Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – July 08, 2026

Transcript

AGI Dreams Podcast – July 08, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  🏎️
• 0:00
  Local
• 0:00
  AI
• 0:00
  Speed
• 0:01
  Wars
• 0:01
  DFlash
• 0:01
  just
• 0:02
  landed
• 0:02
  in
• 0:02
  llama.cpp,
• 0:03
  and
• 0:04
  the
• 0:04
  benchmarks
• 0:04
  deserve
• 0:04
  attention
• 0:05
  rather
• 0:05
  than
• 0:05
  hype.
• 0:06
  The
• 0:06
  technique
• 0:06
  pairs
• 0:07
  a
• 0:07
  block
• 0:07
  diffusion
• 0:07
  drafter
• 0:08
  with
• 0:08
  the
• 0:08
  main
• 0:08
  autoregressive
• 0:09
  model,
• 0:09
  allowing
• 0:10
  roughly
• 0:10
  eight
• 0:10
  tokens
• 0:10
  to
• 0:11
  be
• 0:11
  speculatively
• 0:11
  decoded
• 0:12
  in
• 0:12
  parallel.
• 0:12
  On
• 0:13
  Qwen
• 0:13
  3.5
• 0:13
  27B
• 0:14
  running
• 0:14
  on
• 0:14
  an
• 0:15
  RTX
• 0:15
  3090
• 0:16
  at
• 0:16
  36K
• 0:16
  context,
• 0:17
  the
• 0:17
  tester
• 0:18
  measured
• 0:18
  a
• 0:18
  4.44x
• 0:19
  speedup
• 0:19
  —
• 0:20
  from
• 0:20
  around
• 0:20
  440
• 0:21
  tokens
• 0:21
  per
• 0:22
  second
• 0:22
  stock
• 0:22
  to
• 0:22
  1,950
• 0:24
  with
• 0:24
  DFlash
• 0:24
  engaged.
• 0:25
  The
• 0:25
  gains
• 0:25
  scale
• 0:25
  with
• 0:26
  context
• 0:26
  length:
• 0:26
  at
• 0:27
  short
• 0:27
  prompts
• 0:27
  the
• 0:27
  overhead
• 0:28
  barely
• 0:28
  pays
• 0:28
  for
• 0:28
  itself,
• 0:29
  but
• 0:29
  at
• 0:29
  36K
• 0:30
  tokens
• 0:30
  the
• 0:30
  KV
• 0:30
  cache
• 0:31
  penalty
• 0:31
  on
• 0:31
  standard
• 0:31
  generation
• 0:32
  is
• 0:32
  severe
• 0:32
  enough
• 0:33
  that
• 0:33
  speculative
• 0:33
  drafting
• 0:34
  dominates.
• 0:35
  Prior
• 0:35
  coverage
• 0:35
  of
• 0:35
  DFlash
• 0:36
  on
• 0:36
  Apple
• 0:36
  Silicon
• 0:37
  via
• 0:37
  MLX
• 0:37
  reported
• 0:38
  2x
• 0:38
  gains
• 0:39
  on
• 0:39
  similar
• 0:39
  models,
• 0:40
  so
• 0:40
  the
• 0:40
  llama.cpp
• 0:41
  implementation
• 0:42
  is
• 0:42
  delivering
• 0:42
  on
• 0:42
  the
• 0:42
  architecture's
• 0:43
  promise
• 0:43
  rather
• 0:43
  than
• 0:44
  inventing
• 0:44
  new
• 0:44
  magic.
• 0:45
  Ollama
• 0:45
  0.31
• 0:46
  brings
• 0:46
  multi-token
• 0:47
  prediction
• 0:47
  to
• 0:47
  Gemma
• 0:48
  4
• 0:48
  on
• 0:48
  Apple
• 0:48
  Silicon,
• 0:49
  a
• 0:49
  quieter
• 0:50
  but
• 0:50
  broadly
• 0:51
  useful
• 0:51
  speedup.
• 0:52
  MTP
• 0:52
  works
• 0:53
  because
• 0:53
  the
• 0:53
  prediction
• 0:53
  heads
• 0:54
  are
• 0:54
  already
• 0:54
  baked
• 0:54
  into
• 0:55
  the
• 0:55
  model
• 0:55
  weights
• 0:55
  during
• 0:55
  training
• 0:56
  —
• 0:56
  the
• 0:56
  runtime
• 0:57
  just
• 0:57
  needs
• 0:57
  to
• 0:57
  exploit
• 0:58
  them.
• 0:58
  For
• 0:59
  M4
• 0:59
  Pro
• 0:59
  users
• 0:59
  this
• 1:00
  means
• 1:00
  measurable
• 1:00
  throughput
• 1:01
  gains
• 1:01
  without
• 1:01
  touching
• 1:01
  quantization
• 1:02
  or
• 1:02
  swapping
• 1:02
  models,
• 1:03
  the
• 1:03
  kind
• 1:04
  of
• 1:04
  improvement
• 1:04
  that
• 1:04
  compounds
• 1:05
  across
• 1:05
  every
• 1:05
  session
• 1:06
  rather
• 1:06
  than
• 1:06
  headline-grabbing
• 1:07
  on
• 1:07
  a
• 1:07
  benchmark
• 1:07
  chart.
• 1:08
  It
• 1:08
  also
• 1:09
  signals
• 1:09
  a
• 1:09
  broader
• 1:10
  trend:
• 1:10
  model
• 1:11
  authors
• 1:11
  are
• 1:11
  shipping
• 1:11
  inference
• 1:12
  optimizations
• 1:13
  as
• 1:13
  trained-in
• 1:13
  capabilities
• 1:14
  rather
• 1:14
  than
• 1:14
  leaving
• 1:14
  them
• 1:15
  to
• 1:15
  runtime
• 1:15
  hacks.
• 1:16
  The
• 1:16
  hardware
• 1:16
  side
• 1:16
  of
• 1:17
  the
• 1:17
  speed
• 1:17
  war
• 1:17
  got
• 1:17
  more
• 1:17
  interesting.
• 1:18
  AMD's
• 1:19
  Ryzen
• 1:19
  AI
• 1:19
  Halo
• 1:20
  dev
• 1:20
  kit
• 1:20
  arrived
• 1:21
  at
• 1:21
  LTT
• 1:21
  Labs
• 1:22
  for
• 1:22
  a
• 1:22
  full
• 1:22
  review.
• 1:23
  At
• 1:23
  $4,099
• 1:25
  it
• 1:25
  sits
• 1:25
  in
• 1:25
  an
• 1:25
  awkward
• 1:26
  price
• 1:26
  band
• 1:26
  —
• 1:27
  more
• 1:27
  than
• 1:27
  a
• 1:27
  Mac
• 1:27
  Studio,
• 1:28
  less
• 1:28
  than
• 1:28
  a
• 1:28
  dual-GPU
• 1:29
  workstation
• 1:29
  —
• 1:30
  but
• 1:30
  its
• 1:30
  128GB
• 1:32
  of
• 1:33
  unified
• 1:33
  memory
• 1:33
  is
• 1:33
  the
• 1:33
  real
• 1:34
  pitch.
• 1:35
  That
• 1:35
  unified
• 1:35
  pool
• 1:36
  means
• 1:36
  you
• 1:36
  can
• 1:36
  load
• 1:36
  70B-class
• 1:37
  models
• 1:38
  that
• 1:38
  simply
• 1:39
  do
• 1:39
  not
• 1:39
  fit
• 1:39
  on
• 1:40
  consumer
• 1:40
  discrete
• 1:41
  GPUs.
• 1:42
  The
• 1:42
  Max+
• 1:43
  395
• 1:44
  delivers
• 1:44
  256
• 1:45
  GB/s
• 1:46
  memory
• 1:46
  bandwidth,
• 1:47
  enough
• 1:48
  to
• 1:48
  keep
• 1:48
  a
• 1:48
  70B
• 1:48
  quantized
• 1:49
  model
• 1:49
  generating
• 1:50
  at
• 1:50
  practical
• 1:50
  speeds.
• 1:51
  BKC
• 1:52
  playbooks
• 1:52
  ship
• 1:53
  with
• 1:53
  the
• 1:53
  kit
• 1:53
  for
• 1:53
  ROCm
• 1:54
  setup,
• 1:54
  but
• 1:54
  the
• 1:54
  software
• 1:55
  story
• 1:55
  remains
• 1:55
  AMD's
• 1:56
  persistent
• 1:56
  weakness:
• 1:58
  reviewers
• 1:58
  noted
• 1:58
  rough
• 1:59
  edges
• 1:59
  in
• 1:59
  driver
• 2:00
  stability
• 2:00
  and
• 2:00
  vLLM
• 2:01
  compatibility
• 2:02
  that
• 2:02
  NVIDIA
• 2:02
  users
• 2:03
  never
• 2:03
  encounter.
• 2:04
  Community
• 2:05
  reception
• 2:05
  was
• 2:05
  split
• 2:05
  —
• 2:07
  some
• 2:07
  compared
• 2:07
  it
• 2:07
  favorably
• 2:08
  to
• 2:08
  NVIDIA's
• 2:09
  pricing
• 2:09
  on
• 2:09
  equivalent
• 2:09
  VRAM,
• 2:10
  others
• 2:11
  questioned
• 2:11
  whether
• 2:12
  ROCm
• 2:12
  is
• 2:12
  mature
• 2:13
  enough
• 2:13
  to
• 2:13
  justify
• 2:13
  the
• 2:13
  investment
• 2:14
  when
• 2:14
  a
• 2:14
  Mac
• 2:14
  Studio
• 2:15
  with
• 2:15
  192GB
• 2:17
  "just
• 2:17
  works."
• 2:18
  Then
• 2:19
  there
• 2:19
  is
• 2:19
  Huawei.
• 2:20
  A
• 2:20
  benchmark
• 2:21
  showing
• 2:21
  an
• 2:21
  Ascend
• 2:21
  910C
• 2:23
  running
• 2:23
  DeepSeek-V3
• 2:24
  at
• 2:24
  speeds
• 2:25
  competitive
• 2:25
  with
• 2:26
  NVIDIA
• 2:26
  datacenter
• 2:27
  parts
• 2:27
  drew
• 2:27
  genuine
• 2:28
  surprise
• 2:28
  from
• 2:29
  a
• 2:29
  community
• 2:29
  that
• 2:29
  had
• 2:29
  largely
• 2:30
  written
• 2:30
  off
• 2:30
  Chinese
• 2:31
  AI
• 2:31
  silicon
• 2:31
  as
• 2:32
  export-controlled
• 2:32
  and
• 2:33
  software-starved.
• 2:34
  The
• 2:34
  numbers
• 2:35
  are
• 2:35
  real
• 2:35
  enough
• 2:35
  to
• 2:35
  take
• 2:36
  seriously,
• 2:36
  but
• 2:37
  the
• 2:37
  questions
• 2:37
  that
• 2:37
  matter
• 2:38
  —
• 2:38
  developer
• 2:39
  tooling,
• 2:39
  availability
• 2:40
  outside
• 2:41
  PRC,
• 2:41
  inference
• 2:42
  framework
• 2:42
  support
• 2:43
  —
• 2:44
  remain
• 2:44
  unanswered.
• 2:45
  This
• 2:46
  is
• 2:46
  genuinely
• 2:46
  new
• 2:46
  territory
• 2:47
  for
• 2:47
  the
• 2:47
  local
• 2:47
  AI
• 2:48
  community,
• 2:49
  and
• 2:49
  the
• 2:49
  discomfort
• 2:49
  in
• 2:49
  the
• 2:49
  thread
• 2:50
  is
• 2:50
  palpable.
• 2:53
  🤖
• 2:53
  Models
• 2:53
  Go
• 2:53
  Multimodal
• 2:54
  NVIDIA's
• 2:55
  Nemotron
• 2:55
  Audex
• 2:55
  30B-A3B
• 2:58
  marks
• 2:58
  a
• 2:58
  deliberate
• 2:59
  architectural
• 2:59
  choice:
• 3:00
  audio
• 3:00
  is
• 3:00
  not
• 3:01
  a
• 3:01
  preprocessing
• 3:01
  step,
• 3:02
  it
• 3:02
  is
• 3:02
  a
• 3:02
  first-class
• 3:03
  token
• 3:03
  type
• 3:03
  inside
• 3:04
  the
• 3:04
  transformer.
• 3:05
  The
• 3:05
  model
• 3:05
  uses
• 3:06
  a
• 3:06
  mixture-of-experts
• 3:06
  architecture
• 3:07
  with
• 3:07
  30B
• 3:08
  total
• 3:08
  parameters
• 3:09
  but
• 3:09
  only
• 3:09
  3B
• 3:10
  active
• 3:10
  per
• 3:10
  forward
• 3:11
  pass,
• 3:11
  keeping
• 3:12
  inference
• 3:12
  costs
• 3:12
  manageable
• 3:13
  while
• 3:13
  processing
• 3:14
  raw
• 3:14
  audio
• 3:14
  waveforms
• 3:15
  alongside
• 3:15
  text
• 3:16
  tokens
• 3:16
  in
• 3:16
  a
• 3:16
  shared
• 3:17
  sequence.
• 3:18
  For
• 3:18
  tasks
• 3:18
  like
• 3:18
  meeting
• 3:18
  summarization
• 3:19
  or
• 3:19
  spoken
• 3:20
  Q&A,
• 3:20
  this
• 3:20
  is
• 3:21
  the
• 3:21
  difference
• 3:21
  between
• 3:21
  analyzing
• 3:22
  what
• 3:22
  was
• 3:22
  said
• 3:22
  and
• 3:22
  understanding
• 3:23
  how
• 3:23
  it
• 3:23
  was
• 3:23
  said
• 3:23
  —
• 3:24
  temporal
• 3:24
  grounding,
• 3:25
  tone,
• 3:25
  hesitation
• 3:26
  patterns,
• 3:27
  all
• 3:27
  available
• 3:27
  to
• 3:27
  the
• 3:27
  model
• 3:28
  without
• 3:28
  lossy
• 3:28
  transcription
• 3:29
  in
• 3:29
  the
• 3:29
  middle.
• 3:30
  Google's
• 3:31
  Gemma
• 3:31
  4
• 3:31
  technical
• 3:32
  report
• 3:32
  fills
• 3:32
  in
• 3:32
  the
• 3:32
  details
• 3:33
  behind
• 3:33
  the
• 3:33
  family's
• 3:33
  strongest
• 3:34
  release.
• 3:35
  The
• 3:35
  27B-A4B
• 3:37
  MoE
• 3:37
  variant
• 3:38
  hits
• 3:38
  a
• 3:38
  sweet
• 3:38
  spot
• 3:39
  for
• 3:39
  single-GPU
• 3:39
  deployment
• 3:40
  while
• 3:40
  delivering
• 3:40
  multimodal
• 3:41
  capabilities
• 3:41
  —
• 3:43
  vision,
• 3:43
  audio
• 3:44
  understanding,
• 3:44
  and
• 3:44
  strong
• 3:45
  coding
• 3:45
  performance
• 3:45
  —
• 3:46
  under
• 3:46
  an
• 3:47
  Apache
• 3:47
  2.0
• 3:48
  license.
• 3:49
  Community
• 3:49
  response
• 3:50
  focused
• 3:50
  on
• 3:50
  two
• 3:50
  things:
• 3:51
  benchmark
• 3:51
  performance
• 3:52
  that
• 3:52
  competes
• 3:52
  with
• 3:52
  models
• 3:53
  twice
• 3:53
  its
• 3:53
  active
• 3:54
  parameter
• 3:54
  count,
• 3:55
  and
• 3:55
  the
• 3:55
  practical
• 3:56
  discovery
• 3:56
  that
• 3:56
  quantized
• 3:57
  variants
• 3:57
  (NVFP4,
• 3:59
  W4A4)
• 4:00
  run
• 4:00
  comfortably
• 4:01
  on
• 4:01
  consumer
• 4:01
  hardware
• 4:02
  without
• 4:02
  the
• 4:02
  quality
• 4:03
  cliff
• 4:03
  that
• 4:04
  plagued
• 4:04
  earlier
• 4:04
  generations.
• 4:06
  The
• 4:06
  ecosystem
• 4:06
  matured
• 4:07
  fast
• 4:07
  —
• 4:08
  Parlor
• 4:08
  demonstrated
• 4:09
  a
• 4:09
  fully
• 4:09
  on-device
• 4:10
  multimodal
• 4:10
  app
• 4:11
  pairing
• 4:11
  Gemma
• 4:11
  4
• 4:11
  with
• 4:12
  Kokoro
• 4:12
  TTS
• 4:13
  at
• 4:13
  1.5-3
• 4:13
  second
• 4:14
  end-to-end
• 4:14
  latency
• 4:15
  on
• 4:15
  an
• 4:15
  M3
• 4:15
  Pro
• 4:15
  laptop.
• 4:17
  Hugging
• 4:17
  Face
• 4:17
  and
• 4:17
  Cerebras
• 4:18
  turned
• 4:18
  Gemma
• 4:18
  4
• 4:19
  into
• 4:19
  a
• 4:19
  real-time
• 4:19
  voice
• 4:20
  pipeline,
• 4:20
  and
• 4:20
  the
• 4:20
  latency
• 4:21
  numbers
• 4:21
  explain
• 4:21
  why
• 4:22
  it
• 4:22
  matters.
• 4:23
  Cerebras
• 4:23
  inference
• 4:23
  delivers
• 4:24
  hundreds
• 4:24
  of
• 4:24
  tokens
• 4:25
  per
• 4:25
  second,
• 4:26
  which
• 4:26
  means
• 4:26
  the
• 4:26
  full
• 4:26
  cascade
• 4:27
  —
• 4:28
  Whisper
• 4:28
  transcription,
• 4:29
  Gemma
• 4:29
  4
• 4:30
  reasoning,
• 4:30
  TTS
• 4:31
  synthesis
• 4:31
  —
• 4:32
  completes
• 4:32
  fast
• 4:33
  enough
• 4:33
  for
• 4:33
  natural
• 4:33
  conversation.
• 4:35
  Sub-200ms
• 4:36
  first-token
• 4:37
  latency
• 4:37
  makes
• 4:37
  this
• 4:37
  feel
• 4:38
  like
• 4:38
  talking
• 4:38
  to
• 4:38
  a
• 4:38
  person
• 4:39
  rather
• 4:39
  than
• 4:39
  waiting
• 4:40
  for
• 4:40
  a
• 4:40
  computer.
• 4:41
  The
• 4:41
  blog
• 4:42
  provides
• 4:42
  reproducible
• 4:43
  code
• 4:43
  for
• 4:43
  the
• 4:43
  full
• 4:44
  cascade,
• 4:44
  which
• 4:45
  matters
• 4:45
  more
• 4:45
  than
• 4:45
  the
• 4:45
  demo:
• 4:46
  anyone
• 4:47
  with
• 4:47
  API
• 4:47
  access
• 4:48
  can
• 4:48
  stand
• 4:48
  up
• 4:48
  a
• 4:49
  conversational
• 4:49
  voice
• 4:49
  agent
• 4:50
  without
• 4:50
  specialized
• 4:51
  hardware
• 4:51
  or
• 4:51
  vendor
• 4:52
  lock-in.
• 4:54
  🌍
• 4:54
  World
• 4:55
  Models
• 4:55
  and
• 4:55
  Embodied
• 4:56
  AI
• 4:56
  COMAP
• 4:56
  introduces
• 4:57
  a
• 4:57
  training
• 4:57
  loop
• 4:57
  where
• 4:57
  the
• 4:57
  world
• 4:58
  model
• 4:58
  and
• 4:58
  the
• 4:58
  agent
• 4:58
  policy
• 4:59
  evolve
• 4:59
  together
• 5:00
  through
• 5:00
  on-policy
• 5:00
  self-distillation.
• 5:02
  The
• 5:02
  standard
• 5:02
  approach
• 5:02
  —
• 5:03
  train
• 5:03
  a
• 5:03
  world
• 5:03
  model
• 5:04
  on
• 5:04
  demonstrations,
• 5:05
  then
• 5:05
  train
• 5:05
  a
• 5:05
  policy
• 5:06
  against
• 5:06
  it
• 5:06
  —
• 5:07
  suffers
• 5:07
  from
• 5:07
  distribution
• 5:08
  shift:
• 5:08
  the
• 5:08
  world
• 5:09
  model
• 5:09
  learned
• 5:09
  expert
• 5:10
  behavior
• 5:10
  it
• 5:10
  never
• 5:10
  sees
• 5:11
  during
• 5:11
  policy
• 5:11
  execution,
• 5:12
  so
• 5:12
  it
• 5:13
  provides
• 5:13
  unreliable
• 5:13
  predictions
• 5:14
  exactly
• 5:14
  when
• 5:14
  the
• 5:15
  agent
• 5:15
  needs
• 5:15
  them
• 5:15
  most.
• 5:16
  COMAP's
• 5:17
  fix
• 5:17
  is
• 5:17
  future-aware
• 5:18
  reflection:
• 5:19
  the
• 5:19
  world
• 5:19
  model
• 5:20
  predicts
• 5:20
  consequences
• 5:21
  of
• 5:21
  the
• 5:21
  agent's
• 5:21
  actual
• 5:21
  policy,
• 5:22
  the
• 5:23
  policy
• 5:23
  adapts
• 5:23
  to
• 5:24
  the
• 5:24
  world
• 5:24
  model's
• 5:24
  actual
• 5:25
  predictions,
• 5:25
  and
• 5:25
  both
• 5:26
  improve
• 5:26
  in
• 5:26
  tandem.
• 5:27
  Results
• 5:28
  on
• 5:28
  ALFWorld,
• 5:28
  WebShop,
• 5:29
  and
• 5:29
  ScienceWorld
• 5:29
  show
• 5:30
  consistent
• 5:30
  gains
• 5:31
  over
• 5:31
  fixed-world-model
• 5:32
  baselines.
• 5:33
  The
• 5:33
  paper
• 5:33
  is
• 5:33
  dense
• 5:33
  but
• 5:34
  the
• 5:34
  core
• 5:34
  insight
• 5:34
  is
• 5:34
  clean:
• 5:35
  world
• 5:35
  models
• 5:36
  trained
• 5:36
  off-policy
• 5:37
  are
• 5:37
  a
• 5:37
  leaky
• 5:37
  abstraction,
• 5:38
  and
• 5:38
  co-evolution
• 5:39
  is
• 5:39
  how
• 5:39
  you
• 5:39
  seal
• 5:39
  the
• 5:39
  leak.
• 5:40
  NVIDIA's
• 5:41
  SpatialClaw
• 5:41
  reframes
• 5:42
  spatial
• 5:42
  reasoning
• 5:42
  as
• 5:43
  code
• 5:43
  generation
• 5:43
  rather
• 5:44
  than
• 5:44
  coordinate
• 5:44
  prediction.
• 5:45
  Instead
• 5:45
  of
• 5:45
  asking
• 5:46
  a
• 5:46
  vision-language
• 5:46
  model
• 5:47
  to
• 5:47
  output
• 5:47
  raw
• 5:47
  positions,
• 5:48
  SpatialClaw
• 5:49
  has
• 5:49
  the
• 5:49
  model
• 5:49
  write
• 5:49
  short
• 5:50
  programs
• 5:50
  that
• 5:50
  express
• 5:51
  spatial
• 5:51
  relationships
• 5:51
  —
• 5:52
  "place
• 5:52
  the
• 5:52
  cup
• 5:52
  to
• 5:52
  the
• 5:53
  right
• 5:53
  of
• 5:53
  the
• 5:53
  plate,
• 5:53
  10cm
• 5:54
  from
• 5:54
  the
• 5:54
  edge."
• 5:55
  Across
• 5:55
  20
• 5:56
  benchmarks,
• 5:56
  this
• 5:56
  code-as-action
• 5:57
  interface
• 5:57
  improves
• 5:58
  accuracy
• 5:58
  by
• 5:58
  an
• 5:59
  average
• 5:59
  of
• 5:59
  11.2
• 6:00
  points.
• 6:00
  The
• 6:00
  result
• 6:01
  makes
• 6:01
  sense
• 6:01
  if
• 6:01
  you
• 6:01
  think
• 6:01
  about
• 6:02
  it:
• 6:03
  coordinate
• 6:03
  systems
• 6:03
  are
• 6:03
  arbitrary
• 6:04
  and
• 6:04
  models
• 6:04
  struggle
• 6:05
  with
• 6:05
  them,
• 6:05
  but
• 6:06
  spatial
• 6:06
  language
• 6:06
  is
• 6:06
  exactly
• 6:07
  the
• 6:07
  kind
• 6:07
  of
• 6:07
  compositional
• 6:08
  reasoning
• 6:08
  transformers
• 6:09
  handle
• 6:09
  well.
• 6:10
  The
• 6:10
  VLM
• 6:10
  never
• 6:11
  needs
• 6:11
  to
• 6:11
  learn
• 6:11
  coordinate
• 6:12
  systems
• 6:12
  —
• 6:12
  it
• 6:12
  writes
• 6:13
  programs
• 6:13
  that
• 6:13
  already
• 6:14
  encode
• 6:14
  them.
• 6:14
  LeRobot
• 6:15
  v0.6.0
• 6:17
  is
• 6:17
  the
• 6:17
  most
• 6:17
  ambitious
• 6:17
  release
• 6:18
  yet
• 6:18
  from
• 6:18
  Hugging
• 6:18
  Face's
• 6:19
  robot
• 6:19
  learning
• 6:19
  framework.
• 6:20
  World
• 6:21
  model
• 6:21
  policies
• 6:21
  let
• 6:22
  robots
• 6:22
  plan
• 6:22
  by
• 6:22
  imagining
• 6:23
  future
• 6:23
  states
• 6:23
  before
• 6:24
  acting.
• 6:25
  A
• 6:25
  reward
• 6:25
  models
• 6:26
  API
• 6:26
  replaces
• 6:27
  hand-crafted
• 6:27
  reward
• 6:28
  functions
• 6:28
  with
• 6:28
  learned
• 6:28
  preferences.
• 6:30
  DAgger
• 6:30
  interactive
• 6:31
  deployment
• 6:31
  closes
• 6:31
  the
• 6:31
  gap
• 6:32
  between
• 6:32
  demonstration
• 6:32
  and
• 6:33
  deployment
• 6:33
  by
• 6:33
  feeding
• 6:33
  human
• 6:34
  corrections
• 6:34
  back
• 6:35
  into
• 6:35
  the
• 6:35
  policy
• 6:35
  in
• 6:35
  real-time.
• 6:37
  FSDP
• 6:37
  makes
• 6:38
  training
• 6:38
  these
• 6:38
  models
• 6:39
  feasible
• 6:39
  on
• 6:39
  multi-GPU
• 6:40
  setups
• 6:40
  without
• 6:41
  custom
• 6:41
  distributed
• 6:41
  training
• 6:42
  code.
• 6:43
  Each
• 6:43
  feature
• 6:43
  solves
• 6:44
  a
• 6:44
  different
• 6:44
  piece
• 6:44
  of
• 6:44
  the
• 6:44
  sim-to-real
• 6:45
  transfer
• 6:45
  problem,
• 6:46
  and
• 6:46
  the
• 6:46
  framework
• 6:47
  is
• 6:47
  open-source
• 6:47
  with
• 6:48
  real
• 6:48
  documentation
• 6:48
  —
• 6:49
  which
• 6:49
  in
• 6:49
  robotics
• 6:50
  remains
• 6:50
  the
• 6:50
  exception
• 6:51
  rather
• 6:51
  than
• 6:51
  the
• 6:51
  rule.
• 6:55
  🔒
• 6:55
  AI
• 6:55
  Security
• 6:56
  and
• 6:56
  Interpretability
• 6:56
  Anthropic's
• 6:57
  Jacobian
• 6:57
  Lens
• 6:58
  may
• 6:58
  be
• 6:59
  the
• 6:59
  most
• 6:59
  significant
• 6:59
  interpretability
• 7:00
  advance
• 7:01
  since
• 7:01
  sparse
• 7:01
  autoencoder
• 7:02
  features.
• 7:03
  The
• 7:03
  Jacobian
• 7:04
  matrix
• 7:04
  of
• 7:04
  a
• 7:04
  neural
• 7:05
  network
• 7:05
  layer
• 7:05
  captures
• 7:06
  how
• 7:06
  small
• 7:06
  perturbations
• 7:07
  in
• 7:07
  the
• 7:07
  input
• 7:07
  propagate
• 7:07
  to
• 7:08
  the
• 7:08
  output
• 7:08
  —
• 7:09
  a
• 7:09
  linear
• 7:09
  map
• 7:10
  of
• 7:10
  what
• 7:10
  the
• 7:10
  layer
• 7:10
  computes
• 7:10
  locally.
• 7:11
  Analyzing
• 7:12
  these
• 7:12
  matrices
• 7:13
  across
• 7:13
  layers
• 7:13
  lets
• 7:14
  researchers
• 7:14
  trace
• 7:14
  how
• 7:15
  information
• 7:15
  transforms
• 7:16
  as
• 7:16
  it
• 7:16
  flows
• 7:16
  through
• 7:16
  the
• 7:16
  network:
• 7:17
  where
• 7:17
  features
• 7:18
  are
• 7:18
  created,
• 7:19
  combined,
• 7:19
  and
• 7:20
  destroyed.
• 7:21
  The
• 7:21
  practical
• 7:22
  payoff
• 7:22
  is
• 7:22
  a
• 7:22
  principled
• 7:23
  mathematical
• 7:23
  framework
• 7:24
  for
• 7:24
  understanding
• 7:24
  model
• 7:25
  internals,
• 7:26
  replacing
• 7:26
  the
• 7:26
  post-hoc
• 7:27
  feature
• 7:27
  visualization
• 7:28
  that
• 7:28
  has
• 7:28
  dominated
• 7:28
  interpretability
• 7:29
  work.
• 7:30
  For
• 7:30
  safety
• 7:31
  researchers,
• 7:32
  this
• 7:32
  means
• 7:32
  identifying
• 7:33
  where
• 7:33
  a
• 7:33
  model
• 7:33
  encodes
• 7:34
  dangerous
• 7:34
  capabilities
• 7:35
  not
• 7:35
  by
• 7:35
  searching
• 7:36
  after
• 7:36
  the
• 7:36
  fact,
• 7:37
  but
• 7:37
  by
• 7:37
  watching
• 7:37
  them
• 7:37
  form
• 7:38
  during
• 7:38
  forward
• 7:38
  passes.
• 7:39
  Reuven
• 7:40
  Cohen's
• 7:40
  editorial
• 7:41
  frames
• 7:41
  it
• 7:41
  as
• 7:41
  infrastructure,
• 7:42
  not
• 7:42
  research
• 7:42
  —
• 7:43
  the
• 7:43
  kind
• 7:43
  of
• 7:43
  tool
• 7:43
  defenders
• 7:44
  actually
• 7:44
  need.
• 7:45
  On
• 7:45
  the
• 7:45
  adversarial
• 7:46
  side,
• 7:46
  a
• 7:46
  paper
• 7:47
  on
• 7:47
  molecular
• 7:47
  graph
• 7:48
  neural
• 7:48
  network
• 7:48
  backdoors
• 7:49
  reveals
• 7:49
  how
• 7:49
  domain-specific
• 7:50
  constraints
• 7:51
  break
• 7:51
  generic
• 7:51
  defenses.
• 7:53
  ChemBack
• 7:53
  defines
• 7:53
  backdoor
• 7:54
  attacks
• 7:54
  that
• 7:54
  respect
• 7:55
  chemical
• 7:55
  validity
• 7:55
  —
• 7:56
  the
• 7:56
  poisoned
• 7:56
  molecular
• 7:57
  graphs
• 7:57
  are
• 7:57
  still
• 7:57
  valid
• 7:58
  molecules,
• 7:59
  so
• 7:59
  defenses
• 7:59
  based
• 8:00
  on
• 8:00
  detecting
• 8:00
  "unusual"
• 8:01
  graph
• 8:01
  structures
• 8:01
  fail
• 8:02
  completely.
• 8:03
  ChemGuard
• 8:03
  responds
• 8:04
  with
• 8:04
  an
• 8:04
  admission-based
• 8:05
  defense
• 8:05
  that
• 8:05
  filters
• 8:05
  inputs
• 8:06
  through
• 8:06
  a
• 8:06
  chemistry-aware
• 8:07
  plausibility
• 8:07
  check
• 8:08
  before
• 8:08
  they
• 8:08
  reach
• 8:08
  the
• 8:08
  model.
• 8:09
  This
• 8:09
  matters
• 8:10
  because
• 8:10
  molecular
• 8:11
  GNNs
• 8:11
  are
• 8:11
  deployed
• 8:12
  in
• 8:12
  real
• 8:12
  drug
• 8:12
  discovery
• 8:13
  pipelines,
• 8:14
  and
• 8:14
  the
• 8:14
  standard
• 8:15
  ML
• 8:15
  security
• 8:15
  playbook
• 8:16
  assumes
• 8:16
  data
• 8:16
  points
• 8:16
  live
• 8:17
  in
• 8:17
  a
• 8:17
  continuous
• 8:17
  vector
• 8:18
  space
• 8:18
  where
• 8:18
  outlier
• 8:18
  detection
• 8:19
  works.
• 8:20
  Chemistry
• 8:20
  has
• 8:21
  hard
• 8:21
  structural
• 8:21
  constraints
• 8:22
  —
• 8:23
  valence
• 8:23
  rules,
• 8:24
  ring
• 8:24
  strain,
• 8:24
  bond
• 8:25
  angles
• 8:25
  —
• 8:26
  that
• 8:26
  an
• 8:26
  attacker
• 8:26
  can
• 8:27
  exploit
• 8:27
  precisely
• 8:27
  because
• 8:28
  generic
• 8:28
  defenses
• 8:29
  do
• 8:29
  not
• 8:29
  encode
• 8:29
  them.
• 8:32
  🛠️
• 8:33
  Agentic
• 8:33
  Coding
• 8:33
  and
• 8:33
  Developer
• 8:34
  Culture
• 8:34
  The
• 8:34
  AI
• 8:34
  Builder
• 8:35
  Club's
• 8:35
  open-source
• 8:36
  skills
• 8:36
  repository
• 8:37
  and
• 8:37
  a
• 8:37
  detailed
• 8:38
  "agentic
• 8:38
  workbench"
• 8:39
  post
• 8:39
  both
• 8:40
  point
• 8:40
  at
• 8:40
  the
• 8:40
  same
• 8:40
  shift:
• 8:41
  developers
• 8:42
  are
• 8:42
  assembling
• 8:42
  their
• 8:42
  own
• 8:43
  orchestration
• 8:43
  layers
• 8:44
  rather
• 8:44
  than
• 8:44
  waiting
• 8:44
  for
• 8:44
  vendors
• 8:45
  to
• 8:45
  ship
• 8:45
  them.
• 8:46
  The
• 8:46
  skills
• 8:46
  repo
• 8:47
  defines
• 8:47
  modular
• 8:47
  capability
• 8:48
  specifications
• 8:49
  that
• 8:49
  coding
• 8:49
  agents
• 8:50
  can
• 8:50
  invoke
• 8:50
  —
• 8:51
  structured
• 8:51
  as
• 8:51
  composable
• 8:52
  units
• 8:52
  rather
• 8:53
  than
• 8:53
  monolithic
• 8:53
  system
• 8:54
  prompts.
• 8:55
  The
• 8:55
  workbench
• 8:55
  post
• 8:56
  walks
• 8:56
  through
• 8:56
  a
• 8:56
  personal
• 8:56
  setup
• 8:57
  coordinating
• 8:57
  multiple
• 8:58
  AI
• 8:58
  assistants
• 8:59
  through
• 8:59
  custom
• 8:59
  workflows,
• 9:00
  with
• 9:01
  the
• 9:01
  revealing
• 9:01
  observation
• 9:02
  that
• 9:02
  the
• 9:02
  orchestration
• 9:02
  itself
• 9:03
  is
• 9:03
  becoming
• 9:03
  the
• 9:03
  primary
• 9:04
  engineering
• 9:04
  artifact,
• 9:05
  not
• 9:05
  the
• 9:05
  code
• 9:06
  it
• 9:06
  produces.
• 9:07
  A
• 9:07
  608-page
• 9:08
  DreamLab
• 9:08
  thesis,
• 9:08
  VisionFlow,
• 9:09
  attempts
• 9:09
  the
• 9:10
  bigger
• 9:10
  framing.
• 9:11
  Its
• 9:11
  central
• 9:11
  argument:
• 9:12
  AI
• 9:13
  collapses
• 9:13
  a
• 9:13
  2,000-year
• 9:14
  coordination
• 9:15
  tax
• 9:15
  —
• 9:16
  the
• 9:16
  overhead
• 9:16
  of
• 9:16
  aligning
• 9:17
  human
• 9:17
  effort
• 9:17
  across
• 9:18
  organizations.
• 9:19
  When
• 9:20
  intelligence
• 9:20
  is
• 9:20
  cheap,
• 9:21
  the
• 9:21
  organizational
• 9:22
  forms
• 9:22
  that
• 9:22
  evolved
• 9:23
  to
• 9:23
  manage
• 9:23
  expensive
• 9:24
  human
• 9:24
  coordination
• 9:24
  become
• 9:25
  deadweight.
• 9:26
  The
• 9:26
  thesis
• 9:26
  proposes
• 9:27
  three
• 9:27
  successor
• 9:27
  models:
• 9:28
  Autonomous
• 9:29
  Pods
• 9:29
  (small
• 9:29
  teams
• 9:29
  amplified
• 9:30
  by
• 9:30
  AI),
• 9:31
  Network
• 9:31
  Guilds
• 9:31
  (expertise-sharing
• 9:32
  across
• 9:33
  organizations),
• 9:34
  and
• 9:34
  Platform
• 9:34
  Cooperatives
• 9:35
  (democratically
• 9:35
  governed
• 9:35
  AI
• 9:36
  platforms).
• 9:37
  The
• 9:37
  predictions
• 9:38
  are
• 9:38
  speculative,
• 9:38
  but
• 9:39
  the
• 9:39
  question
• 9:39
  is
• 9:39
  right:
• 9:40
  every
• 9:40
  major
• 9:40
  organizational
• 9:41
  form
• 9:41
  in
• 9:41
  history
• 9:42
  emerged
• 9:43
  from
• 9:43
  a
• 9:43
  specific
• 9:43
  coordination
• 9:44
  constraint,
• 9:45
  and
• 9:45
  LLMs
• 9:45
  are
• 9:45
  dissolving
• 9:46
  several
• 9:46
  of
• 9:46
  those
• 9:46
  constraints
• 9:47
  simultaneously.
• 9:49
  Meanwhile,
• 9:49
  a
• 9:49
  Reddit
• 9:50
  thread
• 9:50
  captures
• 9:50
  the
• 9:50
  human
• 9:51
  cost
• 9:51
  of
• 9:51
  the
• 9:51
  transition
• 9:51
  with
• 9:52
  uncomfortable
• 9:52
  clarity.
• 9:53
  "Does
• 9:54
  anybody
• 9:54
  else
• 9:54
  feel
• 9:54
  like
• 9:54
  their
• 9:55
  ability
• 9:55
  to
• 9:55
  write
• 9:55
  and
• 9:55
  read
• 9:55
  code
• 9:56
  is
• 9:56
  going
• 9:56
  down
• 9:56
  day
• 9:56
  by
• 9:56
  day?"
• 9:57
  The
• 9:57
  community
• 9:58
  consensus
• 9:58
  is
• 9:58
  split
• 9:59
  but
• 9:59
  instructive:
• 10:00
  syntax
• 10:00
  memory
• 10:00
  fades,
• 10:01
  yes,
• 10:01
  but
• 10:02
  it
• 10:02
  was
• 10:02
  never
• 10:02
  the
• 10:02
  valuable
• 10:02
  skill.
• 10:03
  The
• 10:04
  top
• 10:04
  analogy
• 10:04
  —
• 10:05
  a
• 10:05
  farmer
• 10:06
  complaining
• 10:06
  that
• 10:06
  a
• 10:06
  tractor
• 10:07
  makes
• 10:07
  hand-tilling
• 10:07
  strength
• 10:08
  obsolete
• 10:08
  —
• 10:09
  is
• 10:09
  glib
• 10:09
  but
• 10:09
  directionally
• 10:10
  correct.
• 10:11
  The
• 10:11
  more
• 10:11
  nuanced
• 10:12
  responses
• 10:12
  cut
• 10:12
  deeper:
• 10:13
  "When
• 10:13
  I
• 10:13
  feel
• 10:13
  myself
• 10:14
  getting
• 10:14
  lazy,
• 10:15
  I
• 10:15
  make
• 10:15
  myself
• 10:15
  write
• 10:16
  the
• 10:16
  acceptance
• 10:16
  criteria
• 10:17
  and
• 10:17
  one
• 10:17
  failure
• 10:17
  case
• 10:17
  first.
• 10:18
  If
• 10:18
  I
• 10:19
  cannot
• 10:19
  do
• 10:19
  that,
• 10:19
  I
• 10:20
  am
• 10:20
  probably
• 10:20
  outsourcing
• 10:20
  the
• 10:21
  thinking,
• 10:21
  not
• 10:21
  the
• 10:22
  typing."
• 10:23
  Prior
• 10:23
  coverage
• 10:23
  tracked
• 10:24
  the
• 10:24
  same
• 10:24
  tension
• 10:24
  through
• 10:25
  DORA
• 10:25
  metrics
• 10:25
  and
• 10:25
  senior
• 10:26
  engineers
• 10:26
  arguing
• 10:26
  that
• 10:27
  agentic
• 10:27
  tools
• 10:27
  are
• 10:27
  seniority-biased.
• 10:29
  The
• 10:29
  tooling
• 10:29
  professionalizes;
• 10:30
  organizations
• 10:31
  have
• 10:31
  not
• 10:31
  yet
• 10:32
  figured
• 10:32
  out
• 10:32
  what
• 10:32
  the
• 10:32
  new
• 10:32
  job
• 10:32
  description
• 10:33
  actually
• 10:33
  is.
• 10:36
  ☁️
• 10:36
  Infrastructure
• 10:37
  at
• 10:37
  Scale
• 10:37
  Cloudflare's
• 10:38
  Meerkat
• 10:38
  is
• 10:38
  their
• 10:38
  internal
• 10:39
  implementation
• 10:39
  of
• 10:40
  QuePaxa,
• 10:41
  a
• 10:41
  consensus
• 10:41
  algorithm
• 10:41
  designed
• 10:42
  for
• 10:42
  the
• 10:42
  exact
• 10:42
  environment
• 10:43
  where
• 10:43
  Raft
• 10:43
  and
• 10:43
  Paxos
• 10:44
  fail:
• 10:44
  wide-area
• 10:45
  networks
• 10:45
  with
• 10:45
  unpredictable
• 10:46
  latency
• 10:46
  and
• 10:46
  no
• 10:46
  stable
• 10:47
  leader.
• 10:47
  Traditional
• 10:48
  consensus
• 10:48
  assumes
• 10:48
  a
• 10:48
  leader
• 10:49
  that
• 10:49
  can
• 10:49
  coordinate
• 10:49
  rounds
• 10:49
  within
• 10:50
  tight
• 10:50
  timeout
• 10:50
  bounds.
• 10:51
  On
• 10:51
  a
• 10:51
  CDN
• 10:51
  spanning
• 10:52
  300+
• 10:52
  cities,
• 10:53
  that
• 10:53
  assumption
• 10:53
  breaks
• 10:54
  constantly
• 10:54
  —
• 10:54
  leaders
• 10:55
  time
• 10:55
  out,
• 10:55
  new
• 10:55
  elections
• 10:56
  disrupt
• 10:56
  in-flight
• 10:56
  transactions,
• 10:57
  and
• 10:57
  throughput
• 10:57
  collapses
• 10:58
  during
• 10:58
  the
• 10:58
  handoff.
• 10:59
  QuePaxa
• 10:59
  sidesteps
• 11:00
  this
• 11:00
  by
• 11:00
  running
• 11:00
  leaderless:
• 11:01
  every
• 11:01
  node
• 11:01
  can
• 11:01
  propose,
• 11:02
  and
• 11:02
  the
• 11:02
  algorithm
• 11:02
  combines
• 11:03
  optimistic
• 11:03
  and
• 11:03
  pessimistic
• 11:04
  paths
• 11:04
  to
• 11:04
  reach
• 11:04
  agreement
• 11:05
  without
• 11:05
  timeout-driven
• 11:05
  failure
• 11:06
  modes.
• 11:06
  The
• 11:06
  Cloudflare
• 11:07
  post
• 11:07
  walks
• 11:07
  through
• 11:07
  both
• 11:08
  the
• 11:08
  theoretical
• 11:08
  foundations
• 11:09
  and
• 11:09
  the
• 11:09
  engineering
• 11:09
  tradeoffs
• 11:10
  of
• 11:10
  deploying
• 11:10
  it
• 11:11
  at
• 11:11
  global
• 11:11
  scale
• 11:11
  for
• 11:12
  configuration
• 11:12
  propagation
• 11:13
  and
• 11:13
  distributed
• 11:13
  KV
• 11:13
  coordination.
• 11:14
  SkyPilot
• 11:15
  and
• 11:15
  Hugging
• 11:15
  Face
• 11:15
  jointly
• 11:16
  tackled
• 11:16
  the
• 11:16
  other
• 11:16
  infrastructure
• 11:17
  lock-in:
• 11:17
  egress
• 11:17
  costs.
• 11:18
  Most
• 11:18
  AI
• 11:18
  teams
• 11:19
  hold
• 11:19
  GPU
• 11:19
  capacity
• 11:20
  across
• 11:20
  multiple
• 11:21
  clouds
• 11:21
  —
• 11:21
  reserved
• 11:22
  H100s
• 11:22
  on
• 11:22
  AWS,
• 11:23
  a
• 11:23
  cluster
• 11:24
  on
• 11:24
  GCP,
• 11:24
  maybe
• 11:25
  an
• 11:25
  on-prem
• 11:25
  rack
• 11:25
  —
• 11:26
  but
• 11:26
  their
• 11:26
  data
• 11:26
  lives
• 11:26
  in
• 11:26
  a
• 11:26
  single-cloud
• 11:27
  bucket.
• 11:27
  Moving
• 11:28
  a
• 11:28
  model
• 11:28
  to
• 11:28
  wherever
• 11:29
  GPUs
• 11:29
  are
• 11:29
  available
• 11:29
  means
• 11:30
  paying
• 11:30
  egress
• 11:30
  ($0.09/GB
• 11:32
  on
• 11:32
  AWS),
• 11:33
  which
• 11:33
  in
• 11:33
  practice
• 11:33
  pins
• 11:33
  workloads
• 11:34
  to
• 11:34
  whichever
• 11:34
  vendor
• 11:34
  holds
• 11:35
  the
• 11:35
  data
• 11:35
  and
• 11:35
  leaves
• 11:35
  reserved
• 11:36
  capacity
• 11:36
  elsewhere
• 11:36
  idle.
• 11:37
  The
• 11:38
  integration
• 11:38
  makes
• 11:38
  Hugging
• 11:39
  Face
• 11:39
  Storage
• 11:39
  a
• 11:39
  first-class
• 11:40
  SkyPilot
• 11:40
  backend
• 11:40
  with
• 11:41
  zero
• 11:41
  egress
• 11:41
  fees.
• 11:43
  Models
• 11:43
  stay
• 11:43
  on
• 11:43
  the
• 11:43
  Hub,
• 11:44
  SkyPilot
• 11:44
  schedules
• 11:44
  compute
• 11:45
  across
• 11:45
  20+
• 11:46
  clouds,
• 11:46
  and
• 11:47
  lazy
• 11:47
  FUSE
• 11:47
  mounts
• 11:47
  pull
• 11:48
  only
• 11:48
  the
• 11:48
  bytes
• 11:48
  actually
• 11:48
  accessed.
• 11:49
  Xet-backed
• 11:50
  content-defined
• 11:51
  chunking
• 11:51
  means
• 11:51
  checkpoint
• 11:52
  uploads
• 11:52
  only
• 11:52
  transfer
• 11:53
  changed
• 11:53
  chunks
• 11:53
  —
• 11:54
  re-uploading
• 11:55
  an
• 11:55
  8.43
• 11:56
  GB
• 11:56
  checkpoint
• 11:56
  already
• 11:57
  stored
• 11:57
  took
• 11:57
  8
• 11:57
  seconds
• 11:58
  versus
• 11:58
  24
• 11:59
  for
• 11:59
  the
• 11:59
  initial
• 11:59
  upload.
• 12:00
  Storage
• 12:00
  stops
• 12:01
  deciding
• 12:01
  where
• 12:01
  you
• 12:01
  run.
• 12:02
  For
• 12:02
  teams
• 12:03
  juggling
• 12:03
  reserved
• 12:03
  capacity
• 12:04
  across
• 12:04
  three
• 12:04
  or
• 12:04
  four
• 12:05
  vendors,
• 12:05
  this
• 12:06
  is
• 12:06
  the
• 12:06
  plumbing
• 12:06
  that
• 12:06
  turns
• 12:06
  multi-cloud
• 12:07
  from
• 12:07
  a
• 12:07
  procurement
• 12:08
  strategy
• 12:08
  into
• 12:08
  an
• 12:08
  operational
• 12:09
  reality.
• 12:11
  🏛️
• 12:11
  Privacy,
• 12:12
  Policy,
• 12:12
  and
• 12:13
  Open
• 12:13
  Source
• 12:13
  Governance
• 12:14
  Flipper
• 12:14
  Zero's
• 12:14
  development
• 12:15
  blog
• 12:15
  is
• 12:16
  a
• 12:16
  quiet
• 12:16
  signal
• 12:16
  worth
• 12:16
  reading
• 12:17
  closely.
• 12:18
  The
• 12:18
  team
• 12:18
  outlines
• 12:19
  a
• 12:19
  future
• 12:19
  focused
• 12:19
  on
• 12:19
  open
• 12:20
  protocol
• 12:20
  analysis,
• 12:21
  education,
• 12:21
  and
• 12:22
  legitimate
• 12:22
  security
• 12:22
  research
• 12:23
  —
• 12:24
  a
• 12:24
  deliberate
• 12:24
  repositioning
• 12:25
  against
• 12:25
  the
• 12:25
  "hacking
• 12:25
  gadget"
• 12:26
  narrative
• 12:26
  that
• 12:26
  has
• 12:26
  drawn
• 12:27
  regulatory
• 12:27
  scrutiny
• 12:27
  in
• 12:27
  multiple
• 12:28
  jurisdictions.
• 12:29
  The
• 12:29
  tension
• 12:29
  is
• 12:30
  real
• 12:30
  and
• 12:30
  unresolved:
• 12:31
  the
• 12:31
  same
• 12:32
  capabilities
• 12:32
  that
• 12:32
  make
• 12:33
  Flipper
• 12:33
  useful
• 12:33
  for
• 12:33
  security
• 12:34
  professionals
• 12:35
  make
• 12:35
  regulators
• 12:35
  nervous,
• 12:36
  and
• 12:36
  no
• 12:36
  amount
• 12:36
  of
• 12:36
  branding
• 12:37
  resolves
• 12:37
  the
• 12:37
  dual-use
• 12:38
  problem
• 12:38
  when
• 12:39
  legislators
• 12:39
  do
• 12:39
  not
• 12:39
  distinguish
• 12:40
  between
• 12:40
  tools
• 12:41
  and
• 12:41
  intent.
• 12:42
  A
• 12:42
  university
• 12:42
  professor
• 12:43
  pushed
• 12:43
  a
• 12:43
  batch
• 12:43
  of
• 12:43
  PRs
• 12:44
  to
• 12:44
  make
• 12:44
  Open
• 12:44
  WebUI
• 12:45
  WCAG
• 12:45
  2.1
• 12:46
  AA
• 12:46
  compliant
• 12:46
  —
• 12:47
  keyboard
• 12:48
  focus,
• 12:48
  skip-to-content
• 12:49
  links,
• 12:49
  dialog
• 12:49
  roles,
• 12:50
  contrast
• 12:50
  ratios,
• 12:51
  accessible
• 12:51
  names
• 12:51
  for
• 12:52
  icon-only
• 12:52
  buttons.
• 12:53
  The
• 12:53
  motivation
• 12:53
  is
• 12:54
  ADA
• 12:54
  Title
• 12:54
  II
• 12:54
  /
• 12:55
  Section
• 12:55
  508:
• 12:56
  any
• 12:56
  software
• 12:56
  used
• 12:57
  in
• 12:57
  university
• 12:57
  classes
• 12:58
  must
• 12:58
  meet
• 12:58
  minimum
• 12:58
  accessibility
• 12:59
  standards,
• 13:00
  and
• 13:00
  most
• 13:00
  AI
• 13:00
  tools
• 13:01
  fail
• 13:01
  basic
• 13:01
  audits.
• 13:02
  Each
• 13:02
  PR
• 13:02
  is
• 13:02
  scoped
• 13:03
  to
• 13:03
  a
• 13:03
  single
• 13:03
  WCAG
• 13:04
  criterion
• 13:04
  and
• 13:04
  tested.
• 13:05
  The
• 13:05
  implicit
• 13:06
  point
• 13:06
  is
• 13:06
  sharp:
• 13:07
  the
• 13:07
  AI
• 13:07
  tools
• 13:07
  community
• 13:08
  builds
• 13:08
  for
• 13:08
  power
• 13:08
  users
• 13:09
  and
• 13:09
  ignores
• 13:09
  the
• 13:10
  legal
• 13:10
  and
• 13:10
  ethical
• 13:10
  obligation
• 13:11
  to
• 13:11
  make
• 13:11
  interfaces
• 13:12
  usable
• 13:12
  by
• 13:12
  everyone.
• 13:13
  Accessibility
• 13:14
  is
• 13:14
  not
• 13:14
  a
• 13:14
  feature
• 13:14
  request
• 13:15
  —
• 13:15
  in
• 13:15
  educational
• 13:16
  and
• 13:16
  government
• 13:17
  deployments,
• 13:17
  it
• 13:17
  is
• 13:17
  a
• 13:17
  compliance
• 13:18
  gate.
• 13:19
  The
• 13:19
  biggest
• 13:19
  policy
• 13:20
  story
• 13:20
  this
• 13:20
  week
• 13:21
  is
• 13:21
  the
• 13:21
  EU
• 13:21
  Chat
• 13:21
  Control
• 13:22
  saga,
• 13:22
  and
• 13:22
  the
• 13:23
  procedural
• 13:23
  details
• 13:23
  matter
• 13:24
  more
• 13:24
  than
• 13:24
  the
• 13:24
  headlines.
• 13:25
  Chat
• 13:25
  Control
• 13:26
  1.0
• 13:26
  —
• 13:27
  the
• 13:27
  "temporary"
• 13:28
  ePrivacy
• 13:28
  derogation
• 13:29
  that
• 13:29
  let
• 13:29
  providers
• 13:30
  voluntarily
• 13:30
  scan
• 13:31
  private
• 13:31
  messages
• 13:31
  —
• 13:32
  expired
• 13:33
  April
• 13:33
  4
• 13:33
  after
• 13:33
  Parliament
• 13:34
  rejected
• 13:34
  its
• 13:34
  extension
• 13:35
  311
• 13:36
  to
• 13:36
  228.
• 13:38
  That
• 13:38
  should
• 13:38
  have
• 13:38
  been
• 13:38
  the
• 13:38
  end.
• 13:39
  Instead,
• 13:40
  the
• 13:40
  Council
• 13:40
  is
• 13:40
  attempting
• 13:41
  an
• 13:41
  unprecedented
• 13:42
  fast-track
• 13:42
  revival:
• 13:43
  a
• 13:43
  formally
• 13:44
  "new"
• 13:44
  regulation
• 13:44
  with
• 13:45
  identical
• 13:45
  content,
• 13:46
  pushed
• 13:46
  through
• 13:46
  expedited
• 13:47
  procedure.
• 13:48
  Parliament
• 13:48
  voted
• 13:48
  last
• 13:49
  week
• 13:49
  to
• 13:49
  skip
• 13:49
  committee
• 13:49
  review
• 13:50
  entirely.
• 13:51
  The
• 13:51
  binding
• 13:51
  vote
• 13:51
  is
• 13:52
  Thursday,
• 13:52
  July
• 13:52
  9,
• 13:53
  and
• 13:53
  the
• 13:54
  procedural
• 13:54
  trap
• 13:54
  is
• 13:54
  brutal:
• 13:55
  an
• 13:56
  absolute
• 13:56
  majority
• 13:56
  of
• 13:57
  361
• 13:58
  MEPs
• 13:58
  is
• 13:58
  needed
• 13:58
  to
• 13:59
  stop
• 13:59
  it
• 13:59
  —
• 14:00
  abstentions
• 14:01
  and
• 14:01
  absences
• 14:01
  count
• 14:01
  as
• 14:02
  votes
• 14:02
  in
• 14:02
  favor.
• 14:03
  Meanwhile,
• 14:04
  Chat
• 14:04
  Control
• 14:05
  2.0
• 14:05
  —
• 14:06
  the
• 14:06
  permanent
• 14:07
  CSAR
• 14:07
  mandating
• 14:08
  CSAM
• 14:08
  detection
• 14:08
  —
• 14:09
  remains
• 14:10
  deadlocked
• 14:10
  after
• 14:10
  five
• 14:11
  failed
• 14:11
  trilogue
• 14:11
  rounds,
• 14:12
  with
• 14:12
  encryption
• 14:13
  scanning
• 14:13
  as
• 14:13
  the
• 14:13
  unresolved
• 14:14
  red
• 14:14
  line.
• 14:15
  The
• 14:15
  Council's
• 14:16
  own
• 14:16
  legal
• 14:16
  service
• 14:17
  stated
• 14:17
  that
• 14:17
  "voluntary"
• 14:18
  suspicionless
• 14:18
  scanning
• 14:19
  constitutes
• 14:20
  generalized
• 14:20
  surveillance
• 14:21
  incompatible
• 14:21
  with
• 14:22
  Charter
• 14:22
  Article
• 14:22
  7.
• 14:23
  Both
• 14:24
  fights
• 14:24
  are
• 14:24
  live,
• 14:25
  both
• 14:25
  depend
• 14:25
  on
• 14:26
  whether
• 14:26
  enough
• 14:26
  MEPs
• 14:26
  show
• 14:27
  up
• 14:27
  Thursday,
• 14:28
  and
• 14:28
  the
• 14:28
  temporary
• 14:28
  law
• 14:29
  being
• 14:29
  resurrected
• 14:29
  through
• 14:30
  the
• 14:30
  back
• 14:30
  door
• 14:30
  while
• 14:30
  the
• 14:31
  permanent
• 14:31
  one
• 14:31
  remains
• 14:31
  on
• 14:31
  the
• 14:32
  table
• 14:32
  is
• 14:32
  exactly
• 14:33
  the
• 14:33
  kind
• 14:33
  of
• 14:33
  procedural
• 14:34
  maneuver
• 14:34
  that
• 14:34
  erodes
• 14:35
  trust
• 14:35
  in
• 14:35
  EU
• 14:36
  legislative
• 14:36
  institutions.