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

Transcript

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

• 0:00
  The
• 0:00
  open-weight
• 0:00
  model
• 0:01
  community
• 0:01
  has
• 0:01
  turned
• 0:02
  VRAM
• 0:02
  optimization
• 0:03
  into
• 0:03
  a
• 0:03
  competitive
• 0:04
  sport,
• 0:04
  and
• 0:04
  the
• 0:04
  Qwen
• 0:05
  3.6
• 0:05
  series
• 0:06
  is
• 0:06
  the
• 0:06
  current
• 0:06
  arena.
• 0:07
  A
• 0:07
  detailed
• 0:07
  breakdown
• 0:08
  of
• 0:08
  running
• 0:08
  Qwen3.6-27B
• 0:11
  at
• 0:12
  IQ4
• 0:12
  XS
• 0:12
  quantization
• 0:13
  reveals
• 0:13
  that
• 0:14
  llama.cpp
• 0:15
  commit
• 0:15
  1dab5f5a44
• 0:17
  hardcodes
• 0:17
  the
• 0:18
  attn
• 0:18
  qkv
• 0:18
  tensor
• 0:19
  to
• 0:19
  a
• 0:19
  Q5
• 0:19
  K
• 0:19
  minimum,
• 0:20
  inflating
• 0:21
  the
• 0:21
  model
• 0:21
  from
• 0:21
  14.7
• 0:22
  GB
• 0:22
  to
• 0:22
  15.1
• 0:23
  GB
• 0:23
  with
• 0:24
  no
• 0:24
  meaningful
• 0:24
  perplexity
• 0:25
  gain.
• 0:26
  The
• 0:26
  fix
• 0:26
  is
• 0:26
  straightforward
• 0:26
  —
• 0:27
  patch
• 0:28
  the
• 0:28
  importance
• 0:28
  matrix
• 0:28
  to
• 0:28
  let
• 0:29
  IQ4
• 0:29
  XS
• 0:30
  apply
• 0:30
  uniformly
• 0:30
  —
• 0:31
  and
• 0:31
  the
• 0:31
  reward
• 0:32
  is
• 0:32
  110K
• 0:32
  context
• 0:33
  on
• 0:33
  16
• 0:34
  GB
• 0:34
  VRAM
• 0:34
  using
• 0:35
  turbo3
• 0:35
  KV
• 0:35
  cache
• 0:36
  quantization.
• 0:37
  That
• 0:37
  is
• 0:37
  a
• 0:37
  27-billion
• 0:38
  parameter
• 0:39
  model
• 0:39
  fitting
• 0:39
  entirely
• 0:40
  in
• 0:40
  a
• 0:40
  single
• 0:40
  consumer
• 0:41
  GPU's
• 0:41
  memory
• 0:41
  with
• 0:41
  a
• 0:41
  context
• 0:42
  window
• 0:42
  that
• 0:42
  rivals
• 0:43
  many
• 0:43
  API
• 0:43
  providers.
• 0:45
  Speaking
• 0:45
  of
• 0:45
  KV
• 0:45
  cache
• 0:46
  tricks,
• 0:46
  Google's
• 0:47
  TurboQuant
• 0:47
  paper
• 0:47
  generated
• 0:48
  excitement
• 0:49
  but
• 0:49
  the
• 0:49
  community
• 0:49
  reception
• 0:50
  has
• 0:50
  been
• 0:50
  lukewarm.
• 0:51
  The
• 0:52
  technique
• 0:52
  is
• 0:52
  not
• 0:52
  officially
• 0:53
  merged
• 0:53
  into
• 0:53
  llama.cpp,
• 0:55
  and
• 0:55
  forks
• 0:55
  that
• 0:56
  implement
• 0:56
  it
• 0:56
  show
• 0:56
  marginal
• 0:57
  improvement
• 0:57
  over
• 0:57
  the
• 0:57
  existing
• 0:58
  q4
• 0:58
  0
• 0:59
  KV
• 0:59
  cache
• 0:59
  quantization
• 1:00
  that
• 1:00
  has
• 1:00
  been
• 1:00
  available
• 1:01
  for
• 1:01
  months.
• 1:02
  Several
• 1:02
  community
• 1:03
  members
• 1:03
  tested
• 1:03
  the
• 1:04
  available
• 1:04
  forks
• 1:05
  and
• 1:05
  found
• 1:05
  the
• 1:05
  memory
• 1:05
  savings
• 1:06
  negligible
• 1:06
  for
• 1:06
  the
• 1:07
  added
• 1:07
  complexity
• 1:07
  —
• 1:08
  a
• 1:08
  few
• 1:08
  hundred
• 1:09
  megabytes
• 1:09
  at
• 1:09
  best
• 1:09
  on
• 1:10
  typical
• 1:10
  workloads,
• 1:11
  which
• 1:11
  matters
• 1:11
  far
• 1:11
  less
• 1:12
  than
• 1:12
  the
• 1:12
  multi-gigabyte
• 1:13
  savings
• 1:13
  from
• 1:13
  proper
• 1:14
  weight
• 1:14
  quantization.
• 1:15
  The
• 1:15
  consensus
• 1:16
  from
• 1:16
  practitioners:
• 1:17
  overhyped
• 1:17
  for
• 1:17
  real-world
• 1:18
  use,
• 1:19
  at
• 1:19
  least
• 1:19
  in
• 1:19
  its
• 1:19
  current
• 1:19
  form.
• 1:21
  The
• 1:21
  gap
• 1:21
  between
• 1:21
  an
• 1:21
  interesting
• 1:22
  research
• 1:22
  paper
• 1:23
  and
• 1:23
  a
• 1:23
  practical
• 1:23
  llama-server
• 1:24
  flag
• 1:24
  remains
• 1:25
  wide.
• 1:26
  The
• 1:26
  model
• 1:26
  leaderboard
• 1:27
  itself
• 1:27
  keeps
• 1:27
  shuffling.
• 1:29
  Updated
• 1:29
  benchmarks
• 1:30
  show
• 1:30
  Gemma
• 1:30
  4
• 1:30
  pulling
• 1:31
  ahead
• 1:31
  of
• 1:31
  Qwen
• 1:31
  3.5
• 1:32
  on
• 1:32
  several
• 1:32
  tasks,
• 1:33
  but
• 1:33
  the
• 1:33
  more
• 1:34
  interesting
• 1:34
  development
• 1:35
  is
• 1:35
  the
• 1:35
  pairing
• 1:35
  of
• 1:35
  Qwen
• 1:35
  3.6
• 1:36
  27B
• 1:37
  with
• 1:37
  MiniMax
• 1:38
  M2.7
• 1:40
  as
• 1:40
  a
• 1:40
  local
• 1:40
  coding
• 1:41
  stack
• 1:41
  via
• 1:41
  OpenCode.
• 1:42
  Users
• 1:43
  report
• 1:43
  this
• 1:43
  combination
• 1:44
  as
• 1:44
  a
• 1:44
  credible
• 1:44
  replacement
• 1:45
  for
• 1:45
  Claude
• 1:45
  Code
• 1:45
  on
• 1:45
  tasks
• 1:46
  that
• 1:46
  do
• 1:46
  not
• 1:47
  require
• 1:47
  frontier
• 1:47
  reasoning
• 1:48
  —
• 1:49
  at
• 1:49
  zero
• 1:49
  API
• 1:49
  cost.
• 1:51
  The
• 1:51
  open-weight
• 1:51
  models
• 1:52
  are
• 1:52
  no
• 1:52
  longer
• 1:52
  just
• 1:52
  catching
• 1:53
  up;
• 1:54
  they
• 1:54
  are
• 1:54
  carving
• 1:54
  out
• 1:54
  practical
• 1:55
  niches
• 1:55
  where
• 1:55
  API
• 1:56
  models
• 1:56
  cannot
• 1:57
  compete
• 1:57
  on
• 1:57
  economics.
• 1:58
  One
• 1:59
  user
• 1:59
  running
• 1:59
  Qwen3.6-27B-UD-Q6
• 2:06
  K
• 2:06
  XL
• 2:06
  on
• 2:06
  an
• 2:07
  RTX
• 2:07
  5090
• 2:08
  reports
• 2:09
  200K
• 2:09
  context
• 2:10
  at
• 2:10
  50
• 2:10
  tokens
• 2:11
  per
• 2:11
  second,
• 2:12
  calling
• 2:12
  it
• 2:12
  "actually
• 2:12
  usable"
• 2:13
  for
• 2:13
  coding
• 2:13
  workflows
• 2:14
  with
• 2:14
  Claude
• 2:14
  Code
• 2:14
  as
• 2:15
  the
• 2:15
  interface.
• 2:16
  The
• 2:16
  Q6
• 2:17
  K
• 2:17
  XL
• 2:17
  quantization
• 2:18
  sits
• 2:18
  at
• 2:18
  a
• 2:18
  higher
• 2:19
  fidelity
• 2:19
  point
• 2:19
  than
• 2:19
  IQ4
• 2:20
  XS
• 2:21
  but
• 2:21
  demands
• 2:21
  more
• 2:22
  VRAM
• 2:22
  —
• 2:23
  a
• 2:23
  tradeoff
• 2:23
  that
• 2:24
  the
• 2:24
  5090's
• 2:24
  32
• 2:25
  GB
• 2:26
  handles
• 2:26
  comfortably.
• 2:27
  The
• 2:27
  practical
• 2:28
  implication:
• 2:29
  the
• 2:29
  latest
• 2:30
  consumer
• 2:30
  hardware
• 2:31
  can
• 2:31
  now
• 2:31
  run
• 2:31
  a
• 2:31
  near-full-quality
• 2:32
  27B
• 2:33
  model
• 2:33
  with
• 2:34
  context
• 2:34
  windows
• 2:34
  that
• 2:35
  would
• 2:35
  have
• 2:35
  required
• 2:35
  API
• 2:36
  access
• 2:36
  six
• 2:36
  months
• 2:37
  ago.
• 2:38
  At
• 2:38
  the
• 2:38
  extreme
• 2:38
  end
• 2:39
  of
• 2:39
  the
• 2:39
  compression
• 2:39
  spectrum,
• 2:40
  BitNet's
• 2:40
  ternary
• 2:41
  quantization
• 2:41
  architecture
• 2:42
  (weights
• 2:42
  constrained
• 2:42
  to
• 2:42
  -1,
• 2:43
  0,
• 2:44
  +1)
• 2:45
  continues
• 2:45
  to
• 2:45
  attract
• 2:46
  speculative
• 2:46
  interest.
• 2:47
  The
• 2:47
  Ternary
• 2:48
  Bonsai
• 2:48
  8B
• 2:49
  model
• 2:49
  fits
• 2:49
  in
• 2:49
  2.2
• 2:50
  GB,
• 2:51
  which
• 2:51
  is
• 2:51
  remarkable
• 2:51
  —
• 2:52
  that
• 2:52
  is
• 2:52
  an
• 2:53
  8-billion
• 2:53
  parameter
• 2:54
  model
• 2:54
  in
• 2:54
  less
• 2:54
  space
• 2:55
  than
• 2:55
  a
• 2:55
  typical
• 2:55
  movie
• 2:55
  file.
• 2:56
  The
• 2:57
  appeal
• 2:57
  is
• 2:57
  obvious:
• 2:58
  ternary
• 2:58
  weights
• 2:59
  eliminate
• 2:59
  multiply-accumulate
• 3:00
  operations
• 3:01
  in
• 3:01
  favor
• 3:01
  of
• 3:01
  simple
• 3:01
  additions
• 3:02
  and
• 3:02
  subtractions,
• 3:03
  potentially
• 3:04
  enabling
• 3:04
  inference
• 3:04
  on
• 3:04
  hardware
• 3:05
  that
• 3:05
  cannot
• 3:06
  run
• 3:06
  traditional
• 3:06
  floating-point
• 3:07
  models
• 3:07
  at
• 3:07
  all.
• 3:08
  But
• 3:08
  quality
• 3:09
  at
• 3:09
  this
• 3:09
  compression
• 3:09
  level
• 3:10
  remains
• 3:10
  the
• 3:10
  open
• 3:10
  question
• 3:11
  that
• 3:11
  proponents
• 3:11
  tend
• 3:12
  to
• 3:12
  wave
• 3:12
  past.
• 3:13
  BitNet
• 3:13
  is
• 3:14
  architecturally
• 3:14
  elegant,
• 3:15
  but
• 3:15
  until
• 3:15
  ternary
• 3:16
  models
• 3:16
  can
• 3:16
  match
• 3:16
  even
• 3:17
  Q4
• 3:17
  quality
• 3:17
  on
• 3:18
  meaningful
• 3:18
  benchmarks,
• 3:19
  it
• 3:19
  stays
• 3:19
  in
• 3:19
  the
• 3:19
  "promising
• 3:20
  research"
• 3:20
  category
• 3:21
  rather
• 3:21
  than
• 3:21
  the
• 3:21
  "practical
• 3:22
  tool"
• 3:25
  A
• 3:25
  new
• 3:25
  open-source
• 3:25
  coding
• 3:26
  agent
• 3:26
  called
• 3:26
  Dirac
• 3:27
  topped
• 3:27
  the
• 3:27
  TerminalBench-2
• 3:27
  leaderboard
• 3:28
  using
• 3:28
  Gemini-3-flash-preview,
• 3:30
  scoring
• 3:30
  65.2%
• 3:31
  —
• 3:32
  beating
• 3:32
  both
• 3:33
  Google's
• 3:33
  official
• 3:33
  baseline
• 3:34
  at
• 3:34
  47.6%
• 3:35
  and
• 3:35
  the
• 3:35
  top
• 3:36
  closed-source
• 3:36
  agent
• 3:36
  Junie
• 3:37
  CLI
• 3:37
  at
• 3:37
  64.3%.
• 3:39
  The
• 3:39
  technical
• 3:39
  approach
• 3:39
  is
• 3:40
  interesting:
• 3:40
  hash-anchored
• 3:41
  edits
• 3:41
  that
• 3:41
  use
• 3:41
  stable
• 3:42
  line
• 3:42
  hashes
• 3:42
  instead
• 3:42
  of
• 3:42
  fragile
• 3:43
  line
• 3:43
  numbers,
• 3:44
  AST-native
• 3:44
  manipulation
• 3:45
  for
• 3:45
  structural
• 3:45
  refactoring,
• 3:46
  and
• 3:46
  multi-file
• 3:47
  batching
• 3:47
  that
• 3:47
  processes
• 3:48
  several
• 3:48
  files
• 3:49
  in
• 3:49
  a
• 3:49
  single
• 3:49
  LLM
• 3:49
  roundtrip.
• 3:51
  On
• 3:51
  eight
• 3:51
  real-world
• 3:51
  refactoring
• 3:52
  tasks
• 3:52
  across
• 3:53
  transformers,
• 3:53
  VS
• 3:54
  Code,
• 3:54
  and
• 3:54
  Django
• 3:55
  codebases,
• 3:56
  Dirac
• 3:56
  achieved
• 3:56
  100%
• 3:57
  accuracy
• 3:57
  at
• 3:58
  an
• 3:58
  average
• 3:58
  cost
• 3:58
  of
• 3:58
  $0.18
• 3:59
  per
• 3:59
  task
• 4:00
  —
• 4:01
  64.8%
• 4:01
  cheaper
• 4:02
  than
• 4:02
  the
• 4:02
  competition.
• 4:03
  The
• 4:03
  agent
• 4:04
  is
• 4:04
  a
• 4:04
  fork
• 4:04
  of
• 4:04
  Cline
• 4:05
  with
• 4:05
  a
• 4:05
  fundamentally
• 4:05
  reworked
• 4:06
  editing
• 4:06
  strategy.
• 4:07
  No
• 4:07
  MCP
• 4:08
  support,
• 4:08
  which
• 4:09
  is
• 4:09
  a
• 4:09
  deliberate
• 4:09
  choice:
• 4:10
  native
• 4:10
  tool
• 4:10
  calling
• 4:11
  only,
• 4:11
  for
• 4:11
  reliability.
• 4:12
  The
• 4:12
  configuration
• 4:13
  management
• 4:13
  problem
• 4:14
  for
• 4:14
  coding
• 4:14
  agents
• 4:15
  gets
• 4:15
  a
• 4:15
  direct
• 4:15
  answer
• 4:15
  from
• 4:16
  anywhere-agents,
• 4:17
  a
• 4:17
  project
• 4:17
  that
• 4:17
  provides
• 4:18
  a
• 4:18
  unified
• 4:18
  AGENTS.md
• 4:19
  configuration
• 4:20
  format
• 4:21
  that
• 4:21
  works
• 4:21
  across
• 4:21
  Claude
• 4:22
  Code,
• 4:22
  Codex,
• 4:23
  Gemini
• 4:23
  CLI,
• 4:24
  and
• 4:24
  other
• 4:24
  agent
• 4:24
  frameworks.
• 4:25
  It
• 4:25
  includes
• 4:26
  a
• 4:26
  pack
• 4:26
  system
• 4:27
  for
• 4:27
  composable
• 4:27
  rulesets
• 4:28
  and
• 4:28
  a
• 4:28
  writing
• 4:28
  style
• 4:29
  guard
• 4:29
  that
• 4:29
  blocks
• 4:29
  roughly
• 4:30
  45
• 4:30
  AI-tell
• 4:31
  words.
• 4:32
  The
• 4:32
  underlying
• 4:32
  problem
• 4:32
  is
• 4:32
  real:
• 4:33
  as
• 4:33
  developers
• 4:34
  run
• 4:34
  multiple
• 4:35
  agent
• 4:35
  frameworks
• 4:35
  in
• 4:35
  the
• 4:36
  same
• 4:36
  repository,
• 4:37
  maintaining
• 4:38
  separate
• 4:38
  configuration
• 4:39
  files
• 4:39
  for
• 4:39
  each
• 4:39
  becomes
• 4:40
  unsustainable.
• 4:41
  A
• 4:41
  single
• 4:42
  declarative
• 4:42
  format
• 4:43
  that
• 4:43
  all
• 4:43
  agents
• 4:43
  can
• 4:43
  read
• 4:44
  is
• 4:44
  the
• 4:44
  kind
• 4:45
  of
• 4:45
  boring
• 4:45
  infrastructure
• 4:45
  work
• 4:46
  that
• 4:46
  eventually
• 4:46
  becomes
• 4:47
  essential.
• 4:48
  Meanwhile,
• 4:49
  LlamaIndex
• 4:49
  released
• 4:50
  ParseBench,
• 4:51
  a
• 4:51
  benchmark
• 4:51
  that
• 4:52
  tests
• 4:52
  whether
• 4:52
  document
• 4:53
  parsing
• 4:53
  tools
• 4:54
  produce
• 4:54
  output
• 4:55
  that
• 4:55
  AI
• 4:55
  agents
• 4:55
  can
• 4:56
  actually
• 4:56
  act
• 4:56
  on.
• 4:57
  The
• 4:57
  benchmark
• 4:58
  covers
• 4:58
  roughly
• 4:59
  2,000
• 4:59
  human-verified
• 5:00
  pages
• 5:00
  from
• 5:01
  real
• 5:01
  enterprise
• 5:01
  documents
• 5:02
  across
• 5:02
  five
• 5:02
  capability
• 5:03
  dimensions:
• 5:04
  tables,
• 5:05
  charts,
• 5:05
  content
• 5:06
  faithfulness,
• 5:06
  semantic
• 5:07
  formatting,
• 5:07
  and
• 5:07
  visual
• 5:08
  grounding.
• 5:09
  LlamaParse
• 5:09
  Agentic
• 5:10
  leads
• 5:10
  the
• 5:10
  leaderboard
• 5:10
  at
• 5:11
  84.88%
• 5:12
  overall,
• 5:13
  with
• 5:13
  Gemini
• 5:13
  3
• 5:13
  Flash
• 5:14
  at
• 5:14
  75.05%
• 5:16
  and
• 5:16
  Reducto
• 5:17
  at
• 5:17
  72.97%.
• 5:19
  The
• 5:19
  evaluation
• 5:20
  is
• 5:20
  entirely
• 5:21
  deterministic
• 5:21
  —
• 5:22
  no
• 5:22
  LLM-as-a-judge.
• 5:24
  The
• 5:24
  most
• 5:24
  revealing
• 5:24
  dimension
• 5:25
  may
• 5:25
  be
• 5:25
  semantic
• 5:25
  formatting,
• 5:26
  which
• 5:26
  tests
• 5:27
  whether
• 5:27
  parsers
• 5:27
  preserve
• 5:28
  meaning-carrying
• 5:28
  formatting
• 5:29
  like
• 5:29
  strikethrough
• 5:30
  text
• 5:30
  (a
• 5:30
  superseded
• 5:30
  price
• 5:31
  is
• 5:31
  not
• 5:31
  the
• 5:31
  current
• 5:32
  price).
• 5:33
  Most
• 5:33
  parsers
• 5:34
  fail
• 5:34
  badly
• 5:34
  here,
• 5:35
  and
• 5:35
  any
• 5:35
  agent
• 5:35
  downstream
• 5:36
  inherits
• 5:36
  that
• 5:37
  failure
• 5:40
  OpenAI
• 5:41
  released
• 5:41
  GPT-5.5,
• 5:42
  priced
• 5:43
  at
• 5:43
  $5
• 5:43
  per
• 5:44
  million
• 5:44
  input
• 5:44
  tokens
• 5:45
  and
• 5:45
  $30
• 5:45
  per
• 5:45
  million
• 5:46
  output
• 5:46
  tokens.
• 5:47
  The
• 5:47
  community
• 5:48
  response
• 5:48
  has
• 5:48
  been
• 5:48
  measured
• 5:49
  rather
• 5:49
  than
• 5:49
  enthusiastic.
• 5:50
  On
• 5:50
  SWE
• 5:51
  Bench
• 5:51
  Pro,
• 5:52
  GPT-5.5
• 5:53
  scored
• 5:53
  58.6%,
• 5:55
  trailing
• 5:55
  Anthropic's
• 5:56
  Opus
• 5:56
  4.7
• 5:57
  at
• 5:58
  The
• 5:59
  pricing
• 5:59
  positions
• 6:00
  it
• 6:00
  as
• 6:00
  a
• 6:00
  mid-tier
• 6:00
  offering
• 6:01
  —
• 6:01
  cheaper
• 6:02
  than
• 6:02
  frontier
• 6:02
  models
• 6:03
  but
• 6:03
  more
• 6:03
  expensive
• 6:03
  than
• 6:04
  the
• 6:04
  flash-tier
• 6:04
  alternatives
• 6:05
  that
• 6:05
  have
• 6:05
  been
• 6:05
  eating
• 6:05
  into
• 6:05
  API
• 6:06
  revenue.
• 6:07
  OpenAI
• 6:07
  appears
• 6:08
  to
• 6:08
  be
• 6:08
  filling
• 6:08
  out
• 6:08
  its
• 6:08
  model
• 6:09
  lineup
• 6:09
  across
• 6:09
  price
• 6:10
  points
• 6:10
  rather
• 6:10
  than
• 6:10
  pushing
• 6:10
  the
• 6:11
  frontier
• 6:11
  forward
• 6:11
  with
• 6:11
  this
• 6:11
  release.
• 6:12
  For
• 6:12
  coding-heavy
• 6:12
  workloads
• 6:13
  where
• 6:13
  SWE
• 6:13
  Bench
• 6:14
  Pro
• 6:14
  performance
• 6:14
  correlates
• 6:15
  with
• 6:15
  real
• 6:15
  productivity,
• 6:16
  the
• 6:16
  6-point
• 6:16
  benchmark
• 6:17
  gap
• 6:17
  matters
• 6:17
  more
• 6:18
  than
• 6:18
  the
• 6:18
  price
• 6:18
  gap.
• 6:19
  In
• 6:19
  a
• 6:19
  story
• 6:19
  that
• 6:20
  writes
• 6:20
  its
• 6:20
  own
• 6:20
  punchline,
• 6:21
  Tools
• 6:21
  for
• 6:21
  Humanity
• 6:21
  —
• 6:22
  Sam
• 6:22
  Altman's
• 6:23
  identity
• 6:23
  verification
• 6:24
  company
• 6:24
  —
• 6:25
  announced
• 6:25
  a
• 6:25
  partnership
• 6:26
  with
• 6:26
  Bruno
• 6:26
  Mars
• 6:26
  that
• 6:26
  turned
• 6:27
  out
• 6:27
  to
• 6:27
  be
• 6:27
  The
• 6:28
  company
• 6:28
  had
• 6:28
  actually
• 6:29
  partnered
• 6:29
  with
• 6:29
  Thirty
• 6:29
  Seconds
• 6:30
  to
• 6:30
  Mars
• 6:30
  and
• 6:30
  confused
• 6:31
  the
• 6:31
  two.
• 6:32
  An
• 6:32
  identity
• 6:32
  verification
• 6:33
  company
• 6:33
  that
• 6:34
  cannot
• 6:34
  verify
• 6:34
  the
• 6:35
  identity
• 6:35
  of
• 6:35
  its
• 6:35
  own
• 6:35
  partners
• 6:36
  is
• 6:36
  a
• 6:36
  comedy
• 6:37
  of
• 6:37
  errors,
• 6:37
  but
• 6:38
  it
• 6:38
  also
• 6:38
  raises
• 6:38
  a
• 6:38
  real
• 6:38
  question
• 6:39
  about
• 6:39
  the
• 6:39
  operational
• 6:40
  rigor
• 6:40
  of
• 6:40
  companies
• 6:41
  asking
• 6:41
  billions
• 6:41
  of
• 6:41
  people
• 6:42
  to
• 6:42
  scan
• 6:42
  their
• 6:42
  irises.
• 6:43
  If
• 6:43
  your
• 6:44
  verification
• 6:44
  process
• 6:44
  cannot
• 6:45
  distinguish
• 6:45
  between
• 6:46
  two
• 6:46
  musicians
• 6:46
  with
• 6:46
  "Mars"
• 6:47
  in
• 6:47
  their
• 6:47
  name,
• 6:48
  what
• 6:48
  confidence
• 6:49
  should
• 6:49
  users
• 6:49
  place
• 6:49
  in
• 6:49
  the
• 6:49
  rest
• 6:50
  of
• 6:50
  your
• 6:50
  identity
• 6:50
  infrastructure?
• 6:51
  The
• 6:51
  political
• 6:52
  environment
• 6:52
  around
• 6:52
  AI
• 6:53
  continues
• 6:53
  to
• 6:53
  shift.
• 6:54
  The
• 6:54
  Trump
• 6:55
  administration,
• 6:55
  which
• 6:56
  had
• 6:56
  picked
• 6:56
  a
• 6:56
  public
• 6:56
  fight
• 6:56
  with
• 6:57
  Anthropic
• 6:57
  over
• 6:57
  safety
• 6:58
  policy
• 6:58
  disagreements,
• 6:59
  appears
• 6:59
  to
• 6:59
  be
• 6:59
  backing
• 7:00
  off.
• 7:00
  The
• 7:01
  details
• 7:01
  remain
• 7:01
  thin,
• 7:02
  but
• 7:02
  the
• 7:02
  pattern
• 7:02
  is
• 7:03
  familiar:
• 7:04
  aggressive
• 7:04
  posturing
• 7:04
  followed
• 7:05
  by
• 7:05
  quiet
• 7:05
  retreat
• 7:06
  when
• 7:06
  the
• 7:06
  economic
• 7:07
  and
• 7:07
  diplomatic
• 7:07
  implications
• 7:08
  of
• 7:08
  antagonizing
• 7:08
  a
• 7:09
  major
• 7:09
  AI
• 7:09
  company
• 7:10
  become
• 7:10
  apparent.
• 7:11
  AI
• 7:11
  policy
• 7:11
  in
• 7:11
  the
• 7:12
  current
• 7:12
  administration
• 7:12
  has
• 7:13
  been
• 7:13
  more
• 7:13
  reactive
• 7:13
  than
• 7:14
  strategic,
• 7:14
  and
• 7:15
  this
• 7:15
  reversal
• 7:15
  fits
• 7:15
  that
• 7:15
  pattern.
• 7:17
  Underneath
• 7:17
  the
• 7:17
  daily
• 7:17
  news
• 7:18
  cycle,
• 7:18
  Max
• 7:19
  Tegmark's
• 7:19
  Life
• 7:19
  3.0
• 7:20
  framework
• 7:20
  continues
• 7:21
  to
• 7:21
  shape
• 7:21
  how
• 7:22
  researchers
• 7:22
  think
• 7:22
  about
• 7:22
  longer-term
• 7:23
  AI
• 7:23
  trajectories.
• 7:25
  A
• 7:25
  recent
• 7:26
  editorial
• 7:26
  revisits
• 7:26
  the
• 7:27
  book's
• 7:27
  taxonomy
• 7:27
  of
• 7:28
  twelve
• 7:28
  possible
• 7:28
  AI
• 7:29
  futures
• 7:29
  —
• 7:30
  from
• 7:30
  benevolent
• 7:31
  superintelligence
• 7:31
  to
• 7:31
  civilizational
• 7:32
  extinction
• 7:32
  —
• 7:34
  and
• 7:34
  notes
• 7:34
  that
• 7:34
  surveyed
• 7:35
  AI
• 7:35
  researchers
• 7:35
  estimate
• 7:36
  roughly
• 7:36
  a
• 7:36
  1-in-6
• 7:37
  probability
• 7:37
  of
• 7:37
  an
• 7:37
  extinction-level
• 7:38
  outcome.
• 7:39
  Whether
• 7:39
  that
• 7:39
  number
• 7:40
  is
• 7:40
  calibrated
• 7:40
  or
• 7:41
  performative,
• 7:42
  it
• 7:42
  represents
• 7:42
  a
• 7:42
  nontrivial
• 7:43
  fraction
• 7:43
  of
• 7:43
  the
• 7:43
  people
• 7:44
  building
• 7:44
  these
• 7:44
  systems
• 7:45
  expressing
• 7:45
  deep
• 7:46
  uncertainty
• 7:46
  about
• 7:47
  where
• 7:47
  they
• 7:47
  lead.
• 7:48
  The
• 7:48
  gap
• 7:48
  between
• 7:49
  building
• 7:49
  faster
• 7:49
  and
• 7:49
  thinking
• 7:50
  harder
• 7:50
  about
• 7:50
  where
• 7:51
  we
• 7:51
  are
• 7:51
  building
• 7:51
  toward
• 7:52
  is
• 7:52
  not
• 7:52
  closing.
• 7:54
  A
• 7:54
  new
• 7:55
  paper
• 7:55
  introduces
• 7:55
  ArmSSL,
• 7:56
  a
• 7:56
  framework
• 7:57
  for
• 7:57
  embedding
• 7:57
  adversarial-robust
• 7:58
  watermarks
• 7:58
  into
• 7:59
  self-supervised
• 7:59
  learning
• 8:00
  pre-trained
• 8:00
  encoders.
• 8:01
  The
• 8:01
  problem
• 8:01
  is
• 8:01
  significant:
• 8:02
  SSL
• 8:03
  encoders
• 8:03
  like
• 8:03
  SimCLR,
• 8:04
  MoCo
• 8:04
  v2,
• 8:05
  BYOL,
• 8:05
  SimSiam,
• 8:06
  and
• 8:06
  DINOv2
• 8:06
  are
• 8:06
  expensive
• 8:07
  to
• 8:07
  train
• 8:07
  and
• 8:07
  easy
• 8:08
  to
• 8:08
  steal
• 8:08
  —
• 8:08
  just
• 8:09
  fine-tune
• 8:09
  the
• 8:09
  last
• 8:09
  layers
• 8:09
  and
• 8:10
  the
• 8:10
  original
• 8:10
  weights
• 8:10
  carry
• 8:10
  over
• 8:11
  unmarked.
• 8:11
  ArmSSL
• 8:12
  uses
• 8:12
  three
• 8:12
  paired
• 8:14
  discrepancy
• 8:14
  enlargement
• 8:15
  to
• 8:15
  create
• 8:15
  detectable
• 8:16
  differences
• 8:16
  between
• 8:16
  watermarked
• 8:17
  and
• 8:17
  unwatermarked
• 8:18
  encoders,
• 8:19
  latent
• 8:19
  representation
• 8:19
  entanglement
• 8:20
  to
• 8:20
  survive
• 8:20
  fine-tuning
• 8:21
  attacks,
• 8:22
  and
• 8:22
  distribution
• 8:22
  alignment
• 8:23
  to
• 8:23
  maintain
• 8:23
  the
• 8:23
  encoder's
• 8:24
  utility
• 8:24
  on
• 8:24
  clean
• 8:24
  data.
• 8:25
  The
• 8:26
  approach
• 8:26
  is
• 8:26
  black-box,
• 8:27
  meaning
• 8:27
  verification
• 8:28
  does
• 8:28
  not
• 8:28
  require
• 8:28
  access
• 8:29
  to
• 8:29
  model
• 8:29
  internals.
• 8:30
  IP
• 8:30
  protection
• 8:31
  for
• 8:31
  foundation
• 8:32
  models
• 8:32
  has
• 8:32
  been
• 8:32
  an
• 8:32
  afterthought
• 8:33
  in
• 8:33
  the
• 8:33
  rush
• 8:33
  to
• 8:33
  train
• 8:34
  and
• 8:34
  deploy;
• 8:35
  work
• 8:35
  like
• 8:35
  this
• 8:36
  starts
• 8:36
  building
• 8:36
  the
• 8:36
  forensic
• 8:37
  infrastructure
• 8:37
  that
• 8:37
  will
• 8:38
  eventually
• 8:38
  be
• 8:38
  necessary
• 8:39
  when
• 8:39
  model
• 8:39
  theft
• 8:39
  disputes
• 8:40
  reach
• 8:40
  courtrooms.
• 8:41
  On
• 8:41
  the
• 8:41
  offensive
• 8:42
  security
• 8:42
  side,
• 8:43
  GTFOBins
• 8:43
  —
• 8:45
  the
• 8:45
  curated
• 8:45
  database
• 8:46
  of
• 8:46
  Unix
• 8:46
  executables
• 8:47
  that
• 8:47
  can
• 8:47
  be
• 8:47
  exploited
• 8:47
  for
• 8:48
  privilege
• 8:48
  escalation,
• 8:49
  shell
• 8:49
  escapes,
• 8:49
  and
• 8:49
  file
• 8:50
  operations
• 8:50
  —
• 8:51
  resurfaced
• 8:52
  on
• 8:52
  Hacker
• 8:52
  News
• 8:52
  this
• 8:52
  week.
• 8:54
  The
• 8:54
  project
• 8:54
  catalogs
• 8:55
  "living
• 8:55
  off
• 8:55
  the
• 8:55
  land"
• 8:56
  techniques:
• 8:57
  ways
• 8:57
  attackers
• 8:57
  use
• 8:58
  legitimate
• 8:58
  system
• 8:59
  binaries
• 8:59
  (awk,
• 9:00
  python,
• 9:00
  vim,
• 9:01
  tar,
• 9:02
  and
• 9:02
  dozens
• 9:02
  more)
• 9:02
  to
• 9:02
  bypass
• 9:03
  security
• 9:03
  controls
• 9:04
  without
• 9:04
  installing
• 9:04
  malicious
• 9:05
  software.
• 9:06
  Each
• 9:06
  entry
• 9:06
  documents
• 9:07
  the
• 9:07
  specific
• 9:07
  capabilities
• 9:08
  of
• 9:08
  a
• 9:08
  given
• 9:09
  binary
• 9:09
  —
• 9:10
  whether
• 9:10
  it
• 9:10
  can
• 9:10
  read
• 9:10
  files,
• 9:11
  write
• 9:11
  files,
• 9:11
  spawn
• 9:12
  shells,
• 9:12
  or
• 9:12
  escalate
• 9:12
  privileges
• 9:13
  —
• 9:13
  along
• 9:14
  with
• 9:14
  exact
• 9:14
  command
• 9:15
  invocations.
• 9:16
  For
• 9:16
  defenders,
• 9:17
  it
• 9:17
  is
• 9:17
  an
• 9:17
  essential
• 9:18
  reference
• 9:18
  for
• 9:18
  audit
• 9:18
  and
• 9:19
  hardening.
• 9:20
  For
• 9:20
  anyone
• 9:20
  building
• 9:20
  AI
• 9:21
  agents
• 9:21
  that
• 9:21
  execute
• 9:22
  shell
• 9:22
  commands,
• 9:23
  it
• 9:23
  is
• 9:23
  a
• 9:23
  sobering
• 9:23
  reminder
• 9:24
  that
• 9:24
  the
• 9:24
  attack
• 9:25
  surface
• 9:25
  of
• 9:25
  "run
• 9:25
  this
• 9:25
  bash
• 9:26
  command"
• 9:26
  is
• 9:26
  vastly
• 9:27
  larger
• 9:27
  than
• 9:27
  the
• 9:27
  command
• 9:28
  itself.
• 9:29
  Every
• 9:29
  binary
• 9:30
  in
• 9:30
  PATH
• 9:30
  is
• 9:30
  a
• 9:31
  potential
• 9:31
  escalation
• 9:31
  vector,
• 9:32
  and
• 9:32
  the
• 9:32
  agent's
• 9:33
  sandbox
• 9:33
  is
• 9:33
  only
• 9:33
  as
• 9:34
  strong
• 9:34
  as
• 9:34
  its
• 9:34
  least-restricted
• 9:35
  executable.
• 9:38
  A
• 9:38
  paper
• 9:38
  on
• 9:38
  generalization
• 9:39
  at
• 9:39
  the
• 9:39
  edge
• 9:39
  of
• 9:39
  stability
• 9:40
  introduces
• 9:40
  a
• 9:40
  new
• 9:41
  geometric
• 9:41
  quantity
• 9:41
  called
• 9:42
  "sharpness
• 9:42
  dimension"
• 9:43
  (SD)
• 9:43
  that
• 9:43
  may
• 9:43
  explain
• 9:44
  why
• 9:44
  neural
• 9:44
  networks
• 9:44
  generalize
• 9:45
  well
• 9:45
  despite
• 9:46
  operating
• 9:46
  in
• 9:46
  a
• 9:46
  regime
• 9:46
  where
• 9:47
  traditional
• 9:47
  theory
• 9:47
  predicts
• 9:48
  they
• 9:48
  should
• 9:48
  not.
• 9:48
  When
• 9:49
  the
• 9:49
  learning
• 9:49
  rate
• 9:49
  is
• 9:49
  large
• 9:49
  enough
• 9:50
  that
• 9:50
  training
• 9:50
  loss
• 9:50
  oscillates
• 9:51
  rather
• 9:51
  than
• 9:52
  monotonically
• 9:52
  decreasing
• 9:53
  —
• 9:53
  the
• 9:53
  "edge
• 9:54
  of
• 9:54
  stability"
• 9:54
  regime
• 9:55
  —
• 9:55
  the
• 9:55
  authors
• 9:56
  model
• 9:56
  the
• 9:56
  optimizer's
• 9:57
  trajectory
• 9:57
  as
• 9:57
  a
• 9:57
  random
• 9:58
  dynamical
• 9:58
  system
• 9:59
  on
• 9:59
  the
• 9:59
  loss
• 9:59
  landscape's
• 9:59
  pullback
• 10:00
  Using
• 10:01
  Lyapunov
• 10:02
  theory
• 10:02
  from
• 10:02
  dynamical
• 10:03
  systems,
• 10:03
  they
• 10:03
  prove
• 10:04
  a
• 10:04
  generalization
• 10:04
  bound
• 10:05
  that
• 10:05
  depends
• 10:05
  on
• 10:05
  the
• 10:05
  sharpness
• 10:06
  dimension
• 10:06
  of
• 10:06
  this
• 10:06
  attractor
• 10:07
  rather
• 10:07
  than
• 10:07
  on
• 10:07
  the
• 10:07
  number
• 10:08
  of
• 10:08
  model
• 10:08
  parameters.
• 10:09
  The
• 10:09
  bound
• 10:09
  is
• 10:09
  validated
• 10:10
  on
• 10:10
  MLPs,
• 10:11
  GPT-2,
• 10:12
  and
• 10:12
  grokking
• 10:12
  experiments.
• 10:13
  If
• 10:14
  the
• 10:14
  result
• 10:14
  holds
• 10:14
  up
• 10:14
  under
• 10:15
  further
• 10:15
  scrutiny,
• 10:16
  it
• 10:16
  offers
• 10:16
  a
• 10:16
  principled
• 10:17
  explanation
• 10:18
  for
• 10:18
  why
• 10:18
  overparameterized
• 10:19
  models
• 10:19
  trained
• 10:20
  with
• 10:20
  large
• 10:20
  learning
• 10:21
  rates
• 10:21
  do
• 10:21
  not
• 10:21
  simply
• 10:22
  memorize
• 10:22
  their
• 10:22
  training
• 10:23
  data
• 10:23
  —
• 10:24
  the
• 10:24
  geometry
• 10:25
  of
• 10:25
  the
• 10:25
  optimization
• 10:25
  path
• 10:26
  constrains
• 10:26
  effective
• 10:27
  capacity
• 10:28
  in
• 10:28
  ways
• 10:28
  that
• 10:28
  parameter
• 10:28
  counting
• 10:29
  misses.
• 10:30
  The
• 10:30
  mathematical
• 10:31
  foundations
• 10:31
  underlying
• 10:32
  these
• 10:32
  training
• 10:32
  dynamics
• 10:33
  continue
• 10:34
  to
• 10:34
  attract
• 10:34
  educational
• 10:35
  attention.
• 10:36
  A
• 10:36
  recent
• 10:37
  video
• 10:37
  breakdown
• 10:37
  walks
• 10:38
  through
• 10:38
  the
• 10:38
  calculus
• 10:39
  of
• 10:39
  gradient-based
• 10:39
  optimization
• 10:40
  from
• 10:40
  first
• 10:41
  principles,
• 10:42
  covering
• 10:42
  the
• 10:42
  chain
• 10:42
  rule
• 10:43
  mechanics,
• 10:43
  loss
• 10:44
  surface
• 10:44
  geometry,
• 10:45
  and
• 10:45
  the
• 10:45
  specific
• 10:46
  derivative
• 10:46
  computations
• 10:47
  that
• 10:47
  make
• 10:47
  backpropagation
• 10:48
  work.
• 10:49
  The
• 10:49
  presentation
• 10:50
  builds
• 10:50
  intuition
• 10:51
  for
• 10:51
  why
• 10:51
  gradient
• 10:52
  descent
• 10:52
  works
• 10:52
  at
• 10:52
  all
• 10:53
  —
• 10:53
  and
• 10:53
  more
• 10:54
  importantly,
• 10:54
  why
• 10:55
  it
• 10:55
  sometimes
• 10:55
  does
• 10:55
  not
• 10:56
  —
• 10:57
  by
• 10:57
  grounding
• 10:57
  abstract
• 10:58
  optimization
• 10:58
  concepts
• 10:59
  in
• 10:59
  concrete
• 10:59
  geometric
• 11:00
  reasoning.
• 11:01
  For
• 11:01
  practitioners
• 11:02
  who
• 11:02
  use
• 11:02
  these
• 11:02
  tools
• 11:03
  daily
• 11:03
  without
• 11:03
  fully
• 11:04
  understanding
• 11:04
  the
• 11:04
  mathematics
• 11:05
  beneath
• 11:05
  them,
• 11:06
  this
• 11:06
  kind
• 11:06
  of
• 11:06
  foundational
• 11:07
  content
• 11:07
  fills
• 11:08
  a
• 11:08
  gap
• 11:08
  that
• 11:08
  blog
• 11:08
  posts
• 11:09
  and
• 11:09
  API
• 11:10
  documentation
• 11:10
  leave
• 11:10
  open.
• 11:13
  Google
• 11:14
  published
• 11:14
  the
• 11:14
  engineering
• 11:15
  details
• 11:15
  behind
• 11:16
  TorchTPU,
• 11:17
  its
• 11:18
  integration
• 11:18
  layer
• 11:18
  for
• 11:19
  running
• 11:19
  PyTorch
• 11:19
  natively
• 11:20
  on
• 11:20
  Tensor
• 11:20
  Processing
• 11:21
  Units.
• 11:22
  The
• 11:22
  architecture
• 11:22
  is
• 11:23
  built
• 11:23
  on
• 11:23
  an
• 11:23
  "Eager
• 11:23
  First"
• 11:24
  philosophy
• 11:24
  with
• 11:24
  three
• 11:25
  execution
• 11:25
  modes:
• 11:26
  Debug
• 11:27
  Eager
• 11:27
  for
• 11:27
  single-op
• 11:28
  synchronous
• 11:28
  execution,
• 11:29
  Strict
• 11:29
  Eager
• 11:30
  for
• 11:30
  asynchronous
• 11:30
  single-op
• 11:31
  dispatch,
• 11:32
  and
• 11:32
  Fused
• 11:32
  Eager
• 11:32
  which
• 11:33
  automatically
• 11:33
  fuses
• 11:34
  streams
• 11:34
  of
• 11:34
  operations
• 11:35
  into
• 11:35
  dense
• 11:35
  computational
• 11:36
  chunks
• 11:36
  —
• 11:37
  delivering
• 11:37
  50-100%
• 11:40
  speedups
• 11:41
  over
• 11:41
  Strict
• 11:41
  Eager
• 11:41
  with
• 11:41
  no
• 11:42
  user
• 11:42
  configuration.
• 11:43
  For
• 11:43
  full-graph
• 11:44
  compilation,
• 11:45
  TorchTPU
• 11:46
  routes
• 11:46
  through
• 11:46
  XLA
• 11:47
  via
• 11:47
  StableHLO
• 11:48
  rather
• 11:48
  than
• 11:48
  Torch
• 11:48
  Inductor,
• 11:49
  a
• 11:49
  deliberate
• 11:50
  choice
• 11:50
  because
• 11:50
  XLA
• 11:51
  already
• 11:51
  understands
• 11:52
  how
• 11:52
  to
• 11:52
  optimize
• 11:53
  communication-computation
• 11:54
  overlap
• 11:54
  across
• 11:54
  the
• 11:54
  TPU's
• 11:55
  Inter-Chip
• 11:55
  Interconnect
• 11:56
  torus
• 11:56
  topology.
• 11:57
  The
• 11:58
  MPMD
• 11:58
  support
• 11:59
  is
• 11:59
  particularly
• 11:59
  notable:
• 12:00
  prior
• 12:00
  PyTorch/XLA
• 12:02
  only
• 12:03
  supported
• 12:03
  pure
• 12:03
  SPMD
• 12:04
  code,
• 12:05
  which
• 12:05
  meant
• 12:05
  developers
• 12:06
  had
• 12:06
  to
• 12:06
  carefully
• 12:06
  remove
• 12:06
  any
• 12:07
  rank-divergent
• 12:07
  behavior
• 12:08
  (even
• 12:08
  logging).
• 12:09
  TorchTPU
• 12:10
  isolates
• 12:10
  divergent
• 12:11
  executions
• 12:11
  automatically.
• 12:13
  The
• 12:13
  2026
• 12:13
  roadmap
• 12:14
  includes
• 12:14
  bounded
• 12:14
  dynamism
• 12:15
  for
• 12:15
  variable
• 12:15
  sequence
• 12:16
  lengths
• 12:16
  and
• 12:16
  integration
• 12:17
  with
• 12:17
  vLLM.
• 12:18
  At
• 12:18
  the
• 12:18
  opposite
• 12:19
  end
• 12:19
  of
• 12:19
  the
• 12:19
  compute
• 12:19
  spectrum,
• 12:20
  DSPi
• 12:21
  turns
• 12:21
  the
• 12:21
  $4
• 12:22
  Raspberry
• 12:22
  Pi
• 12:22
  Pico
• 12:23
  into
• 12:23
  a
• 12:23
  fully
• 12:23
  featured
• 12:24
  audio
• 12:24
  DSP
• 12:24
  processor.
• 12:26
  The
• 12:26
  firmware
• 12:26
  implements
• 12:26
  parametric
• 12:27
  equalization,
• 12:28
  active
• 12:28
  crossover
• 12:29
  filters,
• 12:29
  time
• 12:30
  alignment,
• 12:30
  loudness
• 12:31
  compensation,
• 12:32
  and
• 12:32
  headphone
• 12:32
  crossfeed
• 12:32
  —
• 12:33
  capabilities
• 12:34
  that
• 12:34
  typically
• 12:34
  require
• 12:35
  dedicated
• 12:35
  hardware
• 12:36
  costing
• 12:36
  hundreds
• 12:37
  of
• 12:37
  dollars.
• 12:38
  The
• 12:38
  Pico
• 12:38
  operates
• 12:39
  as
• 12:39
  a
• 12:39
  USB
• 12:39
  sound
• 12:40
  card,
• 12:40
  receiving
• 12:41
  audio
• 12:41
  over
• 12:41
  USB
• 12:42
  and
• 12:42
  processing
• 12:43
  it
• 12:43
  in
• 12:43
  real-time
• 12:44
  before
• 12:44
  output.
• 12:45
  Room
• 12:45
  correction,
• 12:46
  a
• 12:46
  feature
• 12:46
  that
• 12:47
  usually
• 12:47
  demands
• 12:47
  proprietary
• 12:48
  measurement
• 12:48
  software
• 12:49
  and
• 12:49
  expensive
• 12:49
  hardware,
• 12:50
  is
• 12:50
  implemented
• 12:51
  entirely
• 12:51
  in
• 12:51
  the
• 12:52
  Pico's
• 12:52
  dual
• 12:52
  ARM
• 12:53
  Cortex-M0+
• 12:54
  cores.
• 12:55
  For
• 12:55
  the
• 12:55
  audio
• 12:56
  engineering
• 12:56
  community,
• 12:57
  this
• 12:57
  democratizes
• 12:58
  DSP
• 12:58
  processing
• 12:59
  in
• 12:59
  a
• 12:59
  way
• 12:59
  that
• 12:59
  commercial
• 13:00
  products
• 13:00
  have
• 13:00
  resisted.
• 13:01
  Apple
• 13:01
  Silicon
• 13:02
  users
• 13:02
  running
• 13:02
  3D
• 13:03
  generation
• 13:04
  workflows
• 13:04
  now
• 13:04
  have
• 13:05
  a
• 13:05
  3x
• 13:05
  speedup
• 13:06
  path:
• 13:07
  Hunyuan's
• 13:07
  Image-to-3D
• 13:08
  texture
• 13:08
  pipeline
• 13:09
  ported
• 13:09
  to
• 13:09
  MLX
• 13:10
  runs
• 13:10
  three
• 13:10
  times
• 13:11
  faster
• 13:11
  than
• 13:11
  the
• 13:11
  PyTorch
• 13:12
  equivalent
• 13:12
  while
• 13:12
  using
• 13:13
  half
• 13:13
  the
• 13:13
  VRAM.
• 13:14
  The
• 13:14
  optimization
• 13:15
  leverages
• 13:15
  MLX's
• 13:16
  unified
• 13:17
  memory
• 13:17
  architecture
• 13:18
  and
• 13:18
  Metal
• 13:18
  compute
• 13:19
  shaders,
• 13:19
  turning
• 13:20
  what
• 13:20
  was
• 13:20
  a
• 13:20
  multi-minute
• 13:21
  generation
• 13:21
  process
• 13:22
  into
• 13:22
  something
• 13:22
  approaching
• 13:22
  interactive
• 13:23
  speed.
• 13:24
  For
• 13:24
  creative
• 13:25
  workflows
• 13:25
  that
• 13:25
  iterate
• 13:26
  rapidly
• 13:26
  on
• 13:26
  3D
• 13:27
  asset
• 13:27
  generation,
• 13:28
  this
• 13:28
  is
• 13:28
  the
• 13:28
  difference
• 13:29
  between
• 13:29
  a
• 13:29
  tool
• 13:29
  you
• 13:29
  wait
• 13:30
  for
• 13:30
  and
• 13:30
  a
• 13:30
  tool
• 13:30
  you
• 13:30
  use.
• 13:32
  Embedded
• 13:32
  systems
• 13:33
  development
• 13:33
  is
• 13:33
  getting
• 13:33
  more
• 13:34
  interesting
• 13:34
  too.
• 13:35
  A
• 13:35
  detailed
• 13:36
  writeup
• 13:36
  demonstrates
• 13:37
  running
• 13:37
  bare-metal
• 13:37
  Rust
• 13:38
  on
• 13:38
  the
• 13:38
  ESP32-S3's
• 13:41
  second
• 13:41
  core
• 13:41
  while
• 13:41
  ESP-IDF
• 13:43
  and
• 13:44
  FreeRTOS
• 13:44
  run
• 13:44
  on
• 13:44
  the
• 13:44
  first.
• 13:46
  The
• 13:46
  approach
• 13:46
  uses
• 13:46
  a
• 13:46
  separate
• 13:47
  flash
• 13:47
  partition
• 13:48
  and
• 13:48
  MMU
• 13:49
  mapping
• 13:49
  to
• 13:49
  load
• 13:49
  the
• 13:49
  Rust
• 13:50
  binary,
• 13:51
  making
• 13:51
  it
• 13:51
  hot-swappable
• 13:52
  without
• 13:52
  reflashing
• 13:53
  the
• 13:53
  FreeRTOS
• 13:53
  side.
• 13:54
  Core
• 13:54
  0
• 13:54
  handles
• 13:55
  WiFi,
• 13:55
  Bluetooth,
• 13:56
  and
• 13:56
  OS
• 13:56
  tasks;
• 13:57
  Core
• 13:57
  1
• 13:58
  runs
• 13:58
  deterministic
• 13:59
  Rust
• 13:59
  code
• 13:59
  with
• 13:59
  no
• 13:59
  RTOS
• 14:00
  jitter.
• 14:01
  For
• 14:01
  robotics
• 14:02
  and
• 14:02
  real-time
• 14:02
  control
• 14:03
  applications,
• 14:04
  this
• 14:04
  dual-personality
• 14:05
  architecture
• 14:06
  gives
• 14:06
  you
• 14:06
  the
• 14:06
  connectivity
• 14:07
  stack
• 14:07
  of
• 14:07
  ESP-IDF
• 14:10
  with
• 14:10
  the
• 14:10
  timing
• 14:10
  guarantees
• 14:11
  of
• 14:11
  bare-metal
• 14:12
  execution.
• 14:13
  The
• 14:13
  concept
• 14:13
  of
• 14:14
  "dark
• 14:14
  factories"
• 14:14
  —
• 14:15
  fully
• 14:15
  automated
• 14:16
  manufacturing
• 14:16
  facilities
• 14:17
  with
• 14:17
  minimal
• 14:17
  human
• 14:18
  presence
• 14:18
  —
• 14:19
  is
• 14:19
  starting
• 14:19
  to
• 14:19
  intersect
• 14:20
  with
• 14:20
  AI
• 14:20
  deployment
• 14:21
  in
• 14:21
  physical
• 14:21
  infrastructure.
• 14:22
  Early
• 14:23
  demonstrations
• 14:23
  of
• 14:23
  shipping
• 14:24
  AI-driven
• 14:24
  factory
• 14:25
  automation
• 14:25
  into
• 14:25
  production
• 14:26
  highlight
• 14:26
  both
• 14:27
  the
• 14:27
  potential
• 14:27
  and
• 14:28
  the
• 14:28
  integration
• 14:28
  complexity
• 14:29
  of
• 14:29
  moving
• 14:29
  from
• 14:29
  cloud-hosted
• 14:30
  models
• 14:31
  to
• 14:31
  embedded
• 14:31
  industrial
• 14:32
  control
• 14:32
  systems
• 14:33
  where
• 14:33
  latency
• 14:34
  tolerance
• 14:34
  is
• 14:34
  measured
• 14:35
  in
• 14:35
  milliseconds
• 14:36
  and
• 14:36
  failure
• 14:36
  modes
• 14:36
  involve
• 14:37
  physical
• 14:37
  machinery,
• 14:38
  not
• 14:38
  just
• 14:38
  HTTP
• 14:39
  500
• 14:40
  errors.
• 14:41
  From
• 14:41
  orbit,
• 14:41
  DRISH-X
• 14:43
  exploits
• 14:44
  an
• 14:44
  overlooked
• 14:44
  property
• 14:45
  of
• 14:45
  Sentinel-2
• 14:45
  satellite
• 14:46
  imagery:
• 14:47
  the
• 14:47
  1.01-second
• 14:48
  timing
• 14:48
  offset
• 14:49
  between
• 14:49
  the
• 14:49
  satellite's
• 14:50
  RGB
• 14:50
  band
• 14:50
  captures
• 14:51
  creates
• 14:51
  spectral
• 14:52
  smear
• 14:52
  artifacts
• 14:52
  on
• 14:52
  moving
• 14:53
  vehicles.
• 14:54
  By
• 14:54
  training
• 14:54
  a
• 14:54
  random
• 14:55
  forest
• 14:55
  classifier
• 14:56
  on
• 14:56
  a
• 14:56
  7-feature
• 14:56
  pixel
• 14:57
  stack
• 14:57
  derived
• 14:57
  from
• 14:58
  these
• 14:58
  temporal
• 14:58
  band
• 14:58
  offsets,
• 14:59
  the
• 15:00
  system
• 15:00
  detects
• 15:00
  and
• 15:00
  counts
• 15:01
  trucks
• 15:01
  on
• 15:01
  highways
• 15:02
  without
• 15:02
  any
• 15:02
  high-resolution
• 15:03
  commercial
• 15:03
  imagery.
• 15:04
  The
• 15:04
  approach
• 15:05
  is
• 15:05
  clever
• 15:05
  —
• 15:06
  it
• 15:06
  turns
• 15:06
  what
• 15:06
  most
• 15:06
  remote
• 15:07
  sensing
• 15:07
  pipelines
• 15:07
  treat
• 15:08
  as
• 15:08
  a
• 15:08
  calibration
• 15:08
  artifact
• 15:09
  into
• 15:09
  a
• 15:09
  signal.
• 15:10
  Free
• 15:11
  10-meter
• 15:11
  resolution
• 15:11
  satellite
• 15:12
  data,
• 15:12
  refreshed
• 15:13
  every
• 15:13
  five
• 15:13
  days,
• 15:13
  becomes
• 15:14
  a
• 15:14
  freight
• 15:14
  intelligence
• 15:14
  feed.
• 15:15
  For
• 15:15
  supply
• 15:16
  chain
• 15:16
  monitoring
• 15:16
  and
• 15:17
  infrastructure
• 15:17
  planning,
• 15:18
  this
• 15:18
  is
• 15:18
  surveillance-grade
• 15:19
  capability
• 15:19
  at
• 15:19
  science-grade
• 15:20
  pricing.