Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – April 27, 2026

Transcript

AGI Dreams Podcast – April 27, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  OX
• 0:00
  Security
• 0:01
  dropped
• 0:01
  a
• 0:01
  sobering
• 0:02
  cross-language
• 0:02
  audit
• 0:02
  of
• 0:03
  Anthropic's
• 0:03
  Model
• 0:03
  Context
• 0:04
  Protocol
• 0:04
  this
• 0:04
  week,
• 0:05
  and
• 0:05
  the
• 0:05
  findings
• 0:06
  land
• 0:06
  exactly
• 0:06
  where
• 0:06
  anyone
• 0:07
  who
• 0:07
  has
• 0:07
  built
• 0:07
  a
• 0:07
  tool-integration
• 0:08
  layer
• 0:08
  would
• 0:08
  expect:
• 0:09
  the
• 0:09
  command
• 0:10
  arguments
• 0:10
  passed
• 0:10
  to
• 0:10
  spin
• 0:11
  up
• 0:11
  a
• 0:11
  new
• 0:11
  local
• 0:11
  MCP
• 0:12
  server
• 0:12
  instance
• 0:12
  are
• 0:12
  executed
• 0:13
  in
• 0:13
  a
• 0:13
  server-side
• 0:14
  shell
• 0:14
  with
• 0:14
  no
• 0:14
  input
• 0:15
  sanitization.
• 0:16
  Across
• 0:16
  real-world
• 0:17
  exploitation
• 0:18
  attempts
• 0:18
  against
• 0:18
  LettaAI,
• 0:19
  LangFlow,
• 0:19
  Flowise,
• 0:20
  and
• 0:20
  the
• 0:20
  Windsurf
• 0:20
  IDE,
• 0:21
  researchers
• 0:21
  achieved
• 0:22
  remote
• 0:22
  code
• 0:22
  execution
• 0:23
  —
• 0:23
  and
• 0:24
  in
• 0:24
  Flowise's
• 0:24
  case,
• 0:25
  where
• 0:25
  some
• 0:25
  sanitization
• 0:26
  existed
• 0:26
  (command
• 0:26
  allowlists,
• 0:27
  special-character
• 0:28
  stripping),
• 0:28
  standard
• 0:28
  command
• 0:29
  flags
• 0:29
  bypassed
• 0:30
  it
• 0:30
  trivially.
• 0:31
  The
• 0:31
  most
• 0:31
  telling
• 0:31
  detail
• 0:32
  is
• 0:32
  Anthropic's
• 0:32
  response:
• 0:33
  no
• 0:34
  design
• 0:34
  flaw,
• 0:35
  works
• 0:35
  as
• 0:35
  designed,
• 0:36
  input
• 0:36
  sanitization
• 0:37
  is
• 0:37
  the
• 0:37
  developer's
• 0:38
  responsibility.
• 0:39
  That
• 0:39
  framing
• 0:40
  treats
• 0:40
  MCP
• 0:41
  like
• 0:41
  a
• 0:41
  low-level
• 0:42
  syscall
• 0:42
  when
• 0:43
  the
• 0:43
  ecosystem
• 0:43
  markets
• 0:44
  it
• 0:44
  as
• 0:44
  a
• 0:44
  plug-and-play
• 0:44
  integration
• 0:45
  layer
• 0:45
  for
• 0:45
  developers
• 0:46
  who
• 0:46
  may
• 0:46
  never
• 0:46
  have
• 0:46
  written
• 0:47
  a
• 0:47
  shell
• 0:47
  escape
• 0:47
  filter.
• 0:48
  The
• 0:48
  Hackaday
• 0:49
  comment
• 0:49
  section,
• 0:50
  rarely
• 0:50
  a
• 0:50
  source
• 0:51
  of
• 0:51
  nuance,
• 0:51
  actually
• 0:52
  nailed
• 0:52
  it:
• 0:53
  "Apache
• 0:53
  doesn't
• 0:53
  come
• 0:53
  with
• 0:54
  preconfigured
• 0:54
  setups
• 0:55
  to
• 0:55
  format
• 0:55
  your
• 0:55
  C:."
• 0:56
  The
• 0:56
  difference
• 0:57
  is
• 0:57
  that
• 0:57
  MCP
• 0:58
  ships
• 0:58
  reference
• 0:58
  implementations
• 0:59
  that
• 0:59
  make
• 0:59
  the
• 1:00
  unsafe
• 1:00
  path
• 1:00
  the
• 1:00
  easy
• 1:01
  path.
• 1:02
  Meanwhile,
• 1:03
  the
• 1:03
  exploit-velocity
• 1:04
  clock
• 1:04
  keeps
• 1:04
  accelerating.
• 1:06
  A
• 1:06
  LinkedIn
• 1:06
  post
• 1:06
  from
• 1:07
  a
• 1:07
  security
• 1:07
  lead
• 1:07
  documented
• 1:08
  that
• 1:08
  CVE-2026-41179
• 1:10
  —
• 1:13
  an
• 1:13
  unauthenticated
• 1:14
  single-request
• 1:15
  RCE
• 1:15
  in
• 1:16
  rclone's
• 1:16
  Remote
• 1:16
  Control
• 1:17
  daemon
• 1:17
  via
• 1:17
  the
• 1:17
  WebDAV
• 1:18
  bearer
• 1:18
  token
• 1:18
  command
• 1:19
  parameter
• 1:19
  —
• 1:20
  went
• 1:20
  from
• 1:20
  GitHub
• 1:21
  advisory
• 1:21
  publication
• 1:22
  to
• 1:22
  in-the-wild
• 1:22
  exploit
• 1:23
  in
• 1:23
  18
• 1:23
  hours.
• 1:25
  The
• 1:25
  attacker,
• 1:26
  operating
• 1:26
  from
• 1:26
  a
• 1:26
  Hong
• 1:26
  Kong
• 1:27
  VPS,
• 1:28
  sent
• 1:28
  two
• 1:28
  surgical
• 1:28
  POSTs:
• 1:29
  /rc/noop
• 1:32
  to
• 1:32
  confirm
• 1:32
  the
• 1:32
  daemon
• 1:32
  was
• 1:33
  alive,
• 1:33
  then
• 1:34
  the
• 1:34
  payload
• 1:34
  within
• 1:34
  the
• 1:34
  same
• 1:35
  second.
• 1:36
  Rclone
• 1:36
  sits
• 1:36
  in
• 1:37
  56,000+
• 1:39
  GitHub
• 1:39
  stars
• 1:40
  worth
• 1:40
  of
• 1:40
  backup
• 1:40
  pipelines,
• 1:41
  NAS
• 1:41
  appliances,
• 1:42
  and
• 1:42
  multi-cloud
• 1:43
  sync
• 1:43
  jobs.
• 1:44
  This
• 1:44
  is
• 1:44
  the
• 1:45
  third
• 1:45
  internet-facing
• 1:46
  service
• 1:46
  this
• 1:46
  month
• 1:46
  with
• 1:46
  sub-24-hour
• 1:47
  time-to-exploit
• 1:48
  on
• 1:49
  a
• 1:49
  critical
• 1:49
  CVE.
• 1:50
  On
• 1:50
  the
• 1:51
  defensive
• 1:51
  tooling
• 1:51
  side,
• 1:52
  the
• 1:52
  OpenSourceMalware
• 1:53
  team
• 1:53
  published
• 1:54
  PolinRider,
• 1:55
  a
• 1:55
  comprehensive
• 1:56
  analysis
• 1:56
  of
• 1:56
  a
• 1:56
  DPRK-attributed
• 1:58
  (Lazarus
• 1:58
  group)
• 1:58
  supply
• 1:59
  chain
• 1:59
  campaign
• 2:00
  that
• 2:00
  has
• 2:00
  compromised
• 2:00
  over
• 2:01
  1,950
• 2:02
  public
• 2:03
  GitHub
• 2:03
  repositories
• 2:04
  belonging
• 2:04
  to
• 2:04
  more
• 2:05
  than
• 2:05
  1,047
• 2:06
  unique
• 2:06
  owners.
• 2:07
  The
• 2:08
  attack
• 2:08
  implants
• 2:08
  heavily
• 2:09
  obfuscated
• 2:09
  JavaScript
• 2:10
  payloads
• 2:10
  into
• 2:11
  everyday
• 2:11
  config
• 2:11
  files
• 2:13
  —
• 2:15
  postcss.config.mjs,
• 2:17
  tailwind.config.js,
• 2:19
  eslint.config.mjs
• 2:21
  —
• 2:23
  by
• 2:23
  appending
• 2:23
  malicious
• 2:24
  code
• 2:24
  after
• 2:24
  legitimate
• 2:25
  content,
• 2:26
  making
• 2:26
  it
• 2:26
  invisible
• 2:26
  during
• 2:27
  casual
• 2:27
  code
• 2:27
  review.
• 2:28
  The
• 2:29
  most
• 2:29
  technically
• 2:29
  distinctive
• 2:30
  feature
• 2:30
  is
• 2:31
  a
• 2:31
  blockchain-based
• 2:31
  dead-drop
• 2:32
  C2
• 2:32
  architecture:
• 2:33
  the
• 2:33
  final
• 2:34
  payload
• 2:34
  fetches
• 2:35
  XOR-encrypted
• 2:36
  JavaScript
• 2:36
  from
• 2:36
  immutable
• 2:37
  TRON,
• 2:38
  Aptos,
• 2:38
  and
• 2:38
  Binance
• 2:39
  Smart
• 2:39
  Chain
• 2:39
  transactions,
• 2:41
  making
• 2:41
  the
• 2:41
  command-and-control
• 2:42
  infrastructure
• 2:42
  virtually
• 2:43
  impossible
• 2:43
  to
• 2:44
  take
• 2:44
  down.
• 2:45
  Two
• 2:45
  weaponized
• 2:46
  fake
• 2:46
  job-interview
• 2:47
  projects
• 2:47
  —
• 2:48
  ShoeVista
• 2:49
  and
• 2:49
  StakingGame
• 2:49
  —
• 2:50
  have
• 2:50
  been
• 2:50
  used
• 2:50
  to
• 2:51
  socially
• 2:51
  engineer
• 2:51
  developers,
• 2:52
  compromising
• 2:53
  at
• 2:53
  least
• 2:53
  88
• 2:54
  individuals.
• 2:55
  The
• 2:55
  Neutralinojs
• 2:56
  project
• 2:56
  (8,400
• 2:58
  stars,
• 2:58
  495
• 2:59
  forks)
• 3:00
  was
• 3:00
  among
• 3:01
  the
• 3:01
  highest-profile
• 3:02
  victims.
• 3:03
  OSM
• 3:03
  published
• 3:04
  multi-variant
• 3:05
  YARA
• 3:05
  rules,
• 3:06
  a
• 3:06
  scanner
• 3:06
  script,
• 3:07
  and
• 3:07
  detailed
• 3:08
  IOCs
• 3:08
  including
• 3:09
  blockchain
• 3:09
  addresses
• 3:10
  and
• 3:10
  XOR
• 3:10
  keys.
• 3:16
  Both
• 3:16
  llama.cpp
• 3:18
  and
• 3:18
  ik
• 3:18
  llama.cpp
• 3:19
  now
• 3:19
  ship
• 3:20
  FP4
• 3:20
  inference
• 3:20
  support,
• 3:21
  but
• 3:21
  the
• 3:21
  implementations
• 3:22
  diverge
• 3:22
  in
• 3:22
  ways
• 3:23
  that
• 3:23
  matter.
• 3:23
  llama.cpp
• 3:25
  merged
• 3:25
  NVFP4
• 3:25
  —
• 3:26
  Nvidia's
• 3:27
  block-scaled
• 3:27
  E4M3
• 3:28
  format
• 3:28
  —
• 3:29
  with
• 3:29
  CUDA
• 3:29
  kernels
• 3:30
  in
• 3:30
  mmq.cuh,
• 3:31
  mmvq.cu,
• 3:33
  and
• 3:33
  convert.cu.
• 3:35
  ik
• 3:35
  llama.cpp
• 3:36
  went
• 3:36
  with
• 3:36
  MXFP4,
• 3:37
  the
• 3:37
  MX
• 3:38
  consortium
• 3:38
  standard,
• 3:39
  and
• 3:39
  has
• 3:39
  broader
• 3:39
  backend
• 3:40
  coverage:
• 3:41
  CPU
• 3:41
  (AVX2,
• 3:42
  NEON,
• 3:42
  Zen4)
• 3:43
  and
• 3:43
  CUDA.
• 3:43
  The
• 3:44
  community
• 3:44
  reality
• 3:44
  check
• 3:45
  is
• 3:45
  important
• 3:45
  here:
• 3:46
  one
• 3:46
  commenter
• 3:46
  noted
• 3:47
  this
• 3:47
  is
• 3:47
  "just
• 3:47
  compatibility,
• 3:48
  not
• 3:48
  speedup,
• 3:49
  yet,"
• 3:50
  with
• 3:50
  a
• 3:50
  known
• 3:50
  2%
• 3:51
  perplexity
• 3:51
  loss
• 3:52
  bug,
• 3:52
  and
• 3:53
  real
• 3:53
  NVFP4
• 3:55
  acceleration
• 3:55
  only
• 3:55
  kicking
• 3:56
  in
• 3:56
  on
• 3:56
  prefill
• 3:56
  at
• 3:56
  30-40k+
• 3:57
  context.
• 3:58
  Benchmarks
• 3:59
  on
• 3:59
  the
• 3:59
  Qwen3-1.7B-NVFP4A16
• 4:04
  model
• 4:04
  showed
• 4:04
  85.8%
• 4:05
  recovery
• 4:06
  on
• 4:06
  MMLU-Redux
• 4:08
  and
• 4:09
  a
• 4:09
  rough
• 4:09
  53.8%
• 4:10
  on
• 4:10
  LiveCodeBench
• 4:11
  Pass@1
• 4:11
  —
• 4:12
  functional
• 4:12
  but
• 4:13
  clearly
• 4:13
  lossy.
• 4:14
  The
• 4:14
  real
• 4:15
  question,
• 4:16
  as
• 4:16
  one
• 4:16
  user
• 4:16
  framed
• 4:17
  it:
• 4:17
  "does
• 4:18
  NVFP4
• 4:19
  actually
• 4:20
  beat
• 4:20
  MXFP4
• 4:22
  in
• 4:22
  end-to-end
• 4:22
  latency
• 4:23
  or
• 4:23
  is
• 4:23
  this
• 4:23
  just
• 4:23
  Nvidia
• 4:24
  lock-in
• 4:24
  with
• 4:24
  marginal
• 4:25
  upside?"
• 4:26
  At
• 4:26
  the
• 4:26
  extreme
• 4:27
  end
• 4:27
  of
• 4:27
  the
• 4:27
  compression
• 4:28
  spectrum,
• 4:28
  PrismML
• 4:29
  released
• 4:29
  Bonsai-8B,
• 4:30
  an
• 4:31
  end-to-end
• 4:31
  1-bit
• 4:32
  language
• 4:32
  model
• 4:32
  built
• 4:32
  on
• 4:33
  the
• 4:33
  Qwen3-8B
• 4:33
  dense
• 4:34
  architecture.
• 4:35
  Every
• 4:35
  weight
• 4:35
  maps
• 4:36
  to
• 4:36
  +1
• 4:36
  or
• 4:36
  -1,
• 4:37
  with
• 4:37
  one
• 4:37
  FP16
• 4:38
  scale
• 4:38
  factor
• 4:39
  shared
• 4:39
  across
• 4:39
  every
• 4:39
  128
• 4:40
  weights.
• 4:41
  The
• 4:41
  result:
• 4:42
  1.28
• 4:43
  GB
• 4:43
  parameter
• 4:44
  memory
• 4:44
  (down
• 4:44
  from
• 4:44
  16.38
• 4:46
  GB
• 4:46
  at
• 4:46
  FP16)
• 4:46
  —
• 4:48
  a
• 4:48
  12.8x
• 4:49
  reduction
• 4:49
  that
• 4:49
  fits
• 4:49
  comfortably
• 4:50
  on
• 4:50
  any
• 4:50
  Mac
• 4:50
  or
• 4:51
  iPhone.
• 4:52
  On
• 4:52
  Apple's
• 4:52
  M4
• 4:53
  Pro,
• 4:53
  Bonsai
• 4:54
  generates
• 4:54
  at
• 4:54
  64.2
• 4:55
  tok/s
• 4:56
  versus
• 4:56
  22.9
• 4:57
  tok/s
• 4:58
  for
• 4:58
  4-bit,
• 4:59
  and
• 4:59
  the
• 4:59
  energy-per-token
• 5:00
  drops
• 5:00
  4-6x
• 5:01
  despite
• 5:01
  higher
• 5:01
  instantaneous
• 5:02
  power
• 5:02
  draw.
• 5:03
  The
• 5:03
  benchmark
• 5:04
  story
• 5:04
  is
• 5:04
  the
• 5:04
  real
• 5:05
  surprise:
• 5:06
  a
• 5:06
  70.5
• 5:07
  average
• 5:07
  score
• 5:08
  across
• 5:08
  six
• 5:08
  categories
• 5:09
  on
• 5:09
  EvalScope,
• 5:10
  competitive
• 5:11
  with
• 5:11
  full-precision
• 5:12
  8B
• 5:12
  models
• 5:12
  at
• 5:13
  one-fourteenth
• 5:13
  the
• 5:13
  size.
• 5:14
  PrismML
• 5:15
  quantifies
• 5:16
  this
• 5:16
  with
• 5:16
  an
• 5:16
  "intelligence
• 5:17
  density"
• 5:17
  metric
• 5:17
  —
• 5:18
  alpha
• 5:19
  =
• 5:20
  -ln(1
• 5:20
  -
• 5:21
  score/100)
• 5:22
  /
• 5:23
  size
• 5:23
  GB
• 5:23
  —
• 5:24
  where
• 5:24
  Bonsai
• 5:25
  achieves
• 5:25
  dramatically
• 5:26
  higher
• 5:26
  density
• 5:27
  than
• 5:27
  full-precision
• 5:27
  Qwen3-8B.
• 5:29
  The
• 5:29
  catch:
• 5:30
  this
• 5:30
  requires
• 5:31
  PrismML's
• 5:31
  fork
• 5:31
  of
• 5:32
  MLX
• 5:32
  with
• 5:32
  custom
• 5:33
  1-bit
• 5:33
  kernel
• 5:33
  the
• 5:34
  upstream
• 5:34
  PR
• 5:35
  is
• 5:35
  still
• 5:35
  pending.
• 5:36
  No
• 5:36
  native
• 5:37
  1-bit
• 5:37
  hardware
• 5:37
  exists
• 5:38
  yet,
• 5:38
  so
• 5:39
  all
• 5:39
  gains
• 5:39
  are
• 5:39
  software-kernel
• 5:40
  optimizations
• 5:41
  on
• 5:41
  general-purpose
• 5:42
  silicon.
• 5:43
  For
• 5:43
  anyone
• 5:43
  trying
• 5:44
  to
• 5:44
  figure
• 5:44
  out
• 5:44
  whether
• 5:44
  a
• 5:44
  specific
• 5:45
  GGUF
• 5:46
  quant
• 5:46
  will
• 5:46
  actually
• 5:47
  run
• 5:47
  on
• 5:47
  their
• 5:47
  hardware,
• 5:48
  VRAM.cpp
• 5:51
  takes
• 5:52
  the
• 5:52
  most
• 5:52
  practical
• 5:52
  approach
• 5:53
  possible:
• 5:54
  it
• 5:54
  compiles
• 5:54
  llama.cpp's
• 5:56
  own
• 5:56
  fit.cpp
• 5:58
  layer-assignment
• 5:58
  algorithm
• 5:59
  to
• 5:59
  WebAssembly
• 6:00
  and
• 6:00
  runs
• 6:00
  it
• 6:00
  directly
• 6:01
  in
• 6:01
  the
• 6:01
  browser
• 6:01
  against
• 6:02
  a
• 6:02
  dragged-in
• 6:02
  model
• 6:03
  file.
• 6:03
  No
• 6:04
  estimation
• 6:04
  heuristics,
• 6:05
  no
• 6:05
  outdated
• 6:06
  model
• 6:06
  databases
• 6:06
  —
• 6:07
  just
• 6:07
  the
• 6:07
  actual
• 6:08
  fit
• 6:08
  algorithm
• 6:09
  that
• 6:09
  llama.cpp
• 6:10
  itself
• 6:10
  uses,
• 6:11
  executed
• 6:12
  client-side.
• 6:13
  Multi-GPU
• 6:14
  scenarios
• 6:14
  still
• 6:14
  behave
• 6:15
  strangely,
• 6:16
  and
• 6:16
  MoE
• 6:16
  fitting
• 6:17
  is
• 6:17
  wonky,
• 6:18
  but
• 6:18
  for
• 6:18
  the
• 6:18
  single-GPU
• 6:19
  question
• 6:19
  of
• 6:19
  "can
• 6:20
  my
• 6:20
  16GB
• 6:21
  card
• 6:21
  run
• 6:21
  this
• 6:21
  Q3
• 6:22
  quant
• 6:22
  from
• 6:22
  Bartowski,"
• 6:23
  it
• 6:23
  beats
• 6:24
  every
• 6:24
  calculator
• 6:25
  out
• 6:25
  there.
• 6:26
  The
• 6:26
  hardware
• 6:26
  experimentation
• 6:27
  continues
• 6:28
  to
• 6:28
  push
• 6:28
  into
• 6:28
  exotic
• 6:29
  territory.
• 6:30
  A
• 6:30
  Reddit
• 6:30
  user
• 6:31
  revived
• 6:31
  the
• 6:31
  FPGA
• 6:32
  inference
• 6:33
  idea
• 6:33
  from
• 6:33
  the
• 6:33
  crypto
• 6:34
  mining
• 6:34
  era,
• 6:35
  asking
• 6:35
  whether
• 6:35
  an
• 6:35
  AMD
• 6:36
  Alveo
• 6:37
  V80
• 6:38
  ($9,500,
• 6:39
  32GB
• 6:40
  HBM2e,
• 6:41
  820
• 6:43
  GB/s
• 6:44
  bandwidth,
• 6:45
  ~673
• 6:46
  Mb
• 6:47
  on-chip
• 6:47
  SRAM)
• 6:48
  could
• 6:48
  approximate
• 6:48
  the
• 6:48
  Taalas
• 6:49
  HC1's
• 6:50
  15,000
• 6:51
  tok/s
• 6:52
  by
• 6:52
  treating
• 6:52
  the
• 6:53
  FPGA
• 6:53
  as
• 6:53
  a
• 6:53
  spatial
• 6:54
  compute
• 6:54
  fabric
• 6:55
  rather
• 6:55
  than
• 6:55
  a
• 6:55
  GPU
• 6:56
  substitute.
• 6:57
  The
• 6:57
  speculative
• 6:58
  architecture
• 6:58
  —
• 6:59
  a
• 6:59
  "Dual-Tier
• 7:00
  Speculative
• 7:00
  Fabric"
• 7:01
  with
• 7:01
  an
• 7:01
  ultra-quantized
• 7:02
  draft
• 7:03
  model
• 7:03
  living
• 7:03
  entirely
• 7:04
  in
• 7:04
  SRAM
• 7:05
  and
• 7:05
  a
• 7:05
  DARF
• 7:05
  (Dynamic
• 7:06
  Activation-Routed
• 7:06
  Fetching)
• 7:07
  system
• 7:08
  exploiting
• 7:08
  activation
• 7:09
  sparsity
• 7:09
  to
• 7:10
  fetch
• 7:10
  only
• 7:10
  15-25%
• 7:14
  of
• 7:14
  weights
• 7:14
  from
• 7:14
  HBM
• 7:15
  per
• 7:15
  token
• 7:15
  —
• 7:17
  yields
• 7:17
  theoretical
• 7:17
  estimates
• 7:18
  of
• 7:18
  ~3,200
• 7:19
  tok/s
• 7:21
  on
• 7:21
  Qwen
• 7:21
  4B
• 7:22
  and
• 7:22
  ~1,400
• 7:24
  tok/s
• 7:25
  on
• 7:25
  9B.
• 7:26
  Community
• 7:26
  response
• 7:27
  was
• 7:27
  appropriately
• 7:27
  skeptical:
• 7:29
  "the
• 7:29
  whole
• 7:29
  Gemini
• 7:29
  output
• 7:30
  sounds
• 7:30
  like
• 7:30
  'That's
• 7:31
  a
• 7:31
  great
• 7:31
  idea
• 7:31
  for
• 7:31
  a
• 7:32
  blender!
• 7:33
  Now
• 7:33
  let's
• 7:33
  build
• 7:33
  the
• 7:33
  cold-fusion
• 7:34
  reactor
• 7:34
  it
• 7:34
  needs.'"
• 7:35
  One
• 7:36
  commenter
• 7:36
  pointed
• 7:36
  out
• 7:37
  the
• 7:37
  V80's
• 7:37
  SRAM
• 7:38
  is
• 7:38
  673
• 7:39
  megabits,
• 7:40
  not
• 7:40
  megabytes
• 7:41
  —
• 7:42
  84MB
• 7:43
  after
• 7:43
  dividing
• 7:43
  by
• 7:44
  8.
• 7:45
  Programming
• 7:45
  the
• 7:45
  custom
• 7:45
  memory
• 7:46
  controllers
• 7:46
  in
• 7:47
  Verilog/VHDL
• 7:49
  would
• 7:49
  be
• 7:49
  the
• 7:49
  real
• 7:50
  bottleneck,
• 7:51
  not
• 7:51
  the
• 7:51
  math.
• 7:52
  Still,
• 7:53
  someone
• 7:53
  is
• 7:53
  already
• 7:54
  attempting
• 7:54
  a
• 7:54
  similar
• 7:55
  project
• 7:55
  with
• 7:55
  a
• 7:55
  $250-350
• 7:58
  FPGA
• 7:59
  and
• 7:59
  a
• 7:59
  distilled
• 8:00
  Gemma
• 8:00
  4.
• 8:01
  For
• 8:01
  the
• 8:01
  growing
• 8:02
  population
• 8:02
  running
• 8:03
  heterogeneous
• 8:03
  GPU
• 8:04
  setups
• 8:05
  for
• 8:05
  local
• 8:05
  inference,
• 8:06
  ProjectPhysX
• 8:07
  released
• 8:07
  hw-smi,
• 8:09
  a
• 8:09
  cross-platform
• 8:09
  CPU/GPU
• 8:11
  telemetry
• 8:12
  monitor
• 8:12
  that
• 8:12
  pulls
• 8:13
  data
• 8:13
  directly
• 8:13
  from
• 8:14
  vendor
• 8:14
  APIs
• 8:14
  —
• 8:15
  NVML
• 8:16
  for
• 8:16
  Nvidia,
• 8:17
  ADLX/AMDSMI
• 8:23
  for
• 8:23
  AMD,
• 8:24
  SYSMAN
• 8:26
  for
• 8:26
  Intel
• 8:26
  —
• 8:27
  with
• 8:27
  ASCII
• 8:28
  bar
• 8:28
  and
• 8:28
  graph
• 8:29
  visualization.
• 8:30
  It
• 8:30
  covers
• 8:30
  per-core
• 8:31
  usage,
• 8:32
  memory
• 8:32
  bandwidth,
• 8:33
  temperature,
• 8:34
  power,
• 8:34
  fan
• 8:35
  speed,
• 8:35
  clocks,
• 8:36
  and
• 8:36
  PCIe
• 8:37
  bandwidth
• 8:38
  across
• 8:38
  Windows
• 8:39
  and
• 8:39
  Linux,
• 8:40
  with
• 8:40
  a
• 8:40
  detailed
• 8:40
  compatibility
• 8:41
  matrix
• 8:42
  documenting
• 8:42
  which
• 8:42
  metrics
• 8:43
  are
• 8:43
  supported,
• 8:44
  require
• 8:44
  admin
• 8:45
  privileges,
• 8:46
  need
• 8:46
  workarounds,
• 8:47
  or
• 8:47
  are
• 8:47
  simply
• 8:47
  broken
• 8:48
  in
• 8:48
  vendor
• 8:48
  APIs.
• 8:53
  DeepSeek
• 8:54
  released
• 8:54
  its
• 8:54
  V4
• 8:54
  technical
• 8:55
  paper
• 8:55
  this
• 8:55
  week,
• 8:56
  detailing
• 8:56
  two
• 8:57
  MoE
• 8:57
  models
• 8:57
  that
• 8:57
  push
• 8:58
  the
• 8:58
  efficiency
• 8:58
  frontier
• 8:59
  hard.
• 8:59
  DeepSeek-V4-Pro
• 9:00
  packs
• 9:01
  1.6
• 9:01
  trillion
• 9:02
  total
• 9:02
  parameters
• 9:02
  with
• 9:03
  only
• 9:03
  49
• 9:03
  billion
• 9:04
  activated
• 9:04
  per
• 9:04
  token
• 9:05
  across
• 9:05
  61
• 9:05
  Transformer
• 9:06
  layers;
• 9:07
  DeepSeek-V4-Flash
• 9:08
  is
• 9:08
  the
• 9:08
  lighter
• 9:09
  sibling
• 9:09
  at
• 9:09
  284
• 9:10
  billion
• 9:10
  total,
• 9:11
  13
• 9:11
  billion
• 9:12
  activated,
• 9:12
  43
• 9:13
  layers.
• 9:14
  Both
• 9:14
  support
• 9:14
  one-million-token
• 9:15
  context.
• 9:16
  The
• 9:16
  architectural
• 9:16
  headline
• 9:17
  is
• 9:17
  a
• 9:17
  hybrid
• 9:17
  attention
• 9:18
  mechanism
• 9:18
  combining
• 9:18
  Compressed
• 9:19
  Sparse
• 9:19
  Attention
• 9:19
  (CSA)
• 9:20
  —
• 9:20
  which
• 9:20
  compresses
• 9:21
  KV
• 9:21
  caches
• 9:22
  4x
• 9:22
  along
• 9:22
  the
• 9:22
  sequence
• 9:23
  dimension
• 9:23
  before
• 9:23
  applying
• 9:24
  sparse
• 9:24
  attention
• 9:24
  with
• 9:25
  a
• 9:25
  "lightning
• 9:25
  indexer"
• 9:25
  —
• 9:26
  and
• 9:26
  Heavily
• 9:27
  Compressed
• 9:27
  Attention
• 9:27
  (HCA),
• 9:28
  which
• 9:28
  compresses
• 9:29
  128x
• 9:30
  but
• 9:30
  retains
• 9:30
  dense
• 9:30
  attention.
• 9:31
  At
• 9:31
  1M
• 9:32
  tokens,
• 9:32
  V4-Pro
• 9:33
  requires
• 9:33
  only
• 9:33
  27%
• 9:34
  of
• 9:34
  V3.2's
• 9:36
  single-token
• 9:36
  inference
• 9:37
  FLOPs
• 9:37
  and
• 9:37
  10%
• 9:37
  of
• 9:38
  its
• 9:38
  KV
• 9:38
  cache.
• 9:39
  Infrastructure
• 9:39
  innovations
• 9:40
  include
• 9:40
  fused
• 9:40
  MoE
• 9:41
  kernels
• 9:41
  (1.5-1.96x
• 9:42
  speedup),
• 9:44
  FP4
• 9:46
  quantization-aware
• 9:46
  training
• 9:47
  for
• 9:47
  expert
• 9:47
  weights,
• 9:48
  and
• 9:48
  the
• 9:48
  Muon
• 9:48
  optimizer
• 9:49
  replacing
• 9:49
  AdamW
• 9:50
  with
• 9:50
  hybrid
• 9:50
  Newton-Schulz
• 9:50
  orthogonalization
• 9:51
  for
• 9:51
  faster
• 9:52
  convergence.
• 9:53
  The
• 9:53
  benchmark
• 9:53
  numbers
• 9:53
  are
• 9:53
  striking:
• 9:54
  90.2%
• 9:55
  on
• 9:55
  SimpleQA
• 9:55
  Verified
• 9:56
  (20
• 9:56
  points
• 9:56
  above
• 9:57
  prior
• 9:57
  open
• 9:57
  models),
• 9:58
  a
• 9:58
  Codeforces
• 9:59
  rating
• 9:59
  of
• 9:59
  3206
• 10:00
  (ranking
• 10:00
  23rd
• 10:01
  among
• 10:01
  humans),
• 10:02
  90.1%
• 10:03
  on
• 10:03
  GPQA
• 10:04
  Diamond,
• 10:04
  and
• 10:04
  120/120
• 10:06
  on
• 10:06
  Putnam-2025
• 10:07
  formal
• 10:07
  proofs.
• 10:08
  DeepSeek
• 10:09
  claims
• 10:09
  V4-Pro-Max
• 10:10
  outperforms
• 10:10
  GPT-5.2
• 10:11
  and
• 10:11
  Gemini-3.0-Pro
• 10:13
  on
• 10:13
  standard
• 10:13
  reasoning
• 10:14
  benchmarks
• 10:14
  while
• 10:14
  trailing
• 10:15
  GPT-5.4
• 10:16
  and
• 10:16
  Gemini-3.1-Pro
• 10:17
  by
• 10:17
  an
• 10:17
  estimated
• 10:18
  3-6
• 10:18
  months.
• 10:19
  Post-training
• 10:20
  uses
• 10:20
  a
• 10:20
  two-stage
• 10:20
  pipeline:
• 10:21
  independent
• 10:22
  specialist
• 10:22
  training
• 10:22
  via
• 10:23
  SFT
• 10:23
  plus
• 10:23
  GRPO
• 10:24
  reinforcement
• 10:25
  learning
• 10:25
  across
• 10:26
  domains,
• 10:26
  followed
• 10:27
  by
• 10:27
  on-policy
• 10:27
  distillation
• 10:28
  from
• 10:28
  10+
• 10:28
  teacher
• 10:29
  models
• 10:29
  into
• 10:30
  a
• 10:30
  unified
• 10:30
  student.
• 10:31
  The
• 10:31
  paper
• 10:31
  acknowledges
• 10:32
  the
• 10:32
  architecture's
• 10:32
  complexity
• 10:33
  and
• 10:33
  commits
• 10:33
  to
• 10:33
  simplification
• 10:34
  in
• 10:34
  future
• 10:34
  iterations.
• 10:36
  On
• 10:36
  the
• 10:36
  research
• 10:36
  frontier,
• 10:37
  a
• 10:37
  new
• 10:37
  arxiv
• 10:38
  paper
• 10:38
  proposes
• 10:38
  Agentic
• 10:39
  Federated
• 10:39
  Learning
• 10:39
  (Agentic-FL),
• 10:41
  which
• 10:41
  replaces
• 10:41
  static
• 10:42
  FL
• 10:42
  orchestration
• 10:43
  with
• 10:43
  LLM-based
• 10:43
  autonomous
• 10:44
  agents.
• 10:45
  Server-side
• 10:46
  agents
• 10:46
  use
• 10:46
  ReAct
• 10:46
  reasoning
• 10:47
  to
• 10:47
  handle
• 10:47
  client
• 10:48
  selection,
• 10:48
  adaptive
• 10:49
  aggregation,
• 10:49
  and
• 10:50
  bias
• 10:50
  mitigation;
• 10:51
  client-side
• 10:51
  agents
• 10:52
  manage
• 10:52
  privacy
• 10:53
  budgets
• 10:53
  and
• 10:53
  adapt
• 10:53
  model
• 10:54
  complexity
• 10:54
  to
• 10:54
  local
• 10:55
  hardware.
• 10:56
  The
• 10:56
  proof-of-concept,
• 10:57
  K-Agent
• 10:57
  (built
• 10:57
  with
• 10:57
  LangGraph
• 10:58
  and
• 10:58
  Ollama),
• 10:59
  was
• 10:59
  tested
• 10:59
  on
• 10:59
  CIFAR-10
• 11:03
  and
• 11:03
  MNIST
• 11:03
  under
• 11:04
  severely
• 11:04
  non-IID
• 11:05
  conditions
• 11:05
  with
• 11:05
  25
• 11:06
  clients.
• 11:07
  The
• 11:07
  interesting
• 11:07
  finding:
• 11:08
  the
• 11:08
  tool-agent
• 11:09
  approach
• 11:09
  scaled
• 11:09
  far
• 11:10
  more
• 11:10
  efficiently
• 11:10
  than
• 11:10
  raw
• 11:11
  LLM
• 11:11
  selection
• 11:11
  —
• 11:12
  at
• 11:12
  50
• 11:13
  clients,
• 11:14
  token
• 11:14
  cost
• 11:14
  grew
• 11:14
  slower
• 11:15
  and
• 11:15
  performance
• 11:15
  improved
• 11:16
  as
• 11:16
  the
• 11:16
  LLM
• 11:17
  was
• 11:17
  less
• 11:17
  confused
• 11:18
  by
• 11:18
  reduced
• 11:18
  context.
• 11:19
  The
• 11:19
  paper
• 11:20
  honestly
• 11:20
  flags
• 11:20
  the
• 11:21
  risks:
• 11:21
  LLM
• 11:22
  hallucinations
• 11:23
  causing
• 11:23
  client
• 11:23
  misselection,
• 11:24
  prompt
• 11:25
  injection
• 11:25
  as
• 11:26
  a
• 11:26
  new
• 11:26
  attack
• 11:26
  surface
• 11:26
  for
• 11:27
  federated
• 11:27
  systems,
• 11:28
  and
• 11:28
  context
• 11:29
  window
• 11:29
  limits
• 11:29
  constraining
• 11:30
  scalability.
• 11:31
  The
• 11:31
  authors
• 11:32
  envision
• 11:32
  a
• 11:32
  future
• 11:33
  where
• 11:33
  client-side
• 11:33
  agents
• 11:34
  autonomously
• 11:34
  initiate
• 11:35
  federation
• 11:35
  rounds
• 11:36
  based
• 11:36
  on
• 11:36
  local
• 11:37
  model
• 11:37
  staleness
• 11:37
  —
• 11:38
  transforming
• 11:39
  FL
• 11:39
  from
• 11:39
  centralized
• 11:40
  orchestration
• 11:41
  into
• 11:41
  a
• 11:41
  decentralized
• 11:42
  knowledge
• 11:42
  marketplace.
• 11:47
  The
• 11:47
  hackIDLE
• 11:48
  project
• 11:48
  released
• 11:48
  a
• 11:48
  complete
• 11:49
  NIST
• 11:49
  cybersecurity
• 11:50
  fine-tuning
• 11:50
  pipeline
• 11:51
  purpose-built
• 11:51
  for
• 11:51
  Apple
• 11:52
  Silicon.
• 11:52
  Starting
• 11:53
  from
• 11:53
  596
• 11:54
  raw
• 11:54
  NIST
• 11:55
  PDF
• 11:55
  publications,
• 11:56
  the
• 11:56
  pipeline
• 11:57
  extracts
• 11:57
  structured
• 11:57
  text
• 11:58
  with
• 11:58
  Docling
• 11:58
  (with
• 11:58
  MarkItDown
• 11:59
  fallback
• 11:59
  for
• 11:59
  problematic
• 12:00
  PDFs),
• 12:01
  generates
• 12:01
  chat-style
• 12:02
  training
• 12:02
  data
• 12:02
  with
• 12:02
  sentence-aware
• 12:03
  chunking,
• 12:03
  fine-tunes
• 12:04
  a
• 12:04
  Qwen2.5-Coder-7B-Instruct-4bit
• 12:08
  model
• 12:08
  via
• 12:08
  MLX
• 12:08
  LoRA,
• 12:09
  and
• 12:09
  packages
• 12:10
  the
• 12:10
  result
• 12:10
  for
• 12:10
  Ollama
• 12:11
  through
• 12:11
  GGUF
• 12:11
  conversion.
• 12:12
  The
• 12:12
  published
• 12:12
  training
• 12:13
  dataset
• 12:13
  contains
• 12:14
  530,912
• 12:16
  examples.
• 12:16
  What
• 12:17
  makes
• 12:17
  this
• 12:17
  notable
• 12:17
  beyond
• 12:18
  the
• 12:18
  cybersecurity
• 12:18
  domain
• 12:19
  is
• 12:19
  the
• 12:19
  end-to-end
• 12:20
  reproducibility:
• 12:21
  the
• 12:21
  repo
• 12:21
  includes
• 12:21
  every
• 12:22
  step
• 12:22
  from
• 12:22
  make
• 12:22
  download
• 12:23
  through
• 12:23
  make
• 12:23
  deploy,
• 12:24
  with
• 12:24
  branching
• 12:24
  paths
• 12:25
  into
• 12:25
  CMMC
• 12:25
  and
• 12:26
  HIPAA
• 12:26
  compliance
• 12:26
  variants
• 12:27
  from
• 12:27
  the
• 12:27
  same
• 12:27
  extracted
• 12:28
  corpus.
• 12:28
  A
• 12:28
  smoke
• 12:29
  eval
• 12:29
  suite
• 12:29
  validates
• 12:30
  the
• 12:30
  deployed
• 12:30
  Ollama
• 12:30
  model
• 12:31
  against
• 12:31
  known
• 12:31
  NIST
• 12:32
  questions.
• 12:33
  For
• 12:33
  anyone
• 12:33
  building
• 12:33
  domain-specific
• 12:34
  models
• 12:35
  for
• 12:35
  regulated
• 12:35
  industries,
• 12:36
  this
• 12:36
  is
• 12:36
  a
• 12:36
  concrete
• 12:37
  reference
• 12:37
  architecture
• 12:37
  —
• 12:38
  not
• 12:38
  a
• 12:38
  blog
• 12:39
  post,
• 12:39
  but
• 12:39
  working
• 12:40
  code
• 12:40
  with
• 12:40
  published
• 12:41
  artifacts
• 12:41
  on
• 12:41
  Hugging
• 12:41
  Face.
• 12:42
  The
• 12:42
  Jackrong
• 12:43
  LLM
• 12:43
  Fine-Tuning
• 12:44
  Guide
• 12:44
  takes
• 12:45
  the
• 12:45
  complementary
• 12:45
  approach
• 12:46
  of
• 12:46
  making
• 12:46
  the
• 12:46
  general
• 12:46
  process
• 12:47
  Designed
• 12:48
  as
• 12:48
  a
• 12:49
  "zero
• 12:49
  to
• 12:49
  one"
• 12:49
  learning
• 12:50
  platform,
• 12:50
  it
• 12:50
  provides
• 12:51
  interactive
• 12:51
  Colab
• 12:52
  and
• 12:52
  Kaggle
• 12:52
  notebooks
• 12:53
  covering
• 12:53
  SFT
• 12:54
  and
• 12:54
  GRPO
• 12:54
  reinforcement
• 12:55
  learning
• 12:55
  pipelines
• 12:56
  across
• 12:56
  Qwen3.5,
• 12:58
  Llama3.2,
• 13:00
  and
• 13:00
  upcoming
• 13:01
  Gemma
• 13:01
  4
• 13:01
  and
• 13:01
  DeepSeek
• 13:02
  architectures.
• 13:03
  The
• 13:03
  project
• 13:04
  leverages
• 13:04
  Unsloth
• 13:05
  for
• 13:05
  resource-efficient
• 13:06
  training
• 13:06
  within
• 13:06
  single-GPU
• 13:07
  constraints,
• 13:08
  and
• 13:08
  ships
• 13:08
  24
• 13:09
  curated
• 13:09
  distillation
• 13:10
  datasets
• 13:10
  from
• 13:11
  flagship
• 13:11
  models
• 13:12
  including
• 13:12
  DeepSeek-V3.2,
• 13:14
  Qwen3-235B,
• 13:16
  GLM-4.7,
• 13:17
  and
• 13:18
  GPT-OSS-120B.
• 13:21
  The
• 13:21
  Qwen3.5
• 13:24
  fine-tunes
• 13:25
  have
• 13:25
  crossed
• 13:25
  a
• 13:25
  million
• 13:25
  downloads
• 13:26
  on
• 13:26
  Hugging
• 13:26
  Face
• 13:26
  —
• 13:27
  a
• 13:27
  quiet
• 13:28
  signal
• 13:28
  that
• 13:28
  the
• 13:28
  democratization
• 13:29
  of
• 13:29
  fine-tuning
• 13:30
  is
• 13:30
  no
• 13:30
  longer
• 13:30
  aspirational.
• 13:32
  Rounding
• 13:32
  out
• 13:32
  the
• 13:32
  open-source
• 13:33
  resource
• 13:33
  landscape,
• 13:34
  the
• 13:34
  Open
• 13:35
  Generative
• 13:35
  AI
• 13:35
  project
• 13:36
  aggregates
• 13:36
  200+
• 13:37
  state-of-the-art
• 13:38
  image,
• 13:38
  video,
• 13:39
  lip
• 13:39
  sync,
• 13:40
  and
• 13:40
  cinema
• 13:40
  generation
• 13:41
  models
• 13:41
  into
• 13:42
  a
• 13:42
  single
• 13:42
  desktop
• 13:42
  application
• 13:43
  and
• 13:43
  web
• 13:43
  interface
• 13:44
  —
• 13:44
  explicitly
• 13:45
  uncensored
• 13:46
  and
• 13:46
  self-hostable.
• 13:47
  It
• 13:47
  supports
• 13:48
  two
• 13:48
  local
• 13:48
  inference
• 13:49
  engines:
• 13:50
  sd.cpp
• 13:51
  (bundled,
• 13:52
  runs
• 13:53
  on
• 13:53
  Metal
• 13:53
  GPU
• 13:54
  for
• 13:54
  Apple
• 13:54
  Silicon)
• 13:55
  and
• 13:55
  Wan2GP
• 13:55
  (BYO
• 13:56
  server
• 13:56
  for
• 13:57
  CUDA/ROCm
• 14:01
  video
• 14:01
  models).
• 14:02
  The
• 14:02
  "Generative
• 14:02
  Media
• 14:03
  Skills"
• 14:03
  companion
• 14:04
  library
• 14:04
  lets
• 14:04
  coding
• 14:04
  agents
• 14:05
  like
• 14:05
  Claude
• 14:05
  Code
• 14:06
  drive
• 14:06
  the
• 14:06
  entire
• 14:07
  media
• 14:07
  pipeline
• 14:07
  from
• 14:07
  terminal.
• 14:11
  Anthropic
• 14:11
  published
• 14:11
  a
• 14:12
  detailed
• 14:12
  postmortem
• 14:13
  tracing
• 14:13
  recent
• 14:14
  Claude
• 14:14
  Code
• 14:14
  quality
• 14:14
  complaints
• 14:15
  to
• 14:15
  three
• 14:15
  separate
• 14:16
  changes.
• 14:17
  First,
• 14:17
  on
• 14:17
  March
• 14:17
  4,
• 14:18
  the
• 14:18
  default
• 14:18
  reasoning
• 14:18
  effort
• 14:19
  was
• 14:19
  dropped
• 14:19
  from
• 14:19
  high
• 14:20
  to
• 14:20
  medium
• 14:20
  to
• 14:20
  reduce
• 14:21
  latency
• 14:21
  —
• 14:22
  the
• 14:22
  wrong
• 14:22
  tradeoff,
• 14:23
  reverted
• 14:23
  April
• 14:23
  7
• 14:24
  after
• 14:24
  users
• 14:24
  reported
• 14:25
  noticeably
• 14:25
  worse
• 14:25
  intelligence.
• 14:26
  Second,
• 14:27
  on
• 14:27
  March
• 14:27
  26,
• 14:28
  a
• 14:28
  caching
• 14:28
  optimization
• 14:29
  intended
• 14:29
  to
• 14:30
  clear
• 14:30
  old
• 14:30
  thinking
• 14:30
  from
• 14:30
  stale
• 14:31
  sessions
• 14:31
  had
• 14:31
  a
• 14:31
  bug
• 14:32
  that
• 14:32
  cleared
• 14:32
  thinking
• 14:32
  on
• 14:32
  every
• 14:33
  subsequent
• 14:33
  turn
• 14:33
  for
• 14:34
  the
• 14:34
  rest
• 14:34
  of
• 14:34
  the
• 14:34
  session,
• 14:35
  making
• 14:35
  Claude
• 14:36
  "increasingly
• 14:36
  without
• 14:36
  memory
• 14:37
  of
• 14:37
  why
• 14:37
  it
• 14:37
  had
• 14:37
  chosen
• 14:38
  to
• 14:38
  do
• 14:38
  what
• 14:38
  it
• 14:38
  was
• 14:38
  doing."
• 14:39
  This
• 14:39
  compounded
• 14:40
  into
• 14:40
  cache
• 14:41
  misses
• 14:41
  that
• 14:41
  also
• 14:41
  drained
• 14:42
  usage
• 14:42
  limits
• 14:42
  faster.
• 14:43
  Third,
• 14:44
  on
• 14:44
  April
• 14:44
  16,
• 14:45
  a
• 14:45
  system
• 14:45
  prompt
• 14:46
  instruction
• 14:46
  —
• 14:47
  "keep
• 14:47
  text
• 14:48
  between
• 14:48
  tool
• 14:48
  calls
• 14:49
  to
• 14:49
  ≤25
• 14:49
  words"
• 14:49
  —
• 14:50
  hurt
• 14:50
  coding
• 14:51
  quality
• 14:51
  by
• 14:51
  3%
• 14:52
  in
• 14:52
  ablation
• 14:52
  testing
• 14:52
  and
• 14:53
  was
• 14:53
  reverted
• 14:53
  April
• 14:53
  20.
• 14:55
  The
• 14:55
  postmortem's
• 14:55
  most
• 14:55
  telling
• 14:56
  admission:
• 14:57
  Opus
• 14:57
  4.7
• 14:58
  found
• 14:58
  the
• 14:58
  caching
• 14:58
  bug
• 14:59
  when
• 14:59
  provided
• 14:59
  sufficient
• 14:59
  repository
• 15:00
  context,
• 15:01
  while
• 15:01
  Opus
• 15:01
  4.6
• 15:02
  didn't.
• 15:03
  Anthropic
• 15:03
  is
• 15:03
  resetting
• 15:04
  usage
• 15:04
  limits
• 15:04
  for
• 15:04
  all
• 15:04
  subscribers
• 15:05
  and
• 15:05
  committing
• 15:06
  to
• 15:06
  broader
• 15:06
  eval
• 15:07
  suites,
• 15:07
  soak
• 15:08
  periods,
• 15:08
  and
• 15:09
  gradual
• 15:09
  rollouts
• 15:09
  for
• 15:09
  system
• 15:10
  prompt
• 15:10
  changes.
• 15:11
  The
• 15:11
  trust
• 15:11
  question
• 15:12
  around
• 15:12
  AI-generated
• 15:13
  code
• 15:13
  surfaced
• 15:13
  from
• 15:13
  a
• 15:13
  different
• 15:14
  angle
• 15:14
  with
• 15:14
  the
• 15:14
  MeshCore
• 15:15
  The
• 15:16
  MeshCore
• 15:16
  mesh
• 15:16
  networking
• 15:17
  project
• 15:17
  —
• 15:18
  38,000+
• 15:19
  nodes,
• 15:20
  100,000+
• 15:21
  active
• 15:22
  mobile
• 15:22
  users
• 15:22
  —
• 15:23
  fractured
• 15:23
  after
• 15:23
  team
• 15:24
  member
• 15:24
  Andy
• 15:24
  Kirby
• 15:24
  began
• 15:25
  extensively
• 15:25
  using
• 15:25
  Claude
• 15:26
  Code
• 15:26
  to
• 15:26
  rebuild
• 15:26
  core
• 15:27
  components
• 15:27
  (companion
• 15:28
  app,
• 15:28
  web
• 15:28
  flasher,
• 15:29
  config
• 15:29
  tools)
• 15:29
  without
• 15:30
  disclosing
• 15:30
  the
• 15:30
  AI
• 15:30
  origin.
• 15:31
  A
• 15:31
  Discord
• 15:32
  poll
• 15:32
  showed
• 15:32
  the
• 15:32
  community
• 15:33
  was
• 15:33
  wary
• 15:33
  of
• 15:33
  AI-generated
• 15:34
  code.
• 15:35
  The
• 15:35
  breaking
• 15:36
  point
• 15:36
  wasn't
• 15:36
  the
• 15:36
  AI
• 15:37
  usage
• 15:37
  itself
• 15:37
  but
• 15:38
  the
• 15:38
  combination
• 15:38
  of
• 15:38
  secrecy
• 15:39
  and
• 15:39
  a
• 15:39
  unilateral
• 15:40
  trademark
• 15:40
  filing.
• 15:41
  The
• 15:42
  remaining
• 15:42
  team
• 15:42
  —
• 15:43
  the
• 15:43
  original
• 15:44
  firmware
• 15:44
  developers
• 15:44
  —
• 15:45
  launched
• 15:46
  a
• 15:46
  new
• 15:46
  site
• 15:46
  and
• 15:46
  Discord,
• 15:47
  emphasizing
• 15:48
  that
• 15:48
  their
• 15:48
  releases
• 15:48
  are
• 15:48
  "hand-crafted,
• 15:49
  by
• 15:49
  humans."
• 15:51
  It's
• 15:51
  a
• 15:51
  case
• 15:51
  study
• 15:51
  in
• 15:51
  how
• 15:52
  AI-assisted
• 15:52
  development
• 15:53
  collides
• 15:53
  with
• 15:53
  open-source
• 15:54
  community
• 15:54
  norms
• 15:54
  around
• 15:55
  attribution
• 15:55
  and
• 15:55
  trust.
• 15:56
  On
• 15:56
  the
• 15:56
  constructive
• 15:57
  side,
• 15:58
  MemPalace
• 15:58
  claims
• 15:58
  the
• 15:58
  highest-scoring
• 15:59
  AI
• 15:59
  memory
• 16:00
  system
• 16:00
  ever
• 16:00
  benchmarked:
• 16:01
  96.6%
• 16:02
  R@5
• 16:03
  on
• 16:03
  LongMemEval
• 16:04
  with
• 16:04
  zero
• 16:05
  API
• 16:05
  calls,
• 16:06
  no
• 16:06
  cloud,
• 16:06
  and
• 16:07
  no
• 16:07
  LLM
• 16:07
  at
• 16:07
  any
• 16:08
  stage.
• 16:08
  The
• 16:09
  architecture
• 16:09
  stores
• 16:10
  conversation
• 16:10
  history
• 16:11
  as
• 16:11
  verbatim
• 16:11
  text
• 16:12
  (no
• 16:12
  summarization
• 16:13
  or
• 16:13
  paraphrasing)
• 16:14
  in
• 16:14
  a
• 16:14
  structured
• 16:14
  index
• 16:15
  —
• 16:15
  "wings"
• 16:16
  for
• 16:16
  people/projects,
• 16:17
  "rooms"
• 16:18
  for
• 16:18
  topics,
• 16:19
  "drawers"
• 16:19
  for
• 16:20
  original
• 16:20
  content
• 16:20
  —
• 16:21
  with
• 16:21
  semantic
• 16:21
  search
• 16:22
  scoped
• 16:22
  to
• 16:22
  this
• 16:22
  hierarchy.
• 16:24
  A
• 16:24
  hybrid
• 16:24
  pipeline
• 16:24
  adds
• 16:25
  keyword
• 16:25
  and
• 16:25
  temporal-proximity
• 16:26
  boosting
• 16:27
  to
• 16:27
  reach
• 16:27
  98.4%
• 16:29
  on
• 16:29
  a
• 16:29
  held-out
• 16:29
  set,
• 16:30
  and
• 16:30
  an
• 16:30
  LLM
• 16:31
  rerank
• 16:31
  path
• 16:32
  pushes
• 16:32
  past
• 16:32
  99%.
• 16:34
  The
• 16:34
  system
• 16:34
  includes
• 16:34
  a
• 16:34
  temporal
• 16:35
  entity-relationship
• 16:36
  knowledge
• 16:36
  graph
• 16:37
  backed
• 16:37
  by
• 16:37
  SQLite
• 16:38
  and
• 16:38
  29
• 16:38
  MCP
• 16:39
  tools.
• 16:40
  AgentBox
• 16:41
  from
• 16:41
  DreamLab-AI
• 16:41
  takes
• 16:42
  agent
• 16:42
  infrastructure
• 16:43
  in
• 16:43
  a
• 16:43
  different
• 16:43
  direction:
• 16:44
  a
• 16:44
  manifest-driven,
• 16:45
  reproducible
• 16:46
  runtime
• 16:46
  where
• 16:46
  a
• 16:46
  single
• 16:47
  agentbox.toml
• 16:49
  drives
• 16:49
  the
• 16:50
  Nix
• 16:50
  package
• 16:50
  graph,
• 16:51
  generated
• 16:51
  runtime
• 16:52
  image,
• 16:52
  compose
• 16:52
  file,
• 16:53
  supervisor
• 16:53
  config,
• 16:54
  and
• 16:54
  health/readiness
• 16:55
  contract.
• 16:56
  The
• 16:56
  design
• 16:57
  philosophy
• 16:57
  is
• 16:57
  the
• 16:57
  inverse
• 16:58
  of
• 16:58
  most
• 16:58
  agent
• 16:58
  containers
• 16:59
  —
• 16:59
  immutable
• 17:00
  bootstrap
• 17:00
  (no
• 17:01
  npm
• 17:01
  install
• 17:01
  at
• 17:01
  startup),
• 17:02
  hardened
• 17:02
  baseline
• 17:03
  (non-root,
• 17:03
  read-only
• 17:04
  filesystem,
• 17:05
  all
• 17:05
  capabilities
• 17:06
  dropped),
• 17:06
  and
• 17:06
  a
• 17:06
  sovereign
• 17:07
  data
• 17:07
  stack
• 17:07
  built
• 17:08
  on
• 17:08
  Solid
• 17:08
  pods,
• 17:09
  did:nostr
• 17:11
  identity,
• 17:11
  and
• 17:12
  an
• 17:12
  embedded
• 17:12
  Nostr
• 17:12
  relay.
• 17:14
  It
• 17:14
  supports
• 17:14
  Claude,
• 17:15
  Codex,
• 17:16
  Gemini,
• 17:16
  and
• 17:16
  claude-flow
• 17:17
  as
• 17:17
  agent
• 17:17
  backends
• 17:18
  with
• 17:18
  built-in
• 17:18
  MCP
• 17:19
  service
• 17:19
  support.
• 17:20
  Two
• 17:20
  Claude
• 17:21
  Code
• 17:21
  ecosystem
• 17:22
  projects
• 17:22
  show
• 17:22
  where
• 17:22
  multi-agent
• 17:23
  tooling
• 17:23
  is
• 17:24
  heading.
• 17:25
  Design
• 17:25
  Council
• 17:25
  spawns
• 17:26
  parallel
• 17:26
  Claude
• 17:26
  agents
• 17:27
  with
• 17:27
  independent
• 17:27
  contexts
• 17:28
  —
• 17:29
  a
• 17:29
  security
• 17:29
  engineer,
• 17:30
  principal
• 17:31
  engineer,
• 17:31
  performance
• 17:32
  engineer,
• 17:32
  etc.
• 17:34
  —
• 17:34
  to
• 17:34
  argue
• 17:34
  cross-cutting
• 17:35
  design
• 17:35
  decisions
• 17:35
  from
• 17:36
  genuinely
• 17:36
  separate
• 17:36
  vantage
• 17:37
  points,
• 17:37
  then
• 17:38
  arbitrates
• 17:38
  disagreements
• 17:39
  with
• 17:39
  written
• 17:39
  rationale.
• 17:40
  The
• 17:40
  token
• 17:41
  cost
• 17:41
  is
• 17:41
  10-20x
• 17:42
  a
• 17:42
  single-context
• 17:43
  review,
• 17:43
  but
• 17:44
  the
• 17:44
  structural
• 17:44
  independence
• 17:45
  produces
• 17:45
  disagreements
• 17:46
  that
• 17:46
  sequential
• 17:46
  turns
• 17:47
  cannot.
• 17:48
  Claude
• 17:48
  Code
• 17:49
  Game
• 17:49
  Studios
• 17:49
  goes
• 17:49
  further,
• 17:50
  packaging
• 17:51
  49
• 17:51
  specialized
• 17:52
  agents
• 17:52
  organized
• 17:53
  into
• 17:53
  a
• 17:53
  three-tier
• 17:54
  studio
• 17:54
  hierarchy
• 17:54
  (directors
• 17:55
  on
• 17:55
  Opus,
• 17:56
  department
• 17:56
  leads
• 17:56
  on
• 17:57
  Sonnet,
• 17:57
  specialists
• 17:58
  on
• 17:58
  Sonnet/Haiku)
• 17:59
  with
• 17:59
  72
• 18:00
  slash
• 18:00
  commands,
• 18:01
  12
• 18:01
  automated
• 18:02
  hooks,
• 18:02
  and
• 18:03
  path-scoped
• 18:03
  coding
• 18:03
  standards.
• 18:05
  It's
• 18:05
  collaborative,
• 18:06
  not
• 18:06
  autonomous
• 18:06
  —
• 18:07
  agents
• 18:08
  present
• 18:08
  options,
• 18:08
  you
• 18:09
  decide
• 18:09
  —
• 18:10
  but
• 18:10
  the
• 18:10
  structural
• 18:10
  ambition
• 18:11
  of
• 18:11
  treating
• 18:11
  a
• 18:11
  Claude
• 18:11
  Code
• 18:12
  session
• 18:12
  as
• 18:12
  a
• 18:12
  full
• 18:13
  game
• 18:13
  development
• 18:13
  organization
• 18:14
  with
• 18:14
  escalation
• 18:14
  paths
• 18:15
  and
• 18:15
  quality
• 18:15
  gates
• 18:16
  is
• 18:16
  a
• 18:16
  notable
• 18:16
  step
• 18:17
  beyond
• 18:17
  "chat
• 18:17
  with
• 18:18
  your
• 18:18
  code."
• 18:23
  A
• 18:23
  report
• 18:24
  synthesized
• 18:24
  in
• 18:24
  the
• 18:24
  Unhyped
• 18:25
  AI
• 18:25
  newsletter
• 18:26
  makes
• 18:26
  a
• 18:26
  deceptively
• 18:27
  simple
• 18:27
  observation:
• 18:28
  the
• 18:29
  same
• 18:29
  AI
• 18:29
  use
• 18:29
  case
• 18:30
  moves
• 18:30
  quickly
• 18:31
  in
• 18:31
  one
• 18:31
  organization
• 18:32
  and
• 18:32
  stalls
• 18:32
  in
• 18:33
  another,
• 18:34
  and
• 18:34
  the
• 18:34
  difference
• 18:34
  is
• 18:34
  almost
• 18:35
  never
• 18:35
  the
• 18:35
  technology.
• 18:36
  Across
• 18:37
  the
• 18:37
  cases
• 18:37
  studied,
• 18:38
  77%
• 18:39
  of
• 18:39
  the
• 18:39
  hardest
• 18:40
  challenges
• 18:40
  sat
• 18:41
  in
• 18:41
  surrounding
• 18:41
  systems
• 18:41
  —
• 18:42
  change
• 18:43
  management,
• 18:44
  process
• 18:44
  redesign,
• 18:45
  data
• 18:45
  quality,
• 18:46
  adoption
• 18:46
  —
• 18:47
  not
• 18:47
  the
• 18:47
  technical
• 18:48
  layer.
• 18:49
  In
• 18:49
  61%
• 18:50
  of
• 18:50
  successful
• 18:51
  cases,
• 18:52
  there
• 18:52
  had
• 18:52
  already
• 18:52
  been
• 18:52
  a
• 18:53
  failed
• 18:53
  attempt
• 18:53
  before
• 18:53
  the
• 18:54
  one
• 18:54
  that
• 18:54
  worked.
• 18:55
  The
• 18:55
  practical
• 18:56
  upshot
• 18:56
  for
• 18:56
  boards:
• 18:57
  stop
• 18:57
  asking
• 18:58
  whether
• 18:58
  the
• 18:58
  technology
• 18:59
  is
• 18:59
  impressive
• 19:00
  and
• 19:00
  start
• 19:00
  asking
• 19:00
  five
• 19:01
  grounded
• 19:01
  questions
• 19:01
  —
• 19:02
  what
• 19:03
  changed
• 19:03
  when
• 19:03
  this
• 19:03
  met
• 19:03
  real
• 19:04
  work,
• 19:04
  what
• 19:04
  new
• 19:04
  capability
• 19:05
  did
• 19:05
  it
• 19:05
  unlock,
• 19:06
  how
• 19:06
  quickly
• 19:06
  did
• 19:07
  the
• 19:07
  organization
• 19:07
  learn
• 19:07
  from
• 19:08
  failure,
• 19:08
  did
• 19:09
  the
• 19:09
  work
• 19:09
  get
• 19:09
  redesigned
• 19:10
  or
• 19:10
  just
• 19:10
  layered
• 19:10
  with
• 19:10
  AI,
• 19:11
  and
• 19:11
  is
• 19:11
  the
• 19:11
  organization
• 19:12
  getting
• 19:12
  stronger
• 19:12
  through
• 19:13
  using
• 19:13
  it.
• 19:14
  The
• 19:14
  on-device
• 19:14
  AI
• 19:15
  thesis
• 19:15
  got
• 19:15
  its
• 19:15
  most
• 19:16
  comprehensive
• 19:16
  articulation
• 19:17
  yet
• 19:17
  in
• 19:17
  a
• 19:17
  video
• 19:18
  analysis
• 19:18
  of
• 19:18
  Apple's
• 19:18
  CEO
• 19:19
  transition.
• 19:20
  With
• 19:20
  Tim
• 19:21
  Cook
• 19:21
  stepping
• 19:21
  down
• 19:22
  and
• 19:22
  hardware
• 19:22
  engineer
• 19:22
  John
• 19:23
  Turnis
• 19:23
  taking
• 19:23
  over
• 19:24
  alongside
• 19:24
  chip
• 19:24
  designer
• 19:25
  John
• 19:25
  Suji
• 19:25
  as
• 19:25
  chief
• 19:26
  hardware
• 19:26
  officer,
• 19:27
  the
• 19:27
  argument
• 19:28
  is
• 19:28
  that
• 19:28
  Apple
• 19:28
  is
• 19:28
  conceding
• 19:29
  the
• 19:29
  cloud
• 19:29
  AI
• 19:29
  velocity
• 19:30
  race
• 19:30
  and
• 19:30
  betting
• 19:31
  on
• 19:31
  a
• 19:31
  fundamentally
• 19:31
  different
• 19:32
  game.
• 19:33
  The
• 19:33
  economic
• 19:33
  core:
• 19:34
  cloud
• 19:35
  inference
• 19:35
  has
• 19:35
  variable
• 19:36
  cost
• 19:36
  that
• 19:36
  someone
• 19:36
  pays
• 19:37
  every
• 19:37
  time
• 19:37
  you
• 19:37
  ask
• 19:38
  a
• 19:38
  question,
• 19:39
  and
• 19:39
  every
• 19:39
  major
• 19:40
  frontier
• 19:40
  lab
• 19:40
  is
• 19:40
  losing
• 19:41
  money
• 19:41
  on
• 19:41
  top-tier
• 19:42
  consumer
• 19:42
  subscriptions.
• 19:44
  On-device
• 19:44
  inference
• 19:44
  has
• 19:45
  fixed
• 19:45
  cost
• 19:45
  —
• 19:46
  you
• 19:46
  paid
• 19:46
  for
• 19:46
  the
• 19:46
  chip
• 19:47
  when
• 19:47
  you
• 19:47
  bought
• 19:47
  the
• 19:47
  phone,
• 19:48
  and
• 19:48
  a
• 19:48
  thousand
• 19:48
  queries
• 19:49
  cost
• 19:49
  the
• 19:49
  same
• 19:49
  as
• 19:49
  one.
• 19:50
  The
• 19:51
  analysis
• 19:51
  identifies
• 19:52
  a
• 19:52
  specific
• 19:52
  underserved
• 19:53
  market:
• 19:54
  law
• 19:54
  firms,
• 19:55
  medical
• 19:55
  practices,
• 19:56
  accounting
• 19:57
  firms,
• 19:57
  and
• 19:57
  other
• 19:57
  professionals
• 19:58
  whose
• 19:58
  data
• 19:59
  confidentiality
• 20:00
  requirements
• 20:00
  (attorney-client
• 20:01
  privilege,
• 20:02
  HIPAA,
• 20:03
  fiduciary
• 20:03
  duty)
• 20:03
  structurally
• 20:04
  prevent
• 20:04
  cloud
• 20:05
  AI
• 20:05
  adoption.
• 20:06
  These
• 20:07
  firms
• 20:07
  are
• 20:07
  reportedly
• 20:08
  buying
• 20:08
  clusters
• 20:08
  of
• 20:08
  Mac
• 20:09
  Minis
• 20:09
  to
• 20:09
  run
• 20:10
  local
• 20:10
  inference,
• 20:11
  improvising
• 20:12
  their
• 20:12
  own
• 20:12
  orchestration
• 20:13
  because
• 20:14
  Apple
• 20:14
  hasn't
• 20:14
  built
• 20:14
  the
• 20:15
  enterprise
• 20:15
  stack
• 20:15
  they
• 20:16
  need
• 20:16
  —
• 20:17
  no
• 20:17
  rackable
• 20:17
  form
• 20:17
  factor,
• 20:18
  no
• 20:18
  clustering
• 20:19
  software,
• 20:20
  no
• 20:20
  HIPAA
• 20:20
  BAAs.
• 20:21
  The
• 20:21
  historical
• 20:22
  parallel
• 20:22
  to
• 20:23
  the
• 20:23
  Apple
• 20:23
  II
• 20:23
  disrupting
• 20:24
  the
• 20:24
  mainframe
• 20:24
  time-sharing
• 20:25
  model
• 20:25
  is
• 20:25
  apt:
• 20:26
  a
• 20:26
  metered
• 20:27
  service
• 20:27
  model
• 20:28
  where
• 20:28
  heavy
• 20:28
  users
• 20:28
  are
• 20:29
  a
• 20:29
  cost
• 20:29
  problem,
• 20:30
  versus
• 20:30
  an
• 20:30
  ownership
• 20:31
  model
• 20:31
  where
• 20:31
  marginal
• 20:32
  cost
• 20:32
  drops
• 20:33
  to
• 20:33
  near
• 20:33
  zero.
• 20:34
  The
• 20:34
  model
• 20:35
  alignment
• 20:35
  debate
• 20:35
  surfaced
• 20:36
  through
• 20:36
  a
• 20:36
  security
• 20:37
  researcher's
• 20:37
  account
• 20:38
  of
• 20:38
  starting
• 20:38
  to
• 20:38
  uncensor
• 20:39
  models
• 20:39
  for
• 20:39
  vulnerability
• 20:40
  research
• 20:41
  and
• 20:41
  discovering
• 20:41
  they
• 20:42
  reason
• 20:42
  better
• 20:42
  about
• 20:43
  facts
• 20:43
  without
• 20:43
  safety
• 20:44
  filters.
• 20:45
  The
• 20:45
  core
• 20:45
  "debiasing"
• 20:47
  in
• 20:47
  practice
• 20:48
  almost
• 20:49
  always
• 20:49
  means
• 20:49
  aligning
• 20:50
  outputs
• 20:50
  to
• 20:50
  the
• 20:50
  ideological
• 20:51
  priors
• 20:52
  of
• 20:52
  the
• 20:52
  debiasers,
• 20:53
  and
• 20:53
  asymmetric
• 20:54
  application
• 20:54
  —
• 20:55
  where
• 20:55
  mitigation
• 20:56
  consistently
• 20:56
  pushes
• 20:57
  in
• 20:57
  one
• 20:57
  direction
• 20:58
  on
• 20:58
  contested
• 20:58
  questions
• 20:59
  —
• 21:00
  is
• 21:00
  value
• 21:00
  imposition,
• 21:01
  not
• 21:02
  bias
• 21:02
  reduction.
• 21:03
  The
• 21:04
  technically
• 21:04
  serious
• 21:05
  objection
• 21:05
  is
• 21:05
  that
• 21:06
  safety
• 21:06
  training
• 21:06
  degrades
• 21:07
  capability:
• 21:08
  "a
• 21:08
  model
• 21:08
  trained
• 21:09
  to
• 21:09
  flinch
• 21:09
  or
• 21:09
  equivocate
• 21:10
  on
• 21:10
  factual
• 21:10
  questions
• 21:11
  because
• 21:12
  the
• 21:12
  answers
• 21:12
  are
• 21:12
  politically
• 21:13
  inconvenient
• 21:14
  is
• 21:14
  a
• 21:14
  worse
• 21:14
  reasoning
• 21:15
  engine,
• 21:16
  full
• 21:16
  stop.
• 21:17
  Sycophancy
• 21:18
  and
• 21:18
  evasion
• 21:18
  in
• 21:19
  one
• 21:19
  domain
• 21:19
  leak
• 21:19
  into
• 21:20
  all
• 21:20
  of
• 21:20
  them."
• 21:21
  The
• 21:21
  practical
• 21:21
  recommendation:
• 21:23
  open,
• 21:23
  uncensored,
• 21:24
  and
• 21:24
  local
• 21:24
  models
• 21:25
  like
• 21:25
  Ministral,
• 21:26
  Gemma
• 21:26
  4,
• 21:27
  or
• 21:27
  Qwen