1 00:00:02,070 --> 00:00:06,510 Every time software repeats a job, it uses time and resources. 2 00:00:06,550 --> 00:00:08,660 Beyond Brute Force develops software 3 00:00:08,660 --> 00:00:11,150 that reuses what is still useful, checks the 4 00:00:11,190 --> 00:00:12,760 result, and helps you get more 5 00:00:12,760 --> 00:00:15,790 from the storage and computing you already pay for. 6 00:00:15,830 --> 00:00:17,890 Start with the files you keep. 7 00:00:17,930 --> 00:00:19,490 Backups and project revisions often 8 00:00:19,490 --> 00:00:22,220 contain information you have already saved. 9 00:00:22,260 --> 00:00:25,320 GLYX reuses compatible information 10 00:00:25,320 --> 00:00:27,340 across an archive while preserving what is 11 00:00:27,380 --> 00:00:29,430 needed to recover the exact files. 12 00:00:29,470 --> 00:00:31,640 That gives you a way to keep more 13 00:00:31,640 --> 00:00:33,490 history in the space available, and reduce 14 00:00:33,530 --> 00:00:36,880 storage charges where less billable space is needed. 15 00:00:36,920 --> 00:00:39,150 In our exact-repeat test, adding 16 00:00:39,150 --> 00:00:41,820 almost two gigabytes again needed less than one 17 00:00:41,860 --> 00:00:44,900 megabyte of extra archive and recovery records. 18 00:00:44,940 --> 00:00:47,180 Every file came back exactly. 19 00:00:47,220 --> 00:00:50,190 What you keep is only part of the cost. 20 00:00:50,230 --> 00:00:53,900 There is also the work your applications send to an AI service. 21 00:00:53,940 --> 00:00:56,080 For supported routine operations, 22 00:00:56,080 --> 00:00:58,950 TokenHoloFold can interpret the request, do the 23 00:00:58,990 --> 00:01:01,310 job locally and check the result. 24 00:01:01,350 --> 00:01:03,160 In the recorded test, all three 25 00:01:03,160 --> 00:01:06,060 hundred supported requests completed exactly 26 00:01:06,100 --> 00:01:08,150 without a model-provider call. 27 00:01:08,150 --> 00:01:11,070 If you are paying for that call, removing it removes 28 00:01:11,110 --> 00:01:15,020 its charge. It also removes the wait for that remote response. 29 00:01:15,060 --> 00:01:16,680 We then count the local running 30 00:01:16,680 --> 00:01:20,460 and integration costs to work out what you actually save. 31 00:01:20,500 --> 00:01:23,440 The same principle applies when information changes. 32 00:01:23,480 --> 00:01:26,800 If you update one small part of a large dataset, 33 00:01:26,800 --> 00:01:28,380 Seed State is designed to keep 34 00:01:28,420 --> 00:01:31,630 state that is still valid and update the affected part. 35 00:01:31,670 --> 00:01:33,700 For supported work, that can mean 36 00:01:33,700 --> 00:01:36,620 less waiting for the new result and more capacity 37 00:01:36,660 --> 00:01:39,590 for other jobs. In a recorded graphics-processor 38 00:01:39,590 --> 00:01:41,370 study, the measured update step 39 00:01:41,410 --> 00:01:43,260 ran about twenty times faster. 40 00:01:43,300 --> 00:01:47,510 Finishing that step sooner lets the work after it move forward. 41 00:01:47,550 --> 00:01:49,260 The broader goal is more useful 42 00:01:49,260 --> 00:01:52,560 work from the hardware and power budget you already have. 43 00:01:52,742 --> 00:01:56,207 Expanding Worlds, from Beyond Brute Force, is exploring 44 00:01:56,222 --> 00:02:00,477 that opportunity through compact descriptions, reusable structure and 45 00:02:00,492 --> 00:02:02,237 selective reconstruction. 46 00:02:02,502 --> 00:02:05,937 Sometimes a small description can refer to another 47 00:02:05,952 --> 00:02:08,767 reusable description: a room inside a building, or 48 00:02:08,782 --> 00:02:10,487 a building inside a district. 49 00:02:10,502 --> 00:02:13,007 That is what we mean by nested structure. 50 00:02:13,022 --> 00:02:15,467 It lets useful patterns be shared at more 51 00:02:15,482 --> 00:02:16,517 than one level. 52 00:02:16,782 --> 00:02:20,277 In one recorded study, the expanded information averaged 53 00:02:20,292 --> 00:02:22,657 about five hundred and thirty-one thousand bytes per 54 00:02:22,672 --> 00:02:23,647 session. 55 00:02:23,662 --> 00:02:26,137 Our nested packet averaged four hundred and sixty-two 56 00:02:26,152 --> 00:02:26,777 bytes. 57 00:02:26,792 --> 00:02:31,247 All forty-eight tested state and rendered-frame comparisons matched 58 00:02:31,262 --> 00:02:31,907 exactly. 59 00:02:31,922 --> 00:02:33,947 That is more than ninety-nine point nine percent 60 00:02:33,962 --> 00:02:37,887 less information in that particular packet comparison. 61 00:02:38,162 --> 00:02:40,997 The shared software and assets were already present. 62 00:02:41,012 --> 00:02:43,597 The small packet told that compatible system what 63 00:02:43,612 --> 00:02:44,337 to rebuild. 64 00:02:45,600 --> 00:02:47,920 A world also has to be drawn. 65 00:02:47,960 --> 00:02:51,610 When one object moves, much of the scene may still be usable. 66 00:02:51,650 --> 00:02:53,720 Retained Graphics is designed to 67 00:02:53,720 --> 00:02:56,590 rebuild the image regions affected by a supported 68 00:02:56,630 --> 00:02:59,200 change and check the resulting picture. 69 00:02:59,240 --> 00:03:01,970 The aim is more responsive graphics: 70 00:03:01,970 --> 00:03:04,600 a shorter wait for a designer to see an edit, 71 00:03:04,640 --> 00:03:08,850 or more room for a complex scene within the same rendering budget. 72 00:03:08,890 --> 00:03:12,615 To make systems like these useful, their parts need to fit together. 73 00:03:12,655 --> 00:03:15,240 Glyphtronix is the structured foundation 74 00:03:15,240 --> 00:03:18,050 for supported operations, their roles, 75 00:03:18,090 --> 00:03:20,220 relationships and sources. The 76 00:03:20,220 --> 00:03:23,220 aim is to make workflows easier to compose and 77 00:03:23,260 --> 00:03:25,820 investigate. When something changes, 78 00:03:25,820 --> 00:03:27,940 a developer should be able to trace what 79 00:03:27,980 --> 00:03:30,540 depends on it. When a result looks wrong, 80 00:03:30,540 --> 00:03:32,470 they need to see where it came from. 81 00:03:32,510 --> 00:03:35,100 And when software says the job is finished, 82 00:03:35,100 --> 00:03:37,130 it needs to account for the whole 83 00:03:37,170 --> 00:03:39,380 request. Ask for three outputs 84 00:03:39,380 --> 00:03:42,190 and receive only two, and the job is still 85 00:03:42,230 --> 00:03:44,620 incomplete. Prooftronix is designed 86 00:03:44,620 --> 00:03:47,050 to check the required parts and the combined 87 00:03:47,090 --> 00:03:49,440 response, making missing or mismatched 88 00:03:49,440 --> 00:03:52,310 work visible before somebody relies on it. 89 00:03:52,350 --> 00:03:54,520 Catching that gap earlier can prevent 90 00:03:54,520 --> 00:03:56,900 it becoming rework or a bad decision further 91 00:03:56,940 --> 00:03:58,210 down the line. 92 00:03:58,250 --> 00:04:00,100 Those questions about context, 93 00:04:00,100 --> 00:04:03,440 evidence and completion also shape Cog, which is 94 00:04:03,480 --> 00:04:05,600 coming soon and still in training. 95 00:04:05,640 --> 00:04:07,680 We are developing it to keep possible 96 00:04:07,680 --> 00:04:09,810 answers connected to their sources and 97 00:04:09,850 --> 00:04:12,000 supporting information. The aim 98 00:04:12,000 --> 00:04:15,090 is to make it clearer why a response is supported, 99 00:04:15,130 --> 00:04:17,760 where it came from, and what still needs checking. 100 00:04:17,800 --> 00:04:21,130 This is the research direction we are working towards. 101 00:04:21,170 --> 00:04:25,280 Across the eight technologies, the practical questions stay the same. 102 00:04:25,320 --> 00:04:27,820 Can you keep more useful history? 103 00:04:27,860 --> 00:04:30,885 Get the result sooner? Remove an avoidable charge? 104 00:04:30,925 --> 00:04:33,720 Know whether the job is complete? 105 00:04:33,760 --> 00:04:35,350 We compare the relevant product 106 00:04:35,350 --> 00:04:37,740 with the best practical alternative for your 107 00:04:37,780 --> 00:04:40,720 workload, counting the full result and the full cost. 108 00:04:40,760 --> 00:04:45,690 Beyond Brute Force. Reuse structure. 109 00:04:45,730 --> 00:04:47,109 Create more.