Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – June 04, 2026

Transcript

AGI Dreams Podcast – June 04, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  A
• 0:00
  security
• 0:00
  researcher
• 0:00
  wanted
• 0:01
  to
• 0:01
  write
• 0:01
  a
• 0:01
  Linux
• 0:01
  tool
• 0:01
  for
• 0:02
  his
• 0:02
  Creative
• 0:02
  Sound
• 0:03
  Blaster
• 0:03
  Katana
• 0:03
  V2X
• 0:04
  soundbar.
• 0:05
  What
• 0:05
  he
• 0:05
  found
• 0:05
  instead
• 0:06
  was
• 0:06
  a
• 0:06
  chain
• 0:06
  of
• 0:06
  vulnerabilities
• 0:07
  that
• 0:07
  lets
• 0:07
  any
• 0:07
  attacker
• 0:08
  within
• 0:08
  Bluetooth
• 0:08
  range
• 0:09
  —
• 0:09
  roughly
• 0:09
  15
• 0:10
  meters
• 0:10
  —
• 0:10
  silently
• 0:11
  upload
• 0:11
  custom
• 0:11
  firmware
• 0:12
  to
• 0:12
  the
• 0:12
  speaker,
• 0:12
  turn
• 0:13
  it
• 0:13
  into
• 0:13
  a
• 0:13
  covert
• 0:13
  listening
• 0:14
  device,
• 0:14
  and
• 0:15
  inject
• 0:15
  keystrokes
• 0:15
  into
• 0:15
  the
• 0:16
  connected
• 0:16
  PC
• 0:16
  as
• 0:17
  if
• 0:17
  someone
• 0:17
  plugged
• 0:17
  in
• 0:17
  a
• 0:17
  USB
• 0:18
  Rubber
• 0:18
  Ducky.
• 0:19
  No
• 0:19
  pairing
• 0:19
  required.
• 0:20
  No
• 0:20
  authentication.
• 0:21
  No
• 0:21
  physical
• 0:21
  contact
• 0:22
  with
• 0:22
  the
• 0:22
  device
• 0:22
  whatsoever.
• 0:23
  The
• 0:23
  attack
• 0:24
  is
• 0:24
  possible
• 0:24
  because
• 0:24
  Creative
• 0:25
  bridged
• 0:25
  its
• 0:25
  internal
• 0:25
  CTP
• 0:26
  command
• 0:26
  protocol
• 0:27
  to
• 0:27
  Bluetooth
• 0:27
  Low
• 0:27
  Energy
• 0:28
  without
• 0:28
  requiring
• 0:28
  the
• 0:28
  same
• 0:29
  challenge-response
• 0:29
  authentication
• 0:30
  enforced
• 0:30
  over
• 0:30
  USB.
• 0:31
  Anyone
• 0:32
  can
• 0:32
  connect
• 0:32
  and
• 0:32
  start
• 0:33
  issuing
• 0:33
  commands,
• 0:33
  including
• 0:34
  firmware
• 0:34
  updates.
• 0:35
  The
• 0:35
  firmware
• 0:35
  container
• 0:36
  uses
• 0:36
  a
• 0:36
  SHA-256
• 0:38
  checksum
• 0:39
  but
• 0:39
  no
• 0:39
  cryptographic
• 0:40
  signature
• 0:40
  verification,
• 0:41
  so
• 0:41
  patched
• 0:42
  images
• 0:42
  are
• 0:42
  accepted
• 0:43
  without
• 0:43
  complaint.
• 0:44
  The
• 0:44
  researcher's
• 0:44
  proof-of-concept
• 0:45
  replaces
• 0:46
  a
• 0:46
  dormant
• 0:46
  diagnostic
• 0:47
  task
• 0:47
  in
• 0:47
  the
• 0:47
  FreeRTOS-based
• 0:48
  firmware
• 0:49
  with
• 0:49
  102
• 0:50
  bytes
• 0:50
  of
• 0:50
  hand-written
• 0:50
  ARM/Thumb
• 0:51
  assembly
• 0:52
  that
• 0:52
  types
• 0:52
  arbitrary
• 0:53
  commands
• 0:53
  into
• 0:53
  the
• 0:53
  host
• 0:54
  PC's
• 0:54
  terminal
• 0:54
  after
• 0:55
  boot
• 0:55
  —
• 0:55
  all
• 0:56
  delivered
• 0:56
  wirelessly
• 0:56
  over
• 0:57
  BLE
• 0:57
  in
• 0:57
  about
• 0:57
  ten
• 0:58
  minutes.
• 0:58
  Bluetooth
• 0:59
  is
• 0:59
  always
• 0:59
  on,
• 1:00
  even
• 1:00
  in
• 1:00
  sleep
• 1:00
  mode,
• 1:00
  with
• 1:01
  no
• 1:01
  apparent
• 1:01
  way
• 1:01
  to
• 1:01
  disable
• 1:02
  it.
• 1:03
  Creative,
• 1:03
  contacted
• 1:04
  via
• 1:04
  SingCERT
• 1:04
  after
• 1:05
  two
• 1:05
  ignored
• 1:05
  support
• 1:05
  tickets,
• 1:06
  responded
• 1:07
  that
• 1:07
  they
• 1:07
  "do
• 1:07
  not
• 1:07
  consider
• 1:08
  this
• 1:08
  to
• 1:08
  be
• 1:08
  a
• 1:08
  vulnerability,
• 1:09
  as
• 1:09
  it
• 1:09
  does
• 1:09
  not
• 1:09
  present
• 1:10
  a
• 1:10
  cybersecurity
• 1:10
  risk."
• 1:11
  The
• 1:11
  researcher
• 1:12
  released
• 1:12
  a
• 1:12
  community
• 1:12
  patch
• 1:13
  that
• 1:13
  blocks
• 1:13
  CTP-over-Bluetooth
• 1:14
  entirely
• 1:15
  .
• 1:17
  This
• 1:17
  is
• 1:17
  not
• 1:17
  an
• 1:17
  isolated
• 1:18
  class
• 1:18
  of
• 1:18
  neglect.
• 1:20
  The
• 1:20
  Berkeley
• 1:20
  Vulnerability
• 1:21
  Initiative,
• 1:21
  an
• 1:22
  open
• 1:22
  record
• 1:22
  of
• 1:22
  CVEs
• 1:23
  discovered
• 1:23
  by
• 1:23
  agentic
• 1:24
  systems,
• 1:25
  paints
• 1:25
  a
• 1:25
  structural
• 1:25
  picture
• 1:26
  across
• 1:26
  the
• 1:26
  past
• 1:26
  90
• 1:27
  days.
• 1:28
  Total
• 1:28
  CVEs
• 1:29
  grew
• 1:29
  59.2%
• 1:30
  year-over-year,
• 1:31
  yet
• 1:32
  traditional
• 1:32
  web
• 1:32
  vulnerability
• 1:33
  classes
• 1:33
  —
• 1:34
  XSS,
• 1:35
  SQLi,
• 1:36
  CSRF
• 1:36
  —
• 1:37
  dropped
• 1:37
  by
• 1:38
  320
• 1:38
  filings.
• 1:40
  Agentic
• 1:40
  CVEs
• 1:41
  skew
• 1:41
  high
• 1:41
  or
• 1:42
  critical
• 1:42
  more
• 1:42
  often
• 1:43
  (58.2%
• 1:44
  versus
• 1:45
  51.3%
• 1:46
  for
• 1:46
  the
• 1:46
  rest
• 1:46
  of
• 1:46
  2026),
• 1:48
  but
• 1:48
  severity
• 1:48
  has
• 1:49
  become
• 1:49
  a
• 1:49
  weaker
• 1:49
  predictor
• 1:50
  of
• 1:50
  actual
• 1:50
  exploit
• 1:50
  risk:
• 1:51
  the
• 1:51
  Spearman
• 1:52
  correlation
• 1:52
  between
• 1:53
  CVSS
• 1:53
  and
• 1:53
  EPSS
• 1:54
  scores
• 1:55
  essentially
• 1:55
  collapses
• 1:56
  for
• 1:56
  agentic
• 1:56
  findings.
• 1:57
  Memory-safety
• 1:58
  CVEs
• 1:58
  found
• 1:59
  by
• 1:59
  agents
• 1:59
  cross
• 2:00
  the
• 2:00
  threshold
• 2:00
  for
• 2:00
  likely
• 2:01
  in-the-wild
• 2:01
  exploitation
• 2:02
  4×
• 2:02
  less
• 2:02
  often
• 2:03
  than
• 2:03
  in
• 2:03
  2025.
• 2:04
  The
• 2:05
  takeaway
• 2:05
  is
• 2:05
  that
• 2:05
  AI-driven
• 2:06
  discovery
• 2:06
  is
• 2:07
  changing
• 2:07
  the
• 2:07
  shape
• 2:07
  of
• 2:08
  the
• 2:08
  vulnerability
• 2:08
  landscape,
• 2:09
  not
• 2:09
  just
• 2:09
  its
• 2:10
  volume
• 2:10
  .
• 2:11
  Meanwhile,
• 2:11
  a
• 2:11
  practitioner
• 2:12
  spent
• 2:12
  $1,500
• 2:13
  pitting
• 2:14
  LLMs
• 2:14
  against
• 2:15
  a
• 2:15
  deliberately
• 2:15
  vulnerable
• 2:16
  React
• 2:16
  Native
• 2:16
  app
• 2:16
  with
• 2:17
  a
• 2:17
  wide-open
• 2:17
  Firebase
• 2:18
  backend
• 2:18
  —
• 2:19
  the
• 2:19
  exact
• 2:19
  class
• 2:20
  of
• 2:20
  broken
• 2:20
  access
• 2:20
  control
• 2:21
  that
• 2:21
  plagues
• 2:21
  real
• 2:21
  production
• 2:22
  apps.
• 2:23
  GPT
• 2:23
  o3-mini
• 2:24
  solved
• 2:24
  it
• 2:24
  7
• 2:24
  out
• 2:25
  of
• 2:25
  10
• 2:25
  times;
• 2:26
  DeepSeek
• 2:26
  V4
• 2:27
  Pro
• 2:27
  managed
• 2:27
  3
• 2:27
  out
• 2:28
  of
• 2:28
  10;
• 2:29
  both
• 2:29
  Claude
• 2:29
  models
• 2:29
  hit
• 2:30
  2
• 2:30
  out
• 2:30
  of
• 2:30
  10,
• 2:31
  with
• 2:31
  Opus
• 2:31
  hampered
• 2:32
  by
• 2:32
  late-firing
• 2:32
  safety
• 2:33
  guardrails.
• 2:34
  Gemini
• 2:34
  3.1
• 2:35
  Pro
• 2:35
  refused
• 2:36
  immediately
• 2:36
  every
• 2:36
  time.
• 2:37
  The
• 2:38
  Chinese
• 2:38
  models
• 2:38
  were
• 2:38
  notably
• 2:39
  more
• 2:39
  comfortable
• 2:39
  directly
• 2:40
  attacking
• 2:40
  the
• 2:40
  database,
• 2:41
  while
• 2:41
  Western
• 2:42
  models
• 2:42
  occasionally
• 2:42
  hesitated
• 2:43
  about
• 2:43
  affecting
• 2:44
  "live"
• 2:44
  data.
• 2:45
  At
• 2:45
  $10
• 2:46
  max
• 2:46
  and
• 2:46
  two
• 2:46
  hours
• 2:47
  per
• 2:47
  run,
• 2:48
  the
• 2:48
  cost-per-success
• 2:48
  ranged
• 2:49
  from
• 2:49
  $4.29
• 2:50
  (GPT
• 2:51
  o3-mini)
• 2:52
  to
• 2:53
  never
• 2:53
  (several
• 2:53
  models).
• 2:55
  The
• 2:55
  experiment
• 2:55
  is
• 2:55
  small
• 2:56
  and
• 2:56
  unscientific,
• 2:57
  as
• 2:57
  the
• 2:57
  author
• 2:58
  readily
• 2:58
  admits,
• 2:59
  but
• 2:59
  it
• 2:59
  adds
• 2:59
  a
• 2:59
  useful
• 3:00
  data
• 3:00
  point
• 3:00
  to
• 3:00
  the
• 3:00
  growing
• 3:01
  body
• 3:01
  of
• 3:01
  evidence
• 3:01
  that
• 3:02
  AI
• 3:02
  offensive
• 3:02
  capability
• 3:03
  is
• 3:03
  real,
• 3:04
  uneven,
• 3:04
  and
• 3:04
  heavily
• 3:05
  model-dependent
• 3:05
  .
• 3:09
  Harvard's
• 3:09
  MIMS
• 3:10
  lab
• 3:10
  has
• 3:10
  released
• 3:11
  AutoScientists,
• 3:12
  a
• 3:12
  decentralized
• 3:13
  team
• 3:13
  of
• 3:13
  AI
• 3:13
  agents
• 3:14
  built
• 3:14
  on
• 3:14
  Claude
• 3:14
  Code
• 3:15
  subagents
• 3:15
  that
• 3:15
  self-organize
• 3:16
  around
• 3:16
  promising
• 3:17
  hypotheses
• 3:17
  for
• 3:17
  long-running
• 3:18
  computational
• 3:19
  experiments.
• 3:20
  Unlike
• 3:20
  prior
• 3:21
  multi-agent
• 3:21
  systems
• 3:21
  that
• 3:22
  follow
• 3:22
  a
• 3:22
  single
• 3:22
  research
• 3:23
  trajectory
• 3:23
  or
• 3:23
  coordinate
• 3:24
  through
• 3:24
  a
• 3:24
  central
• 3:24
  planner,
• 3:25
  AutoScientists
• 3:26
  agents
• 3:26
  form
• 3:26
  teams
• 3:27
  around
• 3:27
  hypotheses,
• 3:27
  critique
• 3:28
  each
• 3:28
  other's
• 3:28
  proposals
• 3:29
  before
• 3:29
  spending
• 3:29
  compute,
• 3:30
  and
• 3:30
  share
• 3:30
  successes
• 3:31
  and
• 3:31
  failures
• 3:31
  to
• 3:31
  avoid
• 3:32
  redundant
• 3:32
  exploration.
• 3:33
  The
• 3:33
  system
• 3:34
  coordinates
• 3:34
  through
• 3:34
  a
• 3:34
  local
• 3:35
  ClawInstitute
• 3:35
  server
• 3:36
  handling
• 3:36
  workshops,
• 3:37
  workspaces,
• 3:37
  and
• 3:37
  message-board
• 3:38
  posts,
• 3:38
  while
• 3:39
  the
• 3:39
  orchestrator
• 3:39
  never
• 3:40
  trains
• 3:40
  anything
• 3:40
  itself
• 3:40
  —
• 3:41
  it
• 3:41
  launches
• 3:42
  agents
• 3:42
  and
• 3:42
  harvests
• 3:42
  results.
• 3:43
  On
• 3:43
  BioML-Bench
• 3:44
  (24
• 3:46
  biomedical
• 3:46
  ML
• 3:46
  tasks
• 3:47
  spanning
• 3:47
  drug
• 3:47
  discovery,
• 3:48
  protein
• 3:49
  engineering,
• 3:49
  and
• 3:49
  single-cell
• 3:50
  omics),
• 3:51
  AutoScientists
• 3:51
  achieved
• 3:52
  74.4%
• 3:53
  mean
• 3:53
  leaderboard
• 3:53
  percentile,
• 3:54
  outperforming
• 3:55
  the
• 3:55
  strongest
• 3:55
  prior
• 3:56
  AI
• 3:56
  agent
• 3:56
  by
• 3:56
  8.33
• 3:57
  percentage
• 3:58
  points.
• 3:59
  On
• 3:59
  nanoGPT
• 4:00
  training
• 4:00
  optimization
• 4:01
  it
• 4:01
  reached
• 4:01
  the
• 4:01
  target
• 4:01
  validation
• 4:02
  metric
• 4:02
  1.9×
• 4:03
  faster,
• 4:03
  and
• 4:04
  on
• 4:04
  ProteinGym's
• 4:04
  ACE2-Spike
• 4:05
  binding
• 4:05
  assay
• 4:06
  it
• 4:06
  gained
• 4:06
  12.5%
• 4:07
  over
• 4:07
  the
• 4:07
  baseline
• 4:07
  .
• 4:08
  The
• 4:08
  agentic
• 4:09
  tooling
• 4:09
  ecosystem
• 4:10
  around
• 4:10
  these
• 4:10
  workflows
• 4:11
  is
• 4:11
  consolidating
• 4:11
  fast.
• 4:13
  RuvOS
• 4:13
  packages
• 4:14
  the
• 4:14
  entire
• 4:14
  orchestration
• 4:15
  layer
• 4:15
  —
• 4:16
  semantic
• 4:16
  memory
• 4:16
  with
• 4:16
  HNSW/RaBitQ
• 4:21
  search,
• 4:22
  resumable
• 4:22
  signed
• 4:22
  sessions,
• 4:23
  GOAP-planned
• 4:25
  multi-agent
• 4:25
  pipelines,
• 4:26
  cross-terminal
• 4:27
  coordination,
• 4:28
  and
• 4:28
  a
• 4:28
  signed
• 4:28
  audit
• 4:28
  log
• 4:28
  —
• 4:30
  into
• 4:30
  a
• 4:30
  single
• 4:30
  static
• 4:31
  Rust
• 4:31
  binary
• 4:31
  with
• 4:31
  24
• 4:32
  MCP
• 4:32
  tools.
• 4:33
  No
• 4:34
  Node.js,
• 4:35
  no
• 4:35
  SQLite,
• 4:36
  no
• 4:36
  background
• 4:36
  service.
• 4:37
  It
• 4:37
  consolidates
• 4:38
  ideas
• 4:38
  from
• 4:39
  several
• 4:39
  earlier
• 4:39
  open-source
• 4:40
  projects
• 4:41
  (Ruflo,
• 4:41
  RuVector,
• 4:42
  SONA)
• 4:43
  into
• 4:43
  one
• 4:43
  coherent
• 4:44
  package.
• 4:45
  The
• 4:45
  design
• 4:46
  philosophy
• 4:46
  is
• 4:46
  that
• 4:46
  your
• 4:47
  coding
• 4:47
  assistant
• 4:47
  should
• 4:47
  remember
• 4:48
  decisions
• 4:48
  across
• 4:49
  days,
• 4:49
  resume
• 4:50
  where
• 4:50
  it
• 4:50
  left
• 4:50
  off,
• 4:51
  and
• 4:51
  spin
• 4:51
  up
• 4:51
  specialist
• 4:52
  agents,
• 4:52
  all
• 4:53
  stored
• 4:53
  on
• 4:53
  your
• 4:53
  own
• 4:53
  disk
• 4:54
  with
• 4:54
  tamper-evident
• 4:54
  provenance
• 4:55
  .
• 4:56
  A
• 4:56
  different
• 4:56
  approach
• 4:56
  to
• 4:56
  the
• 4:57
  orchestration
• 4:57
  problem
• 4:58
  comes
• 4:58
  from
• 4:58
  a
• 4:58
  mixed-provider
• 4:59
  workflow
• 5:00
  that
• 5:00
  routes
• 5:00
  each
• 5:01
  phase
• 5:01
  of
• 5:01
  a
• 5:01
  full-stack
• 5:02
  web
• 5:02
  build
• 5:02
  to
• 5:02
  the
• 5:02
  model
• 5:03
  that
• 5:03
  earns
• 5:04
  its
• 5:04
  tokens
• 5:04
  at
• 5:04
  that
• 5:04
  step:
• 5:05
  Sonnet
• 5:06
  explores
• 5:06
  the
• 5:06
  repo,
• 5:07
  Opus
• 5:07
  plans
• 5:08
  the
• 5:08
  architecture
• 5:08
  and
• 5:08
  page
• 5:09
  copy,
• 5:09
  Gemini
• 5:10
  3.5
• 5:11
  Flash
• 5:11
  designs
• 5:11
  the
• 5:12
  UI,
• 5:12
  Opus
• 5:13
  wires
• 5:13
  integrations,
• 5:14
  and
• 5:14
  Sonnet
• 5:14
  validates
• 5:15
  and
• 5:15
  smoke-tests.
• 5:17
  Each
• 5:17
  step
• 5:17
  runs
• 5:17
  as
• 5:18
  a
• 5:18
  separate
• 5:18
  coding
• 5:18
  agent
• 5:19
  session,
• 5:19
  communicating
• 5:20
  via
• 5:20
  a
• 5:21
  single
• 5:21
  markdown
• 5:21
  handoff
• 5:22
  document
• 5:22
  on
• 5:22
  disk.
• 5:23
  The
• 5:23
  rationale
• 5:24
  is
• 5:24
  twofold
• 5:24
  —
• 5:25
  you
• 5:26
  cannot
• 5:26
  switch
• 5:26
  providers
• 5:27
  mid-conversation,
• 5:28
  and
• 5:28
  LLMs
• 5:29
  perform
• 5:29
  better
• 5:30
  when
• 5:30
  each
• 5:30
  session
• 5:30
  focuses
• 5:31
  on
• 5:31
  one
• 5:31
  task.
• 5:32
  The
• 5:32
  workflow
• 5:33
  handles
• 5:33
  everything
• 5:34
  from
• 5:34
  exploration
• 5:35
  through
• 5:35
  deployment
• 5:35
  and
• 5:36
  browser-based
• 5:36
  smoke
• 5:37
  testing
• 5:37
  .
• 5:38
  A
• 5:38
  detailed
• 5:38
  video
• 5:39
  walkthrough
• 5:39
  demonstrates
• 5:40
  the
• 5:40
  approach
• 5:40
  end-to-end,
• 5:41
  showing
• 5:42
  that
• 5:42
  Gemini
• 5:42
  3.5
• 5:43
  Flash
• 5:43
  consistently
• 5:44
  produces
• 5:44
  better-looking
• 5:45
  UIs
• 5:46
  than
• 5:46
  Claude
• 5:46
  while
• 5:47
  Opus
• 5:47
  avoids
• 5:47
  the
• 5:47
  hallucinated
• 5:48
  page
• 5:48
  copy
• 5:49
  that
• 5:49
  plagues
• 5:49
  Gemini
• 5:50
  on
• 5:50
  factual
• 5:50
  content
• 5:51
  .
• 5:52
  The
• 5:52
  broader
• 5:52
  vision
• 5:52
  —
• 5:53
  that
• 5:53
  daily
• 5:54
  workflows
• 5:54
  should
• 5:54
  be
• 5:54
  codified
• 5:55
  into
• 5:55
  skills
• 5:56
  and
• 5:56
  then
• 5:56
  skills
• 5:56
  into
• 5:56
  automations
• 5:57
  —
• 5:58
  is
• 5:58
  gaining
• 5:58
  traction
• 5:59
  among
• 5:59
  power
• 5:59
  users
• 6:00
  who
• 6:00
  see
• 6:00
  Claude
• 6:00
  Code
• 6:00
  not
• 6:01
  as
• 6:01
  an
• 6:01
  assistant
• 6:02
  but
• 6:02
  as
• 6:02
  the
• 6:02
  kernel
• 6:02
  of
• 6:03
  an
• 6:03
  agentic
• 6:03
  operating
• 6:04
  system
• 6:04
  .
• 6:07
  Uber
• 6:07
  blew
• 6:08
  its
• 6:08
  2026
• 6:09
  AI
• 6:09
  budget
• 6:09
  in
• 6:10
  four
• 6:10
  months
• 6:10
  and
• 6:10
  responded
• 6:11
  by
• 6:11
  capping
• 6:11
  every
• 6:12
  employee
• 6:12
  at
• 6:12
  $1,500
• 6:14
  per
• 6:14
  month
• 6:14
  per
• 6:14
  AI
• 6:15
  coding
• 6:15
  tool.
• 6:16
  As
• 6:16
  Simon
• 6:16
  Willison
• 6:17
  points
• 6:17
  out,
• 6:17
  this
• 6:17
  is
• 6:17
  actually
• 6:18
  a
• 6:18
  rational
• 6:18
  signal
• 6:18
  about
• 6:19
  what
• 6:19
  these
• 6:19
  tools
• 6:19
  are
• 6:19
  worth.
• 6:20
  Assuming
• 6:21
  two
• 6:21
  actively
• 6:21
  used
• 6:22
  tools
• 6:22
  per
• 6:22
  engineer,
• 6:23
  the
• 6:23
  cap
• 6:23
  works
• 6:23
  out
• 6:23
  to
• 6:24
  $36,000
• 6:24
  per
• 6:25
  year
• 6:25
  —
• 6:26
  roughly
• 6:26
  11%
• 6:26
  of
• 6:26
  the
• 6:26
  median
• 6:27
  $330,000
• 6:28
  total
• 6:28
  compensation
• 6:29
  package
• 6:29
  for
• 6:30
  Uber
• 6:30
  software
• 6:30
  engineers.
• 6:31
  Willison
• 6:32
  notes
• 6:32
  his
• 6:32
  own
• 6:32
  token
• 6:33
  usage
• 6:33
  runs
• 6:33
  about
• 6:34
  $1,000
• 6:35
  per
• 6:35
  month
• 6:35
  per
• 6:35
  provider,
• 6:36
  which
• 6:36
  he
• 6:36
  currently
• 6:37
  pays
• 6:37
  just
• 6:37
  $100
• 6:38
  for
• 6:38
  thanks
• 6:38
  to
• 6:38
  subsidized
• 6:39
  individual
• 6:39
  plans
• 6:40
  that
• 6:40
  companies
• 6:40
  like
• 6:40
  Uber
• 6:41
  cannot
• 6:41
  access.
• 6:42
  The
• 6:42
  implication:
• 6:44
  even
• 6:44
  at
• 6:44
  list
• 6:44
  prices,
• 6:45
  a
• 6:45
  heavy
• 6:45
  user
• 6:46
  stays
• 6:46
  under
• 6:46
  Uber's
• 6:47
  cap
• 6:47
  with
• 6:47
  $500
• 6:48
  per
• 6:48
  month
• 6:48
  to
• 6:48
  spare.
• 6:49
  The
• 6:49
  spending
• 6:50
  blowout
• 6:50
  was
• 6:50
  not
• 6:50
  about
• 6:51
  individual
• 6:51
  profligacy
• 6:52
  but
• 6:52
  about
• 6:52
  thousands
• 6:53
  of
• 6:53
  engineers
• 6:54
  simultaneously
• 6:55
  discovering
• 6:55
  token-burning
• 6:56
  agentic
• 6:56
  workflows
• 6:57
  that
• 6:57
  nobody
• 6:57
  had
• 6:57
  budgeted
• 6:58
  for
• 6:58
  .
• 6:59
  The
• 6:59
  engineering
• 6:59
  response
• 7:00
  to
• 7:00
  runaway
• 7:00
  AI
• 7:01
  bills
• 7:01
  is
• 7:01
  arriving
• 7:02
  in
• 7:02
  the
• 7:02
  form
• 7:02
  of
• 7:02
  billing
• 7:02
  circuit
• 7:03
  breakers.
• 7:04
  Loopers
• 7:04
  is
• 7:04
  a
• 7:04
  Go-based
• 7:05
  reverse
• 7:05
  proxy
• 7:06
  that
• 7:06
  intercepts
• 7:06
  AI
• 7:06
  API
• 7:07
  requests
• 7:07
  with
• 7:07
  atomic
• 7:08
  Redis
• 7:08
  Lua
• 7:08
  transactions,
• 7:10
  enforcing
• 7:10
  budget
• 7:11
  limits
• 7:11
  across
• 7:11
  five
• 7:11
  granular
• 7:12
  time
• 7:12
  windows
• 7:12
  (minute,
• 7:13
  hourly,
• 7:13
  daily,
• 7:14
  weekly,
• 7:14
  monthly).
• 7:15
  It
• 7:15
  claims
• 7:16
  zero
• 7:16
  budget
• 7:16
  leakage
• 7:16
  under
• 7:17
  a
• 7:17
  1,000-request
• 7:19
  flood
• 7:19
  test
• 7:19
  versus
• 7:19
  215%
• 7:20
  leakage
• 7:21
  for
• 7:21
  LiteLLM,
• 7:22
  25×
• 7:22
  higher
• 7:22
  throughput,
• 7:23
  190×
• 7:24
  lower
• 7:24
  P99
• 7:24
  latency,
• 7:25
  and
• 7:25
  a
• 7:25
  23×
• 7:26
  smaller
• 7:26
  memory
• 7:26
  footprint.
• 7:28
  The
• 7:28
  fail-closed
• 7:28
  design
• 7:29
  blocks
• 7:29
  all
• 7:29
  requests
• 7:30
  if
• 7:30
  Redis
• 7:30
  goes
• 7:30
  down,
• 7:31
  and
• 7:31
  mid-stream
• 7:31
  SSE
• 7:32
  cutoffs
• 7:32
  sever
• 7:33
  streaming
• 7:33
  connections
• 7:33
  the
• 7:33
  moment
• 7:34
  a
• 7:34
  token
• 7:34
  budget
• 7:34
  is
• 7:34
  exceeded.
• 7:36
  The
• 7:36
  architecture
• 7:36
  is
• 7:36
  pass-through:
• 7:37
  API
• 7:38
  keys
• 7:38
  are
• 7:38
  kept
• 7:38
  only
• 7:39
  in
• 7:39
  memory
• 7:39
  during
• 7:39
  request
• 7:40
  lifecycles,
• 7:40
  never
• 7:41
  persisted
• 7:41
  to
• 7:41
  disk.
• 7:42
  It
• 7:42
  supports
• 7:43
  ten
• 7:43
  providers
• 7:43
  and
• 7:43
  ships
• 7:44
  with
• 7:44
  Python
• 7:44
  and
• 7:44
  TypeScript
• 7:45
  SDKs
• 7:45
  .
• 7:46
  Underneath
• 7:47
  the
• 7:47
  token
• 7:47
  economics
• 7:48
  sits
• 7:48
  the
• 7:48
  hardware
• 7:49
  cost
• 7:49
  question.
• 7:50
  NVIDIA's
• 7:51
  new
• 7:51
  Grace,
• 7:51
  Vera,
• 7:52
  and
• 7:52
  RTX
• 7:52
  Spark
• 7:53
  processors
• 7:53
  are
• 7:54
  not
• 7:54
  simply
• 7:54
  faster
• 7:54
  —
• 7:55
  they
• 7:55
  are
• 7:56
  architectures
• 7:56
  optimized
• 7:57
  for
• 7:57
  persistent
• 7:57
  context,
• 7:58
  memory
• 7:59
  locality,
• 8:00
  and
• 8:00
  minimized
• 8:00
  data
• 8:00
  movement.
• 8:02
  Massive
• 8:02
  unified
• 8:03
  memory,
• 8:03
  tight
• 8:03
  CPU-GPU
• 8:04
  integration,
• 8:05
  and
• 8:05
  reduced
• 8:06
  data
• 8:06
  shuttling
• 8:06
  reflect
• 8:07
  a
• 8:07
  thesis
• 8:07
  that
• 8:07
  intelligence
• 8:08
  becomes
• 8:08
  more
• 8:08
  useful
• 8:09
  when
• 8:09
  it
• 8:09
  operates
• 8:10
  continuously
• 8:10
  on
• 8:10
  local
• 8:11
  context
• 8:11
  rather
• 8:12
  than
• 8:12
  constantly
• 8:12
  moving
• 8:13
  information
• 8:13
  between
• 8:13
  disconnected
• 8:14
  systems.
• 8:15
  The
• 8:15
  observation
• 8:16
  that
• 8:16
  most
• 8:16
  AI
• 8:17
  applications
• 8:18
  still
• 8:18
  shoehorn
• 8:18
  models
• 8:19
  designed
• 8:19
  for
• 8:19
  large
• 8:19
  CUDA
• 8:20
  clusters
• 8:20
  into
• 8:20
  architectures
• 8:21
  built
• 8:21
  for
• 8:22
  something
• 8:22
  fundamentally
• 8:23
  different
• 8:23
  suggests
• 8:24
  the
• 8:24
  hardware
• 8:24
  is
• 8:24
  evolving
• 8:25
  toward
• 8:25
  contextual
• 8:26
  intelligence
• 8:27
  while
• 8:27
  much
• 8:27
  of
• 8:27
  the
• 8:27
  software
• 8:28
  ecosystem
• 8:28
  remains
• 8:29
  stuck
• 8:29
  on
• 8:29
  batch
• 8:29
  inference
• 8:30
  and
• 8:30
  token
• 8:30
  generation
• 8:30
  .
• 8:33
  A
• 8:34
  team
• 8:34
  from
• 8:34
  Shanghai
• 8:34
  Jiao
• 8:35
  Tong
• 8:35
  University
• 8:35
  and
• 8:36
  the
• 8:36
  Syft
• 8:36
  news
• 8:36
  platform
• 8:37
  introduces
• 8:37
  DynaTree,
• 8:38
  a
• 8:38
  two-stage
• 8:39
  framework
• 8:39
  that
• 8:39
  decouples
• 8:40
  the
• 8:40
  expensive
• 8:40
  semantic
• 8:40
  expansion
• 8:41
  work
• 8:41
  in
• 8:41
  agentic
• 8:42
  RAG
• 8:42
  from
• 8:42
  the
• 8:42
  real-time
• 8:43
  retrieval
• 8:43
  that
• 8:43
  downstream
• 8:44
  systems
• 8:44
  actually
• 8:44
  need.
• 8:45
  In
• 8:45
  Stage
• 8:46
  I,
• 8:46
  coordinated
• 8:46
  agents
• 8:46
  —
• 8:47
  a
• 8:47
  planner,
• 8:48
  retriever,
• 8:49
  augmenter,
• 8:50
  and
• 8:50
  reflector
• 8:50
  —
• 8:51
  collaboratively
• 8:52
  build
• 8:52
  a
• 8:52
  reusable
• 8:53
  retrieval
• 8:53
  tree
• 8:53
  that
• 8:53
  maps
• 8:54
  the
• 8:54
  semantic
• 8:54
  space
• 8:54
  of
• 8:55
  a
• 8:55
  query
• 8:55
  topic.
• 8:56
  Each
• 8:56
  branch
• 8:56
  captures
• 8:57
  a
• 8:57
  distinct
• 8:57
  subtopic
• 8:58
  or
• 8:58
  refinement
• 8:58
  trajectory,
• 8:59
  and
• 8:59
  a
• 8:59
  Shapley-value
• 9:02
  analysis
• 9:02
  confirms
• 9:03
  that
• 9:03
  the
• 9:03
  Planning
• 9:03
  agent
• 9:04
  contributes
• 9:04
  most
• 9:05
  to
• 9:05
  recall
• 9:05
  while
• 9:05
  Reflection
• 9:06
  contributes
• 9:06
  most
• 9:06
  to
• 9:06
  ranking
• 9:07
  quality.
• 9:08
  In
• 9:08
  Stage
• 9:08
  II,
• 9:09
  a
• 9:09
  lightweight
• 9:09
  daily
• 9:10
  subtree
• 9:10
  selection
• 9:10
  mechanism
• 9:11
  adapts
• 9:11
  to
• 9:11
  evolving
• 9:12
  news
• 9:12
  distributions
• 9:12
  without
• 9:13
  re-running
• 9:13
  any
• 9:13
  agentic
• 9:13
  reasoning
• 9:14
  —
• 9:15
  just
• 9:15
  LLM-based
• 9:17
  weak
• 9:17
  labeling
• 9:17
  over
• 9:17
  a
• 9:18
  compact
• 9:18
  evaluation
• 9:18
  proxy
• 9:19
  of
• 9:19
  about
• 9:19
  200
• 9:19
  examples.
• 9:21
  Deployed
• 9:21
  in
• 9:21
  Syft's
• 9:21
  production
• 9:22
  pipeline
• 9:22
  and
• 9:22
  evaluated
• 9:23
  through
• 9:23
  online
• 9:24
  A/B
• 9:24
  testing
• 9:25
  from
• 9:25
  January
• 9:25
  28
• 9:26
  to
• 9:26
  February
• 9:26
  6,
• 9:27
  2026,
• 9:28
  DynaTree's
• 9:28
  dynamically
• 9:29
  adapted
• 9:29
  variant
• 9:30
  consistently
• 9:30
  outperformed
• 9:31
  all
• 9:31
  existing
• 9:31
  production
• 9:32
  recallers
• 9:32
  on
• 9:32
  every
• 9:33
  evaluation
• 9:33
  day,
• 9:33
  with
• 9:34
  survival
• 9:34
  rate
• 9:34
  improvements
• 9:35
  of
• 9:35
  up
• 9:35
  to
• 9:35
  8%
• 9:36
  over
• 9:36
  a
• 9:36
  fixed
• 9:36
  offline
• 9:37
  subtree.
• 9:37
  The
• 9:37
  paper
• 9:38
  is
• 9:38
  accepted
• 9:38
  at
• 9:38
  KDD
• 9:39
  2026
• 9:39
  .
• 9:40
  A
• 9:40
  separate
• 9:41
  line
• 9:41
  of
• 9:41
  research
• 9:41
  tackles
• 9:42
  a
• 9:42
  fundamental
• 9:43
  weakness
• 9:43
  of
• 9:43
  LLM-based
• 9:44
  program
• 9:45
  synthesis:
• 9:46
  most
• 9:46
  approaches
• 9:46
  use
• 9:47
  simple
• 9:47
  numeric
• 9:47
  scores
• 9:48
  to
• 9:48
  rank
• 9:48
  candidates,
• 9:49
  which
• 9:49
  tell
• 9:49
  you
• 9:49
  nothing
• 9:50
  about
• 9:50
  why
• 9:50
  a
• 9:50
  program
• 9:50
  failed.
• 9:51
  Researchers
• 9:52
  at
• 9:52
  UFRGS,
• 9:53
  Oxford,
• 9:53
  and
• 9:53
  Linköping
• 9:54
  propose
• 9:54
  property-guided
• 9:55
  synthesis
• 9:55
  with
• 9:55
  counterexample-driven
• 9:56
  repair.
• 9:57
  Given
• 9:57
  a
• 9:58
  PDDL
• 9:58
  planning
• 9:58
  domain,
• 9:59
  they
• 9:59
  ask
• 10:00
  the
• 10:00
  LLM
• 10:00
  to
• 10:00
  synthesize
• 10:01
  a
• 10:01
  heuristic
• 10:01
  function
• 10:02
  in
• 10:02
  Python,
• 10:02
  then
• 10:03
  validate
• 10:03
  a
• 10:03
  formally
• 10:03
  defined
• 10:04
  property:
• 10:05
  every
• 10:05
  state
• 10:05
  reachable
• 10:06
  by
• 10:06
  strictly
• 10:06
  improving
• 10:07
  transitions
• 10:07
  must
• 10:07
  have
• 10:08
  a
• 10:08
  strictly
• 10:08
  improving
• 10:08
  successor
• 10:09
  (making
• 10:10
  hill
• 10:10
  climbing
• 10:10
  go
• 10:10
  directly
• 10:11
  to
• 10:11
  a
• 10:11
  goal).
• 10:12
  When
• 10:12
  the
• 10:12
  property
• 10:12
  is
• 10:12
  violated,
• 10:13
  the
• 10:13
  concrete
• 10:14
  failing
• 10:14
  state
• 10:14
  and
• 10:15
  its
• 10:15
  successor
• 10:15
  values
• 10:16
  are
• 10:16
  fed
• 10:16
  back
• 10:16
  to
• 10:16
  the
• 10:16
  LLM.
• 10:17
  Across
• 10:18
  ten
• 10:18
  IPC
• 10:18
  2023
• 10:19
  domains,
• 10:20
  this
• 10:20
  approach
• 10:20
  generates
• 10:21
  7×
• 10:21
  fewer
• 10:21
  programs
• 10:22
  than
• 10:22
  the
• 10:22
  best
• 10:22
  prior
• 10:23
  method,
• 10:23
  reduces
• 10:24
  evaluation
• 10:25
  cost
• 10:25
  by
• 10:25
  roughly
• 10:25
  1000×,
• 10:26
  and
• 10:27
  solves
• 10:27
  more
• 10:27
  test
• 10:27
  tasks
• 10:28
  —
• 10:28
  all
• 10:29
  via
• 10:29
  hill
• 10:29
  climbing
• 10:29
  without
• 10:30
  any
• 10:30
  combinatorial
• 10:31
  search.
• 10:32
  The
• 10:32
  synthesized
• 10:32
  heuristics
• 10:33
  remain
• 10:33
  effectively
• 10:33
  direct
• 10:34
  on
• 10:34
  virtually
• 10:34
  all
• 10:35
  out-of-distribution
• 10:36
  test
• 10:36
  tasks,
• 10:37
  a
• 10:37
  striking
• 10:37
  generalization
• 10:38
  result
• 10:38
  .
• 10:39
  In
• 10:39
  a
• 10:39
  completely
• 10:39
  different
• 10:40
  application
• 10:40
  domain,
• 10:41
  researchers
• 10:42
  at
• 10:42
  TU
• 10:42
  Wien
• 10:43
  introduce
• 10:43
  Conversational
• 10:44
  Demand
• 10:44
  Response
• 10:45
  (CDR),
• 10:46
  where
• 10:46
  aggregators
• 10:46
  and
• 10:47
  energy
• 10:47
  prosumers
• 10:47
  coordinate
• 10:48
  through
• 10:48
  bidirectional
• 10:49
  natural
• 10:49
  language
• 10:50
  powered
• 10:50
  by
• 10:50
  agentic
• 10:50
  AI.
• 10:51
  A
• 10:52
  two-tier
• 10:52
  multi-agent
• 10:53
  system
• 10:53
  pairs
• 10:53
  an
• 10:53
  aggregator
• 10:54
  agent
• 10:54
  with
• 10:54
  a
• 10:54
  prosumer
• 10:55
  Home
• 10:55
  Energy
• 10:55
  Management
• 10:56
  System
• 10:56
  (HEMS)
• 10:57
  that
• 10:57
  calls
• 10:58
  a
• 10:58
  MILP
• 10:58
  optimizer
• 10:59
  as
• 10:59
  a
• 10:59
  tool.
• 11:00
  When
• 11:00
  a
• 11:00
  grid
• 11:00
  operator
• 11:01
  needs
• 11:01
  3
• 11:02
  kW
• 11:02
  of
• 11:02
  flexibility
• 11:03
  for
• 11:03
  an
• 11:03
  evening
• 11:03
  window,
• 11:04
  the
• 11:04
  aggregator
• 11:05
  dispatches
• 11:05
  the
• 11:05
  request;
• 11:07
  the
• 11:07
  HEMS
• 11:07
  evaluates
• 11:08
  feasibility
• 11:08
  via
• 11:08
  dual-solve
• 11:09
  optimization
• 11:10
  (baseline
• 11:10
  versus
• 11:11
  DR-committed),
• 11:12
  translates
• 11:12
  the
• 11:12
  cost-benefit
• 11:13
  tradeoff
• 11:13
  into
• 11:13
  plain
• 11:14
  language,
• 11:15
  and
• 11:15
  presents
• 11:15
  it
• 11:15
  to
• 11:15
  the
• 11:15
  homeowner
• 11:16
  before
• 11:16
  any
• 11:16
  commitment.
• 11:17
  Downstream
• 11:18
  interactions
• 11:18
  complete
• 11:19
  in
• 11:19
  under
• 11:19
  12
• 11:20
  seconds.
• 11:21
  The
• 11:21
  architecture
• 11:21
  is
• 11:21
  fully
• 11:22
  open
• 11:22
  source
• 11:22
  .
• 11:26
  Creating
• 11:26
  images
• 11:27
  from
• 11:27
  noise
• 11:27
  is
• 11:27
  generation;
• 11:28
  reconstructing
• 11:29
  fine
• 11:29
  details
• 11:29
  from
• 11:30
  coarse
• 11:30
  inputs
• 11:30
  is
• 11:31
  super-resolution.
• 11:32
  A
• 11:32
  team
• 11:32
  at
• 11:32
  MIT
• 11:33
  argues
• 11:33
  these
• 11:33
  are
• 11:34
  the
• 11:34
  same
• 11:34
  problem
• 11:34
  —
• 11:35
  reversing
• 11:35
  information
• 11:36
  loss
• 11:36
  across
• 11:36
  scales
• 11:37
  —
• 11:37
  and
• 11:37
  introduces
• 11:38
  SKILD,
• 11:38
  a
• 11:38
  scale-invariant
• 11:39
  frequency-space
• 11:40
  diffusion
• 11:40
  model
• 11:40
  that
• 11:40
  unifies
• 11:41
  both
• 11:41
  tasks
• 11:42
  within
• 11:42
  a
• 11:42
  single
• 11:42
  unconditional
• 11:43
  framework.
• 11:44
  The
• 11:44
  forward
• 11:44
  process
• 11:44
  attenuates
• 11:45
  image
• 11:45
  content
• 11:46
  from
• 11:46
  fine
• 11:46
  to
• 11:46
  coarse
• 11:46
  scales
• 11:47
  while
• 11:47
  injecting
• 11:48
  spectrum-matched
• 11:48
  Gaussian
• 11:49
  noise
• 11:49
  (noise
• 11:50
  that
• 11:50
  carries
• 11:50
  the
• 11:50
  statistical
• 11:50
  profile
• 11:51
  of
• 11:51
  the
• 11:51
  dataset
• 11:51
  itself),
• 11:52
  making
• 11:52
  scale
• 11:52
  an
• 11:53
  explicit
• 11:53
  coordinate
• 11:53
  of
• 11:53
  the
• 11:53
  diffusion
• 11:54
  dynamics.
• 11:55
  The
• 11:55
  same
• 11:55
  trained
• 11:55
  reverse
• 11:56
  process
• 11:56
  handles
• 11:56
  generation
• 11:57
  and
• 11:57
  continuous
• 11:58
  super-resolution
• 11:59
  by
• 11:59
  varying
• 11:59
  only
• 11:59
  the
• 11:59
  starting
• 12:00
  timestep:
• 12:00
  no
• 12:01
  conditioning
• 12:01
  branch,
• 12:02
  no
• 12:02
  classifier-free
• 12:02
  guidance,
• 12:03
  no
• 12:03
  retraining
• 12:04
  per
• 12:04
  scale
• 12:04
  factor.
• 12:05
  SKILD
• 12:05
  reaches
• 12:06
  FID
• 12:06
  1.73
• 12:07
  and
• 12:07
  Inception
• 12:08
  Score
• 12:08
  10.22
• 12:09
  on
• 12:09
  unconditional
• 12:10
  CIFAR-10,
• 12:13
  performs
• 12:13
  2×–4×
• 12:13
  super-resolution
• 12:14
  on
• 12:14
  ImageNet
• 12:15
  from
• 12:15
  a
• 12:15
  single
• 12:15
  checkpoint
• 12:16
  while
• 12:16
  outperforming
• 12:17
  conditional
• 12:17
  baselines
• 12:18
  on
• 12:18
  perceptual
• 12:18
  metrics,
• 12:19
  and
• 12:19
  reconstructs
• 12:20
  critical
• 12:20
  Ising
• 12:20
  model
• 12:21
  fields
• 12:21
  whose
• 12:21
  connected
• 12:22
  four-point
• 12:22
  correlations
• 12:23
  closely
• 12:23
  track
• 12:23
  the
• 12:23
  ground
• 12:23
  truth
• 12:24
  —
• 12:24
  a
• 12:25
  stronger
• 12:25
  test
• 12:25
  of
• 12:25
  scale
• 12:25
  invariance
• 12:26
  than
• 12:26
  any
• 12:26
  perceptual
• 12:27
  quality
• 12:27
  metric
• 12:27
  .
• 12:28
  At
• 12:28
  the
• 12:28
  rendering
• 12:29
  end
• 12:29
  of
• 12:29
  the
• 12:29
  visual
• 12:29
  pipeline,
• 12:30
  Gaussian
• 12:31
  Point
• 12:31
  Splatting
• 12:31
  offers
• 12:31
  a
• 12:31
  stochastic
• 12:32
  method
• 12:32
  for
• 12:32
  rendering
• 12:33
  Gaussian
• 12:33
  splats
• 12:34
  that
• 12:34
  scales
• 12:34
  to
• 12:34
  hundreds
• 12:35
  of
• 12:35
  millions
• 12:35
  of
• 12:35
  Gaussians
• 12:36
  in
• 12:36
  real
• 12:36
  time.
• 12:37
  The
• 12:37
  core
• 12:37
  idea
• 12:37
  is
• 12:37
  to
• 12:38
  sample
• 12:38
  pixel-sized
• 12:38
  opaque
• 12:39
  points
• 12:39
  from
• 12:39
  the
• 12:39
  Gaussians
• 12:40
  and
• 12:40
  splat
• 12:40
  them
• 12:41
  to
• 12:41
  a
• 12:41
  framebuffer
• 12:41
  using
• 12:42
  64-bit
• 12:42
  atomics,
• 12:43
  achieving
• 12:44
  even
• 12:44
  workload
• 12:44
  distribution
• 12:45
  across
• 12:45
  millions
• 12:45
  of
• 12:46
  threads.
• 12:47
  The
• 12:47
  authors
• 12:47
  formalize
• 12:48
  and
• 12:48
  solve
• 12:48
  the
• 12:48
  nontrivial
• 12:49
  problem
• 12:49
  of
• 12:49
  determining
• 12:50
  how
• 12:50
  many
• 12:50
  points
• 12:51
  to
• 12:51
  splat
• 12:51
  per
• 12:51
  Gaussian
• 12:52
  and
• 12:52
  how
• 12:52
  to
• 12:52
  distribute
• 12:53
  them
• 12:53
  to
• 12:53
  achieve
• 12:53
  the
• 12:54
  desired
• 12:54
  opacity,
• 12:55
  keeping
• 12:55
  renders
• 12:56
  faithful
• 12:56
  to
• 12:56
  the
• 12:56
  original
• 12:57
  Gaussian
• 12:57
  splatting.
• 12:58
  The
• 12:58
  work
• 12:58
  is
• 12:58
  presented
• 12:59
  at
• 12:59
  SIGGRAPH
• 13:00
  2026
• 13:01
  .
• 13:03
  DaVinci
• 13:04
  Resolve
• 13:04
  21
• 13:04
  brings
• 13:05
  AI
• 13:05
  tools
• 13:05
  to
• 13:05
  a
• 13:06
  new
• 13:06
  domain:
• 13:06
  still
• 13:07
  photography.
• 13:08
  The
• 13:08
  new
• 13:08
  Photo
• 13:08
  page
• 13:09
  integrates
• 13:09
  Hollywood's
• 13:10
  node-based
• 13:10
  color
• 13:11
  grading
• 13:11
  workflow
• 13:12
  with
• 13:12
  a
• 13:12
  dedicated
• 13:12
  image
• 13:13
  editing
• 13:13
  environment,
• 13:14
  while
• 13:14
  AI
• 13:14
  IntelliSearch
• 13:15
  lets
• 13:15
  users
• 13:16
  find
• 13:16
  specific
• 13:16
  objects,
• 13:17
  scenes,
• 13:18
  or
• 13:18
  individual
• 13:18
  faces
• 13:19
  across
• 13:19
  entire
• 13:19
  projects
• 13:20
  using
• 13:20
  plain
• 13:20
  language.
• 13:22
  AI
• 13:22
  text-to-speech
• 13:23
  generates
• 13:23
  unique
• 13:23
  voices
• 13:24
  from
• 13:24
  10-second
• 13:24
  clips,
• 13:25
  AI
• 13:26
  CineFocus
• 13:26
  simulates
• 13:27
  adjustable
• 13:27
  depth
• 13:27
  of
• 13:28
  field,
• 13:28
  AI
• 13:29
  Face
• 13:29
  Age
• 13:29
  Transformer
• 13:30
  changes
• 13:30
  an
• 13:30
  actor's
• 13:30
  apparent
• 13:31
  age,
• 13:31
  and
• 13:32
  AI
• 13:32
  UltraSharpen
• 13:33
  recovers
• 13:33
  detail
• 13:33
  from
• 13:34
  upscaled
• 13:34
  or
• 13:34
  slightly
• 13:35
  out-of-focus
• 13:35
  footage.
• 13:36
  The
• 13:37
  release
• 13:37
  also
• 13:37
  adds
• 13:38
  native
• 13:38
  OGraf
• 13:38
  HTML
• 13:39
  graphics
• 13:39
  and
• 13:40
  Lottie
• 13:40
  animation
• 13:40
  support,
• 13:41
  Krokodove's
• 13:42
  70+
• 13:43
  Fusion
• 13:43
  compositing
• 13:44
  tools,
• 13:44
  MultiMaster
• 13:45
  HDR/SDR
• 13:47
  trim
• 13:47
  passes,
• 13:48
  and
• 13:48
  expanded
• 13:49
  immersive/VR
• 13:50
  workflows
• 13:50
  including
• 13:51
  foveated
• 13:51
  rendering
• 13:52
  for
• 13:52
  Apple
• 13:52
  Immersive
• 13:52
  .
• 13:53
  For
• 13:53
  those
• 13:53
  building
• 13:54
  the
• 13:54
  models
• 13:54
  that
• 13:54
  power
• 13:55
  these
• 13:55
  tools,
• 13:56
  Hugging
• 13:56
  Face
• 13:56
  has
• 13:56
  published
• 13:57
  a
• 13:57
  thorough
• 13:57
  beginner's
• 13:58
  guide
• 13:58
  to
• 13:58
  torch.profiler
• 13:59
  that
• 14:00
  walks
• 14:00
  through
• 14:00
  reading
• 14:00
  profiler
• 14:01
  tables
• 14:02
  and
• 14:02
  traces
• 14:02
  from
• 14:02
  a
• 14:02
  single
• 14:03
  matrix
• 14:03
  multiplication,
• 14:04
  identifying
• 14:05
  overhead-bound
• 14:05
  versus
• 14:06
  compute-bound
• 14:06
  regimes,
• 14:07
  understanding
• 14:08
  the
• 14:08
  CPU-to-GPU
• 14:09
  dispatch
• 14:09
  chain
• 14:10
  (including
• 14:11
  why
• 14:11
  cudaOccupancyMaxActiveBlocksPerMultiprocessor
• 14:14
  fires
• 14:14
  before
• 14:14
  matmul
• 14:15
  but
• 14:15
  not
• 14:15
  add),
• 14:16
  and
• 14:16
  what
• 14:16
  torch.compile
• 14:17
  actually
• 14:17
  changes
• 14:18
  under
• 14:18
  the
• 14:18
  hood
• 14:18
  —
• 14:19
  the
• 14:19
  fusion
• 14:19
  is
• 14:19
  at
• 14:19
  the
• 14:19
  dispatcher
• 14:20
  level,
• 14:21
  not
• 14:21
  the
• 14:21
  kernel
• 14:21
  level,
• 14:22
  with
• 14:22
  the
• 14:22
  bias
• 14:22
  add
• 14:22
  folded
• 14:23
  into
• 14:23
  a
• 14:23
  GEMM
• 14:23
  epilogue
• 14:24
  prefixed
• 14:24
  by
• 14:24
  a
• 14:24
  Device-to-Device
• 14:25
  memcpy
• 14:25
  .
• 14:26
  On
• 14:26
  the
• 14:26
  model
• 14:27
  release
• 14:27
  front,
• 14:28
  Microsoft
• 14:28
  has
• 14:28
  published
• 14:29
  Harrier
• 14:29
  OSS
• 14:29
  v1,
• 14:30
  a
• 14:30
  270M-parameter
• 14:31
  model
• 14:32
  ,
• 14:33
  and
• 14:33
  Cohere
• 14:34
  Labs
• 14:34
  has
• 14:34
  released
• 14:34
  tiny-aya-global,
• 14:36
  a
• 14:36
  small
• 14:36
  multilingual
• 14:37
  model
• 14:37
  .
• 14:39
  Sixteen
• 14:40
  mathematics
• 14:40
  specialists
• 14:41
  have
• 14:41
  turned
• 14:41
  the
• 14:41
  discipline's
• 14:42
  growing
• 14:42
  unease
• 14:43
  with
• 14:43
  AI
• 14:43
  into
• 14:43
  a
• 14:43
  public
• 14:44
  declaration.
• 14:45
  The
• 14:45
  Leiden
• 14:45
  Declaration
• 14:46
  on
• 14:46
  Artificial
• 14:47
  Intelligence
• 14:47
  and
• 14:47
  Mathematics,
• 14:48
  endorsed
• 14:49
  by
• 14:49
  the
• 14:49
  International
• 14:49
  Mathematical
• 14:50
  Union,
• 14:51
  argues
• 14:51
  that
• 14:51
  mathematics
• 14:52
  is
• 14:52
  more
• 14:52
  than
• 14:52
  a
• 14:52
  machine
• 14:52
  for
• 14:53
  producing
• 14:53
  correct
• 14:53
  answers
• 14:54
  —
• 14:54
  it
• 14:54
  is
• 14:55
  a
• 14:55
  deeply
• 14:55
  human
• 14:55
  endeavor
• 14:56
  built
• 14:56
  on
• 14:56
  creativity,
• 14:57
  understanding,
• 14:58
  collaboration,
• 14:59
  and
• 14:59
  pursuit
• 14:59
  of
• 14:59
  knowledge
• 15:00
  for
• 15:00
  its
• 15:00
  own
• 15:00
  sake.
• 15:01
  Those
• 15:01
  values
• 15:02
  clash
• 15:02
  directly
• 15:02
  with
• 15:02
  the
• 15:02
  incentives
• 15:03
  driving
• 15:03
  AI
• 15:04
  development.
• 15:05
  The
• 15:05
  authors
• 15:06
  warn
• 15:06
  that
• 15:06
  AI-generated
• 15:07
  papers
• 15:07
  could
• 15:07
  overwhelm
• 15:08
  peer
• 15:08
  review
• 15:08
  with
• 15:08
  low-quality
• 15:09
  work,
• 15:10
  make
• 15:10
  it
• 15:10
  difficult
• 15:10
  to
• 15:10
  assign
• 15:11
  proper
• 15:11
  credit
• 15:11
  for
• 15:11
  discoveries,
• 15:12
  and
• 15:13
  disadvantage
• 15:13
  researchers
• 15:14
  who
• 15:14
  choose
• 15:14
  not
• 15:14
  to
• 15:14
  use
• 15:15
  AI
• 15:15
  tools.
• 15:16
  They
• 15:16
  also
• 15:17
  raise
• 15:17
  concerns
• 15:17
  about
• 15:17
  mathematical
• 15:18
  work
• 15:18
  being
• 15:18
  used
• 15:19
  to
• 15:19
  train
• 15:19
  AI
• 15:19
  systems
• 15:20
  for
• 15:20
  military
• 15:20
  and
• 15:20
  surveillance
• 15:21
  purposes.
• 15:22
  The
• 15:22
  declaration
• 15:22
  will
• 15:22
  be
• 15:23
  discussed
• 15:23
  at
• 15:23
  next
• 15:23
  month's
• 15:24
  International
• 15:24
  Congress
• 15:24
  of
• 15:25
  Mathematicians
• 15:25
  in
• 15:25
  Philadelphia.
• 15:26
  The
• 15:26
  timing
• 15:27
  is
• 15:27
  pointed:
• 15:28
  OpenAI
• 15:28
  recently
• 15:28
  reported
• 15:29
  that
• 15:29
  one
• 15:29
  of
• 15:29
  its
• 15:29
  newest
• 15:30
  reasoning
• 15:30
  models
• 15:30
  independently
• 15:31
  solved
• 15:31
  an
• 15:31
  80-year-old
• 15:32
  conjecture,
• 15:33
  the
• 15:33
  latest
• 15:33
  in
• 15:33
  a
• 15:33
  series
• 15:33
  of
• 15:34
  breakthroughs
• 15:34
  pushing
• 15:34
  AI
• 15:35
  to
• 15:35
  the
• 15:35
  frontiers
• 15:35
  of
• 15:35
  the
• 15:35
  subject
• 15:36
  .
• 15:37
  The
• 15:37
  deeper
• 15:37
  question
• 15:38
  underneath
• 15:38
  the
• 15:38
  mathematicians'
• 15:39
  alarm
• 15:39
  is
• 15:39
  one
• 15:40
  that
• 15:40
  a
• 15:40
  commentary
• 15:40
  on
• 15:41
  the
• 15:41
  "solipsistic
• 15:41
  superintelligence"
• 15:42
  paper
• 15:43
  articulates
• 15:43
  well:
• 15:44
  AI
• 15:44
  does
• 15:44
  not
• 15:45
  just
• 15:45
  enter
• 15:45
  an
• 15:45
  operating
• 15:46
  environment
• 15:46
  —
• 15:47
  it
• 15:47
  changes
• 15:47
  the
• 15:47
  operating
• 15:48
  environment.
• 15:49
  Humans
• 15:49
  respond.
• 15:50
  Institutions
• 15:51
  respond.
• 15:52
  Markets
• 15:53
  respond.
• 15:54
  Other
• 15:54
  AI
• 15:55
  systems
• 15:55
  respond.
• 15:56
  Performance
• 15:57
  is
• 15:57
  not
• 15:57
  just
• 15:57
  a
• 15:57
  property
• 15:58
  of
• 15:58
  the
• 15:58
  model;
• 15:59
  it
• 15:59
  becomes
• 15:59
  a
• 15:59
  property
• 16:00
  of
• 16:00
  the
• 16:00
  coupled
• 16:00
  system
• 16:00
  the
• 16:00
  model
• 16:01
  enters.
• 16:02
  The
• 16:02
  practical
• 16:02
  implication
• 16:03
  for
• 16:03
  boards
• 16:03
  and
• 16:04
  executives
• 16:04
  is
• 16:04
  to
• 16:05
  stop
• 16:05
  asking
• 16:05
  only
• 16:06
  "does
• 16:06
  the
• 16:06
  system
• 16:06
  perform?"
• 16:07
  and
• 16:08
  start
• 16:08
  asking
• 16:08
  "what
• 16:09
  behaviors,
• 16:10
  incentives,
• 16:11
  dependencies,
• 16:12
  skills,
• 16:13
  workarounds,
• 16:13
  and
• 16:14
  second-order
• 16:14
  adaptations
• 16:15
  will
• 16:15
  this
• 16:15
  system
• 16:16
  create
• 16:16
  once
• 16:16
  people
• 16:16
  start
• 16:17
  responding
• 16:17
  to
• 16:17
  it?"
• 16:18
  That
• 16:18
  is
• 16:19
  where
• 16:19
  the
• 16:19
  real
• 16:19
  risk
• 16:19
  sits
• 16:20
  —
• 16:21
  but
• 16:21
  also
• 16:21
  where
• 16:21
  the
• 16:21
  real
• 16:22
  value
• 16:22
  lies.
• 16:23
  Not
• 16:23
  in
• 16:23
  treating
• 16:24
  AI
• 16:24
  as
• 16:24
  intelligence
• 16:25
  dropped
• 16:25
  into
• 16:25
  a
• 16:26
  passive,
• 16:26
  frozen
• 16:27
  world,
• 16:27
  but
• 16:27
  in
• 16:27
  understanding
• 16:28
  the
• 16:28
  new
• 16:28
  operating
• 16:29
  environment
• 16:29
  that
• 16:29
  forms
• 16:30
  around
• 16:30
  it
• 16:30
  .