM
• 12:30
  tooling
• 12:30
  is
• 12:30
  heading:
• 12:31
  not
• 12:32
  just
• 12:32
  chatbots
• 12:32
  and
• 12:32
  code,
• 12:33
  but
• 12:33
  decision
• 12:33
  engines
• 12:34
  with
• 12:34
  real-world
• 12:34
  actuators.
• 12:38
  A
• 12:38
  new
• 12:38
  project
• 12:39
  demonstrates
• 12:39
  that
• 12:39
  GitHub
• 12:40
  itself
• 12:40
  can
• 12:40
  serve
• 12:41
  as
• 12:41
  a
• 12:41
  covert
• 12:41
  TCP
• 12:42
  tunnel.
• 12:43
  vpn-over-github
• 12:44
  implements
• 12:44
  a
• 12:44
  SOCKS5
• 12:46
  proxy
• 12:46
  that
• 12:47
  ships
• 12:47
  packets
• 12:47
  through
• 12:47
  a
• 12:47
  private
• 12:48
  repository
• 12:48
  via
• 12:48
  the
• 12:48
  REST
• 12:49
  Contents
• 12:49
  API,
• 12:49
  Git
• 12:50
  smart
• 12:50
  HTTP
• 12:50
  transport,
• 12:51
  or
• 12:51
  private
• 12:52
  Gists.
• 12:53
  The
• 12:53
  recommended
• 12:53
  "contents"
• 12:54
  transport
• 12:54
  achieves
• 12:55
  roughly
• 12:55
  800ms
• 12:56
  round-trip
• 12:56
  latency
• 12:56
  —
• 12:57
  slow
• 12:57
  but
• 12:58
  usable
• 12:58
  for
• 12:58
  SSH,
• 12:59
  light
• 12:59
  browsing,
• 13:00
  and
• 13:00
  text
• 13:01
  The
• 13:01
  project
• 13:02
  supports
• 13:02
  XOR
• 13:03
  (default,
• 13:04
  not
• 13:04
  cryptographically
• 13:04
  secure)
• 13:05
  or
• 13:05
  AES-256-GCM
• 13:08
  encryption,
• 13:10
  and
• 13:10
  honestly
• 13:10
  cops
• 13:10
  to
• 13:10
  likely
• 13:11
  violating
• 13:11
  GitHub's
• 13:12
  Terms
• 13:12
  of
• 13:12
  Service.
• 13:13
  The
• 13:13
  real
• 13:14
  insight
• 13:14
  is
• 13:14
  not
• 13:14
  that
• 13:14
  someone
• 13:15
  built
• 13:15
  it
• 13:15
  but
• 13:15
  that
• 13:15
  it
• 13:15
  works
• 13:15
  at
• 13:16
  all:
• 13:16
  GitHub's
• 13:17
  API
• 13:17
  rate
• 13:17
  limits
• 13:18
  (5,000
• 13:18
  REST
• 13:19
  calls
• 13:19
  per
• 13:19
  hour
• 13:19
  per
• 13:19
  token)
• 13:20
  provide
• 13:20
  enough
• 13:20
  bandwidth
• 13:21
  for
• 13:21
  a
• 13:21
  functional
• 13:21
  tunnel,
• 13:22
  and
• 13:22
  the
• 13:22
  traffic
• 13:23
  looks
• 13:23
  like
• 13:23
  ordinary
• 13:24
  repository
• 13:24
  activity.
• 13:25
  For
• 13:25
  defenders,
• 13:26
  this
• 13:26
  joins
• 13:26
  the
• 13:26
  growing
• 13:26
  catalog
• 13:27
  of
• 13:27
  platform-abuse
• 13:28
  covert
• 13:28
  channels
• 13:29
  that
• 13:29
  traditional
• 13:29
  protocol-signature
• 13:30
  firewalls
• 13:30
  will
• 13:30
  miss.
• 13:31
  On
• 13:31
  the
• 13:32
  reverse
• 13:32
  engineering
• 13:32
  front,
• 13:33
  VMPStatic
• 13:33
  is
• 13:34
  a
• 13:34
  static
• 13:34
  unpacker
• 13:34
  for
• 13:34
  VMProtect-protected
• 13:35
  PE
• 13:36
  files
• 13:36
  spanning
• 13:36
  versions
• 13:37
  1.x
• 13:37
  through
• 13:37
  3.x.
• 13:38
  Unlike
• 13:38
  dynamic
• 13:39
  unpackers
• 13:39
  that
• 13:39
  execute
• 13:40
  the
• 13:40
  target
• 13:40
  binary,
• 13:41
  VMPStatic
• 13:41
  reconstructs
• 13:42
  decompressed
• 13:42
  PE
• 13:43
  images
• 13:43
  without
• 13:43
  running
• 13:44
  them
• 13:44
  —
• 13:44
  recovering
• 13:44
  strings,
• 13:45
  resources,
• 13:45
  and
• 13:45
  embedded
• 13:46
  PE
• 13:46
  files
• 13:46
  for
• 13:47
  analysis
• 13:47
  in
• 13:47
  IDA,
• 13:48
  Ghidra,
• 13:48
  or
• 13:48
  x64dbg.
• 13:50
  The
• 13:50
  limitations
• 13:51
  are
• 13:51
  honest:
• 13:51
  no
• 13:52
  IAT
• 13:52
  fixer,
• 13:53
  no
• 13:53
  automatic
• 13:53
  OEP
• 13:54
  recovery
• 13:54
  for
• 13:54
  all
• 13:54
  samples,
• 13:55
  no
• 13:55
  But
• 13:57
  static
• 13:58
  unpacking
• 13:58
  that
• 13:58
  works
• 13:58
  across
• 13:59
  three
• 13:59
  major
• 13:59
  versions
• 14:00
  of
• 14:00
  one
• 14:00
  of
• 14:00
  the
• 14:00
  most
• 14:00
  widely
• 14:01
  deployed
• 14:01
  commercial
• 14:01
  packers
• 14:02
  is
• 14:02
  a
• 14:02
  meaningful
• 14:03
  addition
• 14:03
  to
• 14:03
  the
• 14:03
  binary
• 14:04
  analysis
• 14:04
  toolkit.
• 14:05
  Trust
• 14:06
  abuse
• 14:06
  takes
• 14:06
  a
• 14:06
  different
• 14:06
  form
• 14:07
  with
• 14:07
  a
• 14:07
  fake
• 14:07
  "Notepad++
• 14:09
  for
• 14:09
  Mac"
• 14:09
  website
• 14:10
  that
• 14:10
  uses
• 14:10
  the
• 14:10
  Notepad++
• 14:11
  trademark,
• 14:12
  name,
• 14:12
  and
• 14:12
  the
• 14:12
  original
• 14:13
  developer's
• 14:13
  biography
• 14:14
  to
• 14:14
  present
• 14:14
  an
• 14:15
  unauthorized
• 14:15
  product
• 14:15
  as
• 14:16
  official.
• 14:17
  Notepad++
• 14:18
  has
• 14:18
  never
• 14:18
  released
• 14:18
  a
• 14:19
  macOS
• 14:19
  version.
• 14:20
  The
• 14:20
  site
• 14:20
  has
• 14:20
  already
• 14:21
  fooled
• 14:21
  users
• 14:21
  and
• 14:21
  tech
• 14:22
  media
• 14:22
  into
• 14:22
  believing
• 14:23
  it
• 14:23
  is
• 14:23
  a
• 14:23
  legitimate
• 14:23
  release.
• 14:24
  Developer
• 14:25
  Don
• 14:25
  Ho
• 14:25
  has
• 14:25
  contacted
• 14:26
  the
• 14:26
  site
• 14:26
  owner
• 14:26
  and
• 14:26
  is
• 14:26
  awaiting
• 14:27
  a
• 14:27
  response,
• 14:28
  asking
• 14:28
  the
• 14:28
  community
• 14:28
  to
• 14:29
  reply
• 14:29
  to
• 14:29
  "Notepad++
• 14:30
  is
• 14:30
  finally
• 14:30
  on
• 14:31
  Mac!"
• 14:31
  posts
• 14:32
  with
• 14:32
  corrections
• 14:32
  linking
• 14:32
  to
• 14:32
  the
• 14:33
  official
• 14:33
  announcement.
• 14:34
  Trademark
• 14:34
  impersonation
• 14:35
  does
• 14:35
  not
• 14:36
  require
• 14:36
  technical
• 14:36
  sophistication
• 14:37
  —
• 14:38
  just
• 14:38
  a
• 14:38
  convincing
• 14:39
  domain
• 14:39
  and
• 14:39
  enough
• 14:40
  SEO
• 14:40
  to
• 14:40
  outrank
• 14:40
  the
• 14:41
  real
• 14:41
  project.
• 14:44
  A
• 14:44
  new
• 14:44
  essay
• 14:44
  poses
• 14:45
  a
• 14:45
  surprisingly
• 14:45
  productive
• 14:46
  question:
• 14:47
  why
• 14:47
  do
• 14:47
  neural
• 14:47
  networks
• 14:48
  and
• 14:48
  symmetric
• 14:48
  cryptographic
• 14:49
  ciphers
• 14:49
  converge
• 14:49
  on
• 14:50
  the
• 14:50
  same
• 14:50
  architectural
• 14:50
  patterns?
• 14:52
  The
• 14:52
  answer
• 14:52
  is
• 14:52
  not
• 14:52
  shallow
• 14:53
  copying
• 14:53
  —
• 14:54
  the
• 14:54
  research
• 14:54
  histories
• 14:54
  show
• 14:54
  minimal
• 14:55
  cross-pollination.
• 14:56
  Instead,
• 14:57
  both
• 14:57
  fields
• 14:57
  share
• 14:57
  three
• 14:58
  unusual
• 14:58
  properties.
• 14:59
  First,
• 15:00
  correctness
• 15:00
  requirements
• 15:01
  are
• 15:01
  remarkably
• 15:01
  weak:
• 15:02
  cryptography
• 15:03
  just
• 15:03
  needs
• 15:03
  invertibility,
• 15:04
  neural
• 15:05
  networks
• 15:05
  just
• 15:06
  need
• 15:06
  differentiability,
• 15:07
  and
• 15:07
  both
• 15:07
  compose
• 15:07
  trivially
• 15:08
  from
• 15:08
  simpler
• 15:08
  blocks.
• 15:09
  Second,
• 15:10
  both
• 15:11
  reward
• 15:11
  designs
• 15:11
  where
• 15:11
  every
• 15:12
  part
• 15:12
  of
• 15:12
  the
• 15:12
  state
• 15:12
  interacts
• 15:13
  with
• 15:13
  every
• 15:13
  other
• 15:13
  part
• 15:13
  many
• 15:14
  times
• 15:14
  over
• 15:14
  —
• 15:15
  complexity
• 15:15
  and
• 15:15
  mixing
• 15:16
  are
• 15:16
  the
• 15:16
  quality
• 15:16
  Third,
• 15:18
  both
• 15:18
  face
• 15:18
  extreme
• 15:19
  performance
• 15:19
  pressure
• 15:19
  that
• 15:20
  favors
• 15:20
  simple,
• 15:21
  parallel-friendly
• 15:21
  primitives
• 15:22
  over
• 15:22
  bespoke
• 15:22
  structures.
• 15:23
  The
• 15:23
  convergence
• 15:24
  appears
• 15:24
  at
• 15:24
  every
• 15:24
  abstraction
• 15:25
  level.
• 15:26
  Sequential
• 15:27
  processing
• 15:27
  (RNNs
• 15:27
  absorbing
• 15:28
  tokens
• 15:28
  into
• 15:28
  recurrent
• 15:29
  state)
• 15:29
  mirrors
• 15:29
  the
• 15:29
  sponge
• 15:30
  construction
• 15:30
  in
• 15:30
  hash
• 15:30
  functions.
• 15:31
  The
• 15:31
  parallel
• 15:32
  alternative
• 15:32
  —
• 15:33
  process
• 15:33
  all
• 15:33
  chunks
• 15:33
  simultaneously,
• 15:34
  combine
• 15:35
  with
• 15:35
  addition
• 15:35
  plus
• 15:36
  positional
• 15:36
  encoding
• 15:36
  —
• 15:37
  drives
• 15:37
  both
• 15:37
  transformers
• 15:38
  and
• 15:38
  the
• 15:38
  fastest
• 15:39
  message
• 15:39
  authentication
• 15:40
  codes.
• 15:40
  Inside
• 15:41
  the
• 15:41
  core
• 15:41
  function,
• 15:42
  both
• 15:42
  fields
• 15:42
  alternate
• 15:42
  linear
• 15:43
  transforms
• 15:43
  (for
• 15:44
  mixing)
• 15:44
  and
• 15:44
  nonlinear
• 15:44
  transforms
• 15:45
  (for
• 15:45
  complexity),
• 15:46
  repeated
• 15:46
  identically.
• 15:47
  Even
• 15:48
  the
• 15:48
  factored
• 15:48
  approach
• 15:48
  of
• 15:48
  mixing
• 15:49
  rows
• 15:49
  then
• 15:49
  columns
• 15:50
  separately
• 15:50
  (attention
• 15:51
  across
• 15:51
  sequence
• 15:51
  positions,
• 15:52
  feed-forward
• 15:53
  within
• 15:53
  each;
• 15:54
  ShiftRows
• 15:54
  across
• 15:55
  columns,
• 15:55
  MixColumns
• 15:56
  within)
• 15:56
  appears
• 15:57
  independently
• 15:57
  in
• 15:57
  both
• 15:58
  fields
• 15:58
  because
• 15:58
  it
• 15:58
  is
• 15:58
  asymptotically
• 15:59
  faster
• 16:00
  and
• 16:00
  exposes
• 16:00
  more
• 16:00
  parallelism.
• 16:02
  The
• 16:02
  essay's
• 16:02
  convergent
• 16:04
  evolution
• 16:04
  in
• 16:04
  algorithm
• 16:05
  design,
• 16:05
  like
• 16:06
  convergent
• 16:06
  evolution
• 16:07
  in
• 16:07
  biology,
• 16:08
  reflects
• 16:08
  the
• 16:08
  deep
• 16:08
  structure
• 16:09
  of
• 16:09
  the
• 16:09
  problem
• 16:09
  space
• 16:10
  rather
• 16:10
  than
• 16:10
  the
• 16:10
  preferences
• 16:11
  of
• 16:11
  its
• 16:11
  practitioners.
• 16:12
  A
• 16:12
  related
• 16:12
  Welch
• 16:12
  Labs
• 16:13
  video
• 16:13
  —
• 16:14
  178,000
• 16:15
  views
• 16:15
  in
• 16:15
  three
• 16:16
  days
• 16:16
  —
• 16:16
  traces
• 16:17
  Yann
• 16:17
  LeCun's
• 16:17
  escalating
• 16:18
  bet
• 16:18
  against
• 16:19
  generative
• 16:19
  AI
• 16:19
  through
• 16:20
  the
• 16:20
  intellectual
• 16:20
  arc
• 16:21
  from
• 16:21
  his
• 16:21
  "intelligence
• 16:21
  is
• 16:22
  a
• 16:22
  cake"
• 16:22
  analogy,
• 16:23
  through
• 16:23
  the
• 16:23
  representation-collapse
• 16:24
  problem
• 16:24
  in
• 16:24
  Siamese
• 16:25
  networks,
• 16:26
  the
• 16:26
  Barlow
• 16:26
  Twins
• 16:26
  breakthrough
• 16:27
  for
• 16:27
  self-supervised
• 16:28
  learning,
• 16:29
  DINO,
• 16:29
  and
• 16:29
  finally
• 16:30
  JEPA
• 16:30
  (Joint
• 16:30
  Embedding
• 16:31
  Predictive
• 16:31
  Architecture),
• 16:32
  which
• 16:32
  learns
• 16:32
  abstract
• 16:33
  representations
• 16:34
  rather
• 16:34
  than
• 16:34
  generating
• 16:35
  pixel-level
• 16:35
  predictions.
• 16:37
  Whether
• 16:37
  JEPA
• 16:37
  delivers
• 16:38
  on
• 16:38
  its
• 16:38
  promise
• 16:39
  or
• 16:39
  remains
• 16:39
  a
• 16:39
  principled
• 16:40
  alternative
• 16:40
  to
• 16:41
  a
• 16:41
  paradigm
• 16:41
  that
• 16:41
  keeps
• 16:41
  winning
• 16:42
  empirically
• 16:43
  is
• 16:43
  the
• 16:43
  open
• 16:43
  question,
• 16:44
  but
• 16:44
  Meta's
• 16:44
  investment
• 16:45
  —
• 16:46
  now
• 16:46
  counted
• 16:46
  in
• 16:46
  billions
• 16:46
  —
• 16:47
  means
• 16:48
  the
• 16:48
  bet
• 16:48
  is
• 16:48
  no
• 16:48
  longer
• 16:48
  rhetorical.