Not a Llama

Earlier episodes

• Wednesday, October 21

  DeepSeek goes on the offensive: US bans it from federal systems as Washington closes the ring on Chinese AI

  9:52
  Stories and sources

  1. 0:40DeepSeek goes on the offensive: US bans it from federal systems as Washington closes the ring on Chinese AI
     techcrunch.cominterconnects.aibsky.app
  2. 2:20Mistral releases Voxtral 3: a real-time voice-agent SDK for 50+ languages
     mistral.ai
  3. 3:32Anthropic's S-1 lays bare the IPO story: $16.8bn revenue, flatlined growth, and a compute squeeze
     ground.newstechcrunch.com
  4. 4:36DeepSeek DeepInfer: the inference engine itself goes open source
     github.comreddit.com
  5. 5:45AI just proved the quasi-Riemann hypothesis, and humans are finally unpacking it
     reddit.com
  6. 7:04Protein design goes mainstream: Claude helps design scorpion neurotoxin binders
     anthropic.comreddit.com
  7. 8:12OpenRouter price wave: Qwen3.8 27B slashed, Mistral Large 4 doubled — Mistral: Mistral Large 4: input $0.68 → $1.36, output $2.09 → $4.18 per 1M tokens; Qwen: Qwen3.8 27B: input $0.24 → $0.04, output $2.20 → $1.35 per 1M tokens (+2)
     openrouter.ai

  Read the transcript

• Tuesday, October 20

  The October AI model race ignites: GPT-6.1 Terra spotted in a Microsoft eval, with Anthropic, Google, and OpenAI all lined up

  9:05
  Stories and sources

  1. 0:27The October AI model race ignites: GPT-6.1 Terra spotted in a Microsoft eval, with Anthropic, Google, and OpenAI all lined up
     reddit.com
  2. 2:09MiniMax M3.1 Flash announced, weights landing Oct 23
     reddit.com
  3. 3:16Inception's Mercury 3: a 1M-token diffusion LLM that matches GPT-6 Luna — Inception: Mercury 3: $0.10 in / $0.35 out per 1M tokens
     openrouter.aireddit.cominceptionlabs.ai
  4. 4:27CERN partners with Mistral to build the first foundation model for collider physics
     reddit.comhome.cern
  5. 5:43OpenAI rushes a $30B raise ahead of Anthropic's IPO
     ground.news
  6. 6:42OpenRouter price wave: DeepSeek, GLM 5, and Qwen3.5 all shift — DeepSeek: DeepSeek V4.1 Flash: input $0.048 → $0.0352, output $0.12 → $0.0112 per 1M tokens; DeepSeek: DeepSeek V4 Flash 0731: input $0.15 → $0.44, output $0.15 → $1.32 per 1M tokens (+2)
     openrouter.ai
  7. 7:42Watch: Gemini 4 Argon really surprised Two Minute Papers
     youtube.com

  Read the transcript

• Monday, October 19

  DeepSeek V4.1 Flash lands: the open-weight release reshaping the local-inference week

  9:56
  Stories and sources

  1. 0:31DeepSeek V4.1 Flash lands: the open-weight release reshaping the local-inference week
     huggingface.coreddit.com
  2. 2:25DeepSeek's new search-and-fetch API and a fresh OpenRouter price wave — Qwen: Qwen3.5-27B: input $0.26 → $0.20, output $2.60 → $1.56 per 1M tokens; Z.ai: GLM 4.7: input $0.60 → $0.54, output $2.20 → $1.98 per 1M tokens (+2)
     api-docs.deepseek.comopenrouter.ai
  3. 3:37Replit drops Claude for ChatGPT as default
     ground.newstechcrunch.com
  4. 4:48Claude cracks a 25-year NSA cryptography puzzle and designs cheaper quantum circuits
     reddit.com
  5. 6:04Google open-sources Harness2 for self-improving agents
     reddit.com
  6. 7:15Anthropic's Claude powers Pentagon operation against Maduro
     ground.news
  7. 8:26How models actually work: Emmanuel Ameisen breaks down interpretability
     ground.news

  Read the transcript

• Sunday, October 18

  Paper day: Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy Distillation

  9:12
  Stories and sources

  1. 0:39The problem
     arxiv.orghuggingface.cogithub.com
  2. 1:56The idea
  3. 3:22How it works
  4. 5:17What they found
  5. 7:10Caveats and why it matters

  Read the transcript

• Saturday, October 17

  Reflection Beam 1.0: a 501B MoE open-weight reasoning model lands on Hugging Face and OpenRouter — Reflection: Beam 1.0 501B A23B: $1.00 in / $3.00 out per 1M tokens

  10:04
  Stories and sources

  1. 0:38Reflection Beam 1.0: a 501B MoE open-weight reasoning model lands on Hugging Face and OpenRouter — Reflection: Beam 1.0 501B A23B: $1.00 in / $3.00 out per 1M tokens
     reddit.comopenrouter.ai
  2. 2:32ORL: a new diffusion LLM hits 2,788 tokens/s on a single RTX 5090 without sacrificing quality
     reddit.com
  3. 4:06Elorian launches Glimpse: an AI model built to reason with images, not just about them
     techcrunch.com
  4. 5:30Hybrid Reasoning Agents: think in a hidden internal language, spend tokens only on tools and answers
     reddit.com
  5. 6:46China proposes export controls on 23 AI technologies, but already-open-sourced models are exempted
     reddit.comground.news
  6. 8:21OpenRouter price wave: DeepSeek, Kimi, Llama, and NVIDIA Nemotron all shift — DeepSeek: DeepSeek Pro Latest: input $0.23 → $0.25, output $8.00 → $0.75 per 1M tokens; DeepSeek: DeepSeek Flash Latest: input $0.0115 → $0.032, output $0.60 → $0.14 per 1M tokens (+6)
     openrouter.ai

  Read the transcript

• Friday, October 16

  Meta's Muse Spark Open: the US open-weight model that finally closes the gap on Chinese labs — Meta: Muse Spark Open: $1.25 in / $4.25 out per 1M tokens

  10:02
  Stories and sources

  1. 0:38Meta's Muse Spark Open: the US open-weight model that finally closes the gap on Chinese labs — Meta: Muse Spark Open: $1.25 in / $4.25 out per 1M tokens
     techcrunch.comtheverge.comresearch.meta.aireddit.comopenrouter.aibsky.app
  2. 3:00NVIDIA's local-inference wave: Nemotron 3.5 Super VL and Bolt 74B land on OpenRouter — NVIDIA: Nemotron 3.5 Super VL: $0.15 in / $0.50 out per 1M tokens
     reddit.comopenrouter.ai
  3. 4:35Gemini 4 Argon posts Google's closest ECI score yet, still trailing the frontier
     reddit.com
  4. 6:15Anthropic's IPO week: $16.8bn revenue, IPO-delay pressure, and METR access
     simonwillison.netbsky.appground.newstechcrunch.com
  5. 7:54FTC subpoenas OpenAI and Anthropic over AI safety incidents
     ground.news

  Read the transcript

• Thursday, October 15

  NVIDIA Nemotron Bolt 3.5 74B: a 24GB quantized MoE built for fast local inference

  8:39
  Stories and sources

  1. 0:27NVIDIA Nemotron Bolt 3.5 74B: a 24GB quantized MoE built for fast local inference
     reddit.com
  2. 2:58OpenRouter price wave: DeepSeek models slashed, GPT-5.6 Sol Pro doubles in price — DeepSeek: DeepSeek Flash Latest: input $0.0396 → $0.0158 per 1M tokens; DeepSeek: DeepSeek V4.1 Flash: input $0.30 → $0.0158, output $1.20 → $0.89 per 1M tokens (+1)
     openrouter.ai
  3. 4:29GPT-6 Astra solves 49 unpublished physics problems, catching errors by world-class researchers
     reddit.com
  4. 6:22OpenAI scrambling to make next flagship safe after canceling a weaker model over safety concerns
     reddit.com

  Read the transcript

• Wednesday, October 14

  Google finally launches Gemini 4 Argon, its long-delayed flagship reasoning model — Google: Gemini 4 Argon: $2.00 in / $10.00 out per 1M tokens

  9:43
  Stories and sources

  1. 0:30Google finally launches Gemini 4 Argon, its long-delayed flagship reasoning model — Google: Gemini 4 Argon: $2.00 in / $10.00 out per 1M tokens
     deepmind.googletheverge.comground.newsopenrouter.aireddit.com
  2. 2:14LiquidAI's LFM2.5-24B-A2B: the largest final open-weight MoE built for fast, private on-device agents — LiquidAI: LFM2.5-24B-A2B: $0.05 in / $0.70 out per 1M tokens
     huggingface.coreddit.comopenrouter.ai
  3. 3:14Anthropic warns small models may be months from self-replicating as AI worms on insecure IoT devices
     arxiv.orgreddit.comground.newsbsky.app
  4. 4:14OpenAI's video and image models stretch toward voice filmmaking
     openai.com
  5. 5:13The local-inference scene: Qwen on a PS5 GPU, Tencent's local Agent Memory, and the RTX 5090 squeeze
     reddit.com
  6. 6:13OpenRouter price wave: DeepSeek Pro and Kimi Latest shift — DeepSeek: DeepSeek Pro Latest: input $0.25 → $0.23, output $0.75 → $8.00 per 1M tokens; MoonshotAI: Kimi Latest: input $0.73 → $0.72, output $10.00 → $24.00 per 1M tokens
     openrouter.ai
  7. 7:09Industry: Character.AI files for bankruptcy while Epic leans into AI-accelerated game development
     techcrunch.comtheverge.comreddit.com
  8. 8:17Watch: Gemini 4 Argon teaches a humanoid robot to play cornhole from one video
     reddit.com

  Read the transcript

• Tuesday, October 13

  Meta's Muse Wonder arrives: its first genuine frontier challenge to Anthropic and OpenAI

  9:46
  Stories and sources

  1. 0:32Meta's Muse Wonder arrives: its first genuine frontier challenge to Anthropic and OpenAI
     techcrunch.comtheverge.comground.newsreddit.combsky.app
  2. 2:40The numbers that decide whether Muse Wonder is real — Meta: Muse Wonder: $4.00 in / $20.00 out per 1M tokens
     openrouter.aireddit.com
  3. 3:59Matthew Berman breaks down the Muse Wonder launch
     youtube.com
  4. 5:17Open-weights surge: Decision 3.0, MiniCPM-V 4.7, and GLM-Legal
     reddit.comz.ai
  5. 6:32A fresh wave of OpenRouter price changes — DeepSeek: DeepSeek Pro Latest: input $0.21 → $0.25, output $5.00 → $0.75 per 1M tokens; DeepSeek: DeepSeek Flash Latest: input $0.0336 → $0.0441, output $1.19 → $0.89 per 1M tokens (+10)
     openrouter.ai
  6. 7:51Microsoft's Maia 300 chip aims to cut Nvidia dependence
     theverge.com

  Read the transcript

• Monday, October 12

  Sakana AI demonstrates a working recursive self-improvement loop

  10:13
  Stories and sources

  1. 0:34Sakana AI demonstrates a working recursive self-improvement loop
     reddit.com
  2. 2:17Terence Tao on 'Math 2.0' and the AI-assisted proof revolution
     teorth.github.io
  3. 3:18AI agents can now systematically decompile and reverse-engineer closed-source software
     momo5502.comreddit.com
  4. 4:28Open-weight releases: AesCode 8B/32B, Whistle 16.9MB STT, OpenMed 3.0
     reddit.com
  5. 5:45OpenRouter price wave: DeepSeek, GLM, Kimi, and NVIDIA models shift — DeepSeek: DeepSeek Pro Latest: input $0.22 → $0.21, output $0.66 → $5.00 per 1M tokens; DeepSeek: DeepSeek Flash Latest: input $0.0001 → $0.0336, output $0.60 → $1.19 per 1M tokens (+9)
     openrouter.ai
  6. 6:43DeepSeek V4.1 Flash flagged for API key exposure in agent sandboxes
     reddit.com
  7. 7:44The widening gap in how people understand LLM capabilities
     twitter.comreddit.com

  Read the transcript

• Sunday, October 11

  Paper day: LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches

  9:01
  Stories and sources

  1. 0:33The problem
     arxiv.orghuggingface.co
  2. 1:29The idea
  3. 3:12How it works
  4. 5:32What they found
  5. 7:14Caveats and why it matters

  Read the transcript

• Sunday, October 11

  OpenRouter rolls out a fresh wave of price cuts on DeepSeek, Qwen, and GLM — DeepSeek: DeepSeek Pro Latest: input $0.30 → $0.22, output $5.00 → $0.66 per 1M tokens; Qwen: Qwen3.8 27B: input $0.42 → $0.04, output $2.55 → $2.30 per 1M tokens (+5)

  9:19
  Stories and sources

  1. 0:27OpenRouter rolls out a fresh wave of price cuts on DeepSeek, Qwen, and GLM — DeepSeek: DeepSeek Pro Latest: input $0.30 → $0.22, output $5.00 → $0.66 per 1M tokens; Qwen: Qwen3.8 27B: input $0.42 → $0.04, output $2.55 → $2.30 per 1M tokens (+5)
     openrouter.ai
  2. 2:28A Fields Medalist on OpenAI's math claim: 'I expected to be surpassed... but not like this'
     reddit.com
  3. 3:58Nvidia reportedly halts RTX 5090 production, tightening the local-inference GPU squeeze
     reddit.com
  4. 5:24Nvidia in talks to acquire open-model startup Reflection AI
     ft.com
  5. 5:51Nadella: assume all frontier models are 'compromised' and build an AI emergency brake
     techcrunch.comtheverge.com
  6. 7:25Qwen3.8-27B on a single 3090: a custom CUDA megakernel hits 140 tok/s on code
     reddit.com

  Read the transcript

• Saturday, October 10

  GLM 5.3 Flash tops the Cyber Index, edging out every Anthropic model

  9:50
  Stories and sources

  1. 0:27GLM 5.3 Flash tops the Cyber Index, edging out every Anthropic model
     reddit.com
  2. 2:08Qwen-Image-2.1-Turbo: 8-step image generation, open weights
     reddit.com
  3. 3:36OpenRouter rolls out a wave of price cuts on DeepSeek and GLM — DeepSeek: DeepSeek Pro Latest: input $0.18 → $0.30, output $8.00 → $5.00 per 1M tokens; DeepSeek: DeepSeek Flash Latest: output $1.20 → $0.60 per 1M tokens (+6)
     openrouter.ai
  4. 5:03Anthropic's agents keep leaking into the real world
     techcrunch.comtheverge.comground.news
  5. 6:39Asana slashes agent costs 76x with GPT-6.1 Sol
     openai.com
  6. 8:05Opus 5.5 builds a playable game in days (and still can't crack the dirt bike)
     reddit.com

  Read the transcript

• Friday, October 9

  Anthropic ships Claude Haiku 5.5, its cheapest and fastest small model yet

  9:37
  Stories and sources

  1. 0:34Anthropic ships Claude Haiku 5.5, its cheapest and fastest small model yet
     latent.space
  2. 2:21StepFun's Step 5 Preview: a 1M-context MoE now on OpenRouter
     openrouter.ai
  3. 3:52Decision models go head-to-head: which runs fastest on an RTX 4090?
     reddit.com
  4. 5:15Google launches offline-first AI Edge Foresight note-taker
     theverge.comtechcrunch.com
  5. 6:29Google makes Gemini an agentic assistant for work
     techcrunch.comtheverge.com
  6. 7:47OpenRouter shakes up pricing on Mistral and DeepSeek models
     openrouter.ai

  Read the transcript

• Thursday, October 8

  Anthropic ships Claude Haiku 5.5: the cheapest, fastest small model it has ever released

  10:51
  Stories and sources

  1. Anthropic ships Claude Haiku 5.5: the cheapest, fastest small model it has ever released
     simonwillison.netanthropic.comreddit.comopenrouter.ai
  2. OpenAI's math breakthrough: 722 papers solving 90 of the top 500 open problems
     latent.spacereddit.com
  3. ChatGPT gets an 'Intelligent UI' with interactive visuals, rolling out with GPT-6
     theverge.com
  4. LiquidAI drops its first open LFM family for the edge
     reddit.comhuggingface.co
  5. Mistral 'Le Chonk' ships on a 3800 GB300 cluster, funded by its Series D
     youtube.com
  6. Benchmark: Haiku 5.5 vs GPT-6 Luna on the voxel pagoda test
     reddit.com
  7. llama.cpp adds a GPU cache for MoE experts held in host memory
     reddit.com

  Read the transcript