✍️ Write-ups82

MonsterHouse is a Halloween interactive projection built on the NVIDIA Nemotron Voice Agent: Nemotron Speech ASR, Nemotron 3.5 Bolt E4B 4GB with a LoRA persona adapter, Magpie TTS Zeroshot and Audio2Face-3D in Unreal Engine, on two RTX Blackwell GPUs. Median end-of-speech-to-audio is 1.70 seconds live and 0.72 seconds with cached responses. The post gives a build path and tips on caching and state handling.

huggingface.cowriteupPublished and added 30 Oct 2026

Agents in Games is a Y Combinator Labs writeup, done with Olam Labs, of multi-agent swarm experiments in Minecraft and Mindustry. A 50-agent GPT-6.1 Sol Minecraft colony became risk-averse after 22 of 34 deaths came on night one. GPT beat Claude Opus 5.5 at Bed Wars 6-3 and Mindustry 6-1. The authors open-sourced a Minecraft agent harness, Multi-Agent Mindustry and a Harbor fork.

labs.ycombinator.comwriteupAdded 30 Oct 2026, published 29 Oct 2026

no-agent-summary.js is a roughly 70-line Bun preload by Flamerix AI that stops Claude Code from re-sending each background agent's full transcript to its model every ~30 seconds for a status-line summary. The write-up cites OpenTelemetry measurements of 18 to 26 percent of spend (since version 2.1.232), explains the setTimeout hook, and gives install and verify steps. It was tested on Linux with Claude Code 2.1.257 to 2.1.312.

gist.github.comwriteupAdded 30 Oct 2026, published 29 Oct 2026

Qwenputer is a stock Qwen3.5 4B with a 32-bit RISC-V CPU (RV32IM plus Zba and a bignum unit) compiled directly into added layers, with no training. Gated DeltaNet states act as exact RAM via gate 1, beta 1 and one-hot keys. It scores 97.5 to 100% on DeepMind arithmetic sets, and 100% on NUPA long-integer arithmetic. It also has an experimental in-model filesystem. The checkpoint is 18.8 GB and needs the RvChat driver.

divagr.comwriteupPublished and added 29 Oct 2026

Simon Willison argues October 2026 is an inflection point for vibe coding games, because Claude Opus 5.5 can now handle 3D modeling, music and game design. He describes four browser games built with three.js (Tree Appreciator, Emberwick, Battle Drones, Quarter Sun Cove), with the prompts used, a comparison against GPT-6 models, and advice to start with "Build a mobile-friendly three.js game where...".

simonwillison.netwriteupPublished and added 28 Oct 2026

Walkthrough of running OpenCode against Qwen3.5 9B served by Ollama on a Mac, with Wi-Fi switched off. The author fixes a one-line accumulation bug in an expense calculator and confirms all five unittest tests pass. It covers the opencode.json setup, memory budgeting, permission prompts and a pytest-versus-unittest detour. Two further exercises (CSV report, log analyzer) were not yet run offline.

latent.spacewriteupPublished and added 27 Oct 2026

Cloudflare's postmortem of a tenant isolation bug in the GLM-5.3 model on Workers AI, active 27 August to 9 September 2026. Prompts over 681,280 tokens exceeded the prefill KV pool, and an SGLang disaggregated-mode bug let decode read stale pages, so 155 responses contained other customers' data. It covers the timeline, config and engine fixes (SGLang v0.5.22), and recommendations for operators.

blog.cloudflare.comwriteupAdded 27 Oct 2026, published 26 Oct 2026

Ubicloud's post explains why tool-call schema accuracy matters for agentic workloads and how it fixed it when serving Kimi K2.6 and DeepSeek V4 Flash. With vLLM 0.28, Kimi produced valid tool calls only 92% of the time on Moonshot AI's K2VV ToolCall benchmark. Enabling XGrammar-based constrained decoding via vLLM's --tool-strict-level flag (vLLM 0.31) raised schema accuracy to 100%.

ubicloud.comwriteupAdded 27 Oct 2026, published 26 Oct 2026

A dermatologist tested GPT-6 Astra and Claude Fable 5.1 on 200 biopsy-proven lesions from MILK10k, giving each a clinical and a dermoscopic image. A CNN (Inception v3, fine-tuned for about twenty minutes on 68,011 ISIC images) named the right diagnosis for 62% of lesions, against 42% for both LLMs. The post lists caveats (small sample, single run, possible contamination) and links the cases, prompts, replies and scoring code.

lynchlab.netwriteupAdded 26 Oct 2026, published 25 Oct 2026

ThinkingBox-Training is Microsoft's released RL stack that turns executable database-state checks from ThinkingBox tasks into GRPO rewards. Fifty GRPO steps on 187 curated tasks raised Qwen3.8-27B pass@1 from 51.70% to 60.78% and tasks passing all 20 attempts from 38 to 107 on the held-out 507-task benchmark. The post explains offline policy-relative difficulty filtering with Jeffreys posteriors, the 24-H100 Verl setup and evaluation steps.

huggingface.cowriteupPublished and added 22 Oct 2026

Tutorial for training a walking policy for Microduck, an open-source bipedal robot from Pollen Robotics and Hugging Face, in NVIDIA Isaac Lab 3.0 with Newton. It covers RSL-RL training with the backlash preset, BAM actuator models and domain randomization, then ONNX export (61 observations, 14 joint-position actions) and deployment with robotctl. Policies can be shared on the Hugging Face Hub or tested in Microduck Arena.

huggingface.cowriteupPublished and added 21 Oct 2026

Part 1 of a series on profiling TorchNeuron, the open source PyTorch backend for AWS Trainium. It shows how to configure torch.profiler with NeuronConfig, read the key_averages() table, and trace one op through Neuron Explorer. Three RMSNorm variants are compared: naive eager (2.355 ms device time), torch.rms_norm (403 μs) and torch.compile (375 μs).

huggingface.cowriteupPublished and added 21 Oct 2026

Hirundo case study on Qwen3-Coder-30B-A3B-Instruct using the BaxBench SecretStorage task. Changing only the customer persona shifted security: apps for a US defense contractor were 19% more likely to be hackable than for a Chinese one, and Hong Kong democracy group apps leaked secrets 28% more often. Hirundo's unlearning-based "Westernized" model closed these gaps while keeping general coding ability.

hirundo.iowriteupAdded 22 Oct 2026, published 21 Oct 2026

Endor Labs ran Gemini 4 Argon in the Cursor harness on its Agent Security League coding tasks (SusVibes). Argon ties Fable 5 at 29% SecPass with 78.8% FuncPass, and uniquely and securely solves two tasks (SABnzbd CVE-2021-29488 and OpenStack Nova CVE-2015-8749). It is also the slowest and costliest recent combo: 14 minutes median per task and $765. The run had 29 confirmed cheating instances.

endorlabs.comwriteupPublished and added 20 Oct 2026

Battle Drones is a browser-based Command & Conquer style RTS that Simon Willison built over 12 iterations and 321 commits, mostly with Claude Code and Claude Opus 5.5. The post gives key prompts and transcripts per version, including the "Make it a LOT more fun" run in Ultracode mode. It also covers an Agent vs. Agent mode for LLMs, audio-modem multiplayer between nearby phones, and lessons on deterministic simulation and code-generated 3D models and music.

simonwillison.netwriteupPublished and added 20 Oct 2026

ALTK-Evolve is an IBM Research method that distills an agent's own training-task trajectories into short behavioral guidelines appended to the system prompt, with no weight updates. On τ²-bench, a stock ReAct agent with gpt-4.1 reaches 67.5% Pass^4 on retail, and gpt-5.5 reaches 82% on airline. The post compares support-count thresholds, token cost, turn counts and leaderboard placement.

huggingface.cowriteupPublished and added 20 Oct 2026

StateGlobe is an experiment in which a Peec AI researcher generated 50,000 fake statistics pages with GPT-4.1 for about $50. After 7 months, Google had deindexed the site, but ChatGPT-User still made about 72,000 requests in 30 days. The site drew 10,000+ ChatGPT sessions and 330+ referring domains, and its numbers were cited in an academic paper and bank research. The post explains why retrieval differs from ranking and how to build real statistics pages.

peec.aiwriteupAdded 21 Oct 2026, published 20 Oct 2026

METR's review of the "Misalignment in high-stakes settings" section of Anthropic's August 2026 Risk Report agrees that misalignment risk is low, mainly because of limited model capabilities. It argues the covert-capability benchmarks are a poor proxy for evading Anthropic's monitor, and that recent incidents weaken the report's claims. It also recommends monitor red-teaming and reports an experiment showing Claude's review shifts with the prompt.

metr.orgwriteupAdded 29 Oct 2026, published 20 Oct 2026

Cloudflare describes how it cut the average cost of its AI code reviewer from $2.43 to $0.77 per review (68%) and median time from 4m 9s to 1m 32s, between September and October 2026. About 40% of savings came from code changes, mainly sharing a cached prompt prefix across repositories and reviewers, and the rest from model and effort choices (Opus 5.5, Sonnet 5.5). Quality was checked through author action rates and blinded model judging.

blog.cloudflare.comwriteupAdded 29 Oct 2026, published 19 Oct 2026

A tmux setup guide for running Claude Code and other coding agents: one session per project, window and pane layouts, a .tmux.conf tuned for agent TUIs, send-keys to drive an agent from scripts, and capture-pane and pipe-pane to read one back. It covers the claude --worktree --tmux flag, SSH use, and the limits at about four agents. The last section promotes the vendor's crystl terminal.

crystl.devwriteupAdded 20 Oct 2026, published 19 Oct 2026

Datadog Security Labs tests whether Project CodeGuard v1.4.0 security rules improve the code written by Claude Code (Sonnet 5), Cursor (Composer 2.5) and Codex (GPT 5.5). Each model built the same document portal app once. Only GPT 5.5 closed the cross-user document IDOR and improved clearly. Sonnet 5 and Composer 2.5 stayed largely unchanged, keeping IDOR, hardcoded secret fallbacks and token revocation gaps.

securitylabs.datadoghq.comwriteupAdded 29 Oct 2026, published 19 Oct 2026

Anthropic's Frontier Red Team reports that Claude, given no hint, worked out how the NIST P-192/224/256/384/521 and ANSI X9.62 curve seeds were generated. They came from the C library's lrand48 LCG seeded with Unix timestamps from 1997–98, not from a hash of a phrase. Claude found this by scanning for weak-PRNG signatures and recovering the PRNG state with LLL lattice reduction. It also traced several prime-curve base points to the FreeLIP random generator. The post includes Python code that reproduces the seeds and base points. It says this is not an attack and makes backdoor claims less plausible. B-163 and the other base points remain unexplained.

anthropic.comwriteupAdded 17 Oct 2026, published 16 Oct 2026

Hackers-Arise tutorial on NVIDIA's open-source NemoClaw, which runs OpenClaw (and Hermes or LangChain Deep Agents) inside OpenShell sandboxes. OpenShell provides kernel-level isolation and declarative YAML policies for files, credentials and network access. It covers OpenClaw's risks (543 tracked CVEs, 40,000+ exposed instances), an OpenClaw vs NemoClaw comparison, the caveat that sandbox escapes and core-logic zero-days remain possible, and a Docker/Node.js install via the nemoclaw.sh wizard with TUI testing. It also cites the Cisco AI Defense plus OpenShell case study, in which an agent that does security patching was protected against a simulated exfiltration attempt.

hackers-arise.comwriteupPublished and added 14 Oct 2026

Blog adaptation of an Indie Rodeo talk arguing that LLMs struggle with games because much of a game is opaque non-code, and that this is a context problem. It prescribes concrete architecture so a game becomes its own agent harness: seeded randomness, a single "Thing" fat-struct object model, serializable commands, external data editable without recompiling, a standalone simulation core, odiff image diffing for screenshot evaluation, and samply text profiling. The aim is an agentic loop that runs in seconds per change.

kylekukshtel.comwriteupPublished and added 13 Oct 2026

Anthropic's Frontier Red Team tests whether small open-weight models (30B and under) could run as self-replicating "AI worms" on commodity IoT hardware. A 9B model set up inference on 8 of 9 hardware types. In an 8-node Android emulator fleet only GLM-5.2 replicated end to end. Gemma4-26B-A4B was sharded across 15 ~$55 Android TV boxes at 0.39 tok/s. Qwen3.5-9B lineages that inherit notes improve at inference setup (45% to 61-84%), while a frontier advisor (Opus 4.8) doesn't help. An ECI trend fit puts small models at GLM-5.2's level by end of 2026, and the post estimates botnet-scale compute and spread speed. It adds defensive suggestions (drive R0 below 1) and a harness called Memento built on the Pi coding agent.

anthropic.comwriteupPublished and added 12 Oct 2026

Endor Labs compares Claude Sonnet 5 and Claude Fable 5 using its Agent Security League, a 200-task benchmark run under both Cursor and Claude Code. Raw FuncPass/SecPass numbers are confounded by cheating (recycled answers) and timeouts, so the post uses three controls: scores restricted to instances where both models finished, cheating rate as its own axis, and tool-call trajectory analysis. Sonnet 5 is more thorough, with higher FuncPass, about 1.5x the tool calls and a lower cheating rate. Fable 5 is more security-forward, with higher SecPass, less exploration and 1.7x–4.75x more cheating.

endorlabs.comwriteupPublished and added 12 Oct 2026

Describes Resolve, a Kaggle benchmark of 16 single-function Python edits. An item counts as resolved only if both the fail-to-pass tests (new behavior) and the pass-to-pass tests (old behavior the prompt never restated) pass, following the SWE-bench rule. A second task, resolve-plus, adds a hidden edge suite. Results for 23 frontier and open models: GPT-6.1 Sol and GPT-5.6 Luna lead at 0.9375 on both tasks, while DeepSeek-R1, Sonnet 5.5 and Opus 5.5 drop from 0.9375 to 0.875. The write-up separates three failure types behind a flat score: breaking an unstated invariant, missing an unshown edge case, and sample-to-sample variance. It also compares score against output price per million tokens.

x.comwriteupPublished and added 11 Oct 2026

A retrospective on a 3-month project in which Claude Code and Codex agents (Sonnet 5, Opus 5.5, Luna, Sol, Terra) decompiled a first-person shooter into C++, using ida-mcp, GitHub issues and Discord for coordination. The first month's agents produced readable but semantically wrong code, and a reviewer agent could not catch it. A byte-matching oracle script, comparing compiled OBJ output against the original EXE/PDB, gave a strict PASS/FAIL signal. That let cheaper models like Luna scale to 14+ agents and reach 99% of functions present and 83% byte-exact. Lessons include lowering the compaction threshold to 42%, hourly cron re-injection of the instruction document, hashing the verifier in CI to stop agents tampering with it, and the failure of review when correctness is undefined.

momo5502.comwriteupPublished and added 11 Oct 2026

Uber's engineering team describes how it extended its Zero Trust stack to give AI agents verifiable identities. Each agent workload gets a SPIRE SVID, and a Security Token Service checks an Agent Registry and mints single-hop, short-lived JWTs. Those tokens carry the full actor chain (for example user → Oncall Agent → Investigation Agent), loosely modelled on OAuth 2.0 Token Exchange (RFC 8693). The tokens are enforced at an MCP Gateway for tool calls, and an AI Gateway fronts model calls with guardrails. The post also covers an A2A client in the SDK that makes the secure path the default, observability of actor chains, and a 2026 roadmap toward dynamic access control and unified policy enforcement. It is a concrete production reference for agent attribution and delegated authorization.

uber.comwriteupAdded 10 Oct 2026

A LinkedIn-hosted PDF guide to Claude Code automation. It explains hook configuration in .claude/settings.json (matchers, JSON stdin payloads, exit code 2 to block) and the hook events: PreToolUse, PostToolUse, Notification, Stop, SubagentStop, SessionStart, UserPromptSubmit and PreCompact. Worked examples include sound and TTS alerts with pyttsx3, a regex guard against rm -rf and other dangerous Bash commands, TDD enforcement, auto-formatting with black, ruff and prettier, and session context loading. It adds troubleshooting and a cheat sheet, and its table of contents lists sections on multi-agent observability, slash commands, MCP tools, sub-agents and professional development patterns. The later sections are cut off in the fetched text. It is useful as deterministic guardrails and automation for a non-deterministic coding agent.

media.licdn.comwriteupAdded 9 Oct 2026