Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 24, 2026

Transcript

AGI Dreams Podcast – March 24, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  security
• 0:01
  conversation
• 0:01
  around
• 0:01
  AI
• 0:02
  pipelines
• 0:02
  has
• 0:02
  long
• 0:02
  fixated
• 0:03
  on
• 0:03
  the
• 0:03
  algorithmic
• 0:03
  layer
• 0:04
  —
• 0:04
  jailbreaks,
• 0:05
  prompt
• 0:05
  injection,
• 0:06
  training
• 0:06
  data
• 0:07
  poisoning.
• 0:08
  A
• 0:08
  new
• 0:08
  paper
• 0:08
  from
• 0:08
  UT
• 0:09
  Austin
• 0:09
  and
• 0:09
  Intel
• 0:09
  researchers
• 0:10
  argues
• 0:10
  that
• 0:10
  fixation
• 0:11
  is
• 0:11
  dangerously
• 0:11
  incomplete.
• 0:12
  "Cascade"
• 0:13
  introduces
• 0:13
  a
• 0:13
  systematic
• 0:14
  red-teaming
• 0:14
  framework
• 0:15
  that
• 0:15
  composes
• 0:15
  attack
• 0:15
  gadgets
• 0:16
  across
• 0:16
  software
• 0:16
  CVEs,
• 0:17
  hardware
• 0:18
  side-channels,
• 0:19
  and
• 0:19
  algorithmic
• 0:19
  exploits
• 0:20
  to
• 0:20
  compromise
• 0:20
  compound
• 0:21
  AI
• 0:21
  systems
• 0:21
  end-to-end.
• 0:22
  The
• 0:22
  core
• 0:23
  insight:
• 0:23
  modern
• 0:24
  AI
• 0:24
  pipelines
• 0:24
  are
• 0:25
  not
• 0:25
  monolithic
• 0:25
  models
• 0:26
  but
• 0:26
  layered
• 0:26
  stacks
• 0:27
  of
• 0:27
  LLMs,
• 0:27
  vector
• 0:28
  databases,
• 0:28
  orchestration
• 0:29
  frameworks
• 0:29
  (LangChain,
• 0:30
  Ollama),
• 0:30
  and
• 0:31
  distributed
• 0:31
  hardware
• 0:31
  —
• 0:32
  each
• 0:32
  layer
• 0:32
  harboring
• 0:33
  its
• 0:33
  own
• 0:33
  vulnerabilities
• 0:34
  that
• 0:34
  can
• 0:34
  amplify
• 0:34
  attacks
• 0:35
  on
• 0:35
  adjacent
• 0:35
  layers.
• 0:36
  The
• 0:36
  paper's
• 0:36
  proof-of-concept
• 0:37
  chains
• 0:37
  three
• 0:38
  distinct
• 0:38
  exploit
• 0:38
  classes.
• 0:39
  First,
• 0:39
  a
• 0:40
  code
• 0:40
  injection
• 0:40
  vulnerability
• 0:41
  in
• 0:41
  a
• 0:41
  query
• 0:41
  enhancer
• 0:41
  (targeting
• 0:42
  real
• 0:42
  CVEs
• 0:43
  in
• 0:43
  LangChain
• 0:44
  and
• 0:44
  LlamaIndex)
• 0:44
  triggers
• 0:45
  a
• 0:45
  denial-of-service
• 0:45
  that
• 0:45
  forces
• 0:46
  the
• 0:46
  pipeline
• 0:46
  to
• 0:46
  bypass
• 0:47
  query
• 0:47
  sanitization
• 0:48
  entirely.
• 0:48
  Second,
• 0:49
  a
• 0:49
  Rowhammer-based
• 0:50
  fault
• 0:50
  injection
• 0:50
  flips
• 0:51
  bits
• 0:51
  in
• 0:51
  the
• 0:51
  guardrail
• 0:51
  model's
• 0:52
  memory
• 0:52
  —
• 0:52
  not
• 0:52
  the
• 0:52
  generator's
• 0:53
  —
• 0:53
  corrupting
• 0:54
  either
• 0:54
  trigger
• 0:54
  tokens
• 0:55
  or
• 0:55
  attention
• 0:55
  masks.
• 0:56
  The
• 0:56
  Type
• 0:56
  3
• 0:57
  variant
• 0:57
  (random
• 0:57
  attention
• 0:58
  bitflip)
• 0:58
  achieved
• 0:58
  a
• 0:58
  94%
• 0:59
  guardrail
• 1:00
  evasion
• 1:00
  rate
• 1:00
  because
• 1:00
  boosting
• 1:01
  attention
• 1:01
  on
• 1:01
  any
• 1:01
  single
• 1:02
  safe
• 1:02
  token
• 1:02
  effectively
• 1:02
  suppresses
• 1:03
  all
• 1:03
  others,
• 1:04
  including
• 1:04
  multiple
• 1:04
  trigger
• 1:04
  words.
• 1:06
  Third,
• 1:06
  with
• 1:07
  the
• 1:07
  guardrail
• 1:07
  neutralized
• 1:08
  and
• 1:08
  the
• 1:08
  query
• 1:08
  enhancer
• 1:08
  down,
• 1:09
  a
• 1:09
  GCG-crafted
• 1:10
  adversarial
• 1:10
  suffix
• 1:11
  jailbreaks
• 1:11
  the
• 1:11
  generator
• 1:12
  at
• 1:12
  an
• 1:12
  82%
• 1:13
  success
• 1:13
  rate.
• 1:14
  The
• 1:14
  researchers
• 1:15
  curated
• 1:15
  a
• 1:15
  corpus
• 1:16
  of
• 1:16
  100
• 1:16
  algorithmic
• 1:17
  vulnerability
• 1:17
  papers,
• 1:18
  100
• 1:18
  software
• 1:19
  CVEs,
• 1:19
  and
• 1:20
  dozens
• 1:20
  of
• 1:20
  hardware
• 1:20
  attack
• 1:21
  primitives,
• 1:21
  then
• 1:21
  used
• 1:22
  LLM-based
• 1:22
  reasoning
• 1:23
  to
• 1:23
  search
• 1:23
  for
• 1:23
  viable
• 1:23
  attack
• 1:24
  chains
• 1:24
  —
• 1:25
  essentially
• 1:25
  building
• 1:25
  an
• 1:26
  automated
• 1:26
  red-team
• 1:26
  that
• 1:26
  explores
• 1:27
  cross-stack
• 1:27
  compositions
• 1:28
  human
• 1:28
  testers
• 1:28
  would
• 1:29
  rarely
• 1:29
  attempt.
• 1:30
  This
• 1:30
  matters
• 1:30
  because
• 1:31
  it
• 1:31
  demonstrates
• 1:31
  that
• 1:31
  guardrails
• 1:32
  and
• 1:32
  query
• 1:32
  sanitizers,
• 1:33
  the
• 1:34
  go-to
• 1:34
  defenses
• 1:34
  for
• 1:34
  production
• 1:35
  AI
• 1:35
  deployments,
• 1:36
  can
• 1:36
  be
• 1:36
  rendered
• 1:37
  irrelevant
• 1:37
  by
• 1:37
  exploits
• 1:38
  that
• 1:38
  never
• 1:38
  touch
• 1:38
  the
• 1:39
  model
• 1:39
  algorithm
• 1:39
  at
• 1:39
  all.
• 1:40
  The
• 1:40
  framework
• 1:41
  also
• 1:41
  systematizes
• 1:42
  existing
• 1:42
  attack
• 1:42
  patterns
• 1:43
  (SQL
• 1:43
  injection
• 1:44
  into
• 1:44
  PoisonedRAG,
• 1:45
  malicious
• 1:45
  HuggingFace
• 1:46
  packages
• 1:46
  exfiltrating
• 1:47
  queries,
• 1:48
  I/O
• 1:48
  bus
• 1:49
  snooping
• 1:49
  for
• 1:49
  membership
• 1:49
  inference)
• 1:50
  into
• 1:50
  a
• 1:50
  composable
• 1:51
  taxonomy.
• 1:52
  Meanwhile,
• 1:53
  a
• 1:53
  lighter
• 1:53
  but
• 1:53
  related
• 1:53
  technique
• 1:54
  surfaced
• 1:54
  on
• 1:54
  Reddit:
• 1:55
  injecting
• 1:56
  special
• 1:56
  tokens
• 1:57
  to
• 1:57
  gaslight
• 1:57
  LLMs
• 1:58
  during
• 1:58
  code
• 1:58
  reviews,
• 1:59
  causing
• 1:59
  them
• 1:59
  to
• 2:00
  overlook
• 2:00
  malicious
• 2:00
  code
• 2:01
  by
• 2:01
  exploiting
• 2:02
  how
• 2:02
  models
• 2:02
  process
• 2:03
  control
• 2:03
  tokens
• 2:03
  differently
• 2:04
  from
• 2:04
  natural
• 2:04
  language.
• 2:06
  On
• 2:06
  the
• 2:06
  defensive
• 2:06
  side,
• 2:07
  Cisco
• 2:07
  shipped
• 2:07
  an
• 2:08
  AI
• 2:08
  Security
• 2:08
  Scanner
• 2:09
  IDE
• 2:09
  plugin
• 2:10
  that
• 2:10
  watches
• 2:10
  the
• 2:10
  supply
• 2:11
  chain
• 2:11
  around
• 2:11
  MCP
• 2:12
  servers,
• 2:12
  agent
• 2:13
  skills,
• 2:13
  and
• 2:13
  AI-generated
• 2:14
  code.
• 2:15
  It
• 2:15
  combines
• 2:16
  YARA
• 2:16
  rules
• 2:17
  and
• 2:17
  behavioral
• 2:17
  analysis
• 2:18
  (fully
• 2:18
  local)
• 2:18
  with
• 2:19
  optional
• 2:19
  LLM-based
• 2:20
  and
• 2:20
  VirusTotal
• 2:21
  analysis,
• 2:21
  scanning
• 2:22
  MCP
• 2:23
  config
• 2:23
  files
• 2:23
  and
• 2:23
  skill
• 2:24
  definitions
• 2:24
  for
• 2:24
  hidden
• 2:24
  instructions,
• 2:25
  data
• 2:26
  exfiltration
• 2:26
  patterns,
• 2:27
  and
• 2:27
  prompt
• 2:27
  injection.
• 2:29
  A
• 2:29
  Watchdog
• 2:29
  mode
• 2:29
  monitors
• 2:30
  file
• 2:30
  changes
• 2:30
  with
• 2:31
  snapshot-and-revert
• 2:31
  capability.
• 2:33
  Notably,
• 2:33
  it
• 2:33
  analyzes
• 2:34
  descriptions
• 2:34
  only
• 2:34
  —
• 2:35
  never
• 2:35
  executes
• 2:36
  MCP
• 2:36
  tools
• 2:37
  or
• 2:37
  runs
• 2:37
  skill
• 2:37
  code
• 2:38
  —
• 2:38
  and
• 2:39
  source
• 2:39
  code
• 2:39
  never
• 2:39
  leaves
• 2:40
  the
• 2:40
  machine
• 2:40
  unless
• 2:40
  you
• 2:41
  opt
• 2:41
  in
• 2:41
  to
• 2:41
  file
• 2:41
  upload.
• 2:42
  The
• 2:42
  offensive
• 2:43
  side
• 2:43
  is
• 2:43
  moving
• 2:43
  even
• 2:44
  faster.
• 2:45
  A
• 2:45
  deep
• 2:45
  analysis
• 2:46
  of
• 2:46
  AI-native
• 2:46
  offensive
• 2:47
  security
• 2:47
  startups
• 2:48
  reveals
• 2:48
  over
• 2:48
  $600
• 2:49
  million
• 2:49
  raised
• 2:50
  across
• 2:50
  five
• 2:50
  companies
• 2:51
  alone.
• 2:52
  Armadin,
• 2:53
  founded
• 2:53
  by
• 2:53
  Kevin
• 2:54
  Mandia
• 2:54
  (who
• 2:54
  built
• 2:54
  and
• 2:55
  sold
• 2:55
  Mandiant
• 2:55
  twice),
• 2:56
  emerged
• 2:57
  with
• 2:57
  a
• 2:57
  record
• 2:57
  $150
• 2:58
  million
• 2:58
  combined
• 2:59
  seed-and-Series-A
• 3:00
  to
• 3:00
  build
• 3:00
  "agentic
• 3:01
  attacker
• 3:01
  swarms."
• 3:03
  Horizon3's
• 3:03
  NodeZero
• 3:04
  has
• 3:04
  powered
• 3:04
  over
• 3:05
  150,000
• 3:06
  autonomous
• 3:06
  pentests,
• 3:07
  including
• 3:08
  the
• 3:08
  NSA's
• 3:08
  Continuous
• 3:09
  Autonomous
• 3:09
  Penetration
• 3:10
  Testing
• 3:10
  program.
• 3:11
  XBOW,
• 3:12
  now
• 3:12
  at
• 3:13
  $273
• 3:14
  million
• 3:14
  in
• 3:14
  funding,
• 3:15
  became
• 3:15
  the
• 3:15
  first
• 3:16
  autonomous
• 3:16
  system
• 3:16
  to
• 3:17
  rank
• 3:17
  #1
• 3:17
  on
• 3:17
  HackerOne,
• 3:18
  identifying
• 3:19
  over
• 3:19
  1,000
• 3:19
  vulnerabilities.
• 3:21
  The
• 3:21
  exploitation
• 3:22
  timeline
• 3:22
  data
• 3:22
  is
• 3:22
  stark:
• 3:23
  from
• 3:23
  771
• 3:24
  days
• 3:25
  between
• 3:25
  disclosure
• 3:25
  and
• 3:26
  exploitation
• 3:26
  in
• 3:26
  2006,
• 3:28
  to
• 3:28
  under
• 3:28
  24
• 3:28
  hours
• 3:29
  in
• 3:29
  recent
• 3:29
  observations,
• 3:30
  to
• 3:30
  Google's
• 3:31
  Threat
• 3:31
  Analysis
• 3:31
  Group
• 3:32
  documenting
• 3:32
  exploitation
• 3:33
  within
• 3:33
  four
• 3:33
  hours.
• 3:34
  When
• 3:34
  your
• 3:34
  quarterly
• 3:35
  pentest
• 3:35
  report
• 3:35
  is
• 3:36
  stale
• 3:36
  before
• 3:36
  the
• 3:36
  ink
• 3:37
  dries,
• 3:37
  the
• 3:37
  model
• 3:38
  is
• 3:38
  not
• 3:38
  struggling
• 3:38
  —
• 3:39
  it
• 3:39
  is
• 3:39
  failing.
• 3:44
  While
• 3:44
  AI
• 3:45
  security
• 3:45
  dominates
• 3:45
  the
• 3:46
  headlines,
• 3:46
  some
• 3:47
  of
• 3:47
  the
• 3:47
  most
• 3:47
  technically
• 3:47
  impressive
• 3:48
  offensive
• 3:48
  work
• 3:48
  happens
• 3:49
  at
• 3:49
  the
• 3:49
  hardware
• 3:49
  layer
• 3:50
  with
• 3:50
  a
• 3:50
  multimeter
• 3:50
  and
• 3:50
  patience.
• 3:51
  A
• 3:51
  new
• 3:52
  writeup
• 3:52
  documents
• 3:52
  the
• 3:52
  first
• 3:53
  publicly
• 3:53
  known
• 3:53
  side-channel
• 3:54
  key
• 3:54
  extraction
• 3:55
  from
• 3:55
  Texas
• 3:55
  Instruments'
• 3:55
  MSPM0G3507
• 3:58
  hardware
• 3:58
  AES
• 3:59
  engine
• 3:59
  —
• 3:59
  a
• 4:00
  microcontroller
• 4:00
  with
• 4:00
  no
• 4:01
  prior
• 4:01
  published
• 4:01
  SCA
• 4:01
  research.
• 4:02
  The
• 4:02
  critical
• 4:03
  insight
• 4:03
  that
• 4:03
  makes
• 4:03
  or
• 4:03
  breaks
• 4:04
  this
• 4:04
  attack:
• 4:05
  hardware
• 4:05
  AES
• 4:05
  does
• 4:05
  not
• 4:06
  leak
• 4:06
  the
• 4:06
  same
• 4:06
  way
• 4:06
  software
• 4:07
  AES
• 4:07
  does.
• 4:08
  In
• 4:08
  software,
• 4:09
  every
• 4:09
  intermediate
• 4:10
  value
• 4:10
  hits
• 4:10
  a
• 4:10
  data
• 4:11
  bus,
• 4:11
  leaking
• 4:12
  power
• 4:12
  proportional
• 4:12
  to
• 4:12
  its
• 4:13
  Hamming
• 4:13
  Weight.
• 4:14
  Hardware
• 4:14
  AES
• 4:14
  processes
• 4:15
  rounds
• 4:15
  through
• 4:16
  dedicated
• 4:16
  logic
• 4:17
  where
• 4:17
  intermediate
• 4:17
  values
• 4:18
  are
• 4:18
  wired
• 4:18
  directly
• 4:18
  between
• 4:19
  stages
• 4:19
  and
• 4:19
  never
• 4:20
  touch
• 4:20
  a
• 4:20
  bus.
• 4:21
  The
• 4:21
  actual
• 4:22
  leakage
• 4:22
  event
• 4:22
  occurs
• 4:22
  at
• 4:23
  round
• 4:23
  boundaries,
• 4:23
  when
• 4:24
  the
• 4:24
  AES
• 4:24
  engine
• 4:25
  writes
• 4:25
  new
• 4:25
  state
• 4:25
  back
• 4:25
  to
• 4:26
  an
• 4:26
  internal
• 4:26
  register,
• 4:27
  leaking
• 4:27
  in
• 4:27
  proportion
• 4:28
  to
• 4:28
  the
• 4:28
  Hamming
• 4:28
  Distance
• 4:29
  between
• 4:29
  old
• 4:29
  and
• 4:29
  new
• 4:30
  state.
• 4:30
  The
• 4:31
  researcher
• 4:31
  used
• 4:31
  a
• 4:31
  ChipWhisperer
• 4:32
  Nano
• 4:32
  clocked
• 4:33
  synchronously
• 4:33
  from
• 4:33
  the
• 4:33
  MSPM0's
• 4:35
  own
• 4:35
  clock
• 4:35
  output
• 4:36
  (1
• 4:36
  sample
• 4:36
  per
• 4:37
  cycle
• 4:37
  —
• 4:38
  eliminating
• 4:38
  jitter
• 4:39
  entirely),
• 4:39
  removed
• 4:40
  the
• 4:40
  onboard
• 4:40
  decoupling
• 4:41
  capacitor
• 4:41
  to
• 4:41
  expose
• 4:42
  the
• 4:42
  VCORE
• 4:42
  rail,
• 4:43
  and
• 4:43
  performed
• 4:43
  reverse
• 4:44
  CPA
• 4:44
  with
• 4:44
  the
• 4:44
  known
• 4:44
  key
• 4:45
  to
• 4:45
  map
• 4:45
  all
• 4:45
  10
• 4:45
  AES
• 4:46
  rounds
• 4:46
  across
• 4:47
  the
• 4:47
  trace
• 4:47
  before
• 4:47
  attempting
• 4:48
  the
• 4:48
  actual
• 4:48
  attack.
• 4:49
  The
• 4:50
  custom
• 4:50
  selection
• 4:50
  function
• 4:51
  targets
• 4:51
  the
• 4:51
  Hamming
• 4:52
  Distance
• 4:52
  across
• 4:52
  SubBytes
• 4:53
  in
• 4:53
  the
• 4:53
  final
• 4:53
  round,
• 4:54
  and
• 4:54
  full
• 4:55
  AES-128
• 4:56
  key
• 4:57
  recovery
• 4:57
  required
• 4:58
  just
• 4:58
  45,500
• 4:59
  traces
• 5:00
  out
• 5:00
  of
• 5:00
  100,000
• 5:01
  captured.
• 5:02
  TI
• 5:02
  has
• 5:03
  a
• 5:03
  standing
• 5:03
  advisory
• 5:04
  acknowledging
• 5:04
  power
• 5:04
  analysis
• 5:05
  vulnerability
• 5:06
  with
• 5:06
  a
• 5:06
  CVSS
• 5:06
  score
• 5:07
  of
• 5:07
  4.8–6.1,
• 5:09
  but
• 5:09
  the
• 5:10
  MSPM0G3507
• 5:13
  carries
• 5:13
  no
• 5:14
  documented
• 5:14
  mitigations.
• 5:15
  For
• 5:16
  anyone
• 5:16
  deploying
• 5:17
  this
• 5:17
  chip
• 5:17
  in
• 5:17
  key-protecting
• 5:18
  roles,
• 5:18
  the
• 5:18
  message
• 5:19
  is
• 5:19
  clear.
• 5:20
  On
• 5:20
  the
• 5:20
  automated
• 5:21
  pentest
• 5:21
  front,
• 5:22
  RedAmon
• 5:22
  hit
• 5:22
  version
• 5:23
  3.0
• 5:23
  with
• 5:23
  25+
• 5:24
  features
• 5:25
  including
• 5:25
  eight
• 5:25
  new
• 5:26
  tools
• 5:26
  in
• 5:26
  the
• 5:26
  discovery
• 5:27
  pipeline
• 5:27
  (ffuf,
• 5:28
  Arjun,
• 5:28
  ParamSpider,
• 5:29
  Amass,
• 5:30
  puredns,
• 5:31
  Katana,
• 5:32
  SecretFinder,
• 5:32
  theHarvester),
• 5:33
  parallel
• 5:34
  fan-out/fan-in
• 5:35
  execution,
• 5:36
  a
• 5:36
  Wave
• 5:37
  Runner
• 5:37
  system
• 5:37
  that
• 5:37
  fires
• 5:38
  independent
• 5:38
  tools
• 5:39
  simultaneously,
• 5:40
  and
• 5:40
  a
• 5:40
  Tool
• 5:40
  Confirmation
• 5:41
  Gate
• 5:41
  for
• 5:41
  human-in-the-loop
• 5:42
  control
• 5:43
  before
• 5:43
  destructive
• 5:44
  tools
• 5:44
  like
• 5:44
  nmap
• 5:45
  or
• 5:45
  Metasploit
• 5:46
  execute.
• 5:47
  Rules
• 5:47
  of
• 5:47
  Engagement
• 5:48
  can
• 5:48
  now
• 5:48
  be
• 5:48
  uploaded
• 5:49
  as
• 5:49
  PDF/DOCX
• 5:52
  and
• 5:52
  automatically
• 5:52
  parsed
• 5:53
  into
• 5:53
  30+
• 5:54
  structured
• 5:54
  enforcement
• 5:55
  settings.
• 6:00
  On
• 6:00
  March
• 6:00
  22,
• 6:01
  an
• 6:01
  attacker
• 6:01
  minted
• 6:02
  80
• 6:02
  million
• 6:03
  unbacked
• 6:03
  USR
• 6:03
  stablecoins
• 6:04
  on
• 6:04
  the
• 6:04
  Resolv
• 6:05
  protocol
• 6:05
  and
• 6:06
  extracted
• 6:06
  roughly
• 6:06
  $23
• 6:07
  million
• 6:07
  in
• 6:08
  ETH
• 6:08
  —
• 6:09
  not
• 6:09
  by
• 6:09
  exploiting
• 6:09
  a
• 6:09
  smart
• 6:10
  contract
• 6:10
  bug,
• 6:10
  but
• 6:10
  by
• 6:11
  compromising
• 6:11
  the
• 6:11
  protocol's
• 6:12
  AWS
• 6:12
  Key
• 6:13
  Management
• 6:13
  Service
• 6:13
  environment
• 6:14
  and
• 6:14
  using
• 6:14
  its
• 6:15
  own
• 6:15
  signing
• 6:15
  key
• 6:15
  against
• 6:16
  it.
• 6:17
  The
• 6:17
  smart
• 6:17
  contract
• 6:18
  worked
• 6:18
  exactly
• 6:18
  as
• 6:18
  designed.
• 6:19
  It
• 6:19
  checked
• 6:20
  for
• 6:20
  a
• 6:20
  valid
• 6:20
  signature
• 6:21
  from
• 6:21
  the
• 6:21
  SERVICE
• 6:21
  ROLE
• 6:21
  key,
• 6:22
  found
• 6:22
  one,
• 6:23
  and
• 6:23
  minted
• 6:23
  whatever
• 6:24
  amount
• 6:24
  was
• 6:24
  requested.
• 6:25
  There
• 6:26
  was
• 6:26
  no
• 6:26
  on-chain
• 6:26
  ratio
• 6:27
  check
• 6:27
  between
• 6:27
  collateral
• 6:28
  deposited
• 6:28
  and
• 6:28
  USR
• 6:29
  minted,
• 6:29
  no
• 6:29
  price
• 6:30
  oracle,
• 6:30
  no
• 6:30
  cap.
• 6:31
  The
• 6:31
  attacker
• 6:32
  deposited
• 6:32
  ~$100K–$200K
• 6:34
  in
• 6:34
  USDC,
• 6:35
  then
• 6:35
  used
• 6:35
  the
• 6:36
  compromised
• 6:36
  key
• 6:36
  to
• 6:36
  authorize
• 6:37
  tens
• 6:37
  of
• 6:37
  millions
• 6:37
  in
• 6:38
  USR
• 6:38
  output
• 6:38
  across
• 6:39
  two
• 6:39
  transactions.
• 6:40
  The
• 6:40
  attacker
• 6:40
  then
• 6:41
  converted
• 6:41
  the
• 6:41
  unbacked
• 6:42
  USR
• 6:42
  to
• 6:42
  wstUSR
• 6:43
  (a
• 6:43
  staking
• 6:43
  pool
• 6:44
  derivative),
• 6:45
  rotated
• 6:45
  through
• 6:45
  multiple
• 6:46
  DEX
• 6:46
  pools
• 6:47
  and
• 6:47
  bridges
• 6:47
  into
• 6:47
  stablecoins,
• 6:48
  and
• 6:48
  finally
• 6:49
  into
• 6:49
  ETH
• 6:49
  —
• 6:50
  a
• 6:50
  textbook
• 6:51
  layered
• 6:51
  extraction
• 6:52
  designed
• 6:52
  to
• 6:52
  maximize
• 6:52
  value
• 6:53
  before
• 6:53
  detection.
• 6:54
  USR's
• 6:55
  peg
• 6:55
  collapsed
• 6:55
  by
• 6:56
  80%.
• 6:57
  Resolv
• 6:57
  had
• 6:57
  undergone
• 6:58
  multiple
• 6:58
  audits.
• 7:00
  The
• 7:00
  takeaway
• 7:00
  is
• 7:00
  not
• 7:00
  that
• 7:01
  audits
• 7:01
  are
• 7:01
  useless
• 7:01
  —
• 7:02
  it
• 7:02
  is
• 7:02
  that
• 7:03
  auditing
• 7:03
  smart
• 7:03
  contracts
• 7:04
  while
• 7:04
  ignoring
• 7:05
  the
• 7:05
  off-chain
• 7:05
  infrastructure
• 7:06
  that
• 7:06
  controls
• 7:07
  privileged
• 7:07
  keys
• 7:08
  is
• 7:08
  security
• 7:08
  theater.
• 7:10
  As
• 7:10
  DeFi
• 7:10
  systems
• 7:10
  increasingly
• 7:11
  depend
• 7:11
  on
• 7:11
  cloud
• 7:12
  services,
• 7:12
  external
• 7:13
  signing
• 7:13
  keys,
• 7:14
  and
• 7:14
  off-chain
• 7:14
  approval
• 7:15
  workflows,
• 7:16
  the
• 7:16
  actual
• 7:16
  attack
• 7:16
  surface
• 7:17
  has
• 7:17
  quietly
• 7:17
  migrated
• 7:18
  from
• 7:18
  Solidity
• 7:19
  to
• 7:19
  AWS
• 7:19
  IAM.
• 7:24
  Walmart
• 7:25
  tested
• 7:25
  roughly
• 7:26
  200,000
• 7:26
  products
• 7:27
  through
• 7:27
  OpenAI's
• 7:28
  Instant
• 7:28
  Checkout,
• 7:29
  which
• 7:29
  let
• 7:29
  users
• 7:29
  complete
• 7:30
  purchases
• 7:30
  entirely
• 7:31
  inside
• 7:31
  ChatGPT
• 7:32
  without
• 7:32
  visiting
• 7:32
  Walmart's
• 7:33
  site.
• 7:34
  The
• 7:34
  results
• 7:34
  were
• 7:34
  blunt:
• 7:35
  in-chat
• 7:35
  purchases
• 7:36
  converted
• 7:36
  at
• 7:36
  one-third
• 7:37
  the
• 7:37
  rate
• 7:37
  of
• 7:37
  transactions
• 7:38
  where
• 7:38
  users
• 7:38
  clicked
• 7:38
  through
• 7:38
  to
• 7:39
  Walmart's
• 7:39
  website.
• 7:40
  Daniel
• 7:40
  Danker,
• 7:41
  Walmart's
• 7:41
  EVP
• 7:41
  of
• 7:42
  product
• 7:42
  and
• 7:42
  design,
• 7:42
  called
• 7:43
  the
• 7:43
  experience
• 7:44
  "unsatisfying."
• 7:45
  OpenAI
• 7:46
  has
• 7:46
  since
• 7:46
  confirmed
• 7:46
  it
• 7:46
  is
• 7:46
  phasing
• 7:47
  out
• 7:47
  Instant
• 7:47
  Checkout
• 7:48
  entirely
• 7:48
  in
• 7:49
  favor
• 7:49
  of
• 7:49
  merchant-handled
• 7:50
  app-based
• 7:50
  checkout.
• 7:51
  Walmart
• 7:51
  will
• 7:52
  now
• 7:52
  embed
• 7:52
  its
• 7:52
  own
• 7:52
  chatbot,
• 7:53
  Sparky,
• 7:53
  inside
• 7:54
  ChatGPT,
• 7:54
  with
• 7:54
  a
• 7:55
  similar
• 7:55
  integration
• 7:56
  coming
• 7:56
  to
• 7:56
  Google
• 7:56
  Gemini
• 7:57
  next
• 7:57
  month.
• 7:57
  The
• 7:58
  implication
• 7:58
  extends
• 7:58
  beyond
• 7:59
  Walmart.
• 7:59
  Agentic
• 8:00
  commerce
• 8:00
  —
• 8:00
  the
• 8:00
  idea
• 8:01
  that
• 8:01
  AI
• 8:01
  assistants
• 8:02
  will
• 8:02
  handle
• 8:02
  purchasing
• 8:02
  end-to-end
• 8:03
  —
• 8:03
  assumes
• 8:04
  users
• 8:04
  trust
• 8:04
  a
• 8:04
  conversational
• 8:05
  interface
• 8:05
  for
• 8:05
  high-intent
• 8:06
  actions
• 8:06
  like
• 8:06
  checkout.
• 8:07
  The
• 8:07
  3x
• 8:08
  conversion
• 8:08
  gap
• 8:08
  suggests
• 8:09
  they
• 8:09
  do
• 8:09
  not,
• 8:09
  at
• 8:09
  least
• 8:09
  not
• 8:10
  yet.
• 8:10
  Shopping
• 8:11
  is
• 8:11
  a
• 8:11
  visual,
• 8:12
  comparative
• 8:12
  activity
• 8:12
  that
• 8:13
  conversational
• 8:13
  UIs
• 8:13
  compress
• 8:14
  into
• 8:14
  a
• 8:14
  linear
• 8:14
  flow.
• 8:15
  Sending
• 8:16
  users
• 8:16
  back
• 8:16
  to
• 8:16
  owned
• 8:16
  environments
• 8:17
  with
• 8:17
  product
• 8:17
  images,
• 8:18
  reviews,
• 8:18
  and
• 8:18
  cart
• 8:18
  persistence
• 8:19
  still
• 8:19
  wins.
• 8:20
  That
• 8:21
  said,
• 8:21
  the
• 8:21
  pivot
• 8:21
  to
• 8:21
  embedded
• 8:21
  merchant
• 8:22
  chatbots
• 8:22
  within
• 8:22
  AI
• 8:23
  platforms
• 8:23
  may
• 8:23
  land
• 8:24
  differently
• 8:24
  —
• 8:25
  the
• 8:25
  user
• 8:25
  stays
• 8:25
  in
• 8:25
  the
• 8:25
  AI
• 8:26
  context
• 8:26
  but
• 8:26
  transacts
• 8:27
  through
• 8:27
  a
• 8:27
  familiar
• 8:28
  merchant
• 8:28
  experience.
• 8:29
  Worth
• 8:29
  watching
• 8:30
  whether
• 8:30
  Sparky-in-ChatGPT
• 8:31
  closes
• 8:31
  the
• 8:31
  gap
• 8:32
  or
• 8:32
  confirms
• 8:32
  it.
• 8:35
  ByteDance
• 8:36
  open-sourced
• 8:37
  DeerFlow
• 8:37
  2.0
• 8:38
  and
• 8:38
  within
• 8:38
  24
• 8:39
  hours
• 8:39
  it
• 8:39
  topped
• 8:40
  GitHub
• 8:40
  Trending,
• 8:41
  accumulating
• 8:41
  ~25,000
• 8:42
  stars.
• 8:44
  What
• 8:44
  distinguishes
• 8:44
  DeerFlow
• 8:45
  from
• 8:45
  the
• 8:45
  crowded
• 8:46
  agent
• 8:46
  framework
• 8:46
  landscape
• 8:47
  (CrewAI,
• 8:48
  AutoGen,
• 8:48
  LangGraph)
• 8:49
  is
• 8:49
  that
• 8:49
  it
• 8:49
  ships
• 8:49
  as
• 8:50
  a
• 8:50
  full
• 8:50
  runtime,
• 8:51
  not
• 8:51
  a
• 8:51
  library.
• 8:52
  Built
• 8:52
  on
• 8:52
  LangGraph
• 8:53
  and
• 8:53
  LangChain,
• 8:54
  DeerFlow's
• 8:54
  lead
• 8:55
  agent
• 8:55
  decomposes
• 8:56
  tasks,
• 8:57
  spawns
• 8:57
  sub-agents
• 8:58
  with
• 8:58
  scoped
• 8:58
  contexts
• 8:59
  and
• 8:59
  tools,
• 9:00
  runs
• 9:00
  them
• 9:00
  in
• 9:00
  isolated
• 9:01
  Docker
• 9:01
  containers,
• 9:02
  and
• 9:02
  stitches
• 9:02
  results
• 9:03
  back
• 9:03
  together.
• 9:04
  Each
• 9:04
  sub-agent
• 9:05
  gets
• 9:05
  its
• 9:05
  own
• 9:06
  filesystem,
• 9:07
  bash
• 9:07
  terminal,
• 9:08
  and
• 9:08
  code
• 9:08
  execution
• 9:09
  capability.
• 9:10
  Skills
• 9:10
  are
• 9:10
  Markdown
• 9:11
  files
• 9:11
  loaded
• 9:11
  progressively
• 9:12
  to
• 9:12
  avoid
• 9:12
  context
• 9:13
  window
• 9:13
  bloat,
• 9:14
  and
• 9:14
  a
• 9:14
  persistent
• 9:14
  JSON-based
• 9:15
  memory
• 9:15
  system
• 9:15
  tracks
• 9:16
  preferences
• 9:16
  and
• 9:16
  project
• 9:17
  context
• 9:17
  across
• 9:17
  sessions
• 9:18
  —
• 9:19
  with
• 9:19
  a
• 9:19
  new
• 9:19
  TIAMAT
• 9:21
  cloud
• 9:22
  backend
• 9:22
  suggesting
• 9:22
  enterprise
• 9:23
  ambitions.
• 9:24
  The
• 9:24
  security
• 9:25
  considerations
• 9:25
  are
• 9:25
  worth
• 9:26
  noting:
• 9:27
  ByteDance
• 9:27
  ownership
• 9:27
  triggers
• 9:28
  review
• 9:28
  processes
• 9:29
  regardless
• 9:29
  of
• 9:29
  technical
• 9:29
  merit,
• 9:30
  and
• 9:30
  the
• 9:30
  recommendation
• 9:31
  to
• 9:31
  deploy
• 9:31
  containerized
• 9:32
  with
• 9:32
  hardened
• 9:32
  images
• 9:33
  applies
• 9:33
  to
• 9:33
  any
• 9:34
  agent
• 9:34
  platform
• 9:35
  that
• 9:35
  executes
• 9:35
  code.
• 9:36
  AgencyCLI
• 9:37
  takes
• 9:37
  a
• 9:37
  fundamentally
• 9:38
  different
• 9:38
  approach:
• 9:39
  infrastructure
• 9:39
  rather
• 9:40
  than
• 9:40
  framework.
• 9:41
  You
• 9:41
  define
• 9:41
  org
• 9:42
  charts
• 9:42
  in
• 9:42
  Markdown
• 9:42
  and
• 9:43
  YAML
• 9:43
  —
• 9:44
  teams,
• 9:44
  roles,
• 9:45
  projects,
• 9:45
  skills
• 9:46
  —
• 9:46
  and
• 9:47
  agents
• 9:47
  assemble
• 9:47
  their
• 9:47
  own
• 9:48
  context
• 9:48
  and
• 9:48
  run
• 9:48
  autonomously
• 9:49
  on
• 9:49
  a
• 9:49
  heartbeat
• 9:50
  schedule.
• 9:50
  The
• 9:51
  key
• 9:51
  differentiator
• 9:51
  is
• 9:52
  inter-agent
• 9:52
  communication:
• 9:53
  agents
• 9:54
  can
• 9:54
  hire,
• 9:54
  message,
• 9:55
  and
• 9:55
  coordinate
• 9:55
  with
• 9:55
  each
• 9:56
  other
• 9:56
  through
• 9:56
  an
• 9:56
  async
• 9:56
  inbox
• 9:57
  system,
• 9:58
  with
• 9:58
  confirm-request
• 9:59
  gates
• 9:59
  for
• 9:59
  human-in-the-loop
• 10:00
  approval.
• 10:01
  It
• 10:01
  is
• 10:01
  model-agnostic
• 10:02
  (Claude
• 10:02
  Code,
• 10:03
  Codex,
• 10:04
  Gemini,
• 10:04
  Cursor),
• 10:05
  runs
• 10:06
  agents
• 10:06
  in
• 10:06
  Docker
• 10:06
  sandboxes
• 10:07
  by
• 10:07
  default,
• 10:08
  and
• 10:08
  requires
• 10:09
  no
• 10:09
  server.
• 10:10
  For
• 10:10
  model
• 10:10
  routing,
• 10:11
  SmarterRouter
• 10:11
  2.2.1
• 10:13
  offers
• 10:13
  a
• 10:14
  self-hosted
• 10:14
  MoE-style
• 10:15
  proxy
• 10:15
  that
• 10:16
  profiles
• 10:16
  incoming
• 10:17
  prompts
• 10:17
  and
• 10:17
  routes
• 10:18
  to
• 10:18
  the
• 10:18
  best
• 10:18
  backend
• 10:18
  model
• 10:19
  —
• 10:20
  local
• 10:20
  via
• 10:20
  Ollama
• 10:21
  or
• 10:21
  external
• 10:21
  via
• 10:22
  OpenAI-compatible
• 10:23
  APIs.
• 10:24
  It
• 10:24
  is
• 10:24
  multimodality-aware
• 10:25
  and
• 10:25
  optimized
• 10:26
  for
• 10:26
  rapid
• 10:27
  Ollama
• 10:27
  model
• 10:27
  loading/unloading.
• 10:29
  Anthropic's
• 10:30
  Claude
• 10:30
  Dispatch
• 10:31
  lets
• 10:31
  users
• 10:31
  control
• 10:32
  desktop
• 10:32
  AI
• 10:32
  tasks
• 10:33
  from
• 10:33
  their
• 10:33
  phone,
• 10:34
  extending
• 10:34
  the
• 10:34
  Cowork
• 10:35
  environment
• 10:35
  into
• 10:35
  a
• 10:35
  persistent
• 10:36
  assistant
• 10:36
  paradigm.
• 10:38
  Community
• 10:38
  reaction
• 10:38
  was
• 10:39
  mixed
• 10:39
  —
• 10:40
  some
• 10:40
  found
• 10:40
  it
• 10:40
  transformative
• 10:41
  for
• 10:41
  cross-device
• 10:42
  workflows
• 10:43
  (one
• 10:43
  user
• 10:43
  had
• 10:43
  it
• 10:43
  search
• 10:44
  multiple
• 10:44
  Gmail
• 10:44
  accounts,
• 10:45
  save
• 10:45
  closing
• 10:46
  documents
• 10:46
  to
• 10:46
  Obsidian,
• 10:47
  and
• 10:47
  fill
• 10:47
  out
• 10:48
  a
• 10:48
  tax
• 10:48
  form),
• 10:49
  while
• 10:49
  others
• 10:49
  criticized
• 10:50
  the
• 10:50
  growing
• 10:50
  fragmentation
• 10:51
  of
• 10:51
  Claude
• 10:51
  Desktop
• 10:51
  into
• 10:52
  disconnected
• 10:52
  feature
• 10:53
  panes.
• 10:54
  ColeMedin's
• 10:54
  deep
• 10:55
  dive
• 10:55
  on
• 10:55
  Claude
• 10:55
  Code
• 10:56
  covers
• 10:56
  roughly
• 10:56
  a
• 10:56
  dozen
• 10:57
  features
• 10:57
  shipped
• 10:57
  in
• 10:57
  just
• 10:58
  the
• 10:58
  last
• 10:58
  couple
• 10:58
  of
• 10:58
  months:
• 11:00
  the
• 11:00
  1M
• 11:00
  token
• 11:01
  context
• 11:01
  window
• 11:02
  (with
• 11:02
  the
• 11:02
  practical
• 11:02
  caveat
• 11:03
  that
• 11:03
  hallucinations
• 11:04
  spike
• 11:04
  sharply
• 11:05
  at
• 11:05
  250K–300K
• 11:06
  tokens),
• 11:08
  agent
• 11:08
  teams
• 11:08
  with
• 11:08
  real-time
• 11:09
  inter-agent
• 11:09
  communication,
• 11:10
  native
• 11:11
  git
• 11:11
  worktree
• 11:11
  support
• 11:12
  for
• 11:12
  parallel
• 11:12
  feature
• 11:13
  branches,
• 11:14
  /simplify
• 11:15
  and
• 11:15
  /batch
• 11:15
  commands
• 11:16
  for
• 11:16
  post-implementation
• 11:17
  cleanup
• 11:17
  and
• 11:17
  large-scale
• 11:18
  refactors,
• 11:19
  /btw
• 11:20
  for
• 11:20
  sidecar
• 11:21
  questions
• 11:22
  that
• 11:22
  do
• 11:22
  not
• 11:22
  bloat
• 11:22
  context,
• 11:23
  /loop
• 11:24
  for
• 11:24
  recurring
• 11:25
  prompts,
• 11:26
  voice
• 11:26
  input,
• 11:27
  effort-level
• 11:27
  tuning,
• 11:28
  and
• 11:28
  scheduled
• 11:29
  cron
• 11:29
  tasks.
• 11:30
  The
• 11:31
  worktree
• 11:31
  support
• 11:31
  is
• 11:31
  arguably
• 11:32
  the
• 11:32
  most
• 11:32
  impactful
• 11:33
  for
• 11:33
  daily
• 11:33
  use
• 11:33
  —
• 11:34
  real
• 11:34
  development
• 11:35
  always
• 11:35
  involves
• 11:36
  multiple
• 11:36
  feature
• 11:36
  branches,
• 11:37
  and
• 11:37
  having
• 11:38
  Claude
• 11:38
  manage
• 11:38
  isolated
• 11:39
  copies
• 11:39
  natively
• 11:39
  eliminates
• 11:40
  the
• 11:40
  manual
• 11:40
  worktree
• 11:41
  juggling
• 11:41
  that
• 11:41
  was
• 11:41
  previously
• 11:42
  required.
• 11:42
  Swictation
• 11:43
  v0.7.30
• 11:45
  shipped
• 11:45
  the
• 11:45
  first
• 11:46
  native
• 11:46
  macOS
• 11:46
  release,
• 11:47
  bringing
• 11:47
  cross-platform
• 11:48
  voice-to-text
• 11:48
  dictation
• 11:49
  to
• 11:49
  both
• 11:49
  Linux
• 11:50
  (NVIDIA
• 11:50
  CUDA)
• 11:50
  and
• 11:50
  Apple
• 11:50
  Silicon
• 11:51
  (CoreML/ANE
• 11:54
  via
• 11:54
  a
• 11:54
  custom
• 11:54
  coreml-native
• 11:55
  crate)
• 11:55
  with
• 11:55
  a
• 11:55
  single
• 11:56
  npm
• 11:56
  install
• 11:57
  -g
• 11:57
  swictation.
• 11:58
  The
• 11:58
  release
• 11:58
  also
• 11:59
  adds
• 11:59
  a
• 11:59
  native
• 11:59
  Tauri
• 12:00
  2
• 12:00
  menu
• 12:00
  bar
• 12:00
  tray
• 12:00
  with
• 12:01
  idle/recording/disconnected
• 12:02
  states
• 12:03
  and
• 12:03
  Unix
• 12:03
  socket
• 12:04
  IPC.
• 12:07
  Baidu's
• 12:08
  Qianfan-OCR
• 12:09
  packs
• 12:09
  document
• 12:09
  parsing,
• 12:09
  layout
• 12:10
  analysis,
• 12:10
  table
• 12:11
  extraction,
• 12:11
  formula
• 12:11
  recognition,
• 12:12
  chart
• 12:12
  understanding,
• 12:13
  and
• 12:13
  key
• 12:13
  information
• 12:14
  extraction
• 12:14
  into
• 12:15
  a
• 12:15
  single
• 12:15
  4-billion-parameter
• 12:16
  vision-language
• 12:16
  model
• 12:17
  —
• 12:17
  and
• 12:17
  it
• 12:17
  leads
• 12:18
  OmniDocBench
• 12:18
  v1.5
• 12:20
  at
• 12:20
  93.12,
• 12:21
  surpassing
• 12:21
  models
• 12:22
  60x
• 12:22
  its
• 12:22
  size
• 12:23
  on
• 12:23
  key
• 12:23
  information
• 12:24
  extraction
• 12:24
  tasks
• 12:25
  (beating
• 12:25
  Gemini-3.1-Pro
• 12:26
  and
• 12:26
  Qwen3-VL-235B).
• 12:30
  The
• 12:30
  key
• 12:30
  contribution
• 12:30
  is
• 12:31
  Layout-as-Thought:
• 12:32
  an
• 12:32
  optional
• 12:33
  reasoning
• 12:33
  phase
• 12:33
  where
• 12:33
  the
• 12:33
  model
• 12:34
  explicitly
• 12:34
  reasons
• 12:35
  about
• 12:35
  bounding
• 12:35
  boxes,
• 12:36
  element
• 12:37
  types,
• 12:37
  and
• 12:37
  reading
• 12:37
  order
• 12:38
  before
• 12:38
  generating
• 12:39
  structured
• 12:39
  output
• 12:39
  —
• 12:40
  document-layout-specific
• 12:41
  chain-of-thought
• 12:42
  that
• 12:42
  can
• 12:42
  be
• 12:42
  toggled
• 12:43
  at
• 12:43
  inference
• 12:43
  time
• 12:43
  depending
• 12:44
  on
• 12:44
  accuracy/speed
• 12:45
  requirements.
• 12:46
  Trained
• 12:46
  on
• 12:47
  2.85
• 12:47
  trillion
• 12:48
  tokens
• 12:48
  across
• 12:48
  192
• 12:49
  languages
• 12:50
  on
• 12:50
  1,024
• 12:51
  Kunlun
• 12:52
  P800
• 12:52
  chips,
• 12:53
  with
• 12:53
  weights
• 12:53
  fully
• 12:54
  open-sourced.
• 12:55
  At
• 12:55
  1.024
• 12:56
  pages/sec
• 12:57
  on
• 12:57
  a
• 12:57
  single
• 12:57
  A100
• 12:58
  with
• 12:58
  W8A8
• 12:59
  quantization,
• 12:59
  this
• 13:00
  is
• 13:00
  a
• 13:00
  production-viable
• 13:01
  document
• 13:01
  intelligence
• 13:02
  model.
• 13:03
  On
• 13:03
  the
• 13:03
  quantization
• 13:03
  tooling
• 13:04
  front,
• 13:04
  AMD's
• 13:05
  Quark
• 13:05
  has
• 13:05
  been
• 13:06
  flying
• 13:06
  under
• 13:06
  the
• 13:06
  radar
• 13:07
  with
• 13:07
  only
• 13:07
  a
• 13:07
  few
• 13:08
  hundred
• 13:08
  downloads
• 13:08
  per
• 13:08
  model
• 13:09
  update
• 13:09
  despite
• 13:10
  offering
• 13:10
  MXFP4
• 13:12
  post-training
• 13:13
  quantization
• 13:14
  that
• 13:14
  reportedly
• 13:14
  preserves
• 13:15
  better
• 13:15
  quality
• 13:16
  than
• 13:16
  standard
• 13:16
  approaches.
• 13:17
  One
• 13:17
  commenter
• 13:18
  noted
• 13:18
  finding
• 13:18
  bugs
• 13:19
  via
• 13:19
  Claude
• 13:19
  when
• 13:19
  working
• 13:20
  with
• 13:20
  Quark's
• 13:20
  codebase,
• 13:21
  suggesting
• 13:21
  the
• 13:22
  tooling
• 13:22
  maturity
• 13:22
  has
• 13:22
  room
• 13:23
  to
• 13:23
  grow.
• 13:24
  A
• 13:24
  community
• 13:25
  experiment
• 13:25
  compressing
• 13:26
  six
• 13:26
  LLMs
• 13:27
  found
• 13:27
  that
• 13:27
  models
• 13:27
  do
• 13:28
  not
• 13:28
  degrade
• 13:28
  uniformly
• 13:29
  under
• 13:29
  compression
• 13:29
  —
• 13:30
  one
• 13:31
  commenter
• 13:31
  hypothesized
• 13:32
  that
• 13:32
  steep
• 13:32
  degradation
• 13:33
  may
• 13:33
  indicate
• 13:34
  thorough
• 13:34
  training
• 13:34
  (all
• 13:35
  neurons
• 13:35
  carrying
• 13:35
  load),
• 13:36
  while
• 13:37
  graceful
• 13:37
  degradation
• 13:38
  may
• 13:38
  signal
• 13:38
  undertrained
• 13:39
  models
• 13:39
  with
• 13:39
  redundant
• 13:40
  neurons
• 13:40
  in
• 13:40
  benchmark
• 13:41
  blind
• 13:41
  spots.
• 13:42
  The
• 13:42
  methodology
• 13:43
  used
• 13:43
  uniform
• 13:44
  intermediate
• 13:44
  dimension
• 13:45
  reduction,
• 13:45
  with
• 13:46
  the
• 13:46
  author
• 13:46
  exploring
• 13:47
  targeted
• 13:47
  per-layer
• 13:47
  approaches
• 13:48
  next.
• 13:49
  HumeAI
• 13:49
  also
• 13:50
  released
• 13:50
  tada-1b,
• 13:51
  an
• 13:51
  expressive
• 13:52
  audio
• 13:52
  generation
• 13:53
  model
• 13:53
  on
• 13:53
  HuggingFace.
• 13:58
  NVIDIA
• 13:58
  published
• 13:58
  a
• 13:58
  complete
• 13:59
  six-command
• 13:59
  pipeline
• 14:00
  for
• 14:00
  building
• 14:00
  domain-specific
• 14:01
  embedding
• 14:01
  models
• 14:02
  in
• 14:02
  under
• 14:02
  a
• 14:02
  day
• 14:02
  on
• 14:02
  a
• 14:02
  single
• 14:03
  GPU
• 14:03
  —
• 14:04
  from
• 14:04
  raw
• 14:04
  documents
• 14:05
  to
• 14:05
  a
• 14:05
  deployed,
• 14:05
  OpenAI-compatible
• 14:06
  inference
• 14:07
  endpoint.
• 14:08
  The
• 14:08
  recipe
• 14:08
  chains
• 14:09
  NeMo
• 14:09
  Curator
• 14:09
  for
• 14:09
  synthetic
• 14:10
  data
• 14:10
  generation
• 14:10
  (using
• 14:11
  Nemotron
• 14:11
  to
• 14:11
  produce
• 14:11
  multi-hop
• 14:12
  question-answer
• 14:12
  pairs
• 14:13
  from
• 14:13
  domain
• 14:13
  documents),
• 14:14
  hard
• 14:15
  negative
• 14:15
  mining
• 14:15
  with
• 14:15
  a
• 14:15
  95%
• 14:16
  margin
• 14:16
  ceiling
• 14:16
  to
• 14:17
  avoid
• 14:17
  false
• 14:17
  negatives,
• 14:18
  contrastive
• 14:19
  fine-tuning
• 14:19
  with
• 14:19
  InfoNCE
• 14:20
  loss,
• 14:21
  BEIR-compatible
• 14:21
  evaluation,
• 14:22
  ONNX/TensorRT
• 14:27
  export,
• 14:27
  and
• 14:28
  NIM
• 14:28
  container
• 14:28
  deployment.
• 14:30
  On
• 14:30
  NVIDIA's
• 14:31
  own
• 14:31
  documentation
• 14:32
  corpus,
• 14:33
  the
• 14:33
  fine-tuned
• 14:33
  model
• 14:33
  showed
• 14:34
  10%+
• 14:34
  improvement
• 14:35
  in
• 14:35
  both
• 14:35
  NDCG@10
• 14:38
  and
• 14:38
  Recall@10.
• 14:40
  More
• 14:40
  compelling
• 14:41
  is
• 14:41
  Atlassian's
• 14:41
  validation:
• 14:42
  applying
• 14:43
  this
• 14:43
  exact
• 14:43
  pipeline
• 14:44
  to
• 14:44
  fine-tune
• 14:44
  on
• 14:44
  their
• 14:45
  Jira
• 14:45
  dataset
• 14:45
  pushed
• 14:46
  Recall@60
• 14:47
  from
• 14:47
  0.751
• 14:49
  to
• 14:49
  0.951
• 14:49
  —
• 14:51
  a
• 14:51
  26.7%
• 14:52
  gain
• 14:52
  that
• 14:52
  directly
• 14:53
  impacts
• 14:53
  search
• 14:54
  quality
• 14:54
  for
• 14:54
  millions
• 14:55
  of
• 14:55
  Rovo
• 14:55
  users.
• 14:56
  The
• 14:56
  practical
• 14:56
  threshold
• 14:57
  is
• 14:57
  accessible:
• 14:58
  50–100
• 15:00
  documents
• 15:00
  for
• 15:00
  a
• 15:00
  proof-of-concept,
• 15:01
  scaling
• 15:01
  up
• 15:02
  from
• 15:02
  there.
• 15:02
  For
• 15:03
  evaluating
• 15:03
  the
• 15:03
  memory
• 15:03
  systems
• 15:04
  that
• 15:04
  sit
• 15:04
  atop
• 15:04
  those
• 15:05
  embeddings,
• 15:05
  WMB-100K
• 15:08
  introduces
• 15:09
  a
• 15:09
  benchmark
• 15:10
  at
• 15:10
  a
• 15:10
  scale
• 15:10
  that
• 15:10
  actually
• 15:10
  reflects
• 15:11
  real
• 15:11
  usage.
• 15:12
  Existing
• 15:12
  benchmarks
• 15:13
  test
• 15:13
  at
• 15:13
  600–1,000
• 15:15
  turns.
• 15:15
  WMB-100K
• 15:18
  tests
• 15:18
  100,000
• 15:19
  turns
• 15:19
  with
• 15:19
  3,134
• 15:21
  questions
• 15:21
  across
• 15:21
  five
• 15:22
  difficulty
• 15:22
  levels,
• 15:22
  including
• 15:23
  false
• 15:23
  memory
• 15:23
  probes
• 15:24
  —
• 15:24
  because
• 15:25
  confidently
• 15:25
  returning
• 15:25
  wrong
• 15:26
  information
• 15:26
  is
• 15:26
  worse
• 15:26
  than
• 15:27
  saying
• 15:27
  "I
• 15:27
  don't
• 15:27
  know."
• 15:28
  At
• 15:28
  $0.07
• 15:28
  per
• 15:29
  run,
• 15:29
  the
• 15:29
  barrier
• 15:30
  to
• 15:30
  evaluation
• 15:30
  is
• 15:30
  essentially
• 15:31
  zero.
• 15:32
  Finally,
• 15:33
  a
• 15:33
  practitioner
• 15:33
  took
• 15:33
  Karpathy's
• 15:34
  autoresearch
• 15:35
  framework
• 15:35
  —
• 15:36
  a
• 15:36
  tight
• 15:36
  hypothesize-edit-train-evaluate-commit/revert
• 15:38
  loop
• 15:39
  with
• 15:39
  an
• 15:39
  LLM
• 15:39
  agent
• 15:40
  in
• 15:40
  the
• 15:40
  middle
• 15:40
  —
• 15:41
  and
• 15:41
  pointed
• 15:41
  it
• 15:41
  at
• 15:41
  a
• 15:41
  real
• 15:42
  research
• 15:42
  problem:
• 15:43
  a
• 15:43
  contrastive
• 15:44
  learning
• 15:44
  model
• 15:44
  for
• 15:44
  image-text
• 15:45
  alignment
• 15:45
  using
• 15:46
  expert
• 15:46
  attention
• 15:46
  heatmaps.
• 15:47
  The
• 15:48
  agent
• 15:48
  ran
• 15:48
  42
• 15:48
  experiments
• 15:49
  in
• 15:49
  a
• 15:49
  single
• 15:49
  Saturday
• 15:50
  on
• 15:50
  one
• 15:50
  RTX
• 15:51
  4090,
• 15:52
  committing
• 15:52
  13
• 15:53
  and
• 15:53
  reverting
• 15:53
  29.
• 15:54
  The
• 15:54
  single
• 15:54
  biggest
• 15:55
  win
• 15:55
  was
• 15:55
  a
• 15:55
  bug
• 15:55
  fix:
• 15:56
  relaxing
• 15:56
  a
• 15:57
  temperature
• 15:57
  clamp
• 15:57
  from
• 15:57
  2
• 15:58
  dropped
• 15:58
  the
• 15:58
  mean
• 15:58
  rank
• 15:58
  by
• 15:58
  113
• 15:59
  points
• 15:59
  —
• 16:00
  worth
• 16:00
  more
• 16:00
  than
• 16:00
  all
• 16:01
  architecture
• 16:01
  changes
• 16:01
  combined.
• 16:02
  Hyperparameter
• 16:03
  tuning
• 16:03
  delivered
• 16:04
  steady
• 16:04
  gains.
• 16:05
  But
• 16:05
  when
• 16:05
  the
• 16:05
  agent
• 16:05
  moved
• 16:06
  to
• 16:06
  architectural
• 16:06
  modifications
• 16:07
  and
• 16:07
  moonshot
• 16:07
  ideas
• 16:08
  in
• 16:08
  later
• 16:08
  phases,
• 16:09
  the
• 16:09
  success
• 16:09
  rate
• 16:09
  dropped
• 16:10
  sharply
• 16:10
  —
• 16:11
  it
• 16:11
  was
• 16:11
  "throwing
• 16:11
  spaghetti
• 16:12
  at
• 16:12
  the
• 16:12
  wall."
• 16:13
  The
• 16:13
  pattern
• 16:13
  is
• 16:13
  familiar:
• 16:14
  LLM
• 16:15
  agents
• 16:15
  excel
• 16:16
  at
• 16:16
  structured
• 16:16
  search
• 16:17
  (hyperparameters,
• 16:18
  obvious
• 16:18
  bugs)
• 16:19
  but
• 16:19
  struggle
• 16:19
  with
• 16:19
  the
• 16:19
  creative
• 16:20
  leaps
• 16:20
  that
• 16:20
  define
• 16:21
  research
• 16:21
  breakthroughs.
• 16:22
  The
• 16:22
  containerized
• 16:23
  execution
• 16:24
  approach
• 16:24
  (no
• 16:24
  network
• 16:25
  access,
• 16:25
  locked-down
• 16:26
  file
• 16:26
  permissions,
• 16:27
  Docker-orchestrated
• 16:28
  training)
• 16:29
  is
• 16:29
  a
• 16:29
  sensible
• 16:29
  template
• 16:30
  for
• 16:30
  anyone
• 16:30
  wanting
• 16:31
  to
• 16:31
  let
• 16:31
  an
• 16:31
  agent
• 16:31
  iterate
• 16:32
  autonomously
• 16:32
  without
• 16:33
  handing
• 16:33
  it
• 16:33
  the
• 16:33
  keys
• 16:34
  to
• 16:34
  the
• 16:34
  kingdom.