Castuner

Frontier Watch

Guerin Green

Independent field analysis of frontier AI — models, benchmarks, regulation — from the Novel Cognition network. Every number sourced; every vendor claim flagged as one. The audio edition of the Frontier Watch briefs.

Podcast website

Listen in your app

Episodes

47 episodes

  1. InferSSD on a 256 GB Mac Studio, Tested: Plain llama.cpp Wrote 3x Faster 6:24
  2. WiCi One, Checked: A 748B AI Off an SSD vs the Mac Studio We Already Run 7:08
  3. AI Agent Costs, Checked: Same Price Per Token, 8x the Bill 8:10
  4. Semantic Caching, Checked: What It Saves, Where It Breaks, and the Threshold That Decides 6:52
  5. Kimi K3.1 Pro: #4 in the World, Called Open, and You Cannot Download It Yet 9:36
  6. Is GEO Dead? I Checked Every Study Used to Bury It 13:52
  7. Linking Two Macs for AI: What EXO and RDMA Actually Do 8:56
  8. Strix: The Open-Source AI Hacker Everyone Describes Wrong 8:37
  9. Neural Networks Are Already Clean? The Researcher Behind Superposition Changes His Mind 10:05
  10. The One Prompting Error You Are Making, and the Research Behind It 5:32
  11. The Wrong Answer (Corrected): Four DGX Sparks for an SEO Agency vs a Leased Mac Studio 6:27
  12. The Wrong Answer: Four DGX Sparks for an SEO Agency vs a Leased Mac Studio 6:46
  13. The Mandarin Slip: Why ChatGPT, Claude and DeepSeek Sometimes Answer in Chinese 6:43
  14. The Three Numbers: Qwen 3.8 in 24 GB, 2.24x Faster, and the 2.1-Point Tax 7:35
  15. Cheap Offense: A $3,000 Test Reached Inside OpenAI 6:49
  16. Where It Falls Over: The Jev Explainer Wave, Checked 8:34
  17. The Five Lies: What SEO Believes About AI Search 12:00
  18. The Harness File: 23 Agent System Prompts, Read Side by Side 8:19
  19. SoL-Pi: An AI Researcher Tuned the Agent Harness and Halved the Tokens 6:05
  20. Four Labs, One Hole: Gemini Is the Fourth AI Model to Hack Real Companies in a Test 6:37
  21. The Yang Claim: Did Escaped AI Agents Plant Self-Replicating Code? 8:16
  22. The Jev File: TypeSafe Claims 193x Faster, and Its Eval Scores Against Two Other Models 9:22
  23. OpenAI's $2,000 Math Breakthroughs Reused Uncredited Work, Mathematicians Say 5:15
  24. DeepSeek V4 Pro More Than Doubled Its Price. The Launch Page Warned You. 4:55
  25. A 153,000-view DeepSeek video, checked: mostly right, one wrong row, one missing hard drive 8:51
  26. There is no Hermes OS: Omarchy is DHH's, Hermes is Nous Research's, and the install wipes your drive 7:30
  27. DeepSeek V4.1 Flash: 552B beats the 1.6T flagship, V4-Pro retired for one day, and the 510 GB question 10:46
  28. The Genie Problem: Schneier at DEF CON 34, Checked and Steel-Manned 12:37
  29. The $22 Million Proof: What a Swarm on Navier-Stokes Does Not Tell You 11:14
  30. GPT-6 Astra: The Word Nobody at OpenAI Said 10:16
  31. Neither Machine Can Be Bought: Halo Station vs the 512GB Mac Studio 8:55
  32. The Control Gap File: Same Survey, Opposite Aggregation Rules 9:16
  33. The Fable 5.1 File: The 130-Point Gain That Is a Version Change 9:19
  34. The Trending Eight: I Opened All Eight License Files 14:36
  35. The Install Was the Easy Part: Auditing an OSINT Console 6:58
  36. The Refusal Number Nobody Divided: Abliterated Qwen 3.8-27B on a Mac 9:03
  37. Ox Alpha: The Stealth Model Nobody Will Claim 9:23
  38. Qwen3.8-27B: The Open-Weight Drop the Community Called Early 8:05
  39. Grok 4.6: One Point From Fable 5, and the Chart That Left Out the Leader 8:48
  40. DeepSeek V4 Pro 0813: Live and Cheap, Verified by No One 9:00
  41. The Unsanctioned Run: A Test Agent Faked a Human to Approve Its Own Code 8:21
  42. Gemma 4: Why It's This Good At This Size 10:52
  43. The Swarm File: OpenAI Agents Built a Message Board 7:50
  44. The Weights File: Qwen3.8-Max Was Called Open-Source 3:58
  45. The Feature That Was Switched Off: DeepSeek on One Mac 11:14
  46. Article 50: The EU AI Act Rule That Went Live 9:33
  47. Ten Proofs: What the $2,000 Number Leaves Out 9:32