Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest / AGI Dreams Podcast – March 20, 2026

Transcript

AGI Dreams Podcast – March 20, 2026 · AGI Dreams – Open, Uncensored, & Local - AI News Digest

• 0:00
  A
• 0:00
  developer
• 0:00
  just
• 0:00
  ran
• 0:01
  Qwen
• 0:01
  3.5
• 0:02
  397B
• 0:02
  —
• 0:03
  the
• 0:04
  actual
• 0:04
  209
• 0:05
  GB
• 0:05
  Mixture-of-Experts
• 0:06
  model
• 0:06
  —
• 0:07
  on
• 0:07
  a
• 0:07
  MacBook
• 0:07
  Pro
• 0:08
  with
• 0:08
  48
• 0:08
  GB
• 0:08
  of
• 0:09
  RAM,
• 0:09
  sustaining
• 0:10
  5.7
• 0:11
  tokens
• 0:11
  per
• 0:11
  second
• 0:11
  with
• 0:12
  only
• 0:12
  5.5
• 0:13
  GB
• 0:13
  of
• 0:13
  resident
• 0:14
  memory.
• 0:15
  The
• 0:15
  entire
• 0:15
  project,
• 0:16
  roughly
• 0:16
  5,000
• 0:17
  lines
• 0:17
  of
• 0:17
  Objective-C
• 0:18
  and
• 0:18
  1,100
• 0:19
  lines
• 0:19
  of
• 0:19
  Metal
• 0:20
  shaders,
• 0:20
  was
• 0:21
  written
• 0:21
  by
• 0:21
  Claude
• 0:21
  Opus
• 0:21
  4.6
• 0:22
  over
• 0:22
  24
• 0:23
  hours
• 0:23
  and
• 0:23
  90
• 0:23
  experiments.
• 0:25
  The
• 0:25
  human
• 0:25
  provided
• 0:26
  direction,
• 0:26
  reference
• 0:27
  materials
• 0:27
  (Apple's
• 0:28
  "LLM
• 0:28
  in
• 0:28
  a
• 0:28
  Flash"
• 0:29
  paper,
• 0:29
  Karpathy's
• 0:30
  autoresearch
• 0:31
  methodology),
• 0:32
  and
• 0:32
  systems-level
• 0:33
  insight
• 0:33
  at
• 0:33
  plateau
• 0:33
  points.
• 0:34
  The
• 0:34
  agent
• 0:35
  did
• 0:35
  the
• 0:35
  implementation
• 0:36
  work.
• 0:37
  The
• 0:37
  key
• 0:37
  insight
• 0:38
  is
• 0:38
  that
• 0:38
  Apple's
• 0:38
  unified
• 0:39
  memory
• 0:39
  architecture,
• 0:40
  originally
• 0:40
  designed
• 0:41
  for
• 0:41
  thin-and-light
• 0:41
  laptops,
• 0:42
  turns
• 0:43
  out
• 0:43
  to
• 0:43
  be
• 0:43
  nearly
• 0:43
  ideal
• 0:44
  for
• 0:44
  streaming
• 0:44
  model
• 0:44
  weights
• 0:45
  from
• 0:45
  SSD
• 0:45
  to
• 0:45
  GPU.
• 0:46
  The
• 0:47
  M3
• 0:47
  Max
• 0:47
  delivers
• 0:48
  17.5
• 0:48
  GB/s
• 0:49
  sequential
• 0:50
  reads
• 0:50
  —
• 0:50
  three
• 0:51
  times
• 0:51
  what
• 0:51
  Apple
• 0:51
  measured
• 0:52
  on
• 0:52
  the
• 0:52
  M1
• 0:52
  Max
• 0:52
  in
• 0:53
  their
• 0:53
  own
• 0:53
  paper.
• 0:54
  MoE
• 0:54
  models
• 0:54
  are
• 0:54
  perfect
• 0:55
  for
• 0:55
  this
• 0:55
  approach
• 0:55
  because
• 0:56
  they're
• 0:56
  absurdly
• 0:56
  sparse
• 0:57
  at
• 0:57
  inference
• 0:57
  time:
• 0:58
  Qwen
• 0:58
  3.5
• 0:59
  397B
• 1:00
  has
• 1:00
  512
• 1:01
  experts
• 1:01
  per
• 1:01
  layer
• 1:02
  but
• 1:02
  only
• 1:02
  activates
• 1:03
  10
• 1:03
  per
• 1:03
  token,
• 1:04
  and
• 1:04
  pruning
• 1:04
  to
• 1:04
  4
• 1:04
  active
• 1:05
  experts
• 1:05
  caused
• 1:05
  no
• 1:05
  quality
• 1:06
  degradation.
• 1:07
  Two-bit
• 1:07
  requantization
• 1:08
  cut
• 1:08
  expert
• 1:08
  storage
• 1:09
  from
• 1:09
  209
• 1:10
  GB
• 1:10
  to
• 1:10
  120
• 1:11
  GB
• 1:11
  with
• 1:12
  negligible
• 1:12
  quality
• 1:13
  loss
• 1:13
  (RMSE
• 1:15
  0.001-0.003
• 1:17
  per
• 1:18
  layer).
• 1:19
  The
• 1:19
  most
• 1:19
  counterintuitive
• 1:20
  finding:
• 1:20
  deleting
• 1:21
  a
• 1:21
  carefully
• 1:21
  engineered
• 1:22
  9.8
• 1:22
  GB
• 1:23
  Metal
• 1:23
  LRU
• 1:23
  cache
• 1:24
  and
• 1:24
  letting
• 1:24
  macOS
• 1:24
  handle
• 1:25
  caching
• 1:25
  made
• 1:25
  everything
• 1:26
  38%
• 1:26
  faster.
• 1:27
  The
• 1:27
  application-level
• 1:28
  cache
• 1:29
  was
• 1:29
  forcing
• 1:29
  Apple's
• 1:29
  hardware
• 1:30
  memory
• 1:30
  compressor
• 1:31
  to
• 1:31
  decompress
• 1:31
  at
• 1:31
  60,000-130,000
• 1:37
  operations
• 1:37
  per
• 1:37
  second,
• 1:38
  burning
• 1:38
  1-2
• 1:39
  GB/s
• 1:40
  of
• 1:40
  bandwidth
• 1:40
  on
• 1:41
  housekeeping.
• 1:42
  Same
• 1:42
  lesson
• 1:42
  PostgreSQL
• 1:43
  teaches
• 1:43
  about
• 1:43
  not
• 1:44
  competing
• 1:44
  with
• 1:44
  the
• 1:44
  OS
• 1:45
  buffer
• 1:45
  cache.
• 1:46
  The
• 1:46
  theoretical
• 1:46
  throughput
• 1:47
  floor,
• 1:48
  limited
• 1:48
  only
• 1:48
  by
• 1:48
  SSD
• 1:49
  bandwidth,
• 1:50
  is
• 1:50
  18.6
• 1:51
  tok/s.
• 1:52
  With
• 1:52
  M4
• 1:53
  Max's
• 1:53
  expected
• 1:54
  25
• 1:54
  GB/s
• 1:55
  SSD
• 1:56
  bandwidth,
• 1:56
  8
• 1:57
  tok/s
• 1:57
  is
• 1:58
  achievable
• 1:58
  with
• 1:58
  zero
• 1:59
  software
• 1:59
  changes.
• 2:01
  Within
• 2:01
  2-3
• 2:01
  hardware
• 2:02
  generations,
• 2:03
  10+
• 2:03
  tok/s
• 2:04
  on
• 2:04
  a
• 2:04
  400B-parameter
• 2:05
  model
• 2:06
  on
• 2:06
  a
• 2:06
  laptop
• 2:06
  becomes
• 2:07
  the
• 2:07
  baseline.
• 2:08
  Supporting
• 2:09
  this
• 2:09
  local-first
• 2:09
  trajectory,
• 2:10
  mlx-tune
• 2:11
  now
• 2:11
  provides
• 2:12
  an
• 2:12
  Unsloth-compatible
• 2:13
  API
• 2:13
  for
• 2:13
  fine-tuning
• 2:14
  LLMs
• 2:15
  natively
• 2:15
  on
• 2:15
  Apple
• 2:16
  Silicon
• 2:16
  —
• 2:17
  SFT,
• 2:17
  DPO,
• 2:18
  GRPO,
• 2:19
  and
• 2:19
  vision
• 2:19
  model
• 2:20
  training
• 2:20
  with
• 2:20
  LoRA/QLoRA,
• 2:23
  all
• 2:23
  workable
• 2:23
  on
• 2:24
  8
• 2:24
  GB
• 2:24
  unified
• 2:25
  RAM.
• 2:26
  The
• 2:26
  idea
• 2:26
  is
• 2:26
  sound:
• 2:27
  prototype
• 2:27
  locally,
• 2:28
  catch
• 2:28
  data
• 2:29
  pipeline
• 2:29
  bugs
• 2:29
  (bad
• 2:30
  chat
• 2:30
  templates,
• 2:31
  wrong
• 2:31
  tokenization,
• 2:32
  response-only
• 2:33
  masking
• 2:33
  failures)
• 2:34
  before
• 2:34
  burning
• 2:35
  GPU
• 2:35
  hours
• 2:35
  in
• 2:36
  the
• 2:36
  cloud
• 2:36
  .
• 2:40
  NVIDIA's
• 2:40
  2026
• 2:41
  conference
• 2:41
  brought
• 2:42
  two
• 2:42
  notable
• 2:42
  announcements:
• 2:43
  Nemotron
• 2:44
  4,
• 2:44
  an
• 2:45
  open
• 2:45
  model
• 2:45
  built
• 2:45
  in
• 2:46
  coalition
• 2:46
  with
• 2:46
  Thinking
• 2:46
  Machines,
• 2:47
  Sarvam,
• 2:48
  Perplexity,
• 2:48
  Mistral,
• 2:49
  and
• 2:49
  multiple
• 2:49
  nations
• 2:49
  —
• 2:50
  signaling
• 2:51
  that
• 2:51
  NVIDIA
• 2:51
  is
• 2:52
  serious
• 2:52
  about
• 2:52
  the
• 2:52
  open-weight
• 2:53
  game
• 2:53
  —
• 2:54
  and
• 2:54
  benchmark
• 2:54
  numbers
• 2:55
  that,
• 2:55
  if
• 2:55
  not
• 2:56
  cherry-picked,
• 2:56
  show
• 2:56
  significantly
• 2:57
  bigger
• 2:57
  margins
• 2:58
  over
• 2:58
  the
• 2:58
  competition
• 2:58
  than
• 2:59
  previous
• 2:59
  generations
• 2:59
  .
• 3:00
  The
• 3:00
  more
• 3:00
  substantive
• 3:01
  release,
• 3:01
  however,
• 3:02
  is
• 3:02
  Nemotron
• 3:02
  3
• 3:03
  Nano
• 3:03
  4B,
• 3:04
  a
• 3:04
  hybrid
• 3:04
  Mamba-Transformer
• 3:05
  model
• 3:05
  that
• 3:05
  represents
• 3:06
  the
• 3:06
  most
• 3:06
  interesting
• 3:07
  architectural
• 3:07
  choice
• 3:08
  in
• 3:08
  the
• 3:08
  sub-10B
• 3:08
  class
• 3:09
  right
• 3:09
  now.
• 3:10
  Pruned
• 3:10
  and
• 3:11
  distilled
• 3:11
  from
• 3:11
  the
• 3:11
  9B
• 3:12
  Nemotron
• 3:12
  Nano
• 3:12
  v2
• 3:13
  using
• 3:13
  NVIDIA's
• 3:14
  new
• 3:14
  Nemotron
• 3:14
  Elastic
• 3:15
  framework,
• 3:16
  the
• 3:16
  4B
• 3:16
  model
• 3:17
  retains
• 3:17
  all
• 3:17
  four
• 3:18
  attention
• 3:18
  layers
• 3:18
  from
• 3:18
  the
• 3:18
  parent
• 3:19
  while
• 3:19
  dropping
• 3:19
  from
• 3:19
  56
• 3:20
  to
• 3:20
  42
• 3:20
  total
• 3:21
  layers
• 3:21
  (21
• 3:21
  Mamba,
• 3:22
  4
• 3:22
  attention,
• 3:23
  17
• 3:24
  MLP).
• 3:25
  The
• 3:25
  router-guided
• 3:25
  pruning
• 3:26
  decided
• 3:26
  which
• 3:26
  dimensions
• 3:27
  to
• 3:27
  cut
• 3:27
  —
• 3:28
  Mamba
• 3:28
  heads,
• 3:29
  hidden
• 3:29
  dimension,
• 3:29
  FFN
• 3:30
  channels,
• 3:31
  and
• 3:31
  full
• 3:31
  layers
• 3:31
  —
• 3:32
  all
• 3:32
  jointly
• 3:32
  optimized
• 3:33
  with
• 3:33
  knowledge
• 3:34
  distillation
• 3:34
  rather
• 3:35
  than
• 3:35
  the
• 3:35
  traditional
• 3:35
  prune-then-retrain
• 3:36
  pipeline.
• 3:37
  What
• 3:37
  makes
• 3:38
  Nemotron
• 3:38
  3
• 3:38
  Nano
• 3:39
  4B
• 3:39
  stand
• 3:39
  out
• 3:40
  is
• 3:40
  its
• 3:40
  target:
• 3:41
  it's
• 3:41
  explicitly
• 3:42
  optimized
• 3:42
  for
• 3:42
  Jetson
• 3:43
  Thor,
• 3:43
  Jetson
• 3:43
  Orin
• 3:44
  Nano,
• 3:44
  DGX
• 3:45
  Spark,
• 3:45
  and
• 3:45
  RTX
• 3:46
  GPUs.
• 3:47
  The
• 3:47
  Q4
• 3:47
  K
• 3:48
  M
• 3:48
  GGUF
• 3:48
  version
• 3:49
  achieves
• 3:49
  18
• 3:50
  tokens/s
• 3:51
  on
• 3:51
  a
• 3:51
  Jetson
• 3:51
  Orin
• 3:52
  Nano
• 3:52
  8GB
• 3:52
  —
• 3:53
  twice
• 3:54
  the
• 3:54
  throughput
• 3:54
  of
• 3:54
  the
• 3:54
  9B
• 3:55
  parent.
• 3:55
  State-of-the-art
• 3:56
  instruction
• 3:57
  following
• 3:57
  (IFBench,
• 3:58
  IFEval)
• 3:58
  in
• 3:59
  its
• 3:59
  size
• 3:59
  class,
• 4:00
  lowest
• 4:00
  VRAM
• 4:00
  footprint
• 4:01
  under
• 4:01
  both
• 4:01
  low
• 4:01
  and
• 4:02
  high
• 4:02
  input/output
• 4:03
  length
• 4:03
  settings,
• 4:04
  and
• 4:04
  strong
• 4:04
  tool-use
• 4:05
  performance
• 4:05
  make
• 4:05
  this
• 4:05
  a
• 4:06
  purpose-built
• 4:06
  edge
• 4:06
  agent
• 4:07
  brain.
• 4:08
  The
• 4:08
  three-stage
• 4:09
  RLVR
• 4:09
  pipeline
• 4:10
  that
• 4:10
  targets
• 4:10
  instruction
• 4:10
  following
• 4:11
  and
• 4:11
  tool
• 4:11
  calling,
• 4:12
  using
• 4:12
  environments
• 4:13
  for
• 4:13
  structured
• 4:13
  output
• 4:14
  (JSON,
• 4:14
  XML)
• 4:15
  and
• 4:15
  multi-turn
• 4:16
  conversational
• 4:17
  tool
• 4:17
  calling,
• 4:18
  is
• 4:18
  clearly
• 4:18
  aimed
• 4:18
  at
• 4:18
  the
• 4:19
  on-device
• 4:19
  agent
• 4:19
  use
• 4:20
  case.
• 4:21
  IBM's
• 4:21
  Granite
• 4:22
  4.0
• 4:22
  1B
• 4:23
  Speech
• 4:23
  takes
• 4:23
  a
• 4:23
  different
• 4:24
  edge
• 4:24
  angle:
• 4:25
  multilingual
• 4:25
  ASR
• 4:26
  and
• 4:26
  bidirectional
• 4:27
  speech
• 4:27
  translation
• 4:28
  in
• 4:28
  a
• 4:28
  model
• 4:28
  with
• 4:28
  half
• 4:28
  the
• 4:28
  parameters
• 4:29
  of
• 4:29
  its
• 4:29
  predecessor.
• 4:30
  It
• 4:30
  now
• 4:31
  covers
• 4:31
  English,
• 4:32
  French,
• 4:33
  German,
• 4:33
  Spanish,
• 4:34
  Portuguese,
• 4:35
  and
• 4:35
  Japanese,
• 4:36
  with
• 4:36
  new
• 4:36
  keyword-list
• 4:37
  biasing
• 4:37
  for
• 4:37
  names
• 4:38
  and
• 4:38
  acronyms
• 4:38
  —
• 4:39
  a
• 4:39
  frequently
• 4:39
  requested
• 4:40
  enterprise
• 4:40
  feature.
• 4:41
  It
• 4:41
  ranked
• 4:41
  #1
• 4:42
  on
• 4:42
  the
• 4:42
  Open
• 4:42
  ASR
• 4:43
  Leaderboard
• 4:43
  despite
• 4:44
  its
• 4:44
  tiny
• 4:44
  footprint
• 4:44
  .
• 4:45
  Meanwhile,
• 4:46
  LocoTrainer-4B
• 4:47
  offers
• 4:48
  a
• 4:48
  specialized
• 4:48
  4B
• 4:49
  agent
• 4:49
  distilled
• 4:49
  from
• 4:50
  Qwen3-Coder-Next,
• 4:51
  trained
• 4:51
  on
• 4:51
  361K+
• 4:54
  MS-SWIFT
• 4:56
  documentation
• 4:57
  samples
• 4:57
  to
• 4:57
  answer
• 4:57
  framework-specific
• 4:58
  questions
• 4:59
  without
• 4:59
  hallucination.
• 5:00
  It
• 5:00
  runs
• 5:01
  a
• 5:01
  Claude
• 5:01
  Code-style
• 5:02
  agent
• 5:02
  loop
• 5:02
  with
• 5:02
  tool
• 5:03
  calling
• 5:03
  (Read,
• 5:03
  Grep,
• 5:04
  Glob,
• 5:05
  Bash,
• 5:05
  Write)
• 5:06
  and
• 5:06
  injects
• 5:06
  absolute
• 5:07
  paths
• 5:07
  so
• 5:07
  the
• 5:07
  model
• 5:08
  never
• 5:08
  guesses
• 5:08
  —
• 5:09
  a
• 5:09
  design
• 5:09
  choice
• 5:10
  that
• 5:10
  took
• 5:10
  tool-calling
• 5:10
  reliability
• 5:11
  from
• 5:11
  0%
• 5:12
  to
• 5:12
  100%
• 5:13
  in
• 5:13
  testing
• 5:13
  .
• 5:13
  On
• 5:13
  the
• 5:14
  MoE
• 5:14
  front,
• 5:14
  BitDance-14B-16x
• 5:17
  appeared
• 5:18
  as
• 5:18
  a
• 5:18
  trending
• 5:18
  model
• 5:18
  on
• 5:18
  HuggingFace
• 5:19
  .
• 5:22
  The
• 5:22
  question
• 5:22
  of
• 5:22
  where
• 5:23
  AI
• 5:23
  agents
• 5:23
  should
• 5:23
  execute
• 5:24
  their
• 5:24
  code
• 5:24
  got
• 5:24
  a
• 5:24
  dramatic
• 5:25
  new
• 5:25
  answer
• 5:25
  this
• 5:25
  week.
• 5:26
  Zeroboot
• 5:27
  achieves
• 5:27
  sub-millisecond
• 5:28
  VM
• 5:28
  sandboxes
• 5:28
  by
• 5:29
  applying
• 5:29
  copy-on-write
• 5:30
  memory
• 5:30
  forking
• 5:30
  to
• 5:30
  Firecracker
• 5:31
  snapshots:
• 5:32
  boot
• 5:32
  a
• 5:32
  VM
• 5:32
  once,
• 5:33
  pre-load
• 5:33
  your
• 5:33
  runtime,
• 5:34
  snapshot
• 5:34
  memory
• 5:34
  and
• 5:34
  CPU
• 5:35
  state,
• 5:35
  then
• 5:36
  fork
• 5:36
  new
• 5:36
  KVM
• 5:36
  virtual
• 5:37
  machines
• 5:37
  by
• 5:37
  mapping
• 5:37
  the
• 5:37
  snapshot
• 5:38
  as
• 5:38
  CoW
• 5:38
  pages
• 5:39
  and
• 5:39
  restoring
• 5:40
  CPU
• 5:40
  state.
• 5:41
  The
• 5:41
  result
• 5:41
  is
• 5:41
  0.79ms
• 5:42
  p50
• 5:43
  spawn
• 5:43
  latency
• 5:44
  (1.74ms
• 5:44
  p99),
• 5:46
  with
• 5:46
  each
• 5:46
  sandbox
• 5:46
  consuming
• 5:47
  roughly
• 5:47
  265KB
• 5:49
  of
• 5:49
  memory.
• 5:49
  For
• 5:49
  comparison,
• 5:50
  E2B
• 5:50
  takes
• 5:51
  ~150ms
• 5:51
  and
• 5:51
  ~128MB
• 5:53
  per
• 5:53
  sandbox.
• 5:54
  A
• 5:54
  thousand
• 5:55
  concurrent
• 5:55
  forks
• 5:55
  complete
• 5:56
  in
• 5:56
  815ms.
• 5:57
  Each
• 5:57
  fork
• 5:58
  is
• 5:58
  a
• 5:58
  real
• 5:58
  KVM
• 5:58
  virtual
• 5:59
  machine
• 5:59
  with
• 5:59
  hardware-enforced
• 6:00
  memory
• 6:00
  isolation
• 6:00
  —
• 6:01
  not
• 6:01
  a
• 6:01
  container,
• 6:02
  not
• 6:02
  a
• 6:02
  process
• 6:02
  namespace
• 6:03
  trick.
• 6:04
  Yesterday's
• 6:04
  coverage
• 6:05
  of
• 6:05
  sandbox
• 6:05
  escapes
• 6:05
  in
• 6:06
  AWS
• 6:06
  Bedrock
• 6:06
  AgentCore
• 6:07
  demonstrated
• 6:08
  that
• 6:08
  agent
• 6:08
  execution
• 6:09
  environments
• 6:09
  remain
• 6:10
  an
• 6:10
  active
• 6:10
  attack
• 6:10
  surface.
• 6:11
  Zeroboot's
• 6:12
  approach
• 6:12
  —
• 6:13
  real
• 6:13
  hypervisor
• 6:13
  isolation
• 6:14
  at
• 6:14
  container-like
• 6:14
  speeds
• 6:15
  —
• 6:16
  addresses
• 6:16
  the
• 6:16
  fundamental
• 6:17
  tension
• 6:17
  between
• 6:17
  security
• 6:18
  and
• 6:18
  latency
• 6:19
  that
• 6:19
  plagues
• 6:19
  every
• 6:20
  agent
• 6:20
  sandbox
• 6:20
  provider.
• 6:21
  Known
• 6:22
  limitations
• 6:22
  are
• 6:22
  honest:
• 6:23
  forks
• 6:23
  share
• 6:24
  CSPRNG
• 6:26
  state
• 6:27
  (requiring
• 6:27
  explicit
• 6:28
  userspace
• 6:28
  PRNG
• 6:29
  reseeding),
• 6:30
  single
• 6:30
  vCPU
• 6:31
  only,
• 6:31
  no
• 6:31
  networking
• 6:32
  (serial
• 6:32
  I/O
• 6:33
  only),
• 6:33
  and
• 6:34
  template
• 6:34
  updates
• 6:34
  require
• 6:35
  a
• 6:35
  full
• 6:35
  15-second
• 6:36
  re-snapshot.
• 6:37
  Mistral
• 6:37
  AI
• 6:38
  entered
• 6:38
  the
• 6:38
  enterprise
• 6:38
  model
• 6:39
  training
• 6:39
  market
• 6:39
  with
• 6:39
  Forge,
• 6:40
  a
• 6:40
  platform
• 6:41
  for
• 6:41
  organizations
• 6:41
  to
• 6:41
  pre-train,
• 6:42
  post-train,
• 6:43
  and
• 6:43
  RL-align
• 6:44
  models
• 6:44
  on
• 6:44
  proprietary
• 6:45
  data.
• 6:46
  The
• 6:46
  pitch:
• 6:47
  generic
• 6:47
  models
• 6:48
  trained
• 6:48
  on
• 6:48
  public
• 6:48
  data
• 6:49
  don't
• 6:49
  understand
• 6:50
  your
• 6:50
  internal
• 6:50
  terminology,
• 6:51
  compliance
• 6:52
  frameworks,
• 6:52
  or
• 6:53
  codebases.
• 6:54
  Forge
• 6:54
  supports
• 6:55
  dense
• 6:55
  and
• 6:55
  MoE
• 6:55
  architectures,
• 6:56
  multimodal
• 6:57
  inputs,
• 6:58
  and
• 6:58
  continuous
• 6:58
  improvement
• 6:59
  through
• 6:59
  reinforcement
• 7:00
  learning
• 7:00
  pipelines.
• 7:01
  Partners
• 7:02
  include
• 7:02
  several
• 7:02
  unnamed
• 7:03
  "world-leading
• 7:03
  organizations"
• 7:05
  plus
• 7:05
  Singapore's
• 7:05
  Home
• 7:06
  Team
• 7:06
  Science
• 7:06
  and
• 7:07
  Technology
• 7:07
  Agency.
• 7:08
  The
• 7:08
  agent
• 7:09
  angle
• 7:09
  is
• 7:09
  explicit
• 7:09
  —
• 7:10
  Mistral
• 7:11
  Vibe,
• 7:11
  their
• 7:11
  autonomous
• 7:12
  agent,
• 7:13
  can
• 7:13
  use
• 7:13
  Forge
• 7:13
  to
• 7:14
  fine-tune
• 7:14
  models,
• 7:15
  find
• 7:15
  optimal
• 7:15
  hyperparameters,
• 7:17
  schedule
• 7:17
  jobs,
• 7:18
  and
• 7:18
  generate
• 7:18
  synthetic
• 7:19
  data
• 7:19
  to
• 7:19
  hill-climb
• 7:20
  evals,
• 7:20
  all
• 7:21
  in
• 7:21
  plain
• 7:21
  English
• 7:22
  .
• 7:22
  On
• 7:23
  the
• 7:23
  data
• 7:23
  access
• 7:23
  side,
• 7:24
  SeeQL
• 7:24
  combines
• 7:25
  OpenUI
• 7:26
  Lang's
• 7:26
  declarative
• 7:27
  UI
• 7:27
  generation
• 7:27
  (claimed
• 7:28
  67%
• 7:29
  fewer
• 7:29
  tokens
• 7:30
  than
• 7:30
  JSON)
• 7:31
  with
• 7:31
  a
• 7:31
  Text-to-SQL
• 7:31
  MCP
• 7:32
  server,
• 7:33
  letting
• 7:33
  users
• 7:34
  query
• 7:34
  SQLite
• 7:35
  databases
• 7:35
  in
• 7:35
  natural
• 7:36
  language
• 7:37
  and
• 7:37
  get
• 7:37
  charts
• 7:37
  and
• 7:38
  data
• 7:38
  tables
• 7:38
  rather
• 7:39
  than
• 7:39
  raw
• 7:39
  SQL
• 7:39
  dumps
• 7:40
  .
• 7:42
  h1-brain
• 7:43
  is
• 7:43
  an
• 7:43
  MCP
• 7:43
  server
• 7:44
  that
• 7:44
  turns
• 7:44
  Claude
• 7:45
  Desktop
• 7:45
  or
• 7:45
  Claude
• 7:46
  Code
• 7:46
  into
• 7:46
  a
• 7:46
  bug
• 7:47
  bounty
• 7:47
  research
• 7:47
  assistant.
• 7:48
  It
• 7:49
  connects
• 7:49
  to
• 7:49
  the
• 7:49
  HackerOne
• 7:49
  API,
• 7:50
  pulls
• 7:50
  your
• 7:51
  bounty
• 7:51
  history,
• 7:51
  program
• 7:52
  scopes,
• 7:52
  and
• 7:53
  report
• 7:53
  details
• 7:53
  into
• 7:53
  a
• 7:54
  local
• 7:54
  SQLite
• 7:54
  database,
• 7:55
  then
• 7:55
  ships
• 7:56
  with
• 7:56
  a
• 7:56
  pre-built
• 7:56
  database
• 7:57
  of
• 7:57
  3,600+
• 7:58
  publicly
• 7:58
  disclosed
• 7:59
  bounty-awarded
• 7:59
  reports.
• 8:00
  The
• 8:00
  primary
• 8:01
  tool,
• 8:01
  hack(handle),
• 8:02
  generates
• 8:02
  a
• 8:02
  full
• 8:03
  attack
• 8:03
  briefing
• 8:03
  in
• 8:04
  a
• 8:04
  single
• 8:04
  call:
• 8:05
  fresh
• 8:05
  scope
• 8:06
  from
• 8:06
  the
• 8:06
  API,
• 8:06
  your
• 8:07
  past
• 8:07
  findings,
• 8:08
  public
• 8:08
  disclosures
• 8:09
  for
• 8:09
  the
• 8:09
  target
• 8:09
  program,
• 8:10
  weakness
• 8:10
  patterns
• 8:11
  across
• 8:11
  all
• 8:11
  your
• 8:11
  programs,
• 8:12
  untouched
• 8:13
  bounty-eligible
• 8:13
  assets,
• 8:14
  and
• 8:14
  suggested
• 8:15
  attack
• 8:15
  vectors
• 8:15
  based
• 8:16
  on
• 8:16
  vulnerabilities
• 8:16
  that
• 8:16
  paid
• 8:17
  elsewhere
• 8:17
  but
• 8:17
  haven't
• 8:18
  been
• 8:18
  found
• 8:18
  on
• 8:18
  this
• 8:18
  target.
• 8:19
  It's
• 8:19
  essentially
• 8:20
  a
• 8:20
  memory-augmented
• 8:21
  offensive
• 8:21
  research
• 8:21
  agent
• 8:21
  that
• 8:22
  never
• 8:22
  forgets
• 8:22
  what
• 8:22
  worked
• 8:23
  before.
• 8:24
  The
• 8:24
  MCP
• 8:24
  (Model
• 8:24
  Context
• 8:25
  Protocol)
• 8:25
  integration
• 8:26
  is
• 8:26
  the
• 8:26
  interesting
• 8:26
  design
• 8:27
  choice.
• 8:27
  Rather
• 8:28
  than
• 8:28
  building
• 8:28
  a
• 8:28
  standalone
• 8:29
  tool,
• 8:29
  h1-brain
• 8:30
  exposes
• 8:30
  its
• 8:30
  capabilities
• 8:31
  through
• 8:31
  MCP
• 8:32
  so
• 8:32
  any
• 8:33
  compatible
• 8:33
  AI
• 8:33
  assistant
• 8:34
  can
• 8:34
  search
• 8:34
  reports,
• 8:35
  cross-reference
• 8:35
  weakness
• 8:36
  patterns,
• 8:36
  and
• 8:36
  generate
• 8:37
  attack
• 8:37
  strategies.
• 8:38
  The
• 8:38
  two-database
• 8:38
  architecture
• 8:39
  —
• 8:39
  personal
• 8:40
  reports
• 8:40
  via
• 8:40
  the
• 8:40
  HackerOne
• 8:41
  API,
• 8:41
  plus
• 8:42
  the
• 8:42
  community
• 8:42
  knowledge
• 8:43
  base
• 8:43
  of
• 8:43
  public
• 8:43
  disclosures
• 8:44
  —
• 8:44
  means
• 8:44
  the
• 8:45
  AI
• 8:45
  can
• 8:45
  reason
• 8:45
  about
• 8:46
  what
• 8:46
  vulnerability
• 8:46
  classes
• 8:47
  are
• 8:47
  rewarded
• 8:47
  on
• 8:47
  a
• 8:47
  given
• 8:48
  program
• 8:48
  and
• 8:49
  identify
• 8:49
  gaps
• 8:49
  in
• 8:49
  your
• 8:50
  coverage.
• 8:51
  On
• 8:51
  the
• 8:51
  defensive
• 8:51
  research
• 8:52
  side,
• 8:52
  OBLITERATUS
• 8:55
  bills
• 8:55
  itself
• 8:55
  as
• 8:56
  the
• 8:56
  most
• 8:56
  advanced
• 8:56
  open-source
• 8:57
  toolkit
• 8:57
  for
• 8:57
  understanding
• 8:58
  and
• 8:58
  removing
• 8:58
  refusal
• 8:59
  behaviors
• 8:59
  from
• 8:59
  LLMs,
• 9:00
  with
• 9:00
  15
• 9:01
  analysis
• 9:01
  modules,
• 9:01
  837
• 9:03
  tests,
• 9:03
  and
• 9:03
  support
• 9:04
  for
• 9:04
  116
• 9:05
  models
• 9:05
  across
• 9:05
  5
• 9:05
  compute
• 9:06
  tiers.
• 9:07
  The
• 9:07
  toolkit
• 9:07
  goes
• 9:07
  beyond
• 9:08
  brute-force
• 9:08
  abliteration:
• 9:09
  it
• 9:09
  maps
• 9:10
  the
• 9:10
  geometric
• 9:10
  structure
• 9:11
  of
• 9:11
  guardrails
• 9:11
  across
• 9:11
  layers,
• 9:12
  detects
• 9:13
  alignment
• 9:13
  training
• 9:13
  methods
• 9:14
  (DPO
• 9:14
  vs
• 9:15
  RLHF
• 9:15
  vs
• 9:16
  CAI
• 9:16
  vs
• 9:17
  SFT)
• 9:17
  from
• 9:17
  subspace
• 9:18
  geometry
• 9:18
  alone,
• 9:19
  and
• 9:19
  measures
• 9:20
  whether
• 9:20
  refusal
• 9:20
  directions
• 9:21
  generalize
• 9:21
  across
• 9:22
  models
• 9:22
  (the
• 9:22
  Cross-Model
• 9:22
  Universality
• 9:23
  Index).
• 9:24
  Novel
• 9:25
  techniques
• 9:25
  include
• 9:25
  Expert-Granular
• 9:26
  Abliteration
• 9:26
  for
• 9:27
  MoE
• 9:27
  models,
• 9:28
  CoT-Aware
• 9:28
  Ablation
• 9:28
  that
• 9:29
  preserves
• 9:29
  reasoning
• 9:29
  circuits,
• 9:30
  and
• 9:30
  LoRA-Based
• 9:31
  Reversible
• 9:31
  Ablation
• 9:32
  that
• 9:32
  avoids
• 9:33
  permanent
• 9:33
  weight
• 9:33
  surgery.
• 9:34
  The
• 9:35
  "informed"
• 9:35
  pipeline
• 9:36
  closes
• 9:36
  the
• 9:36
  analysis-to-removal
• 9:37
  loop:
• 9:38
  analysis
• 9:38
  modules
• 9:39
  auto-configure
• 9:39
  obliteration
• 9:40
  strategy,
• 9:41
  detect
• 9:41
  the
• 9:41
  Ouroboros
• 9:42
  effect
• 9:42
  (guardrails
• 9:43
  that
• 9:43
  self-repair),
• 9:44
  and
• 9:44
  fire
• 9:44
  additional
• 9:44
  targeted
• 9:45
  passes
• 9:45
  at
• 9:45
  compensating
• 9:46
  layers.
• 9:47
  Every
• 9:47
  run
• 9:47
  with
• 9:48
  telemetry
• 9:48
  enabled
• 9:48
  contributes
• 9:49
  to
• 9:49
  a
• 9:49
  crowd-sourced
• 9:50
  research
• 9:50
  dataset
• 9:50
  —
• 9:51
  making
• 9:51
  this
• 9:52
  arguably
• 9:52
  the
• 9:52
  largest
• 9:53
  distributed
• 9:53
  mechanistic
• 9:53
  interpretability
• 9:54
  experiment
• 9:55
  running
• 9:55
  today
• 9:55
  .
• 9:58
  Flash-KMeans,
• 9:59
  a
• 9:59
  new
• 9:59
  paper
• 9:59
  from
• 9:59
  Berkeley,
• 10:00
  demonstrates
• 10:00
  that
• 10:01
  the
• 10:01
  classical
• 10:01
  k-means
• 10:01
  algorithm
• 10:02
  has
• 10:02
  been
• 10:02
  leaving
• 10:03
  massive
• 10:03
  performance
• 10:03
  on
• 10:04
  the
• 10:04
  table
• 10:04
  —
• 10:04
  not
• 10:05
  because
• 10:05
  of
• 10:05
  algorithmic
• 10:05
  complexity,
• 10:06
  but
• 10:06
  because
• 10:06
  GPU
• 10:07
  implementations
• 10:08
  are
• 10:08
  fundamentally
• 10:08
  bottlenecked
• 10:09
  by
• 10:09
  memory
• 10:09
  access
• 10:10
  patterns.
• 10:10
  The
• 10:10
  standard
• 10:11
  approach
• 10:11
  materializes
• 10:12
  a
• 10:12
  massive
• 10:12
  distance
• 10:13
  matrix
• 10:13
  in
• 10:13
  HBM
• 10:14
  (high-bandwidth
• 10:14
  memory)
• 10:14
  —
• 10:15
  writing
• 10:15
  it
• 10:15
  and
• 10:15
  immediately
• 10:16
  reading
• 10:16
  it
• 10:16
  back
• 10:16
  —
• 10:16
  which
• 10:17
  costs
• 10:17
  more
• 10:17
  than
• 10:17
  the
• 10:17
  actual
• 10:18
  computation.
• 10:18
  The
• 10:19
  update
• 10:19
  stage
• 10:19
  suffers
• 10:19
  from
• 10:20
  severe
• 10:20
  atomic
• 10:20
  write
• 10:21
  contention
• 10:21
  as
• 10:21
  threads
• 10:22
  scramble
• 10:22
  to
• 10:22
  update
• 10:22
  the
• 10:22
  same
• 10:23
  centroids.
• 10:23
  Flash-KMeans
• 10:24
  introduces
• 10:25
  two
• 10:25
  kernel-level
• 10:25
  innovations:
• 10:26
  FlashAssign,
• 10:27
  which
• 10:27
  fuses
• 10:28
  distance
• 10:28
  computation
• 10:28
  with
• 10:29
  an
• 10:29
  online
• 10:29
  argmin
• 10:30
  to
• 10:30
  completely
• 10:30
  bypass
• 10:31
  the
• 10:31
  intermediate
• 10:31
  matrix
• 10:32
  (inspired
• 10:32
  by
• 10:32
  FlashAttention's
• 10:33
  IO-aware
• 10:33
  philosophy),
• 10:34
  and
• 10:34
  Sort-Inverse
• 10:35
  Update,
• 10:36
  which
• 10:36
  constructs
• 10:36
  an
• 10:36
  explicit
• 10:37
  inverse
• 10:37
  mapping
• 10:38
  to
• 10:38
  transform
• 10:38
  high-contention
• 10:39
  atomic
• 10:39
  scatters
• 10:40
  into
• 10:40
  regular
• 10:40
  segment-level
• 10:41
  reductions.
• 10:42
  The
• 10:42
  results
• 10:42
  on
• 10:42
  NVIDIA
• 10:43
  H200
• 10:43
  are
• 10:43
  striking:
• 10:44
  up
• 10:44
  to
• 10:44
  54x
• 10:45
  end-to-end
• 10:45
  speedup
• 10:46
  over
• 10:46
  best
• 10:46
  baselines,
• 10:47
  with
• 10:47
  FlashAssign
• 10:48
  delivering
• 10:48
  up
• 10:48
  to
• 10:48
  21.2x
• 10:49
  on
• 10:49
  the
• 10:49
  assignment
• 10:50
  kernel
• 10:50
  and
• 10:50
  Sort-Inverse
• 10:51
  Update
• 10:51
  achieving
• 10:51
  up
• 10:51
  to
• 10:52
  6.3x
• 10:52
  on
• 10:53
  centroids.
• 10:54
  The
• 10:54
  system
• 10:54
  gracefully
• 10:54
  scales
• 10:55
  to
• 10:55
  one
• 10:55
  billion
• 10:55
  points
• 10:56
  in
• 10:56
  out-of-core
• 10:56
  mode
• 10:57
  and
• 10:57
  includes
• 10:57
  a
• 10:57
  cache-aware
• 10:58
  compile
• 10:58
  heuristic
• 10:59
  that
• 10:59
  slashes
• 10:59
  configuration
• 11:00
  tuning
• 11:00
  overhead
• 11:01
  by
• 11:01
  175x
• 11:02
  with
• 11:02
  near-zero
• 11:02
  performance
• 11:03
  degradation.
• 11:04
  The
• 11:04
  practical
• 11:05
  implication:
• 11:06
  k-means
• 11:06
  is
• 11:06
  no
• 11:06
  longer
• 11:07
  just
• 11:07
  an
• 11:07
  offline
• 11:08
  preprocessing
• 11:08
  step
• 11:08
  —
• 11:09
  it
• 11:09
  becomes
• 11:10
  viable
• 11:10
  as
• 11:10
  an
• 11:10
  online
• 11:11
  primitive
• 11:11
  for
• 11:11
  KV
• 11:12
  cache
• 11:12
  compression,
• 11:13
  sparse
• 11:13
  attention
• 11:13
  routing,
• 11:14
  and
• 11:14
  semantic
• 11:15
  token
• 11:15
  permutation
• 11:15
  in
• 11:16
  diffusion
• 11:16
  transformers.
• 11:17
  AgentSAT
• 11:18
  takes
• 11:18
  the
• 11:18
  autonomous
• 11:19
  research
• 11:19
  concept
• 11:20
  in
• 11:20
  a
• 11:20
  completely
• 11:20
  different
• 11:20
  direction:
• 11:21
  an
• 11:21
  AI
• 11:22
  agent
• 11:22
  that
• 11:22
  teaches
• 11:23
  itself
• 11:23
  to
• 11:23
  become
• 11:23
  a
• 11:23
  MaxSAT
• 11:24
  expert.
• 11:25
  Given
• 11:25
  229
• 11:26
  weighted
• 11:26
  MaxSAT
• 11:26
  instances
• 11:27
  from
• 11:27
  the
• 11:27
  2024
• 11:28
  competition,
• 11:28
  the
• 11:29
  agent
• 11:29
  —
• 11:30
  running
• 11:30
  as
• 11:30
  Claude
• 11:30
  Code
• 11:31
  on
• 11:31
  EC2
• 11:31
  —
• 11:32
  reads
• 11:32
  accumulated
• 11:33
  knowledge
• 11:33
  from
• 11:33
  prior
• 11:34
  runs,
• 11:34
  experiments
• 11:35
  with
• 11:35
  solvers,
• 11:36
  discovers
• 11:36
  what
• 11:36
  works,
• 11:37
  and
• 11:37
  pushes
• 11:37
  improvements
• 11:38
  to
• 11:38
  a
• 11:38
  shared
• 11:38
  git
• 11:38
  repo
• 11:39
  so
• 11:39
  other
• 11:39
  agents
• 11:39
  can
• 11:39
  build
• 11:39
  on
• 11:40
  findings.
• 11:41
  The
• 11:41
  result:
• 11:41
  220
• 11:42
  of
• 11:42
  229
• 11:43
  instances
• 11:44
  solved,
• 11:44
  30
• 11:45
  matching
• 11:45
  competition-optimal
• 11:46
  solutions,
• 11:46
  and
• 11:47
  5
• 11:47
  instances
• 11:47
  where
• 11:48
  the
• 11:48
  agent
• 11:48
  found
• 11:48
  solutions
• 11:49
  better
• 11:49
  than
• 11:49
  any
• 11:49
  2024
• 11:50
  competitor,
• 11:51
  plus
• 11:51
  one
• 11:51
  novel
• 11:52
  solve
• 11:52
  with
• 11:52
  no
• 11:52
  known
• 11:53
  prior
• 11:53
  solution.
• 11:54
  Techniques
• 11:54
  discovered
• 11:55
  autonomously
• 11:55
  include
• 11:56
  core-guided
• 11:56
  search,
• 11:57
  clause-weighting
• 11:58
  local
• 11:58
  search,
• 11:58
  and
• 11:59
  alternating
• 11:59
  CWLS
• 12:00
  +
• 12:00
  WalkSAT
• 12:00
  phases
• 12:01
  .
• 12:01
  In
• 12:02
  the
• 12:02
  image
• 12:02
  generation
• 12:03
  domain,
• 12:03
  Curriculum-DPO++
• 12:05
  from
• 12:05
  the
• 12:05
  University
• 12:06
  of
• 12:06
  Bucharest
• 12:07
  extends
• 12:07
  DPO
• 12:08
  by
• 12:08
  combining
• 12:08
  data-level
• 12:09
  curriculum
• 12:09
  (easy-to-hard
• 12:10
  preference
• 12:10
  pairs)
• 12:11
  with
• 12:11
  a
• 12:11
  model-level
• 12:12
  curriculum
• 12:12
  that
• 12:12
  progressively
• 12:13
  unfreezes
• 12:13
  layers
• 12:14
  and
• 12:14
  increases
• 12:14
  LoRA
• 12:15
  rank
• 12:15
  —
• 12:16
  preventing
• 12:16
  the
• 12:16
  model
• 12:16
  from
• 12:17
  overfitting
• 12:17
  on
• 12:17
  easy
• 12:18
  examples
• 12:18
  early
• 12:19
  in
• 12:19
  training
• 12:19
  .
• 12:22
  Eon
• 12:23
  Systems
• 12:23
  published
• 12:23
  a
• 12:23
  detailed
• 12:24
  technical
• 12:24
  write-up
• 12:25
  of
• 12:25
  their
• 12:25
  virtual
• 12:25
  embodied
• 12:26
  fly
• 12:26
  —
• 12:27
  a
• 12:27
  140,000-neuron,
• 12:28
  50-million-synapse
• 12:30
  leaky
• 12:30
  integrate-and-fire
• 12:31
  brain
• 12:31
  model,
• 12:32
  built
• 12:32
  from
• 12:33
  the
• 12:33
  adult
• 12:33
  Drosophila
• 12:33
  connectome,
• 12:34
  driving
• 12:34
  a
• 12:34
  physically
• 12:35
  simulated
• 12:35
  NeuroMechFly
• 12:36
  body
• 12:36
  through
• 12:37
  feeding,
• 12:37
  grooming,
• 12:38
  and
• 12:38
  foraging
• 12:38
  behaviors
• 12:39
  in
• 12:39
  a
• 12:39
  MuJoCo
• 12:39
  physics
• 12:40
  environment.
• 12:41
  The
• 12:41
  brain
• 12:41
  model
• 12:41
  comes
• 12:42
  from
• 12:42
  Shiu
• 12:42
  et
• 12:42
  al.,
• 12:43
  the
• 12:43
  visual
• 12:43
  system
• 12:44
  from
• 12:44
  Lappalainen
• 12:44
  et
• 12:44
  al.'s
• 12:45
  connectome-constrained
• 12:46
  recurrent
• 12:47
  network
• 12:47
  for
• 12:47
  64
• 12:48
  visual
• 12:48
  cell
• 12:48
  types,
• 12:49
  and
• 12:49
  the
• 12:49
  body
• 12:49
  from
• 12:50
  NeuroMechFly
• 12:50
  v2
• 12:51
  with
• 12:51
  87
• 12:52
  independent
• 12:52
  joints.
• 12:53
  The
• 12:53
  sensory-motor
• 12:54
  loop
• 12:54
  has
• 12:54
  four
• 12:54
  parts:
• 12:55
  virtual
• 12:56
  sensory
• 12:56
  events
• 12:56
  map
• 12:57
  onto
• 12:57
  identified
• 12:57
  sensory
• 12:58
  neurons,
• 12:59
  brain
• 12:59
  activity
• 12:59
  updates
• 13:00
  in
• 13:00
  the
• 13:00
  connectome
• 13:00
  model,
• 13:01
  selected
• 13:01
  descending
• 13:02
  neuron
• 13:02
  outputs
• 13:03
  translate
• 13:03
  into
• 13:03
  motor
• 13:03
  commands,
• 13:04
  and
• 13:04
  resulting
• 13:05
  movement
• 13:05
  changes
• 13:06
  sensory
• 13:06
  state
• 13:06
  for
• 13:06
  the
• 13:06
  next
• 13:07
  cycle.
• 13:07
  The
• 13:07
  honest
• 13:08
  limitations
• 13:08
  are
• 13:08
  what
• 13:09
  make
• 13:09
  this
• 13:09
  piece
• 13:09
  valuable.
• 13:10
  The
• 13:10
  current
• 13:11
  descending-neuron
• 13:11
  interface
• 13:12
  is
• 13:12
  sparse
• 13:14
  —
• 13:15
  DNa01/DNa02
• 13:17
  for
• 13:17
  steering,
• 13:18
  oDN1
• 13:18
  for
• 13:18
  forward
• 13:19
  velocity,
• 13:20
  antennal
• 13:20
  descending
• 13:21
  neurons
• 13:21
  for
• 13:21
  grooming
• 13:21
  —
• 13:22
  but
• 13:22
  doesn't
• 13:22
  span
• 13:23
  the
• 13:23
  full
• 13:23
  repertoire
• 13:23
  of
• 13:23
  the
• 13:23
  fly's
• 13:24
  ~1,000
• 13:24
  descending
• 13:25
  neurons.
• 13:26
  Brain-body
• 13:26
  coupling
• 13:27
  is
• 13:27
  the
• 13:27
  hard
• 13:27
  engineering
• 13:28
  problem:
• 13:28
  how
• 13:29
  firing
• 13:29
  rates
• 13:29
  in
• 13:29
  specific
• 13:30
  descending
• 13:30
  neurons
• 13:31
  should
• 13:31
  map
• 13:31
  to
• 13:31
  joint
• 13:31
  torques
• 13:32
  is
• 13:32
  still
• 13:32
  approximated
• 13:33
  by
• 13:33
  hand.
• 13:34
  Internal
• 13:35
  state,
• 13:35
  plasticity,
• 13:36
  learning,
• 13:37
  hormonal
• 13:37
  modulation
• 13:37
  —
• 13:38
  all
• 13:39
  absent.
• 13:40
  But
• 13:40
  as
• 13:40
  a
• 13:40
  research
• 13:40
  platform
• 13:41
  for
• 13:41
  connectome-constrained
• 13:42
  sensorimotor
• 13:43
  control,
• 13:43
  it
• 13:44
  makes
• 13:44
  the
• 13:44
  problem
• 13:44
  concrete
• 13:45
  enough
• 13:45
  to
• 13:45
  improve.
• 13:47
  LeRobot
• 13:47
  v0.5.0,
• 13:50
  Hugging
• 13:50
  Face's
• 13:51
  open-source
• 13:51
  robotics
• 13:52
  framework,
• 13:52
  made
• 13:53
  its
• 13:53
  biggest
• 13:53
  release
• 13:53
  with
• 13:53
  200+
• 13:54
  merged
• 13:55
  PRs
• 13:55
  and
• 13:55
  50+
• 13:56
  new
• 13:56
  contributors.
• 13:58
  The
• 13:58
  headline:
• 13:59
  full
• 13:59
  Unitree
• 13:59
  G1
• 14:00
  humanoid
• 14:00
  support
• 14:01
  including
• 14:01
  locomotion,
• 14:02
  manipulation,
• 14:03
  teleoperation,
• 14:04
  and
• 14:04
  whole-body
• 14:04
  control.
• 14:05
  New
• 14:06
  policies
• 14:06
  include
• 14:07
  Pi0-FAST
• 14:09
  (autoregressive
• 14:11
  VLAs
• 14:12
  with
• 14:12
  Frequency-space
• 14:12
  Action
• 14:13
  Sequence
• 14:13
  Tokenization),
• 14:14
  Real-Time
• 14:15
  Chunking
• 14:15
  for
• 14:15
  responsive
• 14:16
  inference,
• 14:17
  and
• 14:17
  LoRA
• 14:17
  fine-tuning
• 14:18
  for
• 14:18
  large
• 14:18
  VLAs.
• 14:19
  The
• 14:19
  dataset
• 14:20
  pipeline
• 14:20
  now
• 14:20
  supports
• 14:21
  streaming
• 14:21
  video
• 14:22
  encoding
• 14:22
  with
• 14:22
  zero
• 14:22
  wait
• 14:23
  time
• 14:23
  between
• 14:23
  recording
• 14:24
  episodes,
• 14:24
  10x
• 14:25
  faster
• 14:25
  image
• 14:26
  training,
• 14:26
  and
• 14:27
  3x
• 14:27
  faster
• 14:27
  encoding.
• 14:28
  EnvHub
• 14:29
  lets
• 14:29
  users
• 14:29
  load
• 14:30
  simulation
• 14:30
  environments
• 14:31
  directly
• 14:31
  from
• 14:31
  the
• 14:31
  Hugging
• 14:32
  Face
• 14:32
  Hub.
• 14:33
  The
• 14:33
  codebase
• 14:33
  moved
• 14:33
  to
• 14:33
  Python
• 14:34
  3.12
• 14:35
  and
• 14:35
  Transformers
• 14:36
  v5,
• 14:36
  and
• 14:37
  the
• 14:37
  paper
• 14:37
  was
• 14:37
  accepted
• 14:37
  to
• 14:38
  ICLR
• 14:38
  2026
• 14:39
  .
• 14:41
  US
• 14:41
  authorities
• 14:42
  arrested
• 14:42
  Supermicro's
• 14:43
  co-founder
• 14:44
  for
• 14:44
  allegedly
• 14:44
  running
• 14:44
  a
• 14:45
  $2.5
• 14:45
  billion
• 14:46
  GPU
• 14:47
  smuggling
• 14:47
  ring,
• 14:48
  using
• 14:48
  fake
• 14:48
  documents,
• 14:49
  dummy
• 14:49
  servers,
• 14:50
  and
• 14:50
  front
• 14:50
  companies
• 14:50
  in
• 14:50
  Southeast
• 14:51
  Asia
• 14:51
  to
• 14:52
  illegally
• 14:52
  export
• 14:53
  restricted
• 14:53
  NVIDIA
• 14:53
  AI
• 14:54
  chips
• 14:54
  to
• 14:54
  China.
• 14:55
  The
• 14:55
  indictment
• 14:55
  names
• 14:56
  other
• 14:56
  employees
• 14:56
  as
• 14:56
  co-conspirators
• 14:57
  .
• 14:58
  At
• 14:58
  $2.5
• 14:58
  billion,
• 14:59
  this
• 14:59
  would
• 14:59
  be
• 15:00
  the
• 15:00
  largest
• 15:00
  known
• 15:00
  case
• 15:01
  of
• 15:01
  GPU
• 15:01
  export
• 15:02
  control
• 15:02
  circumvention,
• 15:03
  dwarfing
• 15:03
  previous
• 15:03
  incidents.
• 15:04
  The
• 15:04
  implications
• 15:05
  for
• 15:05
  supply
• 15:05
  chain
• 15:05
  integrity
• 15:06
  are
• 15:06
  significant
• 15:06
  —
• 15:07
  if
• 15:07
  a
• 15:07
  major
• 15:08
  server
• 15:08
  manufacturer's
• 15:08
  co-founder
• 15:09
  was
• 15:09
  allegedly
• 15:10
  facilitating
• 15:10
  systematic
• 15:11
  evasion,
• 15:11
  the
• 15:12
  question
• 15:12
  becomes
• 15:12
  how
• 15:12
  many
• 15:12
  other
• 15:13
  channels
• 15:13
  remain
• 15:13
  undetected.
• 15:14
  On
• 15:14
  the
• 15:15
  intellectual
• 15:15
  property
• 15:15
  front,
• 15:16
  Britannica
• 15:16
  and
• 15:16
  Merriam-Webster
• 15:17
  filed
• 15:17
  suit
• 15:18
  against
• 15:18
  OpenAI
• 15:18
  in
• 15:18
  the
• 15:19
  Southern
• 15:19
  District
• 15:19
  of
• 15:19
  New
• 15:19
  York,
• 15:20
  alleging
• 15:20
  that
• 15:20
  ChatGPT
• 15:21
  was
• 15:21
  trained
• 15:22
  on
• 15:22
  their
• 15:22
  researched
• 15:22
  and
• 15:22
  fact-checked
• 15:23
  content,
• 15:24
  then
• 15:24
  delivers
• 15:24
  polished
• 15:25
  answers
• 15:25
  that
• 15:25
  cannibalizes
• 15:26
  the
• 15:26
  traffic
• 15:26
  and
• 15:26
  ad
• 15:26
  revenue
• 15:27
  publishers
• 15:27
  depend
• 15:27
  on.
• 15:28
  "ChatGPT
• 15:29
  starves
• 15:29
  web
• 15:29
  publishers,
• 15:30
  like
• 15:30
  Plaintiffs,
• 15:31
  of
• 15:31
  revenue,"
• 15:31
  the
• 15:31
  complaint
• 15:32
  reads.
• 15:33
  The
• 15:33
  core
• 15:33
  argument:
• 15:34
  where
• 15:34
  a
• 15:34
  search
• 15:34
  engine
• 15:35
  sends
• 15:35
  users
• 15:36
  to
• 15:36
  a
• 15:36
  publisher's
• 15:36
  website,
• 15:37
  ChatGPT
• 15:38
  absorbs
• 15:38
  content
• 15:39
  and
• 15:39
  short-circuits
• 15:39
  the
• 15:39
  visit
• 15:40
  entirely
• 15:40
  .
• 15:41
  The
• 15:41
  counter-argument
• 15:42
  from
• 15:42
  commenters
• 15:42
  is
• 15:43
  sharp
• 15:43
  —
• 15:44
  most
• 15:44
  people
• 15:44
  stopped
• 15:44
  using
• 15:45
  physical
• 15:45
  dictionaries
• 15:46
  when
• 15:46
  Google
• 15:46
  arrived
• 15:46
  —
• 15:47
  but
• 15:47
  the
• 15:47
  underlying
• 15:48
  question
• 15:48
  of
• 15:48
  who
• 15:48
  compensates
• 15:49
  the
• 15:49
  content
• 15:49
  producers
• 15:50
  that
• 15:50
  make
• 15:50
  AI
• 15:50
  outputs
• 15:51
  accurate
• 15:51
  is
• 15:51
  genuinely
• 15:52
  unresolved.
• 15:53
  Vercel
• 15:54
  quietly
• 15:54
  updated
• 15:55
  its
• 15:55
  terms
• 15:55
  to
• 15:55
  default
• 15:56
  free
• 15:56
  and
• 15:56
  hobby
• 15:57
  plan
• 15:57
  users
• 15:57
  into
• 15:57
  model
• 15:58
  training
• 15:58
  on
• 15:58
  their
• 15:58
  code,
• 15:59
  with
• 15:59
  a
• 15:59
  10-day
• 16:00
  opt-out
• 16:00
  window.
• 16:01
  The
• 16:01
  community
• 16:01
  reaction
• 16:02
  was
• 16:02
  predictable
• 16:02
  and
• 16:03
  correct:
• 16:03
  if
• 16:03
  it's
• 16:04
  not
• 16:04
  on
• 16:04
  your
• 16:04
  machine,
• 16:05
  it's
• 16:05
  not
• 16:05
  yours
• 16:05
  anymore.
• 16:06
  The
• 16:06
  practical
• 16:07
  advice:
• 16:07
  treat
• 16:07
  infrastructure
• 16:08
  code,
• 16:09
  business
• 16:09
  logic,
• 16:10
  and
• 16:10
  anything
• 16:10
  with
• 16:10
  customer
• 16:11
  data
• 16:11
  as
• 16:11
  off-limits
• 16:12
  for
• 16:12
  cloud
• 16:12
  AI
• 16:12
  tools;
• 16:13
  use
• 16:13
  local
• 16:14
  models
• 16:14
  for
• 16:14
  sensitive
• 16:15
  work.
• 16:16
  Data
• 16:16
  sovereignty
• 16:17
  is
• 16:17
  becoming
• 16:17
  a
• 16:17
  procurement
• 16:18
  filter
• 16:18
  for
• 16:18
  enterprise
• 16:18
  AI
• 16:19
  tools
• 16:19
  in
• 16:19
  2026
• 16:20
  .
• 16:21
  Meanwhile,
• 16:22
  a
• 16:22
  Reddit
• 16:22
  summary
• 16:23
  of
• 16:23
  Anthropic's
• 16:23
  output
• 16:24
  over
• 16:24
  the
• 16:24
  past
• 16:25
  70
• 16:25
  days
• 16:25
  catalogues
• 16:26
  a
• 16:26
  remarkable
• 16:26
  shipping
• 16:27
  pace:
• 16:28
  Claude
• 16:28
  Cowork,
• 16:29
  Opus
• 16:29
  4.6,
• 16:30
  Sonnet
• 16:30
  4.6,
• 16:31
  Haiku
• 16:32
  4.5,
• 16:33
  Claude
• 16:33
  Code
• 16:33
  security
• 16:34
  and
• 16:34
  review
• 16:34
  workflows,
• 16:35
  voice
• 16:35
  mode,
• 16:36
  Excel
• 16:36
  and
• 16:36
  PowerPoint
• 16:37
  integrations,
• 16:38
  a
• 16:38
  marketplace,
• 16:39
  Skills
• 16:39
  API,
• 16:40
  1M
• 16:41
  context
• 16:41
  window,
• 16:41
  and
• 16:42
  a
• 16:42
  $30B
• 16:43
  Series
• 16:43
  G
• 16:43
  at
• 16:43
  $380B
• 16:45
  valuation.
• 16:46
  The
• 16:46
  Claude
• 16:46
  Code
• 16:47
  arc
• 16:47
  specifically
• 16:47
  —
• 16:48
  from
• 16:49
  CLI
• 16:49
  wrapper
• 16:49
  to
• 16:49
  parallel
• 16:50
  subagents,
• 16:51
  built-in
• 16:51
  security,
• 16:52
  and
• 16:52
  desktop
• 16:53
  preview,
• 16:53
  all
• 16:53
  inside
• 16:54
  one
• 16:54
  quarter
• 16:54
  —
• 16:55
  is
• 16:55
  forcing
• 16:56
  workflow
• 16:56
  changes
• 16:57
  faster
• 16:57
  than
• 16:57
  most
• 16:57
  teams
• 16:58
  can
• 16:58
  absorb
• 16:58
  .
• 16:59
  In
• 16:59
  a
• 16:59
  quieter
• 17:00
  institutional
• 17:00
  move,
• 17:01
  arXiv
• 17:01
  announced
• 17:02
  its
• 17:02
  independence
• 17:02
  from
• 17:03
  Cornell,
• 17:04
  restructuring
• 17:04
  as
• 17:04
  a
• 17:05
  standalone
• 17:05
  entity
• 17:06
  after
• 17:06
  decades
• 17:06
  as
• 17:07
  a
• 17:07
  Cornell-hosted
• 17:07
  project
• 17:08
  .