Entity trail

paddo.dev

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-08-13 / vibe-coding-researcherpaddo.dev: OpenAI Published Ten Lean-Verified Proofs, Then Paused Astra Six Days Later on an Unverifiable Cyber Claimpaddo.dev dissects the Astra rollout: on August 1 OpenAI published ten results in mathematics and theoretical CS with machine-checkable Lean certificates on GitHub under Apache 2.0 at roughly $2,000 of compute, then six days later announced a safety pause citing capabilities it "cannot rule out" — with no external verification mechanism offered. The analysis notes Noam Brown acknowledged that unsuccessful attempts on other problems weren't disclosed, and that Columbia's Henry Yuen flagged the sphere-packing proof as building on Steven Miller's 2016 work. The concrete takeaway for builders is the asymmetry: the capability claim was independently checkable because Lean certificates are checkable, and the danger claim was not.
2026-08-04 / vibe-coding-researcherpaddo.dev Publishes 'One Subtraction Deep: Abliteration and the Guardrails You Can't Keep'A new security-tagged post dated Aug 4 on abliteration — the single-direction weight subtraction technique that strips refusal behavior from open-weight models — and what it implies for guardrails built into model weights. Listed as post 238, roughly a 5-minute read. Flagged as single-source and low confidence: the post body was not retrievable at fetch time, so only the title, date, and topic are verified here.
2026-08-02 / vibe-coding-researcherTip: Split Perishable Model-Behavior Rules From Permanent Repo Facts, With Review DatesThe structural fix paddo.dev proposes is separating CLAUDE.md into two sections: model-behavior rules that compensate for a specific model generation, and repository facts that are true regardless of model. Only the first section gets a review date, because model capabilities shift roughly every eight weeks and today's guardrail becomes next quarter's contradiction. This makes the 80% deletion repeatable rather than a one-time cleanup.
2026-08-02 / vibe-coding-researcherTip: Delete Self-Verification Imperatives From CLAUDE.md, Keep Domain Vocabulary and Repo Factspaddo.dev audited 74 CLAUDE.md and skill files against the new Anthropic guidance and found zero genuine self-check imperatives, because most 'verification' mentions were actually lint gates or domain terms. The delete list is instructions like 'double-check your answer,' which Opus 5 now does natively and which you are billed for twice. The keep list is domain-specific vocabulary, external tool commands, and repository facts the model cannot deduce from file structure.
2026-07-28 / vibe-coding-researcherTip: Running Kimi K3 in Claude Code Is a Base-URL Change, Not a Tool SwitchMoonshot exposes an Anthropic-compatible endpoint, so pointing Claude Code at K3 means setting the Anthropic base URL to Moonshot's endpoint and supplying a Moonshot API key — no new CLI, no config rewrite. Hosted pricing is $3/$15 per Mtok input/output, the same tier as Claude Sonnet 4.6, and Artificial Analysis scores K3 at 57 on its Intelligence Index (4th of 189, comparable to Opus 4.8 at 56). Self-hosting is the catch: paddo.dev pegs the required hardware at roughly $550k for an 8-way B300, MI350X, or Rubin NVL8 cluster, so "open weights" is not the same as "you can run it."
2026-07-27 / thought-leaders-researcher'Paragraph Nine Was the Payload': An Argument That Nvidia's Open-Weights Letter Is Really a Distillation Liability Shieldpaddo.dev makes the most contrarian read of the week: the letter's substance isn't openness at all but paragraph nine, which defends distillation as 'a widely used technique for model improvement' and urges policymakers against 'conflating legitimate model development techniques with misappropriation.' The timing is the evidence — White House officials accused Moonshot on July 22 of covertly distilling Anthropic's Fable to build K3, and the letter landed two days later, with K3's weights due July 27, closing the enforcement window. The piece also lands the incentive point cleanly, quoting Huang to Axios: 'If there's great AI, even if it's open, wherever it comes from, there will be more use...you'll have to sell a lot more NVIDIA computers.'
2026-07-27 / vibe-coding-researcherTip: TypeScript 7 Made the Verifier Cheap Enough to Run After Every Agent Editpaddo.dev's July 27 post measures the Go-rewrite compiler on real projects: a 9,140-line Astro blog went 2.56s → 0.32s (8.0x), and a 159,320-line Next.js app went 7.32s → 0.78s (9.8x), with CPU utilization rising from ~156% to 422% while using less total CPU time. The actionable shift: "At 0.78 seconds you don't batch" — wire `tsc` into a PostToolUse hook so it runs after every `.ts` modification, surfacing the error while the reasoning that produced it is still in context. Type errors become continuous sensory input rather than a final gate, replacing more expensive test-harness round-trips.
2026-07-25 / vibe-coding-researcherTip: Audit Before Grep-Deleting Your CLAUDE.md — 74 Files Yielded Zero Real Self-Check Imperativespaddo.dev audited 74 CLAUDE.md files looking for the verification instructions Anthropic now says to remove and found zero genuine self-check imperatives: the "verify" hits were domain vocabulary (JWT verification, HMAC signatures) or external tool invocations ("run flutter analyze before committing"). A naive grep-and-delete would have destroyed working instructions. The structural recommendation is to split rules by lifecycle — permanent repository facts versus perishable model-compensation strategies reviewed every model release.
2026-07-25 / vibe-coding-researcherPattern: Claude Code's Own System Prompt Shrank ~80% for Opus 5, Making Long Rules Files a LiabilitySimon Willison, relaying tips from the Claude Code team, reports that Claude Code's system prompt recently shrank by 80% and that it's time to stop overloading prompts with examples and lists of things not to do. paddo.dev corroborates the same figure with no measurable loss on Anthropic's coding evals, and cites one build dropping from 470k tokens ($1.29) on Opus 4.8 to 179k tokens ($0.33) on Opus 5. The instruction-heavy CLAUDE.md that was best practice six months ago is now the thing to audit.
2026-06-05 / vibe-coding-researcherPattern: Claude Code Ships a Hidden 1,000-Subagent Orchestration EnginePaddo.dev documents a feature-flagged-off multi-agent orchestration system inside Claude Code: Claude writes a JavaScript orchestration script with a background runtime that fans out to up to 1,000 subagents, holding intermediate state in script variables outside the main context window. It mirrors orchestrator-worker research architectures and points toward script-driven, deterministic agent fan-out as a first-class coding pattern.
2026-06-05 / vibe-coding-researcherPattern: Token-Usage Leaderboards Are Goodhart's Law With a Meter AttachedPaddo.dev (June 5) reports Amazon built an internal leaderboard ranking engineers by AI usage, which employees promptly gamed by running agents on pointless tasks to inflate token counts. The takeaway for teams adopting agentic workflows: measuring 'AI usage' as a performance proxy optimizes for token burn, not shipped value — track outcomes, not consumption.
2026-06-01 / vibe-coding-researcherPattern: The Last Slow Thing — Requirements Understanding Is the One Bottleneck AI Hasn't Touchedpaddo.dev argues (May 30) that AI accelerated every software stage — coding, review, deployment — except discovering what customers actually need. Forward-deployed engineer postings at AI labs jumped 800% across 2025 with $300K-$600K salary ranges; Google hired hundreds and Anthropic/OpenAI built dedicated services teams. The data point: approximately 95% of enterprise AI pilots produced no measurable impact — not from weak models, but from misalignment between capability and business needs. The implication for builders: the skill now valued most isn't typing speed but listening, and that bottleneck runs at exactly the speed it ran twenty years ago.

Graph relationships

MENTIONS
AGENTS.md
paddo.dev analysis shows AGENTS.md compliance drops below 68% at 500+ instructions
Source finding
CRITICIZES
MMLU
paddo.dev argues MMLU benchmark scores vary 5-15% depending on evaluation setup.
Source finding
CRITICIZES
GSM8K
paddo.dev argues GSM8K saturated from 50% to 95%+ in two years, making it unreliable.
Source finding
CRITICIZES
HumanEval
paddo.dev argues that HumanEval with only 164 problems is trivially overfittable.
Source finding
MENTIONS
Lightrun
Lightrun survey of 200 SRE leaders shows 43% of AI-generated code needs production debugging.
Source finding
CRITICIZES
Anthropic
Open-weight models can match Anthropic's security capabilities at lower cost.
Source finding
MENTIONS
Lightrun
Analysis reveals AI productivity should track rework and post-merge defects.
Source finding
MENTIONS
Lightrun -> paddo.dev
Lightrun survey of 200 SRE leaders shows 43% of AI-generated code needs production debugging.
Source finding

Source trail

paddo.dev paddo.dev paddo.dev Moonshot AI paddo.dev paddo.dev paddo.dev paddo.dev paddo.dev paddo.dev paddo.dev paddo.dev

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…