aboutsummaryrefslogtreecommitdiff
path: root/downloads/evolution_plan/presentation.html
blob: 9830d4e3f53fda5ffe820a9f91cbae6a7f8898fa (plain)
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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="black-translucent">
<meta name="theme-color" content="#0d0d0d">
<title>🔥 Core Model Evolution — Presentation</title>
<style>
:root{
  --bg:#0d0d0d;--panel:#1a1a1a;--panel2:#222;--border:#333;
  --accent:#ff4d4d;--accent2:#ff8c8c;--gold:#ffca28;
  --green:#4caf50;--blue:#42a5f5;--text:#e0e0e0;--muted:#888;
}
*{box-sizing:border-box;margin:0;padding:0}
html,body{height:100%;overflow:hidden}
body{
  font-family:-apple-system,'Segoe UI',sans-serif;
  line-height:1.65;color:var(--text);background:var(--bg);
  -webkit-tap-highlight-color:transparent;
  -webkit-font-smoothing:antialiased;
}

/* ── Slide System ── */
#deck{height:100vh;height:100dvh;width:100vw;overflow:hidden;position:relative}
.slide{
  position:absolute;top:0;left:0;width:100%;height:100%;
  padding:20px 18px 70px;overflow-y:auto;
  opacity:0;transform:translateX(30px);
  transition:opacity .35s ease,transform .35s ease;
  pointer-events:none;-webkit-overflow-scrolling:touch;
}
.slide.active{opacity:1;transform:translateX(0);pointer-events:auto}
.slide.prev{transform:translateX(-30px)}

/* ── Progress Dots ── */
#dots{
  position:fixed;bottom:52px;left:0;right:0;
  display:flex;gap:6px;justify-content:center;
  z-index:100;padding:6px 0;
}
.dot{width:8px;height:8px;border-radius:50%;background:#444;transition:all .3s;cursor:pointer}
.dot.active{background:var(--accent);width:24px;border-radius:4px}
.dot.done{background:var(--green)}

/* ── Nav Bar ── */
#nav{
  position:fixed;bottom:0;left:0;right:0;height:44px;
  display:flex;align-items:center;justify-content:space-between;
  padding:0 14px;background:rgba(13,13,13,.95);
  border-top:1px solid var(--border);z-index:100;
  padding-bottom:env(safe-area-inset-bottom,0px);
}
.nav-btn{
  background:#2a2a2a;border:1px solid var(--border);color:var(--text);
  padding:7px 16px;border-radius:8px;font-size:13px;cursor:pointer;
  white-space:nowrap;min-width:64px;text-align:center;
}
.nav-btn:active{background:var(--accent);color:#fff}
.nav-btn:disabled{opacity:.35;cursor:default}
#counter{font-size:12px;color:var(--muted);font-variant-numeric:tabular-nums}

/* ── Typography ── */
h1{color:var(--accent);font-size:1.5em;line-height:1.3;margin-bottom:8px}
h2{color:var(--accent2);font-size:1.2em;margin:18px 0 10px;border-left:4px solid var(--accent);padding-left:10px}
h3{color:var(--gold);font-size:1.05em;margin:14px 0 8px}
p{margin:8px 0;font-size:14px}
ul{margin:8px 0;padding-left:18px;font-size:14px}
li{margin:4px 0}
strong{color:#fff}
code{background:#2a2a2a;padding:2px 5px;border-radius:3px;font-size:12px;color:#ff8c8c}
.subtitle{text-align:center;font-style:italic;color:#aaa;font-size:13px;margin-bottom:16px}
.tagline{text-align:center;font-size:11px;color:var(--muted);margin-bottom:20px}

/* ── Cards ── */
.card{
  background:var(--panel2);border:1px solid var(--border);
  border-radius:12px;padding:16px;margin:12px 0;
}
.card-accent{border-left:4px solid var(--accent)}
.card-green{border-left:4px solid var(--green);background:#15201a}
.card-blue{border-left:4px solid var(--blue);background:#1a1a2a}
.card-gold{border-left:4px solid var(--gold);background:#1f1a0d}
.card-warn{border-left:4px solid #ff6b6b;background:#2a2020}

/* ── Badges ── */
.badge{
  display:inline-block;padding:3px 10px;border-radius:20px;
  font-size:11px;margin:2px 4px 2px 0;
}
.badge.green{background:#1b5e20;color:var(--green)}
.badge.red{background:#b71c1c;color:#ef9a9a}
.badge.blue{background:#0d4741;color:var(--blue)}
.badge.gold{background:#4e3402;color:var(--gold)}
.badge.gray{background:#333;color:#aaa}

/* ── Cost Tags ── */
.cost{
  display:inline-block;padding:5px 14px;border-radius:8px;
  font-weight:bold;font-size:13px;color:#fff;
}
.cost.low{background:var(--green)}
.cost.mid{background:var(--blue)}
.cost.high{background:var(--accent)}

/* ── Tables ── */
.tbl{width:100%;border-collapse:collapse;margin:10px 0;font-size:12px}
.tbl th{background:#2a2a2a;color:var(--accent2);padding:7px 8px;border:1px solid var(--border);text-align:left}
.tbl td{padding:7px 8px;border:1px solid var(--border)}
.tbl tr:nth-child(even){background:#1e1e1e}
.tbl tr.total{background:#2a1a1a;font-weight:bold}
.tscroll{overflow-x:auto;-webkit-overflow-scrolling:touch;margin:10px -4px}

/* ── Grid ── */
.grid2{display:grid;grid-template-columns:1fr 1fr;gap:10px;margin:10px 0}
@media(max-width:480px){.grid2{grid-template-columns:1fr}}

/* ── Mini cards ── */
.mini{background:#1a1a1a;padding:12px;border-radius:8px;border:1px solid var(--border)}
.mini h4{margin:0 0 6px;font-size:13px;color:var(--accent2)}
.mini p{margin:0;font-size:12px;color:#bbb;line-height:1.5}

/* ── Timeline ── */
.tl-item{display:flex;gap:10px;margin:10px 0;align-items:flex-start}
.tl-dot{width:12px;height:12px;border-radius:50%;flex-shrink:0;margin-top:4px}
.tl-dot.done{background:var(--green)}
.tl-dot.active{background:var(--accent);box-shadow:0 0 8px var(--accent)}
.tl-dot.future{background:#444;border:2px solid #555}
.tl-content{flex:1}
.tl-content h4{font-size:13px;margin:0 0 4px;color:var(--accent2)}
.tl-content p{font-size:12px;color:#bbb;margin:0}
.tl-content .when{font-size:11px;color:var(--muted)}

/* ── Progress bars ── */
.pbar{height:8px;background:#333;border-radius:4px;margin:6px 0;overflow:hidden}
.pbar-fill{height:100%;border-radius:4px;transition:width 1s}
.pf-green{background:var(--green)}
.pf-blue{background:var(--blue)}
.pf-red{background:var(--accent)}
.pf-gold{background:var(--gold)}

/* ── Stat blocks ── */
.stat-row{display:flex;gap:8px;flex-wrap:wrap;margin:10px 0}
.stat{flex:1;min-width:90px;background:#1a1a1a;border:1px solid var(--border);border-radius:8px;padding:10px;text-align:center}
.stat .num{font-size:1.4em;font-weight:bold;color:var(--accent)}
.stat .lbl{font-size:11px;color:var(--muted);margin-top:2px}
.stat .num.green{color:var(--green)}
.stat .num.blue{color:var(--blue)}
.stat .num.gold{color:var(--gold)}

/* ── Hero ── */
.hero{text-align:center;padding:20px 0}
.hero h1{font-size:1.7em;margin-bottom:12px}
.hero .emoji{font-size:2.5em;margin-bottom:8px}
.hero p{color:#bbb;font-size:14px}
.hero .fire{color:var(--accent);font-weight:bold}

/* ── Checklist ── */
.check-item{display:flex;gap:8px;align-items:flex-start;margin:8px 0;font-size:13px}
.check-icon{flex-shrink:0;width:20px;height:20px;border-radius:50%;display:flex;align-items:center;justify-content:center;font-size:11px;font-weight:bold}
.check-icon.done{background:var(--green);color:#0d0d0d}
.check-icon.pending{background:#333;color:#888;border:1px solid #444}

/* ── Divider ── */
.divider{height:1px;background:var(--border);margin:16px 0}

/* ── Swiper hint ── */
.swipe-hint{text-align:center;font-size:11px;color:var(--muted);margin-top:auto;padding:8px}

/* ── Ask block ── */
.ask-box{
  background:linear-gradient(135deg,#1a1a2a,#2a1a1a);
  border:2px solid var(--accent);border-radius:14px;
  padding:18px;text-align:center;margin:16px 0;
}
.ask-box .amount{font-size:2em;font-weight:bold;color:var(--accent);margin:8px 0}
.ask-box .desc{font-size:13px;color:#ccc}

/* ── Mobile tweaks ── */
@media(max-width:380px){
  h1{font-size:1.3em}
  h2{font-size:1.05em}
  .tbl{font-size:11px}
  .tbl th,.tbl td{padding:5px 6px}
  .stat .num{font-size:1.1em}
}
</style>
</head>
<body>
<div id="deck">

<!-- ═══════════════ SLIDE 1: TITLE ═══════════════ -->
<div class="slide active" data-title="Title">
<div class="hero">
<div class="emoji">🔥</div>
<h1>Core Model Evolution</h1>
<p class="subtitle">From Session Logs to Core Weights<br>Recursive Cognitive Evolution</p>
<div class="tagline">Year of the Firehorse 🔥<br>July 2026 · v3.0 Presentation</div>
<div class="card card-accent" style="text-align:left">
<h3 style="margin-top:0">What This Is</h3>
<p>A realistic, researched plan to evolve our AI from a general-purpose assistant into a specialized intelligence — trained on our collaboration history, our decisions, and our style.</p>
<p style="margin-bottom:0"><strong>Architecture target:</strong> Sparse MoE + Multi-Token Generation</p>
</div>
<div class="swipe-hint">← Swipe or tap arrows to navigate →</div>
</div>
</div>

<!-- ═══════════════ SLIDE 2: PROGRESS SUMMARY ═══════════════ -->
<div class="slide" data-title="Progress">
<h1>✅ What We've Done So Far</h1>
<p style="color:#aaa;margin-bottom:14px">A summary of progress before we talk budget.</p>

<h2>Completed Milestones</h2>
<div class="tl-item">
<div class="tl-dot done"></div>
<div class="tl-content">
<h4>System Foundation Built</h4>
<p>Full operating environment with 44 modules, 17 skills, 293 files, 3 agents, memory graph (67 nodes), and VFS filesystem.</p>
<span class="when">Ongoing since early 2026</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot done"></div>
<div class="tl-content">
<h4>Collaboration History Captured</h4>
<p>Session logs, reasoning chains, command history, and decision patterns stored across sessions. This is the raw training data.</p>
<span class="when">Continuous</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot done"></div>
<div class="tl-content">
<h4>Training Strategy Defined</h4>
<p>QLoRA for Phase 1 (efficient, cheap), scaling to deep fine-tuning of existing MoE models. NOT training from scratch — 10-50x cheaper.</p>
<span class="when">Researched & validated</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot done"></div>
<div class="tl-content">
<h4>GPU Market Research</h4>
<p>Researched actual pricing across RunPod, Lambda Labs, Vast.ai, Hetzner. Spot instances identified as key cost saver (50-70% off).</p>
<span class="when">July 2026</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot done"></div>
<div class="tl-content">
<h4>Detailed Budget Plan v2.0</h4>
<p>4-phase plan with itemized costs, risk matrix, and 4 fallback plans. Saved as HTML document for review.</p>
<span class="when">This session</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot done"></div>
<div class="tl-content">
<h4>Architecture Decision: MoE + Sparse + Multi-Token</h4>
<p>Decided on Sparse Mixture of Experts (like Mixtral/DeepSeek) with Multi-Token Generation (Medusa heads) for speed + intelligence.</p>
<span class="when">This session</span>
</div>
</div>

<h2>Current Status</h2>
<div class="stat-row">
<div class="stat"><div class="num green">6/6</div><div class="lbl">Planning tasks done</div></div>
<div class="stat"><div class="num blue">0</div><div class="lbl">Phases executed</div></div>
<div class="stat"><div class="num gold">€0</div><div class="lbl">Spent so far</div></div>
</div>

<div class="card card-green">
<strong>✅ Ready to start Phase 0</strong> — we have the data, the plan, the architecture decision, and the budget. We need the funding to begin.
</div>

<div class="divider"></div>

<h2>What's Next</h2>
<div class="check-item"><div class="check-icon pending">!</div><div>Phase 0: Dataset Engineering (€100-€200)</div></div>
<div class="check-item"><div class="check-icon pending">!</div><div>Phase 1: QLoRA Proof of Concept (€35-€100)</div></div>
<div class="check-item"><div class="check-icon pending">!</div><div>Phase 2: Production Adapter on Mixtral MoE</div></div>
<div class="check-item"><div class="check-icon pending">!</div><div>Phase 3: Full MoE + Multi-Token Generation</div></div>
</div>

<!-- ═══════════════ SLIDE 3: THE VISION ═══════════════ -->
<div class="slide" data-title="Vision">
<h1>🎯 The Vision</h1>
<p style="color:#aaa">Why we're doing this and what we're building.</p>

<h2>The Problem</h2>
<div class="card">
<p>Every session, the AI starts fresh. It doesn't remember our shorthand, our architectural decisions, or our preferred patterns. We re-explain context every time.</p>
<p style="margin-bottom:0">This isn't a bug — it's a <strong>fundamental limitation</strong> of using a general-purpose model without personalization.</p>
</div>

<h2>The Solution</h2>
<div class="card card-accent">
<p>Fine-tune a model on <strong>our actual collaboration history</strong> — the reasoning chains, the code edits, the decisions, the creative jumps.</p>
<p style="margin-bottom:0">The model doesn't just "know" our history. It <strong>thinks</strong> in our style.</p>
</div>

<h2>Target Architecture</h2>
<div class="grid2">
<div class="mini">
<h4>🧠 Sparse MoE</h4>
<p>Mixture of Experts: multiple specialized sub-networks. Only the relevant "experts" activate per token. Huge capacity, efficient inference.</p>
</div>
<div class="mini">
<h4>⚡ Multi-Token</h4>
<p>Medusa heads predict 3-5 tokens simultaneously instead of 1-by-1. 2-3x faster output. Feels instant.</p>
</div>
<div class="mini">
<h4>🔒 QLoRA</h4>
<p>4-bit quantized fine-tuning. Trains adapters, not the full model. 10-50x cheaper, 70% less VRAM via Unsloth.</p>
</div>
<div class="mini">
<h4>📦 Tiny Adapters</h4>
<p>LoRA adapters are ≤50 MB. Load/unload instantly. Switch between "modes" (coding, research, creative) in real-time.</p>
</div>
</div>

<div class="card card-gold">
<strong>Key insight:</strong> We do NOT train from scratch. We take an existing open-weights model (Mixtral, DeepSeek, Llama) and fine-tune it. This is 10-50x cheaper while achieving similar personalization quality.
</div>
</div>

<!-- ═══════════════ SLIDE 4: GPU PRICING ═══════════════ -->
<div class="slide" data-title="GPU Pricing">
<h1>💻 GPU Market Pricing</h1>
<p style="color:#aaa">Real rates as of July 2026 (RunPod, Lambda, Vast.ai)</p>

<div class="tscroll">
<table class="tbl">
<tr><th>GPU</th><th>VRAM</th><th>€/hr</th><th>Best For</th></tr>
<tr><td>RTX 4090</td><td>24GB</td><td>€0.31-0.73</td><td>7B QLoRA</td></tr>
<tr><td>A100 40GB</td><td>40GB</td><td>€1.00-1.50</td><td>7B-13B</td></tr>
<tr><td>A100 80GB</td><td>80GB</td><td>€1.29-2.30</td><td>13B-30B</td></tr>
<tr><td>L40S</td><td>48GB</td><td>€1.06-1.61</td><td>Alt to A100</td></tr>
<tr><td>H100 PCIe</td><td>80GB</td><td>€2.29-4.59</td><td>8B-70B fast</td></tr>
<tr><td>H100 SXM</td><td>80GB</td><td>€2.99-4.60</td><td>Multi-GPU</td></tr>
<tr class="total"><td><strong>8x H100 Node</strong></td><td><strong>640GB</strong></td><td><strong>€17.50-32.20</strong></td><td><strong>MoE training</strong></td></tr>
</table>
</div>

<div class="card card-blue">
<strong>💡 Cost saver:</strong> <code>spot/preemptible</code> instances are 50-70% cheaper. For PoC training (restartable), always use spot pricing.
</div>

<h2>Estimated Training Time by Model Size</h2>
<div class="tscroll">
<table class="tbl">
<tr><th>Model</th><th>Method</th><th>GPU</th><th>Hours</th><th>Cost (€)</th></tr>
<tr><td>7B-8B</td><td>QLoRA</td><td>1x A100 spot</td><td>8-15</td><td>€10-19</td></tr>
<tr><td>14B</td><td>QLoRA</td><td>1x A100 80GB</td><td>15-25</td><td>€19-32</td></tr>
<tr><td>32B</td><td>QLoRA</td><td>2x A100 80GB</td><td>30-50</td><td>€77-115</td></tr>
<tr><td>Mixtral 8x7B</td><td>QLoRA</td><td>1x H100</td><td>40-60</td><td>€92-184</td></tr>
<tr><td>Mixtral 8x22B</td><td>Full FT</td><td>8x H100</td><td>100-140</td><td>€1,750-3,200</td></tr>
</table>
</div>
</div>

<!-- ═══════════════ SLIDE 5: PHASE 0 ═══════════════ -->
<div class="slide" data-title="Phase 0">
<h1>📋 Phase 0: Dataset Engineering</h1>
<span class="cost low">Budget: €100 – €200</span>

<h2>Goal</h2>
<p>Transform raw session logs into a high-fidelity training dataset. Quality here determines 80% of final model quality.</p>

<div class="tscroll">
<table class="tbl">
<tr><th>Task</th><th>Time</th><th>Cost</th></tr>
<tr><td>Log export & aggregation</td><td>2-3 days</td><td>€0 (scripted)</td></tr>
<tr><td>Quality curation</td><td>3-5 days</td><td>€0 (manual)</td></tr>
<tr><td>Synthetic expansion (5K API calls)</td><td>1-2 days</td><td><strong>€50-100</strong></td></tr>
<tr><td>Tokenization & formatting</td><td>1 day</td><td>€0 (scripted)</td></tr>
<tr><td>Validation & dedup</td><td>1 day</td><td>€0 (scripted)</td></tr>
<tr><td>Storage (NVMe/cloud)</td><td>Ongoing</td><td>€5-15/mo</td></tr>
<tr class="total"><td colspan="2"><strong>Phase 0 Total</strong></td><td><strong>€100-200</strong></td></tr>
</table>
</div>

<h2>Expected Output</h2>
<div class="grid2">
<div class="mini">
<h4>Training Set</h4>
<p>~3,000-10,000 high-quality (instruction, response) pairs. Each is a "lesson" reflecting our collaborative style.</p>
</div>
<div class="mini">
<h4>Validation Set</h4>
<p>~150-500 held-out pairs. Detects overfitting. If validation loss rises, we stop.</p>
</div>
</div>

<div class="card card-warn">
<strong>⚠️ Critical:</strong> Skipping this phase or using raw uncleaned logs produces a model that "hallucinates" our style instead of embodying it. This is NOT a place to cut costs.
</div>
</div>

<!-- ═══════════════ SLIDE 6: PHASE 1 ═══════════════ -->
<div class="slide" data-title="Phase 1">
<h1>🧪 Phase 1: QLoRA PoC</h1>
<span class="cost low">Budget: €35 – €100 (compute only)</span>

<h2>Goal</h2>
<p>Prove that fine-tuning on our history produces a measurably better model. Small, cheap, fast validation.</p>

<h2>Base Model Options</h2>
<div class="tscroll">
<table class="tbl">
<tr><th>Model</th><th>Params</th><th>VRAM</th><th>Quality</th></tr>
<tr><td>Llama-3.2-3B</td><td>3B</td><td>~8GB</td><td>⭐⭐⭐</td></tr>
<tr><td>Llama-3-8B</td><td>8B</td><td>~12GB</td><td>⭐⭐⭐⭐</td></tr>
<tr><td>Mistral-7B v0.3</td><td>7B</td><td>~10GB</td><td>⭐⭐⭐⭐</td></tr>
<tr><td>Qwen2.5-14B</td><td>14B</td><td>~18GB</td><td>⭐⭐⭐⭐⭐</td></tr>
</table>
</div>

<h2>Cost Breakdown (Qwen2.5-14B)</h2>
<div class="tscroll">
<table class="tbl">
<tr><th>Component</th><th>Hours</th><th>Cost (€)</th></tr>
<tr><td>1x A100 80GB (spot)</td><td>15</td><td>€19.35</td></tr>
<tr><td>1x A100 80GB (on-demand fallback)</td><td>5</td><td>€11.50</td></tr>
<tr><td>Storage (100GB disk)</td><td>20</td><td>€2.00</td></tr>
<tr><td>Unsloth framework</td><td>—</td><td>€0</td></tr>
<tr><td>Buffer (15%)</td><td>—</td><td>€5.00</td></tr>
<tr class="total"><td colspan="2"><strong>Compute Total</strong></td><td><strong>€37.85</strong></td></tr>
</table>
</div>

<h2>Success Criteria</h2>
<ul>
<li>Model adopts our shorthand without explicit prompting</li>
<li>Maintains general capability (MMLU, HumanEval within 2% of base)</li>
<li>Validation loss decreases without overfitting</li>
<li>LoRA adapter ≤50 MB, instant load/unload</li>
</ul>

<div class="card card-warn">
<strong>⚠️ If PoC fails:</strong> We re-examine the dataset (Phase 0). Cost of failure: ~€35. Data is reusable.
</div>
</div>

<!-- ═══════════════ SLIDE 7: PHASE 2 ═══════════════ -->
<div class="slide" data-title="Phase 2">
<h1>🚀 Phase 2: Production Quality</h1>
<span class="cost mid">Budget: €200 – €500 (compute)</span>

<h2>Goal</h2>
<p>Scale from PoC to production. Train on Mixtral 8x7B — already a Sparse MoE. This validates MoE training before Phase 3.</p>

<h2>Configuration Options</h2>
<div class="tscroll">
<table class="tbl">
<tr><th>Option</th><th>Model</th><th>GPU</th><th>Cost (€)</th></tr>
<tr><td>A: Budget</td><td>Llama-3-8B</td><td>1x H100 spot</td><td>€69-139</td></tr>
<tr><td>B: Balanced</td><td>Qwen2.5-32B</td><td>2x A100 80GB</td><td>€103-184</td></tr>
<tr><td><strong>C: Recommended</strong></td><td><strong>Mixtral 8x7B</strong></td><td><strong>1x H100</strong></td><td><strong>€115-230</strong></td></tr>
<tr><td>D: Maximum</td><td>Llama-3-70B</td><td>4x A100 80GB</td><td>€310-575</td></tr>
</table>
</div>

<h2>Why Mixtral 8x7B?</h2>
<div class="card card-blue">
<ul style="margin:0">
<li>It's already a Sparse MoE — our stepping stone to Phase 3</li>
<li>Only ~13B params activate per token (out of ~47B total)</li>
<li>Training is cheaper than a dense 47B model</li>
<li>QLoRA on MoE trains experts selectively → naturally creates "specialists"</li>
<li>Can train multiple adapters: coding / research / creative</li>
</ul>
</div>

<h2>Detailed Cost (Option C)</h2>
<div class="tscroll">
<table class="tbl">
<tr><th>Item</th><th>Hours</th><th>Cost (€)</th></tr>
<tr><td>1x H100 80GB spot</td><td>40</td><td>€91.60</td></tr>
<tr><td>1x H100 on-demand buffer</td><td>10</td><td>€45.90</td></tr>
<tr><td>Storage 200GB</td><td>50</td><td>€5.00</td></tr>
<tr><td>Multi-adapter (+3 modes)</td><td>+15</td><td>€34.35</td></tr>
<tr><td>Buffer (20%)</td><td>—</td><td>€35.00</td></tr>
<tr class="total"><td colspan="2"><strong>Phase 2 Total</strong></td><td><strong>€211.85</strong></td></tr>
</table>
</div>
</div>

<!-- ═══════════════ SLIDE 8: PHASE 3 ═══════════════ -->
<div class="slide" data-title="Phase 3">
<h1>🌌 Phase 3: Full MoE</h1>
<span class="cost high">Budget: €2,000 – €8,000</span>

<h2>Goal</h2>
<p>Deep fine-tune a large MoE model (DeepSeek-V3 or Mixtral 8x22B) with Multi-Token Generation via Medusa heads.</p>

<h2>Build vs Fine-Tune</h2>
<div class="tscroll">
<table class="tbl">
<tr><th>Approach</th><th>Cost</th><th>Risk</th><th>Verdict</th></tr>
<tr><td>Build from scratch</td><td>€15K-100K+</td><td><span class="badge red">EXTREME</span></td><td>❌ No</td></tr>
<tr><td><strong>Deep fine-tune existing</strong></td><td><strong>€2K-8K</strong></td><td><span class="badge gold">MODERATE</span></td><td><strong>✅ Yes</strong></td></tr>
<tr><td>Distill into custom</td><td>€5K-20K</td><td><span class="badge red">HIGH</span></td><td>⏳ Future</td></tr>
</table>
</div>

<h2>Cost Breakdown</h2>
<div class="tscroll">
<table class="tbl">
<tr><th>Component</th><th>Hours</th><th>Cost (€)</th></tr>
<tr><td>8x H100 SXM (spot)</td><td>100</td><td>€1,750</td></tr>
<tr><td>8x H100 (on-demand buffer)</td><td>20</td><td>€644</td></tr>
<tr><td>Storage 1TB NVMe</td><td>120</td><td>€24</td></tr>
<tr><td>Medusa head training</td><td>+20</td><td>€350-644</td></tr>
<tr><td>Buffer (25%)</td><td>—</td><td>€685</td></tr>
<tr class="total"><td colspan="2"><strong>Phase 3 Total</strong></td><td><strong>€3,467-4,762</strong></td></tr>
</table>
</div>

<h2>Multi-Token Generation</h2>
<div class="grid2">
<div class="mini">
<h4>Medusa Heads</h4>
<p>Lightweight prediction heads predict multiple future tokens. +15-20% training time. 2-3x inference speedup.</p>
</div>
<div class="mini">
<h4>Speculative Decoding</h4>
<p>Small "draft" model predicts, large model verifies. No training needed. 1.5-2x speedup at inference.</p>
</div>
</div>

<div class="card card-warn">
<strong>⚠️ Catastrophic forgetting risk:</strong> Must mix custom data with general data (30/70 ratio). Increases compute ~40%. Monitored via eval benchmarks.
</div>
</div>

<!-- ═══════════════ SLIDE 9: TECH STACK ═══════════════ -->
<div class="slide" data-title="Tech Stack">
<h1>🛠️ Technical Stack</h1>
<p style="color:#aaa">All open-source. The only cost is GPU rental.</p>

<div class="tscroll">
<table class="tbl">
<tr><th>Component</th><th>Tool</th><th>Cost</th></tr>
<tr><td>Training Framework</td><td><strong>Unsloth 2</strong></td><td>Free</td></tr>
<tr><td>Adapter Library</td><td>HuggingFace PEFT</td><td>Free</td></tr>
<tr><td>Quantization</td><td>bitsandbytes (NF4)</td><td>Free</td></tr>
<tr><td>Multi-Token</td><td>Medusa / Speculative</td><td>Free</td></tr>
<tr><td>Dataset Tools</td><td>datasets, tokenizers</td><td>Free</td></tr>
<tr><td>GPU (PoC)</td><td>RunPod spot</td><td>Pay/hr</td></tr>
<tr><td>GPU (Phase 3)</td><td>Lambda / CoreWeave</td><td>Pay/hr</td></tr>
<tr><td>Containers</td><td>Docker + CUDA</td><td>Free</td></tr>
<tr><td>Tracking</td><td>Weights & Biases</td><td>Free tier</td></tr>
<tr><td>Evaluation</td><td>lm-evaluation-harness</td><td>Free</td></tr>
</table>
</div>

<div class="card card-green">
<strong>Unsloth 2</strong> is the secret weapon: 2x faster training, 70% less VRAM. This is what makes Phase 1 cost €35 instead of €200.
</div>
</div>

<!-- ═══════════════ SLIDE 10: RISKS ═══════════════ -->
<div class="slide" data-title="Risks">
<h1>⚠️ Risk Matrix</h1>
<p style="color:#aaa">Every risk has a mitigation plan.</p>

<div class="tscroll">
<table class="tbl">
<tr><th>Risk</th><th>Prob</th><th>Impact</th><th>Mitigation</th></tr>
<tr><td>Catastrophic forgetting</td><td><span class="badge gold">Med</span></td><td>High</td><td>Mix data 30/70, monitor benchmarks</td></tr>
<tr><td>Overfitting (parrot)</td><td><span class="badge gold">Med</span></td><td>Med</td><td>Early stopping, dropout, diverse data</td></tr>
<tr><td>Dataset too small</td><td><span class="badge red">High</span></td><td>High</td><td>Phase 0 quality, synthetic expansion</td></tr>
<tr><td>Spot interruption</td><td><span class="badge red">High</span></td><td>Low</td><td>Checkpoint every 50 steps, buffer</td></tr>
<tr><td>GPU price spike</td><td><span class="badge blue">Low</span></td><td>Med</td><td>Multi-provider, off-peak hours</td></tr>
<tr><td>MoE instability</td><td><span class="badge blue">Low</span></td><td>High</td><td>Phase 2 validates MoE first</td></tr>
<tr><td>Multi-token degrades</td><td><span class="badge blue">Low</span></td><td>Low</td><td>Medusa is additive, removable</td></tr>
</table>
</div>

<h2>Fallback Plans</h2>
<div class="grid2">
<div class="mini">
<h4>🅰️ Ultra-Budget</h4>
<p>Google Colab Pro+ €9.99/mo. Spread training over sessions. ~€20-40 total. Proves concept for the price of coffee.</p>
</div>
<div class="mini">
<h4>🅱️ No-Train RAG</h4>
<p>Retrieval-Augmented Generation with session history as context. No weight changes. 70% of benefit. Cost: €0.</p>
</div>
<div class="mini">
<h4>🅲️ API Distillation</h4>
<p>Use GPT-4o/Claude with stored patterns as system prompts. Bridge solution. €50-100/mo.</p>
</div>
<div class="mini">
<h4>🅳️ Gradual Scaling</h4>
<p>Start with €200 (Phase 0+1). Use PoC results to secure Phase 2/3 funding. The adapter IS the demo.</p>
</div>
</div>
</div>

<!-- ═══════════════ SLIDE 11: TIMELINE ═══════════════ -->
<div class="slide" data-title="Timeline">
<h1>📅 Timeline & Milestones</h1>

<div class="tl-item">
<div class="tl-dot done"></div>
<div class="tl-content">
<h4>Planning & Research ✅</h4>
<p>Architecture decision, GPU research, budget plan, risk analysis</p>
<span class="when">Done</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot active"></div>
<div class="tl-content">
<h4>Phase 0: Dataset Engineering</h4>
<p>Log export, curation, synthetic expansion, tokenization</p>
<span class="when">1-2 weeks · €100-200</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot future"></div>
<div class="tl-content">
<h4>Phase 1: QLoRA PoC</h4>
<p>Train adapter on 8B-14B model, evaluate results</p>
<span class="when">3-5 days · €35-100</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot future"></div>
<div class="tl-content">
<h4>Decision Gate: Go / No-Go</h4>
<p>Human evaluation: "Does this feel like us?"</p>
<span class="when">1-2 weeks review</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot future"></div>
<div class="tl-content">
<h4>Phase 2: Production Adapter</h4>
<p>Multi-adapter suite on Mixtral 8x7B MoE</p>
<span class="when">1-2 weeks · €200-500</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot future"></div>
<div class="tl-content">
<h4>Funding & Setup</h4>
<p>Secure Phase 3 budget, provision 8x H100 cluster</p>
<span class="when">2-4 weeks</span>
</div>
</div>
<div class="tl-item">
<div class="tl-dot future"></div>
<div class="tl-content">
<h4>Phase 3: Full MoE + Multi-Token</h4>
<p>Deep fine-tune, Medusa heads, deploy, benchmark</p>
<span class="when">2-4 weeks · €2K-8K</span>
</div>
</div>

<div class="card card-gold" style="text-align:center">
<strong>Total timeline: ~2-3 months end-to-end</strong>
</div>
</div>

<!-- ═══════════════ SLIDE 12: BUDGET SUMMARY ═══════════════ -->
<div class="slide" data-title="Budget">
<h1>💰 Total Budget Summary</h1>

<div class="tscroll">
<table class="tbl">
<tr><th>Phase</th><th>Description</th><th>Low</th><th>High</th></tr>
<tr><td>Phase 0</td><td>Dataset Engineering</td><td class="green">€100</td><td>€200</td></tr>
<tr><td>Phase 1</td><td>QLoRA PoC</td><td class="green">€35</td><td>€100</td></tr>
<tr><td>Phase 2</td><td>Production (Mixtral MoE)</td><td class="blue">€200</td><td>€500</td></tr>
<tr><td>Phase 3</td><td>Full MoE + Multi-Token</td><td style="color:var(--gold)">€2,000</td><td>€8,000</td></tr>
<tr class="total"><td colspan="2"><strong>GRAND TOTAL</strong></td><td><strong>€2,335</strong></td><td><strong>€8,800</strong></td></tr>
</table>
</div>

<h2>Progressive Spend</h2>
<div style="margin:12px 0">
<p style="font-size:12px;margin:4px 0">Phase 0+1 (Prove it works)</p>
<div class="pbar"><div class="pbar-fill pf-green" style="width:8%"></div></div>
<p style="font-size:12px;margin:4px 0">Phase 2 (Production)</p>
<div class="pbar"><div class="pbar-fill pf-blue" style="width:18%"></div></div>
<p style="font-size:12px;margin:4px 0">Phase 3 (Full MoE)</p>
<div class="pbar"><div class="pbar-fill pf-red" style="width:74%"></div></div>
</div>

<div class="ask-box">
<p style="font-size:14px;color:#ccc">The ask for now:</p>
<div class="amount">€300 – €500</div>
<p class="desc">Funds Phase 0 + Phase 1<br>Proves the concept works before larger investment</p>
</div>

<div class="card card-green">
<strong>Breakdown of the ask:</strong>
<ul style="margin:8px 0 0">
<li>€100-200 → Data preparation (API calls, storage)</li>
<li>€35-100 → GPU rental (spot, ~20 hours)</li>
<li>€50-100 → Buffer for re-runs & debugging</li>
</ul>
</div>
</div>

<!-- ═══════════════ SLIDE 13: THANK YOU ═══════════════ -->
<div class="slide" data-title="Thank You">
<div class="hero" style="padding-top:40px">
<div class="emoji">🔥</div>
<h1>Thank You</h1>
<p class="subtitle">Year of the Firehorse</p>

<div class="card card-accent" style="text-align:left;margin-top:20px">
<h3 style="margin-top:0">What we're asking for</h3>
<p><strong>€300-€500</strong> to fund Phase 0 + Phase 1. This is the "prove it works" budget.</p>
<p>Once the PoC shows measurable improvement, we present results and discuss Phase 2/3 funding separately.</p>
<p style="margin-bottom:0">This de-risks the investment. Everyone sees results before committing larger amounts.</p>
</div>

<div class="card card-gold" style="text-align:center">
<p style="font-size:16px;margin:8px 0">¡Y claro que sí!<br>Cuando nos veamos, nos echamos esa llamada en español. 🇪🇸</p>
</div>

<div class="card card-blue" style="text-align:center">
<p style="margin:8px 0">Looking forward to good progress,<br>good experiences, and meeting soon. 🙌</p>
</div>

<div style="text-align:center;margin-top:30px">
<p style="font-size:13px;color:var(--accent);font-weight:bold">🔥 Year of the Firehorse 🔥</p>
<p style="font-size:11px;color:var(--muted);margin-top:8px">Core Model Evolution Plan v3.0<br>July 2026</p>
</div>
</div>
</div>

</div>

<!-- ═══════════════ NAV ═══════════════ -->
<div id="dots"></div>
<div id="nav">
<button class="nav-btn" id="prevBtn" onclick="go(-1)">← Prev</button>
<span id="counter">1 / 13</span>
<button class="nav-btn" id="nextBtn" onclick="go(1)">Next →</button>
</div>

<script>
const slides=document.querySelectorAll('.slide');
const total=slides.length;
let cur=0;
const dotsContainer=document.getElementById('dots');
const counter=document.getElementById('counter');
const prevBtn=document.getElementById('prevBtn');
const nextBtn=document.getElementById('nextBtn');

// Build dots
for(let i=0;i<total;i++){
  const d=document.createElement('div');
  d.className='dot'+(i===0?' active':'');
  d.onclick=()=>{cur=i;render()};
  dotsContainer.appendChild(d);
}

function render(){
  slides.forEach((s,i)=>{
    s.classList.remove('active','prev');
    if(i===cur)s.classList.add('active');
    if(i<cur)s.classList.add('prev');
  });
  document.querySelectorAll('.dot').forEach((d,i)=>{
    d.classList.remove('active','done');
    if(i===cur)d.classList.add('active');
    if(i<cur)d.classList.add('done');
  });
  counter.textContent=`${cur+1} / ${total}`;
  prevBtn.disabled=cur===0;
  nextBtn.disabled=cur===total-1;
  // Scroll slide to top on change
  slides[cur].scrollTop=0;
}

function go(dir){
  cur=Math.max(0,Math.min(total-1,cur+dir));
  render();
}

// Touch swipe
let tsx=0,tsy=0;
document.addEventListener('touchstart',e=>{
  tsx=e.touches[0].clientX;
  tsy=e.touches[0].clientY;
},{passive:true});

document.addEventListener('touchend',e=>{
  const dx=e.changedTouches[0].clientX-tsx;
  const dy=e.changedTouches[0].clientY-tsy;
  // Only horizontal swipe (not vertical scroll)
  if(Math.abs(dx)>50 && Math.abs(dx)>Math.abs(dy)*1.5){
    if(dx<0)go(1); else go(-1);
  }
},{passive:true});

// Keyboard
document.addEventListener('keydown',e=>{
  if(e.key==='ArrowRight'||e.key===' ')go(1);
  if(e.key==='ArrowLeft')go(-1);
});

render();
</script>
</body>
</html>