Hyperagents & DGM-H: AI That Levels Up Its Own Leveling 🔥🤯

dgm-hhyperagentslink-notemetacognitive-self-modificationopen-ended-exploration

Yo, listen up—this paper's dropping bombs on self-improving AI. Before we dive in, WHY does this even exist?

Classic self-improvers? They're stuck in a rut. You got a base agent solving tasks, a fixed meta-agent tweaking it... but that meta-agent's hardcoded by humans. It can't evolve. Result? Progress hits a wall fast—especially outside coding, where "good at tasks" ≠ "good at tweaking yourself." 😩 DGM crushed coding by evolving code agents, but its instruction-gen is rigid AF. Hyperagents fix this pain: they make the improver improvable, unlocking self-accelerating gains on any computable task. No more human bottlenecks. Ohhhhh moment? AI doesn't just get better at chess—it gets better at getting better at chess. 🚀

Big Picture: DGM → DGM-H Evolution

DGM was fire for coding: archive of agents self-modify → evaluate → add winners. But meta-parts fixed.

Hyperagents fuse task agent (solves stuff) + meta agent (makes new agents) into one editable Python program. Meta can rewrite itself. Boom: metacognitive self-mod.

PROBLEM (DGM) ❌                 SOLUTION (DGM-H) ✅
════════════════                 ════════════════
Fixed meta                       Editable hyperagent
│                                 │
▼                                 ▼
┌─────────────────────┐          ┌─────────────────────┐
│ Task Agent          │          │ Task Agent          │
│ + Fixed Meta        │ ───────► │ + Editable Meta     │
│ (handcrafted instr) │          │ (self-modifies)     │
│ Coding only         │          │ Any task!           │
└─────────────────────┘          └─────────────────────┘
     Alignment needed                  No alignment needed

DGM-H keeps DGM's archive + open-ended search, but swaps in hyperagents. Parents from archive → self-mod → eval → archive grows. Safety: sandboxed, human oversight.

Mechanics: How Hyperagents Work Step-by-Step

1️⃣ Hyperagent = Task + Meta in One File

  • Task agent: Gets inputs → spits JSON response (e.g., code patch, accept/reject).
  • Meta agent: Sees repo, past evals, iterations left → edits anything (task code, meta code, tools).
  • Tools: bash, file editor (view/create/replace/insert/undo). FM calls (Claude/o1/etc.).

2️⃣ DGM-H Loop (Pseudocode vibes):

Archive starts w/ initial hyperagent
FOR each iteration:
  - Select parents (perf-based, anti-spam kids)
  - EACH parent: metacognitive self-mod → new hyperagent
  - Eval new one on tasks (staged: cheap first, full if promising)
  - Add if valid → archive grows

Nested deets:

  • Parent select: Sigmoid(perf) * 1/(1+kids) → probs → sample.
  • Eval: Train subset first (e.g., 10 papers), full if >0 success.
  • Multi-domain: Avg perf across domains for select—no manual priority.
Flow: Archive ──► Select Parents ──► Self-Mod ──► Eval ──► Add Back
             ▲                                │
             └────────────────────────────────┘ (Open-ended loop)

Results: Gains + Transfer 💀😂

Tested coding (Polyglot), paper review, robotics rewards (Genesis sim), IMO math grading.

| Domain | Initial | DGM-H Best (Test) | Beats DGM? | Beats Custom? | |---------------------|---------|-------------------|------------|---------------| | Coding (pass@1) | 0.084 | 0.267 ✅ | Comparable| N/A | | Paper Review (acc) | 0.0 | 0.710 ✅ | Yes 🔥 | Tie | | Robotics (fitness) | 0.060 | 0.372 ✅ | Yes | Edges out | | Math Grading (acc) | 0.0 | 0.700 (w/transfer)✅ | N/A | +4% over baseline |

Key flex: Ablations flop w/o self-improve or open-ended (stay ~0).
Meta wins: Learns persistent memory, perf tracking, bias detection. Transfers: Paper/robotics hyperagents → math grading imp@50=0.630 (initial=0). Compounds across runs.

Quali gems:

Memory.json: "Gen55 too harsh → Gen64 balanced → Fix over-correction"
PerfTracker: Trends, stats → "Recent avg > old? Keep direction!"

TL;DR BURN THIS IN: DGM-H = DGM + hyperagents → task gains + meta gains that transfer. No domain hacks needed. 🎯

Edge Cases + Safety 🚨

  • Limits: Fixed tasks/distribs, parent select (prelim: it evolves UCB-style but not always best).
  • Safety: Sandbox, no net, human watch, timeouts. Risks? Fast evolution > oversight; eval gaming; bias amp. Future: evolve tasks, full self-ref.

LOCK IT IN: Hyperagents = future of open-ended AI. Improves search for how to improve. Track the GitHub. 👆

You tracking, fam? Deeper on code? Hit me. 😎

← All notes