-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathrl-environments.html
More file actions
769 lines (647 loc) · 34.5 KB
/
Copy pathrl-environments.html
File metadata and controls
769 lines (647 loc) · 34.5 KB
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
<!DOCTYPE html>
<html lang="en">
<head>
<!-- Google tag (gtag.js) -->
<script async src="https://www.googletagmanager.com/gtag/js?id=G-62YF7Y81BS"></script>
<script>
window.dataLayer = window.dataLayer || [];
function gtag(){dataLayer.push(arguments);}
gtag('js', new Date());
gtag('config', 'G-62YF7Y81BS');
</script>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>RL Environments: The New Data Moat of AI — Omkar Ray</title>
<link href="https://fonts.googleapis.com/css2?family=Newsreader:ital,opsz,wght@0,6..72,400;0,6..72,600;0,6..72,700;1,6..72,400;1,6..72,600&family=JetBrains+Mono:wght@400;500&family=DM+Sans:wght@400;500;600;700&display=swap" rel="stylesheet">
<style>
:root {
--ink: #1a1a1a;
--paper: #faf8f4;
--accent: #c4421a;
--accent-light: #f0d8cf;
--mid: #6b6560;
--rule: #d4cfc8;
--sidebar-bg: #f0ede7;
--code-bg: #2d2926;
--highlight: #fff3cd;
}
* { margin: 0; padding: 0; box-sizing: border-box; }
body {
font-family: 'Newsreader', Georgia, serif;
background: var(--paper);
color: var(--ink);
font-size: 19px;
line-height: 1.72;
-webkit-font-smoothing: antialiased;
}
.back-nav {
max-width: 900px;
margin: 0 auto;
padding: 28px 40px 0;
}
.back-nav a {
font-family: 'DM Sans', sans-serif;
font-size: 13px;
color: var(--mid);
text-decoration: none;
letter-spacing: 0.3px;
}
.back-nav a:hover { color: var(--ink); }
/* ─── HERO ─── */
.hero {
padding: 64px 40px 80px;
max-width: 900px;
margin: 0 auto;
position: relative;
}
.hero::before {
content: '';
position: absolute;
top: 0;
left: 40px;
width: 60px;
height: 4px;
background: var(--accent);
}
.hero .kicker {
font-family: 'DM Sans', sans-serif;
font-size: 13px;
font-weight: 600;
text-transform: uppercase;
letter-spacing: 2.5px;
color: var(--accent);
margin-bottom: 24px;
}
.hero h1 {
font-size: clamp(36px, 5.5vw, 58px);
font-weight: 700;
line-height: 1.12;
letter-spacing: -0.02em;
margin-bottom: 24px;
max-width: 760px;
}
.hero .subtitle {
font-size: 22px;
font-weight: 400;
font-style: italic;
color: var(--mid);
max-width: 620px;
line-height: 1.55;
}
.hero .meta {
margin-top: 40px;
font-family: 'DM Sans', sans-serif;
font-size: 13px;
color: var(--mid);
letter-spacing: 0.3px;
display: flex;
gap: 24px;
align-items: center;
}
.hero .meta span { display: flex; align-items: center; gap: 6px; }
/* ─── CONTENT ─── */
.content {
max-width: 720px;
margin: 0 auto;
padding: 0 40px 120px;
}
.content p { margin-bottom: 22px; }
.content h2 {
font-size: 30px;
font-weight: 700;
margin: 64px 0 20px;
letter-spacing: -0.015em;
line-height: 1.25;
position: relative;
padding-left: 20px;
}
.content h2::before {
content: '';
position: absolute;
left: 0;
top: 4px;
bottom: 4px;
width: 4px;
background: var(--accent);
border-radius: 2px;
}
.content h3 {
font-family: 'DM Sans', sans-serif;
font-size: 18px;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 1.2px;
margin: 48px 0 16px;
color: var(--accent);
}
/* ─── PULLQUOTE ─── */
.pullquote {
margin: 48px -40px;
padding: 36px 40px;
border-top: 1px solid var(--rule);
border-bottom: 1px solid var(--rule);
font-size: 24px;
font-style: italic;
line-height: 1.55;
color: var(--ink);
position: relative;
}
.pullquote::before {
content: '"';
font-size: 72px;
font-weight: 700;
color: var(--accent);
position: absolute;
top: 20px;
left: 40px;
line-height: 1;
opacity: 0.3;
}
.pullquote .attr {
display: block;
font-family: 'DM Sans', sans-serif;
font-size: 13px;
font-style: normal;
font-weight: 600;
color: var(--mid);
margin-top: 12px;
letter-spacing: 0.3px;
}
/* ─── CALLOUT ─── */
.callout {
background: var(--sidebar-bg);
border-left: 3px solid var(--accent);
padding: 28px 32px;
margin: 36px 0;
border-radius: 0 6px 6px 0;
}
.callout .callout-title {
font-family: 'DM Sans', sans-serif;
font-size: 13px;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 1.8px;
color: var(--accent);
margin-bottom: 12px;
}
.callout p {
font-size: 17px;
line-height: 1.65;
margin-bottom: 10px;
}
.callout p:last-child { margin-bottom: 0; }
/* ─── FRAMEWORK DIAGRAM ─── */
.framework {
margin: 48px 0;
background: var(--code-bg);
border-radius: 8px;
padding: 40px;
color: #e8e2da;
}
.framework .fw-title {
font-family: 'DM Sans', sans-serif;
font-size: 13px;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 2px;
color: var(--accent);
margin-bottom: 28px;
opacity: 0.9;
}
.fw-grid { display: grid; grid-template-columns: 1fr; gap: 20px; }
.fw-layer {
display: grid;
grid-template-columns: 140px 1fr;
gap: 16px;
align-items: start;
padding: 16px 0;
border-bottom: 1px solid rgba(255,255,255,0.08);
}
.fw-layer:last-child { border-bottom: none; }
.fw-label {
font-family: 'JetBrains Mono', monospace;
font-size: 12px;
font-weight: 500;
color: var(--accent);
text-transform: uppercase;
letter-spacing: 1px;
padding-top: 2px;
}
.fw-desc {
font-family: 'DM Sans', sans-serif;
font-size: 15px;
line-height: 1.6;
color: #c4bfb7;
}
.fw-desc strong { color: #e8e2da; font-weight: 600; }
.fw-tags { display: flex; flex-wrap: wrap; gap: 6px; margin-top: 8px; }
.fw-tag {
font-family: 'JetBrains Mono', monospace;
font-size: 11px;
background: rgba(196, 66, 26, 0.2);
color: #e8a88a;
padding: 3px 10px;
border-radius: 3px;
border: 1px solid rgba(196, 66, 26, 0.25);
}
/* ─── TABLE ─── */
.data-table {
width: 100%;
margin: 36px 0;
border-collapse: collapse;
font-family: 'DM Sans', sans-serif;
font-size: 15px;
}
.data-table thead th {
text-align: left;
padding: 12px 16px;
font-size: 11px;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 1.5px;
color: var(--mid);
border-bottom: 2px solid var(--ink);
}
.data-table tbody td {
padding: 14px 16px;
border-bottom: 1px solid var(--rule);
vertical-align: top;
line-height: 1.5;
}
.data-table tbody tr:hover { background: var(--sidebar-bg); }
/* ─── NUMBERED ARGUMENTS ─── */
.argument {
display: grid;
grid-template-columns: 48px 1fr;
gap: 16px;
margin: 32px 0;
align-items: start;
}
.argument-num {
font-family: 'Newsreader', serif;
font-size: 36px;
font-weight: 700;
color: var(--accent);
line-height: 1;
opacity: 0.6;
}
.argument-body h4 {
font-size: 20px;
font-weight: 600;
margin-bottom: 8px;
line-height: 1.3;
}
.argument-body p {
font-size: 17px;
color: var(--mid);
margin-bottom: 0;
}
/* ─── SECTION DIVIDER ─── */
.divider {
text-align: center;
margin: 56px 0;
color: var(--rule);
font-size: 20px;
letter-spacing: 12px;
}
/* ─── EMPHASIS ─── */
strong { font-weight: 600; }
em { font-style: italic; }
.highlight {
background: linear-gradient(180deg, transparent 60%, var(--accent-light) 60%);
padding: 0 2px;
}
/* ─── FOOTER ─── */
.essay-footer {
max-width: 720px;
margin: 0 auto;
padding: 40px 40px 80px;
border-top: 1px solid var(--rule);
font-family: 'DM Sans', sans-serif;
font-size: 13px;
color: var(--mid);
display: flex;
justify-content: space-between;
}
/* ─── RESPONSIVE ─── */
@media (max-width: 680px) {
.back-nav { padding: 20px 24px 0; }
.hero { padding: 48px 24px 56px; }
.content { padding: 0 24px 80px; }
.pullquote { margin: 36px -24px; padding: 28px 24px; }
.framework { padding: 28px 24px; }
.fw-layer { grid-template-columns: 1fr; }
.essay-footer { flex-direction: column; gap: 8px; padding: 32px 24px 60px; }
}
/* ─── Scroll Reveal ─── */
.sr { opacity: 0; transition: opacity 0.7s cubic-bezier(0.16, 1, 0.3, 1), transform 0.7s cubic-bezier(0.16, 1, 0.3, 1); }
.sr.sr-up { transform: translateY(32px); }
.sr.sr-left { transform: translateX(-28px); }
.sr.sr-right { transform: translateX(28px); }
.sr.sr-scale { transform: scale(0.96); }
.sr.sr-visible { opacity: 1; transform: none; }
.content h2.sr { transition-duration: 0.6s; }
.pullquote.sr { transition-duration: 0.8s; }
.framework.sr { transition-duration: 0.9s; }
.argument.sr { transition-delay: calc(var(--sr-i, 0) * 0.08s); }
.fw-layer.sr { transition-delay: calc(var(--sr-i, 0) * 0.1s); }
.content p.sr-visible,
.argument.sr-visible,
.callout.sr-visible { transition-timing-function: cubic-bezier(0.16, 1, 0.3, 1); }
@media (prefers-reduced-motion: reduce) {
.sr { transition: opacity 0.3s ease !important; transform: none !important; }
}
.share-bar {
position: fixed; bottom: 0; left: 0; right: 0;
background: rgba(255,255,255,0.95); backdrop-filter: blur(8px); -webkit-backdrop-filter: blur(8px);
border-top: 1px solid #E2E0DC; display: flex; justify-content: center; align-items: center;
gap: 28px; padding: 14px 0; z-index: 100;
}
.share-bar a, .share-bar button {
color: #999; text-decoration: none; background: none; border: none; cursor: pointer;
padding: 4px; transition: color 0.2s; display: flex; align-items: center; gap: 6px;
font-family: 'IBM Plex Mono', monospace; font-size: 13px;
}
.share-bar a:hover, .share-bar button:hover { color: #1a1a2e; }
.share-bar svg { width: 22px; height: 22px; fill: currentColor; }
.share-bar .view-count {
color: #999; display: flex; align-items: center; gap: 6px;
font-family: 'IBM Plex Mono', monospace; font-size: 13px;
}
</style>
</head>
<body>
<div class="back-nav">
<a href="index.html">← Back</a>
</div>
<div class="hero">
<div class="kicker">Investment Thesis · AI Infrastructure</div>
<h1>RL Environments Are the New Data Moat</h1>
<p class="subtitle">Why the companies building the "gyms" where AI agents train may capture more durable value than those building the agents themselves.</p>
<div class="meta">
<span>February 2026</span>
<span>·</span>
<span>15 min read</span>
</div>
</div>
<div class="content">
<p>For a decade, the AI industry ran on a simple formula: <strong>more data + more compute = better models.</strong> The data was static — scraped web pages, annotated images, expert-curated text. The moat belonged to whoever had the biggest dataset or the deepest labeling pipeline.</p>
<p>That era is ending. The frontier has shifted from <em>learning from human artifacts</em> to <em>learning from action and consequence</em>. In this new paradigm, what matters is not the volume of tokens you can feed a model, but the <span class="highlight">quality of the environments you can build for it to practice in.</span></p>
<p>This essay argues that <strong>RL environments</strong> — the sandboxes, simulators, and instrumented workflows where agents act, fail, and improve — are emerging as the most consequential and investable layer of the AI stack. They are to the "Era of Experience" what ImageNet was to the deep learning revolution: the enabling substrate without which progress stalls.</p>
<div class="divider">· · ·</div>
<h2>I. The Shift: From Tokens to Trajectories</h2>
<p>The "pre-training on internet text" paradigm hit diminishing returns around 2024. Not because language models stopped improving, but because the <em>marginal value</em> of another trillion tokens became vanishingly small compared to a well-designed RL training run. OpenAI's progression from o1 to o3 told the story: each generation allocated more compute to reinforcement learning and less to static pre-training.</p>
<p>This shift has a clean theoretical grounding. Silver and Sutton's "Era of Experience" paper from DeepMind frames it precisely: the next generation of AI agents will learn from <em>streams of interaction</em>, take <em>grounded actions</em> in real action spaces, receive <em>grounded rewards</em> from consequences, and plan in the <em>currency of experience.</em></p>
<div class="pullquote">
We added some reinforcement learning compute for o1… o3 maybe had a little bit more RL compute. At some point in the future maybe we'll have a lot of RL compute and then at some far point in the future maybe we'll be totally dominated and crushed by RL compute.
<span class="attr">— Dan Roberts, OpenAI Researcher, at Sequoia Ascent</span>
</div>
<p>What does "RL compute" actually consume? Not text. It consumes <strong>environment interactions</strong> — hundreds of thousands of rollouts where an agent attempts a task, receives a reward signal, and updates its policy. The raw material of RL scaling is not a dataset. It is a <em>world</em> the agent can act in.</p>
<p>This creates a fundamentally different supply chain. In the token era, data companies sold annotated text. In the experience era, they sell <strong>instrumented environments</strong> — interactive sandboxes that package a workflow surface, a task distribution, an evaluation function, and full trajectory logging. The product isn't a file. It's a living system.</p>
<div class="divider">· · ·</div>
<h2>II. Anatomy of an RL Environment</h2>
<p>An RL environment is deceptively simple in concept but ferociously hard to build well. At its core, it has four components — and <em>each</em> one is a source of competitive advantage or failure.</p>
<div class="framework">
<div class="fw-title">The Four Pillars of an RL Environment</div>
<div class="fw-grid">
<div class="fw-layer">
<div class="fw-label">Surface</div>
<div class="fw-desc">
<strong>The action space the agent inhabits.</strong> A UI it can click, an API it can call, a codebase it can modify, a terminal it can type into. The surface must faithfully reproduce the real-world task environment — otherwise the agent learns to exploit the simulation, not solve the problem.
<div class="fw-tags"><span class="fw-tag">Browser</span><span class="fw-tag">IDE</span><span class="fw-tag">API</span><span class="fw-tag">Shell</span><span class="fw-tag">GUI</span></div>
</div>
</div>
<div class="fw-layer">
<div class="fw-label">Tasks</div>
<div class="fw-desc">
<strong>A distribution of problems at calibrated difficulty.</strong> Not one task, but a task <em>library</em> — parameterized, varied, and tunable. Too easy and the agent saturates without learning transfer. Too hard and gradients vanish. The art is in the curriculum: starting simple, escalating complexity, and introducing adversarial variation.
<div class="fw-tags"><span class="fw-tag">Task generation</span><span class="fw-tag">Difficulty curves</span><span class="fw-tag">Curriculum</span></div>
</div>
</div>
<div class="fw-layer">
<div class="fw-label">Eval</div>
<div class="fw-desc">
<strong>A scoring function that maps outcomes to rewards.</strong> The eval defines what "good" means. Binary pass/fail is brittle. Dense reward shaping is powerful but can introduce reward hacking. The best evals blend automated checks with LLM-as-judge and, where available, human ground-truth. They are the loss function of the experience era.
<div class="fw-tags"><span class="fw-tag">Automated checks</span><span class="fw-tag">LLM-as-judge</span><span class="fw-tag">Human ground truth</span></div>
</div>
</div>
<div class="fw-layer">
<div class="fw-label">Logging</div>
<div class="fw-desc">
<strong>Full trajectory capture for learning.</strong> Every action, observation, reward, and intermediate state — recorded. This is the data the RL algorithm actually trains on. Incomplete trajectories mean incomplete gradients. Logging quality directly determines training signal quality.
<div class="fw-tags"><span class="fw-tag">Traces</span><span class="fw-tag">State snapshots</span><span class="fw-tag">Reward signals</span></div>
</div>
</div>
</div>
</div>
<p>The critical insight is that <strong>these four components are deeply interdependent.</strong> A beautiful UI surface paired with a weak eval produces agents that look competent in demos and fail in production. A brilliant task library logged with lossy trajectories wastes compute. Building environments is systems engineering, not a wrapper around an existing application.</p>
<div class="divider">· · ·</div>
<h2>III. Why Environments Are the Bottleneck</h2>
<p>Brendan Foody of Mercor makes a provocative claim: <em>"RL is becoming so effective that models will be able to saturate any evaluation."</em> If that's true — and the evidence increasingly supports it — then the binding constraint on AI progress is no longer algorithmic. It's environmental.</p>
<p>Consider the implications. If RL can reliably improve a model on any eval it's trained against, then the frontier of AI capability is defined entirely by the <strong>breadth, fidelity, and difficulty of available environments.</strong> The labs don't need better optimizers. They need harder gyms.</p>
<p>This is why frontier labs have, in the past 18 months, put dozens of environment vendors into business. They are buying the raw material of RL scaling, and the demand is growing faster than supply.</p>
<div class="argument">
<div class="argument-num">01</div>
<div class="argument-body">
<h4>Environments are harder to commoditize than data</h4>
<p>Static datasets can be copied, leaked, or re-created. An environment is a running system — a live surface, a task generator, an eval pipeline, a logging infrastructure. Replicating it requires engineering the entire stack, not just downloading a file. The defensibility is operational, not legal.</p>
</div>
</div>
<div class="argument">
<div class="argument-num">02</div>
<div class="argument-body">
<h4>Domain expertise creates a compounding advantage</h4>
<p>Building a high-fidelity environment for chip design (Phinity), finance (Isidor), or computer use (Fleet AI, Matrices) requires deep domain knowledge — not just of the task, but of the failure modes, edge cases, and difficulty gradients that make RL training productive. This expertise compounds: each iteration reveals what the model finds hard, which informs the next generation of tasks.</p>
</div>
</div>
<div class="argument">
<div class="argument-num">03</div>
<div class="argument-body">
<h4>The "eval gap" is the real AGI bottleneck</h4>
<p>If RL can saturate any eval, then progress reduces to: <em>can we write evals fast enough for everything humans do?</em> This reframes the AGI question. The barrier isn't intelligence — it's instrumentation. Every un-eval'd workflow is a capability the model can't acquire through experience. Environment builders are, in a very real sense, converting human economic activity into AI-legible training signal.</p>
</div>
</div>
<div class="argument">
<div class="argument-num">04</div>
<div class="argument-body">
<h4>Labs are compute-rich but environment-poor</h4>
<p>The asymmetry is structural. Labs have billions in GPU compute and world-class RL researchers. What they lack — and cannot build internally at the rate they need — is coverage across the long tail of real-world workflows. This creates a natural market: labs buy environments the way cloud customers buy SaaS. The vendor ecosystem isn't incidental. It's load-bearing.</p>
</div>
</div>
<div class="divider">· · ·</div>
<h2>IV. The Competitive Landscape</h2>
<p>The environment-building market is nascent but structuring quickly into clear verticals. Each vertical reflects a domain where RL training demand is acute and the task surface is complex enough to sustain a standalone company.</p>
<table class="data-table">
<thead>
<tr>
<th>Domain</th>
<th>Key Players</th>
<th>Why It's Hard</th>
<th>Moat Potential</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Coding</strong></td>
<td>Preference Model, Proximal, Mechanize, AfterQuery</td>
<td>Requires realistic repos, dependency resolution, test suites that catch subtle bugs — not just syntax</td>
<td>High — code environments improve with real codebase diversity</td>
</tr>
<tr>
<td><strong>Computer Use</strong></td>
<td>Fleet AI, Matrices, DeepTune</td>
<td>Pixel-level fidelity, latency simulation, handling non-deterministic UI states across OSes</td>
<td>Very High — cross-application workflows are combinatorially complex</td>
</tr>
<tr>
<td><strong>Chip Design</strong></td>
<td>Phinity AI</td>
<td>EDA tool integration, PPA (power/performance/area) evaluation, multi-step synthesis flows</td>
<td>Very High — extreme domain specialization</td>
</tr>
<tr>
<td><strong>Finance</strong></td>
<td>Isidor</td>
<td>Regulatory constraints, temporal data leakage, market simulation fidelity</td>
<td>High — compliance requirements create natural barriers</td>
</tr>
<tr>
<td><strong>General Knowledge Work</strong></td>
<td>Turing, Mercor, Surge AI, Handshake</td>
<td>Breadth vs. depth tradeoff; must cover many workflows without sacrificing fidelity</td>
<td>Medium — breadth is defensible, but depth may matter more</td>
</tr>
</tbody>
</table>
<p>The incumbents — Turing, Mercor, Surge AI — have the advantage of existing lab relationships and data annotation infrastructure. But the specialized players have something potentially more durable: <strong>task design expertise</strong> that directly translates into model capability improvement. The question is whether labs will consolidate around a few deep vendors or maintain a broad portfolio.</p>
<div class="divider">· · ·</div>
<h2>V. The Sim-to-Real Problem</h2>
<p>Every environment company faces the same existential question: <strong>does training in our sandbox transfer to the real world?</strong> This is the sim-to-real gap, borrowed from robotics, and it is the single largest technical risk in the environment business.</p>
<p>Consider a coding environment. If the tasks are too synthetic — clean functions, isolated logic, perfect test suites — the model learns to ace the gym but stumbles on real codebases with messy dependencies, legacy patterns, and ambiguous requirements. The environment vendor's value is precisely in closing this gap: making their sandbox realistic enough that RL training transfers.</p>
<div class="callout">
<div class="callout-title">The Difficulty Calibration Problem</div>
<p>An underappreciated challenge: environments must stay <em>ahead</em> of the model. As RL training improves the agent, the environment must produce harder tasks, or the model saturates and stops learning. This creates a treadmill — environment companies must continuously expand their task distribution, or their product becomes obsolete as models improve.</p>
<p>The best environment companies will build <strong>procedural task generators</strong> that can scale difficulty programmatically, rather than relying on hand-crafted task libraries that get exhausted.</p>
</div>
<p>Jerry Dworek, former OpenAI research lead, frames the broader concern precisely: <em>"How do those models generalize? How do those models perform outside of what they've been trained for?"</em> If RL environments produce narrow specialists rather than general improvers, the entire value proposition weakens. The counter-argument is that with enough <em>diverse</em> environments — covering enough of the skill surface — generalization emerges from breadth. This is an empirical question, and its answer will determine whether the environment market is a $1B or $100B opportunity.</p>
<div class="divider">· · ·</div>
<h2>VI. Beyond Labs: The Enterprise RL Opportunity</h2>
<p>Today, the primary buyers of RL environments are frontier labs. But the next wave of demand will come from <strong>enterprises and AI-native application companies</strong> that want to fine-tune agents for their specific workflows.</p>
<p>Cursor's use of online RL to improve their tab completion model — using real user acceptance/rejection signals as the reward — is a template. Any company with an AI agent in production is sitting on a natural RL environment: the agent acts, the user provides implicit feedback (acceptance, correction, abandonment), and the trajectory is logged. The missing piece is the infrastructure to close the loop — to convert these production traces into RL training runs.</p>
<div class="callout">
<div class="callout-title">Application: Voice AI at Scale</div>
<p>Consider a company operating AI voice bots handling millions of customer calls. Every call is a trajectory: the agent speaks, the customer responds, the call resolves or escalates. Success metrics are clear — resolution rate, customer satisfaction, handle time. The environment is <em>already built</em>; it's the production system itself. What's needed is the RL infrastructure to train on these trajectories and the eval framework to measure improvement without regressing on edge cases.</p>
<p>This is where "RL-as-a-service" companies (Applied Compute, CGFT Labs, Osmosis AI) enter: they bring the RL expertise and tooling to companies that have the environments but lack the ML infrastructure to exploit them.</p>
</div>
<p>The enterprise opportunity has different dynamics than the lab market. Labs buy environments for <em>general</em> model improvement. Enterprises need environments that mirror their <em>specific</em> workflows — their CRM, their codebase, their customer demographics. This suggests a market where vertical specialization wins, and where the data generated by production deployment feeds back into environment quality in a flywheel that's very hard to replicate from outside.</p>
<div class="divider">· · ·</div>
<h2>VII. Open Questions and Risks</h2>
<h3>Generalization vs. Narrow Skill</h3>
<p>The most important open question in AI: does RL training produce transferable intelligence, or just task-specific muscle memory? If narrow, the environment market fragments into thousands of niche verticals. If broad, a smaller number of high-quality environments might drive general capability — concentrating value in the best gym builders.</p>
<h3>Continual Learning</h3>
<p>Ilya Sutskever's metaphor of the "superintelligent 15-year-old" points at a gap RL alone may not fill. Current RL produces a model that is frozen after training. Real intelligence requires continual adaptation — learning on deployment, updating beliefs, acquiring new skills in context. If continual learning proves necessary, environment companies will need to evolve from batch-training sandboxes to always-on learning systems.</p>
<h3>Reward Hacking at Scale</h3>
<p>As RL scales, so does the attack surface for reward hacking — agents finding ways to maximize the eval score without actually solving the intended problem. This is the environment builder's deepest technical challenge: designing evals robust enough that optimization pressure doesn't find shortcuts. The arms race between RL algorithms and eval robustness will define the quality bar for environment companies.</p>
<h3>Lab Consolidation Risk</h3>
<p>If one or two labs pull decisively ahead in RL scaling, they may verticalize — building environments in-house rather than buying from vendors. The environment market's growth depends on RL scaling being a broad industry trend, not a single-lab phenomenon. The emergence of open-source RL infrastructure (Prime Intellect, Thinky Machines) is a positive signal here.</p>
<div class="divider">· · ·</div>
<h2>VIII. The Thesis, Compressed</h2>
<p>In the experience era, <strong>the quality of the environment determines the ceiling of the agent.</strong> RL can optimize any objective you give it — which means the binding constraint is having the right objectives, in the right action spaces, at the right difficulty level. Environment builders are the supply-side infrastructure of AI progress.</p>
<p>The most investable positions in this stack are:</p>
<div class="argument">
<div class="argument-num">A</div>
<div class="argument-body">
<h4>Deep-domain environment specialists</h4>
<p>Companies that own the task design, eval logic, and fidelity requirements for a specific high-value domain — and whose environments measurably improve model performance on real-world tasks in that domain. The closer the gym to the game, the more valuable the gym.</p>
</div>
</div>
<div class="argument">
<div class="argument-num">B</div>
<div class="argument-body">
<h4>Environment-building infrastructure</h4>
<p>Tools that let <em>anyone</em> create high-quality RL environments — HUD Evals and equivalents that democratize environment construction. As RL moves from labs to enterprises, the demand for "environment IDEs" will explode.</p>
</div>
</div>
<div class="argument">
<div class="argument-num">C</div>
<div class="argument-body">
<h4>Application companies with natural RL flywheels</h4>
<p>Companies where production usage generates trajectories and reward signals that can be fed back into RL training — creating a self-improving loop that widens the gap with competitors who train only on static data. Cursor is the archetype. Every AI-native company should be asking: <em>is my production system also an RL environment?</em></p>
</div>
</div>
<div class="pullquote">
The primary barrier to applying agents to the entire economy is building evals for everything.
<span class="attr">— Brendan Foody, CEO of Mercor</span>
</div>
<p>If Foody is right, then the companies that can convert human economic activity into well-instrumented, well-evaluated RL environments — at the speed the labs demand — are building one of the most durable and consequential businesses in AI. Not the flashiest. Not the one demoing a chatbot. The one building the gym where the chatbot learned to be good.</p>
</div>
<div class="essay-footer">
<span>Built on analysis from Tanay Jaipuria (Wing VC), Silver & Sutton (DeepMind), and frontier lab disclosures.</span>
<span>Feb 2026</span>
</div>
<script>
(function () {
var content = document.querySelector('.content');
if (!content) return;
var rules = [
{ sel: '.content > p', cls: 'sr-up' },
{ sel: '.content h2', cls: 'sr-left' },
{ sel: '.content h3', cls: 'sr-left' },
{ sel: '.pullquote', cls: 'sr-up' },
{ sel: '.callout', cls: 'sr-right' },
{ sel: '.framework', cls: 'sr-scale' },
{ sel: '.fw-layer', cls: 'sr-up', indexed: true },
{ sel: '.argument', cls: 'sr-up', indexed: true },
{ sel: '.data-table', cls: 'sr-scale' },
{ sel: '.divider', cls: 'sr-up' },
];
var all = [];
rules.forEach(function (r) {
var els = content.querySelectorAll(r.sel);
var idx = 0;
els.forEach(function (el) {
el.classList.add('sr', r.cls);
if (r.indexed) { el.style.setProperty('--sr-i', idx); idx++; }
all.push(el);
});
});
var hero = document.querySelector('.hero');
if (hero) { hero.classList.add('sr', 'sr-up'); all.push(hero); hero.classList.add('sr-visible'); }
var obs = new IntersectionObserver(function (entries) {
entries.forEach(function (e) {
if (e.isIntersecting) { e.target.classList.add('sr-visible'); obs.unobserve(e.target); }
});
}, { threshold: 0.12, rootMargin: '0px 0px -60px 0px' });
all.forEach(function (el) { obs.observe(el); });
})();
</script>
<div class="share-bar">
<a href="https://twitter.com/intent/tweet?url=https://omkarray.com/rl-environments.html&text=RL%20Environments%3A%20The%20New%20Data%20Moat%20of%20AI" target="_blank" rel="noopener" aria-label="Share on X">
<svg viewBox="0 0 24 24"><path d="M18.244 2.25h3.308l-7.227 8.26 8.502 11.24H16.17l-5.214-6.817L4.99 21.75H1.68l7.73-8.835L1.254 2.25H8.08l4.713 6.231zm-1.161 17.52h1.833L7.084 4.126H5.117z"/></svg>
</a>
<a href="https://www.linkedin.com/sharing/share-offsite/?url=https://omkarray.com/rl-environments.html" target="_blank" rel="noopener" aria-label="Share on LinkedIn">
<svg viewBox="0 0 24 24"><path d="M20.5 2h-17A1.5 1.5 0 002 3.5v17A1.5 1.5 0 003.5 22h17a1.5 1.5 0 001.5-1.5v-17A1.5 1.5 0 0020.5 2zM8 19H5v-9h3zM6.5 8.25A1.75 1.75 0 118.3 6.5a1.78 1.78 0 01-1.8 1.75zM19 19h-3v-4.74c0-1.42-.6-1.93-1.38-1.93A1.74 1.74 0 0013 14.19a.66.66 0 000 .14V19h-3v-9h2.9v1.3a3.11 3.11 0 012.7-1.4c1.55 0 3.36.86 3.36 3.66z"/></svg>
</a>
<button onclick="navigator.clipboard.writeText(window.location.href);this.querySelector('svg').style.color='#4A7A5C';setTimeout(()=>this.querySelector('svg').style.color='',1500)" aria-label="Copy link">
<svg viewBox="0 0 24 24"><path d="M5 2h14a1 1 0 011 1v1H4V3a1 1 0 011-1zm-2 4h18v15a1 1 0 01-1 1H4a1 1 0 01-1-1V6zm7 3v2h4V9h-4zm0 4v2h4v-2h-4z"/></svg>
</button>
<a href="mailto:?subject=RL%20Environments%3A%20The%20New%20Data%20Moat&body=RL%20environments%20are%20becoming%20the%20next%20big%20moat%20in%20AI%3A%0A%0Ahttps%3A%2F%2Fomkarray.com%2Frl-environments.html" aria-label="Share via email">
<svg viewBox="0 0 24 24"><path d="M20 4H4c-1.1 0-2 .9-2 2v12c0 1.1.9 2 2 2h16c1.1 0 2-.9 2-2V6c0-1.1-.9-2-2-2zm0 4l-8 5-8-5V6l8 5 8-5v2z"/></svg>
</a>
<span class="view-count">
<svg viewBox="0 0 24 24" style="width:18px;height:18px"><path d="M12 4.5C7 4.5 2.73 7.61 1 12c1.73 4.39 6 7.5 11 7.5s9.27-3.11 11-7.5c-1.73-4.39-6-7.5-11-7.5zM12 17c-2.76 0-5-2.24-5-5s2.24-5 5-5 5 2.24 5 5-2.24 5-5 5zm0-8c-1.66 0-3 1.34-3 3s1.34 3 3 3 3-1.34 3-3-1.34-3-3-3z"/></svg>
<span id="view-count-num">—</span>
</span>
</div>
<script>
(function() {
var page = 'rl-environments';
var key = 'omkarray_views_' + page;
var stored = parseInt(localStorage.getItem(key)) || 0;
var isNew = !sessionStorage.getItem('viewed_' + page);
if (isNew) { stored++; localStorage.setItem(key, stored); sessionStorage.setItem('viewed_' + page, '1'); }
document.getElementById('view-count-num').textContent = stored;
})();
</script>
</body>
</html>