Fernando Gutierrez commited on
Commit
5322285
·
1 Parent(s): 89bb222

SAURON gets its own page: the real self-training story; home carries a teaser

Browse files
assets/sauron-eye.png DELETED

Git LFS Details

  • SHA256: 40d846faf84732431612548963f0c5922a9ae25f6ac7e1710f61889194b137e1
  • Pointer size: 131 Bytes
  • Size of remote file: 514 kB
assets/{sauron-announcement-en.png → sauron-fleet-training.png} RENAMED
File without changes
index.html CHANGED
@@ -18,7 +18,7 @@
18
  <a href="#how">How it works</a>
19
  <a href="#fleet">The fleet</a>
20
  <a href="#demo">Demo</a>
21
- <a href="#sauron">Sauron</a>
22
  <a href="#contact">Contact</a>
23
  </nav>
24
  <a class="btn btn-small" href="https://tibios.ai">tibios.ai</a>
@@ -92,18 +92,19 @@
92
  <section id="sauron" class="section">
93
  <h2>Training TIBI SAURON</h2>
94
  <p class="lede center">
95
- Our north star: what if a frontier model could be trained not in a
96
- data center, but on a network of small machines? TIBI SAURON is a
97
- planned 671B-parameter MoE model — in the spirit of DeepSeek-V3 and
98
- Kimi-K3 — trained across thousands of TibiBoxes.
99
- <em>Theoretical scenario based on our model; not a benchmark.</em>
100
  </p>
101
  <div class="cards">
102
- <div class="card"><h3>671B</h3><p>parameters, MoE architecture inspired by DeepSeek-V3 (671B) and Kimi-K3.</p></div>
103
- <div class="card"><h3>14.8T tokens</h3><p>~3.3 × 10²⁴ total FLOPs of distributed training.</p></div>
104
- <div class="card"><h3>50,000 TibiBoxes</h3><p>~12 months to train — 20,000 nodes for training, 30,000 for inference and services.</p></div>
105
  </div>
106
- <img class="wide sauron-img" src="assets/sauron-announcement-en.png" alt="TIBI SAURON training scenario: estimated time to train a 671B model on a TibiBox network" />
 
 
107
  </section>
108
 
109
  <section class="section">
 
18
  <a href="#how">How it works</a>
19
  <a href="#fleet">The fleet</a>
20
  <a href="#demo">Demo</a>
21
+ <a href="sauron.html">Sauron</a>
22
  <a href="#contact">Contact</a>
23
  </nav>
24
  <a class="btn btn-small" href="https://tibios.ai">tibios.ai</a>
 
92
  <section id="sauron" class="section">
93
  <h2>Training TIBI SAURON</h2>
94
  <p class="lede center">
95
+ Our first model trained by its own fleet: five home machines
96
+ generated their own training data, trained a LoRA adapter on
97
+ Qwen3-4B and deployed it back as a new inference lane —
98
+ zero cloud, zero token spend.
 
99
  </p>
100
  <div class="cards">
101
+ <div class="card"><h3>Qwen3-4B · 4-bit</h3><p>Base model. MLX on Apple Silicon, Unsloth + TRL on CUDA.</p></div>
102
+ <div class="card"><h3>16.5M params</h3><p>LoRA adapter, rank 8 — just 0.74% of the model.</p></div>
103
+ <div class="card"><h3>64 MB</h3><p>Adapter size, served by the head as the @tuned lane.</p></div>
104
  </div>
105
+ <p class="lede center">
106
+ <a class="btn btn-ghost" href="sauron.html">Read the full story →</a>
107
+ </p>
108
  </section>
109
 
110
  <section class="section">
sauron.html ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <meta name="viewport" content="width=device-width, initial-scale=1" />
6
+ <title>SAURON — a model trained by its own fleet · TibiOS</title>
7
+ <meta name="description" content="SAURON: a Qwen3-4B LoRA adapter trained by the TibiOS fleet itself — five home machines, zero cloud, zero token spend." />
8
+ <link rel="icon" href="assets/tibios-logo.png" />
9
+ <link rel="stylesheet" href="style.css" />
10
+ </head>
11
+ <body>
12
+ <header class="nav">
13
+ <div class="brand">
14
+ <img src="assets/tibios-logo.png" alt="TibiOS logo" class="logo" />
15
+ <span>TibiOS</span>
16
+ </div>
17
+ <nav>
18
+ <a href="index.html#how">How it works</a>
19
+ <a href="index.html#fleet">The fleet</a>
20
+ <a href="index.html#demo">Demo</a>
21
+ <a href="sauron.html">Sauron</a>
22
+ <a href="index.html#contact">Contact</a>
23
+ </nav>
24
+ <a class="btn btn-small" href="https://tibios.ai">tibios.ai</a>
25
+ </header>
26
+
27
+ <section class="hero">
28
+ <p class="tagline">Presenting</p>
29
+ <h1>SAURON</h1>
30
+ <p class="tagline">A model trained by its own fleet.</p>
31
+ <p class="lede">
32
+ What if a set of home computers could generate their own training
33
+ data, train a model, deploy it and use it for inference — all within
34
+ the same infrastructure? That is what we achieved with TibiOS.
35
+ <strong>Zero token spend.</strong>
36
+ </p>
37
+ </section>
38
+
39
+ <section class="section">
40
+ <h2>The fleet that trained it</h2>
41
+ <p class="lede center">
42
+ SAURON is based on Qwen3-4B, adapted with LoRA on a heterogeneous
43
+ fleet of five machines:
44
+ </p>
45
+ <div class="cards">
46
+ <div class="card"><h3>🍎 Mac mini M4</h3><p>16 GB — training with MLX.</p></div>
47
+ <div class="card"><h3>💻 MacBook Air M1</h3><p>16 GB — training with MLX.</p></div>
48
+ <div class="card"><h3>⚡ 2 × Jetson Orin Nano</h3><p>8 GB each — training with Unsloth + TRL on CUDA.</p></div>
49
+ <div class="card"><h3>🔌 Orange Pi 6+</h3><p>16 GB — head node and router.</p></div>
50
+ </div>
51
+ </section>
52
+
53
+ <section class="section">
54
+ <h2>The full pipeline</h2>
55
+ <div class="steps">
56
+ <div class="step"><span>1</span><p>The inference lanes generate their own training data.</p></div>
57
+ <div class="step"><span>2</span><p>The Macs train in a ring, splitting each batch by every machine's measured speed.</p></div>
58
+ <div class="step"><span>3</span><p>Training produces the SAURON LoRA adapter.</p></div>
59
+ <div class="step"><span>4</span><p>The head node publishes the adapter as a new inference lane: <strong>@tuned</strong>.</p></div>
60
+ <div class="step"><span>5</span><p>The fleet infers again — and can generate new data for the next cycle.</p></div>
61
+ </div>
62
+ </section>
63
+
64
+ <section class="section">
65
+ <h2>Technical data</h2>
66
+ <div class="cards">
67
+ <div class="card"><h3>Qwen3-4B · 4-bit</h3><p>Base model, quantized.</p></div>
68
+ <div class="card"><h3>MLX + Unsloth/TRL</h3><p>MLX on Apple Silicon; Unsloth + TRL on CUDA.</p></div>
69
+ <div class="card"><h3>LoRA r=8, α=16</h3><p>Rank 8, alpha 16, seven projections.</p></div>
70
+ <div class="card"><h3>16.5M parameters</h3><p>Trainable — about 0.74% of the model.</p></div>
71
+ <div class="card"><h3>64 MB</h3><p>Adapter size.</p></div>
72
+ <div class="card"><h3>Home Wi-Fi LAN</h3><p>Zero electricity cost in our scenario, no cloud needed for this local pipeline.</p></div>
73
+ </div>
74
+ <p class="lede center note">
75
+ Kimi K3 and Claude were used exclusively to draft the glossary data —
76
+ not to train SAURON.
77
+ </p>
78
+ </section>
79
+
80
+ <section class="section">
81
+ <img class="wide" src="assets/sauron-fleet-training.png" alt="SAURON infographic: the fleet, the pipeline and the LoRA adapter data" />
82
+ </section>
83
+
84
+ <section class="section">
85
+ <p class="lede center big-quote">
86
+ The important result is not just a tuned model. It is a working cycle
87
+ in which the same fleet both trains and infers — a fundamental piece
88
+ of the TibiOS vision: distributed AI infrastructure where heterogeneous
89
+ machines collaborate, learn and serve models.
90
+ </p>
91
+ <p class="lede center big-quote">
92
+ <strong>You don't need to start with a data center.<br />Start with what you have.</strong>
93
+ </p>
94
+ </section>
95
+
96
+ <footer id="contact" class="footer">
97
+ <img src="assets/tibios-logo.png" alt="" class="logo" />
98
+ <p><strong>TibiOS</strong> — Decentralized AI for everyone</p>
99
+ <p>
100
+ <a href="https://tibios.ai">https://tibios.ai</a> ·
101
+ <a href="mailto:fgparamio@tibios.ai">fgparamio@tibios.ai</a> ·
102
+ +52 999 122 3748
103
+ </p>
104
+ <p class="tiny">Global community — we collaborate, we grow, we innovate.</p>
105
+ </footer>
106
+ </body>
107
+ </html>
style.css CHANGED
@@ -241,6 +241,55 @@ body {
241
  gap: 1.25rem;
242
  }
243
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
244
  .card {
245
  background: var(--card);
246
  border: 1px solid rgba(53, 199, 232, 0.15);
 
241
  gap: 1.25rem;
242
  }
243
 
244
+ /* Numbered steps (Sauron pipeline) */
245
+
246
+ .steps {
247
+ display: grid;
248
+ grid-template-columns: repeat(auto-fit, minmax(280px, 1fr));
249
+ gap: 1.25rem;
250
+ max-width: 900px;
251
+ margin: 0 auto;
252
+ }
253
+
254
+ .step {
255
+ display: flex;
256
+ gap: 1rem;
257
+ align-items: flex-start;
258
+ background: var(--card);
259
+ border: 1px solid rgba(53, 199, 232, 0.15);
260
+ border-radius: var(--radius);
261
+ padding: 1.2rem 1.4rem;
262
+ }
263
+
264
+ .step span {
265
+ flex: none;
266
+ width: 40px;
267
+ height: 40px;
268
+ border-radius: 50%;
269
+ background: var(--cyan);
270
+ color: #06222e;
271
+ font-weight: 700;
272
+ display: flex;
273
+ align-items: center;
274
+ justify-content: center;
275
+ font-size: 1.15rem;
276
+ }
277
+
278
+ .step p {
279
+ margin: 0;
280
+ color: var(--text);
281
+ }
282
+
283
+ .note {
284
+ font-size: 0.9rem;
285
+ margin-top: 1.5rem;
286
+ }
287
+
288
+ .big-quote {
289
+ font-size: 1.2rem;
290
+ color: var(--text);
291
+ }
292
+
293
  .card {
294
  background: var(--card);
295
  border: 1px solid rgba(53, 199, 232, 0.15);