H
Howardism
Howardism · Vol. 03Masthead · No. 00

A wiki, set in plates.

CuratorHoward TaiIssueVol. 03 · no. 524CadenceBatched, not scheduledBasedTaiwan, 23°NContacthoward@howardism.dev

Working notes on interaction-era AI, kept publicly. Each plate is a domain; each domain carries its brightest notes and the source I'm currently digesting. 524 notes across 383 sources, last filed 5 Oct 2026. Updated whenever a batch finishes — not on a schedule.

Plate · 01
68
filed under agent systems
Activity · 8 wks
Plate · 02
39
filed under agent security
Activity · 8 wks

Agent Security, in order.

Prompt injection, agent identity, and securing autonomous agents.

  1. 01
    AI Control vs. Alignment
    · 16′
  2. 02
    Blocking Monitors Against Malign Coding Agents
    · 18′
  3. 03
    Harness Configuration Defects
    · 17′
  4. 04
    AI-Enabled Influence Operations
    · 15′
  5. All 39 notes in agent security →
Plate · 03
57
filed under ai coding practice
Activity · 8 wks

AI Coding Practice, in order.

Workflow, review discipline, and verification practice for AI-native teams.

  1. 01
    What the Agent-PR Oversight Numbers Can and Cannot Say
    · 14′
  2. 02
    Checkpoint-Gated Convergence
    · 15′
  3. 03
    Context Smells
    · 10′
  4. 04
    Follow-Up Fixes on Agent PRs
    · 17′
  5. All 57 notes in ai coding practice →
Plate · 04
39
filed under evals & benchmarks
Activity · 8 wks

Evals & Benchmarks, in order.

Benchmark design, contamination, LLM judges, and measuring capability.

  1. 01
    Benchmark Task Defects (Spec–Test Mismatch)
    · 13′
  2. 02
    Construct or Roster? What BEI's Entanglement Graph and IRT's Ability Ranking Are Measuring
    · 18′
  3. 03
    The Price of Fixed Capability
    · 15′
  4. 04
    Verbalized-Confidence Soft Scoring for LLM Judges
    · 16′
  5. All 39 notes in evals & benchmarks →
Plate · 05
32
filed under model capability & training
Activity · 8 wks

Model Capability & Training, in order.

Training, scaling laws, test-time compute, and frontier capability.

  1. 01
    Diversity Calibration Under SFT
    · 7′
  2. 02
    Anchored Bellman-Residual Correction (BRACE)
    · 14′
  3. 03
    ExecCritic: Learn to Test, Test to Improve
    · 16′
  4. 04
    Pre-Reasoning Commitment
    · 15′
  5. All 32 notes in model capability & training →
Plate · 06
44
filed under alignment & safety
Activity · 8 wks

Alignment & Safety, in order.

Alignment training, misalignment evals, and reward hacking.

  1. 01
    The Fable Gate: Who It Caps, and What in the Safety Case Is Independent of the Model
    · 11′
  2. 02
    Agentic Self-Modification (Agent-Initiated Weight Updates)
    · 15′
  3. 03
    Corrigibility (and the Shutdown Problem)
    · 17′
  4. 04
    Embedded Evaluation
    · 15′
  5. All 44 notes in alignment & safety →
Plate · 07
13
filed under interpretability
Activity · 8 wks

Interpretability, in order.

Reading model internals: activations and the global workspace.

  1. 01
    Sandbagging Elicitation (Reference & Context Grafting)
    · 13′
  2. 02
    Interference Weights
    · 19′
  3. 03
    Invisible Reasoning (Filler-Token Latent Computation)
    · 24′
  4. 04
    Introspective Coupling
    · 17′
  5. All 13 notes in interpretability →
Plate · 08
11
filed under interaction & multimodal
Activity · 8 wks

Interaction & Multimodal, in order.

Real-time, multimodal, full-duplex, and human-AI collaboration.

  1. 01
    Native Multimodal Modeling: Fusion Depth and I/O Duality
    · 27′
  2. 02
    Content-Driven Intervention
    · 29′
  3. 03
    Live-Path Minimalism
    · 21′
  4. 04
    Why AI Lags at Design
    · 8′
  5. All 11 notes in interaction & multimodal →
Plate · 09
11
filed under formal math
Activity · 8 wks

Formal Math, in order.

Proof search, Lean, and verifier-driven mathematics.

  1. 01
    Statement Drift
    · 19′
  2. 02
    Kernel-Level Proof Auditing
    · 35′
  3. 03
    OEIS Open Benchmark
    · 31′
  4. 04
    Automated Conjecturing
    · 27′
  5. All 11 notes in formal math →
Plate · 10
20
filed under startup & founder
Activity · 8 wks

Startup & Founder, in order.

Building AI-native companies: speed, moats, and lifecycle.

  1. 01
    Forward-Deployed Engineering as a Delivery Layer
    · 18′
  2. 02
    The Solo-Founder Shift
    · 24′
  3. 03
    The 1% Rule for Wedge Selection
    · 8′
  4. 04
    The Orchestrator's Real Workload: Decision Burden, Framing Discipline, and Whether Taste Scales
    · 8′
  5. All 20 notes in startup & founder →
Plate · 11
24
filed under product & org
Activity · 8 wks

Product & Org, in order.

Product cadence, org design, and the AI-native team.

  1. 01
    Build Instead of Buy Under Agentic Coding
    · 14′
  2. 02
    Psychological Costs of AI Adoption
    · 30′
  3. 03
    Pilot-to-Production Gap
    · 47′
  4. 04
    Prototype Fidelity After Cheap Polish
    · 10′
  5. All 24 notes in product & org →
Plate · 12
37
filed under ai economics & labor
Activity · 8 wks
Plate · 13
33
filed under superintelligence trajectory
Activity · 8 wks
Plate · 14
96
filed under entities
Activity · 8 wks

Entities, in order.

Profiles of the people, labs, products, and projects.

  1. 01
    Jev
    · 7′
  2. 02
    Yoshua Bengio
    · 7′
  3. 03
    Claude Opus 5.5
    · 5′
  4. 04
    Google Threat Intelligence Group (GTIG)
    · 5′
  5. All 96 notes in entities →
Plate · ∞

The desk.

The sources behind the wiki, ranked by how often they're cited.

Article
169
Paper
180
Talk
32
Repo
2
Shown
24 / 383