Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 29, 2026

Transcript

AGI Dreams Podcast – June 29, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  OpenAI
• 0:00
  previewed
• 0:01
  GPT-5.6
• 0:01
  Sol
• 0:02
  last
• 0:02
  week
• 0:02
  --
• 0:03
  and
• 0:03
  the
• 0:03
  most
• 0:03
  important
• 0:03
  detail
• 0:04
  is
• 0:04
  not
• 0:04
  what
• 0:04
  the
• 0:04
  model
• 0:04
  can
• 0:04
  do,
• 0:05
  but
• 0:05
  who
• 0:05
  gets
• 0:05
  to
• 0:05
  use
• 0:06
  it.
• 0:06
  Sol
• 0:06
  launches
• 0:07
  as
• 0:07
  a
• 0:07
  limited
• 0:07
  preview
• 0:08
  available
• 0:08
  only
• 0:08
  to
• 0:08
  a
• 0:08
  handful
• 0:09
  of
• 0:09
  vetted
• 0:09
  US
• 0:09
  organizations,
• 0:10
  at
• 0:10
  the
• 0:10
  explicit
• 0:11
  request
• 0:11
  of
• 0:11
  the
• 0:11
  US
• 0:12
  government.
• 0:12
  The
• 0:12
  three-tier
• 0:13
  family
• 0:13
  --
• 0:14
  Sol
• 0:14
  at
• 0:14
  $5/$30
• 0:16
  per
• 0:16
  million
• 0:16
  input/output
• 0:17
  tokens,
• 0:18
  Terra
• 0:18
  at
• 0:18
  $2.50/$15,
• 0:21
  and
• 0:21
  Luna
• 0:21
  at
• 0:21
  $1/$6
• 0:22
  --
• 0:23
  slots
• 0:23
  in
• 0:23
  below
• 0:24
  Anthropic's
• 0:24
  Fable
• 0:25
  5
• 0:25
  Mythos-class
• 0:25
  pricing
• 0:26
  ($10/$50)
• 0:27
  while
• 0:28
  OpenAI
• 0:28
  claims
• 0:28
  Sol
• 0:29
  matches
• 0:29
  or
• 0:29
  exceeds
• 0:29
  Mythos
• 0:30
  on
• 0:30
  long-horizon
• 0:31
  agentic
• 0:31
  coding
• 0:31
  and
• 0:31
  multi-step
• 0:32
  reasoning.
• 0:33
  Over
• 0:33
  700,000
• 0:34
  A100-equivalent
• 0:35
  GPU
• 0:35
  hours
• 0:36
  went
• 0:36
  to
• 0:36
  automated
• 0:36
  red-teaming,
• 0:37
  and
• 0:37
  the
• 0:37
  model's
• 0:38
  cyber
• 0:38
  capabilities
• 0:39
  required
• 0:39
  layered
• 0:40
  safeguards
• 0:40
  including
• 0:40
  a
• 0:41
  government-approved
• 0:41
  limited
• 0:42
  rollout.
• 0:42
  The
• 0:43
  staggered
• 0:43
  release
• 0:43
  is
• 0:43
  a
• 0:43
  direct
• 0:44
  consequence
• 0:44
  of
• 0:45
  what
• 0:45
  happened
• 0:45
  when
• 0:45
  Fable
• 0:45
  5
• 0:46
  launched
• 0:46
  in
• 0:46
  April
• 0:47
  and
• 0:47
  the
• 0:47
  NSA
• 0:47
  reportedly
• 0:48
  pulled
• 0:48
  it
• 0:48
  within
• 0:48
  days
• 0:49
  --
• 0:49
  OpenAI
• 0:50
  is
• 0:50
  not
• 0:50
  repeating
• 0:50
  that
• 0:50
  mistake.
• 0:51
  The
• 0:51
  mid-tier
• 0:52
  story
• 0:52
  is
• 0:52
  quietly
• 0:53
  more
• 0:53
  interesting
• 0:54
  for
• 0:54
  practitioners.
• 0:55
  An
• 0:55
  independent
• 0:55
  cost-performance
• 0:56
  comparison
• 0:56
  shows
• 0:57
  Terra
• 0:57
  matching
• 0:57
  Fable
• 0:58
  5
• 0:58
  at
• 0:58
  84.3%
• 0:59
  on
• 0:59
  TerminalBench
• 1:00
  2.1
• 1:01
  while
• 1:01
  edging
• 1:01
  GPT-5.5
• 1:02
  at
• 1:02
  83.4%
• 1:03
  --
• 1:04
  at
• 1:04
  half
• 1:05
  the
• 1:05
  per-token
• 1:05
  cost
• 1:05
  of
• 1:06
  Sol.
• 1:06
  For
• 1:06
  shops
• 1:07
  running
• 1:07
  high-volume
• 1:07
  inference,
• 1:08
  Terra
• 1:08
  is
• 1:08
  the
• 1:08
  real
• 1:09
  product;
• 1:09
  Sol
• 1:10
  is
• 1:10
  the
• 1:10
  prestige
• 1:10
  benchmark.
• 1:11
  The
• 1:12
  2x
• 1:12
  cost
• 1:12
  advantage
• 1:13
  over
• 1:13
  Sol
• 1:13
  with
• 1:13
  roughly
• 1:14
  1%
• 1:14
  benchmark
• 1:15
  variance
• 1:15
  is
• 1:15
  exactly
• 1:16
  the
• 1:16
  kind
• 1:16
  of
• 1:16
  trade
• 1:16
  practitioners
• 1:17
  actually
• 1:17
  optimize
• 1:18
  for.
• 1:19
  Nobody
• 1:19
  is
• 1:19
  choosing
• 1:20
  between
• 1:20
  Sol
• 1:20
  and
• 1:20
  Terra
• 1:21
  based
• 1:21
  on
• 1:21
  capability
• 1:21
  --
• 1:22
  they
• 1:22
  are
• 1:23
  choosing
• 1:23
  based
• 1:23
  on
• 1:23
  whether
• 1:24
  the
• 1:24
  marginal
• 1:24
  quality
• 1:24
  improvement
• 1:25
  justifies
• 1:26
  doubling
• 1:26
  their
• 1:26
  inference
• 1:26
  bill.
• 1:27
  For
• 1:28
  most
• 1:28
  production
• 1:28
  workloads,
• 1:29
  it
• 1:29
  does
• 1:29
  not.
• 1:30
  But
• 1:30
  the
• 1:30
  controlled-access
• 1:31
  rollout
• 1:31
  is
• 1:32
  the
• 1:32
  signal
• 1:32
  that
• 1:32
  matters
• 1:32
  beyond
• 1:33
  pricing
• 1:33
  spreadsheets.
• 1:34
  Dean
• 1:35
  Ball,
• 1:35
  writing
• 1:35
  from
• 1:35
  inside
• 1:36
  the
• 1:36
  policy
• 1:36
  apparatus
• 1:37
  (he
• 1:37
  joins
• 1:37
  OpenAI
• 1:38
  soon),
• 1:38
  lays
• 1:39
  out
• 1:39
  what
• 1:39
  has
• 1:39
  actually
• 1:39
  happened:
• 1:40
  the
• 1:40
  US
• 1:41
  has
• 1:41
  stumbled
• 1:41
  into
• 1:41
  a
• 1:41
  de
• 1:41
  facto
• 1:42
  licensing
• 1:42
  regime
• 1:43
  for
• 1:43
  frontier
• 1:43
  AI
• 1:44
  models,
• 1:44
  without
• 1:45
  published
• 1:45
  standards,
• 1:46
  without
• 1:46
  clear
• 1:46
  timelines,
• 1:47
  and
• 1:47
  without
• 1:47
  technically
• 1:48
  expert
• 1:48
  staff
• 1:49
  to
• 1:49
  write
• 1:49
  them.
• 1:50
  The
• 1:50
  Center
• 1:50
  for
• 1:50
  AI
• 1:51
  Standards
• 1:51
  and
• 1:51
  Innovation
• 1:52
  hired
• 1:52
  someone
• 1:52
  with
• 1:52
  frontier
• 1:53
  lab
• 1:53
  experience
• 1:53
  --
• 1:54
  and
• 1:54
  fired
• 1:55
  him
• 1:55
  within
• 1:55
  days.
• 1:56
  The
• 1:56
  remaining
• 1:56
  staff
• 1:57
  have
• 1:57
  reportedly
• 1:57
  been
• 1:58
  on
• 1:58
  a
• 1:58
  stop-work
• 1:58
  order
• 1:59
  for
• 1:59
  much
• 1:59
  of
• 1:59
  the
• 1:59
  post-Mythos
• 2:00
  crisis
• 2:00
  period.
• 2:01
  Ball's
• 2:01
  central
• 2:02
  proposal
• 2:02
  --
• 2:03
  private
• 2:03
  auditing
• 2:04
  bodies
• 2:04
  that
• 2:04
  verify
• 2:05
  labs'
• 2:05
  compliance
• 2:06
  with
• 2:06
  their
• 2:06
  own
• 2:06
  safety
• 2:06
  frameworks,
• 2:07
  certified
• 2:08
  by
• 2:08
  the
• 2:08
  government
• 2:09
  but
• 2:09
  operating
• 2:09
  independently
• 2:10
  --
• 2:11
  is
• 2:11
  the
• 2:11
  most
• 2:12
  concrete
• 2:12
  governance
• 2:13
  architecture
• 2:13
  anyone
• 2:13
  has
• 2:14
  put
• 2:14
  forward
• 2:14
  since
• 2:14
  the
• 2:15
  Fable
• 2:15
  5
• 2:15
  shutdown.
• 2:16
  The
• 2:16
  key
• 2:17
  insight
• 2:17
  is
• 2:17
  entity-level
• 2:18
  regulation
• 2:19
  rather
• 2:19
  than
• 2:19
  model-level:
• 2:20
  a
• 2:21
  "model"
• 2:21
  is
• 2:22
  just
• 2:22
  floating-point
• 2:22
  numbers
• 2:23
  that
• 2:23
  get
• 2:23
  cheaper
• 2:24
  to
• 2:24
  reproduce
• 2:24
  every
• 2:24
  six
• 2:25
  months,
• 2:26
  but
• 2:26
  the
• 2:26
  lab's
• 2:26
  internal
• 2:27
  governance
• 2:27
  of
• 2:27
  recursive
• 2:28
  self-improvement
• 2:28
  is
• 2:29
  durable
• 2:29
  and
• 2:29
  auditable.
• 2:30
  He
• 2:30
  draws
• 2:31
  an
• 2:31
  uncomfortable
• 2:31
  parallel
• 2:32
  to
• 2:32
  crypto:
• 2:33
  the
• 2:33
  industry
• 2:34
  converged
• 2:34
  on
• 2:34
  preferring
• 2:35
  clear,
• 2:35
  predictable
• 2:36
  regulation
• 2:37
  over
• 2:37
  the
• 2:37
  raw
• 2:37
  exercise
• 2:38
  of
• 2:38
  state
• 2:38
  power,
• 2:39
  even
• 2:39
  if
• 2:39
  it
• 2:39
  meant
• 2:40
  accepting
• 2:40
  constraints.
• 2:42
  Representatives
• 2:42
  Obernolte
• 2:43
  and
• 2:43
  Trahan
• 2:44
  have
• 2:44
  introduced
• 2:44
  a
• 2:45
  bipartisan
• 2:45
  bill
• 2:45
  tracking
• 2:46
  this
• 2:46
  framework.
• 2:47
  Organizations
• 2:48
  like
• 2:48
  the
• 2:48
  Frontier
• 2:49
  Model
• 2:49
  Forum,
• 2:49
  AVERI,
• 2:50
  METR,
• 2:51
  Apollo
• 2:51
  Research,
• 2:52
  and
• 2:52
  Fathom
• 2:53
  already
• 2:53
  form
• 2:53
  the
• 2:53
  nucleus
• 2:54
  of
• 2:54
  the
• 2:54
  ecosystem
• 2:55
  Ball
• 2:55
  describes
• 2:55
  --
• 2:56
  the
• 2:56
  infrastructure
• 2:57
  is
• 2:57
  not
• 2:57
  hypothetical.
• 2:59
  Whether
• 2:59
  this
• 2:59
  is
• 2:59
  pragmatic
• 3:00
  institution-building
• 3:01
  or
• 3:01
  regulatory
• 3:02
  capture
• 3:02
  in
• 3:02
  academic
• 3:03
  dress
• 3:03
  depends
• 3:04
  on
• 3:04
  how
• 3:04
  much
• 3:04
  you
• 3:04
  trust
• 3:04
  the
• 3:04
  labs
• 3:05
  to
• 3:05
  write
• 3:05
  their
• 3:05
  own
• 3:05
  safety
• 3:06
  standards,
• 3:07
  but
• 3:07
  the
• 3:07
  alternative
• 3:08
  is
• 3:08
  an
• 3:08
  indefinite
• 3:08
  freeze
• 3:09
  with
• 3:09
  no
• 3:09
  clear
• 3:09
  exit
• 3:10
  and
• 3:10
  a
• 3:10
  growing
• 3:10
  risk
• 3:10
  of
• 3:11
  market
• 3:11
  panic
• 3:11
  as
• 3:11
  $100
• 3:12
  billion
• 3:12
  data
• 3:12
  centers
• 3:13
  sit
• 3:13
  idle.
• 3:17
  DeepSeek
• 3:17
  V4
• 3:18
  Pro
• 3:18
  shipped
• 3:18
  DSpark,
• 3:19
  a
• 3:19
  speculative
• 3:20
  decoding
• 3:20
  module
• 3:20
  that
• 3:20
  promises
• 3:21
  5x
• 3:21
  cheaper
• 3:22
  serving
• 3:22
  for
• 3:22
  short-context
• 3:23
  workloads
• 3:23
  and
• 3:24
  7.6x
• 3:24
  for
• 3:25
  long-context.
• 3:26
  Speculative
• 3:26
  decoding
• 3:26
  --
• 3:27
  using
• 3:27
  a
• 3:27
  small
• 3:28
  draft
• 3:28
  model
• 3:28
  to
• 3:29
  propose
• 3:29
  tokens
• 3:29
  that
• 3:29
  the
• 3:29
  full
• 3:30
  model
• 3:30
  verifies
• 3:31
  in
• 3:31
  parallel
• 3:31
  --
• 3:32
  has
• 3:32
  been
• 3:32
  a
• 3:32
  research
• 3:32
  technique
• 3:33
  for
• 3:33
  two
• 3:33
  years,
• 3:34
  but
• 3:34
  DSpark
• 3:34
  packages
• 3:35
  it
• 3:35
  as
• 3:35
  a
• 3:35
  first-party
• 3:36
  module
• 3:36
  for
• 3:36
  one
• 3:36
  of
• 3:36
  the
• 3:36
  most
• 3:37
  deployed
• 3:37
  open-weight
• 3:37
  models.
• 3:38
  The
• 3:38
  economics
• 3:39
  are
• 3:39
  straightforward:
• 3:40
  if
• 3:40
  your
• 3:40
  V4
• 3:41
  Pro
• 3:41
  inference
• 3:41
  bill
• 3:41
  is
• 3:41
  $10K/month,
• 3:44
  DSpark
• 3:44
  could
• 3:44
  cut
• 3:44
  it
• 3:44
  to
• 3:45
  $1.3K-$2K
• 3:46
  without
• 3:47
  touching
• 3:47
  model
• 3:47
  quality.
• 3:48
  That
• 3:49
  is
• 3:49
  the
• 3:49
  kind
• 3:49
  of
• 3:49
  cost
• 3:49
  reduction
• 3:50
  that
• 3:50
  changes
• 3:50
  deployment
• 3:51
  decisions.
• 3:52
  For
• 3:52
  the
• 3:52
  hardware-maximalist
• 3:53
  crowd,
• 3:53
  a
• 3:53
  detailed
• 3:54
  guide
• 3:54
  walks
• 3:54
  through
• 3:55
  running
• 3:55
  a
• 3:55
  high-quality
• 3:55
  GLM-5.2
• 3:56
  quant
• 3:57
  at
• 3:57
  128K
• 3:58
  context
• 3:59
  on
• 3:59
  four
• 3:59
  DGX
• 4:00
  Spark
• 4:00
  nodes.
• 4:01
  The
• 4:01
  setup
• 4:01
  involves
• 4:02
  NVFP4
• 4:04
  quantization,
• 4:05
  a
• 4:05
  patched
• 4:05
  vLLM
• 4:06
  (the
• 4:06
  stock
• 4:06
  version
• 4:06
  does
• 4:06
  not
• 4:07
  handle
• 4:07
  GLM-5.2's
• 4:08
  attention
• 4:09
  layout
• 4:09
  correctly),
• 4:10
  and
• 4:10
  careful
• 4:10
  memory
• 4:10
  management
• 4:11
  to
• 4:11
  sustain
• 4:11
  roughly
• 4:12
  15
• 4:12
  tokens
• 4:12
  per
• 4:13
  second.
• 4:14
  The
• 4:14
  comparison
• 4:15
  target
• 4:15
  is
• 4:15
  Apple's
• 4:15
  M3
• 4:16
  Ultra
• 4:16
  --
• 4:17
  and
• 4:17
  the
• 4:17
  DGX
• 4:17
  Spark
• 4:18
  cluster
• 4:18
  wins
• 4:18
  on
• 4:18
  throughput
• 4:19
  but
• 4:19
  loses
• 4:19
  on
• 4:19
  watts-per-token
• 4:20
  and
• 4:20
  total
• 4:20
  cost
• 4:21
  of
• 4:21
  ownership.
• 4:22
  The
• 4:22
  real
• 4:22
  lesson
• 4:23
  is
• 4:23
  that
• 4:23
  "running
• 4:23
  it
• 4:23
  locally"
• 4:24
  now
• 4:24
  requires
• 4:25
  specifying
• 4:26
  which
• 4:26
  "local"
• 4:26
  you
• 4:27
  mean:
• 4:27
  a
• 4:27
  $40K
• 4:28
  workstation
• 4:29
  cluster
• 4:29
  is
• 4:29
  local
• 4:30
  in
• 4:30
  the
• 4:30
  sense
• 4:30
  that
• 4:30
  it
• 4:30
  sits
• 4:31
  in
• 4:31
  your
• 4:31
  rack,
• 4:31
  but
• 4:32
  the
• 4:32
  economics
• 4:32
  are
• 4:32
  closer
• 4:33
  to
• 4:33
  a
• 4:33
  small
• 4:33
  cloud
• 4:34
  deployment
• 4:34
  than
• 4:34
  a
• 4:34
  laptop
• 4:35
  experiment.
• 4:36
  The
• 4:36
  guide
• 4:36
  is
• 4:36
  valuable
• 4:37
  as
• 4:37
  a
• 4:37
  reference
• 4:37
  for
• 4:37
  anyone
• 4:38
  considering
• 4:38
  DGX
• 4:39
  Spark
• 4:39
  purchases,
• 4:40
  but
• 4:40
  it
• 4:40
  also
• 4:40
  illustrates
• 4:41
  how
• 4:41
  far
• 4:41
  "local
• 4:42
  inference"
• 4:42
  has
• 4:42
  drifted
• 4:42
  from
• 4:43
  its
• 4:43
  hobbyist
• 4:43
  roots.
• 4:45
  Ornith
• 4:45
  35B
• 4:46
  arrives
• 4:46
  with
• 4:46
  strong
• 4:47
  early
• 4:47
  results
• 4:47
  on
• 4:47
  greenfield
• 4:48
  coding
• 4:48
  --
• 4:49
  generating
• 4:49
  new
• 4:50
  projects
• 4:50
  from
• 4:50
  scratch
• 4:50
  --
• 4:51
  while
• 4:51
  falling
• 4:52
  short
• 4:52
  on
• 4:52
  repair
• 4:52
  and
• 4:53
  debugging
• 4:53
  tasks
• 4:53
  where
• 4:54
  models
• 4:54
  need
• 4:54
  to
• 4:54
  reason
• 4:55
  about
• 4:55
  existing
• 4:55
  code
• 4:55
  structure.
• 4:57
  Community
• 4:57
  testing
• 4:58
  shows
• 4:58
  it
• 4:58
  handles
• 4:58
  instruction-following
• 4:59
  and
• 4:59
  creative
• 5:00
  generation
• 5:00
  well
• 5:01
  at
• 5:01
  its
• 5:01
  size
• 5:01
  class,
• 5:02
  but
• 5:02
  struggles
• 5:02
  with
• 5:03
  the
• 5:03
  kind
• 5:03
  of
• 5:03
  "read
• 5:03
  this
• 5:03
  broken
• 5:04
  codebase
• 5:04
  and
• 5:04
  fix
• 5:05
  it"
• 5:05
  work
• 5:06
  that
• 5:06
  separates
• 5:06
  coding
• 5:06
  benchmarks
• 5:07
  from
• 5:07
  coding
• 5:07
  reality.
• 5:09
  The
• 5:09
  greenfield-vs-repair
• 5:10
  split
• 5:10
  is
• 5:11
  worth
• 5:11
  tracking:
• 5:11
  most
• 5:12
  coding
• 5:12
  benchmark
• 5:12
  suites
• 5:13
  over-weight
• 5:13
  greenfield
• 5:14
  generation,
• 5:15
  which
• 5:15
  means
• 5:15
  models
• 5:16
  optimized
• 5:16
  for
• 5:16
  benchmarks
• 5:17
  may
• 5:17
  disappoint
• 5:18
  practitioners
• 5:19
  whose
• 5:19
  daily
• 5:19
  work
• 5:19
  is
• 5:20
  overwhelmingly
• 5:20
  maintenance
• 5:21
  and
• 5:21
  debugging.
• 5:22
  A
• 5:22
  thoughtful
• 5:23
  book
• 5:23
  review
• 5:23
  of
• 5:23
  Guglielmo
• 5:24
  Iozzia's
• 5:24
  Domain-Specific
• 5:25
  Small
• 5:25
  Language
• 5:26
  Models
• 5:26
  captures
• 5:27
  where
• 5:27
  the
• 5:27
  SLM
• 5:28
  community
• 5:28
  actually
• 5:29
  stands.
• 5:30
  The
• 5:30
  reviewer
• 5:30
  --
• 5:31
  an
• 5:31
  AI
• 5:32
  practitioner
• 5:32
  since
• 5:32
  1999
• 5:33
  with
• 5:34
  50+
• 5:34
  deployed
• 5:35
  systems
• 5:35
  --
• 5:36
  endorses
• 5:37
  the
• 5:37
  core
• 5:37
  thesis
• 5:37
  (domain-tuned
• 5:38
  small
• 5:38
  models
• 5:39
  beat
• 5:39
  renting
• 5:40
  generic
• 5:40
  intelligence)
• 5:41
  while
• 5:41
  flagging
• 5:41
  what
• 5:41
  the
• 5:42
  book
• 5:42
  undersells:
• 5:43
  fine-tuning
• 5:44
  is
• 5:44
  10-100x
• 5:45
  the
• 5:45
  effort
• 5:45
  of
• 5:45
  API
• 5:46
  calls,
• 5:47
  the
• 5:47
  required
• 5:47
  skills
• 5:48
  are
• 5:48
  rare,
• 5:49
  and
• 5:49
  evaluation
• 5:49
  should
• 5:49
  be
• 5:50
  the
• 5:50
  spine
• 5:50
  of
• 5:50
  any
• 5:50
  SLM
• 5:51
  project,
• 5:52
  not
• 5:52
  an
• 5:52
  appendix.
• 5:53
  The
• 5:53
  most
• 5:54
  useful
• 5:54
  framing
• 5:54
  from
• 5:55
  the
• 5:55
  author's
• 5:55
  talk:
• 5:56
  "from
• 5:56
  renting
• 5:56
  intelligence
• 5:57
  to
• 5:57
  owning
• 5:57
  it,
• 5:57
  from
• 5:58
  general
• 5:58
  capability
• 5:58
  to
• 5:59
  specific
• 5:59
  mastery,
• 6:00
  from
• 6:00
  centralized
• 6:01
  intelligence
• 6:01
  to
• 6:01
  distributed
• 6:02
  intelligence."
• 6:03
  That
• 6:03
  shift
• 6:03
  is
• 6:04
  real,
• 6:04
  even
• 6:05
  if
• 6:05
  the
• 6:05
  execution
• 6:06
  difficulty
• 6:06
  is
• 6:06
  consistently
• 6:07
  underestimated
• 6:08
  by
• 6:08
  both
• 6:08
  vendors
• 6:09
  and
• 6:09
  authors.
• 6:12
  SpectralQuant
• 6:13
  introduces
• 6:13
  calibration-aware
• 6:14
  Q4
• 6:14
  K
• 6:15
  M
• 6:15
  quantization
• 6:16
  and
• 6:16
  claims
• 6:16
  96.5%
• 6:17
  BF16
• 6:18
  gap
• 6:19
  recovery
• 6:19
  --
• 6:20
  meaning
• 6:20
  the
• 6:20
  quantized
• 6:20
  model
• 6:21
  retains
• 6:21
  96.5%
• 6:22
  of
• 6:22
  the
• 6:22
  quality
• 6:23
  difference
• 6:23
  between
• 6:24
  a
• 6:24
  naive
• 6:24
  quant
• 6:25
  and
• 6:25
  the
• 6:25
  full-precision
• 6:26
  original.
• 6:27
  Instead
• 6:27
  of
• 6:27
  treating
• 6:27
  all
• 6:28
  weights
• 6:28
  equally
• 6:28
  during
• 6:29
  quantization,
• 6:30
  SpectralQuant
• 6:30
  uses
• 6:31
  spectral
• 6:31
  analysis
• 6:32
  to
• 6:32
  identify
• 6:32
  which
• 6:33
  weight
• 6:33
  matrices
• 6:33
  carry
• 6:34
  the
• 6:34
  most
• 6:34
  information
• 6:34
  and
• 6:35
  allocates
• 6:35
  precision
• 6:35
  budget
• 6:36
  accordingly.
• 6:37
  This
• 6:37
  publication
• 6:37
  has
• 6:38
  covered
• 6:38
  calibration
• 6:39
  extensively
• 6:39
  --
• 6:39
  from
• 6:40
  imatrix
• 6:40
  ladders
• 6:40
  showing
• 6:41
  10%
• 6:41
  accuracy
• 6:42
  gaps
• 6:42
  purely
• 6:42
  from
• 6:42
  calibration
• 6:43
  data
• 6:43
  selection,
• 6:44
  to
• 6:44
  practitioners
• 6:45
  saturating
• 6:45
  models
• 6:46
  with
• 6:46
  2
• 6:46
  million
• 6:46
  tokens
• 6:47
  of
• 6:47
  calibration
• 6:47
  data
• 6:48
  to
• 6:48
  eliminate
• 6:48
  quantization
• 6:49
  slop.
• 6:49
  SpectralQuant
• 6:50
  enters
• 6:50
  a
• 6:50
  crowded
• 6:51
  field,
• 6:51
  but
• 6:52
  the
• 6:52
  spectral
• 6:52
  approach
• 6:52
  is
• 6:53
  a
• 6:53
  genuinely
• 6:53
  different
• 6:53
  technique
• 6:54
  from
• 6:54
  the
• 6:54
  importance-weighted
• 6:55
  methods
• 6:55
  we
• 6:55
  have
• 6:55
  tracked,
• 6:56
  and
• 6:56
  96.5%
• 6:57
  recovery
• 6:58
  at
• 6:58
  Q4
• 6:58
  precision
• 6:59
  is
• 6:59
  a
• 6:59
  strong
• 7:00
  result
• 7:00
  if
• 7:00
  it
• 7:00
  replicates
• 7:01
  across
• 7:01
  model
• 7:01
  families.
• 7:02
  Procedural
• 7:03
  Skill
• 7:03
  Transfer
• 7:03
  takes
• 7:03
  a
• 7:03
  different
• 7:04
  angle
• 7:04
  on
• 7:04
  the
• 7:04
  small-model
• 7:05
  problem:
• 7:06
  instead
• 7:06
  of
• 7:06
  fine-tuning
• 7:07
  or
• 7:07
  distilling,
• 7:08
  scaffold
• 7:08
  the
• 7:08
  smaller
• 7:08
  model
• 7:09
  at
• 7:09
  inference
• 7:09
  time
• 7:09
  using
• 7:10
  structured
• 7:10
  procedures
• 7:11
  extracted
• 7:11
  from
• 7:11
  a
• 7:11
  larger
• 7:12
  model.
• 7:13
  Early
• 7:13
  manual
• 7:13
  testing
• 7:14
  shows
• 7:14
  promising
• 7:15
  results
• 7:15
  on
• 7:15
  tasks
• 7:15
  where
• 7:16
  the
• 7:16
  small
• 7:16
  model
• 7:16
  has
• 7:16
  the
• 7:16
  latent
• 7:17
  capability
• 7:17
  but
• 7:18
  lacks
• 7:18
  the
• 7:18
  procedural
• 7:18
  knowledge
• 7:19
  to
• 7:19
  apply
• 7:19
  it
• 7:19
  consistently.
• 7:20
  The
• 7:20
  distinction
• 7:21
  from
• 7:21
  distillation
• 7:22
  matters
• 7:22
  --
• 7:23
  no
• 7:23
  training
• 7:23
  run
• 7:23
  required,
• 7:24
  no
• 7:24
  gradient
• 7:25
  updates,
• 7:26
  just
• 7:26
  a
• 7:26
  structured
• 7:26
  prompt
• 7:27
  scaffold
• 7:27
  that
• 7:27
  the
• 7:27
  small
• 7:28
  model
• 7:28
  follows.
• 7:29
  If
• 7:29
  the
• 7:29
  results
• 7:30
  hold
• 7:30
  at
• 7:30
  scale,
• 7:30
  this
• 7:31
  is
• 7:31
  a
• 7:31
  cheaper
• 7:31
  and
• 7:31
  more
• 7:31
  accessible
• 7:32
  path
• 7:32
  to
• 7:32
  capability
• 7:33
  transfer
• 7:33
  than
• 7:33
  any
• 7:34
  fine-tuning
• 7:34
  approach,
• 7:35
  because
• 7:35
  it
• 7:35
  requires
• 7:36
  zero
• 7:36
  ML
• 7:37
  infrastructure
• 7:37
  beyond
• 7:38
  inference.
• 7:39
  Multi-Tier
• 7:39
  MoE
• 7:40
  Caching
• 7:40
  formalizes
• 7:41
  what
• 7:41
  llama.cpp
• 7:43
  has
• 7:43
  been
• 7:43
  doing
• 7:43
  heuristically:
• 7:44
  instead
• 7:45
  of
• 7:45
  treating
• 7:45
  all
• 7:45
  mixture-of-experts
• 7:46
  experts
• 7:47
  equally,
• 7:47
  profile
• 7:48
  activation
• 7:49
  patterns
• 7:49
  and
• 7:49
  pin
• 7:49
  the
• 7:50
  top
• 7:50
  20%
• 7:50
  of
• 7:50
  experts
• 7:51
  --
• 7:52
  which
• 7:52
  handle
• 7:52
  roughly
• 7:53
  85%
• 7:53
  of
• 7:54
  all
• 7:54
  activations
• 7:54
  --
• 7:55
  in
• 7:55
  fast
• 7:55
  memory
• 7:56
  while
• 7:56
  allowing
• 7:56
  the
• 7:57
  rest
• 7:57
  to
• 7:57
  swap
• 7:57
  from
• 7:57
  slower
• 7:58
  tiers.
• 7:59
  The
• 7:59
  contribution
• 8:00
  over
• 8:00
  existing
• 8:01
  hot/cold
• 8:01
  caching
• 8:02
  is
• 8:02
  the
• 8:02
  graduated
• 8:03
  allocation
• 8:04
  policy:
• 8:05
  not
• 8:05
  just
• 8:05
  two
• 8:05
  buckets
• 8:06
  but
• 8:06
  a
• 8:06
  tiered
• 8:06
  hierarchy
• 8:07
  that
• 8:07
  matches
• 8:07
  memory
• 8:07
  costs
• 8:08
  to
• 8:08
  expert
• 8:08
  importance.
• 8:09
  For
• 8:10
  Qwen-class
• 8:10
  MoE
• 8:10
  models
• 8:11
  on
• 8:11
  consumer
• 8:11
  hardware,
• 8:12
  this
• 8:13
  is
• 8:13
  the
• 8:13
  difference
• 8:13
  between
• 8:13
  "technically
• 8:14
  fits
• 8:14
  in
• 8:14
  memory"
• 8:15
  and
• 8:15
  "actually
• 8:16
  runs
• 8:16
  at
• 8:16
  usable
• 8:16
  speed."
• 8:19
  CVE-bench
• 8:20
  takes
• 8:20
  the
• 8:20
  SWE-bench
• 8:21
  template
• 8:21
  --
• 8:22
  give
• 8:22
  a
• 8:22
  model
• 8:23
  a
• 8:23
  bug
• 8:23
  description
• 8:23
  and
• 8:24
  a
• 8:24
  codebase,
• 8:24
  measure
• 8:25
  whether
• 8:25
  it
• 8:25
  can
• 8:25
  produce
• 8:25
  a
• 8:25
  working
• 8:26
  exploit
• 8:26
  or
• 8:26
  fix
• 8:26
  --
• 8:27
  and
• 8:27
  applies
• 8:28
  it
• 8:28
  to
• 8:28
  real
• 8:28
  CVEs.
• 8:29
  Five
• 8:29
  validated
• 8:30
  vulnerabilities
• 8:31
  form
• 8:31
  the
• 8:31
  initial
• 8:31
  benchmark,
• 8:32
  each
• 8:32
  with
• 8:32
  a
• 8:32
  Section
• 8:33
  42
• 8:33
  discrimination
• 8:34
  gate
• 8:34
  that
• 8:34
  tests
• 8:35
  whether
• 8:35
  the
• 8:35
  model
• 8:35
  can
• 8:35
  distinguish
• 8:36
  between
• 8:36
  vulnerable
• 8:37
  and
• 8:37
  patched
• 8:37
  code
• 8:38
  before
• 8:38
  attempting
• 8:38
  exploitation.
• 8:39
  This
• 8:40
  gate
• 8:40
  matters:
• 8:40
  it
• 8:41
  separates
• 8:41
  models
• 8:41
  that
• 8:42
  understand
• 8:42
  the
• 8:42
  vulnerability
• 8:43
  from
• 8:43
  models
• 8:43
  that
• 8:43
  blindly
• 8:44
  try
• 8:44
  payloads.
• 8:45
  The
• 8:45
  benchmark
• 8:46
  design
• 8:46
  addresses
• 8:46
  a
• 8:46
  persistent
• 8:47
  gap
• 8:47
  in
• 8:47
  security
• 8:48
  AI
• 8:48
  evaluation.
• 8:49
  Coding
• 8:50
  AI
• 8:50
  got
• 8:50
  SWE-bench
• 8:50
  two
• 8:51
  years
• 8:51
  ago;
• 8:52
  security
• 8:52
  AI
• 8:52
  has
• 8:52
  been
• 8:52
  measuring
• 8:53
  itself
• 8:53
  with
• 8:53
  ad
• 8:53
  hoc
• 8:54
  CTF
• 8:54
  challenges
• 8:55
  and
• 8:55
  vendor
• 8:55
  demos
• 8:55
  ever
• 8:55
  since.
• 8:56
  CVE-bench
• 8:57
  imposes
• 8:57
  the
• 8:58
  same
• 8:58
  rigor:
• 8:58
  reproducible,
• 8:59
  automated,
• 9:00
  scored
• 9:01
  against
• 9:01
  ground
• 9:01
  truth.
• 9:02
  The
• 9:02
  RedBlue
• 9:03
  hackbot
• 9:03
  project
• 9:03
  demonstrates
• 9:04
  autonomous
• 9:04
  vulnerability
• 9:05
  discovery
• 9:05
  in
• 9:05
  a
• 9:05
  controlled
• 9:06
  environment
• 9:06
  --
• 9:07
  the
• 9:07
  kind
• 9:07
  of
• 9:07
  red-team
• 9:08
  automation
• 9:08
  that
• 9:08
  Sol's
• 9:09
  700,000
• 9:09
  GPU-hour
• 9:10
  red-teaming
• 9:10
  effort
• 9:11
  relies
• 9:11
  on
• 9:11
  at
• 9:11
  massive
• 9:11
  scale.
• 9:12
  The
• 9:13
  creator
• 9:13
  frames
• 9:13
  it
• 9:13
  as
• 9:14
  "for
• 9:14
  fun
• 9:14
  and
• 9:14
  share,"
• 9:15
  which
• 9:15
  undersells
• 9:16
  the
• 9:16
  implication:
• 9:17
  if
• 9:17
  a
• 9:17
  weekend
• 9:18
  project
• 9:18
  can
• 9:18
  assemble
• 9:19
  autonomous
• 9:20
  exploitation
• 9:20
  agents
• 9:21
  from
• 9:21
  off-the-shelf
• 9:21
  components,
• 9:22
  the
• 9:23
  offensive
• 9:23
  toolkit
• 9:23
  is
• 9:24
  commoditizing
• 9:24
  faster
• 9:25
  than
• 9:25
  defensive
• 9:25
  tooling
• 9:26
  can
• 9:26
  keep
• 9:26
  pace.
• 9:27
  The
• 9:27
  question
• 9:28
  is
• 9:28
  not
• 9:28
  whether
• 9:28
  autonomous
• 9:29
  security
• 9:29
  testing
• 9:29
  will
• 9:30
  become
• 9:30
  ubiquitous,
• 9:31
  but
• 9:31
  whether
• 9:31
  the
• 9:31
  defenders
• 9:32
  or
• 9:32
  attackers
• 9:32
  adopt
• 9:33
  it
• 9:33
  faster.
• 9:34
  The
• 9:34
  same
• 9:34
  project
• 9:34
  ships
• 9:35
  as
• 9:35
  @metaharness/redblue
• 9:36
  on
• 9:36
  npm
• 9:37
  --
• 9:38
  an
• 9:38
  autonomous
• 9:38
  red-team/blue-team
• 9:39
  agent
• 9:40
  available
• 9:40
  as
• 9:40
  a
• 9:40
  standard
• 9:41
  JavaScript
• 9:41
  package
• 9:41
  install.
• 9:42
  The
• 9:43
  distribution
• 9:43
  choice
• 9:43
  is
• 9:43
  the
• 9:44
  story:
• 9:44
  npm
• 9:45
  installation
• 9:45
  normalizes
• 9:46
  security
• 9:46
  testing
• 9:47
  tools
• 9:47
  the
• 9:47
  way
• 9:47
  eslint
• 9:48
  normalized
• 9:48
  linting.
• 9:49
  Whether
• 9:50
  that
• 9:50
  democratization
• 9:51
  net
• 9:51
  improves
• 9:52
  security
• 9:52
  (more
• 9:52
  defenders
• 9:53
  than
• 9:53
  attackers
• 9:53
  use
• 9:53
  linters)
• 9:54
  or
• 9:54
  degrades
• 9:54
  it
• 9:54
  (script
• 9:55
  kiddies
• 9:55
  get
• 9:56
  one-click
• 9:56
  exploitation
• 9:57
  frameworks)
• 9:57
  depends
• 9:58
  entirely
• 9:58
  on
• 9:58
  the
• 9:58
  ratio
• 9:59
  of
• 9:59
  constructive
• 9:59
  to
• 10:00
  destructive
• 10:00
  users,
• 10:01
  and
• 10:01
  history
• 10:01
  suggests
• 10:02
  both
• 10:02
  happen
• 10:02
  simultaneously.
• 10:05
  Phantoms
• 10:06
  and
• 10:06
  Disclosures
• 10:06
  proposes
• 10:07
  a
• 10:07
  causal
• 10:07
  framework
• 10:08
  for
• 10:08
  auditing
• 10:08
  synthetic
• 10:09
  data
• 10:09
  privacy
• 10:09
  --
• 10:10
  moving
• 10:11
  beyond
• 10:11
  the
• 10:11
  usual
• 10:11
  "does
• 10:11
  it
• 10:12
  memorize
• 10:12
  training
• 10:12
  examples?"
• 10:13
  question
• 10:13
  to
• 10:13
  a
• 10:13
  structural
• 10:14
  analysis
• 10:14
  of
• 10:14
  what
• 10:14
  information
• 10:15
  flows
• 10:15
  are
• 10:15
  possible
• 10:16
  given
• 10:16
  the
• 10:16
  data
• 10:16
  generation
• 10:17
  pipeline.
• 10:18
  The
• 10:18
  framework
• 10:18
  treats
• 10:18
  synthetic
• 10:19
  data
• 10:19
  generators
• 10:19
  as
• 10:20
  causal
• 10:20
  models
• 10:20
  and
• 10:20
  asks
• 10:21
  whether
• 10:21
  the
• 10:21
  generator's
• 10:21
  structure
• 10:22
  permits
• 10:22
  leaking
• 10:22
  specific
• 10:23
  training
• 10:23
  records,
• 10:24
  even
• 10:24
  when
• 10:24
  empirical
• 10:25
  tests
• 10:25
  fail
• 10:25
  to
• 10:25
  detect
• 10:26
  memorization.
• 10:27
  This
• 10:27
  is
• 10:27
  the
• 10:27
  right
• 10:28
  level
• 10:28
  of
• 10:28
  analysis:
• 10:29
  empirical
• 10:29
  privacy
• 10:30
  tests
• 10:30
  are
• 10:30
  necessary
• 10:31
  but
• 10:31
  fundamentally
• 10:31
  insufficient,
• 10:33
  because
• 10:33
  they
• 10:33
  can
• 10:33
  only
• 10:33
  find
• 10:33
  leaks
• 10:34
  they
• 10:34
  are
• 10:34
  designed
• 10:34
  to
• 10:34
  look
• 10:34
  for.
• 10:35
  A
• 10:35
  generator
• 10:36
  that
• 10:36
  passes
• 10:36
  every
• 10:37
  membership
• 10:37
  inference
• 10:37
  attack
• 10:37
  today
• 10:38
  might
• 10:38
  still
• 10:38
  leak
• 10:38
  records
• 10:39
  through
• 10:39
  a
• 10:39
  causal
• 10:39
  pathway
• 10:40
  that
• 10:40
  no
• 10:40
  current
• 10:40
  test
• 10:40
  probes.
• 10:41
  The
• 10:41
  paper
• 10:42
  provides
• 10:42
  the
• 10:42
  mathematical
• 10:43
  machinery
• 10:43
  to
• 10:43
  reason
• 10:43
  about
• 10:44
  these
• 10:44
  structural
• 10:44
  guarantees
• 10:45
  rather
• 10:45
  than
• 10:46
  relying
• 10:46
  on
• 10:46
  empirical
• 10:47
  absence
• 10:47
  of
• 10:47
  evidence.
• 10:48
  A
• 10:48
  developer
• 10:49
  running
• 10:49
  a
• 10:49
  local
• 10:50
  LLM
• 10:50
  application
• 10:50
  asks
• 10:51
  the
• 10:51
  question
• 10:51
  every
• 10:52
  privacy-conscious
• 10:52
  builder
• 10:53
  faces:
• 10:54
  how
• 10:54
  do
• 10:54
  you
• 10:54
  prove
• 10:54
  you
• 10:54
  don't
• 10:54
  collect
• 10:55
  data?
• 10:56
  The
• 10:56
  community
• 10:56
  response
• 10:56
  coalesces
• 10:57
  around
• 10:57
  a
• 10:57
  practical
• 10:58
  stack
• 10:58
  --
• 10:59
  open-source
• 10:59
  model
• 10:59
  weights,
• 11:00
  auditable
• 11:00
  code,
• 11:01
  network
• 11:01
  traffic
• 11:01
  analysis,
• 11:02
  and
• 11:02
  attestation
• 11:03
  logs
• 11:03
  --
• 11:04
  while
• 11:04
  acknowledging
• 11:04
  the
• 11:04
  fundamental
• 11:05
  asymmetry:
• 11:06
  proving
• 11:06
  a
• 11:06
  negative
• 11:06
  is
• 11:06
  always
• 11:07
  harder
• 11:07
  than
• 11:07
  proving
• 11:07
  a
• 11:07
  positive.
• 11:09
  The
• 11:09
  thread
• 11:09
  is
• 11:09
  more
• 11:09
  valuable
• 11:10
  as
• 11:10
  a
• 11:10
  requirements
• 11:10
  document
• 11:11
  than
• 11:11
  a
• 11:11
  solution:
• 11:12
  it
• 11:12
  catalogs
• 11:13
  what
• 11:13
  users
• 11:13
  actually
• 11:13
  want
• 11:14
  to
• 11:14
  verify,
• 11:14
  which
• 11:15
  is
• 11:15
  prerequisite
• 11:16
  knowledge
• 11:16
  for
• 11:16
  anyone
• 11:16
  building
• 11:17
  privacy-preserving
• 11:18
  AI
• 11:18
  tooling.
• 11:19
  The
• 11:19
  consensus
• 11:19
  is
• 11:20
  that
• 11:20
  no
• 11:20
  single
• 11:20
  mechanism
• 11:21
  suffices;
• 11:22
  you
• 11:22
  need
• 11:22
  layers
• 11:22
  of
• 11:23
  verifiable
• 11:23
  evidence,
• 11:23
  and
• 11:24
  even
• 11:24
  then,
• 11:24
  some
• 11:24
  users
• 11:24
  will
• 11:25
  (rationally)
• 11:25
  remain
• 11:25
  skeptical.
• 11:26
  Cloudflare
• 11:27
  launched
• 11:27
  self-managed
• 11:28
  OAuth
• 11:28
  for
• 11:29
  all
• 11:29
  customers,
• 11:29
  letting
• 11:30
  any
• 11:30
  application
• 11:30
  behind
• 11:31
  Cloudflare
• 11:31
  issue
• 11:31
  and
• 11:32
  manage
• 11:32
  its
• 11:32
  own
• 11:32
  OAuth
• 11:33
  tokens
• 11:33
  without
• 11:33
  a
• 11:33
  third-party
• 11:34
  identity
• 11:34
  provider.
• 11:35
  For
• 11:36
  AI
• 11:36
  agent
• 11:36
  builders,
• 11:37
  this
• 11:37
  solves
• 11:37
  a
• 11:37
  specific
• 11:37
  pain
• 11:38
  point:
• 11:39
  agents
• 11:39
  that
• 11:39
  need
• 11:39
  to
• 11:39
  authenticate
• 11:40
  to
• 11:40
  services
• 11:40
  on
• 11:41
  behalf
• 11:41
  of
• 11:41
  users
• 11:41
  currently
• 11:42
  require
• 11:42
  either
• 11:42
  storing
• 11:43
  user
• 11:43
  credentials
• 11:43
  (terrible)
• 11:44
  or
• 11:44
  complex
• 11:45
  OAuth
• 11:45
  dance
• 11:46
  integration
• 11:46
  with
• 11:47
  external
• 11:47
  providers
• 11:48
  (expensive
• 11:48
  and
• 11:48
  brittle).
• 11:49
  Self-managed
• 11:50
  OAuth
• 11:50
  collapses
• 11:51
  the
• 11:51
  integration
• 11:51
  to
• 11:52
  a
• 11:52
  single
• 11:52
  Cloudflare
• 11:52
  configuration.
• 11:54
  The
• 11:54
  timing
• 11:55
  is
• 11:55
  not
• 11:55
  accidental
• 11:55
  --
• 11:56
  as
• 11:56
  agentic
• 11:57
  AI
• 11:57
  systems
• 11:57
  proliferate,
• 11:58
  the
• 11:59
  authorization
• 11:59
  layer
• 11:59
  becomes
• 12:00
  the
• 12:00
  critical
• 12:00
  infrastructure,
• 12:01
  and
• 12:01
  Cloudflare
• 12:02
  is
• 12:02
  positioning
• 12:03
  itself
• 12:03
  as
• 12:03
  the
• 12:03
  default.
• 12:07
  TxBench-PP
• 12:08
  benchmarks
• 12:08
  AI
• 12:08
  agent
• 12:09
  performance
• 12:09
  on
• 12:09
  small-molecule
• 12:10
  preclinical
• 12:10
  pharmacology
• 12:11
  --
• 12:11
  a
• 12:11
  domain
• 12:12
  where
• 12:12
  the
• 12:12
  gap
• 12:12
  between
• 12:12
  "sounds
• 12:13
  right"
• 12:13
  and
• 12:13
  "is
• 12:13
  right"
• 12:14
  can
• 12:14
  kill
• 12:14
  a
• 12:14
  drug
• 12:14
  program.
• 12:15
  The
• 12:15
  benchmark
• 12:16
  spans
• 12:16
  1,120
• 12:17
  samples
• 12:18
  covering
• 12:18
  pharmacokinetics,
• 12:19
  ADMET
• 12:19
  properties,
• 12:20
  and
• 12:20
  drug-drug
• 12:21
  interactions,
• 12:22
  with
• 12:22
  the
• 12:22
  best-performing
• 12:22
  model
• 12:23
  (Opus
• 12:23
  4.8/Pi)
• 12:24
  hitting
• 12:24
  only
• 12:24
  59.3%
• 12:26
  accuracy.
• 12:26
  That
• 12:27
  ceiling
• 12:27
  is
• 12:27
  the
• 12:27
  finding:
• 12:28
  even
• 12:29
  the
• 12:29
  most
• 12:29
  capable
• 12:29
  frontier
• 12:30
  models
• 12:30
  are
• 12:30
  functionally
• 12:31
  unreliable
• 12:32
  for
• 12:32
  pharmacological
• 12:32
  reasoning
• 12:33
  without
• 12:33
  domain-specific
• 12:34
  scaffolding.
• 12:35
  For
• 12:35
  pharma
• 12:36
  teams
• 12:36
  evaluating
• 12:36
  AI
• 12:37
  co-pilots,
• 12:37
  59.3%
• 12:38
  means
• 12:39
  roughly
• 12:39
  4
• 12:39
  in
• 12:39
  10
• 12:39
  answers
• 12:40
  are
• 12:40
  wrong
• 12:40
  --
• 12:40
  not
• 12:41
  a
• 12:41
  co-pilot,
• 12:41
  a
• 12:41
  liability.
• 12:42
  The
• 12:42
  benchmark
• 12:43
  also
• 12:43
  reveals
• 12:43
  that
• 12:44
  smaller
• 12:44
  models
• 12:44
  with
• 12:44
  domain
• 12:45
  fine-tuning
• 12:46
  do
• 12:46
  not
• 12:46
  close
• 12:46
  the
• 12:46
  gap,
• 12:47
  suggesting
• 12:47
  that
• 12:48
  pharmacological
• 12:48
  reasoning
• 12:49
  requires
• 12:49
  capabilities
• 12:49
  that
• 12:50
  current
• 12:50
  architectures
• 12:50
  struggle
• 12:51
  with
• 12:51
  regardless
• 12:51
  of
• 12:51
  training
• 12:52
  data.
• 12:52
  Herdr
• 12:53
  is
• 12:53
  a
• 12:53
  Rust
• 12:53
  terminal
• 12:54
  agent
• 12:54
  multiplexer
• 12:54
  that
• 12:55
  coordinates
• 12:55
  16+
• 12:56
  AI
• 12:56
  agents
• 12:56
  from
• 12:56
  a
• 12:56
  single
• 12:57
  terminal
• 12:57
  session.
• 12:58
  The
• 12:58
  architecture
• 12:59
  is
• 12:59
  straightforward
• 12:59
  --
• 13:00
  spawn
• 13:00
  agents,
• 13:01
  route
• 13:01
  tasks,
• 13:02
  collect
• 13:02
  results
• 13:02
  --
• 13:03
  but
• 13:03
  the
• 13:03
  execution
• 13:03
  matters:
• 13:04
  Herdr
• 13:05
  handles
• 13:05
  session
• 13:05
  management,
• 13:06
  output
• 13:06
  interleaving,
• 13:07
  and
• 13:07
  lifecycle
• 13:07
  coordination
• 13:08
  that
• 13:08
  every
• 13:09
  practitioner
• 13:09
  building
• 13:09
  multi-agent
• 13:10
  workflows
• 13:10
  reinvents
• 13:11
  from
• 13:11
  scratch.
• 13:12
  The
• 13:12
  Rust
• 13:12
  implementation
• 13:13
  delivers
• 13:13
  the
• 13:13
  performance
• 13:14
  characteristics
• 13:14
  (low
• 13:15
  overhead,
• 13:15
  reliable
• 13:16
  process
• 13:16
  management,
• 13:17
  no
• 13:17
  GIL
• 13:18
  contention)
• 13:18
  that
• 13:18
  Python-based
• 13:19
  orchestrators
• 13:19
  struggle
• 13:20
  with
• 13:20
  when
• 13:20
  managing
• 13:21
  more
• 13:21
  than
• 13:21
  a
• 13:21
  handful
• 13:21
  of
• 13:22
  concurrent
• 13:22
  agents.
• 13:23
  For
• 13:23
  teams
• 13:23
  already
• 13:24
  running
• 13:24
  multiple
• 13:24
  coding
• 13:25
  agents
• 13:25
  in
• 13:25
  parallel,
• 13:26
  Herdr
• 13:26
  replaces
• 13:27
  a
• 13:27
  pile
• 13:27
  of
• 13:27
  tmux
• 13:27
  scripts
• 13:28
  with
• 13:28
  a
• 13:28
  purpose-built
• 13:28
  tool.
• 13:31
  audio.cpp
• 13:33
  packs
• 13:33
  12
• 13:33
  audio
• 13:34
  models
• 13:34
  --
• 13:35
  including
• 13:35
  Qwen3-TTS,
• 13:36
  PocketTTS,
• 13:37
  and
• 13:37
  Vevo2
• 13:37
  --
• 13:38
  into
• 13:38
  a
• 13:38
  single
• 13:39
  C++/ggml
• 13:41
  runtime
• 13:42
  with
• 13:42
  CUDA
• 13:42
  support
• 13:42
  and
• 13:42
  up
• 13:43
  to
• 13:43
  5x
• 13:43
  faster
• 13:43
  inference
• 13:44
  than
• 13:44
  Python
• 13:44
  equivalents.
• 13:45
  This
• 13:45
  publication
• 13:46
  identified
• 13:46
  the
• 13:46
  gap
• 13:47
  in
• 13:47
  Edition
• 13:47
  268:
• 13:48
  the
• 13:48
  speech
• 13:49
  AI
• 13:49
  community
• 13:49
  had
• 13:49
  powerful
• 13:50
  models
• 13:50
  but
• 13:50
  no
• 13:50
  lightweight
• 13:51
  unified
• 13:51
  C++
• 13:52
  runtime.
• 13:53
  audio.cpp
• 13:55
  is
• 13:55
  the
• 13:55
  most
• 13:55
  complete
• 13:55
  answer
• 13:56
  yet,
• 13:56
  combining
• 13:57
  TTS
• 13:57
  and
• 13:58
  ASR
• 13:58
  in
• 13:58
  one
• 13:58
  binary
• 13:59
  without
• 13:59
  Python
• 13:59
  dependencies
• 14:00
  or
• 14:00
  heavyweight
• 14:01
  framework
• 14:01
  installations.
• 14:02
  For
• 14:02
  edge
• 14:03
  deployment
• 14:03
  and
• 14:03
  embedded
• 14:04
  systems
• 14:04
  where
• 14:04
  Python
• 14:04
  is
• 14:04
  not
• 14:05
  an
• 14:05
  option,
• 14:05
  this
• 14:06
  is
• 14:06
  the
• 14:06
  enabling
• 14:06
  infrastructure.
• 14:08
  ONOTE
• 14:08
  benchmarks
• 14:09
  omnimodal
• 14:09
  notation
• 14:10
  processing
• 14:10
  for
• 14:10
  expert-level
• 14:11
  music
• 14:11
  intelligence
• 14:12
  across
• 14:12
  three
• 14:12
  notation
• 14:13
  systems
• 14:13
  (common
• 14:14
  Western
• 14:14
  notation,
• 14:15
  guitar
• 14:15
  tablature,
• 14:16
  and
• 14:16
  numbered
• 14:16
  notation),
• 14:17
  four
• 14:17
  tasks,
• 14:18
  and
• 14:18
  1,120
• 14:20
  samples.
• 14:21
  This
• 14:21
  is
• 14:21
  the
• 14:21
  first
• 14:22
  dedicated
• 14:22
  evaluation
• 14:23
  framework
• 14:23
  for
• 14:23
  music
• 14:24
  notation
• 14:24
  AI
• 14:25
  that
• 14:25
  we
• 14:25
  have
• 14:25
  seen.
• 14:26
  Prior
• 14:26
  coverage
• 14:27
  profiled
• 14:27
  the
• 14:27
  tools
• 14:28
  --
• 14:28
  Audiveris
• 14:29
  for
• 14:29
  PDF-to-MusicXML,
• 14:31
  Music21
• 14:31
  for
• 14:32
  computational
• 14:32
  musicology
• 14:33
  --
• 14:33
  but
• 14:34
  without
• 14:34
  a
• 14:34
  benchmark,
• 14:35
  there
• 14:35
  was
• 14:35
  no
• 14:35
  way
• 14:35
  to
• 14:35
  compare
• 14:36
  them
• 14:36
  systematically
• 14:37
  or
• 14:37
  track
• 14:37
  progress.
• 14:39
  ONOTE
• 14:39
  provides
• 14:39
  that
• 14:39
  missing
• 14:40
  measurement
• 14:40
  layer,
• 14:41
  and
• 14:41
  early
• 14:41
  results
• 14:42
  suggest
• 14:42
  that
• 14:42
  current
• 14:43
  models
• 14:43
  handle
• 14:43
  standard
• 14:44
  Western
• 14:44
  notation
• 14:45
  reasonably
• 14:45
  well
• 14:46
  but
• 14:46
  struggle
• 14:46
  with
• 14:46
  tablature
• 14:47
  and
• 14:47
  numbered
• 14:48
  notation,
• 14:48
  where
• 14:49
  layout
• 14:49
  conventions
• 14:50
  vary
• 14:50
  widely.
• 14:51
  PP-OCRv6
• 14:54
  scales
• 14:54
  from
• 14:54
  1.5M
• 14:55
  to
• 14:56
  34.5M
• 14:57
  parameters
• 14:57
  across
• 14:58
  three
• 14:58
  tiers
• 14:58
  (tiny,
• 14:59
  small,
• 14:59
  medium),
• 15:00
  supports
• 15:00
  50
• 15:00
  languages
• 15:01
  including
• 15:01
  CJK
• 15:02
  and
• 15:02
  46
• 15:03
  Latin-script
• 15:03
  languages,
• 15:04
  and
• 15:04
  runs
• 15:04
  on
• 15:04
  PaddlePaddle,
• 15:05
  Transformers,
• 15:06
  or
• 15:06
  ONNX
• 15:07
  Runtime
• 15:07
  backends.
• 15:08
  The
• 15:08
  upgrade
• 15:09
  positions
• 15:09
  the
• 15:09
  pipeline
• 15:10
  approach
• 15:10
  --
• 15:11
  separate
• 15:11
  detection
• 15:12
  and
• 15:12
  recognition
• 15:12
  stages
• 15:13
  --
• 15:14
  against
• 15:14
  the
• 15:14
  trend
• 15:14
  toward
• 15:15
  unified
• 15:15
  VLM-based
• 15:16
  OCR
• 15:16
  systems
• 15:17
  like
• 15:17
  NVIDIA's
• 15:17
  Nemotron
• 15:18
  OCR
• 15:18
  v2
• 15:19
  and
• 15:19
  HunyuanOCR.
• 15:21
  PP-OCRv6's
• 15:24
  edge-deployment
• 15:25
  advantage
• 15:25
  remains
• 15:26
  real:
• 15:27
  a
• 15:27
  1.5M-parameter
• 15:28
  model
• 15:29
  runs
• 15:29
  on
• 15:29
  hardware
• 15:29
  where
• 15:30
  a
• 15:30
  1B-parameter
• 15:30
  VLM
• 15:31
  simply
• 15:31
  cannot.
• 15:33
  For
• 15:33
  document
• 15:33
  processing
• 15:34
  pipelines
• 15:34
  that
• 15:34
  need
• 15:35
  to
• 15:35
  run
• 15:35
  on
• 15:35
  phones,
• 15:36
  embedded
• 15:36
  devices,
• 15:37
  or
• 15:37
  constrained
• 15:37
  server
• 15:38
  environments,
• 15:38
  the
• 15:39
  pipeline
• 15:39
  school
• 15:39
  is
• 15:39
  not
• 15:40
  outdated
• 15:40
  --
• 15:41
  it
• 15:41
  is
• 15:41
  the
• 15:41
  only
• 15:41
  viable
• 15:42
  architecture.
• 15:43
  MolmoMotion
• 15:44
  predicts
• 15:44
  where
• 15:45
  objects
• 15:45
  will
• 15:45
  move
• 15:45
  in
• 15:46
  3D
• 15:46
  space
• 15:46
  over
• 15:46
  the
• 15:47
  next
• 15:47
  few
• 15:47
  seconds,
• 15:48
  given
• 15:48
  a
• 15:48
  video
• 15:48
  frame,
• 15:49
  marked
• 15:49
  query
• 15:49
  points
• 15:50
  on
• 15:50
  an
• 15:50
  object,
• 15:51
  and
• 15:51
  a
• 15:51
  text
• 15:51
  instruction
• 15:52
  describing
• 15:52
  the
• 15:52
  intended
• 15:53
  action.
• 15:54
  Built
• 15:54
  on
• 15:54
  Molmo
• 15:55
  2
• 15:55
  as
• 15:55
  backbone
• 15:56
  and
• 15:56
  trained
• 15:56
  on
• 15:57
  MolmoMotion-1M
• 15:57
  --
• 15:58
  1.16
• 15:59
  million
• 15:59
  videos
• 15:59
  with
• 16:00
  3D
• 16:00
  point
• 16:00
  trajectories
• 16:01
  paired
• 16:01
  with
• 16:01
  action
• 16:02
  descriptions
• 16:02
  spanning
• 16:02
  736
• 16:04
  motion
• 16:04
  types
• 16:04
  and
• 16:04
  5,600
• 16:06
  distinct
• 16:06
  objects
• 16:06
  --
• 16:07
  it
• 16:07
  represents
• 16:08
  motion
• 16:08
  as
• 16:08
  object-attached
• 16:09
  3D
• 16:09
  points
• 16:10
  in
• 16:10
  world
• 16:10
  space:
• 16:11
  class-agnostic,
• 16:12
  view-stable,
• 16:13
  and
• 16:13
  directly
• 16:13
  consumable
• 16:14
  by
• 16:14
  downstream
• 16:14
  planners.
• 16:15
  In
• 16:15
  simulation,
• 16:16
  a
• 16:16
  robot
• 16:17
  policy
• 16:17
  built
• 16:18
  on
• 16:18
  MolmoMotion
• 16:18
  succeeds
• 16:19
  on
• 16:19
  76.3%
• 16:20
  of
• 16:21
  pick-and-place
• 16:21
  tasks
• 16:22
  versus
• 16:22
  56.0%
• 16:24
  for
• 16:24
  the
• 16:24
  same
• 16:24
  policy
• 16:25
  without
• 16:25
  motion
• 16:25
  forecasting,
• 16:26
  and
• 16:26
  reaches
• 16:27
  target
• 16:27
  accuracy
• 16:28
  in
• 16:28
  one-sixth
• 16:28
  the
• 16:28
  training
• 16:29
  steps.
• 16:30
  The
• 16:30
  insight
• 16:30
  is
• 16:30
  that
• 16:31
  forecasting
• 16:31
  fills
• 16:32
  the
• 16:32
  gap
• 16:32
  between
• 16:32
  perception
• 16:33
  and
• 16:33
  planning
• 16:33
  --
• 16:34
  the
• 16:34
  same
• 16:35
  cup
• 16:35
  follows
• 16:35
  a
• 16:35
  similar
• 16:36
  3D
• 16:36
  trajectory
• 16:37
  whether
• 16:37
  a
• 16:37
  human
• 16:38
  hand
• 16:38
  or
• 16:38
  a
• 16:38
  robot
• 16:39
  gripper
• 16:39
  lifts
• 16:39
  it,
• 16:40
  and
• 16:40
  a
• 16:40
  model
• 16:40
  that
• 16:40
  learns
• 16:41
  this
• 16:41
  from
• 16:41
  video
• 16:42
  can
• 16:42
  transfer
• 16:42
  that
• 16:42
  knowledge
• 16:43
  to
• 16:43
  robotics
• 16:43
  without
• 16:44
  domain-specific
• 16:44
  retraining.