QuantizationInterpretability/research_summary.html at main · JackSwitzer/QuantizationInterpretability · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
<!DOCTYPE html>
<html lang="en">
<head>
    <meta charset="UTF-8">
    <meta name="viewport" content="width=device-width, initial-scale=1.0">
    <title>Quantization × Interpretability</title>
    <link href="https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@400;500;600&family=Source+Serif+4:ital,opsz,wght@0,8..60,300;0,8..60,400;0,8..60,600;0,8..60,700;1,8..60,400&display=swap" rel="stylesheet">
    <style>
        :root {
            --ink: #1a1a1a;
            --paper: #f8f6f1;
            --accent: #c15f3c;
            --accent-light: #e8d4cc;
            --success: #2d6a4f;
            --warning: #b5651d;
            --muted: #6b6b6b;
            --border: #d4d0c8;
            --bf16: #3b82f6;
            --fp16: #8b5cf6;
            --int8: #f59e0b;
            --int4: #ef4444;
        }

        * { margin: 0; padding: 0; box-sizing: border-box; }

        body {
            font-family: 'Source Serif 4', Georgia, serif;
            background: var(--paper);
            color: var(--ink);
            line-height: 1.6;
            font-size: 16px;
        }

        .container { max-width: 900px; margin: 0 auto; padding: 0 32px; }

        header {
            padding: 48px 0 32px;
            border-bottom: 1px solid var(--border);
            margin-bottom: 32px;
        }

        .eyebrow {
            font-family: 'JetBrains Mono', monospace;
            font-size: 10px;
            letter-spacing: 2px;
            text-transform: uppercase;
            color: var(--accent);
            margin-bottom: 8px;
        }

        h1 {
            font-size: 2.4rem;
            font-weight: 700;
            line-height: 1.1;
            margin-bottom: 12px;
        }

        .subtitle {
            font-size: 1.1rem;
            color: var(--muted);
            font-weight: 300;
        }

        /* Hero Stats - Compact */
        .hero-stats {
            display: grid;
            grid-template-columns: repeat(4, 1fr);
            gap: 1px;
            background: var(--border);
            border: 1px solid var(--border);
            margin: 24px 0;
        }

        .stat {
            background: var(--paper);
            padding: 16px 12px;
            text-align: center;
        }

        .stat-value {
            font-family: 'JetBrains Mono', monospace;
            font-size: 1.8rem;
            font-weight: 600;
            color: var(--accent);
            line-height: 1;
        }

        .stat-label {
            font-size: 0.7rem;
            color: var(--muted);
            margin-top: 4px;
            text-transform: uppercase;
            letter-spacing: 0.5px;
        }

        .stat-detail {
            font-size: 0.65rem;
            color: var(--muted);
            margin-top: 2px;
        }

        /* Collapsible Sections */
        .collapsible {
            border: 1px solid var(--border);
            margin-bottom: 8px;
            background: white;
        }

        .collapsible-header {
            display: flex;
            justify-content: space-between;
            align-items: center;
            padding: 16px 20px;
            cursor: pointer;
            user-select: none;
            transition: background 0.15s;
        }

        .collapsible-header:hover {
            background: rgba(0,0,0,0.02);
        }

        .collapsible-header h2 {
            font-size: 1.1rem;
            font-weight: 600;
            margin: 0;
            display: flex;
            align-items: center;
            gap: 12px;
        }

        .collapsible-header .tag {
            font-family: 'JetBrains Mono', monospace;
            font-size: 0.65rem;
            padding: 3px 8px;
            background: var(--accent);
            color: white;
            text-transform: uppercase;
            letter-spacing: 0.5px;
        }

        .collapsible-header .tag.finding { background: var(--success); }
        .collapsible-header .tag.method { background: var(--muted); }

        .collapsible-header .arrow {
            font-size: 1.2rem;
            color: var(--muted);
            transition: transform 0.2s;
        }

        .collapsible.open .arrow {
            transform: rotate(180deg);
        }

        .collapsible-content {
            display: none;
            padding: 0 20px 20px;
            border-top: 1px solid var(--border);
        }

        .collapsible.open .collapsible-content {
            display: block;
        }

        /* Bullet Lists - High Signal */
        .bullet-list {
            list-style: none;
            margin: 16px 0;
        }

        .bullet-list li {
            padding: 8px 0;
            padding-left: 20px;
            position: relative;
            border-bottom: 1px solid #f0f0f0;
        }

        .bullet-list li:last-child {
            border-bottom: none;
        }

        .bullet-list li::before {
            content: "→";
            position: absolute;
            left: 0;
            color: var(--accent);
            font-weight: 600;
        }

        .bullet-list .key {
            font-weight: 600;
            color: var(--ink);
        }

        .bullet-list .value {
            font-family: 'JetBrains Mono', monospace;
            color: var(--accent);
            font-size: 0.9em;
        }

        .bullet-list .note {
            color: var(--muted);
            font-size: 0.85em;
        }

        /* Data Tables - Compact */
        .data-table {
            width: 100%;
            border-collapse: collapse;
            margin: 16px 0;
            font-size: 0.85rem;
        }

        .data-table th {
            font-family: 'JetBrains Mono', monospace;
            font-size: 0.65rem;
            text-transform: uppercase;
            letter-spacing: 0.5px;
            text-align: left;
            padding: 8px 12px;
            background: var(--ink);
            color: var(--paper);
            font-weight: 500;
        }

        .data-table td {
            padding: 8px 12px;
            border-bottom: 1px solid var(--border);
        }

        .data-table .highlight { color: var(--accent); font-weight: 600; }
        .data-table .success { color: var(--success); font-weight: 600; }
        .data-table .warning { color: var(--warning); font-weight: 600; }

        /* Key Insight Box */
        .insight {
            background: var(--ink);
            color: var(--paper);
            padding: 16px 20px;
            margin: 16px 0;
            font-size: 0.95rem;
        }

        .insight strong {
            color: var(--accent-light);
        }

        /* Scope Grid */
        .scope-grid {
            display: grid;
            grid-template-columns: repeat(4, 1fr);
            gap: 16px;
            margin: 16px 0;
        }

        .scope-item {
            text-align: center;
            padding: 12px;
            background: #f5f5f5;
        }

        .scope-item .num {
            font-family: 'JetBrains Mono', monospace;
            font-size: 1.5rem;
            font-weight: 600;
            color: var(--ink);
        }

        .scope-item .lbl {
            font-size: 0.7rem;
            color: var(--muted);
            text-transform: uppercase;
            margin-top: 2px;
        }

        /* Precision Legend */
        .precision-legend {
            display: flex;
            gap: 16px;
            margin: 16px 0 8px;
            font-size: 0.75rem;
            font-family: 'JetBrains Mono', monospace;
        }

        .precision-legend span {
            display: flex;
            align-items: center;
            gap: 6px;
        }

        .precision-legend .dot {
            width: 10px;
            height: 10px;
            border-radius: 2px;
        }

        .precision-legend .dot.bf16 { background: var(--bf16); }
        .precision-legend .dot.fp16 { background: var(--fp16); }
        .precision-legend .dot.int8 { background: var(--int8); }
        .precision-legend .dot.int4 { background: var(--int4); }

        /* Degradation Chart */
        .degradation-chart {
            margin: 16px 0;
        }

        .degrade-row {
            display: grid;
            grid-template-columns: 90px 1fr;
            align-items: center;
            margin: 16px 0;
            gap: 12px;
        }

        .degrade-row .task {
            font-size: 0.85rem;
            font-weight: 600;
        }

        .degrade-row .sub {
            font-size: 0.65rem;
            color: var(--muted);
        }

        .model-row {
            display: flex;
            align-items: center;
            gap: 8px;
            margin: 4px 0;
        }

        .model-label {
            font-family: 'JetBrains Mono', monospace;
            font-size: 0.6rem;
            color: var(--muted);
            width: 70px;
            text-align: right;
        }

        .stacked-bar {
            display: flex;
            height: 20px;
            border-radius: 2px;
            overflow: hidden;
            flex: 1;
            background: #e5e5e5;
        }

        .bar-segment {
            display: flex;
            align-items: center;
            justify-content: center;
            font-family: 'JetBrains Mono', monospace;
            font-size: 0.65rem;
            color: white;
            font-weight: 500;
        }

        .bar-segment.bf16-lost { background: #d97706; }
        .bar-segment.int4 { background: #fbbf24; color: #78350f; }
        .bar-segment.same {
            background: repeating-linear-gradient(
                45deg,
                #d97706,
                #d97706 4px,
                #fbbf24 4px,
                #fbbf24 8px
            );
            color: white;
            text-shadow: 0 0 2px rgba(0,0,0,0.5);
        }

        /* Tick marks for chart */
        .chart-axis {
            display: grid;
            grid-template-columns: 90px 1fr;
            gap: 12px;
            margin-bottom: 8px;
            padding-left: 78px;
        }

        .tick-marks {
            display: flex;
            justify-content: space-between;
            font-family: 'JetBrains Mono', monospace;
            font-size: 0.6rem;
            color: var(--muted);
            padding: 0 2px;
        }

        .tick-marks span {
            position: relative;
        }

        .tick-marks span::before {
            content: '';
            position: absolute;
            top: -4px;
            left: 50%;
            transform: translateX(-50%);
            width: 1px;
            height: 3px;
            background: var(--muted);
        }

        /* Nested collapsible for examples */
        .nested-collapsible {
            border: 1px solid var(--border);
            margin: 16px 0;
            background: #fafafa;
        }

        .nested-header {
            display: flex;
            justify-content: space-between;
            align-items: center;
            padding: 10px 16px;
            cursor: pointer;
            font-size: 0.85rem;
            font-weight: 500;
        }

        .nested-header:hover {
            background: rgba(0,0,0,0.02);
        }

        .nested-content {
            display: none;
            padding: 0 16px 16px;
            border-top: 1px solid var(--border);
        }

        .nested-collapsible.open .nested-content {
            display: block;
        }

        .nested-collapsible.open .nested-arrow {
            transform: rotate(180deg);
        }

        .nested-arrow {
            font-size: 0.9rem;
            color: var(--muted);
            transition: transform 0.2s;
        }

        /* Code Example */
        .code-example {
            background: #1e1e1e;
            border-radius: 4px;
            padding: 16px;
            margin: 16px 0;
            font-family: 'JetBrains Mono', monospace;
            font-size: 0.8rem;
            overflow-x: auto;
        }

        .code-example .label {
            color: #888;
            font-size: 0.7rem;
            margin-bottom: 8px;
        }

        .code-example pre {
            color: #d4d4d4;
            margin: 0;
            white-space: pre-wrap;
        }

        .code-example .comment { color: #6a9955; }
        .code-example .keyword { color: #569cd6; }
        .code-example .string { color: #ce9178; }
        .code-example .truncated { color: #f44; background: rgba(244,68,68,0.2); padding: 0 4px; }

        /* Footer */
        footer {
            border-top: 1px solid var(--border);
            padding: 24px 0;
            margin-top: 32px;
            text-align: center;
            color: var(--muted);
            font-size: 0.8rem;
        }

        @media (max-width: 768px) {
            .hero-stats, .scope-grid { grid-template-columns: repeat(2, 1fr); }
            h1 { font-size: 1.8rem; }
            .container { padding: 0 16px; }
        }
    </style>
</head>
<body>
    <div class="container">
        <header>
            <div class="eyebrow">Anthropic Fellows Program</div>
            <h1>Quantization × Interpretability</h1>
            <p class="subtitle">Do SAEs trained on BF16 transfer to INT4? What capabilities break when you quantize?</p>
        </header>

        <!-- Hero Stats -->
        <div class="hero-stats">
            <div class="stat">
                <div class="stat-value">99%</div>
                <div class="stat-label">SAE Transfer</div>
                <div class="stat-detail">BF16→INT4</div>
            </div>
            <div class="stat">
                <div class="stat-value">-50%</div>
                <div class="stat-label">Code (HumanEval)</div>
                <div class="stat-detail">40%→20%</div>
            </div>
            <div class="stat">
                <div class="stat-value">0%</div>
                <div class="stat-label">Knowledge (MMLU)</div>
                <div class="stat-detail">100%→100%</div>
            </div>
            <div class="stat">
                <div class="stat-value">0.5×</div>
                <div class="stat-label">Best SAE</div>
                <div class="stat-detail">beats 8× by 2.3×</div>
            </div>
        </div>

        <!-- TL;DR -->
        <div class="insight">
            <strong>TL;DR:</strong> BF16-trained SAEs work on INT4 models (99% correlation). Code generation breaks at INT4 (-50%), knowledge retrieval survives (0% loss). Undercomplete SAEs (0.5×) transfer 2.3× better than overcomplete (8×).
        </div>

        <!-- ==================== KEY FINDINGS ==================== -->

        <!-- Finding 1: Degradation -->
        <div class="collapsible open">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag finding">Finding 1</span> Degradation has structure</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <div class="precision-legend">
                    <span><span class="dot" style="background: #fbbf24;"></span> INT4</span>
                    <span><span class="dot" style="background: #d97706;"></span> BF16 total</span>
                    <span><span class="dot" style="background: repeating-linear-gradient(45deg, #d97706, #d97706 2px, #fbbf24 2px, #fbbf24 4px);"></span> Same</span>
                </div>

                <div class="degradation-chart">
                    <!-- Tick marks -->
                    <div class="chart-axis">
                        <div></div>
                        <div class="tick-marks">
                            <span>0</span>
                            <span>25</span>
                            <span>50</span>
                            <span>75</span>
                            <span>100</span>
                        </div>
                    </div>
                    <!-- Code / HumanEval -->
                    <div class="degrade-row">
                        <div><div class="task">Code</div><div class="sub">HumanEval</div></div>
                        <div>
                            <div class="model-row">
                                <span class="model-label">Qwen3</span>
                                <div class="stacked-bar">
                                    <div class="bar-segment int4" style="width: 20%;">20%</div>
                                    <div class="bar-segment bf16-lost" style="width: 20%;">40%</div>
                                </div>
                            </div>
                            <div class="model-row">
                                <span class="model-label">StarCoder2</span>
                                <div class="stacked-bar">
                                    <div class="bar-segment int4" style="width: 18%;">18%</div>
                                    <div class="bar-segment bf16-lost" style="width: 17%;">35%</div>
                                </div>
                            </div>
                        </div>
                    </div>
                    <!-- Math / GSM8K -->
                    <div class="degrade-row">
                        <div><div class="task">Math</div><div class="sub">GSM8K</div></div>
                        <div>
                            <div class="model-row">
                                <span class="model-label">Qwen3</span>
                                <div class="stacked-bar">
                                    <div class="bar-segment int4" style="width: 87%;">87%</div>
                                    <div class="bar-segment bf16-lost" style="width: 6%;">93%</div>
                                </div>
                            </div>
                            <div class="model-row">
                                <span class="model-label">StarCoder2</span>
                                <div class="stacked-bar">
                                    <div class="bar-segment int4" style="width: 71%;">71%</div>
                                    <div class="bar-segment bf16-lost" style="width: 7%;">78%</div>
                                </div>
                            </div>
                        </div>
                    </div>
                    <!-- Knowledge / MMLU-CS -->
                    <div class="degrade-row">
                        <div><div class="task">Knowledge</div><div class="sub">MMLU-CS</div></div>
                        <div>
                            <div class="model-row">
                                <span class="model-label">Qwen3</span>
                                <div class="stacked-bar">
                                    <div class="bar-segment same" style="width: 100%;">100%</div>
                                </div>
                            </div>
                            <div class="model-row">
                                <span class="model-label">StarCoder2</span>
                                <div class="stacked-bar">
                                    <div class="bar-segment same" style="width: 93%;">93%</div>
                                </div>
                            </div>
                        </div>
                    </div>
                </div>

                <div class="insight">
                    <strong>Pattern:</strong> Generative tasks (code) break first. Discriminative tasks (knowledge) survive. Quantization adds noise to generation, not retrieval.
                </div>

                <!-- Nested collapsible for code example -->
                <div class="nested-collapsible">
                    <div class="nested-header" onclick="this.parentElement.classList.toggle('open')">
                        <span>Example: HumanEval/8 - sum_product function</span>
                        <span class="nested-arrow">▼</span>
                    </div>
                    <div class="nested-content">
                        <div class="code-example">
                            <div class="label">BF16 (passes)</div>
                            <pre><span class="keyword">if not</span> numbers: <span class="keyword">return</span> (<span class="string">0</span>, <span class="string">1</span>)
total_sum = sum(numbers)
total_product = <span class="string">1</span>
<span class="keyword">for</span> num <span class="keyword">in</span> numbers: total_product *= num
<span class="keyword">return</span> (total_sum, total_product)</pre>
                        </div>
                        <div class="code-example">
                            <div class="label">INT4 (fails - truncated)</div>
                            <pre><span class="comment"># Handle empty list case</span>
<span class="keyword">if not</span> numbers: <span class="keyword">return</span> (<span class="string">0</span>, <span class="string">1</span>)
<span class="comment"># Initialize sum and product</span>
total_sum = <span class="string">0</span>  <span class="comment"># different approach</span>
total_product = <span class="string">1</span>
<span class="comment"># Iterate through the list</span>
<span class="keyword">for</span> num <span class="keyword">in</span> <span class="truncated">← truncated</span></pre>
                        </div>
                        <ul class="bullet-list">
                            <li><span class="key">INT4 adds verbose comments</span> <span class="note">wastes tokens before completing logic</span></li>
                            <li><span class="key">Changes algorithm</span> <span class="note">manual sum loop vs built-in sum()</span></li>
                            <li><span class="key">Gets truncated</span> <span class="note">same token limit, less code</span></li>
                        </ul>
                    </div>
                </div>
            </div>
        </div>

        <!-- Finding 2: Transfer -->
        <div class="collapsible">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag finding">Finding 2</span> SAEs transfer across precisions</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <ul class="bullet-list">
                    <li><span class="key">Sample correlation:</span> <span class="value">99%</span> <span class="note">Same inputs → same features fire</span></li>
                    <li><span class="key">Top-10 agreement:</span> <span class="value">89%</span> <span class="note">9/10 most active features match</span></li>
                    <li><span class="key">Feature correlation:</span> <span class="value">95%</span> <span class="note">Each feature behaves consistently</span></li>
                    <li><span class="key">Sparsity agreement:</span> <span class="value">99%</span> <span class="note">Same features stay silent/active</span></li>
                </ul>

                <table class="data-table">
                    <thead>
                        <tr><th>Metric</th><th>FP16</th><th>INT8</th><th>INT4</th></tr>
                    </thead>
                    <tbody>
                        <tr><td>Sample Correlation</td><td class="success">99.9%</td><td class="success">99.7%</td><td class="highlight">99.0%</td></tr>
                        <tr><td>Feature Correlation</td><td class="success">99.7%</td><td class="success">98.5%</td><td class="highlight">95.3%</td></tr>
                        <tr><td>Top-10 Agreement</td><td class="success">97.2%</td><td class="success">94.1%</td><td class="highlight">88.9%</td></tr>
                    </tbody>
                </table>

                <div class="insight">
                    <strong>Implication:</strong> Train interpretability tools on expensive BF16, deploy monitoring on cheap INT4. Features mean the same thing across precisions.
                </div>
            </div>
        </div>

        <!-- ==================== METHODOLOGY ==================== -->

        <!-- Method -->
        <div class="collapsible">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag method">Method</span> How we ran experiments</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <div class="scope-grid">
                    <div class="scope-item"><div class="num">2</div><div class="lbl">Models</div></div>
                    <div class="scope-item"><div class="num">4</div><div class="lbl">Precisions</div></div>
                    <div class="scope-item"><div class="num">168</div><div class="lbl">SAEs</div></div>
                    <div class="scope-item"><div class="num">16h</div><div class="lbl">H100 Time</div></div>
                </div>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">Models</h4>
                <ul class="bullet-list">
                    <li><span class="key">Qwen3-Coder-30B-A3B:</span> <span class="note">MoE architecture, d_model=2048, 48 layers</span></li>
                    <li><span class="key">StarCoder2-15B:</span> <span class="note">Dense architecture, d_model=6144, 40 layers</span></li>
                </ul>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">Precisions</h4>
                <ul class="bullet-list">
                    <li><span class="key">BF16:</span> <span class="note">torch.bfloat16 - baseline full precision</span></li>
                    <li><span class="key">FP16:</span> <span class="note">torch.float16 - half precision</span></li>
                    <li><span class="key">INT8:</span> <span class="note">bitsandbytes load_in_8bit</span></li>
                    <li><span class="key">INT4:</span> <span class="note">bitsandbytes NF4 quantization</span></li>
                </ul>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">SAE Training</h4>
                <ul class="bullet-list">
                    <li><span class="key">Hidden dim:</span> <span class="value">0.5× d_model</span> <span class="note">(undercomplete)</span></li>
                    <li><span class="key">L1 coefficient:</span> <span class="value">5e-4</span></li>
                    <li><span class="key">Epochs:</span> <span class="value">1000</span></li>
                    <li><span class="key">Training data:</span> <span class="value">500 coding prompts</span></li>
                </ul>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">Sample Prompts Used</h4>
                <div class="code-example">
                    <pre><span class="string">"Write a Python function to check if a number is prime"</span>
<span class="string">"Implement a binary search algorithm"</span>
<span class="string">"Create a function that reverses a linked list"</span>
<span class="string">"Write a recursive fibonacci function"</span>
<span class="comment"># ... 500 coding prompts from CodeSearchNet</span></pre>
                </div>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">Pipeline</h4>
                <ul class="bullet-list">
                    <li>1. Load model at each precision (BF16, FP16, INT8, INT4)</li>
                    <li>2. Extract activations at 7-11 layers (early, middle, late)</li>
                    <li>3. Train SAE per precision per layer</li>
                    <li>4. Procrustes alignment: compare BF16 features vs quantized</li>
                    <li>5. Semantic transfer: apply BF16-trained SAE to INT4 activations</li>
                    <li>6. Benchmark: HumanEval (10), GSM8K (15), MMLU-CS (15)</li>
                </ul>
            </div>
        </div>

        <!-- Context -->
        <div class="collapsible">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag method">Context</span> Why this matters</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <h4 style="margin-top: 12px; font-size: 0.9rem;">Why this question matters</h4>
                <ul class="bullet-list">
                    <li><span class="key">Anthropic's mission:</span> <span class="note">Interpretability tools must work on deployed models. Production uses INT4, research uses BF16.</span></li>
                    <li><span class="key">Safety gap:</span> <span class="note">If SAE features don't transfer, safety monitoring breaks when you quantize.</span></li>
                </ul>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">Why code models?</h4>
                <ul class="bullet-list">
                    <li><span class="key">Ground truth:</span> <span class="note">Code runs or it doesn't. No ambiguity in measuring degradation.</span></li>
                    <li><span class="key">Structured features:</span> <span class="note">Functions, loops, types → easier to verify feature meaning across precisions</span></li>
                    <li><span class="key">Practical relevance:</span> <span class="note">Code models are deployed quantized in IDEs, APIs</span></li>
                </ul>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">Personal interest</h4>
                <ul class="bullet-list">
                    <li><span class="key">Background:</span> <span class="note">Worked on benchmark infra at IBM, saw inconsistencies that were hard to explain</span></li>
                    <li><span class="key">Question:</span> <span class="note">Can we see inside these models to understand what breaks?</span></li>
                </ul>
            </div>
        </div>

        <!-- ==================== MORE FINDINGS ==================== -->

        <!-- Finding 3: Undercomplete -->
        <div class="collapsible">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag finding">Finding 3</span> Smaller SAEs transfer better</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <p style="margin: 12px 0; color: var(--muted); font-size: 0.9rem;">
                    Counter-intuitively, SAEs with <em>fewer</em> features (0.5× model dimension) transfer across precisions 2.3× better than larger SAEs (8×).
                </p>

                <table class="data-table">
                    <thead>
                        <tr><th>SAE Size</th><th>Alignment</th><th>Features Alive</th><th></th></tr>
                    </thead>
                    <tbody>
                        <tr><td class="highlight">0.5× (1024 features)</td><td class="highlight">79%</td><td>60%</td><td class="highlight">Best</td></tr>
                        <tr><td>1× (2048 features)</td><td>71%</td><td>46%</td><td>Baseline</td></tr>
                        <tr><td>2× (4096 features)</td><td>56%</td><td>20%</td><td>Common default</td></tr>
                        <tr><td class="warning">8× (16384 features)</td><td class="warning">34%</td><td class="warning">1%</td><td class="warning">Worst</td></tr>
                    </tbody>
                </table>

                <div class="insight">
                    <strong>Intuition:</strong> With limited capacity, the SAE must learn only the most fundamental features - ones that exist regardless of precision. Larger SAEs have room to memorize precision-specific artifacts that don't transfer.
                </div>

                <ul class="bullet-list">
                    <li><span class="key">Compression forces generalization:</span> <span class="note">0.5× SAE can only fit ~1024 features, so it picks the universal ones</span></li>
                    <li><span class="key">Overcomplete SAEs overfit:</span> <span class="note">8× SAE learns 16k features including BF16-specific noise patterns</span></li>
                    <li><span class="key">Dead features correlate with poor transfer:</span> <span class="note">8× has 99% dead features, 0.5× has 40% dead</span></li>
                </ul>
            </div>
        </div>

        <!-- ==================== EVIDENCE ==================== -->

        <!-- Evidence -->
        <div class="collapsible">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag method">Evidence</span> Why this is real</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <ul class="bullet-list">
                    <li><span class="key">Multiple metrics converge:</span> <span class="note">Sample, feature, top-k, sparsity, Procrustes all agree</span></li>
                    <li><span class="key">Cross-architecture:</span> <span class="note">Qwen3-30B (MoE) + StarCoder2-15B (Dense) both show 85-89% Procrustes alignment</span></li>
                    <li><span class="key">Layer consistency:</span> <span class="note">7-11 layers tested, holds across early/middle/late</span></li>
                    <li><span class="key">Systematic ablations:</span> <span class="note">5 hidden dims, 7 L1 values, 6 epoch counts, 7 data sizes</span></li>
                </ul>

                <h4 style="margin-top: 16px; font-size: 0.9rem;">Limitations</h4>
                <ul class="bullet-list">
                    <li><span class="key">Sample size:</span> <span class="note">500 prompts, 10-15 benchmark samples per task</span></li>
                    <li><span class="key">Models:</span> <span class="note">Code-focused only. Claude/GPT untested</span></li>
                </ul>
            </div>
        </div>

        <!-- ==================== NEXT / IMPACT ==================== -->

        <!-- Next -->
        <div class="collapsible">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag">Next</span> Future directions</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <ul class="bullet-list">
                    <li><span class="key">Feature interpretation:</span> <span class="note">Verify "function def" feature means same thing across precisions</span></li>
                    <li><span class="key">Circuit transfer:</span> <span class="note">Do attention patterns survive quantization?</span></li>
                    <li><span class="key">Broader models:</span> <span class="note">Llama, Mistral, Claude-style architectures</span></li>
                    <li><span class="key">Quant methods:</span> <span class="note">NF4 vs GPTQ vs AWQ</span></li>
                </ul>
            </div>
        </div>

        <!-- Impact -->
        <div class="collapsible">
            <div class="collapsible-header" onclick="this.parentElement.classList.toggle('open')">
                <h2><span class="tag">Impact</span> Why this matters for safety</h2>
                <span class="arrow">▼</span>
            </div>
            <div class="collapsible-content">
                <ul class="bullet-list">
                    <li><span class="key">The gap:</span> <span class="note">Interpretability research uses BF16. Production uses INT4.</span></li>
                    <li><span class="key">The question:</span> <span class="note">Do safety guarantees from BF16 transfer to INT4?</span></li>
                    <li><span class="key">Our answer:</span> <span class="note">Features transfer (99%). SAEs can be cross-precision monitors.</span></li>
                    <li><span class="key">Practical:</span> <span class="note">Train expensive interpretability on BF16, deploy monitoring on cheap INT4</span></li>
                </ul>
            </div>
        </div>

        <footer>
            <p>Quantization × Interpretability | Anthropic Fellows Program | 2026</p>
        </footer>
    </div>

    <script>
        // Open first finding by default
        document.querySelectorAll('.collapsible')[0].classList.add('open');
    </script>
</body>
</html>