WEBVTT

1
00:00:02.156 --> 00:00:08.828
You're listening to Peeter Levvuls clearing a two-week Windows XP block with a nineteen-dollar Kimi K3 switch.

2
00:00:08.934 --> 00:00:11.520
Written by Dee-aygo Ferrahro on Promptway.

3
00:00:12.053 --> 00:00:14.133
Peeter Levvuls wanted to install Yahoo!

4
00:00:14.373 --> 00:00:19.413
Messenger from two thousand three inside a Windows XP desktop running in the browser.

5
00:00:19.493 --> 00:00:22.933
Claude Code kept treating the emulator like a cybersecurity problem.

6
00:00:24.000 --> 00:00:28.413
He spent two weeks bouncing between model fallbacks and safety blocks.

7
00:00:28.453 --> 00:00:32.933
Then he opened X and asked how to run Kimi K3 through a coding agent.

8
00:00:33.013 --> 00:00:35.253
The answer changed his afternoon.

9
00:00:35.333 --> 00:00:45.173
Levels installed OpenCode, connected it directly to Kimi, paid nineteen dollars for a membership, switched the agent into Build mode, and let it work through the simulator's to-do list.

10
00:00:45.253 --> 00:00:51.253
His report after the switch was short: K3 was "absolutely hammering through" the work.

11
00:00:51.333 --> 00:00:55.293
This is a good founder case study because the task stayed recognizable.

12
00:00:55.333 --> 00:00:57.226
A browser emulator was stuck.

13
00:00:57.280 --> 00:00:58.973
The founder changed the stack.

14
00:00:59.013 --> 00:01:00.213
Work resumed.

15
00:01:00.293 --> 00:01:05.120
It is also a messy model comparison, which is where the useful part begins.

16
00:01:05.146 --> 00:01:06.693
The project K3 walked into

17
00:01:06.733 --> 00:01:09.853
The Windows XP simulator sits at pieter.com.

18
00:01:09.893 --> 00:01:19.533
It is the kind of project Levels keeps returning to: old software, browser emulation, a long list of rough edges, and no client waiting for a compliance memo.

19
00:01:19.573 --> 00:01:22.173
He had already been using Claude Code heavily.

20
00:01:22.213 --> 00:01:28.973
In June, he wrote that he had coded almost entirely on a virtual private server with Claude Code for nearly a year.

21
00:01:29.013 --> 00:01:33.693
The switch did not come from a tourist opening two chat tabs and asking for a snake game.

22
00:01:33.733 --> 00:01:38.973
It came from a paying power user who had run into the same refusal pattern for days.

23
00:01:39.013 --> 00:01:41.706
Levels described the immediate problem in public:

24
00:01:41.760 --> 00:01:44.573
"Claude Code couldn't do this for 2 weeks."

25
00:01:44.613 --> 00:01:47.213
His complaint centered on safety fallbacks.

26
00:01:47.253 --> 00:01:54.586
The model treated requests around the Windows XP environment as risky even though Levels was working on a hobby project he controlled.

27
00:01:54.640 --> 00:01:57.733
The task context changes how I read the refusal.

28
00:01:57.813 --> 00:02:02.093
A guardrail can be reasonable in one environment and maddening in another.

29
00:02:02.133 --> 00:02:04.893
Levels was not asking an agent to probe a bank.

30
00:02:04.933 --> 00:02:06.093
He was trying to make Yahoo!

31
00:02:06.133 --> 00:02:08.253
Messenger work in a browser toy.

32
00:02:08.293 --> 00:02:11.173
He changed four variables, not one

33
00:02:11.253 --> 00:02:15.453
The viral version of this story is Kimi K3 beat Claude.

34
00:02:15.493 --> 00:02:18.613
The public record supports a narrower finding.

35
00:02:18.853 --> 00:02:20.893
Layer, Before, After.

36
00:02:20.933 --> 00:02:25.613
Model, Claude models, with reported safety fallbacks, Kimi K3.

37
00:02:25.653 --> 00:02:28.413
Harness, Claude Code, OpenCode.

38
00:02:28.453 --> 00:02:34.213
Provider path, Anthropic through Claude Code, Direct Kimi connection after an OpenRouter rate limit.

39
00:02:34.293 --> 00:02:40.693
Permissions, Claude's policy and tool gates, OpenCode Build mode with permission bypass enabled.

40
00:02:40.933 --> 00:02:46.053
Task, Windows XP simulator to-do list, The same project and backlog.

41
00:02:46.133 --> 00:02:49.413
K3 deserves credit for completing work that had stalled.

42
00:02:49.493 --> 00:02:53.173
Levels also removed several sources of friction around the model.

43
00:02:53.253 --> 00:02:55.733
He published the setup on July 17.

44
00:02:55.813 --> 00:02:57.253
Install OpenCode.

45
00:02:57.333 --> 00:02:58.853
Create a Kimi account.

46
00:02:58.933 --> 00:03:00.453
Pay nineteen dollars.

47
00:03:00.533 --> 00:03:02.053
Get an A P I key.

48
00:03:02.133 --> 00:03:04.613
Connect Kimi Code inside OpenCode.

49
00:03:04.773 --> 00:03:06.053
Switch to Build mode.

50
00:03:06.133 --> 00:03:11.253
His instructions also recommend bypassing permissions, which he already did in Claude Code.

51
00:03:11.713 --> 00:03:15.453
That final setting makes the run faster and less comparable.

52
00:03:15.493 --> 00:03:21.213
A model with broad shell access can finish jobs that a more constrained agent pauses to confirm.

53
00:03:21.253 --> 00:03:24.013
It can also damage more when it guesses wrong.

54
00:03:24.053 --> 00:03:27.173
Levels made a rational trade for a hobby emulator.

55
00:03:27.253 --> 00:03:30.653
I would not copy that trade onto a production database.

56
00:03:30.693 --> 00:03:32.293
Why K3 fit this job

57
00:03:32.853 --> 00:03:39.053
Moonshot AI released Kimi K3 on July 16, one day before Levels published his switch.

58
00:03:39.093 --> 00:03:47.613
The model has 2.8 trillion total parameters in a mixture-of-experts architecture, with 104 billion activated for each token.

59
00:03:47.653 --> 00:03:52.693
It accepts a one-million-token context and handles text, images, and video.

60
00:03:52.733 --> 00:03:55.933
The specifications matter less here than the training target.

61
00:03:55.973 --> 00:04:02.773
Moonshot built K3 for long coding sessions, large repositories, terminal tools, and visual feedback loops.

62
00:04:02.853 --> 00:04:06.253
A browser operating-system simulator touches all four.

63
00:04:06.293 --> 00:04:13.520
Moonshot's own technical post says an early K3 build handled most of the team's kernel-optimization work late in development.

64
00:04:13.546 --> 00:04:19.933
The company also reports a 48-hour autonomous chip-design run and a compiler project built from scratch.

65
00:04:19.973 --> 00:04:25.773
Those are vendor case studies, so I treat the measurements as claims until independent teams reproduce them.

66
00:04:25.813 --> 00:04:28.973
They still show what Moonshot tuned the model to attempt.

67
00:04:29.013 --> 00:04:33.453
Levels supplied an outside example with a public project and a named workflow.

68
00:04:33.493 --> 00:04:39.293
On the same day, he also published a macOS 27 browser interface created with K3.

69
00:04:39.333 --> 00:04:42.666
That does not prove the model wins every frontend task.

70
00:04:42.720 --> 00:04:46.253
It does show the visual coding loop was more than a benchmark row.

71
00:04:46.293 --> 00:04:49.573
The cost moved from abstract to nineteen dollars

72
00:04:49.653 --> 00:04:53.213
Levels first tried OpenRouter and hit an upstream rate limit.

73
00:04:53.253 --> 00:04:59.333
He then went straight to Kimi, bought the nineteen dollars membership, and used the provider's A P I key with OpenCode.

74
00:04:59.413 --> 00:05:11.333
Moonshot prices the K3 A P I separately at thirty cents per million cache-hit input tokens, three dollars per million cache-miss input tokens, and fifteen dollars per million output tokens.

75
00:05:11.573 --> 00:05:17.333
The company says coding workloads on its official A P I exceed a 90 percent cache-hit rate.

76
00:05:17.453 --> 00:05:20.053
That is vendor-reported cache performance.

77
00:05:20.133 --> 00:05:26.933
Your bill depends on the harness, prompt reuse, context size, and how often the agent rewrites its own plan.

78
00:05:27.013 --> 00:05:32.506
K3 always thinks, and long autonomous runs can burn output tokens quickly.

79
00:05:32.560 --> 00:05:35.573
The relevant founder number remains nineteen dollars.

80
00:05:35.653 --> 00:05:40.133
That was cheap enough for Levels to stop arguing with his old setup and try a new route.

81
00:05:40.213 --> 00:05:42.293
Where the case study gets uncomfortable

82
00:05:42.853 --> 00:05:45.253
K3 has its own failure modes.

83
00:05:45.333 --> 00:05:47.893
Moonshot lists three in the release notes.

84
00:05:47.973 --> 00:05:51.413
First, the model expects preserved thinking history.

85
00:05:51.493 --> 00:05:57.333
A harness that drops earlier reasoning, or a mid-session model swap, can make quality unstable.

86
00:05:57.453 --> 00:06:04.213
Moonshot recommends starting K3 in a compatible harness instead of dropping it into a conversation another model began.

87
00:06:04.413 --> 00:06:07.333
Second, the model can act too aggressively.

88
00:06:07.413 --> 00:06:12.213
Moonshot says K3 may make unexpected decisions when intent is ambiguous.

89
00:06:12.253 --> 00:06:17.133
The company recommends explicit constraints in the system prompt or AGENTS.md.

90
00:06:17.173 --> 00:06:23.626
Third, Moonshot concedes that the overall user experience still trails the strongest proprietary models.

91
00:06:23.680 --> 00:06:31.053
A public Kimi Code issue filed after launch also reports the terminal interface hanging during a trivial prompt at maximum effort.

92
00:06:31.093 --> 00:06:33.813
The failure surface changed with the model.

93
00:06:33.973 --> 00:06:40.253
Levels solved one kind of agent friction by choosing a model trained to keep going and a harness configured to let it.

94
00:06:40.293 --> 00:06:45.466
The same combination can turn a vague instruction into a long, expensive mistake.

95
00:06:45.520 --> 00:06:47.173
What I would copy from the switch

96
00:06:47.840 --> 00:06:52.240
I would copy the routing decision and leave the permission bypass behind.

97
00:06:52.266 --> 00:07:00.653
When a coding agent refuses a legitimate task twice, write down the task, the exact refusal, the branch state, and the acceptance test.

98
00:07:00.693 --> 00:07:04.106
Start a fresh session in a second harness with a second model.

99
00:07:04.160 --> 00:07:07.066
Give it the same repository and the same test.

100
00:07:07.200 --> 00:07:11.413
Compare the completed diff, elapsed time, tool calls, and regressions.

101
00:07:11.533 --> 00:07:14.733
Do not continue the old conversation after the switch.

102
00:07:14.773 --> 00:07:18.093
K3's own documentation warns against that path.

103
00:07:18.133 --> 00:07:22.333
Keep the second agent inside a disposable branch and a scoped environment.

104
00:07:22.373 --> 00:07:26.093
Levels can rebuild a browser simulator when an agent gets inventive.

105
00:07:26.133 --> 00:07:27.973
Your billing system is less funny.

106
00:07:28.053 --> 00:07:28.933
The founder lesson

107
00:07:29.333 --> 00:07:32.053
Levels did not wait for a benchmark committee.

108
00:07:32.133 --> 00:07:36.293
He had a blocked task, spent nineteen dollars, and changed the route.

109
00:07:36.333 --> 00:07:44.053
The result is strong evidence that Kimi K3 belongs in the coding-agent rotation for long, visual, tool-heavy work.

110
00:07:44.093 --> 00:07:52.133
It is weak evidence that K3 is categorically better than Claude because the harness, provider, and permissions changed with the model.

111
00:07:52.213 --> 00:07:59.773
The evidence is enough for a useful operator story: a stuck job, a visible change, and a result someone else can test.

112
00:08:00.293 --> 00:08:02.293
Run K3 on a fresh branch.

113
00:08:02.693 --> 00:08:04.773
Keep the acceptance test fixed.

114
00:08:04.853 --> 00:08:07.170
Keep the permissions narrower than Peeter Levvuls did.

115
00:08:08.381 --> 00:08:09.270
That's the story.

116
00:08:09.510 --> 00:08:12.350
You'll find every source linked at the end of the article.

117
00:08:12.710 --> 00:08:14.790
Thanks for listening, on Promptway.
