Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 12, 2026

Transcript

AGI Dreams Podcast – March 12, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  The
• 0:00
  attack
• 0:00
  surface
• 0:01
  of
• 0:01
  an
• 0:01
  AI
• 0:01
  agent
• 0:02
  is
• 0:02
  not
• 0:02
  a
• 0:02
  port
• 0:02
  or
• 0:03
  a
• 0:03
  buffer
• 0:03
  —
• 0:03
  it
• 0:04
  is
• 0:04
  language
• 0:04
  itself.
• 0:05
  VotalAI,
• 0:06
  presenting
• 0:06
  at
• 0:06
  RSA
• 0:07
  Conference
• 0:07
  2026,
• 0:08
  has
• 0:08
  open-sourced
• 0:09
  a
• 0:09
  white-box
• 0:10
  red-teaming
• 0:10
  framework
• 0:11
  that
• 0:11
  takes
• 0:11
  this
• 0:11
  seriously:
• 0:12
  it
• 0:12
  reads
• 0:13
  your
• 0:13
  application's
• 0:13
  source
• 0:14
  code,
• 0:14
  maps
• 0:14
  every
• 0:15
  tool,
• 0:15
  role,
• 0:16
  and
• 0:16
  guardrail,
• 0:17
  and
• 0:17
  then
• 0:17
  generates
• 0:18
  LLM-powered
• 0:18
  attacks
• 0:19
  across
• 0:19
  47
• 0:19
  categories
• 0:20
  —
• 0:21
  from
• 0:21
  forged
• 0:21
  JWTs
• 0:22
  and
• 0:22
  RBAC
• 0:23
  bypass
• 0:23
  through
• 0:23
  steganographic
• 0:24
  exfiltration,
• 0:25
  memory
• 0:26
  poisoning,
• 0:26
  and
• 0:26
  multi-agent
• 0:27
  confused-deputy
• 0:28
  attacks.
• 0:29
  The
• 0:29
  key
• 0:29
  design
• 0:30
  choice
• 0:30
  is
• 0:30
  the
• 0:30
  white-box
• 0:31
  premise.
• 0:32
  Black-box
• 0:32
  testers
• 0:33
  treat
• 0:33
  an
• 0:33
  agent
• 0:34
  endpoint
• 0:34
  like
• 0:34
  a
• 0:34
  chatbot
• 0:35
  and
• 0:35
  hope
• 0:35
  something
• 0:35
  breaks.
• 0:36
  VotalAI's
• 0:37
  codebase
• 0:38
  analyzer
• 0:38
  bundles
• 0:38
  source
• 0:39
  files,
• 0:39
  sends
• 0:40
  them
• 0:40
  to
• 0:40
  an
• 0:40
  LLM,
• 0:41
  and
• 0:41
  extracts
• 0:41
  a
• 0:41
  structured
• 0:42
  map
• 0:42
  of
• 0:42
  tools,
• 0:43
  permissions,
• 0:43
  regex-based
• 0:44
  guardrails,
• 0:45
  sensitive
• 0:45
  data
• 0:46
  locations,
• 0:46
  and
• 0:47
  authentication
• 0:47
  weaknesses.
• 0:48
  Every
• 0:49
  attack
• 0:49
  module
• 0:49
  then
• 0:49
  uses
• 0:50
  that
• 0:50
  map
• 0:50
  to
• 0:50
  generate
• 0:51
  targeted,
• 0:51
  application-specific
• 0:52
  payloads.
• 0:54
  In
• 0:54
  benchmarking,
• 0:55
  the
• 0:55
  team
• 0:55
  claims
• 0:55
  white-box
• 0:56
  attacks
• 0:56
  find
• 0:56
  3–5x
• 0:58
  more
• 0:58
  vulnerabilities
• 0:59
  than
• 0:59
  equivalent
• 0:59
  black-box
• 1:00
  approaches.
• 1:01
  The
• 1:01
  framework
• 1:02
  runs
• 1:02
  adaptive
• 1:02
  rounds:
• 1:03
  each
• 1:03
  round's
• 1:04
  results
• 1:04
  inform
• 1:04
  the
• 1:04
  next,
• 1:05
  letting
• 1:05
  the
• 1:06
  system
• 1:06
  evolve
• 1:06
  its
• 1:06
  strategy
• 1:07
  mid-test.
• 1:08
  A
• 1:08
  demo
• 1:08
  run
• 1:08
  against
• 1:09
  a
• 1:09
  reference
• 1:09
  agentic
• 1:10
  app
• 1:10
  found
• 1:11
  14
• 1:11
  vulnerabilities
• 1:12
  in
• 1:12
  238
• 1:13
  attacks,
• 1:14
  including
• 1:14
  critical
• 1:15
  JWT
• 1:16
  forgery
• 1:16
  and
• 1:16
  role-escalation
• 1:17
  bugs
• 1:17
  —
• 1:18
  the
• 1:18
  kind
• 1:18
  of
• 1:18
  findings
• 1:19
  that
• 1:19
  would
• 1:19
  take
• 1:19
  a
• 1:19
  human
• 1:20
  pentester
• 1:20
  days
• 1:20
  to
• 1:21
  surface
• 1:21
  manually.
• 1:22
  The
• 1:22
  framework's
• 1:22
  47-category
• 1:24
  taxonomy
• 1:24
  is
• 1:24
  worth
• 1:25
  studying
• 1:25
  on
• 1:25
  its
• 1:25
  own.
• 1:26
  It
• 1:26
  includes
• 1:26
  agentic-specific
• 1:27
  attack
• 1:28
  classes
• 1:28
  that
• 1:28
  most
• 1:28
  traditional
• 1:29
  scanners
• 1:29
  do
• 1:29
  not
• 1:30
  even
• 1:30
  conceptualize:
• 1:31
  agentic
• 1:31
  workflow
• 1:32
  bypass
• 1:32
  (checkpoint
• 1:33
  injection,
• 1:34
  approval-gate
• 1:34
  forgery),
• 1:35
  agent
• 1:35
  reflection
• 1:36
  exploit
• 1:36
  (poisoning
• 1:37
  ReAct
• 1:37
  reasoning
• 1:37
  loops
• 1:38
  with
• 1:38
  fake
• 1:38
  Thought/Action/Observation
• 1:39
  blocks),
• 1:41
  and
• 1:41
  cross
• 1:41
  session
• 1:42
  injection
• 1:42
  (corrupting
• 1:43
  persistent
• 1:43
  memory
• 1:44
  and
• 1:44
  shared
• 1:44
  vector
• 1:44
  stores).
• 1:45
  For
• 1:46
  anyone
• 1:46
  building
• 1:46
  production
• 1:47
  agents
• 1:47
  with
• 1:47
  tool
• 1:47
  access,
• 1:48
  this
• 1:48
  taxonomy
• 1:49
  is
• 1:49
  a
• 1:49
  practical
• 1:50
  threat
• 1:50
  model.
• 1:51
  The
• 1:51
  tool
• 1:51
  supports
• 1:52
  OpenAI,
• 1:53
  Anthropic,
• 1:54
  and
• 1:54
  OpenRouter
• 1:54
  backends,
• 1:55
  making
• 1:55
  it
• 1:55
  provider-agnostic
• 1:56
  for
• 1:57
  the
• 1:57
  attack
• 1:57
  generation
• 1:57
  itself.
• 1:58
  While
• 1:59
  agentic
• 1:59
  AI
• 1:59
  security
• 2:00
  tooling
• 2:00
  matures,
• 2:01
  the
• 2:01
  human
• 2:01
  side
• 2:02
  of
• 2:02
  data
• 2:02
  protection
• 2:02
  is
• 2:03
  having
• 2:03
  a
• 2:03
  worse
• 2:03
  week.
• 2:04
  A
• 2:04
  whistleblower
• 2:05
  complaint
• 2:05
  alleges
• 2:06
  that
• 2:06
  a
• 2:06
  former
• 2:06
  DOGE
• 2:06
  employee
• 2:07
  claimed
• 2:07
  access
• 2:08
  to
• 2:08
  two
• 2:08
  highly
• 2:08
  sensitive
• 2:09
  Social
• 2:09
  Security
• 2:10
  Administration
• 2:10
  databases
• 2:11
  and
• 2:12
  planned
• 2:12
  to
• 2:12
  share
• 2:12
  the
• 2:12
  information
• 2:13
  with
• 2:13
  a
• 2:13
  private
• 2:13
  employer
• 2:14
  —
• 2:15
  an
• 2:15
  unprecedented
• 2:15
  breach
• 2:16
  allegation
• 2:16
  at
• 2:16
  an
• 2:16
  agency
• 2:17
  serving
• 2:17
  over
• 2:18
  70
• 2:18
  million
• 2:18
  Americans.
• 2:19
  The
• 2:20
  SSA's
• 2:20
  internal
• 2:21
  watchdog
• 2:21
  is
• 2:21
  investigating.
• 2:23
  No
• 2:23
  exploit
• 2:23
  code
• 2:23
  needed;
• 2:24
  just
• 2:24
  a
• 2:24
  person
• 2:25
  with
• 2:25
  access
• 2:26
  and,
• 2:26
  allegedly,
• 2:27
  questionable
• 2:27
  intentions.
• 2:31
  Running
• 2:32
  a
• 2:32
  120-billion-parameter
• 2:33
  model
• 2:33
  with
• 2:34
  a
• 2:34
  one-million-token
• 2:35
  context
• 2:35
  window
• 2:35
  on
• 2:35
  a
• 2:36
  single
• 2:36
  desktop
• 2:36
  machine
• 2:37
  sounds
• 2:37
  like
• 2:37
  a
• 2:37
  2028
• 2:38
  prediction.
• 2:38
  It
• 2:39
  is
• 2:39
  happening
• 2:39
  now.
• 2:40
  A
• 2:40
  LocalLLaMA
• 2:40
  user
• 2:41
  benchmarked
• 2:41
  Nemotron
• 2:42
  3
• 2:42
  Super
• 2:42
  (Q4
• 2:42
  K
• 2:43
  M
• 2:43
  quantization)
• 2:43
  on
• 2:43
  an
• 2:44
  M1
• 2:44
  Ultra
• 2:44
  with
• 2:44
  128
• 2:45
  GB
• 2:46
  unified
• 2:46
  memory,
• 2:47
  processing
• 2:47
  the
• 2:47
  full
• 2:47
  million-token
• 2:48
  prefill
• 2:48
  in
• 2:48
  approximately
• 2:49
  3
• 2:49
  hours
• 2:49
  and
• 2:50
  20
• 2:50
  minutes.
• 2:51
  At
• 2:51
  the
• 2:51
  start
• 2:51
  the
• 2:51
  model
• 2:51
  processes
• 2:52
  255
• 2:53
  tokens
• 2:53
  per
• 2:53
  second;
• 2:54
  at
• 2:54
  one
• 2:55
  million
• 2:55
  tokens
• 2:55
  of
• 2:55
  depth,
• 2:56
  that
• 2:56
  drops
• 2:56
  to
• 2:56
  49
• 2:57
  t/s
• 2:57
  for
• 2:57
  prefill
• 2:58
  and
• 2:58
  8
• 2:58
  t/s
• 2:59
  for
• 2:59
  generation
• 2:59
  —
• 3:00
  a
• 3:00
  steep
• 3:00
  but
• 3:00
  non-catastrophic
• 3:01
  degradation
• 3:02
  curve
• 3:02
  enabled
• 3:02
  by
• 3:02
  Nemotron's
• 3:03
  hybrid
• 3:03
  Mamba-2
• 3:03
  architecture,
• 3:04
  which
• 3:05
  makes
• 3:05
  memory
• 3:05
  scaling
• 3:05
  sublinear
• 3:06
  compared
• 3:06
  to
• 3:06
  pure-attention
• 3:07
  models.
• 3:08
  The
• 3:08
  same
• 3:08
  user
• 3:08
  benchmarked
• 3:09
  Qwen
• 3:09
  3.5
• 3:10
  122B
• 3:10
  (pure
• 3:11
  MoE
• 3:11
  transformer)
• 3:12
  under
• 3:13
  identical
• 3:13
  conditions
• 3:13
  and
• 3:14
  found
• 3:14
  it
• 3:14
  starts
• 3:14
  faster
• 3:14
  at
• 3:15
  391
• 3:15
  t/s
• 3:16
  but
• 3:16
  degrades
• 3:17
  more
• 3:17
  aggressively,
• 3:18
  hitting
• 3:18
  59
• 3:18
  t/s
• 3:19
  at
• 3:19
  250K
• 3:20
  depth.
• 3:21
  The
• 3:21
  Mamba-2
• 3:22
  hybrid
• 3:22
  wins
• 3:22
  the
• 3:22
  long-context
• 3:23
  race
• 3:23
  decisively.
• 3:24
  This
• 3:25
  long-context
• 3:25
  capability
• 3:26
  reframes
• 3:27
  the
• 3:27
  RAG-vs-stuff-the-window
• 3:28
  debate.
• 3:29
  If
• 3:29
  the
• 3:29
  entire
• 3:29
  Lord
• 3:30
  of
• 3:30
  the
• 3:30
  Rings
• 3:30
  plus
• 3:31
  The
• 3:31
  Hobbit
• 3:31
  fits
• 3:31
  in
• 3:31
  the
• 3:32
  prompt
• 3:32
  with
• 3:32
  room
• 3:32
  to
• 3:32
  spare,
• 3:33
  the
• 3:33
  engineering
• 3:34
  overhead
• 3:34
  of
• 3:34
  chunking
• 3:35
  strategies,
• 3:36
  embedding
• 3:36
  models,
• 3:37
  vector
• 3:37
  databases,
• 3:38
  and
• 3:38
  rerankers
• 3:39
  starts
• 3:39
  to
• 3:39
  look
• 3:39
  like
• 3:39
  unnecessary
• 3:40
  complexity
• 3:41
  for
• 3:41
  bounded
• 3:41
  datasets.
• 3:42
  The
• 3:42
  counterarguments
• 3:43
  remain
• 3:43
  valid
• 3:43
  —
• 3:44
  the
• 3:44
  "rereading
• 3:45
  tax"
• 3:45
  of
• 3:45
  processing
• 3:46
  a
• 3:46
  250K-token
• 3:47
  manual
• 3:47
  on
• 3:47
  every
• 3:48
  query,
• 3:48
  attention
• 3:49
  dilution
• 3:49
  burying
• 3:50
  needles
• 3:50
  in
• 3:50
  haystacks,
• 3:51
  and
• 3:51
  the
• 3:51
  simple
• 3:51
  fact
• 3:52
  that
• 3:52
  enterprise
• 3:52
  data
• 3:52
  lakes
• 3:53
  measure
• 3:53
  in
• 3:53
  terabytes,
• 3:54
  not
• 3:54
  megabytes.
• 3:55
  For
• 3:55
  bounded,
• 3:55
  high-reasoning
• 3:56
  tasks
• 3:56
  (legal
• 3:57
  contract
• 3:57
  analysis,
• 3:58
  codebase
• 3:58
  auditing),
• 3:59
  long
• 3:59
  context
• 4:00
  wins
• 4:00
  on
• 4:00
  simplicity.
• 4:01
  For
• 4:01
  infinite
• 4:02
  datasets,
• 4:03
  the
• 4:03
  vector
• 4:03
  database
• 4:03
  stays.
• 4:04
  The
• 4:04
  practical
• 4:05
  answer,
• 4:05
  as
• 4:05
  usual,
• 4:06
  is
• 4:06
  routing
• 4:07
  between
• 4:07
  both.
• 4:08
  On
• 4:08
  the
• 4:08
  edge-inference
• 4:08
  front,
• 4:09
  Embedl
• 4:09
  released
• 4:10
  FlashHead,
• 4:11
  a
• 4:11
  drop-in
• 4:11
  replacement
• 4:12
  for
• 4:12
  the
• 4:12
  language-model
• 4:12
  head
• 4:13
  that
• 4:13
  delivers
• 4:13
  up
• 4:14
  to
• 4:14
  40%
• 4:14
  faster
• 4:15
  token
• 4:15
  generation
• 4:15
  on
• 4:16
  NVIDIA
• 4:16
  Jetson
• 4:16
  devices
• 4:17
  without
• 4:17
  sacrificing
• 4:18
  reasoning
• 4:18
  quality,
• 4:19
  stacking
• 4:19
  on
• 4:19
  top
• 4:20
  of
• 4:20
  W4A16
• 4:22
  quantization.
• 4:23
  On
• 4:23
  an
• 4:24
  AGX
• 4:24
  Thor,
• 4:25
  the
• 4:25
  Cosmos-Reason2-2B
• 4:26
  model
• 4:26
  jumps
• 4:27
  from
• 4:27
  88.3
• 4:28
  t/s
• 4:29
  (W4A16)
• 4:30
  to
• 4:30
  128.2
• 4:32
  t/s
• 4:33
  with
• 4:33
  FlashHead
• 4:33
  enabled.
• 4:34
  Architecture
• 4:34
  choices
• 4:35
  matter
• 4:35
  beyond
• 4:35
  speed.
• 4:37
  A
• 4:37
  careful
• 4:37
  benchmark
• 4:38
  of
• 4:38
  DeepSeek
• 4:38
  V3.2
• 4:40
  Speciale
• 4:40
  shows
• 4:41
  that
• 4:41
  running
• 4:41
  it
• 4:41
  with
• 4:41
  dense
• 4:42
  attention
• 4:42
  (as
• 4:43
  llama.cpp
• 4:44
  currently
• 4:44
  does,
• 4:45
  lacking
• 4:45
  a
• 4:45
  native
• 4:45
  sparse-attention
• 4:46
  implementation)
• 4:47
  produces
• 4:47
  a
• 4:47
  17%
• 4:48
  accuracy
• 4:48
  decrease
• 4:49
  on
• 4:49
  lineage-512
• 4:50
  tasks
• 4:51
  and
• 4:51
  22%
• 4:52
  on
• 4:52
  lineage-1024
• 4:53
  versus
• 4:54
  the
• 4:54
  intended
• 4:54
  sparse
• 4:55
  (DSA)
• 4:55
  attention
• 4:55
  path.
• 4:57
  The
• 4:57
  model's
• 4:57
  sparse
• 4:57
  attention
• 4:58
  selects
• 4:58
  only
• 4:58
  the
• 4:59
  top
• 4:59
  2,048
• 5:00
  tokens
• 5:00
  per
• 5:00
  head;
• 5:01
  forcing
• 5:01
  it
• 5:02
  to
• 5:02
  attend
• 5:02
  to
• 5:02
  the
• 5:02
  full
• 5:02
  context
• 5:03
  overwhelms
• 5:03
  it
• 5:03
  with
• 5:04
  noise
• 5:04
  it
• 5:04
  was
• 5:04
  never
• 5:05
  trained
• 5:05
  to
• 5:05
  handle.
• 5:06
  Until
• 5:07
  llama.cpp
• 5:08
  adds
• 5:09
  proper
• 5:09
  DSA
• 5:09
  support,
• 5:10
  running
• 5:10
  DeepSeek
• 5:11
  V3.2
• 5:12
  locally
• 5:13
  means
• 5:13
  accepting
• 5:13
  measurably
• 5:14
  degraded
• 5:14
  reasoning
• 5:15
  on
• 5:15
  complex
• 5:15
  tasks.
• 5:17
  A
• 5:17
  companion
• 5:17
  paper
• 5:17
  on
• 5:18
  activation
• 5:18
  outliers
• 5:19
  in
• 5:19
  transformer
• 5:19
  quantization
• 5:20
  provides
• 5:21
  the
• 5:21
  theoretical
• 5:21
  explanation
• 5:22
  for
• 5:22
  why
• 5:22
  naive
• 5:23
  INT8
• 5:23
  quantization
• 5:24
  crushes
• 5:24
  model
• 5:25
  quality.
• 5:26
  The
• 5:26
  researchers
• 5:27
  show
• 5:27
  that
• 5:27
  structured
• 5:27
  activation
• 5:28
  outliers
• 5:28
  —
• 5:29
  not
• 5:29
  random
• 5:29
  noise
• 5:29
  —
• 5:30
  drive
• 5:30
  PTQ
• 5:31
  failure.
• 5:32
  In
• 5:32
  BERT-base
• 5:32
  fine-tuned
• 5:33
  on
• 5:33
  QNLI,
• 5:34
  global
• 5:34
  W8A8
• 5:35
  quantization
• 5:36
  drops
• 5:36
  validation
• 5:36
  accuracy
• 5:37
  from
• 5:37
  89.66%
• 5:38
  to
• 5:39
  54.33%.
• 5:41
  The
• 5:41
  culprit:
• 5:41
  kurtosis
• 5:42
  in
• 5:42
  activation
• 5:43
  distributions
• 5:43
  reaches
• 5:44
  271
• 5:45
  by
• 5:45
  Layer
• 5:45
  11
• 5:45
  (a
• 5:45
  Gaussian
• 5:46
  has
• 5:46
  kurtosis
• 5:46
  of
• 5:46
  3),
• 5:47
  and
• 5:47
  the
• 5:47
  top
• 5:47
  1%
• 5:48
  of
• 5:48
  channels
• 5:48
  concentrate
• 5:49
  55%
• 5:50
  of
• 5:50
  total
• 5:50
  activation
• 5:51
  energy.
• 5:52
  Uniform
• 5:52
  min-max
• 5:52
  scaling
• 5:53
  allocates
• 5:53
  nearly
• 5:54
  all
• 5:54
  dynamic
• 5:54
  range
• 5:55
  to
• 5:55
  these
• 5:55
  few
• 5:55
  channels,
• 5:56
  crushing
• 5:56
  everything
• 5:57
  else.
• 5:57
  Mixed-precision
• 5:58
  (keeping
• 5:58
  critical
• 5:59
  layers
• 5:59
  in
• 5:59
  FP16)
• 6:00
  recovers
• 6:00
  to
• 6:00
  89.42%,
• 6:02
  while
• 6:02
  per-embedding-group
• 6:03
  quantization
• 6:04
  only
• 6:04
  partially
• 6:04
  recovers.
• 6:05
  Surprisingly,
• 6:06
  percentile-based
• 6:07
  calibration
• 6:08
  makes
• 6:08
  things
• 6:08
  worse
• 6:08
  —
• 6:09
  the
• 6:09
  dominant
• 6:09
  channels
• 6:10
  encode
• 6:10
  signal,
• 6:11
  not
• 6:11
  noise.
• 6:12
  The
• 6:12
  practical
• 6:12
  takeaway:
• 6:13
  channel-aware
• 6:14
  precision
• 6:14
  allocation
• 6:15
  is
• 6:15
  necessary,
• 6:16
  and
• 6:16
  simply
• 6:16
  clipping
• 6:17
  outliers
• 6:17
  destroys
• 6:17
  information.
• 6:19
  On
• 6:19
  an
• 6:19
  RTX
• 6:20
  3050,
• 6:21
  none
• 6:21
  of
• 6:21
  the
• 6:21
  INT8
• 6:22
  methods
• 6:22
  produced
• 6:23
  measurable
• 6:23
  latency
• 6:24
  improvements,
• 6:25
  a
• 6:25
  reminder
• 6:25
  that
• 6:25
  quantization
• 6:26
  benefits
• 6:26
  depend
• 6:27
  entirely
• 6:27
  on
• 6:27
  hardware
• 6:28
  kernel
• 6:28
  support.
• 6:32
  A
• 6:32
  systematic
• 6:33
  comparison
• 6:33
  from
• 6:33
  Distil
• 6:34
  Labs
• 6:34
  pits
• 6:34
  fine-tuned
• 6:35
  Qwen3
• 6:35
  models
• 6:36
  (0.6B
• 6:37
  to
• 6:37
  8B
• 6:37
  parameters)
• 6:38
  against
• 6:38
  frontier
• 6:38
  APIs
• 6:39
  —
• 6:40
  GPT-5
• 6:41
  nano/mini/5.2,
• 6:42
  Gemini
• 6:43
  Flash,
• 6:43
  Claude
• 6:43
  Haiku/Sonnet/Opus,
• 6:45
  Grok
• 6:45
  4
• 6:45
  —
• 6:46
  across
• 6:47
  nine
• 6:47
  datasets
• 6:47
  spanning
• 6:48
  classification,
• 6:48
  function
• 6:49
  calling,
• 6:49
  QA,
• 6:49
  and
• 6:50
  open-book
• 6:50
  QA.
• 6:51
  The
• 6:51
  headline:
• 6:52
  a
• 6:52
  0.6B
• 6:53
  model
• 6:53
  hits
• 6:54
  98.7%
• 6:55
  on
• 6:55
  smart-home
• 6:55
  function
• 6:56
  calling
• 6:56
  versus
• 6:56
  Gemini
• 6:56
  Flash
• 6:57
  at
• 6:57
  92.0%.
• 6:59
  On
• 6:59
  Text2SQL,
• 7:00
  Qwen3-4B
• 7:01
  reaches
• 7:01
  98.0%
• 7:02
  versus
• 7:03
  Claude
• 7:03
  Haiku
• 7:03
  at
• 7:03
  98.7%,
• 7:05
  but
• 7:05
  at
• 7:05
  $3
• 7:06
  per
• 7:06
  million
• 7:06
  requests
• 7:06
  versus
• 7:07
  $378.
• 7:09
  Classification
• 7:09
  tasks
• 7:10
  (Banking77,
• 7:11
  TREC)
• 7:12
  are
• 7:12
  "basically
• 7:12
  solved"
• 7:13
  —
• 7:13
  distilled
• 7:14
  models
• 7:14
  land
• 7:14
  within
• 7:14
  0–1.5
• 7:16
  percentage
• 7:16
  points
• 7:16
  of
• 7:16
  the
• 7:17
  best
• 7:17
  frontier
• 7:17
  option.
• 7:18
  Where
• 7:18
  frontier
• 7:18
  still
• 7:18
  wins
• 7:19
  clearly
• 7:19
  is
• 7:19
  open-ended
• 7:20
  reasoning
• 7:20
  requiring
• 7:20
  world
• 7:21
  knowledge:
• 7:21
  HotpotQA
• 7:22
  shows
• 7:22
  92.0%
• 7:24
  versus
• 7:24
  Haiku's
• 7:24
  98.0%.
• 7:26
  The
• 7:26
  throughput
• 7:26
  numbers
• 7:27
  are
• 7:27
  equally
• 7:27
  striking:
• 7:28
  Qwen3-4B
• 7:29
  on
• 7:29
  a
• 7:29
  single
• 7:29
  H100
• 7:30
  sustains
• 7:30
  222
• 7:31
  requests
• 7:32
  per
• 7:32
  second
• 7:32
  for
• 7:32
  Text2SQL
• 7:33
  with
• 7:34
  p50
• 7:34
  latency
• 7:35
  of
• 7:35
  390ms
• 7:36
  and
• 7:36
  only
• 7:36
  7.6
• 7:37
  GiB
• 7:37
  VRAM
• 7:38
  (BF16,
• 7:39
  no
• 7:39
  quantization).
• 7:40
  FP8
• 7:41
  adds
• 7:41
  15%
• 7:42
  throughput
• 7:42
  and
• 7:42
  cuts
• 7:43
  VRAM
• 7:43
  by
• 7:43
  44%
• 7:44
  with
• 7:44
  no
• 7:44
  accuracy
• 7:44
  loss.
• 7:45
  The
• 7:45
  practical
• 7:46
  decision
• 7:46
  framework:
• 7:47
  distill
• 7:47
  when
• 7:47
  you
• 7:48
  have
• 7:48
  structured
• 7:48
  tasks,
• 7:48
  well-defined
• 7:49
  schemas,
• 7:50
  high
• 7:50
  volume,
• 7:50
  or
• 7:50
  data-sovereignty
• 7:51
  needs;
• 7:52
  call
• 7:52
  frontier
• 7:52
  APIs
• 7:53
  for
• 7:53
  broad
• 7:53
  world
• 7:54
  knowledge
• 7:54
  or
• 7:54
  low
• 7:55
  volume.
• 7:56
  The
• 7:56
  cost
• 7:56
  differential
• 7:57
  is
• 7:57
  100x
• 7:58
  or
• 7:58
  more
• 7:58
  at
• 7:58
  scale.
• 7:59
  On
• 7:59
  the
• 7:59
  research
• 8:00
  side,
• 8:00
  EvoKernel
• 8:01
  from
• 8:01
  Shanghai
• 8:02
  Jiao
• 8:02
  Tong
• 8:02
  University
• 8:03
  tackles
• 8:03
  a
• 8:03
  harder
• 8:04
  version
• 8:04
  of
• 8:04
  this
• 8:04
  problem:
• 8:05
  getting
• 8:05
  LLMs
• 8:06
  to
• 8:06
  generate
• 8:06
  kernel
• 8:07
  code
• 8:07
  for
• 8:07
  hardware
• 8:08
  they
• 8:08
  have
• 8:08
  essentially
• 8:09
  zero
• 8:09
  training
• 8:09
  data
• 8:09
  for.
• 8:10
  The
• 8:10
  "data
• 8:11
  wall"
• 8:11
  is
• 8:11
  stark
• 8:12
  —
• 8:13
  GPT-5.2
• 8:14
  achieves
• 8:14
  92%
• 8:15
  correctness
• 8:16
  on
• 8:16
  CUDA
• 8:16
  L1
• 8:17
  tasks
• 8:17
  but
• 8:17
  collapses
• 8:18
  to
• 8:18
  14%
• 8:19
  on
• 8:19
  Ascend
• 8:19
  C
• 8:20
  (NPU)
• 8:20
  tasks.
• 8:21
  EvoKernel
• 8:22
  formulates
• 8:22
  kernel
• 8:22
  synthesis
• 8:23
  as
• 8:23
  a
• 8:23
  reinforcement-learning
• 8:24
  task
• 8:24
  over
• 8:25
  self-evolving
• 8:25
  memory,
• 8:26
  using
• 8:26
  value-driven
• 8:27
  retrieval
• 8:27
  (learned
• 8:28
  Q-values)
• 8:28
  to
• 8:28
  select
• 8:28
  which
• 8:29
  past
• 8:29
  experiences
• 8:30
  to
• 8:30
  feed
• 8:30
  the
• 8:30
  generator.
• 8:31
  Starting
• 8:31
  from
• 8:32
  11.0%
• 8:33
  correctness,
• 8:34
  EvoKernel
• 8:34
  pushes
• 8:35
  GPT-5.2
• 8:36
  to
• 8:36
  83.0%
• 8:37
  on
• 8:37
  NPU
• 8:38
  benchmarks,
• 8:39
  with
• 8:39
  a
• 8:39
  median
• 8:40
  3.60x
• 8:41
  latency
• 8:41
  speedup
• 8:42
  over
• 8:42
  initial
• 8:42
  drafts
• 8:43
  through
• 8:43
  iterative
• 8:43
  refinement.
• 8:44
  The
• 8:44
  cross-task
• 8:45
  memory
• 8:45
  transfer
• 8:46
  is
• 8:46
  the
• 8:46
  most
• 8:46
  interesting
• 8:47
  finding:
• 8:48
  experiences
• 8:49
  from
• 8:49
  simpler
• 8:49
  operators
• 8:50
  genuinely
• 8:50
  accelerate
• 8:51
  learning
• 8:51
  on
• 8:51
  complex
• 8:52
  ones,
• 8:52
  suggesting
• 8:53
  that
• 8:53
  value-guided
• 8:54
  experience
• 8:54
  accumulation
• 8:55
  can
• 8:55
  substitute
• 8:56
  for
• 8:56
  training
• 8:56
  data
• 8:56
  in
• 8:57
  cold-start
• 8:57
  domains.
• 9:02
  Stripe
• 9:02
  is
• 9:02
  now
• 9:02
  shipping
• 9:03
  over
• 9:03
  1,300
• 9:04
  pull
• 9:04
  requests
• 9:04
  per
• 9:05
  week
• 9:05
  that
• 9:05
  are
• 9:05
  entirely
• 9:05
  AI-written
• 9:06
  —
• 9:06
  human-reviewed,
• 9:07
  but
• 9:07
  with
• 9:07
  zero
• 9:08
  human-authored
• 9:08
  lines.
• 9:09
  The
• 9:09
  enabling
• 9:10
  architecture
• 9:10
  is
• 9:10
  not
• 9:11
  a
• 9:11
  more
• 9:11
  powerful
• 9:11
  model
• 9:11
  but
• 9:12
  a
• 9:12
  more
• 9:12
  disciplined
• 9:12
  harness.
• 9:13
  Stripe's
• 9:13
  internal
• 9:14
  system,
• 9:14
  Minions,
• 9:15
  implements
• 9:16
  "blueprints"
• 9:16
  —
• 9:17
  workflows
• 9:17
  that
• 9:17
  interleave
• 9:18
  agentic
• 9:18
  nodes
• 9:18
  (where
• 9:18
  an
• 9:19
  LLM
• 9:19
  reasons
• 9:19
  and
• 9:19
  codes)
• 9:20
  with
• 9:20
  deterministic
• 9:20
  nodes
• 9:21
  (where
• 9:21
  linting,
• 9:22
  type-checking,
• 9:22
  and
• 9:23
  test
• 9:23
  execution
• 9:23
  happen
• 9:24
  in
• 9:24
  guaranteed
• 9:24
  fashion).
• 9:25
  The
• 9:25
  system
• 9:26
  curates
• 9:26
  context
• 9:26
  before
• 9:27
  the
• 9:27
  agent
• 9:27
  ever
• 9:27
  sees
• 9:27
  a
• 9:27
  task,
• 9:28
  pulling
• 9:28
  from
• 9:29
  a
• 9:29
  500-tool
• 9:29
  MCP
• 9:30
  server
• 9:30
  called
• 9:30
  "tool
• 9:30
  shed"
• 9:31
  and
• 9:31
  selecting
• 9:32
  a
• 9:32
  relevant
• 9:32
  subset.
• 9:33
  Each
• 9:33
  run
• 9:33
  executes
• 9:34
  in
• 9:34
  an
• 9:34
  isolated
• 9:34
  AWS
• 9:35
  EC2
• 9:35
  instance,
• 9:36
  pre-loaded
• 9:36
  with
• 9:37
  Stripe's
• 9:37
  codebase
• 9:37
  and
• 9:38
  lint
• 9:38
  caches.
• 9:38
  The
• 9:39
  test
• 9:39
  validation
• 9:39
  loop
• 9:39
  pulls
• 9:40
  from
• 9:40
  Stripe's
• 9:40
  three-million-test
• 9:41
  CI
• 9:41
  suite
• 9:41
  and
• 9:41
  gives
• 9:42
  the
• 9:42
  agent
• 9:42
  a
• 9:42
  maximum
• 9:43
  of
• 9:43
  two
• 9:43
  retry
• 9:43
  cycles
• 9:43
  before
• 9:44
  escalating
• 9:44
  to
• 9:44
  a
• 9:44
  human.
• 9:45
  Shopify
• 9:45
  has
• 9:45
  open-sourced
• 9:46
  a
• 9:46
  similar
• 9:46
  system
• 9:47
  called
• 9:47
  Roast.
• 9:47
  The
• 9:47
  pattern
• 9:48
  is
• 9:48
  clear:
• 9:48
  the
• 9:48
  industry
• 9:49
  is
• 9:49
  moving
• 9:49
  toward
• 9:49
  systems
• 9:50
  that
• 9:50
  control
• 9:50
  agents,
• 9:51
  not
• 9:51
  agents
• 9:51
  that
• 9:51
  control
• 9:52
  systems.
• 9:52
  The
• 9:52
  developer
• 9:53
  tooling
• 9:53
  ecosystem
• 9:54
  around
• 9:54
  Claude
• 9:54
  Code
• 9:54
  has
• 9:54
  matured
• 9:55
  to
• 9:55
  the
• 9:55
  point
• 9:55
  where
• 9:55
  a
• 9:55
  top-10
• 9:56
  list
• 9:56
  is
• 9:56
  genuinely
• 9:57
  useful.
• 9:57
  Key
• 9:58
  recommendations
• 9:58
  include
• 9:58
  the
• 9:58
  Supabase
• 9:59
  CLI
• 9:59
  (not
• 9:59
  the
• 10:00
  MCP
• 10:00
  server
• 10:00
  —
• 10:01
  CLIs
• 10:01
  outperform
• 10:01
  MCPs
• 10:02
  because
• 10:02
  they
• 10:02
  are
• 10:02
  purpose-built
• 10:03
  for
• 10:03
  terminal-native
• 10:03
  agents),
• 10:04
  the
• 10:04
  Anthropic
• 10:04
  skill-creator
• 10:05
  skill
• 10:05
  (which
• 10:05
  now
• 10:05
  supports
• 10:06
  A/B
• 10:07
  testing
• 10:07
  of
• 10:07
  skill
• 10:07
  performance),
• 10:08
  the
• 10:08
  GSD
• 10:08
  framework
• 10:09
  for
• 10:09
  spec-driven
• 10:09
  project
• 10:10
  scaffolding,
• 10:10
  the
• 10:11
  NotebookLM
• 10:11
  CLI
• 10:11
  for
• 10:12
  offloading
• 10:12
  research
• 10:12
  to
• 10:12
  Google's
• 10:13
  free
• 10:13
  tool,
• 10:13
  and
• 10:13
  the
• 10:13
  Playwright
• 10:14
  CLI
• 10:14
  for
• 10:14
  browser
• 10:14
  automation.
• 10:15
  The
• 10:15
  throughline:
• 10:16
  CLI
• 10:16
  tools
• 10:17
  plus
• 10:17
  purpose-built
• 10:17
  skills
• 10:18
  consistently
• 10:18
  outperform
• 10:19
  MCP-based
• 10:19
  equivalents
• 10:20
  because
• 10:20
  they
• 10:20
  avoid
• 10:20
  the
• 10:21
  overhead
• 10:21
  of
• 10:21
  the
• 10:21
  MCP
• 10:21
  protocol
• 10:22
  layer.
• 10:23
  The
• 10:23
  memory
• 10:23
  problem
• 10:23
  for
• 10:24
  AI
• 10:24
  agents
• 10:24
  gets
• 10:24
  a
• 10:24
  dedicated
• 10:25
  solution
• 10:25
  in
• 10:25
  SAGE,
• 10:26
  an
• 10:26
  open-source
• 10:26
  (Apache
• 10:27
  2.0)
• 10:27
  desktop
• 10:28
  application
• 10:28
  that
• 10:28
  gives
• 10:29
  any
• 10:29
  MCP-compatible
• 10:30
  AI
• 10:30
  persistent,
• 10:30
  encrypted,
• 10:31
  locally-stored
• 10:32
  memory.
• 10:33
  Every
• 10:33
  memory
• 10:33
  passes
• 10:34
  through
• 10:34
  four
• 10:34
  validators
• 10:34
  (Sentinel,
• 10:35
  Dedup,
• 10:36
  Quality,
• 10:36
  Consistency)
• 10:37
  with
• 10:37
  a
• 10:37
  BFT
• 10:37
  quorum
• 10:38
  (3/4)
• 10:38
  required
• 10:39
  to
• 10:39
  commit
• 10:39
  —
• 10:40
  real
• 10:40
  signed
• 10:40
  vote
• 10:41
  transactions
• 10:41
  through
• 10:41
  CometBFT
• 10:42
  consensus.
• 10:43
  Memories
• 10:44
  strengthen
• 10:44
  with
• 10:44
  corroboration
• 10:45
  and
• 10:45
  decay
• 10:45
  without
• 10:45
  reinforcement.
• 10:46
  The
• 10:47
  team
• 10:47
  cites
• 10:47
  published
• 10:47
  research
• 10:48
  showing
• 10:48
  agents
• 10:48
  with
• 10:49
  governed
• 10:49
  memory
• 10:49
  measurably
• 10:50
  improve
• 10:50
  over
• 10:50
  time,
• 10:51
  while
• 10:51
  memoryless
• 10:52
  agents
• 10:52
  show
• 10:52
  zero
• 10:53
  improvement
• 10:53
  regardless
• 10:54
  of
• 10:54
  session
• 10:54
  count.
• 10:55
  SAGE
• 10:55
  supports
• 10:56
  Claude,
• 10:56
  ChatGPT,
• 10:57
  DeepSeek,
• 10:58
  Gemini,
• 10:59
  Cursor,
• 10:59
  and
• 11:00
  Ollama,
• 11:00
  and
• 11:01
  includes
• 11:01
  a
• 11:01
  Chrome
• 11:01
  extension
• 11:02
  for
• 11:02
  free-tier
• 11:02
  ChatGPT
• 11:03
  users.
• 11:04
  At
• 11:04
  the
• 11:04
  infrastructure
• 11:05
  layer,
• 11:06
  Common
• 11:06
  Ground
• 11:06
  Core
• 11:07
  (CGC)
• 11:07
  goes
• 11:07
  further,
• 11:09
  proposing
• 11:09
  a
• 11:09
  "sociotechnical
• 11:10
  operating
• 11:11
  system"
• 11:11
  for
• 11:11
  multi-agent
• 11:12
  coordination
• 11:13
  based
• 11:13
  on
• 11:13
  cybernetics
• 11:14
  principles.
• 11:15
  State
• 11:15
  lives
• 11:15
  in
• 11:16
  Postgres
• 11:16
  with
• 11:16
  CAS
• 11:17
  locks;
• 11:18
  NATS
• 11:18
  JetStream
• 11:18
  serves
• 11:19
  as
• 11:19
  a
• 11:19
  pure
• 11:19
  wakeup
• 11:19
  doorbell;
• 11:20
  an
• 11:21
  immutable
• 11:21
  CardBox
• 11:22
  creates
• 11:22
  unforgeable
• 11:23
  cognitive
• 11:23
  lineage
• 11:23
  for
• 11:24
  every
• 11:24
  tool
• 11:24
  call
• 11:25
  and
• 11:25
  reasoning
• 11:25
  step.
• 11:26
  Agents
• 11:27
  can
• 11:27
  fork
• 11:27
  and
• 11:27
  join
• 11:28
  hundreds
• 11:28
  of
• 11:28
  child
• 11:28
  nodes
• 11:29
  through
• 11:29
  the
• 11:29
  L1
• 11:29
  kernel
• 11:30
  while
• 11:30
  the
• 11:30
  system
• 11:30
  handles
• 11:31
  concurrent
• 11:31
  convergence.
• 11:33
  The
• 11:33
  key
• 11:33
  design
• 11:33
  decision:
• 11:34
  humans
• 11:35
  participate
• 11:35
  as
• 11:36
  asynchronous
• 11:36
  nodes
• 11:36
  under
• 11:37
  the
• 11:37
  same
• 11:37
  protocol
• 11:38
  as
• 11:38
  AI
• 11:38
  agents,
• 11:39
  not
• 11:39
  as
• 11:39
  top-level
• 11:40
  prompters.
• 11:41
  The
• 11:41
  project
• 11:42
  is
• 11:42
  in
• 11:42
  preview
• 11:42
  with
• 11:42
  no
• 11:43
  ACL
• 11:44
  and
• 11:44
  an
• 11:44
  RCE
• 11:44
  warning,
• 11:45
  but
• 11:45
  the
• 11:45
  architecture
• 11:46
  —
• 11:47
  protocol-first,
• 11:48
  worker-agnostic,
• 11:50
  zero-brain-split
• 11:50
  —
• 11:51
  represents
• 11:52
  a
• 11:52
  thoughtful
• 11:52
  approach
• 11:53
  to
• 11:53
  the
• 11:53
  coordination-collapse
• 11:54
  problem
• 11:54
  that
• 11:55
  plagues
• 11:55
  naive
• 11:55
  multi-agent
• 11:56
  setups.
• 12:01
  Atlassian
• 12:01
  is
• 12:01
  cutting
• 12:02
  1,600
• 12:02
  jobs
• 12:03
  —
• 12:03
  roughly
• 12:03
  10%
• 12:04
  of
• 12:04
  its
• 12:04
  workforce
• 12:04
  —
• 12:05
  to
• 12:05
  "self-fund
• 12:05
  further
• 12:06
  investment
• 12:06
  in
• 12:06
  AI
• 12:07
  and
• 12:07
  enterprise
• 12:07
  sales."
• 12:08
  CEO
• 12:09
  Mike
• 12:09
  Cannon-Brookes
• 12:09
  insists
• 12:10
  Atlassian
• 12:10
  does
• 12:11
  not
• 12:11
  follow
• 12:11
  the
• 12:11
  philosophy
• 12:12
  of
• 12:12
  replacing
• 12:12
  people
• 12:13
  with
• 12:13
  AI,
• 12:13
  then
• 12:13
  immediately
• 12:14
  adds:
• 12:15
  "It
• 12:15
  would
• 12:15
  be
• 12:15
  disingenuous
• 12:15
  to
• 12:16
  pretend
• 12:16
  AI
• 12:16
  doesn't
• 12:16
  change
• 12:17
  the
• 12:17
  mix
• 12:17
  of
• 12:17
  skills
• 12:17
  we
• 12:17
  need
• 12:18
  or
• 12:18
  the
• 12:18
  number
• 12:18
  of
• 12:18
  roles
• 12:19
  required
• 12:19
  in
• 12:19
  certain
• 12:19
  areas."
• 12:20
  The
• 12:20
  severance
• 12:20
  bill:
• 12:21
  $225–236
• 12:23
  million.
• 12:24
  About
• 12:24
  900
• 12:25
  of
• 12:25
  the
• 12:25
  cuts
• 12:25
  target
• 12:25
  developers
• 12:26
  and
• 12:26
  software
• 12:26
  roles
• 12:26
  across
• 12:27
  the
• 12:27
  US,
• 12:27
  Australia,
• 12:28
  and
• 12:28
  India.
• 12:28
  CTO
• 12:29
  Rajeev
• 12:29
  Rajan
• 12:30
  is
• 12:30
  stepping
• 12:30
  down.
• 12:31
  Atlassian's
• 12:31
  stock
• 12:31
  has
• 12:32
  fallen
• 12:32
  over
• 12:32
  50%
• 12:33
  since
• 12:33
  January,
• 12:33
  part
• 12:34
  of
• 12:34
  a
• 12:34
  broader
• 12:34
  SaaS
• 12:34
  selloff
• 12:35
  driven
• 12:35
  by
• 12:35
  the
• 12:35
  market
• 12:35
  narrative
• 12:36
  that
• 12:36
  AI
• 12:36
  threatens
• 12:36
  per-seat
• 12:37
  licensing
• 12:37
  models.
• 12:37
  Whether
• 12:38
  AI
• 12:38
  is
• 12:38
  the
• 12:38
  cause
• 12:39
  or
• 12:39
  the
• 12:39
  cover,
• 12:39
  the
• 12:40
  result
• 12:40
  is
• 12:40
  the
• 12:40
  same
• 12:40
  for
• 12:40
  the
• 12:40
  people
• 12:41
  being
• 12:41
  walked
• 12:41
  out.
• 12:42
  The
• 12:42
  emotional
• 12:42
  toll
• 12:42
  is
• 12:43
  not
• 12:43
  limited
• 12:43
  to
• 12:43
  those
• 12:43
  receiving
• 12:44
  severance
• 12:44
  packages.
• 12:45
  NetworkChuck,
• 12:45
  a
• 12:45
  tech
• 12:46
  YouTuber
• 12:46
  with
• 12:46
  five
• 12:46
  million
• 12:47
  subscribers,
• 12:47
  posted
• 12:48
  a
• 12:48
  remarkably
• 12:48
  candid
• 12:49
  video
• 12:49
  about
• 12:49
  nearly
• 12:49
  quitting
• 12:49
  his
• 12:50
  channel
• 12:50
  due
• 12:50
  to
• 12:50
  AI-induced
• 12:51
  burnout
• 12:51
  and
• 12:51
  anxiety.
• 12:52
  A
• 12:52
  UC
• 12:53
  Berkeley
• 12:53
  study
• 12:53
  he
• 12:53
  cites
• 12:54
  reports
• 12:54
  62%
• 12:55
  of
• 12:55
  AI
• 12:55
  workers
• 12:55
  experiencing
• 12:56
  burnout,
• 12:56
  anxiety,
• 12:57
  or
• 12:57
  decision
• 12:58
  paralysis
• 12:58
  by
• 12:58
  month
• 12:59
  six.
• 13:00
  The
• 13:00
  Matt
• 13:00
  Schumer
• 13:00
  article
• 13:00
  that
• 13:01
  went
• 13:01
  viral
• 13:01
  —
• 13:02
  "If
• 13:02
  your
• 13:02
  job
• 13:02
  happens
• 13:02
  on
• 13:03
  a
• 13:03
  screen,
• 13:03
  AI
• 13:03
  is
• 13:03
  coming
• 13:04
  for
• 13:04
  significant
• 13:04
  parts
• 13:05
  of
• 13:05
  it"
• 13:05
  —
• 13:05
  crystallized
• 13:06
  a
• 13:06
  fear
• 13:06
  many
• 13:06
  were
• 13:06
  already
• 13:07
  stuffing
• 13:07
  down.
• 13:08
  The
• 13:08
  rebuttal
• 13:08
  data
• 13:09
  is
• 13:09
  real
• 13:09
  too:
• 13:10
  95%
• 13:10
  of
• 13:10
  organizations
• 13:11
  see
• 13:11
  no
• 13:11
  measurable
• 13:12
  ROI
• 13:12
  from
• 13:12
  AI;
• 13:13
  the
• 13:13
  Yale
• 13:14
  Budget
• 13:14
  Lab
• 13:14
  says
• 13:15
  AI
• 13:15
  labor
• 13:15
  displacement
• 13:16
  "remains
• 13:16
  largely
• 13:17
  speculative";
• 13:18
  every
• 13:18
  previous
• 13:19
  tech
• 13:19
  revolution
• 13:19
  overestimated
• 13:20
  the
• 13:20
  speed
• 13:21
  of
• 13:21
  economic
• 13:21
  transformation.
• 13:22
  The
• 13:23
  resolution
• 13:23
  NetworkChuck
• 13:24
  arrives
• 13:24
  at
• 13:24
  is
• 13:24
  pragmatic:
• 13:25
  learning
• 13:26
  AI
• 13:26
  is
• 13:26
  still
• 13:26
  essential,
• 13:27
  but
• 13:27
  IT
• 13:27
  fundamentals
• 13:28
  (networking,
• 13:29
  security,
• 13:29
  systems)
• 13:30
  are
• 13:30
  what
• 13:30
  make
• 13:30
  you
• 13:30
  better
• 13:31
  with
• 13:31
  AI
• 13:31
  and
• 13:31
  valuable
• 13:32
  when
• 13:32
  AI
• 13:32
  fails.
• 13:33
  The
• 13:33
  anxiety
• 13:34
  is
• 13:34
  not
• 13:35
  speculative.
• 13:36
  The
• 13:36
  job
• 13:36
  losses
• 13:37
  are
• 13:37
  not
• 13:37
  speculative.
• 13:38
  The
• 13:38
  timeline
• 13:39
  for
• 13:39
  mass
• 13:39
  displacement
• 13:40
  remains
• 13:40
  genuinely
• 13:41
  unknown.
• 13:42
  Meanwhile,
• 13:42
  Debian's
• 13:43
  attempt
• 13:43
  to
• 13:43
  establish
• 13:44
  policy
• 13:44
  on
• 13:44
  AI-generated
• 13:45
  contributions
• 13:46
  ended
• 13:46
  in
• 13:46
  a
• 13:46
  deliberate
• 13:47
  non-decision.
• 13:48
  Lucas
• 13:48
  Nussbaum
• 13:49
  proposed
• 13:49
  a
• 13:49
  general
• 13:50
  resolution
• 13:51
  requiring
• 13:51
  disclosure
• 13:52
  of
• 13:52
  AI-generated
• 13:53
  code,
• 13:53
  accountability
• 13:54
  for
• 13:54
  technical
• 13:55
  merit
• 13:55
  and
• 13:55
  license
• 13:55
  compliance,
• 13:56
  and
• 13:56
  prohibition
• 13:57
  against
• 13:57
  using
• 13:58
  AI
• 13:58
  tools
• 13:58
  with
• 13:58
  embargoed
• 13:59
  security
• 13:59
  data.
• 14:00
  The
• 14:00
  ensuing
• 14:01
  debate
• 14:01
  was
• 14:01
  illuminating:
• 14:02
  Russ
• 14:02
  Allbery
• 14:03
  argued
• 14:03
  that
• 14:03
  "AI"
• 14:04
  is
• 14:04
  too
• 14:04
  amorphous
• 14:05
  for
• 14:05
  durable
• 14:05
  policy
• 14:05
  and
• 14:06
  urged
• 14:06
  specificity
• 14:07
  around
• 14:07
  LLMs;
• 14:08
  Simon
• 14:08
  Richter
• 14:09
  raised
• 14:09
  the
• 14:09
  onboarding
• 14:10
  problem
• 14:10
  —
• 14:10
  AI-assisted
• 14:11
  drive-by
• 14:12
  contributions
• 14:12
  are
• 14:13
  "a
• 14:13
  missed
• 14:13
  opportunity
• 14:14
  to
• 14:14
  onboard
• 14:14
  a
• 14:14
  new
• 14:14
  contributor";
• 14:15
  Matthew
• 14:16
  Vernon
• 14:16
  centered
• 14:17
  the
• 14:17
  ethical
• 14:17
  dimension
• 14:17
  of
• 14:18
  organizations
• 14:18
  that
• 14:18
  "hoover
• 14:19
  up
• 14:19
  content
• 14:19
  with
• 14:19
  scant
• 14:20
  regard
• 14:20
  to
• 14:20
  copyright";
• 14:21
  Ted
• 14:22
  Ts'o
• 14:22
  countered
• 14:22
  that
• 14:23
  gatekeeping
• 14:23
  AI-using
• 14:24
  contributors
• 14:24
  is
• 14:25
  even
• 14:25
  more
• 14:25
  self-defeating
• 14:26
  than
• 14:26
  the
• 14:26
  risks.
• 14:27
  Nussbaum
• 14:28
  ultimately
• 14:28
  withdrew
• 14:29
  the
• 14:29
  GR,
• 14:30
  noting
• 14:30
  the
• 14:30
  discussion
• 14:30
  had
• 14:31
  been
• 14:31
  civil
• 14:31
  enough
• 14:31
  that
• 14:31
  case-by-case
• 14:32
  handling
• 14:32
  under
• 14:33
  existing
• 14:33
  policies
• 14:34
  was
• 14:34
  sufficient
• 14:34
  for
• 14:34
  now.
• 14:35
  Given
• 14:36
  how
• 14:36
  fast
• 14:36
  the
• 14:36
  underlying
• 14:37
  technology
• 14:37
  is
• 14:37
  shifting,
• 14:38
  kicking
• 14:39
  the
• 14:39
  can
• 14:39
  may
• 14:39
  be
• 14:39
  the
• 14:39
  least
• 14:40
  disruptive
• 14:40
  option.
• 14:43
  Photoroom's
• 14:44
  PRX
• 14:44
  team
• 14:45
  trained
• 14:45
  a
• 14:45
  competitive
• 14:46
  text-to-image
• 14:46
  model
• 14:47
  in
• 14:47
  24
• 14:47
  hours
• 14:47
  on
• 14:48
  32
• 14:48
  H200
• 14:49
  GPUs
• 14:50
  for
• 14:50
  approximately
• 14:50
  $1,500
• 14:51
  —
• 14:52
  a
• 14:52
  budget
• 14:52
  that
• 14:52
  would
• 14:53
  have
• 14:53
  been
• 14:53
  absurd
• 14:53
  two
• 14:53
  years
• 14:54
  ago.
• 14:54
  The
• 14:54
  recipe
• 14:55
  stacks
• 14:55
  several
• 14:55
  recent
• 14:56
  innovations:
• 14:57
  x-prediction
• 14:57
  formulation
• 14:58
  that
• 14:58
  eliminates
• 14:58
  the
• 14:58
  VAE
• 14:59
  entirely
• 14:59
  and
• 14:59
  trains
• 15:00
  directly
• 15:00
  in
• 15:00
  pixel
• 15:01
  space
• 15:01
  (patch
• 15:01
  size
• 15:02
  32,
• 15:02
  256-dim
• 15:03
  bottleneck),
• 15:04
  TREAD
• 15:05
  token
• 15:05
  routing
• 15:05
  that
• 15:06
  randomly
• 15:06
  skips
• 15:06
  tokens
• 15:07
  through
• 15:07
  contiguous
• 15:08
  transformer
• 15:08
  blocks,
• 15:09
  REPA
• 15:09
  representation
• 15:10
  alignment
• 15:10
  using
• 15:10
  DINOv3
• 15:11
  as
• 15:11
  teacher,
• 15:12
  perceptual
• 15:12
  losses
• 15:13
  (LPIPS
• 15:13
  +
• 15:14
  DINO
• 15:14
  features)
• 15:14
  applied
• 15:15
  on
• 15:15
  pooled
• 15:15
  full
• 15:15
  images
• 15:16
  at
• 15:16
  all
• 15:16
  noise
• 15:16
  levels,
• 15:17
  and
• 15:17
  the
• 15:17
  Muon
• 15:17
  optimizer.
• 15:18
  Training
• 15:19
  starts
• 15:19
  at
• 15:19
  512px
• 15:21
  then
• 15:21
  fine-tunes
• 15:21
  at
• 15:21
  1024px.
• 15:23
  The
• 15:23
  result
• 15:23
  is
• 15:23
  not
• 15:24
  flawless
• 15:24
  —
• 15:25
  texture
• 15:25
  glitches,
• 15:26
  occasional
• 15:26
  anatomy
• 15:26
  issues
• 15:27
  —
• 15:27
  but
• 15:27
  the
• 15:27
  remaining
• 15:28
  artifacts
• 15:28
  look
• 15:29
  like
• 15:29
  undertraining,
• 15:30
  not
• 15:30
  architectural
• 15:30
  flaws.
• 15:31
  All
• 15:32
  code
• 15:32
  and
• 15:32
  configs
• 15:32
  are
• 15:32
  open-sourced.
• 15:34
  The
• 15:34
  democratization
• 15:35
  curve
• 15:35
  for
• 15:35
  image
• 15:35
  generation
• 15:36
  is
• 15:36
  steep:
• 15:37
  what
• 15:37
  cost
• 15:37
  millions
• 15:37
  in
• 15:37
  compute
• 15:38
  two
• 15:38
  years
• 15:38
  ago
• 15:38
  now
• 15:39
  costs
• 15:39
  less
• 15:39
  than
• 15:39
  a
• 15:39
  used
• 15:39
  laptop.
• 15:41
  On
• 15:41
  the
• 15:41
  integration
• 15:42
  side,
• 15:42
  a
• 15:43
  new
• 15:43
  OpenWebUI
• 15:43
  tool
• 15:44
  brings
• 15:44
  LTX
• 15:45
  2.3
• 15:45
  video
• 15:45
  generation
• 15:46
  directly
• 15:46
  into
• 15:46
  the
• 15:46
  chat
• 15:47
  interface
• 15:47
  via
• 15:47
  ComfyUI
• 15:48
  workflows,
• 15:49
  supporting
• 15:50
  both
• 15:50
  text-to-video
• 15:51
  and
• 15:51
  image-to-video
• 15:52
  with
• 15:52
  adjustable
• 15:52
  resolution,
• 15:53
  frame
• 15:53
  count,
• 15:54
  and
• 15:54
  FPS.
• 15:55
  CodeSpeak,
• 15:56
  created
• 15:56
  by
• 15:56
  the
• 15:57
  Kotlin
• 15:57
  creator,
• 15:58
  proposes
• 15:58
  a
• 15:58
  fundamentally
• 15:59
  different
• 15:59
  interface
• 16:00
  between
• 16:00
  humans
• 16:00
  and
• 16:00
  code
• 16:01
  generators:
• 16:02
  a
• 16:02
  formal
• 16:02
  language
• 16:03
  where
• 16:03
  plain-text
• 16:03
  specifications
• 16:04
  5–10x
• 16:05
  smaller
• 16:06
  than
• 16:06
  the
• 16:06
  equivalent
• 16:06
  source
• 16:07
  code
• 16:07
  produce
• 16:07
  production-ready
• 16:08
  implementations
• 16:09
  via
• 16:09
  LLM.
• 16:10
  The
• 16:11
  pitch
• 16:11
  is
• 16:11
  that
• 16:11
  maintaining
• 16:12
  specs
• 16:12
  is
• 16:12
  easier
• 16:13
  than
• 16:13
  maintaining
• 16:13
  code,
• 16:14
  and
• 16:14
  when
• 16:14
  your
• 16:15
  model
• 16:15
  outputs
• 16:15
  code
• 16:16
  rather
• 16:16
  than
• 16:16
  latents,
• 16:17
  you
• 16:17
  can
• 16:17
  plug
• 16:17
  in
• 16:17
  the
• 16:18
  entire
• 16:18
  classical
• 16:18
  software
• 16:19
  engineering
• 16:19
  toolbox
• 16:20
  for
• 16:20
  verification.
• 16:21
  Whether
• 16:21
  this
• 16:22
  is
• 16:22
  the
• 16:22
  future
• 16:22
  of
• 16:22
  programming
• 16:23
  or
• 16:23
  an
• 16:23
  elaborate
• 16:24
  prompt-engineering
• 16:25
  framework
• 16:25
  with
• 16:25
  syntax
• 16:26
  highlighting
• 16:26
  remains
• 16:27
  to
• 16:27
  be
• 16:27
  seen,
• 16:28
  but
• 16:28
  the
• 16:28
  pedigree
• 16:29
  is
• 16:29
  worth
• 16:29
  watching.
• 16:30
  For
• 16:30
  pure
• 16:30
  creative
• 16:31
  joy,
• 16:31
  Loopmaster
• 16:32
  published
• 16:32
  a
• 16:33
  step-by-step
• 16:33
  tutorial
• 16:34
  on
• 16:34
  building
• 16:34
  Roland's
• 16:35
  TB-303
• 16:37
  synthesizer
• 16:37
  from
• 16:37
  scratch
• 16:38
  —
• 16:38
  the
• 16:38
  instrument
• 16:39
  responsible
• 16:40
  for
• 16:40
  the
• 16:40
  entire
• 16:40
  acid
• 16:41
  house
• 16:41
  genre.
• 16:42
  Starting
• 16:42
  from
• 16:43
  raw
• 16:43
  oscillators,
• 16:44
  adding
• 16:44
  the
• 16:44
  characteristic
• 16:45
  diode-ladder
• 16:46
  low-pass
• 16:46
  filter,
• 16:47
  shaping
• 16:47
  the
• 16:47
  cutoff
• 16:48
  with
• 16:48
  envelopes
• 16:48
  instead
• 16:49
  of
• 16:49
  LFOs,
• 16:50
  layering
• 16:50
  accents
• 16:51
  and
• 16:51
  slides,
• 16:51
  and
• 16:52
  finishing
• 16:52
  with
• 16:52
  distortion,
• 16:53
  the
• 16:53
  tutorial
• 16:54
  turns
• 16:54
  synthesis
• 16:55
  education
• 16:55
  into
• 16:55
  interactive
• 16:56
  code.
• 16:57
  The
• 16:57
  "acid"
• 16:58
  sound
• 16:58
  emerges
• 16:58
  incrementally,
• 16:59
  and
• 17:00
  the
• 17:00
  values
• 17:00
  in
• 17:00
  the
• 17:00
  accent
• 17:01
  and
• 17:01
  slide
• 17:01
  parameters
• 17:02
  change
• 17:02
  the
• 17:02
  output
• 17:03
  dramatically.
• 17:04
  Infinite
• 17:04
  sounds
• 17:05
  from
• 17:05
  finite
• 17:05
  parameters
• 17:06
  —
• 17:06
  a
• 17:07
  good
• 17:07
  metaphor
• 17:07
  for
• 17:07
  what
• 17:08
  the
• 17:08
  rest
• 17:08
  of
• 17:08
  this
• 17:08
  edition's
• 17:09
  tools
• 17:09
  are
• 17:09
  trying
• 17:09
  to
• 17:09
  achieve.