1
00:00:00,000 --> 00:00:03,760
Thanks to Jellypod for helping make this
daily show a reality.

2
00:00:03,952 --> 00:00:09,040
Last time we spent our time on the
backend, how fast the model generates tokens.

3
00:00:09,187 --> 00:00:13,920
Today is about the other half of the loop,
which is how fast you can tell the agent

4
00:00:13,984 --> 00:00:15,040
what to do next.

5
00:00:15,333 --> 00:00:22,240
On October 9, OpenAI put Composer
predictions into beta in the Codex desktop app.

6
00:00:22,610 --> 00:00:23,730
Here's the mechanism.

7
00:00:24,390 --> 00:00:29,730
After Codex finishes responding in a
thread, a suggested next message can appear in

8
00:00:29,750 --> 00:00:30,570
the message box.

9
00:00:31,150 --> 00:00:33,410
It's a whole message, not a code snippet.

10
00:00:33,870 --> 00:00:35,610
You press Tab to accept it.

11
00:00:36,150 --> 00:00:41,110
And this is the part to remember:
accepting only puts the text in the box.

12
00:00:41,210 --> 00:00:42,730
It does not send anything.

13
00:00:43,250 --> 00:00:46,770
You read it, edit it, and then you hit
Enter yourself.

14
00:00:47,167 --> 00:00:49,567
That distinction matters more than it
sounds.

15
00:00:49,634 --> 00:00:52,847
Think about what you actually type in a
long agent session.

16
00:00:53,027 --> 00:00:54,527
Run the unit tests.

17
00:00:54,639 --> 00:00:55,887
Show me the diff.

18
00:00:55,994 --> 00:00:57,727
Check the linter output.

19
00:00:57,914 --> 00:01:00,527
Go ahead with the refactor on the second
file.

20
00:01:00,647 --> 00:01:06,287
Those turns are short and predictable, but
each one still costs you a context switch

21
00:01:06,347 --> 00:01:07,407
and a few seconds.

22
00:01:07,567 --> 00:01:11,167
The prediction is trying to take that
typing off your hands,

23
00:01:11,207 --> 00:01:13,647
based on the conversation in the current
thread.

24
00:01:13,987 --> 00:01:17,087
And it's different from inline code
completion in an editor.

25
00:01:17,467 --> 00:01:19,967
It doesn't update word by word as you
type.

26
00:01:20,367 --> 00:01:24,747
It proposes a complete next turn, once,
after the response lands.

27
00:01:25,367 --> 00:01:28,887
If you start typing your own message
instead, you're just writing your own message.

28
00:01:29,327 --> 00:01:33,107
You don't have to wait for the suggestion,
and you can ignore it entirely.

29
00:01:33,500 --> 00:01:37,740
Let's talk about the safety angle, because
this is an agent that can run shell

30
00:01:37,793 --> 00:01:39,500
commands and apply patches.

31
00:01:39,713 --> 00:01:44,220
If Tab sent the message, a wrong guess
could trigger real execution.

32
00:01:44,367 --> 00:01:49,420
Since Tab only fills the composer, the
human review step is still there.

33
00:01:49,567 --> 00:01:52,540
Treat the suggestion like a draft from a
colleague.

34
00:01:52,652 --> 00:01:58,140
Read it before Enter, especially if it
says something like proceed or apply.

35
00:01:58,407 --> 00:01:59,427
Now, who can use it.

36
00:01:59,827 --> 00:02:01,287
This is where people will get tripped up.

37
00:02:01,827 --> 00:02:06,587
During the beta it's for personal ChatGPT
Pro users aged 18 and older,

38
00:02:07,147 --> 00:02:11,127
in all supported regions, in the latest
Codex desktop app.

39
00:02:11,647 --> 00:02:17,967
It works in local threads and SSH threads,
and only with GPT 6 Astra or

40
00:02:18,187 --> 00:02:19,847
GPT 6.1 Sol.

41
00:02:20,507 --> 00:02:24,247
The help center says it isn't available in
Chat on web or desktop,

42
00:02:24,707 --> 00:02:26,707
and not on other plans on the web.

43
00:02:27,267 --> 00:02:31,307
It says that in Work, predictions may
appear in local Work threads.

44
00:02:31,583 --> 00:02:35,743
So if you're on a Business or Enterprise
workspace, or you've picked a different

45
00:02:35,810 --> 00:02:38,063
model, you may simply not see it.

46
00:02:38,189 --> 00:02:43,343
Before you assume something is broken,
check three things from the help center: the

47
00:02:43,423 --> 00:02:47,983
app is up to date, Show predictions is on,
and your account,

48
00:02:48,050 --> 00:02:50,543
model, and thread type are all supported.

49
00:02:50,963 --> 00:02:52,403
Setup is short.

50
00:02:52,663 --> 00:02:54,543
It's on by default for eligible accounts.

51
00:02:55,143 --> 00:03:00,023
Open Settings in the desktop app, go to
General, then the Composer section,

52
00:03:00,203 --> 00:03:02,163
and you'll find Show predictions.

53
00:03:02,783 --> 00:03:06,962
Flip it off if you find the suggestions
distracting, and flip it back on in the same

54
00:03:06,983 --> 00:03:07,243
place.

55
00:03:07,542 --> 00:03:09,502
Next, what it can see.

56
00:03:09,567 --> 00:03:12,662
Predictions use the conversation in your
current thread.

57
00:03:12,758 --> 00:03:17,062
They don't go looking up memories or
pulling from connected apps on their own.

58
00:03:17,242 --> 00:03:22,822
But if memory context or connected app
output was already pulled into the thread

59
00:03:22,872 --> 00:03:25,382
earlier, that can inform the suggestion.

60
00:03:25,562 --> 00:03:28,022
You can use predictions with memory off.

61
00:03:28,235 --> 00:03:34,422
If you want memory on, that's Settings,
Personalization, Enable Codex memories.

62
00:03:34,758 --> 00:03:36,898
On cost, there's good news for the beta.

63
00:03:37,378 --> 00:03:41,318
Generating predictions doesn't count
toward your Codex usage limits and doesn't

64
00:03:41,358 --> 00:03:42,118
consume credits.

65
00:03:42,638 --> 00:03:46,758
But the message you send does, including
an accepted prediction.

66
00:03:47,258 --> 00:03:50,078
So Tab is free, and Enter is normal
billing.

67
00:03:50,375 --> 00:03:51,695
A few caveats.

68
00:03:51,765 --> 00:03:56,375
The composer may skip a prediction on a
given turn, and the help center says that

69
00:03:56,455 --> 00:03:58,615
alone doesn't mean something is wrong.

70
00:03:58,725 --> 00:04:03,495
Predictions can also take several seconds
to show up, and longer in long

71
00:04:03,564 --> 00:04:04,775
conversations.

72
00:04:04,855 --> 00:04:09,495
And they suggest a complete message, so
they won't finish a sentence you've already

73
00:04:09,555 --> 00:04:10,055
started.

74
00:04:10,247 --> 00:04:11,655
Plan around that.

75
00:04:11,835 --> 00:04:15,095
Don't sit and wait for a suggestion, just
type.

76
00:04:15,365 --> 00:04:18,665
I'd add one interpretation, and it's mine,
not OpenAI's.

77
00:04:19,144 --> 00:04:23,164
The suggestion will be most useful when
your next step is obvious from the thread,

78
00:04:23,505 --> 00:04:25,705
like after a test run or a diff.

79
00:04:26,305 --> 00:04:29,865
When the next step is a real design
decision, you'll probably want to write that one

80
00:04:29,905 --> 00:04:30,665
yourself anyway.

81
00:04:31,402 --> 00:04:34,322
The related change shipped a day earlier,
on October 8.

82
00:04:34,862 --> 00:04:39,462
OpenAI rolled out faster steering in Codex
in the ChatGPT desktop app.

83
00:04:39,962 --> 00:04:44,122
When you send a follow up while a task is
running, Codex can respond to your changes

84
00:04:44,202 --> 00:04:44,542
sooner.

85
00:04:45,002 --> 00:04:48,082
You'd use that to correct an approach, add
missing information,

86
00:04:48,542 --> 00:04:50,062
or change direction mid run.

87
00:04:50,333 --> 00:04:52,653
And there's a setting that decides how it
behaves.

88
00:04:52,813 --> 00:04:56,413
Go to Settings, General, Follow up
behavior.

89
00:04:56,613 --> 00:05:01,613
You choose whether a follow up message
steers the current run or waits for the next

90
00:05:01,673 --> 00:05:01,853
one.

91
00:05:02,066 --> 00:05:05,613
If you want to redirect an agent that's
heading the wrong way,

92
00:05:05,661 --> 00:05:06,733
pick steer.

93
00:05:06,893 --> 00:05:10,653
If you'd rather queue thoughts without
interrupting work in progress,

94
00:05:10,699 --> 00:05:11,933
choose the waiting option.

95
00:05:12,109 --> 00:05:17,533
Pair that with predictions and the loop
tightens on both ends: the suggestion fills

96
00:05:17,586 --> 00:05:21,853
in your next message faster, and steering
lets that message land sooner.

97
00:05:22,215 --> 00:05:25,355
A few smaller things worth knowing from
the open source CLI.

98
00:05:25,815 --> 00:05:32,275
Version 0.162.0, released October 8, added
copy support for transcript blocks with

99
00:05:32,315 --> 00:05:33,855
the slash copy command.

100
00:05:34,295 --> 00:05:37,934
Control Insert now copies a selection, and
there's a config key,

101
00:05:38,335 --> 00:05:43,355
tui dot mouse underscore scroll underscore
speed, to tune mouse wheel scrolling.

102
00:05:43,833 --> 00:05:48,713
The same release makes URLs clickable in
approval headers, questions,

103
00:05:48,873 --> 00:05:54,473
MCP prompts, warnings, and banners,
including when a link wraps across lines.

104
00:05:54,617 --> 00:05:59,713
That sounds minor until you're approving
something and need to open the link in it.

105
00:05:59,803 --> 00:06:06,473
And from 0.161.0 on October 7, you can
sign in to an MCP

106
00:06:06,530 --> 00:06:11,513
server from an active terminal session
with slash mcp login,

107
00:06:11,545 --> 00:06:12,473
then the server name.

108
00:06:12,780 --> 00:06:13,960
So here's what to try today.

109
00:06:14,300 --> 00:06:17,440
If you're on a personal Pro account,
update the desktop app,

110
00:06:17,940 --> 00:06:24,740
open a local thread on GPT 6.1 Sol or GPT
6 Astra, and let a turn finish.

111
00:06:25,160 --> 00:06:29,540
When a suggestion appears, press Tab, read
it, change one word,

112
00:06:29,920 --> 00:06:30,460
and send it.

113
00:06:30,980 --> 00:06:35,260
Then check your Follow up behavior
setting, so you know how mid run messages will

114
00:06:35,300 --> 00:06:35,700
behave.

115
00:06:36,000 --> 00:06:40,320
One last habit worth building: always read
before Enter.

116
00:06:40,500 --> 00:06:43,120
The feature is designed so that you can.

117
00:06:43,232 --> 00:06:45,600
That's the whole point of Tab not sending.