1 00:00:00,532 --> 00:00:02,549 This is a language model. 2 00:00:02,549 --> 00:00:04,757 Every point of light is one number, 3 00:00:04,757 --> 00:00:08,299 a weight. There are seven and a half billion of them, 4 00:00:08,299 --> 00:00:11,450 and together they're everything the model knows. 5 00:00:11,838 --> 00:00:15,070 The usual way to teach it a new skill is to let all of 6 00:00:15,070 --> 00:00:17,879 them move, each by a tiny amount. 7 00:00:17,879 --> 00:00:20,747 What comes out is a whole new copy of the model: 8 00:00:20,747 --> 00:00:23,100 twenty-eight gigabytes of numbers. 9 00:00:23,439 --> 00:00:26,160 This paper tries something that sounds like it shouldn't 10 00:00:26,160 --> 00:00:29,042 work. It doesn't write a single new number. 11 00:00:29,042 --> 00:00:32,080 It picks about three million of the existing ones, 12 00:00:32,080 --> 00:00:36,030 fewer than one in two thousand, and sets them to zero. 13 00:00:36,370 --> 00:00:39,330 On six math tests it never trained on, 14 00:00:39,330 --> 00:00:42,958 its average score goes from forty percent to fifty-two. 15 00:00:42,958 --> 00:00:46,320 Letting every weight move gets about the same score. 16 00:00:47,395 --> 00:00:50,370 Teaching a model by deleting part of it. 17 00:00:52,110 --> 00:00:54,783 The method is called Deletion Tuning. 18 00:00:54,783 --> 00:00:57,871 It's from a paper accepted at NeurIPS twenty twenty-six, 19 00:00:57,871 --> 00:01:01,065 from N Y U and the University of Calgary. 20 00:01:01,065 --> 00:01:03,231 We'll see how a deletion can teach, 21 00:01:03,231 --> 00:01:06,234 how you train a choice that can only be yes or no, 22 00:01:06,234 --> 00:01:09,340 what you get for it, and why it works at all. 23 00:01:13,220 --> 00:01:17,104 Under the hood, a language model is a tower of layers. 24 00:01:17,104 --> 00:01:20,580 Your text goes in at the bottom, gets transformed layer 25 00:01:20,580 --> 00:01:23,780 after layer, and at the top comes a guess at the next 26 00:01:23,780 --> 00:01:24,770 word. 27 00:01:25,110 --> 00:01:28,892 Most of each layer is a few big grids of numbers. 28 00:01:28,892 --> 00:01:31,270 Each number in a grid is a weight. 29 00:01:31,270 --> 00:01:34,550 It takes one value coming in, multiplies it, 30 00:01:34,550 --> 00:01:37,827 and adds the result into one value going out. 31 00:01:37,827 --> 00:01:41,110 So each column of the grid handles one input, 32 00:01:41,110 --> 00:01:45,665 each row builds one output, and every tile is one weight. 33 00:01:45,665 --> 00:01:48,790 Stack enough of these grids and you get eight billion 34 00:01:48,790 --> 00:01:52,070 weights. Everything the model does comes from those 35 00:01:52,070 --> 00:01:53,220 values. 36 00:01:53,863 --> 00:01:57,400 A freshly pretrained model has read an enormous amount of 37 00:01:57,400 --> 00:02:00,680 text, but nobody has taught it to do any particular job 38 00:02:00,680 --> 00:02:04,607 yet. That second stage is called post-training, 39 00:02:04,607 --> 00:02:07,080 and this paper tests three kinds of it. 40 00:02:07,080 --> 00:02:09,642 Some of its models have already had a first round, 41 00:02:09,642 --> 00:02:13,073 to follow instructions, and the paper adds one more. 42 00:02:13,073 --> 00:02:16,200 We'll start with the main kind, reinforcement learning on 43 00:02:16,200 --> 00:02:17,190 math. 44 00:02:17,530 --> 00:02:20,682 Give the model a word problem with a checkable answer. 45 00:02:20,682 --> 00:02:24,250 On this test, the final answer has to come after four hash 46 00:02:24,250 --> 00:02:28,085 marks, like this. Let the model write several attempts. 47 00:02:28,085 --> 00:02:29,769 Check each final answer: 48 00:02:29,769 --> 00:02:31,480 right or wrong. 49 00:02:31,820 --> 00:02:34,860 Then nudge the model's weights so it becomes a little more 50 00:02:34,860 --> 00:02:37,420 likely to write the words of the right attempts, 51 00:02:37,420 --> 00:02:40,220 and a little less likely to write the wrong ones. 52 00:02:40,220 --> 00:02:42,451 The checker only says yes or no. 53 00:02:42,451 --> 00:02:45,820 What gets nudged smoothly is the model's own odds of 54 00:02:45,820 --> 00:02:49,260 writing each word. How hard each attempt gets pushed 55 00:02:49,260 --> 00:02:52,059 depends on how it compares with the rest of its group. 56 00:02:52,059 --> 00:02:55,941 A right answer when most attempts failed gets a big push. 57 00:02:55,941 --> 00:02:59,020 Repeat over thousands of problems, and the model gets 58 00:02:59,020 --> 00:03:00,090 better at them. 59 00:03:00,750 --> 00:03:02,923 The standard way to make those nudges, 60 00:03:02,923 --> 00:03:06,227 full fine-tuning, lets every weight move, 61 00:03:06,227 --> 00:03:09,630 by tiny amounts. All eight billion of them are free to 62 00:03:09,630 --> 00:03:13,310 change. Yet earlier research found that reinforcement 63 00:03:13,310 --> 00:03:16,750 learning ends up noticeably changing only a small part of 64 00:03:16,750 --> 00:03:20,048 a model: five to thirty percent of its weights. 65 00:03:20,048 --> 00:03:24,140 This paper pushes that observation as far as it can go. 66 00:03:24,476 --> 00:03:27,190 So the result is a complete new model: 67 00:03:27,190 --> 00:03:30,400 about thirty gigabytes, counted in the thirty-two-bit 68 00:03:30,400 --> 00:03:32,479 format that training uses. 69 00:03:32,479 --> 00:03:35,680 Teach three skills that way, and you have three full 70 00:03:35,680 --> 00:03:38,110 copies to store and to load. 71 00:03:41,809 --> 00:03:44,362 There are cheaper ways to change a model. 72 00:03:44,362 --> 00:03:46,648 The most popular is called LoRA. 73 00:03:46,648 --> 00:03:49,121 It keeps the original weights frozen. 74 00:03:49,121 --> 00:03:52,890 Instead, beside a grid it trains two short lists of 75 00:03:52,890 --> 00:03:55,613 numbers: one with a number for each row, 76 00:03:55,613 --> 00:03:57,849 and one with a number for each column. 77 00:03:57,849 --> 00:04:00,970 Multiply them, each row's number times each column's 78 00:04:00,970 --> 00:04:03,610 number, and you get a full grid of changes, 79 00:04:03,610 --> 00:04:05,599 the same size as the original. 80 00:04:05,599 --> 00:04:09,530 Here, this row's number is two and this column's number is 81 00:04:09,530 --> 00:04:12,568 three, so this weight changes by six. 82 00:04:12,568 --> 00:04:15,690 So every row of changes has the same shape, 83 00:04:15,690 --> 00:04:17,530 just turned up or down. 84 00:04:17,530 --> 00:04:20,968 That grid of changes is added to the original grid. 85 00:04:20,968 --> 00:04:24,970 Real lists hold tiny numbers, so nearly every weight in it 86 00:04:24,970 --> 00:04:26,760 shifts just a little. 87 00:04:27,100 --> 00:04:29,713 And it isn't one pair for the whole model. 88 00:04:29,713 --> 00:04:33,026 Every grid in every layer gets its own pair. 89 00:04:33,026 --> 00:04:36,620 This eight-billion-weight model has seven grids in each of 90 00:04:36,620 --> 00:04:38,221 its thirty-two layers: 91 00:04:38,221 --> 00:04:40,411 two hundred and twenty-four pairs. 92 00:04:40,411 --> 00:04:43,260 Each pair is tiny next to its grid. 93 00:04:43,260 --> 00:04:46,540 A grid of fifty-nine million weights gets two lists with 94 00:04:46,540 --> 00:04:49,074 about eighteen thousand numbers between them. 95 00:04:49,074 --> 00:04:52,460 For more freedom, a grid can get several pairs, 96 00:04:52,460 --> 00:04:55,094 with their grids of changes added together. 97 00:04:55,094 --> 00:04:57,740 The number of pairs is called the rank. 98 00:04:57,740 --> 00:05:01,180 At rank one, the whole change for an eight-billion-weight 99 00:05:01,180 --> 00:05:03,690 model fits in ten megabytes. 100 00:05:04,275 --> 00:05:07,972 Another approach, random sparse fine-tuning, 101 00:05:07,972 --> 00:05:10,990 picks a small random set of the model's weights before 102 00:05:10,990 --> 00:05:14,190 training starts, say one in two thousand, 103 00:05:14,190 --> 00:05:17,790 and lets only those change, to any value they like. 104 00:05:17,790 --> 00:05:21,470 To save the result, you store each changed value and its 105 00:05:21,470 --> 00:05:24,910 position. For the same eight-billion-weight model, 106 00:05:24,910 --> 00:05:27,340 that's about twenty-seven megabytes. 107 00:05:27,977 --> 00:05:31,214 Every way of changing a model makes two choices: 108 00:05:31,214 --> 00:05:34,159 which weights change, and by how much. 109 00:05:34,159 --> 00:05:36,799 Full fine-tuning changes every weight, 110 00:05:36,799 --> 00:05:40,959 by any amount. LoRA changes nearly every weight too, 111 00:05:40,959 --> 00:05:44,560 but each grid's change is built from just a pair or a few 112 00:05:44,560 --> 00:05:48,800 pairs of lists. Random sparse fine-tuning picks its 113 00:05:48,800 --> 00:05:51,760 weights blindly, then sets them to anything. 114 00:05:51,760 --> 00:05:54,750 Every one of them writes new numbers into the model. 115 00:05:55,062 --> 00:05:58,278 Deletion Tuning learns which weights to change, 116 00:05:58,278 --> 00:06:00,755 and allows exactly one change: 117 00:06:00,755 --> 00:06:03,330 to zero. It can't add anything. 118 00:06:03,330 --> 00:06:04,927 It can only remove. 119 00:06:04,927 --> 00:06:08,130 So can removing alone teach a model something it didn't 120 00:06:08,130 --> 00:06:08,880 know? 121 00:06:13,339 --> 00:06:16,301 Picture an editor who's allowed to strike things out, 122 00:06:16,301 --> 00:06:17,491 but never to write. 123 00:06:17,491 --> 00:06:19,494 That sounds like a weak editor. 124 00:06:19,494 --> 00:06:22,374 But strike one comma from "Let's eat, 125 00:06:22,374 --> 00:06:26,060 Grandma," and the sentence means something very different. 126 00:06:26,060 --> 00:06:29,180 Strike words at random and almost nothing changes. 127 00:06:29,180 --> 00:06:32,771 Strike the right ones, and the meaning changes completely. 128 00:06:32,771 --> 00:06:36,090 All of an editor's power is in choosing where. 129 00:06:36,750 --> 00:06:40,584 Inside a model, one deletion can act like that comma. 130 00:06:40,584 --> 00:06:43,788 Every layer has a part with thousands of channels: 131 00:06:43,788 --> 00:06:46,828 little pathways a word's signal can take. 132 00:06:46,828 --> 00:06:50,590 For each word passing through, a gate decides how far each 133 00:06:50,590 --> 00:06:53,791 channel opens. The gate reads the input: 134 00:06:53,791 --> 00:06:56,141 each input value times a gate weight, 135 00:06:56,141 --> 00:06:59,542 all added up. Here the total is positive, 136 00:06:59,542 --> 00:07:02,860 so this channel opens and its signal flows on. 137 00:07:03,225 --> 00:07:05,270 Now delete one gate weight. 138 00:07:05,270 --> 00:07:08,957 Its contribution is gone, the total drops below zero, 139 00:07:08,957 --> 00:07:11,210 and the channel stays nearly shut. 140 00:07:11,210 --> 00:07:14,305 For other words the same deletion barely matters. 141 00:07:14,305 --> 00:07:17,760 One deletion changes which words get through which 142 00:07:17,760 --> 00:07:18,670 channel. 143 00:07:19,009 --> 00:07:21,333 Notice what kind of change that is. 144 00:07:21,333 --> 00:07:25,648 Full fine-tuning can nudge every weight by a tiny amount. 145 00:07:25,648 --> 00:07:27,168 A deletion is the opposite: 146 00:07:27,168 --> 00:07:29,797 a big change, the whole weight gone, 147 00:07:29,797 --> 00:07:31,600 in very few places. 148 00:07:32,180 --> 00:07:35,371 Deletion Tuning only deletes inside these gates, 149 00:07:35,371 --> 00:07:38,420 one gate grid per layer, because the gate is where 150 00:07:38,420 --> 00:07:40,578 channels open and close. 151 00:07:40,578 --> 00:07:44,126 Everything else in the model stays exactly as it was. 152 00:07:44,126 --> 00:07:47,540 In the eight-billion-weight model, the gate grids hold one 153 00:07:47,540 --> 00:07:51,300 point nine billion weights, about a quarter of the model. 154 00:07:51,300 --> 00:07:54,580 The trained deletions take out roughly one gate weight in 155 00:07:54,580 --> 00:07:57,540 five hundred. And since the gates are a quarter of the 156 00:07:57,540 --> 00:08:01,060 model, that's about one in two thousand of all its 157 00:08:01,060 --> 00:08:02,050 weights. 158 00:08:05,771 --> 00:08:09,627 There's a problem. Models learn by small adjustments: 159 00:08:09,627 --> 00:08:13,150 for every weight, training asks which way a tiny change 160 00:08:13,150 --> 00:08:15,870 would help, and moves it a little that way. 161 00:08:15,870 --> 00:08:18,670 But a deletion isn't a small adjustment. 162 00:08:18,670 --> 00:08:21,389 A weight is either there or it's gone. 163 00:08:21,389 --> 00:08:24,349 Flip a switch a tiny amount and nothing happens, 164 00:08:24,349 --> 00:08:27,580 so training gets no signal about which way to go. 165 00:08:28,212 --> 00:08:31,677 Here's the trick. The model's own weights are frozen: 166 00:08:31,677 --> 00:08:35,040 training never changes a single one of their values. 167 00:08:35,040 --> 00:08:37,999 Instead, every gate weight gets a dial, 168 00:08:37,999 --> 00:08:40,196 a number training can turn smoothly, 169 00:08:40,196 --> 00:08:42,880 and the dial controls a switch that sits in front of the 170 00:08:42,880 --> 00:08:47,120 weight. During training the switch is soft near zero: 171 00:08:47,120 --> 00:08:49,920 a dial a little above zero lets most of the weight 172 00:08:49,920 --> 00:08:53,271 through, a dial a little below lets a little through. 173 00:08:53,271 --> 00:08:57,520 So turning a dial a tiny bit changes the model a tiny bit, 174 00:08:57,520 --> 00:08:59,550 and that's what training needs. 175 00:08:59,889 --> 00:09:03,810 When the model practises its math, it uses the clean-cut 176 00:09:03,810 --> 00:09:06,925 version: every switch fully on or off, 177 00:09:06,925 --> 00:09:09,305 the model as it stands at that moment. 178 00:09:09,305 --> 00:09:12,130 Then the verdicts on those attempts push back through the 179 00:09:12,130 --> 00:09:14,610 soft version to turn the dials. 180 00:09:14,610 --> 00:09:17,330 That works because the two versions are almost the same 181 00:09:17,330 --> 00:09:21,410 model. Nearly every dial sits well away from zero, 182 00:09:21,410 --> 00:09:24,889 where the soft switch is already fully on or fully off. 183 00:09:24,889 --> 00:09:28,000 Only the few dials near zero are in between. 184 00:09:28,340 --> 00:09:31,945 And over training, the switch is made ten times sharper, 185 00:09:31,945 --> 00:09:35,250 so the soft version closes in on the clean cut. 186 00:09:35,540 --> 00:09:38,699 The same two dials now let through all of the weight, 187 00:09:38,699 --> 00:09:42,580 and none of it. By the end, a positive dial keeps its 188 00:09:42,580 --> 00:09:45,299 weight and a negative dial deletes it. 189 00:09:45,299 --> 00:09:48,580 Written out, the weight is multiplied by how open its 190 00:09:48,580 --> 00:09:52,158 switch is. What's left is one bit per weight: 191 00:09:52,158 --> 00:09:53,890 keep, or delete. 192 00:09:54,550 --> 00:09:56,132 Where do the dials start? 193 00:09:56,132 --> 00:09:58,470 With an idea called pruning: 194 00:09:58,470 --> 00:10:01,910 deleting weights to make a model smaller and faster while 195 00:10:01,910 --> 00:10:04,149 hurting it as little as possible. 196 00:10:04,149 --> 00:10:06,949 A classic pruning score, called Wanda, 197 00:10:06,949 --> 00:10:09,350 guesses which weights matter least. 198 00:10:09,350 --> 00:10:12,210 A weight probably matters little if it's small, 199 00:10:12,210 --> 00:10:15,266 or if the input it multiplies is usually quiet. 200 00:10:15,266 --> 00:10:18,950 So the score is the weight's size times how active its 201 00:10:18,950 --> 00:10:21,676 input is on the task's own prompts. 202 00:10:21,676 --> 00:10:24,950 The lowest-scoring sliver start just below zero, 203 00:10:24,950 --> 00:10:26,549 already switched off. 204 00:10:26,549 --> 00:10:29,501 Everything else starts just above, switched on, 205 00:10:29,501 --> 00:10:31,990 with a little random jitter on every dial, 206 00:10:31,990 --> 00:10:33,910 so a few of those start off too. 207 00:10:33,910 --> 00:10:37,766 Altogether, about one gate weight in five hundred starts 208 00:10:37,766 --> 00:10:41,270 switched off. There's also a mild pull toward keeping the 209 00:10:41,270 --> 00:10:44,390 bigger weights. From there, training decides what to 210 00:10:44,390 --> 00:10:46,627 delete. There's no target number. 211 00:10:46,627 --> 00:10:49,015 In fact, with reinforcement learning, 212 00:10:49,015 --> 00:10:52,391 the number switched off ends up close to where it started. 213 00:10:52,391 --> 00:10:55,909 Only the dials that start near zero can really move. 214 00:10:55,909 --> 00:10:58,534 The rest sit where the switch is fully on, 215 00:10:58,534 --> 00:11:00,255 and get no signal at all. 216 00:11:00,255 --> 00:11:03,990 So training's real choice is which of those edge cases to 217 00:11:03,990 --> 00:11:05,220 switch off. 218 00:11:05,806 --> 00:11:08,681 What you ship at the end is just the switches. 219 00:11:08,681 --> 00:11:11,800 One bit for each of the one point nine billion gate 220 00:11:11,800 --> 00:11:14,280 weights comes to about two hundred and twenty-four 221 00:11:14,280 --> 00:11:18,120 megabytes. That list of switches is called a mask. 222 00:11:18,120 --> 00:11:21,640 But almost every switch is on, so the stream squeezes down 223 00:11:21,640 --> 00:11:25,400 to almost nothing. Compressed, the whole skill is eight 224 00:11:25,400 --> 00:11:28,680 point four megabytes, against thirty gigabytes for a full 225 00:11:28,680 --> 00:11:32,520 copy. To use it, you take the original model and set those 226 00:11:32,520 --> 00:11:36,119 weights to zero. What runs is an ordinary model, 227 00:11:36,119 --> 00:11:39,350 the same size and the same speed as before. 228 00:11:43,010 --> 00:11:46,370 Now the tests. Six models from two families, 229 00:11:46,370 --> 00:11:49,970 three sizes each. The Llama models train and test on 230 00:11:49,970 --> 00:11:53,490 grade-school math. The Qwen models train on a broader math 231 00:11:53,490 --> 00:11:57,565 collection and are tested on six held-out math benchmarks. 232 00:11:57,565 --> 00:12:00,769 The score is the share of test problems answered right, 233 00:12:00,769 --> 00:12:04,130 with one try each, and every method keeps its best score 234 00:12:04,130 --> 00:12:07,312 along the way. The original model is grey. 235 00:12:07,312 --> 00:12:10,527 Full fine-tuning is coral, LoRA purple, 236 00:12:10,527 --> 00:12:14,127 random sparse green, and Deletion Tuning blue. 237 00:12:14,127 --> 00:12:17,570 One thing to know: the Llama test only counts an answer 238 00:12:17,570 --> 00:12:20,287 written exactly after the four hash marks, 239 00:12:20,287 --> 00:12:23,489 so part of what the Llamas learn here is that format. 240 00:12:23,489 --> 00:12:26,640 The Qwen tests read the answer however it's written. 241 00:12:26,640 --> 00:12:28,929 If deleting really can't teach much, 242 00:12:28,929 --> 00:12:31,840 the blue bars should stay close to the grey ones. 243 00:12:32,500 --> 00:12:34,260 It's nowhere near grey. 244 00:12:34,260 --> 00:12:37,619 Blue climbs nearly as high as coral everywhere. 245 00:12:37,619 --> 00:12:41,300 At the smallest sizes, deletion scores a few points lower 246 00:12:41,300 --> 00:12:42,739 than full fine-tuning: 247 00:12:42,739 --> 00:12:45,620 seventeen and a half against nineteen on the smallest 248 00:12:45,620 --> 00:12:49,940 Qwen. But measured as a share of full fine-tuning's score, 249 00:12:49,940 --> 00:12:53,620 it climbs as models grow, and at the largest of these six 250 00:12:53,620 --> 00:12:57,478 in each family, Deletion Tuning matches full fine-tuning: 251 00:12:57,478 --> 00:13:00,739 eighty-three point five five against eighty-three point 252 00:13:00,739 --> 00:13:04,740 two four on Llama, fifty-two point three against fifty-one 253 00:13:04,740 --> 00:13:06,659 point nine on Qwen. 254 00:13:06,659 --> 00:13:09,620 The authors take this as a sign that deleting works better 255 00:13:09,620 --> 00:13:12,985 as models grow. They don't pin down why. 256 00:13:12,985 --> 00:13:16,580 One natural guess: a bigger model has far more weights to 257 00:13:16,580 --> 00:13:20,660 choose from. LoRA and random sparse fine-tuning land in 258 00:13:20,660 --> 00:13:23,780 the same place too, so all of these are strong ways to 259 00:13:23,780 --> 00:13:27,890 learn. What's surprising is that deletion is one of them. 260 00:13:28,550 --> 00:13:31,384 It holds at fourteen billion weights, 261 00:13:31,384 --> 00:13:34,465 and on a thirty-billion-weight mixture-of-experts model, 262 00:13:34,465 --> 00:13:37,190 where each layer has a hundred and twenty-eight small 263 00:13:37,190 --> 00:13:41,110 expert blocks and a router that picks which ones to use. 264 00:13:41,110 --> 00:13:45,110 Deletion Tuning masks the gate inside every expert and 265 00:13:45,110 --> 00:13:46,614 leaves the router alone. 266 00:13:46,614 --> 00:13:49,910 On the same six math benchmarks, deletion scores 267 00:13:49,910 --> 00:13:53,750 fifty-eight point five, and full fine-tuning fifty-eight 268 00:13:53,750 --> 00:13:57,750 point four. The mask for that model is thirty megabytes. 269 00:13:57,750 --> 00:14:00,550 The full fine-tuned copy is a hundred and sixteen 270 00:14:00,550 --> 00:14:01,540 thousand. 271 00:14:02,200 --> 00:14:05,313 Against the other cheap methods, here are the sizes, 272 00:14:05,313 --> 00:14:08,789 on a scale where each step is ten times bigger. 273 00:14:08,789 --> 00:14:12,440 Rank-one LoRA and the smallest random sparse runs take a 274 00:14:12,440 --> 00:14:14,598 few megabytes to a few dozen. 275 00:14:14,598 --> 00:14:17,399 Deletion Tuning is the smallest at every size: 276 00:14:17,399 --> 00:14:19,037 half a megabyte to eight. 277 00:14:19,037 --> 00:14:22,360 Full copies sit thirty-six hundred to forty-seven hundred 278 00:14:22,360 --> 00:14:25,320 times higher. The others store numbers. 279 00:14:25,320 --> 00:14:27,750 A mask only stores positions. 280 00:14:28,329 --> 00:14:31,449 For this film, we wanted to see it ourselves. 281 00:14:31,449 --> 00:14:33,931 The authors released their trained masks, 282 00:14:33,931 --> 00:14:36,411 so we downloaded the eight-megabyte mask for the 283 00:14:36,411 --> 00:14:39,690 eight-billion-weight Llama, and set its three point six 284 00:14:39,690 --> 00:14:41,689 million weights to zero. 285 00:14:41,689 --> 00:14:44,010 On the first four hundred test problems, 286 00:14:44,010 --> 00:14:47,290 our run went from fifty-nine to eighty-four percent, 287 00:14:47,290 --> 00:14:48,853 in line with the paper. 288 00:14:48,853 --> 00:14:51,930 But look at the kind of problem that flipped most often. 289 00:14:51,930 --> 00:14:54,488 The original model gets the math right, 290 00:14:54,488 --> 00:14:57,452 then writes a dollar sign after the hash marks, 291 00:14:57,452 --> 00:14:59,900 and the strict test marks it wrong. 292 00:14:59,900 --> 00:15:03,450 With the deletions, the model writes the bare number. 293 00:15:03,450 --> 00:15:06,250 Read just the last number of each answer, 294 00:15:06,250 --> 00:15:08,970 and the original was already right eighty-three percent of 295 00:15:08,970 --> 00:15:12,650 the time. So on this test, most of what deletion taught 296 00:15:12,650 --> 00:15:14,466 the Llama was the format. 297 00:15:14,466 --> 00:15:16,967 The smallest Llama tells the same story: 298 00:15:16,967 --> 00:15:21,050 46% right by its last number, under one percent in the 299 00:15:21,050 --> 00:15:24,843 required format. The Qwen results don't have this problem: 300 00:15:24,843 --> 00:15:27,689 their tests read the answer however it's written, 301 00:15:27,689 --> 00:15:29,880 so a stray symbol can't sink it. 302 00:15:30,539 --> 00:15:34,214 A fair question: maybe the starting point does the work. 303 00:15:34,214 --> 00:15:37,260 Pruning scores like Wanda are good at finding weights that 304 00:15:37,260 --> 00:15:38,620 are safe to delete. 305 00:15:38,620 --> 00:15:41,651 So the authors deleted weights by Wanda alone, 306 00:15:41,651 --> 00:15:44,620 with no training, at eight different amounts, 307 00:15:44,620 --> 00:15:47,258 and kept whichever amount scored best. 308 00:15:47,258 --> 00:15:51,020 It barely moves: every model stays within two point four 309 00:15:51,020 --> 00:15:52,530 points of where it started. 310 00:15:52,530 --> 00:15:55,900 On the eight-billion-weight Llama, the best Wanda pruning 311 00:15:55,900 --> 00:16:00,380 scores sixty point six five, against sixty point two seven 312 00:16:00,380 --> 00:16:01,901 for the original model. 313 00:16:01,901 --> 00:16:05,206 Trained deletion, starting from the same pruning scores, 314 00:16:05,206 --> 00:16:08,059 reaches eighty-three point five five. 315 00:16:08,059 --> 00:16:10,890 The skill is in the choices training makes. 316 00:16:11,550 --> 00:16:14,830 Nor is it fragile. The authors varied how many weights the 317 00:16:14,830 --> 00:16:17,254 pruning score switches off at the start, 318 00:16:17,254 --> 00:16:20,910 how fast the dials turn, and how hard the pull toward big 319 00:16:20,910 --> 00:16:23,870 weights is: twelve combinations. 320 00:16:23,870 --> 00:16:26,750 Every one landed within one point six of the best on 321 00:16:26,750 --> 00:16:29,631 Llama, and three point two on Qwen. 322 00:16:29,631 --> 00:16:31,630 The single setting used for every main 323 00:16:31,630 --> 00:16:35,340 reinforcement-learning run was at or near the top. 324 00:16:36,000 --> 00:16:39,520 One more test. Let a model try the same problem many 325 00:16:39,520 --> 00:16:43,120 times, sampling its words freely, and count the problem 326 00:16:43,120 --> 00:16:45,435 solved if any try is right. 327 00:16:45,435 --> 00:16:48,960 Tries add up fast: a model that gets a problem right half 328 00:16:48,960 --> 00:16:52,875 the time solves it three times in four with two tries. 329 00:16:52,875 --> 00:16:54,800 With a hundred and twenty-eight tries, 330 00:16:54,800 --> 00:16:58,554 the original Llama solves ninety-five percent of the test. 331 00:16:58,554 --> 00:17:01,760 So maybe training only makes the model give the answers it 332 00:17:01,760 --> 00:17:04,400 could already find, more reliably. 333 00:17:04,400 --> 00:17:07,440 If that were all, the two curves would close up as the 334 00:17:07,440 --> 00:17:10,992 tries pile up. At a hundred and twenty-eight tries, 335 00:17:10,992 --> 00:17:12,268 they still haven't: 336 00:17:12,268 --> 00:17:15,520 the deletion-tuned model solves ninety-nine point six 337 00:17:15,520 --> 00:17:18,463 percent, more than four points higher. 338 00:17:18,463 --> 00:17:21,440 So there are problems it solves that the original never 339 00:17:21,440 --> 00:17:25,119 solved, not even once in a hundred and twenty-eight tries, 340 00:17:25,119 --> 00:17:27,757 at least by this test's strict rules. 341 00:17:27,757 --> 00:17:31,950 On Qwen, scored more forgivingly, the gap is six points. 342 00:17:36,192 --> 00:17:39,239 Learning something new usually costs something old. 343 00:17:39,239 --> 00:17:42,210 Train a model hard on math and it can get worse at other 344 00:17:42,210 --> 00:17:43,810 things it used to do. 345 00:17:43,810 --> 00:17:46,690 So while the Llama models trained on math, 346 00:17:46,690 --> 00:17:50,291 the authors kept testing them on six earlier skills: 347 00:17:50,291 --> 00:17:53,248 two kinds of common sense, general knowledge, 348 00:17:53,248 --> 00:17:56,412 telling truth from myth, following instructions, 349 00:17:56,412 --> 00:17:59,730 and writing code. This chart shows how the average of 350 00:17:59,730 --> 00:18:01,982 those six changes during training. 351 00:18:01,982 --> 00:18:04,368 The dashed line means no change. 352 00:18:04,368 --> 00:18:06,640 Below it, the model is forgetting. 353 00:18:07,300 --> 00:18:09,602 Full fine-tuning drifts down: 354 00:18:09,602 --> 00:18:12,340 by the end it's one to three points worse at the old 355 00:18:12,340 --> 00:18:15,860 skills. LoRA and random sparse slip too, 356 00:18:15,860 --> 00:18:18,172 by up to five point four points. 357 00:18:18,172 --> 00:18:21,300 Deletion Tuning stays on the line, within a point at every 358 00:18:21,300 --> 00:18:23,650 check, at all three sizes. 359 00:18:24,320 --> 00:18:26,997 So how far did each model move? 360 00:18:26,997 --> 00:18:30,230 That matters, because a recent study found that one 361 00:18:30,230 --> 00:18:33,110 number, measured on the new task alone, 362 00:18:33,110 --> 00:18:36,525 tracks how much a model forgets everywhere else. 363 00:18:36,525 --> 00:18:38,071 Here's how it's measured. 364 00:18:38,071 --> 00:18:40,309 Let the original model write an answer, 365 00:18:40,309 --> 00:18:44,070 and at every word, see how likely the new model finds the 366 00:18:44,070 --> 00:18:45,979 word the original chose. 367 00:18:45,979 --> 00:18:49,526 If the new model still likes that word almost as much, 368 00:18:49,526 --> 00:18:51,509 that word adds almost nothing. 369 00:18:51,509 --> 00:18:53,877 If it now finds it much less likely, 370 00:18:53,877 --> 00:18:55,632 that word adds a lot. 371 00:18:55,632 --> 00:18:58,470 Averaged over every word of many answers, 372 00:18:58,470 --> 00:19:01,940 this is called the forward KL divergence. 373 00:19:02,280 --> 00:19:05,640 A model's answers can change a lot while this number stays 374 00:19:05,640 --> 00:19:08,147 small. Most words stay put. 375 00:19:08,147 --> 00:19:10,840 Shift the odds at one decisive word, 376 00:19:10,840 --> 00:19:13,951 just enough that a different word comes out on top, 377 00:19:13,951 --> 00:19:17,000 and the final answer changes, while the average barely 378 00:19:17,000 --> 00:19:21,000 moves. So a big jump in test score can come with a tiny 379 00:19:21,000 --> 00:19:24,920 average. Full fine-tuning climbs to one on the smallest 380 00:19:24,920 --> 00:19:27,542 Llama, and one and a half on the largest. 381 00:19:27,542 --> 00:19:31,272 LoRA and random sparse reach about point four to one point 382 00:19:31,272 --> 00:19:35,590 three. Deletion Tuning never goes above point zero five. 383 00:19:36,316 --> 00:19:39,885 Is that just because deletion changes so few weights? 384 00:19:39,885 --> 00:19:41,427 The authors checked. 385 00:19:41,427 --> 00:19:45,130 Random sparse fine-tuning at its smallest setting changes 386 00:19:45,130 --> 00:19:48,730 one weight in two thousand, about as many as Deletion 387 00:19:48,730 --> 00:19:52,810 Tuning deletes. It still forgets, by up to five point four 388 00:19:52,810 --> 00:19:56,170 points. So the small count alone doesn't seem to explain 389 00:19:56,170 --> 00:19:59,136 it. Random sparse differs in two ways, 390 00:19:59,136 --> 00:20:01,677 though: its weights are picked at random, 391 00:20:01,677 --> 00:20:03,674 and it can set them to anything. 392 00:20:03,674 --> 00:20:06,570 This test can't say which matters more. 393 00:20:06,570 --> 00:20:09,210 The authors' reading is that deleting itself keeps the 394 00:20:09,210 --> 00:20:11,450 model close to where it started: 395 00:20:11,450 --> 00:20:14,674 a deletion can't move a weight anywhere except to zero, 396 00:20:14,674 --> 00:20:17,240 and training picks only a few to remove. 397 00:20:21,549 --> 00:20:25,133 How can a few zeros do the work of a full rewrite? 398 00:20:25,133 --> 00:20:28,540 To find out, the authors trained all four methods the same 399 00:20:28,540 --> 00:20:31,503 way, by reinforcement learning on math, 400 00:20:31,503 --> 00:20:33,340 and looked at the change itself: 401 00:20:33,340 --> 00:20:36,300 the difference between each grid before training and 402 00:20:36,300 --> 00:20:40,149 after. For LoRA it's each grid's pairs of lists, 403 00:20:40,149 --> 00:20:44,380 multiplied out. For random sparse it's scattered tiles set 404 00:20:44,380 --> 00:20:48,166 to new values. For deletion it's scattered tiles too, 405 00:20:48,166 --> 00:20:51,850 each one changed by exactly minus its old value. 406 00:20:52,189 --> 00:20:56,030 To compare changes this different, they use a standard way 407 00:20:56,030 --> 00:20:58,030 of taking any grid apart. 408 00:20:58,030 --> 00:21:01,630 Remember LoRA's grid of changes, where every row has the 409 00:21:01,630 --> 00:21:04,169 same shape, turned up or down. 410 00:21:04,169 --> 00:21:06,180 Call that a simple pattern. 411 00:21:06,180 --> 00:21:08,910 Any grid at all can be rebuilt by adding up simple 412 00:21:08,910 --> 00:21:11,549 patterns. Each one has a strength: 413 00:21:11,549 --> 00:21:13,390 how much of the grid it carries. 414 00:21:13,390 --> 00:21:15,305 Add them strongest first. 415 00:21:15,305 --> 00:21:17,950 The first gives a rough version of the grid, 416 00:21:17,950 --> 00:21:20,328 and each one added sharpens it. 417 00:21:20,328 --> 00:21:23,580 Call the strongest ones a grid's main patterns. 418 00:21:24,240 --> 00:21:28,480 First: how many patterns does each change really use? 419 00:21:28,480 --> 00:21:31,120 The measure is called stable rank. 420 00:21:31,120 --> 00:21:34,720 A change that's one pattern has a stable rank of one. 421 00:21:34,720 --> 00:21:36,624 Two patterns of equal strength: 422 00:21:36,624 --> 00:21:40,192 two. One strong pattern plus a few faint ones: 423 00:21:40,192 --> 00:21:44,480 still about one. LoRA is held to at most one or four 424 00:21:44,480 --> 00:21:48,800 patterns per grid by design, and scores one and about two. 425 00:21:48,800 --> 00:21:51,985 Full fine-tuning's changes use a dozen or two. 426 00:21:51,985 --> 00:21:55,360 Random sparse changes are scattered like noise, 427 00:21:55,360 --> 00:21:58,960 and need hundreds. For deletion, the paper reports 428 00:21:58,960 --> 00:22:01,200 sixty-six to ninety-three. 429 00:22:01,200 --> 00:22:05,054 But that's an average over all seven grids in each layer, 430 00:22:05,054 --> 00:22:07,200 and deletion changes only one of them, 431 00:22:07,200 --> 00:22:11,360 the gate. So for this film, we measured the released masks 432 00:22:11,360 --> 00:22:13,199 on the gate grids ourselves: 433 00:22:13,199 --> 00:22:15,247 about six hundred patterns. 434 00:22:15,247 --> 00:22:18,240 Count the six unchanged grids as zero, 435 00:22:18,240 --> 00:22:21,597 average all seven, and that comes to eighty-four, 436 00:22:21,597 --> 00:22:23,599 close to the paper's ninety-three. 437 00:22:23,599 --> 00:22:26,398 So the paper's average seems to include them. 438 00:22:26,398 --> 00:22:29,758 Six hundred is as many patterns as random sparse's, 439 00:22:29,758 --> 00:22:31,600 and it fits what a deletion is: 440 00:22:31,600 --> 00:22:33,790 single tiles, scattered. 441 00:22:34,450 --> 00:22:36,095 Now the fingerprint. 442 00:22:36,095 --> 00:22:39,330 Earlier research found that when reinforcement learning is 443 00:22:39,330 --> 00:22:43,490 free to change every weight, its changes have two telltale 444 00:22:43,490 --> 00:22:46,770 habits, as recognisable as a fingerprint. 445 00:22:46,770 --> 00:22:50,050 A method whose changes share both is changing the model 446 00:22:50,050 --> 00:22:53,085 the way reinforcement learning naturally does. 447 00:22:53,085 --> 00:22:55,570 The first habit is about the original grid, 448 00:22:55,570 --> 00:22:59,170 not the change. Take the original grid apart into its 449 00:22:59,170 --> 00:23:02,610 patterns, and measure how strong each one is. 450 00:23:02,610 --> 00:23:04,442 Do it again after training. 451 00:23:04,442 --> 00:23:07,410 With every weight free, reinforcement learning barely 452 00:23:07,410 --> 00:23:09,167 changes those strengths: 453 00:23:09,167 --> 00:23:12,368 the grid's patterns stay about as strong as they were. 454 00:23:12,368 --> 00:23:15,730 Averaged over the whole model, deletion keeps them almost 455 00:23:15,730 --> 00:23:18,098 as still as full fine-tuning does: 456 00:23:18,098 --> 00:23:20,259 within one point four times. 457 00:23:20,259 --> 00:23:22,190 The other methods shift them more: 458 00:23:22,190 --> 00:23:25,570 rank-one LoRA, up to eighty-four times more than full 459 00:23:25,570 --> 00:23:28,610 fine-tuning. But that average, too, 460 00:23:28,610 --> 00:23:31,811 counts the six grids deletion never touches. 461 00:23:31,811 --> 00:23:35,090 On the gate grids it does change, we find it shifts the 462 00:23:35,090 --> 00:23:38,290 strengths about nine times as much as full fine-tuning 463 00:23:38,290 --> 00:23:39,760 does on average. 464 00:23:40,420 --> 00:23:43,873 The second habit is about which weights get changed. 465 00:23:43,873 --> 00:23:46,980 Rebuild the original grid from only its strongest 466 00:23:46,980 --> 00:23:50,100 patterns, the hundred and twenty-eight strongest out of 467 00:23:50,100 --> 00:23:51,939 hundreds or thousands. 468 00:23:51,939 --> 00:23:55,460 That gives a rough version that keeps its main structure. 469 00:23:55,460 --> 00:23:58,162 The half of its tiles with the largest values, 470 00:23:58,162 --> 00:24:01,460 positive or negative, are called the principal weights: 471 00:24:01,460 --> 00:24:04,305 the ones the main patterns lean on most. 472 00:24:04,305 --> 00:24:07,699 Now count only changes big enough to notice, 473 00:24:07,699 --> 00:24:10,422 more than a thousandth of a weight's size. 474 00:24:10,422 --> 00:24:13,380 Changes landing at random would hit principal weights half 475 00:24:13,380 --> 00:24:16,900 the time. With every weight free, reinforcement learning's 476 00:24:16,900 --> 00:24:19,137 changes hit them less often than that: 477 00:24:19,137 --> 00:24:22,420 here, full fine-tuning's land on them forty-one to 478 00:24:22,420 --> 00:24:24,258 forty-six percent of the time. 479 00:24:24,258 --> 00:24:26,339 That's only a few points under half, 480 00:24:26,339 --> 00:24:29,241 but it holds across millions of changed weights, 481 00:24:29,241 --> 00:24:32,420 in both models. Deletion's land on them forty-three to 482 00:24:32,420 --> 00:24:36,215 forty-six percent. LoRA's come close to a coin flip, 483 00:24:36,215 --> 00:24:39,220 and random sparse sits exactly on it. 484 00:24:39,220 --> 00:24:41,940 This measure only counts weights that changed, 485 00:24:41,940 --> 00:24:44,733 so the untouched grids don't water it down. 486 00:24:44,733 --> 00:24:47,780 Our check of the released masks finds forty-six and 487 00:24:47,780 --> 00:24:49,036 forty-seven percent: 488 00:24:49,036 --> 00:24:52,340 still under half. So with every weight free, 489 00:24:52,340 --> 00:24:55,300 reinforcement learning leaves each grid's patterns about 490 00:24:55,300 --> 00:24:58,660 as strong as they were, and leans its changes away from 491 00:24:58,660 --> 00:25:00,210 the weights they rest on. 492 00:25:00,870 --> 00:25:03,990 Put it all on the scorecard, with deletion measured on the 493 00:25:03,990 --> 00:25:07,585 grids it changes. Like full fine-tuning, 494 00:25:07,585 --> 00:25:10,572 it leans its changes away from the principal weights; 495 00:25:10,572 --> 00:25:13,189 LoRA and random sparse don't. 496 00:25:13,189 --> 00:25:16,812 But its changes use as many patterns as random sparse's, 497 00:25:16,812 --> 00:25:19,910 and on its own grids, they shift the strengths more than 498 00:25:19,910 --> 00:25:21,414 full fine-tuning's. 499 00:25:21,414 --> 00:25:24,447 So deletion clearly shares one habit. 500 00:25:24,447 --> 00:25:26,994 The other holds only across the whole model, 501 00:25:26,994 --> 00:25:30,230 where six grids in seven don't change at all. 502 00:25:30,230 --> 00:25:33,026 The authors offer a reason for the resemblance. 503 00:25:33,026 --> 00:25:37,430 LoRA is held to a pair or a few pairs of lists per grid. 504 00:25:37,430 --> 00:25:40,630 Random sparse is held to tiles picked at random before 505 00:25:40,630 --> 00:25:43,748 training began. Deletion has neither limit: 506 00:25:43,748 --> 00:25:46,231 training chooses which tiles to delete. 507 00:25:46,231 --> 00:25:48,220 It just can't choose the values. 508 00:25:48,220 --> 00:25:52,340 And choosing where is exactly what the second habit takes. 509 00:25:52,680 --> 00:25:55,480 So the shape of the change explains only part of the 510 00:25:55,480 --> 00:25:59,240 story. What sets deletion apart most clearly is the 511 00:25:59,240 --> 00:26:01,043 measure from the last chapter: 512 00:26:01,043 --> 00:26:03,991 how far it moves the model's answers. 513 00:26:03,991 --> 00:26:06,360 On the same eight-billion-weight model, 514 00:26:06,360 --> 00:26:10,040 at the same point in training, deletion moves them dozens 515 00:26:10,040 --> 00:26:12,790 of times less than full fine-tuning. 516 00:26:17,048 --> 00:26:20,318 Reinforcement learning is only one way to teach. 517 00:26:20,318 --> 00:26:22,823 The other common one is plain imitation: 518 00:26:22,823 --> 00:26:26,170 show the model a written answer, and push it toward each 519 00:26:26,170 --> 00:26:29,290 word of it. That kind of training is known to change 520 00:26:29,290 --> 00:26:32,490 weights much more broadly, which should make it harder for 521 00:26:32,490 --> 00:26:34,375 a method that can only delete. 522 00:26:34,375 --> 00:26:36,970 The authors tried it on hard knowledge questions in 523 00:26:36,970 --> 00:26:40,330 fourteen subjects, and on twenty-three tough reasoning 524 00:26:40,330 --> 00:26:43,690 tasks, one mask per subject or task. 525 00:26:43,690 --> 00:26:46,650 At three billion weights, deletion matched full 526 00:26:46,650 --> 00:26:50,010 fine-tuning on both, and both clearly beat the original 527 00:26:50,010 --> 00:26:53,850 model. At one billion it won one and lost one: 528 00:26:53,850 --> 00:26:56,912 on the reasoning tasks it reached five point seven, 529 00:26:56,912 --> 00:26:59,976 where full fine-tuning reached eight point four. 530 00:26:59,976 --> 00:27:01,809 It also deletes more here: 531 00:27:01,809 --> 00:27:04,970 for this film we counted a quarter of a percent to one 532 00:27:04,970 --> 00:27:06,840 percent of the gate weights. 533 00:27:07,500 --> 00:27:10,534 The third way is a model teaching itself. 534 00:27:10,534 --> 00:27:13,762 A teacher copy gets to see a worked answer first. 535 00:27:13,762 --> 00:27:15,580 The student writes its own attempt, 536 00:27:15,580 --> 00:27:18,460 and at every word, it's pushed toward what the teacher 537 00:27:18,460 --> 00:27:22,460 would have said. The teacher is a slowly updated average 538 00:27:22,460 --> 00:27:25,420 of the student's recent versions, so it keeps up without 539 00:27:25,420 --> 00:27:29,020 jumping around. Tested on using software tools and on 540 00:27:29,020 --> 00:27:32,620 science questions, deletion matched full fine-tuning on 541 00:27:32,620 --> 00:27:34,860 both, within about a point either way, 542 00:27:34,860 --> 00:27:36,856 starting from far lower. 543 00:27:36,856 --> 00:27:39,740 These masks are bigger, twenty-seven and forty-seven 544 00:27:39,740 --> 00:27:41,984 megabytes, because they delete more, 545 00:27:41,984 --> 00:27:44,860 about point eight to one point eight percent of the gate 546 00:27:44,860 --> 00:27:46,213 weights by our count. 547 00:27:46,213 --> 00:27:49,050 A full copy is still twenty-nine thousand. 548 00:27:53,324 --> 00:27:54,882 So what does this buy you? 549 00:27:54,882 --> 00:27:57,218 Start with one copy of a model. 550 00:27:57,218 --> 00:28:00,191 The authors released thirty-seven masks for one 551 00:28:00,191 --> 00:28:03,630 three-billion-weight Llama, one for each subject and 552 00:28:03,630 --> 00:28:05,573 reasoning task it learned. 553 00:28:05,573 --> 00:28:09,453 Each is a small compressed file, four to eleven megabytes, 554 00:28:09,453 --> 00:28:12,068 where a full copy would be twelve thousand. 555 00:28:12,068 --> 00:28:15,150 Switching skills means restoring the original weights and 556 00:28:15,150 --> 00:28:18,110 zeroing a different set, instead of loading a whole new 557 00:28:18,110 --> 00:28:18,940 model. 558 00:28:19,600 --> 00:28:22,054 Masks are also easy to look inside, 559 00:28:22,054 --> 00:28:24,480 so for this film, we did. 560 00:28:24,480 --> 00:28:27,272 In the math mask for the eight-billion-weight Llama, 561 00:28:27,272 --> 00:28:30,709 the deletions are spread almost evenly across every layer, 562 00:28:30,709 --> 00:28:33,920 much as they started, about one gate weight in five 563 00:28:33,920 --> 00:28:37,520 hundred each, sprinkled across nearly every channel. 564 00:28:37,520 --> 00:28:39,840 Then we compared the fourteen masks that the 565 00:28:39,840 --> 00:28:42,800 one-billion-weight Llama learned for fourteen school 566 00:28:42,800 --> 00:28:46,400 subjects. Take any two subjects, and about a quarter of 567 00:28:46,400 --> 00:28:49,200 one's deletions are also in the other's. 568 00:28:49,200 --> 00:28:51,200 Part of that is a shared start: 569 00:28:51,200 --> 00:28:54,720 every subject began from the same random jitter and nearly 570 00:28:54,720 --> 00:28:56,404 the same pruning score. 571 00:28:56,404 --> 00:28:59,520 A hundred and twenty-seven thousand weights are deleted in 572 00:28:59,520 --> 00:29:02,640 all fourteen. That's almost exactly the size of the 573 00:29:02,640 --> 00:29:06,400 starting sliver, so that shared core is probably just 574 00:29:06,400 --> 00:29:08,121 where every mask began. 575 00:29:08,121 --> 00:29:10,408 How much of the rest is a shared start, 576 00:29:10,408 --> 00:29:13,840 and how much is shared skill, is an open question. 577 00:29:13,840 --> 00:29:16,480 And that's the kind of question the authors hope masks can 578 00:29:16,480 --> 00:29:20,000 answer: comparing what different skills delete could show 579 00:29:20,000 --> 00:29:22,510 where post-training changes a model. 580 00:29:23,170 --> 00:29:27,090 Now the fine print, which the authors state themselves. 581 00:29:27,090 --> 00:29:30,450 Because it forgets so little, deletion looks promising for 582 00:29:30,450 --> 00:29:33,250 learning one skill after another, but that hasn't been 583 00:29:33,250 --> 00:29:37,649 tested yet. Neither has combining masks to merge skills. 584 00:29:37,649 --> 00:29:39,250 And training isn't cheap: 585 00:29:39,250 --> 00:29:42,538 every gate weight needs its own dial while it learns, 586 00:29:42,538 --> 00:29:45,650 so it takes more memory than LoRA, though less than full 587 00:29:45,650 --> 00:29:49,011 fine-tuning. One idea they leave for later: 588 00:29:49,011 --> 00:29:52,130 today's masks are scattered, so they don't make the model 589 00:29:52,130 --> 00:29:56,370 smaller or faster. Masks in regular patterns could teach a 590 00:29:56,370 --> 00:29:59,090 model and compress it at the same time. 591 00:29:59,090 --> 00:30:01,491 Learning a mask isn't new in itself: 592 00:30:01,491 --> 00:30:04,370 it had been used to adapt image models and smaller 593 00:30:04,370 --> 00:30:08,210 language models. What's new is making the mask the whole 594 00:30:08,210 --> 00:30:10,884 of post-training for a large language model. 595 00:30:10,884 --> 00:30:14,121 And each result here comes from a single training run, 596 00:30:14,121 --> 00:30:17,570 so gaps of a few tenths of a point can't be told apart 597 00:30:17,570 --> 00:30:18,640 from noise. 598 00:30:19,300 --> 00:30:21,346 So, back to where we started. 599 00:30:21,346 --> 00:30:25,140 Billions of weights, and a new skill without a single one 600 00:30:25,140 --> 00:30:27,678 of them rewritten. Just a few million, 601 00:30:27,678 --> 00:30:30,624 chosen carefully, and set to zero. 602 00:30:30,624 --> 00:30:33,540 In the paper's words, these results suggest that 603 00:30:33,540 --> 00:30:36,500 post-training can work by selection alone: 604 00:30:36,500 --> 00:30:39,140 keeping nearly every pretrained weight, 605 00:30:39,140 --> 00:30:41,170 and choosing the few to delete. 606 00:30:42,630 --> 00:30:44,693 The new skill wasn't written in. 607 00:30:44,693 --> 00:30:46,500 It was carved out.