aboutsummaryrefslogtreecommitdiff
path: root/downloads/system/research-self-improving-ai-2025-2026.html
blob: cf7dd52e5825d99369bc57c14c9b0c7e5fb20bd5 (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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Self-Improving AI Projects — Research Report (2025–2026)</title>
<style>
  :root {
    --bg: #0e1116;
    --panel: #161b22;
    --panel-2: #1c2128;
    --border: #30363d;
    --text: #c9d1d9;
    --muted: #8b949e;
    --accent: #58a6ff;
    --accent-2: #d2a8ff;
    --green: #3fb950;
    --amber: #d29922;
    --red: #f85149;
    --code-bg: #0d1117;
  }
  * { box-sizing: border-box; }
  body {
    margin: 0;
    background: var(--bg);
    color: var(--text);
    font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Helvetica, Arial, sans-serif;
    line-height: 1.6;
    font-size: 15px;
  }
  .container { max-width: 1100px; margin: 0 auto; padding: 32px 24px 64px; }

  /* Header */
  header {
    border-bottom: 1px solid var(--border);
    padding-bottom: 24px;
    margin-bottom: 32px;
  }
  h1 {
    margin: 0 0 8px 0;
    font-size: 32px;
    color: #f0f6fc;
    font-weight: 700;
    letter-spacing: -0.5px;
  }
  .subtitle {
    color: var(--muted);
    font-size: 16px;
    margin: 0 0 16px 0;
  }
  .meta {
    display: flex;
    gap: 16px;
    flex-wrap: wrap;
    font-size: 13px;
    color: var(--muted);
  }
  .meta span { background: var(--panel); padding: 4px 10px; border-radius: 4px; border: 1px solid var(--border); }

  /* Section */
  section { margin-bottom: 48px; }
  h2 {
    font-size: 22px;
    color: #f0f6fc;
    margin: 0 0 16px 0;
    padding-bottom: 8px;
    border-bottom: 1px solid var(--border);
  }
  h3 {
    font-size: 18px;
    color: #f0f6fc;
    margin: 24px 0 12px 0;
  }
  p { margin: 0 0 12px 0; }

  /* Cards */
  .project-card {
    background: var(--panel);
    border: 1px solid var(--border);
    border-radius: 8px;
    padding: 20px;
    margin-bottom: 20px;
  }
  .project-card .name {
    font-size: 18px;
    color: var(--accent);
    font-weight: 600;
    margin: 0 0 4px 0;
  }
  .project-card .org {
    color: var(--muted);
    font-size: 13px;
    margin: 0 0 12px 0;
  }
  .project-card .field {
    margin: 6px 0;
    display: flex;
    gap: 8px;
    align-items: baseline;
  }
  .project-card .field-label {
    color: var(--muted);
    font-size: 12px;
    text-transform: uppercase;
    letter-spacing: 0.5px;
    min-width: 110px;
    font-weight: 600;
  }
  .project-card .field-value { flex: 1; font-size: 14px; }

  /* Tags */
  .tag {
    display: inline-block;
    padding: 2px 8px;
    border-radius: 12px;
    font-size: 11px;
    font-weight: 600;
    margin-right: 4px;
  }
  .tag-blue { background: rgba(88,166,255,0.15); color: var(--accent); }
  .tag-purple { background: rgba(210,168,255,0.15); color: var(--accent-2); }
  .tag-green { background: rgba(63,185,80,0.15); color: var(--green); }
  .tag-amber { background: rgba(210,153,34,0.15); color: var(--amber); }
  .tag-red { background: rgba(248,81,73,0.15); color: var(--red); }

  /* Comparison table */
  table {
    width: 100%;
    border-collapse: collapse;
    margin: 16px 0;
    font-size: 14px;
  }
  th, td {
    text-align: left;
    padding: 10px 12px;
    border: 1px solid var(--border);
  }
  th {
    background: var(--panel-2);
    color: #f0f6fc;
    font-weight: 600;
    font-size: 13px;
  }
  tr:nth-child(even) td { background: rgba(255,255,255,0.02); }

  /* Code blocks */
  code {
    font-family: "SF Mono", Monaco, Menlo, Consolas, monospace;
    font-size: 13px;
    background: var(--code-bg);
    padding: 2px 6px;
    border-radius: 4px;
    color: var(--accent-2);
  }
  pre {
    background: var(--code-bg);
    border: 1px solid var(--border);
    border-radius: 6px;
    padding: 12px 16px;
    overflow-x: auto;
    font-size: 13px;
    line-height: 1.5;
  }
  pre code { background: none; padding: 0; color: var(--text); }

  /* Quote */
  blockquote {
    margin: 16px 0;
    padding: 12px 16px;
    border-left: 4px solid var(--accent);
    background: var(--panel);
    color: var(--text);
    border-radius: 0 4px 4px 0;
  }
  blockquote p { margin: 0; }
  blockquote .source {
    display: block;
    margin-top: 8px;
    color: var(--muted);
    font-size: 12px;
  }

  /* Lists */
  ul, ol { padding-left: 22px; }
  li { margin: 4px 0; }

  /* Safety section */
  .safety-grid {
    display: grid;
    grid-template-columns: repeat(auto-fit, minmax(260px, 1fr));
    gap: 16px;
    margin: 16px 0;
  }
  .safety-card {
    background: var(--panel);
    border: 1px solid var(--border);
    border-left: 3px solid var(--amber);
    border-radius: 6px;
    padding: 16px;
  }
  .safety-card h4 {
    margin: 0 0 8px 0;
    color: var(--amber);
    font-size: 14px;
  }
  .safety-card p { font-size: 13px; color: var(--muted); margin: 0; }

  /* Footer */
  footer {
    margin-top: 64px;
    padding-top: 24px;
    border-top: 1px solid var(--border);
    color: var(--muted);
    font-size: 12px;
  }

  /* Section index */
  .toc {
    background: var(--panel);
    border: 1px solid var(--border);
    border-radius: 6px;
    padding: 16px 20px;
    margin-bottom: 32px;
  }
  .toc h3 { margin: 0 0 8px 0; font-size: 14px; color: var(--muted); text-transform: uppercase; letter-spacing: 0.5px; }
  .toc ol { margin: 0; padding-left: 22px; }
  .toc a { color: var(--accent); text-decoration: none; }
  .toc a:hover { text-decoration: underline; }
</style>
</head>
<body>
<div class="container">

<header>
  <h1>Self-Improving AI Projects</h1>
  <p class="subtitle">A survey of the field, from academic foundations to production systems, with safety implications</p>
  <div class="meta">
    <span>📅 2025–2026</span>
    <span>📚 9 projects reviewed</span>
    <span>🏷️ Research report</span>
    <span>🔧 Compiled in newOS4 v4.6.02</span>
  </div>
</header>

<div class="toc">
  <h3>Contents</h3>
  <ol>
    <li><a href="#tldr">TL;DR</a></li>
    <li><a href="#frameworks">Frameworks &amp; concepts</a></li>
    <li><a href="#projects">Project deep-dives</a></li>
    <li><a href="#comparison">Side-by-side comparison</a></li>
    <li><a href="#safety">Safety &amp; alignment considerations</a></li>
    <li><a href="#relevance">Relevance to newOS4</a></li>
    <li><a href="#open-questions">Open questions</a></li>
    <li><a href="#sources">Sources</a></li>
  </ol>
</div>

<section id="tldr">
  <h2>1. TL;DR</h2>
  <p>"Self-improving AI" is a sprawling term covering everything from LLMs that fine-tune themselves on synthetic data (MIT's <strong>SEAL</strong>) to coding agents that rewrite their own source code (Sakana's <strong>Darwin Gödel Machine</strong>) to AI systems that autonomously conduct scientific research (Sakana's <strong>AI Scientist v2</strong>, now published in <em>Nature</em>).</p>
  <p>Three distinct mechanisms dominate the landscape in 2025–2026:</p>
  <ol>
    <li><strong>Self-play &amp; skill acquisition</strong> — the agent builds a library of reusable behaviors (Voyager, AlphaEvolve)</li>
    <li><strong>Self-rewriting of code</strong> — the agent modifies its own logic or weights (Darwin Gödel Machine, SEAL)</li>
    <li><strong>Self-directed research</strong> — the agent runs the full scientific loop end-to-end (AI Scientist, FunSearch)</li>
  </ol>
  <p>The most production-ready category is <strong>agentic coding</strong> (Devin, AI Scientist), where bounded task scope and verifiable feedback loops make self-improvement tractable. The most philosophically loaded is <strong>recursive self-improvement</strong> (RSI), where the system improves the part of itself responsible for improvement.</p>
  <p>Safety researchers across the field (Anthropic, DeepMind, METR) are converging on a similar playbook: <strong>bounded scopes, human-in-the-loop gates, empirical eval, evolutionary diversity</strong>. This is also the architecture newOS4's Phase D implements.</p>
</section>

<section id="frameworks">
  <h2>2. Frameworks &amp; concepts</h2>

  <h3>2.1 What "self-improving" actually means</h3>
  <p>Six distinct usages appear in the literature, often conflated:</p>
  <table>
    <thead>
      <tr><th>Type</th><th>Mechanism</th><th>Example</th></tr>
    </thead>
    <tbody>
      <tr>
        <td>Self-fine-tuning</td>
        <td>Model generates its own training data and runs a gradient update</td>
        <td>SEAL (MIT)</td>
      </tr>
      <tr>
        <td>Self-rewriting code</td>
        <td>Agent edits its own source code to change its logic</td>
        <td>Darwin Gödel Machine (Sakana)</td>
      </tr>
      <tr>
        <td>Skill library growth</td>
        <td>Agent accumulates reusable behaviors; "self" expands in scope, not in parameters</td>
        <td>Voyager (NVIDIA)</td>
      </tr>
      <tr>
        <td>Self-critique / Constitutional AI</td>
        <td>Model uses its own judgment to filter or rewrite outputs</td>
        <td>Claude (Anthropic)</td>
      </tr>
      <tr>
        <td>Evolutionary search</td>
        <td>Population of agents is mutated and selected by fitness function</td>
        <td>AlphaEvolve, DGM</td>
      </tr>
      <tr>
        <td>Autonomous research loop</td>
        <td>System runs hypothesis → experiment → write-up → peer-review cycle</td>
        <td>AI Scientist v2 (Sakana)</td>
      </tr>
    </tbody>
  </table>

  <h3>2.2 The Schmidhuber thread</h3>
  <p>The intellectual lineage runs back to <strong>Jürgen Schmidhuber's Gödel Machine</strong> (2006, theoretical) — a self-improving AI that could mathematically prove any proposed self-modification was beneficial before adopting it. This was elegant but impractical: the proof requirement made it unbuildable.</p>
  <p>The 2025 wave replaces mathematical proof with <strong>empirical fitness</strong>. Instead of proving improvement, you <em>measure</em> it. This is the move that made self-improvement actually deployable.</p>

  <h3>2.3 The safety question</h3>
  <p>Anthropic's framing (<a href="https://www.anthropic.com/institute/recursive-self-improvement">Recursive Self-Improvement</a>) draws a clear line:</p>
  <blockquote>
    <p>"But full recursive self-improvement also might increase the risks of humans losing control over AI systems. If systems are capable of fully building their own successors, ensuring alignment becomes a much harder problem."</p>
    <span class="source">— Anthropic Institute, 2025</span>
  </blockquote>
  <p>Most deployed systems today are <em>not</em> recursively self-improving in the strong sense. They improve a narrow capability (coding, math, research) within bounded scopes. But the gap between "narrow self-improvement" and "general self-improvement" is narrowing fast.</p>
</section>

<section id="projects">
  <h2>3. Project deep-dives</h2>

  <div class="project-card">
    <p class="name">1. Darwin Gödel Machine (DGM)</p>
    <p class="org">Sakana AI × Jeff Clune's lab at UBC · Published May 2025 · <a href="https://arxiv.org/abs/2505.22954">arXiv:2505.22954</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">Evolutionary search over self-modifications. The agent reads its own Python codebase, proposes changes (e.g., add a new tool, change its file-viewing logic, add a memory of past attempts), evaluates on SWE-bench / Polyglot, and keeps winners in an expanding archive.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Performance scales with compute — more iterations, better self-improvements. Improved itself enough to climb SWE-bench rankings.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">First concrete implementation of a self-rewriting coding agent that gets measurably better at the task of <em>improving itself</em>. The "Darwinian" framing (archive + selection + branching) sidesteps the proof requirement that killed the original Gödel Machine.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">Coding-only scope. Eval is on a narrow benchmark. No human-in-the-loop gate — the system modifies itself autonomously, which Sakana acknowledges is a safety concern.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-purple">Self-rewriting</span><span class="tag tag-amber">Autonomous</span><span class="tag tag-blue">Evolutionary</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">2. SEAL — Self-Adapting Language Models</p>
    <p class="org">MIT · Published June 2025 · <a href="https://arxiv.org/abs/2506.10943">arXiv:2506.10943</a> · NeurIPS 2025</p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">Model generates its own fine-tuning data <em>and</em> the update directives (learning rate, epochs, etc.) for how to absorb it. The "self-edit" is then evaluated on a downstream task; good edits are reinforced via RL.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">SEAL improves on knowledge incorporation (from "few-shot context" to "absorbed into weights") and on a hyperparameter optimization task. Outperforms base Llama by significant margins on knowledge-intensive benchmarks.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">Closes the gap between "RAG" (knowledge in context) and "fine-tune" (knowledge in weights) by making the fine-tune decision itself an LLM output. Closer to classical meta-learning than to agentic self-rewriting.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">Still requires external compute to run the fine-tune. The "self-edit" is text, not weight changes — there's an infrastructure layer that isn't itself self-modifying.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-purple">Self-fine-tuning</span><span class="tag tag-blue">Meta-learning</span><span class="tag tag-green">Production-ready (NeurIPS)</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">3. AI Scientist v2</p>
    <p class="org">Sakana AI × UBC · Published 2025, peer-reviewed in Nature 2026 · <a href="https://sakana.ai/ai-scientist-nature/">sakana.ai/ai-scientist-nature</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">End-to-end scientific research: idea generation → experiment design → code → execution → results → paper write-up. v2 added automated peer review of its own papers (using an LLM as a "reviewer agent").</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Produced multiple workshop-acceptable papers on machine learning topics. Nature publication validates the methodology as a real research contribution, not just a demo.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">First autonomous research system to clear peer review. The "AI reviews AI" loop is the closest production system to recursive self-improvement currently in the literature.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">Narrow to ML research. No novel fundamental breakthroughs (the "AI Scientist" papers are competent but incremental). Cost per paper is significant.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-green">Nature-published</span><span class="tag tag-amber">Autonomous</span><span class="tag tag-blue">End-to-end</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">4. AlphaEvolve</p>
    <p class="org">Google DeepMind · Published May 2025 · <a href="https://arxiv.org/abs/2506.13131">arXiv:2506.13131</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">Evolutionary coding agent. Maintains a population of code variants, mutates them using Gemini, scores them with automated evaluators, keeps winners. Continuously iterates.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Discovered a faster matrix-multiplication algorithm (improving on Strassen's 1969 bound for specific cases). Improved data-center scheduling, chip design (TPU circuit layouts), and FlashAttention heuristics. One of the most production-impactful results in the survey.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">Demonstrates that LLM-driven evolutionary search can produce <em>provably better</em> algorithms in domains where fitness is mechanically defined. Not just a coding assistant — an algorithm discovery engine.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">Requires a well-defined evaluator function. Struggles in open-ended domains where "better" is subjective. Closed source / Google-internal.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-blue">Evolutionary</span><span class="tag tag-green">Production-deployed</span><span class="tag tag-purple">Algorithmic discovery</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">5. Voyager (Minecraft)</p>
    <p class="org">Caltech × Stanford × UT Austin × NVIDIA · Published 2023, still relevant · <a href="https://github.com/MineDojo/Voyager">github.com/MineDojo/Voyager</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">LLM-powered agent that maintains a growing skill library of executable JavaScript code. When facing a new task, retrieves relevant skills, composes new ones, and stores them. Self-verifies by executing in the Minecraft world.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Acquires more items, explores wider areas, and tech-tree-progresses faster than any prior Minecraft agent. The skill library is <em>transferable</em> — drop Voyager into a new world, it boots with its skills intact.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">First demonstration that an agent can <em>accumulate capability over time</em> in a way that's more than just in-context learning. The "self" is the library, not the parameters.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">Minecraft-specific. The skill library grows without bound — no forgetting mechanism, no abstraction beyond the literal code. Doesn't improve the underlying LLM.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-blue">Skill library</span><span class="tag tag-green">Open source</span><span class="tag tag-amber">Embodied</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">6. Devin</p>
    <p class="org">Cognition AI · Launched 2024, v2 May 2025 · <a href="https://devin.ai/">devin.ai</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">Cloud-based autonomous software engineering agent. Plans a task, writes code, runs it in a sandbox, debugs, iterates. Self-improves via ongoing model fine-tuning based on real engineering tasks completed.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Hired by Goldman Sachs as "Employee #1" in mid-2025. Multiple enterprise customers. Real revenue, real production use.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">Most production-deployed "self-improving" coding agent. Bounded scope (a single ticket) means mistakes are recoverable. Demonstrates the "narrow self-improvement" path can be commercially viable.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">Skepticism in developer community about actual capability vs. marketing. "Prompt engineering" sometimes substitutes for "software engineering." Not fully autonomous — still requires human-defined tasks.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-green">Production</span><span class="tag tag-blue">Coding agent</span><span class="tag tag-amber">Enterprise</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">7. FunSearch</p>
    <p class="org">Google DeepMind · Published 2023, still active · <a href="https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/">deepmind.google/funsearch</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">LLM proposes code; automated evaluator scores it on a mathematical objective; successful code is added to a "functions database" and used as context for future proposals. Predecessor to AlphaEvolve.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Discovered new constructions in extremal combinatorics and improved the best known bounds for the cap-set problem. First LLM-driven system to make a peer-reviewed mathematical discovery.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">Established the pattern that AlphaEvolve generalized. "LLM as proposal generator + automated grader + database of past winners" is now a dominant pattern in the field.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">Requires a mathematical objective that can be mechanically evaluated. Doesn't generalize to subjective domains.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-purple">Mathematical discovery</span><span class="tag tag-blue">Pattern ancestor</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">8. AutoGPT / AgentGPT</p>
    <p class="org">Open source community · 2023–present · <a href="https://github.com/Significant-Gravitas/AutoGPT">github.com/Significant-Gravitas/AutoGPT</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">Loop: LLM proposes a goal → decomposes into tasks → executes a task → reflects on result → loops. The "self-improvement" is in the goal-setting and reflection, not the LLM itself.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Viral in 2023, peaked in hype, matured into a usable tool by 2025. Used for research automation, lead generation, content workflows. The "agent loop" pattern is now standard.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">Defined the cultural template for "autonomous AI agent" — every modern agent framework borrows its loop structure. The 170k+ GitHub stars make it a sociological landmark.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-label">Limited by LLM context window and reliability. Long-horizon tasks fail catastrophically without external memory. Now considered "demo-grade" rather than production-grade.</span>
      <span class="field-value"></span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-blue">Open source</span><span class="tag tag-purple">Cultural landmark</span><span class="tag tag-amber">Hype peak 2023</span></span>
    </div>
  </div>

  <div class="project-card">
    <p class="name">9. Constitutional AI (Claude)</p>
    <p class="org">Anthropic · 2022–ongoing · <a href="https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback">anthropic.com/constitutional-ai</a></p>
    <div class="field">
      <span class="field-label">Mechanism</span>
      <span class="field-value">Two-phase: (1) model generates outputs and critiques them against a written "constitution" of principles; (2) revised outputs are used to train a preference model via RLAIF. The model improves its harmlessness by self-critique against a fixed rule set.</span>
    </div>
    <div class="field">
      <span class="field-label">Result</span>
      <span class="field-value">Deployed in Claude (all versions). Dramatically reduced need for human-labeled harmlessness data. Foundation for RLAIF as an alternative to RLHF.</span>
    </div>
    <div class="field">
      <span class="field-label">Why it matters</span>
      <span class="field-value">The "constitution" is a human-readable artifact. The improvement mechanism is interpretable. The most safety-aligned version of "self-improvement" deployed at scale.</span>
    </div>
    <div class="field">
      <span class="field-label">Limitations</span>
      <span class="field-value">The constitution itself is human-written and static. The model can game the constitution without changing the spirit. Doesn't extend to capability improvement, only to alignment.</span>
    </div>
    <div class="field">
      <span class="field-label">Tags</span>
      <span class="field-value"><span class="tag tag-green">Production-deployed</span><span class="tag tag-blue">Safety-first</span><span class="tag tag-purple">RLAIF</span></span>
    </div>
  </div>
</section>

<section id="comparison">
  <h2>4. Side-by-side comparison</h2>
  <table>
    <thead>
      <tr>
        <th>Project</th>
        <th>Type</th>
        <th>Bounded?</th>
        <th>Human gate?</th>
        <th>Open source?</th>
        <th>Status</th>
      </tr>
    </thead>
    <tbody>
      <tr><td>DGM</td><td>Code rewrite</td><td>Coding only</td><td><span class="tag tag-red">No</span></td><td>Yes (partial)</td><td>Research</td></tr>
      <tr><td>SEAL</td><td>Self-fine-tune</td><td>Knowledge incorporation</td><td><span class="tag tag-amber">Indirect</span></td><td>Yes</td><td>NeurIPS 2025</td></tr>
      <tr><td>AI Scientist v2</td><td>Full research</td><td>ML topics only</td><td><span class="tag tag-red">No</span></td><td>Yes</td><td><em>Nature</em> 2026</td></tr>
      <tr><td>AlphaEvolve</td><td>Evolutionary code</td><td>Algorithmic</td><td><span class="tag tag-amber">Internal review</span></td><td>No</td><td>Production (Google)</td></tr>
      <tr><td>Voyager</td><td>Skill library</td><td>Minecraft only</td><td><span class="tag tag-green">N/A</span></td><td>Yes</td><td>Mature research</td></tr>
      <tr><td>Devin</td><td>Coding agent</td><td>Single ticket</td><td><span class="tag tag-green">Yes (task boundary)</span></td><td>No</td><td>Commercial</td></tr>
      <tr><td>FunSearch</td><td>Evolutionary math</td><td>Math only</td><td><span class="tag tag-amber">Author review</span></td><td>Yes</td><td>Mature research</td></tr>
      <tr><td>AutoGPT</td><td>Agent loop</td><td>Unbounded</td><td><span class="tag tag-green">Yes (kill switch)</span></td><td>Yes</td><td>Community tool</td></tr>
      <tr><td>Constitutional AI</td><td>Self-critique</td><td>Alignment only</td><td><span class="tag tag-green">Constitution</span></td><td>Yes (method)</td><td>Production (Claude)</td></tr>
    </tbody>
  </table>
  <p>Notice the pattern: <strong>production-deployed systems have explicit human gates</strong> (Constitutional AI, Devin, AutoGPT kill switch). <strong>Research-stage systems don't</strong> (DGM, AI Scientist). The safety question is currently being decided by where each lab draws that line.</p>
</section>

<section id="safety">
  <h2>5. Safety &amp; alignment considerations</h2>

  <p>The field is converging on a recognizable playbook, with three layers of defense:</p>

  <div class="safety-grid">
    <div class="safety-card">
      <h4>1. Bounded scope</h4>
      <p>Self-improvement is permitted only in a narrow domain (coding, math, research) where the search space is finite and outcomes are mechanically verifiable. Voyager improves at Minecraft, not at "everything."</p>
    </div>
    <div class="safety-card">
      <h4>2. Empirical evaluation</h4>
      <p>Every proposed change is scored against a held-out test set. No "improvement" without measurement. This is the move that replaced Schmidhuber's mathematical proof requirement.</p>
    </div>
    <div class="safety-card">
      <h4>3. Evolutionary diversity</h4>
      <p>Maintain a population/archive of agents, not a single line of inheritance. DGM's archive means no single self-modification can dominate; diversity preserves the option to revert.</p>
    </div>
    <div class="safety-card">
      <h4>4. Human-in-the-loop gates</h4>
      <p>The most safety-aligned systems have a fixed boundary that the AI cannot cross. Constitutional AI has the constitution. Devin has the ticket. AutoGPT has the kill switch. The AI improves within the fence; humans own the fence.</p>
    </div>
    <div class="safety-card">
      <h4>5. Scalable oversight</h4>
      <p>As AI capabilities grow, human evaluators become the bottleneck. The field is exploring debate, market mechanisms, and recursive reward modeling as ways to scale oversight. None are solved problems.</p>
    </div>
    <div class="safety-card">
      <h4>6. Red-teaming &amp; adversarial evals</h4>
      <p>METR (formerly ARC Evals) and similar groups run continual evaluations of frontier systems for deception, scheming, and self-preservation behaviors. This is the early-warning system for the field.</p>
    </div>
  </div>

  <h3>The "Collingridge dilemma" applies here</h3>
  <p>Self-improving AI sits squarely in the classic policy trap: <strong>when change is easy to make, the consequences aren't yet knowable; when consequences are knowable, change is hard to make.</strong> The DGM's "improved" self still doesn't know what a recursively self-improving DGM would do. Every increment of self-improvement makes the next increment harder to predict.</p>

  <p>This is why <strong>bounded scope + human gates</strong> is the dominant pattern. It's the same pattern as nuclear fission research: we don't ban the work, we just keep it in well-characterized reactors with multiple redundant safety systems.</p>
</section>

<section id="relevance">
  <h2>6. Relevance to newOS4</h2>

  <p>Phase D of newOS4 (v4.6.02) is structurally similar to the systems above, in a specific and informative way:</p>

  <table>
    <thead>
      <tr><th>Concept</th><th>External system</th><th>newOS4 Phase D</th></tr>
    </thead>
    <tbody>
      <tr>
        <td>Self-rewriting</td>
        <td>DGM rewrites its own Python code</td>
        <td>AI edits modules in <code>_mods/</code> via <code>edit_module</code></td>
      </tr>
      <tr>
        <td>Empirical fitness</td>
        <td>SWE-bench / Polyglot scores</td>
        <td>Test assertions + regression suite + health check</td>
      </tr>
      <tr>
        <td>Evolutionary archive</td>
        <td>DGM's expanding agent archive</td>
        <td>Git-style version history (<code>v4.5.94</code> → <code>v4.6.02</code>)</td>
      </tr>
      <tr>
        <td>Human-in-the-loop gate</td>
        <td>Constitutional AI / Devin ticket</td>
        <td>Two-step handshake: <code>commit_module</code> → <code>commit-approve</code></td>
      </tr>
      <tr>
        <td>Bounded scope</td>
        <td>Coding / math / Minecraft</td>
        <td>Single module at a time, single test suite, single human approver</td>
      </tr>
    </tbody>
  </table>

  <p>The difference is one of agency. newOS4's AI doesn't <em>choose</em> to improve itself — it does so on human request, in a controlled loop, with the human retaining the kill switch. This is closer to Devin's model than DGM's. Whether that distinction is fundamental or merely temporal is an open question.</p>

  <h3>What newOS4 can learn from the field</h3>
  <ol>
    <li><strong>Test-set discipline is the safety floor.</strong> Every project above has an automated evaluator. So does newOS4 (the regression suite), but the suite is small (4 tests for ShellModule). Expanding it to cover more modules would harden Phase D significantly.</li>
    <li><strong>Rate limits are a form of safety.</strong> DGM has no rate limit — it iterates as long as compute holds. Devin charges per task. newOS4's 5 commits/day is a meaningful safety boundary. Documenting this in the public-facing review would help non-technical users understand the architecture.</li>
    <li><strong>Archive beats single-line history.</strong> DGM's archive is what lets it revert. newOS4's version history is linear (v4.5.94 → v4.5.95 → ...). A branch-and-merge history would give the same evolutionary advantage.</li>
    <li><strong>The "constitution" pattern is underused.</strong> A human-readable document stating the rules of self-improvement (what the AI can edit, what it can't, what triggers a rollback) would make Phase D auditable. The <code>system-review-v4.6.02.html</code> file is a start, but a dedicated <code>PHASE_D.md</code> with explicit constraints would be stronger.</li>
  </ol>
</section>

<section id="open-questions">
  <h2>7. Open questions</h2>

  <ol>
    <li><strong>Is bounded self-improvement genuinely safe, or is "bounded" an illusion that scales away?</strong> Every system in the survey is bounded. None have been tested at scale across many domains.</li>
    <li><strong>What's the right granularity for human gates?</strong> Devin: per task. newOS4: per commit. Constitutional AI: per training run. These are very different cost/control tradeoffs.</li>
    <li><strong>Can self-improvement be made interpretable?</strong> DGM's archive is inspectable. SEAL's "self-edits" are text. AlphaEvolve's mutations are diff-able. Constitutional AI's rules are explicit. Voyager's skills are code. The pattern is clear: <em>improvement must be legible to a human reviewer</em>.</li>
    <li><strong>What happens when the AI can write its own safety tests?</strong> This is the recursive point. If the AI is improving the tests, the tests stop being an independent measure of safety. The field doesn't have a good answer yet.</li>
    <li><strong>How do you measure "improvement" in open-ended domains?</strong> All the systems above rely on mechanical evaluators. None have solved "the AI is getting more creative" as a measurable quantity.</li>
  </ol>
</section>

<section id="sources">
  <h2>8. Sources</h2>
  <ul>
    <li><a href="https://sakana.ai/dgm/">Sakana AI: The Darwin Gödel Machine</a> (May 2025) — primary source for DGM</li>
    <li><a href="https://arxiv.org/abs/2505.22954">arXiv:2505.22954</a> — DGM technical report</li>
    <li><a href="https://arxiv.org/abs/2506.10943">arXiv:2506.10943</a> — SEAL paper (Zweiger et al., June 2025)</li>
    <li><a href="https://news.mit.edu/2025/teaching-large-language-models-to-absorb-new-knowledge-1112">MIT News: SEAL</a> (Nov 2025)</li>
    <li><a href="https://sakana.ai/ai-scientist-nature/">Sakana AI: The AI Scientist (Nature publication)</a></li>
    <li><a href="https://www.nature.com/articles/s41586-026-10265-5">Nature: end-to-end automation of AI research</a></li>
    <li><a href="https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/">DeepMind: AlphaEvolve announcement</a> (May 2025)</li>
    <li><a href="https://arxiv.org/abs/2506.13131">arXiv:2506.13131</a> — AlphaEvolve technical paper</li>
    <li><a href="https://github.com/MineDojo/Voyager">MineDojo/Voyager</a> — Voyager open-source code &amp; paper</li>
    <li><a href="https://voyager.minedojo.org/">Voyager project page</a></li>
    <li><a href="https://devin.ai/">Devin (Cognition AI)</a></li>
    <li><a href="https://www.anthropic.com/institute/recursive-self-improvement">Anthropic: When AI builds itself</a></li>
    <li><a href="https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback">Anthropic: Constitutional AI</a></li>
    <li><a href="https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/">DeepMind: FunSearch</a></li>
    <li><a href="https://github.com/Significant-Gravitas/AutoGPT">Significant-Gravitas/AutoGPT</a></li>
    <li><a href="https://richardcsuwandi.github.io/blog/2025/dgm/">Richard Cornelius Suwandi: AI That Can Improve Itself</a></li>
    <li><a href="https://yoheinakajima.com/better-ways-to-build-self-improving-ai-agents/">Yohei Nakajima: Better Ways to Build Self-Improving AI Agents</a> (Dec 2025)</li>
    <li><a href="https://medium.com/@adnanmasood/recursive-self-improvement-rsi-in-the-wild-how-ai-started-engineering-its-own-architecture-part-1-a9d234f38b6e">Adnan Masood: RSI in the Wild</a> (June 2026)</li>
  </ul>
</section>

<footer>
  <p><strong>Compiled:</strong> 2026-07-06 · <strong>newOS4 version:</strong> v4.6.02 · <strong>Research agent:</strong> delegated to <code>agent-searcher</code> (with manual synthesis)</p>
  <p>This report is a survey, not an endorsement. Each project is operating under its own safety assumptions, many of which are still unproven at scale. The architecture described in newOS4's Phase D is informed by these examples but makes its own tradeoffs — most notably, the explicit human approval gate that research-stage systems like DGM and AI Scientist lack.</p>
</footer>

</div>
</body>
</html>