WEBVTT

1
00:00:00.080 --> 00:00:05.440
<v Speaker 1>AI Daily Briefing. I'm ed sharp, thanks for joining me today.

2
00:00:05.759 --> 00:00:10.439
<v Speaker 1>Space debris crisis gets commercial fix. The Pentagon just handed

3
00:00:10.439 --> 00:00:14.240
<v Speaker 1>commercial space companies eight point four million dollars to solve

4
00:00:14.279 --> 00:00:18.120
<v Speaker 1>one of the most urgent infrastructure problems in orbit. Firefly,

5
00:00:18.440 --> 00:00:23.239
<v Speaker 1>Deorbit and Catalyst each received contracts to design spacecraft that

6
00:00:23.320 --> 00:00:28.320
<v Speaker 1>can autonomously deorbit unprepared satellites, satellites that weren't built with

7
00:00:28.359 --> 00:00:32.640
<v Speaker 1>any removal mechanism. That's the hard part, and it's never

8
00:00:32.719 --> 00:00:37.159
<v Speaker 1>actually been demonstrated. Here's what matters about this move. The

9
00:00:37.200 --> 00:00:41.000
<v Speaker 1>Pentagon isn't building a government fleet. It's buying a service

10
00:00:41.439 --> 00:00:45.399
<v Speaker 1>paper deorbit, not own and operate. That's a deliberate shift

11
00:00:45.399 --> 00:00:50.079
<v Speaker 1>of operational risk and capital cost onto private contractors. If

12
00:00:50.119 --> 00:00:54.280
<v Speaker 1>this works, it validates an entire commercial space servicing industry

13
00:00:54.640 --> 00:00:58.960
<v Speaker 1>estimated at three trillion dollars. If the autonomous capture fails,

14
00:00:59.320 --> 00:01:02.560
<v Speaker 1>the debris claw keeps running. Right now, it's at two

15
00:01:02.640 --> 00:01:06.920
<v Speaker 1>point eight days between significant conjunction events in low Earth orbit.

16
00:01:07.480 --> 00:01:11.040
<v Speaker 1>That's not a slow moving problem. Shifting to AI infrastructure,

17
00:01:11.480 --> 00:01:14.799
<v Speaker 1>Open AI shipped the GPT five point six family free

18
00:01:14.879 --> 00:01:19.439
<v Speaker 1>variants soul Terror and Lunar the headline isn't raw capability.

19
00:01:19.719 --> 00:01:24.040
<v Speaker 1>Its economics soul now matches GPT five point five extra

20
00:01:24.120 --> 00:01:27.560
<v Speaker 1>high performance at a cost drop from thirty three dollars

21
00:01:27.640 --> 00:01:30.599
<v Speaker 1>to one dollar and thirty three cents per million. Tokens,

22
00:01:31.200 --> 00:01:36.040
<v Speaker 1>persistent reasoning, multi agent orchestration, and programmatic tool calling are

23
00:01:36.079 --> 00:01:40.239
<v Speaker 1>all built in on financial research tasks. That combination is

24
00:01:40.280 --> 00:01:43.959
<v Speaker 1>cutting token costs by around twenty one percent. The signal

25
00:01:44.000 --> 00:01:46.959
<v Speaker 1>here is that model selection is changing. It's no longer

26
00:01:47.079 --> 00:01:51.560
<v Speaker 1>default a flagship decision. Hybrid routing small plus media models

27
00:01:51.560 --> 00:01:54.680
<v Speaker 1>for the right task is now cost competitive with yesterday's

28
00:01:54.719 --> 00:01:58.920
<v Speaker 1>premium tier. That's an economics inversion, and enterprise teams will

29
00:01:58.920 --> 00:02:03.040
<v Speaker 1>notice quickly. Meanwhile, Ali Barber's q N three point eight

30
00:02:03.400 --> 00:02:07.640
<v Speaker 1>twenty seven billion parameter dense model has outperformed clawed opu's

31
00:02:07.680 --> 00:02:10.800
<v Speaker 1>four point six max on s WE bench bro. The

32
00:02:10.840 --> 00:02:13.280
<v Speaker 1>part worth watching is that it runs on a single

33
00:02:13.319 --> 00:02:17.479
<v Speaker 1>twenty four bigabyte GPU and ships under APATCHE two point zero.

34
00:02:18.080 --> 00:02:21.439
<v Speaker 1>Third party verification is still pending, so treat the benchmark

35
00:02:21.520 --> 00:02:25.039
<v Speaker 1>with appropriate caution. But if it holds, the open weight

36
00:02:25.039 --> 00:02:28.800
<v Speaker 1>frontier has reached parity with closed proprietary tiers that cost

37
00:02:28.879 --> 00:02:32.719
<v Speaker 1>multiples more that's not a gradual trend. That's a supply

38
00:02:32.840 --> 00:02:36.960
<v Speaker 1>shock to proprietary model positioning. On the security side, AI

39
00:02:37.080 --> 00:02:41.360
<v Speaker 1>no taking platform TLM, AI no taking platform TLM DV

40
00:02:41.560 --> 00:02:43.960
<v Speaker 1>left one hundred and eighty one thousand, eight hundred and

41
00:02:44.000 --> 00:02:47.400
<v Speaker 1>seventy four meeting records queriable by any signed up user

42
00:02:47.439 --> 00:02:51.599
<v Speaker 1>for six months, two million users affected. The access required

43
00:02:51.639 --> 00:02:55.879
<v Speaker 1>no additional authentication. This isn't a sophisticated attack. It's a

44
00:02:55.879 --> 00:03:00.680
<v Speaker 1>permission's failure, sitting undetected for half a year inside enterprise workflows.

45
00:03:01.000 --> 00:03:05.960
<v Speaker 1>Here's the thing. The liability scope is still unquantified. GDPR exposure,

46
00:03:06.159 --> 00:03:10.800
<v Speaker 1>customer retention impact, incident response adequacy. None of that is resolved.

47
00:03:11.240 --> 00:03:14.759
<v Speaker 1>The broader implication is harder to ignore as AI tools

48
00:03:14.840 --> 00:03:18.360
<v Speaker 1>capture more ambient workplace activity. The blast radius of a

49
00:03:18.400 --> 00:03:22.400
<v Speaker 1>misconfiguration scales with adoption. That's the data security debt the

50
00:03:22.439 --> 00:03:26.439
<v Speaker 1>industry is accumulating. At Crypto twenty twenty six, a new

51
00:03:26.479 --> 00:03:31.240
<v Speaker 1>dedicated track formerly treats trained neural networks as mathematical objects

52
00:03:31.400 --> 00:03:35.319
<v Speaker 1>subject to cryptanalytic attack. One hundred and eighty nine papers

53
00:03:35.479 --> 00:03:39.840
<v Speaker 1>across ten volumes, a record model extraction attacks and water

54
00:03:39.919 --> 00:03:44.479
<v Speaker 1>mark robustness now have rigorous formal frameworks. The practical consequence

55
00:03:44.759 --> 00:03:47.439
<v Speaker 1>is that IP theft from model weights is no longer

56
00:03:47.479 --> 00:03:51.439
<v Speaker 1>a loose threat. It's a defined attack surface with documented methods.

57
00:03:52.159 --> 00:03:55.719
<v Speaker 1>Security teams at model providers need to be watching this closely.

58
00:03:56.439 --> 00:03:59.560
<v Speaker 1>Two financial signals to close on, and Thropic posted ten

59
00:03:59.560 --> 00:04:02.120
<v Speaker 1>point nine nine billion dollars in quarter two revenue, up

60
00:04:02.199 --> 00:04:04.520
<v Speaker 1>one hundred and thirty percent year over year, with five

61
00:04:04.639 --> 00:04:07.639
<v Speaker 1>hundred and fifty nine million dollars in operating profit, two

62
00:04:07.759 --> 00:04:10.840
<v Speaker 1>years ahead of their own schedule. The important caveat is

63
00:04:10.840 --> 00:04:14.439
<v Speaker 1>that a SpaceX bulk discount inflated those results. Q three

64
00:04:14.479 --> 00:04:17.240
<v Speaker 1>will show whether enterprise growth holds without a single META

65
00:04:17.319 --> 00:04:20.639
<v Speaker 1>customer carrying the margin. And open AI's S one filing

66
00:04:20.720 --> 00:04:24.560
<v Speaker 1>is expected imminently, with a September listing target, two billion

67
00:04:24.560 --> 00:04:27.600
<v Speaker 1>dollars per month in revenue, fourteen billion dollar projected loss

68
00:04:27.600 --> 00:04:30.759
<v Speaker 1>for twenty twenty six, and Microsoft revenue share terms all

69
00:04:30.759 --> 00:04:34.240
<v Speaker 1>disclosed ahead of the prospectus. The numbers are out, the

70
00:04:34.319 --> 00:04:39.000
<v Speaker 1>interpretation is just beginning. The real watch points heading into

71
00:04:39.040 --> 00:04:42.879
<v Speaker 1>next week whether the d orbit capture demonstration succeeds, what

72
00:04:43.040 --> 00:04:46.560
<v Speaker 1>open AI's public filing actually reveals about its cost structure,

73
00:04:46.759 --> 00:04:51.040
<v Speaker 1>and whether q wen's benchmark lead survives independent verification. Three

74
00:04:51.120 --> 00:04:55.040
<v Speaker 1>separate proof points. Each one resolves a different kind of uncertainty.

75
00:04:55.560 --> 00:04:59.240
<v Speaker 1>Thanks for listening. This podcast was built using AI technology.

76
00:04:59.639 --> 00:05:00.720
<v Speaker 1>A Yes Sweet production