July 17, 2026 — Transcript

Full transcript

1. 0:00A massive change is taking place, and
2. 0:01most don't even notice because the
3. 0:03content you consume around AI is there
4. 0:05to keep your attention, not to teach you
5. 0:07anything valuable. A lot of you asked me
6. 0:09to talk about Inkling, and we will
7. 0:11because the real story is about the
8. 0:12future of how we use AI, and the world's
9. 0:15largest hedge fund is way ahead of us.
10. 0:17I'm Gigachad, and we need to talk about
11. 0:19Thinking Machines, literally. Mira
12. 0:21Murati was the former CTO of OpenAI. She
13. 0:24was the architect behind ChatGPT. She
14. 0:27walked out of the company and took some
15. 0:29of the best minds there with her. They
16. 0:31then founded Thinking Machines, which
17. 0:33closed the largest seed round in Silicon
18. 0:36Valley history. They gave us a glimpse
19. 0:38of their strategy by launching a
20. 0:40fine-tuning platform called Tinker.
21. 0:42We'll come back to that one. You might
22. 0:44know that they just dropped Inkling.
23. 0:46It's a large language model,
24. 0:47specifically a mixture of experts
25. 0:49transformer. Yes, Stateside MOE. It has
26. 0:52256 specialist brains inside. Six are
27. 0:56turned on per task, and two generalists
28. 0:58are always on. But, here's where it gets
29. 1:00exciting. During reinforcement learning,
30. 1:02something very curious happened. The
31. 1:05model started compressing its own
32. 1:07reasoning. It dropped the grammatical
33. 1:09overhead and connectives while keeping
34. 1:11the logic intact. They call it chain of
35. 1:14thought condensation. Think of it like
36. 1:16the kids who get really good at math and
37. 1:18stop showing their work. The answers are
38. 1:20still all right. And yes, it fine-tuned
39. 1:23itself using Tinker, which we're about
40. 1:25to talk about. Inkling is Apache 2.0
41. 1:27licensed. It's free to download, modify,
42. 1:30and redistribute. And it was explicitly
43. 1:32trained to resist censorship. This is
44. 1:36separate from safety. You can't use this
45. 1:38to do something bad, but if you want to
46. 1:40discuss something politically sensitive
47. 1:41or highly debated, go for it. Inkling's
48. 1:44not crushing the benchmarks, but that's
49. 1:46fine. Thinking Machines didn't try to
50. 1:48build a master. They built a student.
51. 1:51Inkling is the starting line. It's a
52. 1:53foundation built to be remixed,
53. 1:55fine-tuned, owned. They've already
54. 1:58previewed something bigger, their TML
55. 2:00interaction model. It's a full duplex AI
56. 2:03that listens, sees, and replies
57. 2:05simultaneously in 200 millisecond
58. 2:07chunks. I want to emphasize, no
59. 2:09turn-taking, real-time interruption. You
60. 2:12know when you're talking to an LLM and
61. 2:13then realize you should have said
62. 2:14something, but you have to wait for it
63. 2:16to respond? No more. The real business
64. 2:19is Tinker. It gives us ownership,
65. 2:21control. We build what we want. And if
66. 2:24this doesn't sound important to you, Ray
67. 2:26Dalio's Bridgewater Associates is a
68. 2:28prominent user of Tinker. They used it
69. 2:31to fine-tune a Chinese model, Qin 3, and
70. 2:34they're beating the market with it. And
71. 2:36they told us how they did it in a
72. 2:38published research paper. So, would you
73. 2:40rather have the perfect AI locked behind
74. 2:43a paywall that you can't touch, or
75. 2:45something you can download, customize,
76. 2:47and truly own? Bridgewater Associates
77. 2:50manages over a hundred and fifty billion
78. 2:52dollars. I would second-guess myself if
79. 2:55I didn't have the same answer as them. A
80. 2:57massive chip