The US government attempted to blacklist Anthropic after the company refused Pentagon demands for unrestricted Claude access, including potential lethal targeting. A federal judge blocked the ban as likely First Amendment retaliation. The dispute reveals a structural problem: serious AI governance arguments only happen inside frontier labs, creating a form of regulatory capture that predates regulation itself.

A founder recounts rebuilding an entire project solo using AI agents by enforcing discipline through machine-held rules — pre-commit hooks, CI gates, and branch protection — rather than relying on agent memory or long contexts. In 18 days with 655 commits, they shipped a full vertical stack including backend, web, mobile, CI, and deployment. The core insight: when rules are enforced by infrastructure, agent discipline is free, and fresh sessions become a feature rather than a limitation.

A developer built a voice cloning pipeline using Qwen3-TTS to preserve their Nigerian accent, discovering a hardcoded min_new_tokens bug that caused mid-sentence truncation. The deeper realization: the 50MB trained model weights are biometric data — anyone with them can impersonate the speaker for life, and unlike passwords, voices cannot be rotated. The author deliberately withheld the weights from GitHub, treating voice fingerprints as biometric data that should never be committed to public repositories.

A detailed engineering walkthrough of rebuilding a RAG retrieval layer for production, covering document-type-specific chunking strategies (recursive, semantic, agentic), hybrid retrieval combining vector search and BM25 with reciprocal rank fusion, and cross-encoder reranking. The author reports achieving 95% recall@10 and 40% latency reduction by moving beyond naive fixed-token chunking and pure vector search.

A practical guide to building automated LLM evaluation pipelines that replace manual 'looks good to me' reviews. The author shares a production architecture with multiple LLM judges (faithfulness, instruction following, JSON schema, safety, domain expert), golden dataset management, and CI/CD integration that caught 92% of hallucinations before deployment.

A detailed writeup of fine-tuning Google's MuRIL model to classify Indian citizen grievances across 14 government departments in Hindi, Hinglish, and English. The author honestly documents four bugs encountered—including a misleading 100% accuracy from train/test leakage, RNG state contamination across categories, and keyword bleed—achieving 87.3% accuracy on a rigorous holdout set. The article is more valuable for its debugging lessons than the happy path.

Benchmarks PyTorch 2.13's new MPS FlexAttention against standard SDPA on an M1 Max, finding a 7.83x speedup for sparse local-window attention at 32,768 tokens. For standard causal attention, SDPA remains ~19x faster. FlexAttention on Apple Silicon is promising for long, sparse patterns but not a universal win, and the API remains unstable in 2.13.

"Weights to Words" is a method that automatically discovers domain-relevant preference dimensions described in natural language and paired with model vectors, enabling users to inspect and edit preference model inferences. Validated across moral dilemmas, movies, wines, and LLM responses with two pre-registered experiments (N=450, N=449), it shows that regularizing toward learned dimensions and incorporating user edits both improve prediction accuracy. Participants preferred its inferred profiles and endorsed its predictions as more accurate in head-to-head comparisons.

PlanFlip introduces four planning-phase prompt injection attacks against multi-agent LLM systems that corrupt downstream sub-tasks via the Planner agent. Testing nine frontier LLMs across 3,479 episodes reveals that stronger models like GPT-5 are more vulnerable (ASR=0.68), homogeneous pipelines have a correlated-agent blind spot, and reasoning-augmented models like DeepSeek-R1 resist injections. The authors propose two defenses—GoalAnchorCheck and CrossAgentConsensus—achieving detection rates up to 1.00, concluding that heterogeneous model diversity is a security prerequisite.

The openFDA adverse event API hard-caps skip at 25,000, but the Link header's search_after cursor bypasses this limit entirely — measured at 31,968 records across 32 pages with zero duplicates. The critical trap: any non-zero skip silently degrades the cursor, so you must start with skip=0 or absent. Without an API key, limit maxes at 999; limit=1000 returns a misleading 403 API_KEY_MISSING error. The article includes production-ready JavaScript with retry logic, deduplication via safetyreportid+version+date composite keys, and a detailed error-code reference table.

Agent memory systems are uncurated datasets that agents write to themselves and then use as highest-authority context — functionally equivalent to fine-tuning on unreviewed self-generated text. Three critical failure modes emerge: consolidation manufactures false confidence by stripping hedges, repeated summarization causes semantic drift and self-reinforcing hallucinations, and the write path is a proven attack surface with AgentPoison achieving 80%+ success rates. Existing benchmarks like LoCoMo and LongMemEval only test the read path, leaving write-path vulnerabilities completely unmeasured.

Every MCP tool registered in Claude Code injects 80–150 tokens of tool-definition overhead into the context window on every turn, compounding across entire sessions. Measured results show the GitHub MCP server (26 tools) costs ~62,000 tokens per 20-turn session — roughly $0.19 at Sonnet 4 pricing — and scales to ~$18.60 in pure overhead for 2,000-turn autonomous loops. Practical mitigations include project-scoped MCP configs, preferring servers with fewer focused tools, and keeping tool descriptions tight when authoring MCP servers.

A postmortem of a silent data-loss bug in ZVec, an in-process vector database, where calling optimize() on a crash-recovered collection discarded WAL-replayed documents without any error signal. The root cause was a WAL handle lifecycle bug where segment recovery failed to reattach the handle, causing optimize() to skip persisting recovered data. The author argues that fixing individual bugs is insufficient and advocates for durable execution layers like Temporal where state ownership and recovery are explicit.

The author revisits a 2.5-year-old thesis that software development lacks engineering rigor, admitting the proposed cure (measurement) was wrong. The real solution turned out to be determinism: deriving architecture decisions through fixed procedures from stated requirements rather than subjective judgment. A procedure that takes SLOs and derives architecture through rules eliminates taste-based debates entirely.

Every MCP tool registered in Claude Code injects 80–150 tokens of tool definitions into the context window on every turn. The GitHub MCP server alone adds ~62,000 tokens of overhead per 20-turn session, costing ~$0.19 at Sonnet 4 pricing — and in autonomous loops of 2,000 turns, that balloons to ~$18.60 in pure overhead. The author recommends project-scoped MCP configs, fewer tools per server, and tight descriptions to minimize the per-turn tax.

A practical API guide for Kimi K3 covering reasoning_effort settings, thinking history preservation, tool calling with dynamic loading, structured output, context caching, prefix completion, and vision input. The model supports 1M-token context, always-on thinking, and up to 128 tools across Chat Completions, Responses, and Claude-compatible APIs. Includes verified code examples for production integration.

AI coding agents have made writing code cheap but reviewing it the new bottleneck—humans must either babysit every step or drown in unread output. The author compares this to CAD's transformation of engineering: a shift up in abstraction level that's painful mid-transition because the tooling bridge doesn't exist yet. The likely path forward isn't upfront specification but capturing the shape of work done once, so agents only pull humans in at genuine decision points.

A solo open-source maintainer discovered their data-loss bug fix sat in a green PR for four days because merging was only the first step of a five-step manual release ritual. The automated parts—CI tests, packaging checks—held perfectly, while every manual step (merge decision, changelog entry, release cutting) silently drifted. The insight: machine-watched processes are reliable; human-only processes decay silently and consistently.

The author replaced fragile prompt-chained agents with structured workflows using Pydantic schemas, validation guardrails, explicit state machines, and per-step evaluation, improving task success from 60% to 94%. Each step has typed inputs/outputs, blocking guardrails that auto-retry on failure, and an evaluation harness for granular debugging. The approach gives full visibility into which step fails and why, eliminating the cascading-failure problem of naive prompt chains.

A detailed technical guide to the openFDA enforcement API, covering four recall endpoints with different schemas and refresh schedules. Documents critical undocumented traps: unparenthesized OR queries silently discard clauses, skip-based pagination caps at 26,000 records (below total counts for device and food recalls), and generic error messages hide useful details. Includes concrete measured data from live API calls on 2026-07-20.