Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – May 28, 2026

Transcript

AGI Dreams Podcast – May 28, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Andrej
• 0:00
  Karpathy's
• 0:01
  move
• 0:01
  from
• 0:01
  OpenAI
• 0:02
  co-founder
• 0:02
  to
• 0:02
  Anthropic
• 0:03
  hire
• 0:03
  is
• 0:03
  now
• 0:04
  official,
• 0:04
  and
• 0:05
  the
• 0:05
  stated
• 0:05
  mission
• 0:05
  lands
• 0:06
  squarely
• 0:06
  on
• 0:06
  the
• 0:06
  third
• 0:06
  rail
• 0:07
  of
• 0:07
  AI
• 0:07
  safety
• 0:08
  discourse:
• 0:09
  teaching
• 0:09
  Claude
• 0:09
  to
• 0:09
  improve
• 0:10
  itself
• 0:10
  without
• 0:10
  human
• 0:11
  supervision.
• 0:12
  The
• 0:12
  community
• 0:13
  reaction
• 0:13
  oscillates
• 0:14
  between
• 0:14
  awe
• 0:14
  and
• 0:14
  alarm.
• 0:16
  One
• 0:16
  commenter
• 0:17
  captured
• 0:17
  it
• 0:17
  well:
• 0:18
  "Recursive
• 0:18
  self-improvement
• 0:19
  was
• 0:19
  a
• 0:19
  pipedream
• 0:20
  only
• 0:20
  a
• 0:20
  year
• 0:20
  ago.
• 0:21
  Now
• 0:22
  it's
• 0:22
  becoming
• 0:22
  a
• 0:22
  reality...
• 0:23
  What
• 0:24
  does
• 0:24
  it
• 0:24
  mean
• 0:24
  when
• 0:24
  for-profit
• 0:25
  organizations
• 0:25
  not
• 0:26
  only
• 0:26
  have
• 0:26
  the
• 0:26
  keys
• 0:26
  to
• 0:26
  superior
• 0:27
  knowledge
• 0:27
  and
• 0:28
  understanding,
• 0:29
  but
• 0:29
  the
• 0:29
  resources
• 0:30
  for
• 0:30
  learning
• 0:30
  are
• 0:30
  also
• 0:30
  locked
• 0:31
  away?"
• 0:32
  Others
• 0:32
  are
• 0:32
  more
• 0:32
  skeptical,
• 0:33
  noting
• 0:34
  that
• 0:34
  LLMs
• 0:34
  are
• 0:34
  well
• 0:35
  past
• 0:35
  the
• 0:35
  point
• 0:35
  where
• 0:36
  any
• 0:36
  single
• 0:36
  hire
• 0:36
  moves
• 0:37
  the
• 0:37
  needle
• 0:37
  technically,
• 0:38
  and
• 0:38
  questioning
• 0:39
  whether
• 0:39
  Anthropic's
• 0:39
  marketing
• 0:40
  department
• 0:40
  is
• 0:40
  doing
• 0:40
  more
• 0:41
  heavy
• 0:41
  lifting
• 0:41
  than
• 0:41
  the
• 0:41
  research.
• 0:43
  The
• 0:43
  practical
• 0:43
  concern
• 0:44
  worth
• 0:44
  tracking
• 0:44
  is
• 0:44
  behavioral
• 0:45
  continuity
• 0:45
  —
• 0:46
  if
• 0:46
  a
• 0:46
  model
• 0:47
  improves
• 0:47
  itself
• 0:47
  without
• 0:48
  human
• 0:48
  review,
• 0:49
  every
• 0:49
  training
• 0:49
  run
• 0:50
  introduces
• 0:50
  regression
• 0:50
  risk
• 0:51
  that
• 0:51
  supervised
• 0:51
  checkpoints
• 0:52
  would
• 0:52
  normally
• 0:52
  catch.
• 0:54
  The
• 0:54
  talent
• 0:54
  war
• 0:54
  takes
• 0:55
  a
• 0:55
  different
• 0:55
  shape
• 0:55
  in
• 0:55
  Beijing.
• 0:56
  China
• 0:57
  is
• 0:57
  reportedly
• 0:57
  clamping
• 0:58
  down
• 0:58
  on
• 0:58
  overseas
• 0:59
  travel
• 0:59
  for
• 0:59
  AI
• 0:59
  researchers
• 1:00
  at
• 1:00
  Alibaba
• 1:01
  and
• 1:01
  DeepSeek,
• 1:02
  treating
• 1:02
  AI
• 1:02
  expertise
• 1:03
  as
• 1:03
  a
• 1:03
  national
• 1:04
  security
• 1:04
  asset
• 1:04
  subject
• 1:05
  to
• 1:05
  movement
• 1:05
  restrictions.
• 1:07
  The
• 1:07
  comparison
• 1:07
  to
• 1:07
  US
• 1:08
  security
• 1:08
  clearance
• 1:08
  constraints
• 1:09
  is
• 1:09
  not
• 1:09
  unreasonable
• 1:10
  —
• 1:11
  most
• 1:11
  countries
• 1:11
  restrict
• 1:12
  travel
• 1:12
  for
• 1:12
  individuals
• 1:13
  in
• 1:13
  sensitive
• 1:13
  fields
• 1:14
  —
• 1:14
  but
• 1:15
  the
• 1:15
  timing
• 1:15
  is
• 1:15
  instructive.
• 1:17
  The
• 1:17
  Qwen
• 1:17
  team
• 1:17
  lost
• 1:18
  its
• 1:18
  technical
• 1:18
  architect
• 1:19
  under
• 1:19
  murky
• 1:19
  circumstances
• 1:20
  earlier
• 1:20
  this
• 1:20
  year
• 1:21
  after
• 1:21
  Alibaba
• 1:22
  dismantled
• 1:22
  his
• 1:22
  vertically
• 1:23
  integrated
• 1:23
  R&D
• 1:24
  group
• 1:24
  and
• 1:25
  replaced
• 1:25
  it
• 1:25
  with
• 1:25
  horizontal
• 1:26
  modules.
• 1:27
  Meanwhile,
• 1:28
  the
• 1:28
  Stanford
• 1:28
  HAI
• 1:29
  report
• 1:29
  documented
• 1:30
  an
• 1:30
  89%
• 1:31
  collapse
• 1:31
  in
• 1:31
  China-to-US
• 1:32
  researcher
• 1:32
  migration
• 1:33
  since
• 1:33
  2017.
• 1:34
  If
• 1:35
  the
• 1:35
  talent
• 1:35
  is
• 1:35
  already
• 1:35
  not
• 1:36
  leaving
• 1:36
  voluntarily,
• 1:37
  travel
• 1:37
  restrictions
• 1:38
  look
• 1:38
  less
• 1:38
  like
• 1:38
  security
• 1:39
  policy
• 1:39
  and
• 1:40
  more
• 1:40
  like
• 1:40
  signaling.
• 1:41
  The
• 1:41
  deeper
• 1:42
  risk
• 1:42
  is
• 1:42
  self-inflicted:
• 1:43
  making
• 1:44
  China
• 1:44
  less
• 1:44
  attractive
• 1:45
  to
• 1:45
  its
• 1:45
  own
• 1:45
  domestic
• 1:45
  AI
• 1:46
  talent
• 1:46
  who
• 1:46
  might
• 1:46
  reasonably
• 1:47
  wonder
• 1:47
  whether
• 1:47
  building
• 1:48
  cutting-edge
• 1:48
  systems
• 1:48
  also
• 1:49
  means
• 1:49
  surrendering
• 1:49
  freedom
• 1:50
  of
• 1:50
  movement.
• 1:53
  The
• 1:53
  SWE-rebench
• 1:56
  leaderboard
• 1:56
  is
• 1:56
  back
• 1:57
  after
• 1:57
  a
• 1:57
  three-month
• 1:57
  hiatus
• 1:58
  with
• 1:58
  110
• 1:59
  fresh
• 1:59
  Python
• 1:59
  tasks
• 2:00
  drawn
• 2:00
  from
• 2:00
  real
• 2:00
  GitHub
• 2:01
  PRs
• 2:01
  in
• 2:01
  March
• 2:01
  through
• 2:01
  May.
• 2:03
  The
• 2:03
  setup
• 2:03
  follows
• 2:03
  the
• 2:03
  standard
• 2:04
  SWE-bench
• 2:04
  format
• 2:05
  —
• 2:06
  models
• 2:06
  read
• 2:07
  PR
• 2:07
  issues,
• 2:07
  edit
• 2:08
  code,
• 2:08
  run
• 2:08
  tests,
• 2:09
  pass
• 2:09
  the
• 2:09
  full
• 2:09
  suite
• 2:09
  —
• 2:10
  but
• 2:10
  the
• 2:10
  batch
• 2:11
  is
• 2:11
  deliberately
• 2:11
  larger
• 2:12
  to
• 2:12
  smooth
• 2:12
  out
• 2:12
  the
• 2:12
  noise
• 2:12
  that
• 2:13
  plagues
• 2:13
  monthly
• 2:13
  updates.
• 2:14
  GPT-5.5,
• 2:17
  Opus
• 2:17
  4.7,
• 2:18
  Cursor
• 2:18
  Composer
• 2:19
  2.5,
• 2:20
  and
• 2:20
  Kimi
• 2:20
  K2.6
• 2:22
  are
• 2:22
  all
• 2:22
  on
• 2:23
  the
• 2:23
  board,
• 2:23
  with
• 2:23
  Gemini
• 2:23
  Flash
• 2:24
  3.5,
• 2:24
  DeepSeek
• 2:25
  v4
• 2:25
  Pro,
• 2:25
  and
• 2:26
  Qwen3.5-397B
• 2:30
  promised
• 2:31
  for
• 2:31
  the
• 2:31
  next
• 2:31
  wave.
• 2:32
  The
• 2:32
  community
• 2:33
  response
• 2:33
  is
• 2:33
  telling:
• 2:34
  the
• 2:34
  loudest
• 2:34
  requests
• 2:35
  are
• 2:35
  not
• 2:35
  for
• 2:35
  frontier
• 2:35
  models
• 2:36
  but
• 2:36
  for
• 2:36
  local
• 2:36
  ones.
• 2:38
  "We
• 2:38
  already
• 2:38
  know
• 2:38
  Gemini,
• 2:39
  ChatGPT,
• 2:39
  Claude,
• 2:40
  and
• 2:40
  DeepSeek
• 2:40
  are
• 2:40
  good,"
• 2:41
  one
• 2:41
  commenter
• 2:42
  notes.
• 2:42
  "We
• 2:42
  want
• 2:43
  to
• 2:43
  know
• 2:43
  how
• 2:43
  good
• 2:43
  local
• 2:43
  models
• 2:44
  do
• 2:44
  because
• 2:45
  many
• 2:45
  of
• 2:45
  them
• 2:45
  seem
• 2:45
  to
• 2:45
  be
• 2:45
  benchmaxxed."
• 2:46
  Another
• 2:47
  requests
• 2:47
  a
• 2:47
  fixed
• 2:48
  tool-call
• 2:48
  budget
• 2:48
  or
• 2:48
  wall-clock
• 2:49
  column,
• 2:50
  arguing
• 2:50
  that
• 2:50
  "pass
• 2:51
  rate
• 2:51
  without
• 2:51
  cost-to-fix
• 2:52
  is
• 2:52
  incomplete
• 2:53
  because
• 2:53
  a
• 2:53
  14B
• 2:54
  model
• 2:54
  that
• 2:54
  needs
• 2:54
  4
• 2:55
  retries
• 2:55
  is
• 2:55
  a
• 2:55
  totally
• 2:56
  different
• 2:56
  workflow
• 2:56
  than
• 2:56
  a
• 2:56
  70B
• 2:57
  that
• 2:57
  lands
• 2:57
  pass@1."
• 2:58
  The
• 2:58
  broader
• 2:59
  reasoning
• 2:59
  race
• 2:59
  reinforces
• 3:00
  the
• 3:00
  fatigue.
• 3:01
  Tencent's
• 3:02
  Hy3
• 3:02
  Preview
• 3:02
  scores
• 3:03
  87.8
• 3:04
  on
• 3:04
  the
• 3:04
  China
• 3:04
  High
• 3:04
  School
• 3:05
  Biology
• 3:05
  Olympiad,
• 3:06
  topping
• 3:06
  Gemini
• 3:07
  and
• 3:07
  GPT
• 3:07
  —
• 3:08
  and
• 3:08
  nobody
• 3:08
  quite
• 3:09
  knows
• 3:09
  what
• 3:09
  to
• 3:09
  do
• 3:09
  with
• 3:09
  that
• 3:09
  information.
• 3:10
  The
• 3:11
  benchmarks-as-GPU-TFLOP-flexing
• 3:13
  analogy
• 3:14
  is
• 3:14
  apt:
• 3:14
  Hy3
• 3:15
  can
• 3:15
  hit
• 3:15
  that
• 3:15
  level
• 3:16
  on
• 3:16
  math
• 3:16
  and
• 3:16
  code
• 3:16
  in
• 3:16
  real
• 3:17
  testing
• 3:17
  but,
• 3:18
  as
• 3:18
  one
• 3:18
  practitioner
• 3:18
  reports,
• 3:19
  "falls
• 3:19
  on
• 3:20
  its
• 3:20
  face
• 3:20
  in
• 3:20
  weird
• 3:20
  edge
• 3:21
  cases
• 3:21
  that
• 3:21
  never
• 3:21
  show
• 3:22
  up
• 3:22
  in
• 3:22
  charts."
• 3:23
  The
• 3:23
  consensus
• 3:23
  is
• 3:23
  hardening
• 3:24
  that
• 3:24
  you
• 3:24
  need
• 3:24
  your
• 3:24
  own
• 3:25
  benchmarks
• 3:25
  —
• 3:26
  evals
• 3:26
  on
• 3:26
  your
• 3:26
  domain
• 3:27
  data,
• 3:27
  with
• 3:27
  your
• 3:27
  edge
• 3:28
  cases,
• 3:28
  measuring
• 3:29
  what
• 3:29
  actually
• 3:29
  matters
• 3:30
  for
• 3:30
  your
• 3:30
  production
• 3:30
  workflow.
• 3:31
  "Benchmarks
• 3:32
  are
• 3:32
  effectively
• 3:32
  meaningless,"
• 3:33
  another
• 3:33
  commenter
• 3:34
  argues.
• 3:35
  "They
• 3:35
  are
• 3:35
  a
• 3:35
  metric
• 3:35
  that
• 3:35
  has
• 3:36
  become
• 3:36
  the
• 3:36
  goal."
• 3:37
  MiMo-V2.5-coder
• 3:40
  appeared
• 3:40
  this
• 3:40
  week
• 3:41
  billed
• 3:41
  as
• 3:41
  "an
• 3:41
  excellent
• 3:42
  alternative
• 3:42
  to
• 3:42
  Qwen3.6
• 3:45
  and
• 3:45
  DS4,
• 3:46
  especially
• 3:46
  for
• 3:46
  coding."
• 3:47
  The
• 3:47
  community
• 3:48
  was
• 3:48
  unconvinced.
• 3:49
  Multiple
• 3:49
  commenters
• 3:50
  pointed
• 3:50
  out
• 3:50
  it
• 3:50
  is
• 3:50
  not
• 3:50
  a
• 3:51
  fine-tune
• 3:51
  but
• 3:51
  a
• 3:51
  regular
• 3:52
  quantization
• 3:52
  with
• 3:52
  a
• 3:52
  slightly
• 3:53
  customized
• 3:53
  bits-per-layer
• 3:54
  profile
• 3:55
  and
• 3:55
  a
• 3:55
  coding-skewed
• 3:55
  imatrix,
• 3:56
  making
• 3:56
  the
• 3:57
  "-coder"
• 3:57
  suffix
• 3:57
  misleading.
• 3:58
  No
• 3:58
  benchmarks
• 3:59
  were
• 3:59
  provided
• 3:59
  to
• 4:00
  support
• 4:00
  the
• 4:00
  claims.
• 4:03
  Krasis,
• 4:04
  an
• 4:04
  LLM
• 4:04
  runtime
• 4:05
  purpose-built
• 4:05
  for
• 4:05
  models
• 4:06
  that
• 4:06
  exceed
• 4:06
  VRAM,
• 4:07
  hit
• 4:07
  its
• 4:07
  v1.0
• 4:09
  release
• 4:09
  with
• 4:09
  results
• 4:10
  that
• 4:10
  stretch
• 4:10
  what
• 4:10
  a
• 4:10
  single
• 4:11
  consumer
• 4:11
  GPU
• 4:11
  can
• 4:12
  do.
• 4:13
  The
• 4:13
  headline:
• 4:14
  Qwen3.6-35B-A3B
• 4:19
  at
• 4:19
  Q4
• 4:20
  quantization
• 4:20
  running
• 4:20
  at
• 4:21
  reading
• 4:21
  speed
• 4:21
  (12.5
• 4:22
  tokens/second
• 4:23
  decode)
• 4:23
  on
• 4:24
  a
• 4:24
  laptop
• 4:24
  RTX
• 4:24
  3070
• 4:25
  Mobile
• 4:25
  with
• 4:25
  8GB
• 4:26
  VRAM.
• 4:27
  On
• 4:27
  a
• 4:27
  5090,
• 4:28
  the
• 4:28
  same
• 4:28
  model
• 4:28
  hits
• 4:29
  124.9
• 4:30
  tok/s
• 4:31
  decode
• 4:31
  and
• 4:31
  over
• 4:31
  10,000
• 4:32
  tok/s
• 4:33
  prefill.
• 4:33
  The
• 4:34
  v1.0
• 4:35
  update
• 4:36
  is
• 4:36
  substantial
• 4:36
  —
• 4:37
  the
• 4:37
  hot
• 4:37
  path
• 4:37
  is
• 4:37
  now
• 4:38
  100%
• 4:38
  Rust
• 4:39
  (Python's
• 4:39
  GIL
• 4:40
  was
• 4:40
  causing
• 4:40
  frequent
• 4:41
  slowdowns),
• 4:42
  Ampere-generation
• 4:42
  cards
• 4:43
  are
• 4:43
  fully
• 4:43
  supported,
• 4:44
  memory
• 4:44
  requirements
• 4:45
  dropped
• 4:45
  from
• 4:45
  2x
• 4:46
  to
• 4:46
  1x
• 4:46
  the
• 4:46
  quantized
• 4:47
  model
• 4:47
  size
• 4:47
  plus
• 4:47
  overhead,
• 4:48
  and
• 4:48
  the
• 4:49
  old
• 4:49
  AWQ
• 4:49
  attention
• 4:50
  was
• 4:50
  replaced
• 4:50
  with
• 4:50
  sensitivity-aware
• 4:51
  HQQ
• 4:51
  attention
• 4:52
  at
• 4:52
  4,
• 4:52
  6,
• 4:52
  or
• 4:53
  8
• 4:53
  bits.
• 4:54
  HQQ
• 4:54
  assets
• 4:55
  are
• 4:55
  built
• 4:55
  by
• 4:55
  mathematically
• 4:56
  assessing
• 4:56
  the
• 4:56
  model
• 4:56
  rather
• 4:57
  than
• 4:57
  requiring
• 4:57
  a
• 4:57
  pre-built
• 4:58
  BF16
• 4:58
  template,
• 4:59
  and
• 4:59
  Krasis
• 5:00
  can
• 5:00
  mix
• 5:00
  precision
• 5:01
  —
• 5:01
  90%
• 5:02
  HQQ4
• 5:02
  +
• 5:03
  10%
• 5:04
  HQQ6
• 5:04
  for
• 5:04
  sensitive
• 5:05
  layers
• 5:05
  —
• 5:06
  keeping
• 5:06
  memory
• 5:06
  low
• 5:07
  while
• 5:07
  improving
• 5:07
  accuracy.
• 5:08
  A
• 5:08
  llama.cpp
• 5:09
  user
• 5:10
  pushed
• 5:10
  back,
• 5:11
  claiming
• 5:11
  3x
• 5:11
  better
• 5:12
  speeds
• 5:12
  with
• 5:12
  a
• 5:12
  ByteShape
• 5:13
  CPU-5
• 5:13
  quant
• 5:14
  on
• 5:14
  a
• 5:14
  6GB
• 5:14
  3060
• 5:15
  laptop.
• 5:16
  The
• 5:16
  comparison
• 5:17
  is
• 5:17
  not
• 5:17
  quite
• 5:17
  apples-to-apples
• 5:18
  (different
• 5:18
  quantization,
• 5:19
  different
• 5:19
  offload
• 5:20
  strategy),
• 5:21
  but
• 5:21
  it
• 5:21
  keeps
• 5:21
  the
• 5:21
  competitive
• 5:22
  pressure
• 5:22
  on.
• 5:23
  At
• 5:23
  the
• 5:23
  other
• 5:23
  extreme
• 5:23
  of
• 5:24
  the
• 5:24
  size
• 5:24
  spectrum,
• 5:25
  a
• 5:25
  head-to-head
• 5:25
  benchmark
• 5:26
  pitted
• 5:26
  Needle
• 5:26
  (26M
• 5:27
  parameters,
• 5:28
  distilled
• 5:28
  from
• 5:28
  Gemini
• 5:29
  3.1
• 5:29
  for
• 5:30
  function
• 5:30
  calls)
• 5:30
  against
• 5:31
  Qwen3-0.6B
• 5:34
  on
• 5:35
  CPU-only
• 5:35
  tool
• 5:35
  routing
• 5:36
  —
• 5:36
  50
• 5:36
  queries
• 5:37
  across
• 5:37
  five
• 5:37
  difficulty
• 5:38
  tiers.
• 5:39
  Needle
• 5:39
  won
• 5:39
  72%
• 5:40
  tool-match
• 5:40
  accuracy
• 5:41
  versus
• 5:41
  56%,
• 5:42
  at
• 5:43
  4.4x
• 5:44
  lower
• 5:44
  latency.
• 5:45
  The
• 5:45
  interesting
• 5:46
  finding
• 5:46
  is
• 5:46
  not
• 5:46
  the
• 5:46
  headline
• 5:47
  but
• 5:47
  the
• 5:47
  failure
• 5:47
  shapes:
• 5:48
  Needle
• 5:49
  fails
• 5:49
  by
• 5:49
  picking
• 5:49
  the
• 5:49
  wrong
• 5:50
  tool
• 5:50
  but
• 5:50
  gets
• 5:50
  arguments
• 5:51
  right
• 5:51
  97%
• 5:52
  of
• 5:52
  the
• 5:52
  time;
• 5:53
  Qwen3
• 5:53
  fails
• 5:53
  by
• 5:54
  not
• 5:54
  calling
• 5:54
  a
• 5:54
  tool
• 5:54
  at
• 5:54
  all,
• 5:55
  answering
• 5:55
  in
• 5:56
  prose
• 5:56
  instead.
• 5:57
  At
• 5:57
  tier
• 5:57
  3
• 5:58
  (implicit
• 5:58
  queries
• 5:59
  like
• 5:59
  "should
• 5:59
  I
• 6:00
  bring
• 6:00
  an
• 6:00
  umbrella
• 6:00
  in
• 6:00
  Amsterdam?"),
• 6:01
  Needle
• 6:02
  holds
• 6:02
  at
• 6:02
  80%
• 6:03
  while
• 6:03
  Qwen3
• 6:03
  collapses
• 6:04
  to
• 6:04
  10%.
• 6:06
  For
• 6:06
  agent-as-operator
• 6:07
  workloads,
• 6:08
  a
• 6:08
  model
• 6:08
  that
• 6:08
  occasionally
• 6:09
  routes
• 6:09
  wrong
• 6:09
  but
• 6:09
  always
• 6:10
  acts
• 6:10
  is
• 6:10
  debuggable;
• 6:11
  a
• 6:12
  model
• 6:12
  that
• 6:12
  goes
• 6:12
  quiet
• 6:13
  under
• 6:13
  ambiguity
• 6:14
  is
• 6:14
  a
• 6:14
  production
• 6:14
  halt.
• 6:15
  On
• 6:15
  the
• 6:15
  hardware
• 6:16
  side,
• 6:16
  a
• 6:16
  detailed
• 6:17
  comparison
• 6:18
  of
• 6:18
  the
• 6:18
  RTX
• 6:18
  5090
• 6:20
  versus
• 6:20
  RTX
• 6:21
  6000
• 6:21
  PRO
• 6:22
  MaxQ
• 6:23
  on
• 6:23
  diffusion
• 6:23
  workloads
• 6:24
  found
• 6:24
  the
• 6:24
  MaxQ
• 6:25
  matches
• 6:25
  the
• 6:25
  5090
• 6:26
  at
• 6:26
  400W
• 6:27
  while
• 6:28
  consuming
• 6:28
  only
• 6:28
  325W
• 6:29
  —
• 6:30
  75%
• 6:31
  of
• 6:31
  the
• 6:31
  power
• 6:31
  for
• 6:32
  identical
• 6:32
  performance.
• 6:34
  At
• 6:34
  full
• 6:34
  600W,
• 6:35
  the
• 6:35
  5090
• 6:36
  leads
• 6:37
  by
• 6:37
  25%,
• 6:38
  but
• 6:39
  the
• 6:39
  efficiency
• 6:39
  curve
• 6:39
  matters
• 6:40
  for
• 6:40
  anyone
• 6:40
  building
• 6:41
  dense
• 6:41
  multi-GPU
• 6:42
  rigs
• 6:42
  where
• 6:42
  thermal
• 6:43
  headroom
• 6:43
  is
• 6:43
  the
• 6:43
  real
• 6:43
  constraint.
• 6:45
  Chinese
• 6:45
  memory
• 6:46
  manufacturer
• 6:46
  CXMT
• 6:47
  entering
• 6:47
  the
• 6:48
  Corsair
• 6:48
  DDR5
• 6:49
  supply
• 6:49
  chain
• 6:49
  could
• 6:49
  eventually
• 6:50
  help
• 6:50
  the
• 6:50
  memory
• 6:50
  pricing
• 6:51
  picture,
• 6:52
  though
• 6:52
  modules
• 6:52
  are
• 6:52
  currently
• 6:53
  China-only
• 6:54
  and
• 6:54
  the
• 6:54
  real
• 6:54
  pressure
• 6:54
  point
• 6:55
  —
• 6:56
  HBM
• 6:57
  wafer
• 6:57
  allocation
• 6:58
  cannibalizing
• 6:58
  DDR5
• 6:59
  and
• 6:59
  GDDR
• 7:00
  production
• 7:00
  —
• 7:01
  remains
• 7:02
  unresolved.
• 7:03
  Separately,
• 7:04
  a
• 7:04
  community-driven
• 7:05
  TTS
• 7:05
  benchmark
• 7:06
  covering
• 7:06
  all
• 7:06
  major
• 7:07
  engines
• 7:07
  through
• 7:08
  May
• 7:08
  2026
• 7:09
  finally
• 7:09
  provides
• 7:10
  centralized
• 7:10
  speed
• 7:10
  comparisons
• 7:11
  across
• 7:11
  platforms,
• 7:13
  though
• 7:13
  commenters
• 7:13
  rightly
• 7:14
  note
• 7:14
  that
• 7:14
  real-time
• 7:15
  factor,
• 7:15
  memory
• 7:16
  usage,
• 7:16
  and
• 7:17
  quality
• 7:17
  tradeoffs
• 7:17
  matter
• 7:18
  more
• 7:18
  than
• 7:18
  cherry-picked
• 7:19
  demo
• 7:19
  clips.
• 7:22
  "Harness
• 7:22
  engineering"
• 7:23
  is
• 7:23
  solidifying
• 7:24
  as
• 7:24
  the
• 7:24
  term-of-art
• 7:24
  for
• 7:25
  the
• 7:25
  discipline
• 7:25
  of
• 7:25
  wrapping
• 7:26
  AI
• 7:26
  coding
• 7:26
  agents
• 7:26
  in
• 7:27
  rules,
• 7:27
  tools,
• 7:28
  hooks,
• 7:28
  memory,
• 7:28
  and
• 7:28
  orchestration
• 7:29
  scaffolding.
• 7:30
  Cole
• 7:30
  Medin's
• 7:30
  editorial
• 7:31
  lays
• 7:31
  out
• 7:31
  the
• 7:31
  two-level
• 7:32
  framing:
• 7:33
  the
• 7:33
  first
• 7:33
  level
• 7:33
  is
• 7:33
  a
• 7:33
  single
• 7:34
  agent
• 7:34
  session
• 7:35
  (your
• 7:35
  CLAUDE.md
• 7:36
  rules,
• 7:37
  MCP
• 7:37
  servers,
• 7:38
  hooks,
• 7:38
  skills
• 7:38
  —
• 7:39
  the
• 7:39
  "AI
• 7:40
  layer"),
• 7:40
  and
• 7:40
  the
• 7:40
  second
• 7:41
  is
• 7:41
  orchestrating
• 7:42
  multiple
• 7:42
  sessions
• 7:42
  into
• 7:43
  a
• 7:43
  workflow
• 7:43
  where
• 7:43
  one
• 7:44
  plans,
• 7:44
  one
• 7:45
  implements,
• 7:45
  one
• 7:45
  validates,
• 7:46
  and
• 7:46
  handoffs
• 7:47
  are
• 7:47
  automated.
• 7:48
  The
• 7:48
  key
• 7:48
  mindset
• 7:48
  shift:
• 7:49
  "When
• 7:49
  the
• 7:49
  agent
• 7:49
  does
• 7:49
  something
• 7:50
  dumb,
• 7:50
  you
• 7:50
  don't
• 7:51
  just
• 7:51
  blame
• 7:51
  the
• 7:51
  model
• 7:51
  —
• 7:52
  you
• 7:52
  improve
• 7:52
  the
• 7:52
  harness.
• 7:53
  It
• 7:54
  missed
• 7:54
  a
• 7:54
  convention,
• 7:55
  so
• 7:55
  that
• 7:55
  becomes
• 7:55
  a
• 7:55
  rule.
• 7:56
  It
• 7:57
  ran
• 7:57
  something
• 7:57
  destructive,
• 7:58
  so
• 7:58
  a
• 7:58
  hook
• 7:58
  now
• 7:59
  blocks
• 7:59
  it."
• 8:00
  A
• 8:00
  skeptic
• 8:00
  in
• 8:00
  the
• 8:00
  comments
• 8:00
  offers
• 8:01
  the
• 8:01
  necessary
• 8:01
  counterweight:
• 8:02
  "nobody
• 8:03
  could
• 8:03
  tell
• 8:03
  you
• 8:03
  what
• 8:03
  'context
• 8:04
  engineering'
• 8:04
  means,
• 8:05
  and
• 8:05
  they
• 8:05
  couldn't
• 8:05
  tell
• 8:05
  you
• 8:05
  what
• 8:05
  'prompt
• 8:06
  engineering'
• 8:06
  meant
• 8:06
  before
• 8:07
  that...
• 8:08
  we're
• 8:08
  still
• 8:08
  very
• 8:08
  much
• 8:08
  in
• 8:09
  the
• 8:09
  Land
• 8:09
  of
• 8:09
  Feels
• 8:09
  and
• 8:09
  Vibes."
• 8:10
  Fair
• 8:10
  point
• 8:10
  —
• 8:11
  but
• 8:11
  the
• 8:11
  practitioner
• 8:12
  who
• 8:12
  grew
• 8:12
  their
• 8:12
  rules
• 8:13
  file
• 8:13
  from
• 8:13
  10
• 8:13
  lines
• 8:14
  to
• 8:14
  400
• 8:14
  and
• 8:14
  cut
• 8:15
  recurring
• 8:15
  task
• 8:15
  errors
• 8:16
  by
• 8:16
  half
• 8:16
  represents
• 8:17
  real
• 8:17
  signal.
• 8:18
  The
• 8:18
  pacphi
• 8:19
  Gist
• 8:19
  takes
• 8:19
  this
• 8:19
  further
• 8:20
  with
• 8:20
  a
• 8:20
  concrete,
• 8:21
  end-to-end
• 8:21
  autonomous
• 8:22
  workflow.
• 8:23
  The
• 8:23
  setup:
• 8:24
  read
• 8:24
  your
• 8:24
  ADRs
• 8:25
  and
• 8:25
  domain
• 8:25
  models,
• 8:26
  seed
• 8:27
  a
• 8:27
  GitHub
• 8:27
  issue
• 8:27
  backlog
• 8:28
  with
• 8:28
  dependency
• 8:28
  graphs
• 8:29
  and
• 8:29
  success
• 8:29
  criteria,
• 8:30
  then
• 8:31
  launch
• 8:31
  an
• 8:31
  autonomous
• 8:31
  loop
• 8:32
  where
• 8:32
  a
• 8:32
  fresh
• 8:32
  agent
• 8:33
  session
• 8:33
  branches,
• 8:33
  implements
• 8:34
  via
• 8:34
  the
• 8:34
  SPARC
• 8:35
  orchestrator
• 8:35
  skill,
• 8:36
  creates
• 8:36
  a
• 8:36
  PR,
• 8:37
  monitors
• 8:38
  CI
• 8:38
  with
• 8:38
  automatic
• 8:39
  flake
• 8:39
  retries,
• 8:40
  runs
• 8:40
  a
• 8:40
  multi-stage
• 8:41
  automated
• 8:41
  review
• 8:41
  gate
• 8:42
  (code
• 8:42
  review,
• 8:43
  issue-criteria
• 8:44
  matching,
• 8:44
  quality
• 8:45
  scoring),
• 8:45
  and
• 8:45
  auto-merges
• 8:46
  when
• 8:46
  zero
• 8:47
  blocking
• 8:47
  findings
• 8:47
  remain.
• 8:49
  The
• 8:49
  human
• 8:49
  role
• 8:49
  reduces
• 8:50
  to
• 8:50
  reviewing
• 8:50
  PRs
• 8:51
  that
• 8:51
  get
• 8:51
  flagged.
• 8:52
  Whether
• 8:53
  this
• 8:53
  works
• 8:53
  reliably
• 8:53
  at
• 8:54
  scale
• 8:54
  is
• 8:54
  an
• 8:54
  open
• 8:54
  question,
• 8:55
  but
• 8:55
  the
• 8:55
  machinery
• 8:56
  is
• 8:56
  impressively
• 8:56
  complete.
• 8:58
  Three
• 8:58
  new
• 8:58
  tools
• 8:58
  flesh
• 8:59
  out
• 8:59
  the
• 8:59
  ecosystem.
• 9:00
  piia-engram
• 9:01
  addresses
• 9:01
  a
• 9:01
  genuine
• 9:02
  pain
• 9:02
  point:
• 9:03
  every
• 9:03
  time
• 9:03
  you
• 9:03
  switch
• 9:04
  between
• 9:04
  Claude
• 9:04
  Code,
• 9:05
  Cursor,
• 9:05
  and
• 9:06
  Codex,
• 9:06
  your
• 9:07
  preferences,
• 9:08
  code
• 9:08
  standards,
• 9:09
  and
• 9:09
  hard-won
• 9:09
  lessons
• 9:10
  reset
• 9:10
  to
• 9:10
  zero.
• 9:11
  It
• 9:12
  stores
• 9:12
  your
• 9:12
  developer
• 9:12
  identity
• 9:13
  as
• 9:13
  local
• 9:13
  JSON
• 9:14
  files
• 9:14
  under
• 9:14
  ~/.engram/
• 9:16
  and
• 9:16
  exposes
• 9:17
  them
• 9:17
  to
• 9:17
  every
• 9:17
  MCP-compatible
• 9:19
  tool,
• 9:19
  so
• 9:19
  new
• 9:20
  sessions
• 9:20
  start
• 9:20
  with
• 9:20
  full
• 9:21
  context.
• 9:22
  It
• 9:22
  explicitly
• 9:22
  positions
• 9:23
  itself
• 9:23
  as
• 9:24
  storing
• 9:24
  "who
• 9:24
  you
• 9:25
  are
• 9:25
  as
• 9:25
  a
• 9:25
  person"
• 9:26
  rather
• 9:26
  than
• 9:26
  agent
• 9:26
  session
• 9:27
  history
• 9:27
  —
• 9:28
  a
• 9:28
  different
• 9:28
  layer
• 9:29
  than
• 9:29
  Mem0
• 9:29
  or
• 9:29
  Zep.
• 9:30
  Tactile
• 9:31
  inverts
• 9:31
  the
• 9:31
  standard
• 9:32
  computer-use
• 9:32
  approach:
• 9:33
  instead
• 9:34
  of
• 9:34
  screenshot-guess-click
• 9:35
  loops,
• 9:35
  agents
• 9:36
  first
• 9:36
  inspect
• 9:36
  accessibility
• 9:37
  semantics
• 9:37
  (element
• 9:38
  roles,
• 9:39
  names,
• 9:40
  states,
• 9:40
  hierarchy)
• 9:41
  exposed
• 9:42
  by
• 9:42
  the
• 9:42
  OS,
• 9:43
  fall
• 9:43
  back
• 9:43
  to
• 9:43
  OCR-grounded
• 9:44
  coordinates,
• 9:45
  and
• 9:45
  only
• 9:45
  then
• 9:45
  resort
• 9:46
  to
• 9:46
  visual
• 9:46
  reasoning.
• 9:47
  The
• 9:47
  insight
• 9:48
  is
• 9:48
  that
• 9:48
  agent-ready
• 9:49
  software
• 9:49
  should
• 9:49
  also
• 9:50
  be
• 9:50
  human-accessible
• 9:51
  software,
• 9:52
  and
• 9:52
  building
• 9:52
  agents
• 9:53
  that
• 9:53
  depend
• 9:53
  on
• 9:53
  accessibility
• 9:54
  APIs
• 9:55
  creates
• 9:55
  economic
• 9:55
  pressure
• 9:56
  to
• 9:56
  fix
• 9:56
  long-standing
• 9:57
  gaps
• 9:57
  that
• 9:57
  affect
• 9:58
  screen
• 9:58
  reader
• 9:58
  users
• 9:58
  too.
• 10:00
  Hadrian's
• 10:00
  OpenHack
• 10:01
  brings
• 10:01
  agent
• 10:01
  orchestration
• 10:02
  to
• 10:02
  whitebox
• 10:03
  security
• 10:03
  review
• 10:03
  —
• 10:04
  a
• 10:04
  file-based
• 10:05
  state
• 10:05
  machine
• 10:06
  where
• 10:06
  recon
• 10:06
  discovers
• 10:07
  attack
• 10:07
  surfaces,
• 10:08
  a
• 10:08
  router
• 10:08
  generates
• 10:09
  scoped
• 10:09
  vulnerability
• 10:10
  scenarios,
• 10:11
  expert
• 10:11
  agents
• 10:11
  prove
• 10:12
  or
• 10:12
  reject
• 10:12
  each
• 10:13
  one,
• 10:13
  and
• 10:13
  an
• 10:13
  independent
• 10:14
  triage
• 10:14
  agent
• 10:15
  evaluates
• 10:15
  reportability.
• 10:16
  Human
• 10:17
  operators
• 10:17
  approve
• 10:18
  every
• 10:18
  phase
• 10:18
  transition.
• 10:20
  Twelve
• 10:20
  expert
• 10:20
  agent
• 10:20
  families
• 10:21
  cover
• 10:21
  the
• 10:21
  OWASP
• 10:22
  Top
• 10:22
  10:2025.
• 10:28
  A
• 10:28
  security
• 10:28
  researcher
• 10:29
  poking
• 10:29
  at
• 10:29
  a
• 10:29
  fintech's
• 10:30
  mobile
• 10:30
  API
• 10:30
  discovered
• 10:31
  that
• 10:31
  adding
• 10:31
  a
• 10:31
  trailing
• 10:32
  slash
• 10:32
  to
• 10:32
  an
• 10:32
  authenticated
• 10:33
  endpoint
• 10:33
  flipped
• 10:34
  the
• 10:34
  response
• 10:34
  from
• 10:34
  401
• 10:35
  to
• 10:35
  200
• 10:35
  with
• 10:36
  full
• 10:36
  account
• 10:36
  data.
• 10:37
  The
• 10:37
  target
• 10:37
  ran
• 10:37
  AWS
• 10:38
  HTTP
• 10:39
  API
• 10:39
  (the
• 10:40
  newer,
• 10:40
  cheaper
• 10:40
  alternative
• 10:41
  to
• 10:41
  REST
• 10:41
  API)
• 10:42
  with
• 10:42
  a
• 10:42
  Lambda
• 10:42
  authorizer
• 10:43
  checking
• 10:43
  JWTs
• 10:44
  against
• 10:44
  Cognito.
• 10:45
  The
• 10:45
  bug:
• 10:46
  HTTP
• 10:47
  API's
• 10:47
  greedy
• 10:48
  path
• 10:48
  matching
• 10:49
  rewrote
• 10:49
  the
• 10:49
  trailing-slash
• 10:50
  path,
• 10:50
  stripping
• 10:50
  the
• 10:50
  slash
• 10:51
  before
• 10:51
  forwarding
• 10:51
  to
• 10:52
  the
• 10:52
  integration,
• 10:53
  but
• 10:53
  the
• 10:53
  auth
• 10:53
  check
• 10:54
  ran
• 10:54
  on
• 10:54
  the
• 10:54
  original
• 10:54
  path.
• 10:56
  The
• 10:56
  rewrite
• 10:56
  dropped
• 10:56
  the
• 10:56
  auth
• 10:57
  context.
• 10:58
  The
• 10:58
  researcher
• 10:58
  confirmed
• 10:58
  the
• 10:58
  bypass
• 10:59
  extended
• 10:59
  to
• 10:59
  wire
• 11:00
  transfer
• 11:00
  initiation
• 11:00
  —
• 11:01
  a
• 11:01
  $0.01
• 11:02
  test
• 11:02
  transfer
• 11:02
  went
• 11:03
  through
• 11:03
  without
• 11:03
  a
• 11:03
  valid
• 11:03
  JWT.
• 11:05
  The
• 11:05
  fix
• 11:05
  was
• 11:05
  straightforward:
• 11:07
  switch
• 11:07
  from
• 11:07
  HTTP
• 11:08
  API
• 11:09
  to
• 11:09
  REST
• 11:09
  API
• 11:10
  (stricter
• 11:10
  path
• 11:11
  matching)
• 11:12
  and
• 11:12
  add
• 11:12
  userId
• 11:12
  validation
• 11:13
  in
• 11:13
  every
• 11:14
  Lambda,
• 11:14
  not
• 11:15
  just
• 11:15
  the
• 11:15
  authorizer.
• 11:16
  The
• 11:16
  bounty:
• 11:17
  $12,000.
• 11:19
  This
• 11:19
  fits
• 11:19
  a
• 11:20
  pattern
• 11:20
  that
• 11:20
  keeps
• 11:20
  repeating:
• 11:21
  a
• 11:21
  trivial
• 11:22
  input
• 11:22
  normalization
• 11:23
  mismatch
• 11:23
  defeats
• 11:23
  sophisticated
• 11:24
  auth
• 11:24
  middleware.
• 11:26
  A
• 11:26
  malformed
• 11:26
  Host
• 11:26
  header
• 11:27
  bypassing
• 11:27
  Starlette/FastAPI
• 11:29
  auth.
• 11:30
  Two
• 11:30
  missing
• 11:30
  regex
• 11:31
  characters
• 11:31
  exposing
• 11:32
  AWS
• 11:32
  CodeBuild.
• 11:33
  An
• 11:34
  unsanitized
• 11:34
  semicolon
• 11:35
  enabling
• 11:35
  GitHub
• 11:36
  RCE.
• 11:37
  The
• 11:37
  lesson
• 11:37
  is
• 11:38
  structural:
• 11:38
  when
• 11:38
  the
• 11:39
  routing
• 11:39
  layer
• 11:39
  and
• 11:39
  the
• 11:39
  auth
• 11:39
  layer
• 11:40
  disagree
• 11:40
  on
• 11:40
  what
• 11:40
  constitutes
• 11:41
  a
• 11:41
  "match,"
• 11:42
  the
• 11:42
  gap
• 11:42
  between
• 11:42
  them
• 11:42
  becomes
• 11:43
  an
• 11:43
  exploit.
• 11:44
  No
• 11:44
  amount
• 11:45
  of
• 11:45
  JWT
• 11:45
  complexity
• 11:46
  or
• 11:46
  Cognito
• 11:46
  configuration
• 11:47
  can
• 11:48
  compensate
• 11:48
  for
• 11:48
  path-parsing
• 11:49
  inconsistency.
• 11:50
  On
• 11:50
  the
• 11:50
  network
• 11:51
  defense
• 11:51
  side,
• 11:51
  Mullvad
• 11:52
  is
• 11:52
  rolling
• 11:52
  out
• 11:52
  mitigations
• 11:53
  against
• 11:53
  exit
• 11:54
  IP
• 11:54
  fingerprinting
• 11:54
  across
• 11:55
  13
• 11:55
  WireGuard
• 11:56
  servers
• 11:56
  spanning
• 11:56
  Australia,
• 11:57
  Canada,
• 11:58
  Germany,
• 11:58
  Finland,
• 11:59
  France,
• 11:59
  Ireland,
• 12:00
  Norway,
• 12:01
  Sweden,
• 12:01
  and
• 12:01
  the
• 12:01
  US.
• 12:02
  The
• 12:03
  mitigation
• 12:03
  addresses
• 12:04
  research
• 12:04
  showing
• 12:04
  that
• 12:04
  VPN
• 12:05
  exit
• 12:05
  IP
• 12:05
  behavior
• 12:06
  could
• 12:06
  be
• 12:06
  used
• 12:06
  to
• 12:06
  correlate
• 12:07
  traffic
• 12:07
  across
• 12:08
  sessions.
• 12:09
  It
• 12:09
  is
• 12:09
  an
• 12:09
  incremental
• 12:09
  improvement,
• 12:10
  but
• 12:11
  the
• 12:11
  arms
• 12:11
  race
• 12:11
  extends
• 12:12
  well
• 12:12
  beyond
• 12:12
  IP
• 12:13
  addresses
• 12:13
  —
• 12:14
  browser
• 12:14
  fingerprinting
• 12:15
  via
• 12:15
  GPU,
• 12:15
  fonts,
• 12:16
  canvas,
• 12:16
  WebGL,
• 12:17
  and
• 12:17
  behavioral
• 12:18
  biometrics
• 12:18
  has
• 12:19
  been
• 12:19
  demonstrated
• 12:19
  to
• 12:19
  uniquely
• 12:20
  identify
• 12:20
  machines
• 12:21
  even
• 12:21
  through
• 12:21
  VPN
• 12:22
  tunnels.
• 12:24
  Every
• 12:25
  async
• 12:25
  reinforcement
• 12:26
  learning
• 12:26
  library
• 12:26
  trips
• 12:26
  over
• 12:27
  the
• 12:27
  same
• 12:27
  bottleneck:
• 12:28
  shipping
• 12:28
  the
• 12:28
  entire
• 12:28
  model
• 12:29
  from
• 12:29
  trainer
• 12:29
  to
• 12:29
  inference
• 12:30
  engine
• 12:30
  after
• 12:30
  every
• 12:31
  optimizer
• 12:31
  step.
• 12:32
  For
• 12:32
  a
• 12:32
  7B
• 12:33
  model
• 12:33
  in
• 12:33
  bf16,
• 12:34
  that
• 12:34
  is
• 12:34
  14GB.
• 12:35
  For
• 12:36
  a
• 12:36
  frontier
• 12:36
  1T
• 12:36
  model,
• 12:37
  it
• 12:37
  is
• 12:37
  a
• 12:37
  terabyte.
• 12:38
  Per
• 12:39
  step.
• 12:40
  Hugging
• 12:40
  Face's
• 12:41
  new
• 12:41
  delta
• 12:41
  weight
• 12:41
  sync
• 12:41
  in
• 12:41
  TRL
• 12:42
  exploits
• 12:43
  an
• 12:43
  elegant
• 12:43
  property
• 12:43
  of
• 12:44
  bf16
• 12:44
  arithmetic:
• 12:45
  at
• 12:45
  typical
• 12:45
  RL
• 12:46
  learning
• 12:46
  rates
• 12:47
  (~3e-6),
• 12:48
  Adam's
• 12:48
  per-weight
• 12:49
  update
• 12:49
  magnitude
• 12:49
  is
• 12:50
  smaller
• 12:50
  than
• 12:50
  the
• 12:50
  spacing
• 12:51
  between
• 12:51
  adjacent
• 12:51
  bf16
• 12:52
  representable
• 12:52
  values.
• 12:53
  The
• 12:54
  optimizer
• 12:54
  is
• 12:54
  whispering
• 12:55
  below
• 12:55
  bf16's
• 12:56
  hearing
• 12:56
  threshold.
• 12:57
  The
• 12:57
  result:
• 12:58
  roughly
• 12:58
  99%
• 12:59
  of
• 12:59
  weights
• 12:59
  are
• 12:59
  bit-identical
• 13:00
  between
• 13:00
  consecutive
• 13:01
  steps,
• 13:01
  with
• 13:02
  the
• 13:02
  worst
• 13:02
  case
• 13:02
  never
• 13:02
  dropping
• 13:03
  below
• 13:03
  98%.
• 13:04
  Ship
• 13:05
  only
• 13:05
  the
• 13:05
  changed
• 13:05
  elements
• 13:06
  as
• 13:06
  a
• 13:06
  sparse
• 13:06
  safetensors
• 13:07
  file,
• 13:08
  upload
• 13:08
  it
• 13:08
  to
• 13:08
  a
• 13:08
  Hub
• 13:08
  Bucket,
• 13:09
  and
• 13:09
  the
• 13:09
  per-step
• 13:10
  payload
• 13:10
  drops
• 13:10
  from
• 13:10
  1.2GB
• 13:11
  to
• 13:11
  20-35MB
• 13:14
  on
• 13:14
  Qwen3-0.6B.
• 13:18
  The
• 13:18
  team
• 13:18
  demonstrated
• 13:19
  a
• 13:19
  fully
• 13:19
  disaggregated
• 13:20
  training
• 13:20
  where
• 13:20
  the
• 13:20
  trainer
• 13:21
  ran
• 13:21
  on
• 13:21
  one
• 13:21
  box,
• 13:22
  vLLM
• 13:22
  lived
• 13:23
  in
• 13:23
  a
• 13:23
  Hugging
• 13:23
  Face
• 13:23
  Space,
• 13:24
  and
• 13:24
  the
• 13:24
  Wordle
• 13:24
  environment
• 13:25
  lived
• 13:25
  in
• 13:25
  another
• 13:25
  Space
• 13:25
  —
• 13:26
  no
• 13:26
  shared
• 13:26
  cluster,
• 13:27
  no
• 13:27
  RDMA,
• 13:28
  no
• 13:28
  VPN.
• 13:29
  Weights
• 13:29
  flowed
• 13:30
  through
• 13:30
  a
• 13:30
  single
• 13:30
  bucket.
• 13:31
  For
• 13:31
  a
• 13:31
  405B
• 13:32
  model,
• 13:33
  the
• 13:33
  delta
• 13:33
  would
• 13:33
  be
• 13:33
  roughly
• 13:34
  6GB
• 13:34
  versus
• 13:35
  810GB
• 13:37
  for
• 13:37
  a
• 13:37
  full
• 13:37
  sync.
• 13:38
  At
• 13:38
  1GB/s
• 13:40
  of
• 13:40
  usable
• 13:40
  internet
• 13:41
  bandwidth,
• 13:41
  a
• 13:41
  cross-cloud
• 13:42
  full
• 13:42
  broadcast
• 13:43
  takes
• 13:43
  13
• 13:43
  minutes;
• 13:44
  the
• 13:44
  delta
• 13:45
  path
• 13:45
  does
• 13:45
  it
• 13:45
  in
• 13:45
  6
• 13:45
  seconds.
• 13:46
  On
• 13:46
  the
• 13:47
  privacy
• 13:47
  side,
• 13:48
  PACZero
• 13:48
  introduces
• 13:49
  a
• 13:49
  family
• 13:49
  of
• 13:49
  PAC-private
• 13:50
  zeroth-order
• 13:51
  mechanisms
• 13:51
  for
• 13:52
  fine-tuning
• 13:52
  LLMs
• 13:53
  that
• 13:53
  delivers
• 13:54
  usable
• 13:54
  accuracy
• 13:55
  at
• 13:55
  privacy
• 13:55
  levels
• 13:56
  no
• 13:56
  existing
• 13:56
  differential
• 13:57
  privacy
• 13:57
  method
• 13:58
  can
• 13:58
  match.
• 13:59
  The
• 13:59
  core
• 13:59
  insight:
• 13:59
  sign-quantizing
• 14:00
  subset-aggregated
• 14:01
  zeroth-order
• 14:02
  gradients
• 14:02
  creates
• 14:03
  frequent
• 14:03
  "unanimity"
• 14:04
  steps
• 14:04
  where
• 14:04
  every
• 14:05
  candidate
• 14:05
  subset
• 14:05
  agrees
• 14:06
  on
• 14:06
  the
• 14:06
  update
• 14:06
  direction,
• 14:07
  costing
• 14:08
  zero
• 14:08
  mutual
• 14:08
  information
• 14:09
  under
• 14:09
  PAC
• 14:09
  Privacy.
• 14:10
  On
• 14:10
  SST-2
• 14:11
  with
• 14:11
  OPT-1.3B,
• 14:13
  the
• 14:13
  budgeted
• 14:13
  variant
• 14:14
  reaches
• 14:14
  90.8%
• 14:15
  accuracy
• 14:16
  at
• 14:16
  a
• 14:16
  total
• 14:16
  MI
• 14:16
  budget
• 14:17
  of
• 14:17
  0.01
• 14:17
  nats
• 14:17
  —
• 14:18
  within
• 14:19
  1.4
• 14:19
  percentage
• 14:20
  points
• 14:20
  of
• 14:20
  the
• 14:20
  non-private
• 14:21
  baseline.
• 14:22
  The
• 14:22
  zero-privacy-loss
• 14:23
  variant
• 14:24
  (PACZero-ZPL)
• 14:26
  replaces
• 14:26
  noised
• 14:27
  releases
• 14:27
  with
• 14:27
  uniform
• 14:28
  coin
• 14:28
  flips
• 14:28
  on
• 14:28
  disagreement
• 14:29
  steps,
• 14:29
  achieving
• 14:30
  MI=0
• 14:31
  for
• 14:31
  every
• 14:31
  step
• 14:32
  count
• 14:32
  while
• 14:32
  still
• 14:32
  maintaining
• 14:33
  competitive
• 14:33
  accuracy
• 14:34
  (89.1%
• 14:35
  on
• 14:36
  OPT-6.7B).
• 14:38
  By
• 14:38
  contrast,
• 14:39
  all
• 14:39
  DP-ZO
• 14:40
  baselines
• 14:40
  collapse
• 14:41
  to
• 14:41
  chance
• 14:41
  accuracy
• 14:42
  below
• 14:42
  epsilon=2.
• 14:44
  The
• 14:44
  unanimity
• 14:44
  branch
• 14:44
  fires
• 14:45
  on
• 14:45
  34-45%
• 14:47
  of
• 14:47
  SST-2
• 14:48
  training
• 14:48
  steps,
• 14:48
  providing
• 14:49
  directed
• 14:49
  optimization
• 14:50
  at
• 14:50
  literally
• 14:50
  zero
• 14:51
  privacy
• 14:51
  cost.
• 14:52
  Whether
• 14:52
  these
• 14:52
  rates
• 14:53
  persist
• 14:53
  on
• 14:53
  architectures
• 14:54
  beyond
• 14:54
  OPT
• 14:55
  and
• 14:55
  at
• 14:55
  larger
• 14:55
  scales
• 14:56
  is
• 14:56
  the
• 14:56
  open
• 14:56
  question.
• 15:01
  The
• 15:01
  mesh
• 15:02
  networking
• 15:02
  space
• 15:02
  has
• 15:03
  three
• 15:03
  contenders,
• 15:04
  and
• 15:04
  a
• 15:04
  detailed
• 15:04
  comparison
• 15:05
  argues
• 15:05
  only
• 15:05
  one
• 15:05
  has
• 15:06
  a
• 15:06
  real
• 15:06
  future
• 15:06
  at
• 15:06
  scale.
• 15:07
  Meshtastic,
• 15:08
  the
• 15:08
  first-mover
• 15:09
  in
• 15:09
  consumer
• 15:09
  LoRa
• 15:09
  mesh,
• 15:10
  works
• 15:10
  well
• 15:10
  for
• 15:10
  small
• 15:11
  private
• 15:11
  groups
• 15:11
  but
• 15:11
  uses
• 15:12
  flood
• 15:12
  routing
• 15:12
  —
• 15:13
  broadcasting
• 15:13
  every
• 15:14
  message
• 15:14
  across
• 15:14
  the
• 15:14
  entire
• 15:15
  network
• 15:15
  —
• 15:16
  which
• 15:16
  collapses
• 15:16
  under
• 15:16
  load.
• 15:17
  MeshCore
• 15:18
  improves
• 15:18
  on
• 15:18
  this
• 15:18
  with
• 15:19
  actual
• 15:19
  path-based
• 15:19
  routing
• 15:20
  (up
• 15:20
  to
• 15:20
  64
• 15:21
  hops)
• 15:22
  and
• 15:22
  dramatically
• 15:22
  reduced
• 15:23
  radio
• 15:23
  congestion,
• 15:24
  but
• 15:24
  its
• 15:24
  official
• 15:24
  clients
• 15:25
  are
• 15:25
  proprietary
• 15:25
  with
• 15:26
  paywalled
• 15:26
  features.
• 15:27
  "The
• 15:27
  point
• 15:28
  of
• 15:28
  having
• 15:28
  an
• 15:28
  off-grid
• 15:28
  mesh
• 15:29
  network
• 15:29
  in
• 15:29
  the
• 15:29
  first
• 15:29
  place
• 15:30
  is
• 15:30
  total
• 15:30
  freedom
• 15:30
  and
• 15:30
  control,"
• 15:31
  the
• 15:32
  author
• 15:32
  argues,
• 15:33
  "so
• 15:33
  in
• 15:33
  this
• 15:33
  particular
• 15:34
  case
• 15:34
  I
• 15:34
  simply
• 15:34
  cannot
• 15:34
  ever
• 15:35
  support
• 15:35
  a
• 15:35
  closed-source
• 15:35
  solution."
• 15:36
  Reticulum
• 15:37
  is
• 15:37
  the
• 15:37
  clear
• 15:37
  favorite:
• 15:38
  a
• 15:38
  full
• 15:38
  transport-agnostic
• 15:39
  networking
• 15:39
  stack
• 15:40
  that
• 15:40
  routes
• 15:40
  packets
• 15:40
  seamlessly
• 15:41
  across
• 15:41
  LoRa,
• 15:42
  Wi-Fi,
• 15:43
  Ethernet,
• 15:44
  the
• 15:44
  internet,
• 15:44
  Tor,
• 15:45
  I2P,
• 15:45
  or
• 15:45
  any
• 15:46
  medium
• 15:46
  accessible
• 15:46
  via
• 15:46
  TCP,
• 15:47
  UDP,
• 15:48
  or
• 15:48
  serial
• 15:48
  interface.
• 15:49
  Its
• 15:49
  address
• 15:50
  space
• 15:50
  is
• 15:50
  globally
• 15:50
  unique
• 15:51
  through
• 15:51
  cryptographic
• 15:51
  identity,
• 15:52
  requiring
• 15:52
  no
• 15:53
  central
• 15:53
  authority.
• 15:54
  Distinct
• 15:55
  local
• 15:55
  meshes
• 15:55
  merge
• 15:55
  automatically
• 15:56
  when
• 15:56
  any
• 15:57
  connection
• 15:57
  point
• 15:57
  is
• 15:57
  established.
• 15:58
  The
• 15:58
  drawback:
• 15:59
  no
• 16:00
  standalone
• 16:00
  LoRa
• 16:00
  firmware,
• 16:01
  meaning
• 16:01
  remote
• 16:02
  solar-powered
• 16:02
  repeater
• 16:03
  nodes
• 16:03
  need
• 16:03
  a
• 16:03
  Raspberry
• 16:03
  Pi
• 16:04
  in
• 16:04
  addition
• 16:04
  to
• 16:04
  the
• 16:04
  radio.
• 16:05
  A
• 16:05
  MicroReticulum
• 16:06
  port
• 16:06
  for
• 16:06
  ESP32
• 16:08
  is
• 16:08
  in
• 16:08
  development
• 16:09
  and
• 16:09
  could
• 16:09
  be
• 16:09
  transformative
• 16:10
  for
• 16:10
  adoption.
• 16:11
  In
• 16:11
  alternative
• 16:11
  computing,
• 16:12
  a
• 16:12
  multi-institution
• 16:13
  team
• 16:13
  published
• 16:14
  in
• 16:14
  Nature
• 16:14
  Communications
• 16:15
  a
• 16:15
  neuromorphic
• 16:16
  computer
• 16:16
  that
• 16:16
  combines
• 16:17
  quantum-tunnelling
• 16:17
  physics
• 16:18
  with
• 16:18
  brain-inspired
• 16:19
  architecture
• 16:19
  to
• 16:19
  tackle
• 16:20
  combinatorial
• 16:20
  optimization
• 16:21
  problems
• 16:21
  —
• 16:22
  protein
• 16:22
  folding,
• 16:23
  logistics
• 16:23
  networks,
• 16:24
  chip
• 16:24
  routing
• 16:24
  —
• 16:25
  where
• 16:25
  conventional
• 16:26
  AI
• 16:26
  stalls.
• 16:27
  The
• 16:27
  neuromorphic
• 16:28
  autoencoder
• 16:28
  with
• 16:29
  a
• 16:29
  Fowler-Nordheim
• 16:31
  annealer
• 16:32
  searches
• 16:32
  for
• 16:32
  solutions
• 16:33
  "the
• 16:33
  way
• 16:33
  natural
• 16:34
  processes
• 16:34
  navigate
• 16:34
  a
• 16:35
  complex
• 16:35
  energy
• 16:35
  landscape
• 16:36
  to
• 16:36
  settle
• 16:36
  into
• 16:37
  stability,"
• 16:38
  with
• 16:38
  a
• 16:38
  guarantee
• 16:38
  of
• 16:38
  asymptotic
• 16:39
  convergence
• 16:39
  to
• 16:40
  the
• 16:40
  optimal
• 16:40
  solution.
• 16:41
  The
• 16:41
  team
• 16:41
  spans
• 16:42
  Washington
• 16:42
  University,
• 16:43
  IISc
• 16:44
  Bangalore,
• 16:44
  Heidelberg,
• 16:45
  Johns
• 16:45
  Hopkins,
• 16:46
  and
• 16:46
  UC
• 16:46
  Santa
• 16:46
  Cruz.
• 16:48
  And
• 16:48
  in
• 16:48
  a
• 16:48
  reminder
• 16:48
  that
• 16:48
  production
• 16:49
  AI
• 16:49
  still
• 16:49
  fails
• 16:50
  in
• 16:50
  mundane
• 16:50
  ways,
• 16:51
  Starbucks
• 16:51
  is
• 16:52
  scrapping
• 16:52
  its
• 16:52
  AI-powered
• 16:53
  inventory
• 16:53
  management
• 16:54
  tool
• 16:54
  after
• 16:54
  it
• 16:54
  reportedly
• 16:55
  miscounted
• 16:56
  and
• 16:56
  mislabeled
• 16:56
  items
• 16:57
  —
• 16:58
  a
• 16:58
  high-profile
• 16:58
  case
• 16:59
  of
• 16:59
  the
• 16:59
  pattern
• 16:59
  where
• 16:59
  AI
• 17:00
  deployment
• 17:00
  breaks
• 17:00
  not
• 17:01
  on
• 17:01
  intelligence
• 17:02
  but
• 17:02
  on
• 17:02
  coordination
• 17:03
  with
• 17:03
  messy
• 17:03
  real-world
• 17:04
  data.