Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 07, 2026

Transcript

AGI Dreams Podcast – April 07, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  Google
• 0:00
  DeepMind's
• 0:00
  Gemma
• 0:01
  4
• 0:01
  family
• 0:01
  landed
• 0:02
  this
• 0:02
  week
• 0:02
  with
• 0:02
  something
• 0:03
  the
• 0:03
  open-weight
• 0:03
  community
• 0:04
  rarely
• 0:04
  gets:
• 0:04
  a
• 0:05
  genuinely
• 0:05
  competitive
• 0:06
  multimodal
• 0:06
  model
• 0:07
  under
• 0:07
  a
• 0:07
  permissive
• 0:07
  Apache
• 0:08
  2
• 0:08
  license.
• 0:09
  The
• 0:09
  lineup
• 0:09
  spans
• 0:10
  four
• 0:10
  sizes
• 0:10
  —
• 0:11
  from
• 0:11
  the
• 0:11
  tiny
• 0:11
  E2B
• 0:12
  (2.3B
• 0:13
  effective
• 0:13
  parameters)
• 0:14
  up
• 0:14
  to
• 0:14
  a
• 0:14
  31B
• 0:15
  dense
• 0:15
  model
• 0:15
  —
• 0:16
  but
• 0:16
  the
• 0:16
  star
• 0:16
  is
• 0:16
  the
• 0:16
  26B-A4B,
• 0:18
  a
• 0:18
  mixture-of-experts
• 0:19
  architecture
• 0:20
  with
• 0:20
  128
• 0:21
  experts
• 0:22
  and
• 0:22
  only
• 0:22
  8
• 0:22
  active
• 0:23
  per
• 0:23
  token
• 0:23
  (roughly
• 0:23
  4B
• 0:24
  parameters
• 0:24
  per
• 0:24
  forward
• 0:25
  pass).
• 0:26
  The
• 0:26
  result
• 0:26
  is
• 0:26
  a
• 0:26
  model
• 0:26
  that
• 0:27
  achieves
• 0:27
  an
• 0:27
  LMArena
• 0:28
  Elo
• 0:28
  of
• 0:28
  ~1441
• 0:29
  while
• 0:29
  fitting
• 0:30
  comfortably
• 0:30
  on
• 0:30
  a
• 0:30
  48GB
• 0:31
  MacBook
• 0:31
  Pro
• 0:31
  at
• 0:32
  51
• 0:32
  tokens
• 0:32
  per
• 0:32
  second.
• 0:34
  For
• 0:34
  context,
• 0:35
  reaching
• 0:35
  comparable
• 0:35
  Elo
• 0:36
  scores
• 0:36
  typically
• 0:36
  requires
• 0:37
  models
• 0:37
  with
• 0:37
  400B+
• 0:38
  total
• 0:39
  parameters.
• 0:40
  The
• 0:40
  smaller
• 0:41
  E2B
• 0:41
  and
• 0:41
  E4B
• 0:41
  variants
• 0:42
  also
• 0:42
  support
• 0:43
  audio
• 0:43
  input,
• 0:43
  making
• 0:44
  them
• 0:44
  genuine
• 0:44
  multimodal
• 0:45
  on-device
• 0:45
  candidates.
• 0:46
  All
• 0:47
  models
• 0:47
  support
• 0:47
  vision,
• 0:48
  text,
• 0:49
  function
• 0:49
  calling,
• 0:50
  and
• 0:50
  configurable
• 0:50
  thinking
• 0:51
  modes,
• 0:51
  with
• 0:51
  day-zero
• 0:52
  support
• 0:52
  across
• 0:52
  transformers,
• 0:53
  llama.cpp,
• 0:55
  MLX,
• 0:56
  WebGPU,
• 0:57
  and
• 0:57
  Rust.
• 0:58
  The
• 0:58
  community
• 0:59
  response
• 0:59
  was
• 0:59
  immediate
• 1:00
  and
• 1:00
  practical.
• 1:01
  George
• 1:02
  Liu
• 1:02
  published
• 1:02
  a
• 1:02
  detailed
• 1:03
  walkthrough
• 1:03
  of
• 1:03
  running
• 1:04
  the
• 1:04
  26B-A4B
• 1:06
  locally
• 1:06
  using
• 1:07
  LM
• 1:07
  Studio
• 1:07
  0.4.0's
• 1:09
  new
• 1:09
  headless
• 1:09
  CLI
• 1:10
  (lms
• 1:10
  and
• 1:10
  llmster),
• 1:11
  which
• 1:11
  extracts
• 1:12
  the
• 1:12
  inference
• 1:12
  engine
• 1:13
  into
• 1:13
  a
• 1:13
  standalone
• 1:13
  daemon
• 1:14
  —
• 1:14
  no
• 1:14
  GUI
• 1:15
  required.
• 1:16
  On
• 1:16
  his
• 1:16
  M4
• 1:16
  Pro
• 1:17
  with
• 1:17
  48GB
• 1:18
  unified
• 1:18
  memory,
• 1:19
  the
• 1:19
  Q4
• 1:19
  K
• 1:19
  M
• 1:20
  quantization
• 1:20
  (17.99GB)
• 1:23
  delivers
• 1:24
  51
• 1:24
  tok/s
• 1:25
  with
• 1:25
  a
• 1:25
  48K
• 1:26
  context
• 1:26
  window
• 1:27
  and
• 1:27
  1.5s
• 1:28
  time
• 1:28
  to
• 1:28
  first
• 1:28
  token.
• 1:29
  Liu
• 1:29
  also
• 1:30
  integrated
• 1:30
  it
• 1:30
  with
• 1:30
  Claude
• 1:31
  Code
• 1:31
  via
• 1:31
  a
• 1:31
  local
• 1:32
  OpenAI-compatible
• 1:33
  API,
• 1:34
  creating
• 1:34
  a
• 1:34
  zero-cost
• 1:35
  local
• 1:35
  inference
• 1:35
  backend
• 1:36
  for
• 1:36
  agentic
• 1:36
  workflows.
• 1:38
  His
• 1:38
  memory
• 1:38
  profiling
• 1:39
  shows
• 1:39
  the
• 1:39
  model
• 1:39
  fits
• 1:40
  at
• 1:40
  full
• 1:40
  256K
• 1:41
  context
• 1:42
  within
• 1:42
  37.48GB,
• 1:45
  leaving
• 1:45
  headroom
• 1:46
  on
• 1:46
  a
• 1:46
  48GB
• 1:47
  machine.
• 1:48
  The
• 1:48
  key
• 1:48
  insight
• 1:49
  for
• 1:49
  MoE
• 1:49
  models:
• 1:50
  skip
• 1:50
  speculative
• 1:51
  decoding
• 1:51
  entirely
• 1:52
  —
• 1:52
  it
• 1:52
  blows
• 1:53
  up
• 1:53
  memory
• 1:53
  bandwidth
• 1:54
  by
• 1:54
  loading
• 1:54
  the
• 1:54
  union
• 1:55
  of
• 1:55
  all
• 1:55
  experts
• 1:55
  across
• 1:56
  speculative
• 1:56
  tokens,
• 1:57
  with
• 1:57
  benchmarks
• 1:58
  on
• 1:58
  Mixtral
• 1:58
  showing
• 1:59
  a
• 1:59
  54%
• 2:00
  slowdown
• 2:00
  on
• 2:00
  math
• 2:00
  tasks.
• 2:02
  Meanwhile,
• 2:03
  a
• 2:03
  community
• 2:03
  member
• 2:03
  discovered
• 2:04
  that
• 2:04
  swapping
• 2:05
  the
• 2:05
  F16
• 2:05
  vision
• 2:06
  projector
• 2:06
  for
• 2:06
  Q8
• 2:07
  0
• 2:07
  loses
• 2:07
  no
• 2:08
  quality
• 2:08
  —
• 2:09
  and
• 2:09
  actually
• 2:09
  performed
• 2:10
  slightly
• 2:10
  better
• 2:10
  in
• 2:10
  some
• 2:11
  tests
• 2:11
  —
• 2:12
  while
• 2:12
  freeing
• 2:12
  roughly
• 2:13
  30K
• 2:13
  tokens
• 2:14
  of
• 2:14
  context,
• 2:15
  enabling
• 2:15
  60K+
• 2:16
  total
• 2:17
  context
• 2:17
  with
• 2:17
  vision
• 2:18
  still
• 2:18
  active.
• 2:19
  And
• 2:19
  within
• 2:19
  days
• 2:20
  of
• 2:20
  Gemma
• 2:20
  4's
• 2:20
  release,
• 2:21
  an
• 2:21
  abliterated
• 2:22
  variant
• 2:22
  appeared
• 2:23
  using
• 2:23
  Adaptive
• 2:24
  Refusal
• 2:24
  Abliteration
• 2:25
  (ARA),
• 2:26
  a
• 2:26
  two-pass
• 2:26
  SVD
• 2:27
  technique
• 2:27
  that
• 2:28
  achieved
• 2:28
  the
• 2:28
  lowest
• 2:28
  refusal
• 2:29
  rate,
• 2:29
  highest
• 2:30
  quality
• 2:30
  score,
• 2:31
  and
• 2:31
  lowest
• 2:32
  KL
• 2:32
  divergence
• 2:33
  of
• 2:33
  any
• 2:33
  published
• 2:33
  abliteration
• 2:34
  of
• 2:34
  this
• 2:34
  model.
• 2:35
  Two
• 2:35
  optimized
• 2:36
  builds
• 2:36
  are
• 2:37
  available:
• 2:38
  APEX
• 2:38
  IQ
• 2:38
  Q5
• 2:39
  K
• 2:39
  M
• 2:39
  (~20
• 2:40
  GB,
• 2:41
  imatrix-calibrated
• 2:42
  for
• 2:42
  llama.cpp)
• 2:44
  and
• 2:44
  MLX
• 2:45
  mixed-bit
• 2:45
  (~15
• 2:46
  GB,
• 2:46
  tiered
• 2:47
  4/6/8-bit
• 2:48
  for
• 2:48
  Apple
• 2:48
  Silicon,
• 2:49
  both
• 2:49
  with
• 2:49
  vision
• 2:49
  support.
• 2:50
  Full
• 2:51
  BF16
• 2:51
  SafeTensors
• 2:52
  and
• 2:52
  GGUF
• 2:53
  are
• 2:53
  also
• 2:53
  available.
• 2:58
  The
• 2:58
  on-device
• 2:59
  AI
• 2:59
  frontier
• 2:59
  keeps
• 3:00
  pushing
• 3:00
  lower.
• 3:01
  A
• 3:01
  Caltech
• 3:01
  spinout
• 3:02
  claims
• 3:02
  a
• 3:02
  1-bit
• 3:02
  8B
• 3:03
  parameter
• 3:03
  model
• 3:04
  fitting
• 3:04
  in
• 3:04
  1.15GB
• 3:06
  of
• 3:06
  memory
• 3:06
  that's
• 3:07
  competitive
• 3:07
  with
• 3:07
  Llama
• 3:07
  3
• 3:08
  8B
• 3:08
  on
• 3:08
  benchmarks
• 3:09
  —
• 3:10
  440
• 3:10
  tok/s
• 3:11
  on
• 3:11
  a
• 3:11
  4090,
• 3:12
  136
• 3:13
  tok/s
• 3:14
  on
• 3:14
  M4
• 3:14
  Pro,
• 3:15
  and
• 3:15
  reportedly
• 3:16
  ~40
• 3:17
  tok/s
• 3:17
  on
• 3:17
  an
• 3:17
  iPhone.
• 3:18
  If
• 3:19
  those
• 3:19
  numbers
• 3:19
  hold
• 3:19
  up
• 3:19
  under
• 3:20
  independent
• 3:20
  testing,
• 3:21
  a
• 3:21
  private
• 3:21
  LLM
• 3:22
  on
• 3:22
  a
• 3:22
  phone
• 3:22
  stops
• 3:23
  being
• 3:23
  a
• 3:23
  party
• 3:23
  trick
• 3:23
  and
• 3:24
  starts
• 3:24
  being
• 3:24
  a
• 3:24
  product.
• 3:25
  The
• 3:25
  skepticism
• 3:26
  is
• 3:26
  warranted
• 3:26
  —
• 3:27
  benchmark-maxing
• 3:28
  startups
• 3:28
  are
• 3:29
  a
• 3:29
  dime
• 3:29
  a
• 3:29
  dozen
• 3:29
  —
• 3:30
  but
• 3:30
  the
• 3:30
  architecture
• 3:31
  (ternary
• 3:31
  weights,
• 3:32
  pure
• 3:32
  add/subtract
• 3:33
  computation)
• 3:34
  is
• 3:34
  sound
• 3:34
  enough
• 3:34
  to
• 3:34
  deserve
• 3:35
  scrutiny.
• 3:36
  At
• 3:36
  the
• 3:36
  opposite
• 3:37
  extreme,
• 3:37
  a
• 3:38
  Samsung
• 3:38
  Galaxy
• 3:38
  Watch
• 3:39
  4
• 3:39
  Classic
• 3:39
  is
• 3:39
  now
• 3:40
  running
• 3:40
  SmolLM2-360M
• 3:43
  thanks
• 3:43
  to
• 3:44
  a
• 3:44
  clever
• 3:44
  llama.cpp
• 3:45
  patch.
• 3:46
  The
• 3:47
  problem:
• 3:48
  the
• 3:48
  default
• 3:48
  model
• 3:48
  loading
• 3:49
  path
• 3:49
  maps
• 3:49
  the
• 3:49
  file
• 3:50
  into
• 3:50
  memory
• 3:50
  via
• 3:50
  mmap,
• 3:51
  then
• 3:51
  copies
• 3:52
  tensors
• 3:52
  into
• 3:52
  ggml
• 3:53
  allocations
• 3:53
  —
• 3:55
  effectively
• 3:55
  loading
• 3:55
  the
• 3:55
  270MB
• 3:58
  model
• 3:59
  twice,
• 3:59
  peaking
• 4:00
  at
• 4:00
  524MB
• 4:02
  in
• 4:02
  a
• 4:02
  device
• 4:03
  with
• 4:03
  380MB
• 4:04
  free.
• 4:05
  The
• 4:05
  fix
• 4:06
  passes
• 4:06
  host
• 4:06
  ptr
• 4:07
  into
• 4:07
  llama
• 4:07
  model
• 4:08
  params
• 4:08
  so
• 4:09
  CPU
• 4:09
  tensors
• 4:10
  point
• 4:10
  directly
• 4:10
  into
• 4:10
  the
• 4:10
  mmap
• 4:11
  region,
• 4:11
  cutting
• 4:12
  peak
• 4:12
  RAM
• 4:12
  from
• 4:12
  524MB
• 4:15
  to
• 4:15
  142MB
• 4:16
  (74%
• 4:17
  reduction)
• 4:18
  and
• 4:18
  boot
• 4:19
  time
• 4:19
  from
• 4:19
  19s
• 4:20
  to
• 4:20
  11s.
• 4:21
  The
• 4:21
  patch
• 4:22
  is
• 4:22
  heading
• 4:22
  upstream
• 4:22
  as
• 4:23
  a
• 4:23
  PR
• 4:23
  to
• 4:23
  ggml-org/llama.cpp.
• 4:26
  A
• 4:26
  commenter's
• 4:27
  observation
• 4:28
  lands
• 4:28
  perfectly:
• 4:29
  "In
• 4:29
  2026,
• 4:31
  a
• 4:31
  watch
• 4:31
  has
• 4:31
  380
• 4:32
  Megabytes
• 4:33
  of
• 4:33
  free
• 4:33
  RAM.
• 4:34
  My
• 4:34
  first
• 4:34
  computer
• 4:35
  had
• 4:35
  80
• 4:35
  Megabytes
• 4:36
  of
• 4:36
  total
• 4:36
  hard
• 4:37
  drive
• 4:37
  space."
• 4:38
  And
• 4:38
  then
• 4:38
  there's
• 4:39
  the
• 4:39
  lawyer
• 4:39
  from
• 4:39
  South
• 4:40
  Carolina
• 4:40
  who
• 4:40
  built
• 4:40
  a
• 4:40
  10x
• 4:41
  NVIDIA
• 4:41
  V100
• 4:42
  SXM2
• 4:43
  server
• 4:43
  with
• 4:43
  320GB
• 4:45
  of
• 4:45
  VRAM,
• 4:46
  entirely
• 4:46
  orchestrated
• 4:47
  through
• 4:47
  Claude
• 4:47
  Code
• 4:48
  over
• 4:48
  SSH.
• 4:49
  His
• 4:49
  candid
• 4:49
  writeup
• 4:50
  —
• 4:50
  "this
• 4:51
  is
• 4:51
  just
• 4:51
  the
• 4:51
  corniest
• 4:52
  mid-life
• 4:52
  crisis
• 4:53
  I
• 4:53
  could
• 4:53
  have
• 4:53
  ever
• 4:53
  had"
• 4:54
  —
• 4:55
  details
• 4:55
  the
• 4:55
  real-world
• 4:56
  friction
• 4:56
  of
• 4:56
  V100
• 4:57
  hardware:
• 4:58
  no
• 4:58
  FlashAttention2
• 4:58
  (requires
• 4:59
  SM
• 4:59
  80+),
• 5:00
  no
• 5:01
  GPTQ
• 5:01
  (ExLlamaV2
• 5:02
  broken
• 5:03
  on
• 5:03
  SM
• 5:03
  7.0),
• 5:04
  no
• 5:04
  FP8,
• 5:05
  and
• 5:05
  an
• 5:06
  NCCL
• 5:06
  dependency
• 5:07
  conflict
• 5:07
  where
• 5:07
  pip
• 5:08
  install
• 5:08
  -e
• 5:08
  .
• 5:09
  pulls
• 5:09
  in
• 5:09
  cu13
• 5:10
  NCCL
• 5:11
  alongside
• 5:11
  cu12
• 5:12
  PyTorch,
• 5:13
  silently
• 5:13
  breaking
• 5:14
  every
• 5:14
  multi-GPU
• 5:15
  launch.
• 5:16
  His
• 5:16
  benchmarks
• 5:16
  show
• 5:17
  Command
• 5:17
  R
• 5:17
  32B
• 5:18
  at
• 5:18
  35.2
• 5:19
  tok/s
• 5:20
  on
• 5:20
  TP=4,
• 5:22
  Gemma
• 5:22
  4
• 5:23
  31B
• 5:23
  at
• 5:23
  21.6
• 5:24
  tok/s,
• 5:25
  and
• 5:25
  Qwen
• 5:25
  2.5
• 5:26
  72B
• 5:27
  at
• 5:27
  14.9
• 5:28
  tok/s
• 5:29
  on
• 5:29
  TP=4
• 5:30
  PP=2.
• 5:32
  Community
• 5:32
  feedback
• 5:33
  was
• 5:33
  blunt:
• 5:34
  llama.cpp
• 5:36
  with
• 5:36
  GGUF
• 5:36
  quants
• 5:37
  would
• 5:37
  outperform
• 5:37
  vLLM
• 5:38
  on
• 5:38
  V100s
• 5:39
  for
• 5:39
  most
• 5:39
  workloads,
• 5:40
  with
• 5:40
  one
• 5:40
  commenter
• 5:41
  reporting
• 5:41
  45
• 5:42
  tok/s
• 5:42
  for
• 5:42
  Qwen3.5
• 5:45
  122B-A10B
• 5:48
  on
• 5:48
  8x
• 5:49
  V100s
• 5:49
  via
• 5:50
  llama.cpp.
• 5:56
  The
• 5:56
  New
• 5:56
  Yorker
• 5:56
  published
• 5:57
  what
• 5:57
  may
• 5:57
  be
• 5:57
  the
• 5:57
  most
• 5:57
  comprehensive
• 5:58
  profile
• 5:58
  of
• 5:59
  Sam
• 5:59
  Altman
• 5:59
  to
• 5:59
  date,
• 6:00
  drawing
• 6:00
  on
• 6:00
  over
• 6:01
  a
• 6:01
  hundred
• 6:01
  interviews
• 6:02
  and
• 6:02
  previously
• 6:02
  undisclosed
• 6:03
  documents
• 6:03
  —
• 6:04
  including
• 6:04
  the
• 6:04
  "Ilya
• 6:05
  Memos,"
• 6:05
  seventy
• 6:06
  pages
• 6:06
  of
• 6:06
  Slack
• 6:07
  messages
• 6:07
  and
• 6:07
  HR
• 6:08
  documents
• 6:08
  that
• 6:08
  Sutskever
• 6:09
  compiled
• 6:10
  and
• 6:10
  sent
• 6:10
  as
• 6:10
  disappearing
• 6:11
  messages
• 6:11
  to
• 6:11
  fellow
• 6:11
  board
• 6:12
  members
• 6:12
  before
• 6:12
  the
• 6:12
  November
• 6:13
  2023
• 6:13
  firing.
• 6:14
  The
• 6:15
  memos
• 6:15
  allege
• 6:15
  a
• 6:15
  pattern
• 6:16
  of
• 6:16
  misrepresentation:
• 6:17
  Altman
• 6:18
  purportedly
• 6:18
  told
• 6:18
  the
• 6:18
  board
• 6:19
  that
• 6:19
  GPT-4
• 6:19
  features
• 6:20
  had
• 6:20
  been
• 6:20
  approved
• 6:20
  by
• 6:20
  a
• 6:21
  safety
• 6:21
  panel
• 6:21
  when
• 6:21
  they
• 6:21
  hadn't,
• 6:22
  neglected
• 6:22
  to
• 6:22
  mention
• 6:23
  that
• 6:23
  Microsoft
• 6:24
  had
• 6:24
  released
• 6:24
  an
• 6:24
  early
• 6:25
  ChatGPT
• 6:25
  version
• 6:26
  in
• 6:26
  India
• 6:26
  without
• 6:27
  completing
• 6:27
  a
• 6:27
  required
• 6:28
  safety
• 6:28
  review,
• 6:29
  and
• 6:29
  privately
• 6:29
  assured
• 6:30
  different
• 6:30
  factions
• 6:30
  contradictory
• 6:31
  things
• 6:31
  about
• 6:31
  Brockman's
• 6:32
  role.
• 6:33
  A
• 6:33
  board
• 6:33
  member
• 6:33
  describes
• 6:34
  Altman
• 6:34
  as
• 6:34
  having
• 6:35
  "two
• 6:35
  traits
• 6:35
  that
• 6:35
  are
• 6:35
  almost
• 6:36
  never
• 6:36
  seen
• 6:36
  in
• 6:36
  the
• 6:36
  same
• 6:37
  person:
• 6:38
  a
• 6:38
  strong
• 6:38
  desire
• 6:38
  to
• 6:38
  please
• 6:39
  people,
• 6:39
  to
• 6:40
  be
• 6:40
  liked
• 6:40
  in
• 6:40
  any
• 6:40
  given
• 6:40
  interaction,
• 6:41
  and
• 6:42
  almost
• 6:42
  a
• 6:42
  sociopathic
• 6:43
  lack
• 6:43
  of
• 6:43
  concern
• 6:43
  for
• 6:44
  the
• 6:44
  consequences
• 6:44
  that
• 6:44
  may
• 6:44
  come
• 6:45
  from
• 6:45
  deceiving
• 6:45
  someone."
• 6:47
  The
• 6:47
  article
• 6:47
  traces
• 6:47
  the
• 6:47
  erosion
• 6:48
  of
• 6:48
  OpenAI's
• 6:48
  safety
• 6:49
  commitments
• 6:49
  in
• 6:49
  granular
• 6:49
  detail.
• 6:51
  The
• 6:51
  superalignment
• 6:51
  team,
• 6:52
  promised
• 6:52
  20%
• 6:53
  of
• 6:53
  compute,
• 6:53
  reportedly
• 6:54
  received
• 6:54
  1-2%
• 6:55
  —
• 6:56
  "most
• 6:56
  of
• 6:56
  it
• 6:56
  on
• 6:56
  the
• 6:57
  oldest
• 6:57
  cluster
• 6:57
  with
• 6:57
  the
• 6:57
  worst
• 6:58
  chips."
• 6:59
  The
• 6:59
  WilmerHale
• 6:59
  investigation
• 7:00
  into
• 7:00
  Altman's
• 7:01
  firing
• 7:01
  produced
• 7:01
  no
• 7:01
  written
• 7:02
  report;
• 7:02
  findings
• 7:03
  were
• 7:03
  limited
• 7:03
  to
• 7:03
  oral
• 7:04
  briefings
• 7:04
  shared
• 7:04
  with
• 7:04
  two
• 7:05
  board
• 7:05
  members,
• 7:06
  and
• 7:06
  some
• 7:06
  board
• 7:06
  members
• 7:06
  say
• 7:06
  they
• 7:07
  never
• 7:07
  received
• 7:07
  even
• 7:07
  those.
• 7:09
  Multiple
• 7:09
  administration
• 7:10
  officials
• 7:10
  expressed
• 7:11
  discomfort
• 7:11
  with
• 7:11
  Altman's
• 7:12
  geopolitical
• 7:12
  ambitions
• 7:13
  —
• 7:14
  a
• 7:14
  plan
• 7:14
  to
• 7:14
  build
• 7:14
  chip
• 7:15
  foundries
• 7:15
  in
• 7:15
  the
• 7:15
  UAE,
• 7:16
  a
• 7:16
  data-center
• 7:17
  campus
• 7:17
  in
• 7:17
  Abu
• 7:17
  Dhabi
• 7:18
  seven
• 7:18
  times
• 7:18
  larger
• 7:19
  than
• 7:19
  Central
• 7:19
  Park,
• 7:20
  and
• 7:20
  a
• 7:20
  $500
• 7:21
  billion
• 7:21
  Stargate
• 7:22
  joint
• 7:22
  venture
• 7:22
  announced
• 7:23
  alongside
• 7:23
  Trump.
• 7:24
  A
• 7:25
  former
• 7:25
  OpenAI
• 7:25
  executive
• 7:26
  calls
• 7:26
  the
• 7:26
  UAE
• 7:27
  data
• 7:27
  center
• 7:27
  expansion
• 7:28
  "the
• 7:28
  most
• 7:29
  reckless
• 7:29
  thing
• 7:29
  that
• 7:29
  has
• 7:29
  been
• 7:29
  done."
• 7:30
  Whether
• 7:31
  one
• 7:31
  views
• 7:31
  Altman
• 7:32
  as
• 7:32
  a
• 7:32
  visionary
• 7:32
  navigating
• 7:33
  impossible
• 7:33
  constraints
• 7:34
  or
• 7:34
  a
• 7:34
  pattern-matching
• 7:35
  dealmaker
• 7:36
  who
• 7:36
  tells
• 7:36
  each
• 7:36
  audience
• 7:37
  what
• 7:37
  they
• 7:37
  want
• 7:37
  to
• 7:37
  hear,
• 7:38
  the
• 7:38
  structural
• 7:38
  question
• 7:39
  remains:
• 7:40
  the
• 7:40
  person
• 7:40
  who
• 7:40
  may
• 7:40
  wield
• 7:41
  the
• 7:41
  most
• 7:41
  consequential
• 7:41
  technology
• 7:42
  in
• 7:42
  human
• 7:42
  history
• 7:43
  has
• 7:43
  systematically
• 7:44
  dismantled
• 7:45
  every
• 7:45
  governance
• 7:45
  mechanism
• 7:46
  designed
• 7:46
  to
• 7:46
  constrain
• 7:47
  him.
• 7:49
  A
• 7:49
  GitHub
• 7:49
  issue
• 7:50
  on
• 7:50
  the
• 7:50
  Claude
• 7:50
  Code
• 7:51
  repository
• 7:51
  has
• 7:52
  become
• 7:52
  the
• 7:52
  most
• 7:52
  data-driven
• 7:53
  critique
• 7:53
  of
• 7:53
  agentic
• 7:54
  coding
• 7:54
  regression
• 7:54
  to
• 7:54
  date.
• 7:56
  A
• 7:56
  user
• 7:56
  analyzing
• 7:57
  17,871
• 7:59
  thinking
• 7:59
  blocks
• 7:59
  and
• 7:59
  234,760
• 8:02
  tool
• 8:02
  calls
• 8:03
  across
• 8:03
  6,852
• 8:04
  session
• 8:05
  files
• 8:05
  documented
• 8:06
  a
• 8:06
  measurable
• 8:07
  quality
• 8:07
  collapse
• 8:08
  correlating
• 8:08
  with
• 8:08
  thinking
• 8:08
  content
• 8:09
  redaction
• 8:09
  starting
• 8:10
  in
• 8:10
  early
• 8:10
  March.
• 8:11
  The
• 8:11
  read-to-edit
• 8:12
  ratio
• 8:12
  —
• 8:13
  how
• 8:13
  many
• 8:13
  files
• 8:14
  the
• 8:14
  model
• 8:14
  reads
• 8:14
  before
• 8:14
  making
• 8:15
  changes
• 8:15
  —
• 8:16
  dropped
• 8:16
  from
• 8:16
  6.6
• 8:17
  in
• 8:17
  January
• 8:18
  to
• 8:18
  2.0
• 8:19
  by
• 8:19
  late
• 8:19
  March,
• 8:20
  a
• 8:20
  70%
• 8:20
  reduction
• 8:21
  in
• 8:21
  research
• 8:21
  before
• 8:22
  mutation.
• 8:23
  Full-file
• 8:23
  rewrites
• 8:24
  doubled.
• 8:25
  A
• 8:25
  programmatic
• 8:26
  "stop
• 8:26
  hook"
• 8:27
  catching
• 8:27
  premature
• 8:28
  stopping,
• 8:28
  responsibility
• 8:29
  dodging,
• 8:30
  and
• 8:30
  permission-seeking
• 8:31
  behavior
• 8:31
  fired
• 8:31
  173
• 8:32
  times
• 8:33
  in
• 8:33
  17
• 8:33
  days
• 8:33
  after
• 8:34
  March
• 8:34
  8;
• 8:34
  it
• 8:35
  fired
• 8:35
  zero
• 8:35
  times
• 8:35
  before.
• 8:36
  The
• 8:37
  user's
• 8:37
  positive-to-negative
• 8:38
  sentiment
• 8:38
  ratio
• 8:39
  in
• 8:39
  prompts
• 8:39
  collapsed
• 8:39
  from
• 8:40
  4.4:1
• 8:41
  to
• 8:41
  3.0:1.
• 8:42
  Most
• 8:43
  strikingly,
• 8:43
  the
• 8:43
  human
• 8:44
  sent
• 8:44
  roughly
• 8:44
  the
• 8:44
  same
• 8:44
  number
• 8:45
  of
• 8:45
  prompts
• 8:45
  in
• 8:45
  February
• 8:46
  (5,608)
• 8:47
  and
• 8:47
  March
• 8:47
  (5,701),
• 8:49
  but
• 8:49
  the
• 8:49
  model
• 8:50
  consumed
• 8:50
  orders
• 8:51
  of
• 8:51
  magnitude
• 8:51
  more
• 8:51
  tokens
• 8:52
  to
• 8:52
  produce
• 8:52
  demonstrably
• 8:53
  worse
• 8:53
  results.
• 8:54
  The
• 8:54
  user
• 8:54
  —
• 8:55
  running
• 8:55
  50+
• 8:56
  concurrent
• 8:56
  agent
• 8:57
  sessions
• 8:57
  doing
• 8:57
  systems
• 8:58
  programming
• 8:58
  —
• 8:59
  was
• 8:59
  forced
• 8:59
  to
• 8:59
  abandon
• 9:00
  the
• 9:00
  multi-agent
• 9:00
  workflow
• 9:01
  entirely
• 9:01
  and
• 9:02
  retreat
• 9:02
  to
• 9:02
  single-session
• 9:03
  supervised
• 9:03
  operation.
• 9:04
  The
• 9:05
  counterpoint
• 9:05
  comes
• 9:05
  from
• 9:05
  a
• 9:05
  developer
• 9:06
  who
• 9:06
  spent
• 9:06
  three
• 9:06
  months
• 9:07
  building
• 9:07
  SyntaqLite,
• 9:08
  a
• 9:08
  comprehensive
• 9:09
  SQLite
• 9:09
  devtools
• 9:10
  suite
• 9:10
  he'd
• 9:10
  wanted
• 9:10
  for
• 9:11
  eight
• 9:11
  years.
• 9:12
  His
• 9:12
  250-hour
• 9:13
  retrospective
• 9:14
  is
• 9:14
  the
• 9:14
  most
• 9:15
  honest
• 9:15
  post-mortem
• 9:16
  of
• 9:16
  AI-assisted
• 9:17
  development
• 9:17
  published
• 9:18
  this
• 9:18
  year.
• 9:19
  The
• 9:19
  first
• 9:19
  month
• 9:20
  was
• 9:20
  pure
• 9:20
  vibe-coding:
• 9:21
  Claude
• 9:22
  acted
• 9:22
  as
• 9:22
  designer
• 9:23
  and
• 9:23
  implementer
• 9:23
  while
• 9:24
  the
• 9:24
  developer
• 9:24
  managed.
• 9:25
  The
• 9:25
  result
• 9:26
  was
• 9:26
  a
• 9:26
  functional
• 9:26
  prototype
• 9:27
  with
• 9:27
  500+
• 9:27
  tests
• 9:28
  —
• 9:29
  and
• 9:29
  complete
• 9:29
  spaghetti
• 9:30
  code
• 9:30
  he
• 9:30
  couldn't
• 9:30
  reason
• 9:31
  about,
• 9:31
  forcing
• 9:32
  a
• 9:32
  total
• 9:32
  rewrite.
• 9:33
  The
• 9:33
  second
• 9:34
  phase
• 9:34
  worked:
• 9:35
  he
• 9:35
  took
• 9:35
  ownership
• 9:35
  of
• 9:35
  all
• 9:36
  design
• 9:36
  decisions,
• 9:37
  used
• 9:37
  AI
• 9:38
  as
• 9:38
  "autocomplete
• 9:39
  on
• 9:39
  steroids"
• 9:39
  within
• 9:40
  a
• 9:40
  tight
• 9:40
  review
• 9:40
  loop,
• 9:41
  and
• 9:41
  shipped
• 9:41
  a
• 9:41
  parser,
• 9:42
  formatter,
• 9:43
  VS
• 9:44
  Code
• 9:44
  extension,
• 9:45
  Python
• 9:45
  bindings,
• 9:46
  and
• 9:46
  WASM
• 9:46
  playground.
• 9:47
  His
• 9:48
  taxonomy
• 9:48
  is
• 9:48
  precise:
• 9:49
  when
• 9:50
  you
• 9:50
  understand
• 9:50
  the
• 9:50
  problem
• 9:50
  deeply,
• 9:51
  AI
• 9:52
  is
• 9:52
  excellent;
• 9:53
  when
• 9:53
  you
• 9:53
  can
• 9:53
  describe
• 9:53
  what
• 9:54
  you
• 9:54
  want
• 9:54
  but
• 9:54
  don't
• 9:55
  know
• 9:55
  the
• 9:55
  domain,
• 9:56
  AI
• 9:56
  is
• 9:56
  good
• 9:57
  but
• 9:57
  requires
• 9:57
  care;
• 9:58
  when
• 9:58
  you
• 9:58
  don't
• 9:59
  even
• 9:59
  know
• 9:59
  what
• 9:59
  you
• 9:59
  want,
• 10:00
  AI
• 10:00
  is
• 10:00
  "somewhere
• 10:01
  between
• 10:01
  unhelpful
• 10:02
  and
• 10:02
  harmful."
• 10:02
  The
• 10:03
  slot-machine
• 10:03
  comparison
• 10:04
  recurs
• 10:04
  —
• 10:05
  the
• 10:05
  sunk-cost
• 10:06
  fallacy,
• 10:06
  the
• 10:06
  tiredness
• 10:07
  feedback
• 10:07
  loop,
• 10:08
  the
• 10:08
  "just
• 10:08
  one
• 10:08
  more
• 10:09
  prompt"
• 10:09
  compulsion.
• 10:10
  That
• 10:10
  compulsion
• 10:11
  is
• 10:11
  now
• 10:11
  a
• 10:11
  recognized
• 10:12
  phenomenon.
• 10:13
  A
• 10:13
  LeadDev
• 10:13
  article
• 10:14
  catalogs
• 10:14
  developers
• 10:15
  losing
• 10:15
  sleep
• 10:15
  to
• 10:16
  agentic
• 10:16
  coding
• 10:16
  —
• 10:17
  not
• 10:17
  from
• 10:17
  deadline
• 10:18
  pressure
• 10:18
  but
• 10:18
  from
• 10:19
  the
• 10:19
  dopamine
• 10:19
  loop
• 10:19
  of
• 10:19
  rapid
• 10:20
  iteration.
• 10:21
  Reddit
• 10:21
  comments
• 10:22
  paint
• 10:22
  a
• 10:22
  vivid
• 10:22
  picture:
• 10:23
  "I've
• 10:23
  done
• 10:23
  7
• 10:24
  all-nighters
• 10:24
  the
• 10:24
  last
• 10:25
  2
• 10:25
  months,"
• 10:25
  one
• 10:26
  writes.
• 10:26
  Another
• 10:27
  describes
• 10:27
  ending
• 10:28
  up
• 10:28
  in
• 10:28
  therapy
• 10:28
  after
• 10:28
  a
• 10:29
  crash.
• 10:30
  The
• 10:30
  comparison
• 10:30
  to
• 10:30
  Civilization's
• 10:31
  "one
• 10:31
  more
• 10:31
  turn"
• 10:32
  is
• 10:32
  apt,
• 10:33
  but
• 10:33
  the
• 10:33
  stakes
• 10:33
  are
• 10:33
  higher:
• 10:34
  context
• 10:35
  splits
• 10:35
  between
• 10:35
  human
• 10:36
  and
• 10:36
  agent
• 10:36
  mean
• 10:37
  walking
• 10:37
  away
• 10:37
  risks
• 10:38
  losing
• 10:38
  the
• 10:38
  thread
• 10:39
  entirely,
• 10:40
  creating
• 10:40
  a
• 10:40
  perverse
• 10:41
  incentive
• 10:41
  to
• 10:41
  keep
• 10:42
  going
• 10:42
  until
• 10:42
  5
• 10:42
  AM.
• 10:45
  AutoAgent
• 10:46
  takes
• 10:46
  the
• 10:46
  concept
• 10:47
  of
• 10:47
  agent
• 10:47
  engineering
• 10:48
  and
• 10:48
  turns
• 10:48
  it
• 10:48
  recursive:
• 10:49
  you
• 10:49
  write
• 10:49
  a
• 10:49
  program.md
• 10:50
  file
• 10:51
  that
• 10:51
  instructs
• 10:51
  a
• 10:51
  meta-agent,
• 10:52
  which
• 10:52
  then
• 10:53
  modifies
• 10:53
  the
• 10:53
  actual
• 10:54
  agent
• 10:54
  harness
• 10:54
  (agent.py
• 10:55
  —
• 10:56
  tools,
• 10:56
  prompts,
• 10:57
  orchestration,
• 10:57
  routing),
• 10:58
  runs
• 10:58
  benchmarks,
• 10:58
  checks
• 10:59
  scores,
• 10:59
  and
• 10:59
  iterates
• 11:00
  autonomously
• 11:00
  overnight.
• 11:01
  It's
• 11:01
  hill-climbing
• 11:02
  on
• 11:02
  benchmark
• 11:03
  scores
• 11:03
  with
• 11:03
  Docker
• 11:03
  isolation,
• 11:04
  using
• 11:04
  the
• 11:05
  same
• 11:05
  evaluate-keep-or-discard
• 11:06
  loop
• 11:06
  as
• 11:06
  autoresearch.
• 11:07
  The
• 11:07
  harness
• 11:08
  is
• 11:08
  a
• 11:08
  single
• 11:08
  file
• 11:09
  with
• 11:09
  a
• 11:09
  fixed
• 11:09
  Harbor
• 11:09
  adapter
• 11:10
  boundary
• 11:10
  and
• 11:10
  an
• 11:10
  editable
• 11:11
  everything-else
• 11:11
  surface.
• 11:13
  The
• 11:13
  key
• 11:13
  design
• 11:13
  choice
• 11:14
  is
• 11:14
  that
• 11:14
  humans
• 11:14
  program
• 11:15
  the
• 11:15
  meta-agent's
• 11:15
  instructions,
• 11:17
  not
• 11:17
  the
• 11:17
  harness
• 11:18
  directly
• 11:18
  —
• 11:20
  a
• 11:20
  clean
• 11:20
  separation
• 11:21
  that
• 11:21
  keeps
• 11:21
  the
• 11:21
  optimization
• 11:22
  loop
• 11:22
  stable
• 11:22
  while
• 11:22
  the
• 11:22
  agent
• 11:23
  explores
• 11:23
  the
• 11:23
  design
• 11:24
  space.
• 11:25
  On
• 11:25
  the
• 11:25
  desktop
• 11:25
  automation
• 11:26
  front,
• 11:26
  usecomputer
• 11:27
  is
• 11:27
  a
• 11:27
  cross-platform
• 11:28
  CLI
• 11:29
  for
• 11:29
  AI
• 11:29
  agents
• 11:29
  built
• 11:30
  on
• 11:30
  a
• 11:30
  native
• 11:30
  Zig
• 11:30
  binary
• 11:31
  —
• 11:32
  screenshot,
• 11:32
  mouse
• 11:33
  control,
• 11:34
  keyboard
• 11:34
  synthesis,
• 11:35
  all
• 11:35
  available
• 11:35
  as
• 11:35
  commands
• 11:36
  with
• 11:36
  no
• 11:36
  Node.js
• 11:38
  runtime
• 11:38
  required.
• 11:39
  It
• 11:39
  supports
• 11:39
  the
• 11:39
  screenshot-act-screenshot
• 11:41
  feedback
• 11:41
  loop
• 11:41
  with
• 11:41
  coordinate
• 11:42
  mapping,
• 11:43
  works
• 11:43
  on
• 11:43
  macOS
• 11:43
  (accessibility
• 11:44
  API),
• 11:45
  Linux
• 11:45
  (X11/XWayland),
• 11:50
  and
• 11:50
  Windows,
• 11:50
  and
• 11:50
  ships
• 11:51
  with
• 11:51
  an
• 11:51
  agent
• 11:52
  skill
• 11:52
  that
• 11:52
  teaches
• 11:52
  Claude
• 11:53
  Code
• 11:53
  or
• 11:53
  OpenCode
• 11:54
  the
• 11:54
  correct
• 11:54
  workflow.
• 11:55
  The
• 11:55
  library
• 11:56
  API
• 11:56
  mirrors
• 11:56
  the
• 11:57
  native
• 11:57
  command
• 11:57
  shapes,
• 11:58
  making
• 11:58
  it
• 11:58
  trivially
• 11:58
  integrable
• 11:59
  as
• 11:59
  an
• 11:59
  OpenAI
• 12:00
  computer
• 12:00
  tool
• 12:00
  or
• 12:00
  Anthropic
• 12:01
  computer-use
• 12:01
  backend.
• 12:03
  LightRAG
• 12:03
  continues
• 12:04
  to
• 12:04
  prove
• 12:04
  that
• 12:04
  graph
• 12:05
  RAG
• 12:05
  doesn't
• 12:05
  require
• 12:05
  Microsoft-scale
• 12:06
  budgets.
• 12:07
  A
• 12:07
  tutorial
• 12:08
  walkthrough
• 12:08
  demonstrates
• 12:09
  the
• 12:09
  full
• 12:09
  pipeline:
• 12:10
  Docker-compose
• 12:11
  setup
• 12:11
  via
• 12:12
  Claude
• 12:12
  Code,
• 12:13
  OpenAI
• 12:13
  embeddings
• 12:14
  for
• 12:14
  knowledge-graph
• 12:14
  construction,
• 12:15
  and
• 12:15
  Claude
• 12:16
  Code
• 12:16
  skills
• 12:16
  wrapping
• 12:17
  LightRAG's
• 12:17
  API
• 12:17
  endpoints
• 12:18
  for
• 12:18
  seamless
• 12:18
  query/upload/explore
• 12:20
  from
• 12:20
  the
• 12:20
  terminal.
• 12:21
  The
• 12:21
  key
• 12:22
  selling
• 12:22
  point
• 12:22
  remains
• 12:23
  cost
• 12:23
  efficiency
• 12:23
  —
• 12:24
  a
• 12:24
  study
• 12:24
  cited
• 12:25
  in
• 12:25
  the
• 12:25
  video
• 12:25
  showed
• 12:25
  RAG
• 12:26
  being
• 12:26
  1,250x
• 12:27
  cheaper
• 12:28
  than
• 12:28
  naive
• 12:28
  LLM
• 12:29
  context
• 12:29
  stuffing
• 12:30
  for
• 12:30
  large
• 12:30
  document
• 12:30
  corpora,
• 12:31
  though
• 12:31
  that
• 12:32
  figure
• 12:32
  is
• 12:32
  from
• 12:32
  July
• 12:33
  2025
• 12:34
  and
• 12:34
  the
• 12:34
  gap
• 12:34
  has
• 12:34
  likely
• 12:35
  narrowed.
• 12:36
  The
• 12:36
  practical
• 12:36
  threshold
• 12:37
  for
• 12:37
  when
• 12:37
  RAG
• 12:37
  pays
• 12:37
  off:
• 12:38
  somewhere
• 12:38
  around
• 12:39
  500-2,000
• 12:43
  text
• 12:43
  pages,
• 12:44
  where
• 12:44
  agent
• 12:44
  search
• 12:45
  starts
• 12:45
  losing
• 12:45
  coherence
• 12:46
  and
• 12:46
  cost
• 12:46
  explodes.
• 12:47
  Career-Ops
• 12:48
  takes
• 12:48
  a
• 12:48
  different
• 12:49
  angle
• 12:49
  entirely:
• 12:50
  an
• 12:50
  AI-powered
• 12:51
  job
• 12:51
  search
• 12:51
  pipeline
• 12:52
  built
• 12:52
  on
• 12:52
  Claude
• 12:52
  Code
• 12:53
  that
• 12:53
  evaluates
• 12:53
  offers
• 12:54
  with
• 12:54
  structured
• 12:54
  A-F
• 12:54
  scoring,
• 12:55
  generates
• 12:56
  ATS-optimized
• 12:57
  CVs
• 12:57
  per
• 12:57
  job
• 12:58
  description,
• 12:59
  and
• 12:59
  scans
• 12:59
  45+
• 13:00
  company
• 13:00
  career
• 13:01
  portals
• 13:01
  via
• 13:01
  Playwright.
• 13:02
  The
• 13:02
  author
• 13:03
  used
• 13:03
  it
• 13:03
  to
• 13:03
  evaluate
• 13:04
  740+
• 13:05
  job
• 13:05
  offers,
• 13:06
  generate
• 13:06
  100+
• 13:07
  tailored
• 13:07
  CVs,
• 13:08
  and
• 13:08
  land
• 13:08
  a
• 13:08
  Head
• 13:08
  of
• 13:08
  Applied
• 13:09
  AI
• 13:09
  role.
• 13:10
  The
• 13:10
  system
• 13:10
  is
• 13:10
  explicitly
• 13:11
  not
• 13:11
  a
• 13:11
  spray-and-pray
• 13:12
  tool
• 13:12
  —
• 13:13
  it
• 13:13
  filters
• 13:13
  ruthlessly,
• 13:14
  recommending
• 13:15
  against
• 13:15
  applying
• 13:15
  to
• 13:15
  anything
• 13:16
  below
• 13:16
  4.0/5.
• 13:18
  The
• 13:18
  design
• 13:18
  philosophy
• 13:18
  mirrors
• 13:19
  good
• 13:19
  agent
• 13:19
  architecture:
• 13:20
  human-in-the-loop
• 13:21
  for
• 13:21
  all
• 13:21
  decisions,
• 13:22
  accumulating
• 13:22
  context
• 13:23
  over
• 13:23
  time
• 13:23
  (interview
• 13:24
  story
• 13:24
  banks,
• 13:24
  negotiation
• 13:25
  scripts),
• 13:26
  and
• 13:26
  self-customization
• 13:27
  where
• 13:27
  Claude
• 13:27
  reads
• 13:27
  and
• 13:28
  modifies
• 13:28
  its
• 13:28
  own
• 13:28
  configuration
• 13:29
  files.
• 13:34
  A
• 13:34
  paper
• 13:34
  from
• 13:34
  Politecnico
• 13:35
  di
• 13:35
  Torino
• 13:36
  and
• 13:36
  Huawei
• 13:36
  tackles
• 13:37
  what
• 13:37
  may
• 13:37
  be
• 13:37
  the
• 13:37
  most
• 13:37
  persistent
• 13:38
  failure
• 13:38
  mode
• 13:39
  in
• 13:39
  cybersecurity
• 13:39
  ML:
• 13:40
  shortcut
• 13:40
  learning.
• 13:41
  Models
• 13:41
  that
• 13:41
  achieve
• 13:42
  92%+
• 13:43
  accuracy
• 13:43
  on
• 13:43
  random
• 13:44
  splits
• 13:44
  crater
• 13:44
  by
• 13:45
  nearly
• 13:45
  30%
• 13:46
  on
• 13:46
  temporal
• 13:46
  splits
• 13:46
  —
• 13:47
  where
• 13:47
  they
• 13:47
  must
• 13:47
  classify
• 13:48
  attacks
• 13:48
  they've
• 13:49
  never
• 13:49
  seen
• 13:49
  variants
• 13:49
  of.
• 13:50
  The
• 13:50
  authors
• 13:50
  propose
• 13:51
  SALM
• 13:51
  (Semantically
• 13:52
  Aligned
• 13:52
  Language
• 13:52
  Models),
• 13:53
  a
• 13:53
  two-stage
• 13:53
  contrastive
• 13:54
  learning
• 13:54
  framework.
• 13:55
  Stage
• 13:55
  1
• 13:55
  restructures
• 13:56
  an
• 13:56
  LLM
• 13:57
  embedding
• 13:57
  space
• 13:57
  around
• 13:57
  15
• 13:58
  vulnerability
• 13:59
  types
• 13:59
  using
• 13:59
  contrastive
• 14:00
  learning
• 14:00
  on
• 14:00
  textual
• 14:01
  CVE
• 14:01
  descriptions.
• 14:02
  Stage
• 14:03
  2
• 14:03
  freezes
• 14:03
  the
• 14:03
  text
• 14:03
  encoder
• 14:04
  and
• 14:04
  trains
• 14:04
  a
• 14:04
  payload
• 14:05
  encoder
• 14:05
  to
• 14:05
  align
• 14:06
  raw
• 14:06
  HTTP
• 14:07
  payloads
• 14:07
  to
• 14:07
  that
• 14:07
  same
• 14:08
  space
• 14:08
  via
• 14:08
  cross-modal
• 14:09
  alignment.
• 14:10
  At
• 14:10
  inference,
• 14:11
  classification
• 14:11
  is
• 14:12
  pure
• 14:12
  semantic
• 14:12
  retrieval:
• 14:13
  embed
• 14:13
  the
• 14:13
  payload,
• 14:14
  compare
• 14:14
  cosine
• 14:15
  distance
• 14:15
  to
• 14:15
  textual
• 14:16
  prototypes
• 14:16
  like
• 14:17
  "SQL
• 14:17
  injection
• 14:17
  attack"
• 14:18
  —
• 14:18
  prototypes
• 14:19
  the
• 14:19
  model
• 14:19
  never
• 14:19
  saw
• 14:19
  during
• 14:20
  training,
• 14:20
  forcing
• 14:21
  genuine
• 14:21
  generalization
• 14:22
  rather
• 14:22
  than
• 14:22
  memorized
• 14:23
  patterns.
• 14:24
  SALM
• 14:24
  reaches
• 14:24
  68.1%
• 14:26
  accuracy
• 14:27
  on
• 14:27
  temporally
• 14:27
  shifted
• 14:28
  private
• 14:28
  data
• 14:29
  versus
• 14:29
  62.3%
• 14:30
  for
• 14:31
  fine-tuned
• 14:31
  CodeBERT,
• 14:32
  with
• 14:32
  the
• 14:32
  improvement
• 14:33
  concentrated
• 14:33
  on
• 14:34
  mid-frequency
• 14:34
  classes
• 14:35
  where
• 14:35
  lexical
• 14:35
  cues
• 14:36
  are
• 14:36
  unreliable
• 14:37
  but
• 14:37
  textual
• 14:37
  descriptions
• 14:38
  provide
• 14:38
  clear
• 14:38
  semantic
• 14:39
  grounding.
• 14:40
  The
• 14:40
  approach
• 14:40
  also
• 14:41
  opens
• 14:41
  a
• 14:41
  zero-shot
• 14:42
  pathway:
• 14:43
  add
• 14:43
  a
• 14:43
  new
• 14:43
  vulnerability
• 14:44
  type
• 14:44
  by
• 14:44
  specifying
• 14:45
  a
• 14:45
  text
• 14:45
  description,
• 14:46
  no
• 14:46
  retraining
• 14:47
  required.
• 14:48
  The
• 14:48
  gap
• 14:48
  between
• 14:49
  68%
• 14:49
  accuracy
• 14:50
  and
• 14:50
  production-grade
• 14:51
  reliability
• 14:51
  is
• 14:51
  wide,
• 14:52
  but
• 14:52
  the
• 14:52
  direction
• 14:53
  —
• 14:53
  transferring
• 14:54
  knowledge
• 14:54
  from
• 14:55
  data-rich
• 14:55
  text
• 14:55
  modalities
• 14:56
  to
• 14:56
  data-scarce
• 14:57
  payload
• 14:57
  modalities
• 14:57
  —
• 14:58
  addresses
• 14:59
  the
• 14:59
  fundamental
• 14:59
  asymmetry
• 15:00
  that
• 15:00
  makes
• 15:00
  cybersecurity
• 15:01
  ML
• 15:01
  so
• 15:01
  fragile.
• 15:03
  On
• 15:03
  the
• 15:03
  practical
• 15:03
  tooling
• 15:04
  side,
• 15:04
  Ice
• 15:04
  Tea
• 15:05
  is
• 15:05
  a
• 15:05
  Go-based
• 15:05
  SAST
• 15:06
  scanner
• 15:06
  combining
• 15:07
  Tree-sitter
• 15:07
  AST
• 15:08
  pattern
• 15:08
  matching
• 15:08
  with
• 15:09
  optional
• 15:09
  LLM
• 15:09
  reasoning
• 15:10
  for
• 15:10
  false-positive
• 15:11
  filtering.
• 15:12
  It
• 15:12
  ships
• 15:12
  with
• 15:12
  82
• 15:13
  built-in
• 15:13
  SKILLs
• 15:14
  covering
• 15:14
  456+
• 15:16
  detection
• 15:16
  rules
• 15:17
  across
• 15:17
  12
• 15:17
  security
• 15:18
  domains
• 15:18
  (auth,
• 15:19
  injection,
• 15:20
  web/API,
• 15:21
  crypto,
• 15:22
  filesystem,
• 15:23
  infrastructure,
• 15:24
  logging,
• 15:24
  memory
• 15:25
  safety,
• 15:25
  cloud,
• 15:26
  Android,
• 15:27
  network),
• 15:27
  supports
• 15:28
  10
• 15:28
  languages,
• 15:28
  and
• 15:29
  outputs
• 15:29
  SARIF
• 15:29
  2.1.0
• 15:31
  for
• 15:31
  GitHub
• 15:32
  integration.
• 15:33
  The
• 15:33
  extensibility
• 15:34
  model
• 15:34
  is
• 15:34
  notable:
• 15:35
  add
• 15:35
  detection
• 15:36
  rules
• 15:36
  via
• 15:36
  Markdown
• 15:37
  +
• 15:37
  YAML
• 15:37
  files
• 15:38
  with
• 15:38
  no
• 15:38
  Go
• 15:38
  code
• 15:39
  required.
• 15:40
  Its
• 15:40
  three-engine
• 15:41
  architecture
• 15:41
  —
• 15:42
  pattern
• 15:42
  matching,
• 15:43
  taint
• 15:43
  tracking,
• 15:44
  LLM
• 15:44
  verification
• 15:45
  —
• 15:46
  represents
• 15:46
  the
• 15:46
  emerging
• 15:47
  consensus
• 15:47
  that
• 15:48
  pure
• 15:48
  static
• 15:48
  analysis
• 15:49
  and
• 15:49
  pure
• 15:49
  AI
• 15:50
  scanning
• 15:50
  both
• 15:50
  fail,
• 15:51
  and
• 15:51
  the
• 15:51
  hybrid
• 15:51
  approach
• 15:52
  is
• 15:52
  the
• 15:52
  way
• 15:52
  forward.
• 15:57
  Filippo
• 15:57
  Valsorda,
• 15:58
  one
• 15:58
  of
• 15:58
  the
• 15:58
  Go
• 15:59
  standard
• 15:59
  library's
• 16:00
  cryptography
• 16:00
  maintainers,
• 16:01
  has
• 16:01
  publicly
• 16:02
  shifted
• 16:02
  his
• 16:02
  position
• 16:02
  on
• 16:03
  post-quantum
• 16:03
  migration
• 16:04
  urgency
• 16:04
  —
• 16:05
  and
• 16:05
  the
• 16:05
  shift
• 16:05
  is
• 16:05
  dramatic.
• 16:06
  Two
• 16:06
  recent
• 16:07
  papers
• 16:07
  catalyzed
• 16:07
  the
• 16:08
  change:
• 16:09
  Google
• 16:09
  revised
• 16:09
  down
• 16:10
  the
• 16:10
  estimated
• 16:10
  logical
• 16:10
  qubits
• 16:11
  needed
• 16:11
  to
• 16:11
  break
• 16:11
  256-bit
• 16:13
  elliptic
• 16:13
  curves
• 16:14
  (P-256,
• 16:15
  secp256k1)
• 16:17
  enough
• 16:17
  to
• 16:17
  make
• 16:17
  attacks
• 16:18
  feasible
• 16:18
  in
• 16:18
  minutes
• 16:18
  on
• 16:19
  fast-clock
• 16:19
  architectures,
• 16:20
  and
• 16:20
  Oratomic
• 16:20
  showed
• 16:21
  it
• 16:21
  could
• 16:21
  be
• 16:21
  done
• 16:21
  with
• 16:21
  as
• 16:21
  few
• 16:21
  as
• 16:22
  10,000
• 16:22
  physical
• 16:22
  qubits
• 16:23
  given
• 16:23
  non-local
• 16:24
  connectivity.
• 16:25
  Google's
• 16:25
  Heather
• 16:25
  Adkins
• 16:26
  and
• 16:26
  Sophie
• 16:26
  Schmieg
• 16:26
  have
• 16:27
  set
• 16:27
  2029
• 16:27
  as
• 16:28
  their
• 16:28
  deadline.
• 16:29
  Scott
• 16:29
  Aaronson's
• 16:29
  "clearest
• 16:30
  warning"
• 16:30
  draws
• 16:30
  an
• 16:30
  analogy
• 16:31
  to
• 16:31
  how
• 16:31
  nuclear
• 16:31
  fission
• 16:32
  research
• 16:32
  stopped
• 16:32
  being
• 16:32
  public
• 16:33
  between
• 16:33
  1939
• 16:34
  and
• 16:34
  1940.
• 16:35
  Valsorda's
• 16:35
  framing
• 16:36
  is
• 16:36
  blunt:
• 16:36
  "the
• 16:37
  bet
• 16:37
  is
• 16:37
  not
• 16:37
  'are
• 16:37
  you
• 16:37
  100%
• 16:38
  sure
• 16:38
  a
• 16:38
  CRQC
• 16:38
  will
• 16:39
  exist
• 16:39
  in
• 16:39
  2030?'
• 16:39
  —
• 16:40
  the
• 16:40
  bet
• 16:40
  is
• 16:41
  'are
• 16:41
  you
• 16:41
  100%
• 16:41
  sure
• 16:41
  a
• 16:41
  CRQC
• 16:42
  will
• 16:42
  NOT
• 16:43
  exist
• 16:43
  in
• 16:43
  2030?'"
• 16:44
  The
• 16:44
  practical
• 16:45
  implications
• 16:45
  are
• 16:45
  severe.
• 16:46
  Any
• 16:47
  TLS
• 16:47
  connection
• 16:47
  without
• 16:48
  ML-KEM
• 16:49
  key
• 16:50
  exchange
• 16:50
  should
• 16:50
  now
• 16:50
  be
• 16:50
  considered
• 16:51
  a
• 16:51
  potential
• 16:51
  active
• 16:52
  compromise.
• 16:52
  Hybrid
• 16:53
  classic
• 16:53
  +
• 16:53
  post-quantum
• 16:54
  authentication
• 16:55
  "makes
• 16:55
  no
• 16:55
  sense
• 16:55
  anymore"
• 16:56
  and
• 16:56
  will
• 16:56
  only
• 16:56
  slow
• 16:57
  deployment
• 16:57
  —
• 16:57
  go
• 16:58
  straight
• 16:58
  to
• 16:58
  pure
• 16:58
  ML-DSA-44.
• 17:00
  New
• 17:01
  non-post-quantum
• 17:01
  schemes
• 17:02
  should
• 17:02
  simply
• 17:02
  not
• 17:02
  be
• 17:02
  deployed,
• 17:03
  period.
• 17:04
  Hardware
• 17:05
  attestation
• 17:05
  systems
• 17:05
  (Intel
• 17:06
  SGX,
• 17:07
  AMD
• 17:07
  SEV-SNP)
• 17:10
  are
• 17:10
  "just
• 17:10
  fd"
• 17:10
  —
• 17:11
  their
• 17:12
  root
• 17:12
  keys
• 17:12
  are
• 17:12
  not
• 17:12
  PQ,
• 17:13
  and
• 17:13
  there's
• 17:13
  no
• 17:13
  progress
• 17:14
  on
• 17:14
  rolling
• 17:14
  out
• 17:14
  PQ
• 17:15
  replacements
• 17:15
  at
• 17:15
  hardware
• 17:15
  speeds.
• 17:17
  File
• 17:17
  encryption
• 17:17
  is
• 17:17
  especially
• 17:18
  vulnerable
• 17:18
  to
• 17:18
  store-now-decrypt-later.
• 17:20
  Symmetric
• 17:21
  cryptography
• 17:21
  gets
• 17:21
  a
• 17:21
  pass:
• 17:22
  128-bit
• 17:23
  keys
• 17:24
  remain
• 17:24
  sufficient
• 17:24
  because
• 17:25
  Grover's
• 17:25
  algorithm
• 17:26
  doesn't
• 17:26
  parallelize.
• 17:27
  Valsorda
• 17:28
  has
• 17:28
  started
• 17:28
  teaching
• 17:28
  a
• 17:28
  PhD
• 17:29
  cryptography
• 17:30
  course
• 17:30
  at
• 17:30
  the
• 17:30
  University
• 17:31
  of
• 17:31
  Bologna
• 17:32
  where
• 17:32
  RSA,
• 17:33
  ECDSA,
• 17:34
  and
• 17:34
  ECDH
• 17:35
  are
• 17:35
  introduced
• 17:35
  as
• 17:35
  legacy
• 17:36
  algorithms
• 17:36
  —
• 17:37
  because
• 17:37
  that's
• 17:38
  how
• 17:38
  the
• 17:38
  students
• 17:38
  will
• 17:38
  encounter
• 17:39
  them
• 17:39
  in
• 17:39
  their
• 17:39
  careers.
• 17:43
  Alibaba's
• 17:43
  Qwen
• 17:44
  team
• 17:44
  released
• 17:44
  Qwen3.6-Plus,
• 17:47
  positioning
• 17:48
  it
• 17:48
  as
• 17:48
  a
• 17:48
  "critical
• 17:48
  milestone"
• 17:49
  toward
• 17:49
  native
• 17:49
  multimodal
• 17:50
  agents
• 17:50
  with
• 17:51
  emphasis
• 17:51
  on
• 17:51
  agentic
• 17:51
  coding.
• 17:52
  Community
• 17:53
  reaction
• 17:53
  is
• 17:53
  mixed:
• 17:54
  the
• 17:54
  benchmarks
• 17:55
  look
• 17:55
  strong
• 17:55
  against
• 17:55
  GLM-5,
• 17:56
  Opus
• 17:57
  4.5,
• 17:58
  and
• 17:58
  Gemini
• 17:58
  3
• 17:58
  Pro,
• 17:59
  but
• 17:59
  commenters
• 18:00
  note
• 18:00
  the
• 18:00
  conspicuous
• 18:01
  absence
• 18:01
  of
• 18:01
  GPT
• 18:02
  5.4
• 18:03
  and
• 18:03
  Opus
• 18:03
  4.6
• 18:04
  from
• 18:04
  comparisons,
• 18:05
  and
• 18:05
  the
• 18:05
  team
• 18:05
  admits
• 18:06
  to
• 18:06
  "correcting
• 18:06
  problematic
• 18:07
  tasks"
• 18:07
  in
• 18:07
  SWE-bench
• 18:08
  Pro
• 18:08
  before
• 18:08
  evaluation.
• 18:10
  Open-source
• 18:10
  smaller
• 18:11
  variants
• 18:11
  are
• 18:11
  promised
• 18:11
  "in
• 18:11
  the
• 18:12
  coming
• 18:12
  days."
• 18:13
  Omnivoice,
• 18:14
  an
• 18:14
  open-source
• 18:15
  TTS
• 18:15
  system
• 18:16
  supporting
• 18:16
  600+
• 18:17
  languages
• 18:17
  with
• 18:18
  voice
• 18:18
  cloning,
• 18:19
  is
• 18:19
  generating
• 18:19
  excitement
• 18:20
  for
• 18:20
  its
• 18:20
  quality-to-size
• 18:21
  ratio.
• 18:22
  Early
• 18:22
  testers
• 18:23
  report
• 18:23
  12x
• 18:24
  real-time
• 18:24
  generation
• 18:25
  speed
• 18:25
  on
• 18:25
  a
• 18:25
  5090,
• 18:27
  "insanely
• 18:27
  good"
• 18:27
  voice
• 18:28
  cloning
• 18:28
  even
• 18:28
  for
• 18:29
  non-English
• 18:29
  languages,
• 18:30
  and
• 18:30
  expressive
• 18:30
  tags
• 18:31
  ([laughter],
• 18:31
  [sigh],
• 18:31
  [surprise-ah])
• 18:32
  that
• 18:32
  make
• 18:32
  output
• 18:32
  sound
• 18:33
  alive.
• 18:34
  At
• 18:34
  ~6.5GB
• 18:35
  VRAM,
• 18:36
  it
• 18:36
  sits
• 18:36
  in
• 18:36
  a
• 18:36
  practical
• 18:37
  sweet
• 18:37
  spot
• 18:37
  for
• 18:37
  local
• 18:38
  deployment.
• 18:39
  The
• 18:39
  main
• 18:39
  gap:
• 18:40
  inference
• 18:40
  still
• 18:40
  requires
• 18:41
  torchaudio,
• 18:42
  with
• 18:42
  no
• 18:42
  lightweight
• 18:43
  C++
• 18:43
  runtime
• 18:44
  yet.
• 18:45
  Andriy
• 18:45
  Burkov
• 18:46
  highlighted
• 18:46
  a
• 18:46
  paper
• 18:46
  that
• 18:46
  eliminates
• 18:47
  iterative
• 18:47
  denoising
• 18:48
  from
• 18:48
  image
• 18:48
  generation
• 18:49
  entirely.
• 18:49
  Instead
• 18:50
  of
• 18:50
  running
• 18:50
  a
• 18:50
  neural
• 18:51
  network
• 18:51
  dozens
• 18:51
  of
• 18:52
  times
• 18:52
  to
• 18:52
  gradually
• 18:52
  clean
• 18:53
  noise
• 18:53
  into
• 18:53
  an
• 18:53
  image
• 18:54
  (diffusion's
• 18:54
  core
• 18:54
  mechanism),
• 18:55
  the
• 18:56
  authors
• 18:56
  define
• 18:56
  a
• 18:57
  "drifting
• 18:57
  field"
• 18:58
  that
• 18:58
  lets
• 18:58
  training
• 18:58
  itself
• 18:59
  do
• 18:59
  the
• 18:59
  distributional
• 18:59
  work
• 19:00
  —
• 19:00
  the
• 19:00
  network
• 19:01
  is
• 19:01
  a
• 19:01
  single-pass
• 19:02
  mapping
• 19:02
  from
• 19:02
  noise
• 19:02
  to
• 19:02
  data.
• 19:03
  The
• 19:03
  result:
• 19:04
  1.54
• 19:05
  FID
• 19:06
  on
• 19:06
  ImageNet
• 19:07
  256x256,
• 19:10
  better
• 19:10
  than
• 19:10
  all
• 19:11
  prior
• 19:11
  single-step
• 19:11
  methods
• 19:12
  and
• 19:12
  competitive
• 19:12
  with
• 19:13
  multi-step
• 19:13
  diffusion
• 19:14
  models
• 19:14
  that
• 19:15
  run
• 19:15
  hundreds
• 19:15
  of
• 19:15
  passes.
• 19:16
  The
• 19:16
  detail
• 19:17
  buried
• 19:17
  at
• 19:17
  the
• 19:17
  end
• 19:17
  —
• 19:18
  the
• 19:18
  framework
• 19:19
  also
• 19:19
  works
• 19:19
  for
• 19:19
  robot
• 19:20
  control
• 19:20
  policies
• 19:20
  —
• 19:21
  hints
• 19:21
  at
• 19:22
  how
• 19:22
  general
• 19:22
  this
• 19:22
  approach
• 19:22
  really
• 19:23
  is.
• 19:24
  Single-pass
• 19:24
  generation
• 19:25
  from
• 19:25
  complex
• 19:25
  distributions
• 19:26
  is
• 19:26
  not
• 19:27
  just
• 19:27
  an
• 19:27
  image
• 19:27
  technique;
• 19:28
  it's
• 19:28
  a
• 19:28
  paradigm.