Skip to content

Commit 9f38833

Browse files
mstrathmanclaude
andcommitted
site: 'Why not the usual ways?' comparison on the landing page
Four spec-sheet cards in the tabular-FM vendors' own format, with the rows their versions omit: an ML pipeline (months to first prediction), AutoML (automates the fit, keeps the pipeline), a prediction API (seconds to start, but training hidden in every metered call and data leaving on each one), and SELECT forecast(...) (seconds to start, in-process microsecond serving, data never leaves, flat cost). Includes the honest asterisk that zero-shot frontier accuracy is the ceiling and links the measured gap. Styling is variable-driven so both themes inherit correct contrast. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
1 parent cae5de2 commit 9f38833

2 files changed

Lines changed: 127 additions & 0 deletions

File tree

‎website/src/content/docs/index.mdx‎

Lines changed: 63 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -56,6 +56,69 @@ SELECT forecast(ts, value, 24, '{"model":"traffic-v1"}') FROM readings;
5656
</Card>
5757
</CardGrid>
5858

59+
60+
## Why not the usual ways?
61+
62+
Every alternative gets tables right. The differences are where the work
63+
lives, where your data goes, and what each prediction costs at volume.
64+
65+
<div class="cmp-grid">
66+
<div class="cmp-card">
67+
<p class="cmp-kicker">Why not an ML pipeline?</p>
68+
<p class="cmp-sub">Great algorithms. The cost is everything around them:
69+
the export, the feature job, the container, the deploy.</p>
70+
<ul class="cmp-specs">
71+
<li><span>FIRST PREDICTION</span><b class="bad">MONTHS</b></li>
72+
<li><span>TRAINING</span><b class="bad">PER DATASET, YOUR CLUSTER</b></li>
73+
<li><span>SERVING</span><b class="mid">YOUR INFRA, MS</b></li>
74+
<li><span>DATA LEAVES THE DB</span><b class="mid">TO THE PIPELINE</b></li>
75+
<li><span>COST AT VOLUME</span><b class="mid">YOUR CLUSTER</b></li>
76+
</ul>
77+
</div>
78+
<div class="cmp-card">
79+
<p class="cmp-kicker">Why not AutoML?</p>
80+
<p class="cmp-sub">Automates the fit. Keeps the pipeline, the
81+
infrastructure, and the wait.</p>
82+
<ul class="cmp-specs">
83+
<li><span>FIRST PREDICTION</span><b class="bad">WEEKS</b></li>
84+
<li><span>TRAINING</span><b class="bad">PER DATASET, MANAGED</b></li>
85+
<li><span>SERVING</span><b class="mid">DEPLOYED ENDPOINT</b></li>
86+
<li><span>DATA LEAVES THE DB</span><b class="bad">USUALLY, TO A CLOUD</b></li>
87+
<li><span>COST AT VOLUME</span><b class="mid">YOUR CLUSTER</b></li>
88+
</ul>
89+
</div>
90+
<div class="cmp-card">
91+
<p class="cmp-kicker">Why not a prediction API?</p>
92+
<p class="cmp-sub">Seconds to the first answer, and the frontier models
93+
live here. But "no training" means your rows ride along as context on
94+
every call, on a meter, in someone else's cloud.</p>
95+
<ul class="cmp-specs">
96+
<li><span>FIRST PREDICTION</span><b class="good">SECONDS</b></li>
97+
<li><span>TRAINING</span><b class="bad">HIDDEN IN EVERY CALL</b></li>
98+
<li><span>SERVING</span><b class="bad">NETWORK ROUND-TRIP</b></li>
99+
<li><span>DATA LEAVES THE DB</span><b class="bad">EVERY CALL</b></li>
100+
<li><span>COST AT VOLUME</span><b class="bad">LINEAR, THEIR INVOICE</b></li>
101+
</ul>
102+
</div>
103+
<div class="cmp-card cmp-ours">
104+
<p class="cmp-kicker"><code>SELECT forecast(...)</code></p>
105+
<p class="cmp-sub">A SQL function at your data. Zero-shot in seconds;
106+
distill a teacher once and serve its student in microseconds, forever.</p>
107+
<ul class="cmp-specs">
108+
<li><span>FIRST PREDICTION</span><b class="good">SECONDS</b></li>
109+
<li><span>TRAINING</span><b class="good">NONE, OR ONCE IN SECONDS</b></li>
110+
<li><span>SERVING</span><b class="good">IN-PROCESS, MICROSECONDS</b></li>
111+
<li><span>DATA LEAVES THE DB</span><b class="good">NEVER</b></li>
112+
<li><span>COST AT VOLUME</span><b class="good">FLAT, NEAR ZERO</b></li>
113+
</ul>
114+
</div>
115+
</div>
116+
117+
The honest asterisk: a frontier model called zero-shot is still the accuracy
118+
ceiling. The [benchmarks](./benchmarks/) measure the gap; a distilled student
119+
typically gives up about half a point of accuracy against its teacher and
120+
keeps everything in the right column.
121+
59122
:::caution[Pre-alpha]
60123
The API and the SQL surface are unstable and may change without notice. Not yet
61124
recommended for production.

‎website/src/styles/theme.css‎

Lines changed: 64 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -63,3 +63,67 @@ div.hero {
6363
font-size: 0.8em;
6464
padding-block: 0.05em;
6565
}
66+
67+
/* "Why not the usual ways?" comparison cards on the landing page: a
68+
* spec-sheet look (terse caps labels, mono values, verdict colors).
69+
* Grid collapses to one column on narrow screens. */
70+
.cmp-grid {
71+
display: grid;
72+
grid-template-columns: repeat(auto-fit, minmax(19rem, 1fr));
73+
gap: 1rem;
74+
margin-top: 1.5rem;
75+
}
76+
.cmp-card {
77+
border: 1px solid var(--sl-color-gray-5);
78+
border-radius: 0.5rem;
79+
padding: 1.25rem 1.25rem 0.75rem;
80+
background: var(--sl-color-black);
81+
}
82+
.cmp-card.cmp-ours {
83+
border-color: var(--sl-color-accent);
84+
}
85+
.cmp-kicker {
86+
font-size: var(--sl-text-lg);
87+
font-weight: 600;
88+
color: var(--sl-color-white);
89+
margin: 0;
90+
}
91+
.cmp-kicker code {
92+
font-size: 0.85em;
93+
}
94+
.cmp-sub {
95+
font-size: var(--sl-text-sm);
96+
color: var(--sl-color-gray-3);
97+
margin-top: 0.5rem;
98+
}
99+
.cmp-specs {
100+
list-style: none;
101+
padding: 0;
102+
margin: 1rem 0 0.25rem;
103+
border-top: 1px solid var(--sl-color-gray-6);
104+
}
105+
.cmp-specs li {
106+
display: flex;
107+
justify-content: space-between;
108+
gap: 0.75rem;
109+
align-items: baseline;
110+
padding: 0.45rem 0;
111+
border-bottom: 1px solid var(--sl-color-gray-6);
112+
font-family: var(--__sl-font-mono, monospace);
113+
font-size: 0.72rem;
114+
letter-spacing: 0.04em;
115+
}
116+
.cmp-specs li:last-child {
117+
border-bottom: none;
118+
}
119+
.cmp-specs span {
120+
color: var(--sl-color-gray-3);
121+
white-space: nowrap;
122+
}
123+
.cmp-specs b {
124+
font-weight: 600;
125+
text-align: right;
126+
}
127+
.cmp-specs b.good { color: var(--sl-color-green, #4ade80); }
128+
.cmp-specs b.mid { color: var(--sl-color-gray-2); }
129+
.cmp-specs b.bad { color: var(--sl-color-orange, #fb923c); }

0 commit comments

Comments
 (0)