Blame for index.html

<!doctype html>
<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>IRIS · A closer look at AI training | Saccadic</title><style>
:root{--paper:#F0E4BE;--panel:#EBDFB2;--ink:#1A1A1A;--body:#3A2F1F;--muted:#6B5F4F;--line:#8B7E5E;--purple:#5C3A8C;--blue:#1F8FDC}*{box-sizing:border-box}html{scroll-behavior:smooth}body{margin:0;background:var(--paper);color:var(--body);font:18px/1.6 'Latin Modern Roman',Charter,Cambria,Georgia,serif}main{max-width:1160px;margin:auto;padding:32px 44px 54px}a{color:var(--blue);text-underline-offset:4px}a:hover{color:var(--purple)}a:focus-visible,summary:focus-visible{outline:2px solid var(--ink);outline-offset:5px}header{display:flex;align-items:center;justify-content:space-between;gap:30px;border-bottom:1px solid var(--line);padding-bottom:22px}header img{width:104px;height:104px;object-fit:contain}nav{display:flex;gap:24px;flex-wrap:wrap;font-size:16px}nav a{color:var(--body)}.hero{padding:52px 0 38px;display:grid;grid-template-columns:1.5fr 1fr;gap:60px;align-items:end}.eyebrow{margin:0 0 12px;font-size:17px;font-style:italic;color:var(--muted)}h1,h2,h3{color:var(--ink);line-height:1.15}h1{font-size:clamp(40px,5.8vw,70px);font-weight:400;letter-spacing:-.025em;margin:0 0 24px}h1 span{color:var(--purple)}.lede{max-width:680px;font-size:23px;line-height:1.45;margin:0 0 28px}.aside{border-top:1px solid var(--line);padding-top:20px;font-size:19px}.aside strong{display:block;color:var(--ink);font-size:22px;margin-bottom:12px}.cta{display:inline-block;border:1px solid var(--ink);padding:10px 20px;color:var(--ink);text-decoration:none;background:var(--panel);box-shadow:1.5px 1.5px 0 rgba(58,47,31,.55)}.cta:active{transform:translate(1.5px,1.5px);box-shadow:none}.minor{margin-left:20px;font-size:17px}.figure{border:1px solid var(--line);margin:0 0 48px;padding:24px;background:var(--panel)}.figure h2{font-size:27px;margin:0 0 10px}.figure p{margin:0 0 20px;max-width:880px}.figure img{display:block;width:100%;height:auto;border:0}.figure figcaption{font-size:16px;font-style:italic;margin-top:15px}.figure figcaption b{font:700 14px 'Latin Modern Mono','SF Mono',Menlo,Consolas,monospace;color:var(--ink)}section{margin:48px 0}section>h2{font-size:30px;font-weight:700;border-bottom:.6px solid var(--ink);padding-bottom:13px;margin:0 0 12px}.subtitle{font-style:italic;margin:0 0 26px}.grid{display:grid;grid-template-columns:repeat(3,1fr);gap:32px}.grid article{border-top:1px solid var(--line);padding-top:18px}.grid h3{font-size:23px;margin:0 0 12px}.grid p{margin:0}.features{display:grid;grid-template-columns:1fr 1fr;gap:24px 40px;margin:28px 0}.features article{border-bottom:.6px solid var(--line);padding-bottom:22px}.features h3{font-size:22px;margin:0 0 10px}.features p{margin:0}.callout{background:var(--panel);border:.6px solid var(--line);border-left:2px solid var(--purple);padding:22px 28px;margin:28px 0}.callout p{margin:0}.environment{display:grid;grid-template-columns:1.2fr 1fr;gap:42px}details{border-bottom:.6px solid var(--line);padding:16px 0}summary{cursor:pointer;color:var(--ink);font-weight:bold}details p{margin:12px 0 4px}.contact{border:1px solid var(--line);background:var(--panel);padding:30px}.contact h2{border:0;padding:0;font-size:33px}.contact p{max-width:800px}.contact .cta{background:var(--paper)}footer{border-top:1px solid var(--line);padding-top:20px;font-size:15px;color:var(--muted);display:flex;justify-content:space-between;gap:24px}footer p{margin:0}.note{font-size:15px;font-style:italic;color:var(--muted)}@media(max-width:720px){body{font-size:17px}main{padding:24px 20px}.hero{grid-template-columns:1fr;gap:24px;padding:36px 0}.lede{font-size:21px}.grid,.features,.environment{grid-template-columns:1fr;gap:24px}header{align-items:flex-start}header img{width:80px;height:80px}nav{gap:10px 18px;max-width:190px}.figure{padding:14px}.minor{display:inline-block;margin:16px 0 0}footer{display:block}footer p+p{margin-top:10px}}
</style></head><body><main>
<header><a href="https://www.saccadic.ai/" target="_blank" rel="noopener"><img src="https://cdn-avatars.huggingface.co/v1/production/uploads/69d6d16790de87a3b534bf99/SkapdHGpRMem-dCvIvgMx.png" alt="Saccadic: purple lotus with a blue eye"></a><nav aria-label="Page navigation"><a href="#evidence">Measured run</a><a href="#inside">Inside IRIS</a><a href="#workflow">Workflow</a><a href="#evaluation">Evaluation</a></nav></header>
<div class="hero"><div><p class="eyebrow">IRIS™ · Intrinsic Rank Inspection System</p><h1>A closer look<br>at <span>AI training.</span></h1><p class="lede">Observe how a neural network’s internal structure changes while the run is still underway.</p><a class="cta" href="https://www.saccadic.ai/iris" target="_blank" rel="noopener">Explore IRIS ↗</a><a class="minor" href="#evidence">See a measured run ↓</a></div><aside class="aside"><strong>The scoreboard reports the outcome.</strong>A loss curve can show a change without locating it inside the model. IRIS adds a measured view of representations, expert routing, and coordination across training domains.</aside></div>
<figure class="figure" id="evidence"><h2>Same model. Different learning times.</h2><p>In a modular-arithmetic experiment, the operations reached high held-out accuracy at different stages. The overall accuracy curve hides some of that progression.</p><a href="https://cdn-uploads.huggingface.co/production/uploads/69d6d16790de87a3b534bf99/LsCSdafjN_zw5hphoWKHa.png" target="_blank" rel="noopener"><img src="https://cdn-uploads.huggingface.co/production/uploads/69d6d16790de87a3b534bf99/LsCSdafjN_zw5hphoWKHa.png" alt="Accuracy curves for addition, subtraction, multiplication, division, and exponentiation modulo 113; training accuracy rises early and held-out accuracy rises later, at different times for each operation."></a><figcaption><b>FIG 4.4</b> · Training and held-out accuracy in the measured run. <a href="https://huggingface.co/blog/Saccadic-AI/accuracy-plateaued-the-model-kept-changing" target="_blank" rel="noopener">Read the experiment and its conditions ↗</a></figcaption></figure>
<section id="inside"><h2>Look inside the run</h2><p class="subtitle">Measurements that help turn an observation into a testable question.</p><figure class="figure" style="margin:28px 0 32px"><h3 style="font-size:27px;margin:0 0 12px">Follow the signal through the model</h3><p>The accuracy curves show what the model achieves. The Flow Map adds a view of the measured signal carried between experts and layers during training.</p><a href="https://cdn-uploads.huggingface.co/production/uploads/69d6d16790de87a3b534bf99/U2tq3qHgjyswjkRs2gl6R.png" target="_blank" rel="noopener"><img src="https://cdn-uploads.huggingface.co/production/uploads/69d6d16790de87a3b534bf99/U2tq3qHgjyswjkRs2gl6R.png" alt="IRIS signal flow map at training step 49,300. Mechanics in green and arithmetic in orange follow overlapping paths through experts in layers L0, L1, L2, and L3, from input to output."></a><figcaption><b>FIG 10.1</b> · A snapshot at training step 49,300. Green links show mechanics; orange links show arithmetic. Line thickness represents measured signal; expert circles are sized by traffic share. Select the image to enlarge it.</figcaption><p style="margin:20px 0 0">Compare where domain paths overlap and how traffic is distributed across layers. Revisit the recorded run in IRIS to inspect how those patterns change over time.</p><p class="note" style="margin:12px 0 0">Static capture of the IRIS Bench. The controls shown inside the image are part of the captured interface.</p></figure><div class="features"><article><h3>Follow the signal</h3><p>The Flow Map shows measured connections between layers and experts for different input domains.</p></article><article><h3>Inspect representation structure</h3><p>Track how activation variation concentrates and how its dominant directions move during training.</p></article><article><h3>Compare expert routing</h3><p>Inspect which subjects reach each expert, separately for each layer, and how routing changes over time.</p></article><article><h3>Observe collaboration and conflict</h3><p>Compare sharing and training-gradient agreement across domains, alongside loss and task results.</p></article></div><div class="callout"><p><strong>Read the measurement in context.</strong> IRIS records observations. Use them with validation, gradients, and task results to decide what to test next.</p></div></section>
<section id="workflow"><h2>A training movie, built from measurements</h2><p class="subtitle">A consistent view during training, and a record to revisit afterward.</p><div class="grid"><article><h3>Capture</h3><p>The tracker measures the model using consistent probe inputs. Capture intervals can be fixed or adapt to organizational activity.</p></article><article><h3>Record</h3><p>Measurements accumulate in a self-contained <code>.iris</code> file. Training checkpoints keep their usual role.</p></article><article><h3>Inspect</h3><p>The local Bench reads the growing record during training and opens it afterward, without loading the model.</p></article></div></section>
<section><h2>Work in your own environment</h2><div class="environment"><div><p>IRIS fits existing PyTorch training loops. Capture and inspection can run on CPU, Apple MPS, or NVIDIA CUDA.</p><p>Support includes GPT / LLaMA / Mistral, BERT-family encoders, and supported mixture-of-experts models. User-defined architectures use a measurement adapter to declare capture points.</p><p class="note">Coverage depends on model layout. Shared-weight configurations are currently unsupported.</p></div><div><details open><summary>Model and data handling</summary><p>Model weights, activations, and probe text are not sent to Saccadic. Licensing and metering contact Saccadic.</p></details><details><summary>Keep your existing experiment tracker</summary><p>Optional configured relays send selected scalar measurements to Weights &amp; Biases, TensorBoard, MLflow, and Aim.</p></details><details><summary>Read, explain, and export</summary><p>Each graph includes an Explain guide. Define labeled scalar signals, derive formulas from captured channels, and export custom charts as PNG images, Excel workbooks, or JSON data.</p></details></div></div></section>
<section class="contact" id="evaluation"><p class="eyebrow">Research and joint evaluations</p><h2>One model. One run.<br>One concrete investigation.</h2><p>For teams training or fine-tuning models that need a clearer internal view. A joint evaluation can test whether the readings focus a real investigation and measure overhead in your workload.</p><p>Academic researchers are encouraged to request credit grants.</p><p>Commercial IRIS pricing depends on model size, with costs of just pennies per run for models below 2 million parameters. See the product site for pricing and licensing terms.</p><a class="cta" href="mailto:info@saccadic.ai">Discuss an evaluation ↗</a><a class="minor" href="https://www.saccadic.ai/iris" target="_blank" rel="noopener">Product details ↗</a></section>
<footer><p>© Saccadic LLC · <a href="https://www.saccadic.ai/" target="_blank" rel="noopener">saccadic.ai</a></p><p>IRIS beta available for download at <a href="https://www.saccadic.ai/iris" target="_blank" rel="noopener">www.saccadic.ai/iris</a>.</p></footer>
</main></body></html>