- 01 ai
How to Read a Viral AI Doom Take Without Being Captured
A four-question filter for evaluating viral AI risk claims: specificity, incentive, unfalsifiability, and the missing P(boom). Operator framework, not x-risk adjudication.
- 02 ai
Opportunity AI vs Efficiency AI: A Two-Lane Operator Framework
Why 'is this model better?' is the wrong question when the model expands what you can attempt. A two-lane scorecard and a six-step opportunity audit for evaluating capability-expanding AI work.
- 03 ai
Agent Containment: the Boundary That Matters After a Bad Agent Decision
A practical containment checklist for limiting what an AI agent can read, change, reach, and send when prompts and approvals are not enough.
- 04 ai
Verifiable Finish Lines: an Agent Stopping Is Not Proof the Job Is Done
A practical completion contract for AI-agent work: define the artifact, checks, evidence, bounds, and terminal states before the run starts.
- 05 ai
Personal Model Benchmarks: Stop Chasing a Winner and Assign Each Model a Job
Build a small, dated suite from your recurring work so each AI configuration earns a task—not a universal crown.
- 06 ai
AI Operator Shift: Pick the Computer Before You Delegate the Work
AIDB's operator-shift frame turns into a practical choice: use a sandbox, cloud workspace, persistent computer, or your own browser.
- 07 ai
When Labs Control the Model and Harness: Keep Your AI Exit Route
A five-move operator playbook for keeping models, orchestration, and evals portable when vendor terms or access change.
- 08 ai
Multiplayer AI Agents: Shared Sessions Need Real Handoffs
A 10-minute pilot for shared agent sessions: separate visibility from control, make handoffs explicit, and keep authority outside the model.
- 09 ai
GPT-6 Astra Just Launched — and You Still Can't Use It
GPT-6 Astra launched behind the Trusted Access Program and Daybreak cyber defenders; API and ChatGPT plans follow within days. What day one unlocks, and what the benchmarks do not prove.
- 10 ai
Muse Spark 1.3: Cheap Contributor Tokens Are Not a Private-Data Tier
Meta's Muse Spark 1.3 ships two tiers with the same model and different data terms. How to pick Standard or Contributor without leaking work you can't get back.
- 11 computers
What Is Tailcat? The Encrypted Netcat That Trades a Tailnet for a Token
Tailcat is Tailscale's account-free encrypted netcat. Understand the token trade-off, direct-versus-relay path, and why it is not a tailnet.
- 12 ai
Claude Fable 5.1: Pay for the Long Jobs, Not Every Prompt
A practical Fable 5.1 guide for long-running agents: model ID, cache-aware costs, fallback handling, and preserved-thinking migration checks.
- 13 ai
Claude Mythos 5.1: Access Is the Product Constraint
Mythos 5.1 is the same underlying model as Fable 5.1 with different safeguards. It is limited to approved Project Glasswing customers.
- 14 ai
Tests Passed, But the Fix Is Not Live: Verify the Running Agent Service
A green test suite proves an artifact, not a running workload. A platform-neutral verification recipe with five independent checks before you declare a fix live.
- 15 ai
Rollback Before You Deploy: Name the Trigger, Artifact, and Decision Maker
Write the recovery plan before release so a failure does not turn into an hour of guessing which artifact to restore and who can decide.
- 16 ai
The Agent Said It Failed. Check the Work Before You Retry.
A coding agent reports a failure. Before you click retry, inspect the tree, the run log, and the prior commit. A practical inspection playbook for coding agents and scheduled jobs.
- 17 ai
Make Long Agent Jobs Resumable: Checkpoints Instead of Restarts.
Design long agent jobs so they resume from the last completed step instead of redoing everything; separate retry from durable recovery and protect each external side effect with idempotency.
- 18 ai
Silent on Success, Loud on Failure: Alerts Operators Will Not Ignore.
Most operator alert channels fail in one of two ways: they scream on every healthy run, or they go quiet the moment things actually break. Five tips to make alerts trustworthy at 3 a.m.
- 19 ai
Split Your Monitor From Your Worker: Reliable Scheduled Agent Jobs.
Treat the scheduler as a thin trigger and the worker as the owner of an idempotent unit of work. A platform-neutral playbook for reliable scheduled agent jobs.
- 20 ai
Hermes Bot Mode: Keep Specialist Agents in Their Own Lanes
Give Hermes specialists distinct profiles, scoped groups, human-led approvals, and clear limits for routines, model routing, and cross-machine messaging.
- 21 ai
MiniMax H3 Max by fal: A Video Is Not Finished When the API Says It Is
Run paid H3 Max video jobs through fal with a spend cap, retained request IDs, verified MP4 storage, and separate human approval.
- 22 ai
Darkbloom: A Private Inference Network, Not a Model, and It Is Still Alpha
Darkbloom is a private-inference network, not a model — what it routes, what its encryption actually protects, and the alpha caveats before you send it data.
- 23 ai
Darkbloom Setup: The Two-Sided Alpha Risks Before You Join
Both sides of the Darkbloom public-alpha network: developer API setup (consumer) and Apple Silicon Mac provider setup, with hardware, install, and earnings caveats.
- 24 ai
GLM-5.3-Flash on DGX Spark: The Verified 4x GB300 Recipe and the Single-GB10 Reality
The verified GLM-5.3-Flash recipe is SGLang on 4x GB300 (FP8 weights, FP8 KV, TRT-LLM DSA, NEXTN MTP): 1,870 tok/s aggregate. On a single DGX Spark expect far less. Weights are ~306 GiB.
- 25 ai
GLM-5.3-Flash on Mac: Do Not Mistake Hosted Access for Local AI
GLM-5.3-Flash (320B/18B MoE) has no Apple Silicon path today. What works: Ollama :cloud (hosted, not local), the Z.ai API, and watching oMLX. Includes the thinking-mode migration trap.
- 26 ai
GLM-5.3-Flash on RTX 3090: When Local Inference Is the Wrong Call
The only documented local path for GLM-5.3-Flash on a 24 GB card is KTransformers heterogeneous offload, and Ampere is outside its validated list. Realistic ceiling: single-digit tok/s.
- 27 ai
Qwen3.8-Flash-Next on DGX Spark: Fast Only If the Kernels Load
NVFP4 weights plus MTP speculative decoding turn the DGX Spark into the strongest local path for Qwen3.8-Flash-Next, if the b12x kernels are actually loaded.
- 28 ai
Qwen3.8 Flash-Next on Apple Silicon: 4-bit Needs About 110GB of a 128GB Mac
Run Alibaba's 125B/6B MoE on Apple Silicon with Unsloth Dynamic 3.0 GGUFs, llama.cpp Metal, and a raised iogpu.wired_limit_mb — 4-bit at ~110 GB on 128 GB.
- 29 ai
Qwen3.8-Flash-Next on RTX 3090: Why the Smaller 27B Is Faster
Run Qwen3.8-Flash-Next MoE on one RTX 3090 (24 GB + 96 GB RAM): UD-IQ1_S GGUF at 72.5 GB, hybrid GPU/CPU offload, ~3-8 tok/s. Dense 27B still wins on 24 GB cards.
- 30 ai
What Is Cursor Origin? A Beta Git Forge, What It Lacks, and the GitHub Coexistence Tradeoff
What Cursor's git forge shipped in beta, what it lacks, and how GitHub coexistence actually works before you claim a namespace.
- 31 ai
Cursor Origin Setup: Claim the Namespace Before You Create the Repo
Claim the permanent namespace safely, install origin CLI, create or mirror repos, and run your first PR loop on Cursor Origin.
- 32 ai
GMI Cloud Setup: Choose Serverless or Dedicated Before Your First GPU Bill
First OpenAI-compatible call on GMI Cloud serverless vs Dedicated decision table and the GPU Compute path with cost checks.
- 33 ai
GMI Cloud: What You Are Actually Renting—and What to Check First
GMI Cloud's four products date-stamped pricing and which vendor claims deserve skepticism before you rent its GPUs.
- 34 ai
Zide Setup: From First Install to a Real Issue-to-PR Loop
Per-OS install, git-host connection, model setup, and the first agent-assisted issue-to-PR loop in Zide.
- 35 ai
What Is Zide? The Agentic Developer Desktop, the Plan Gates, and the Vendor Claims to Verify
What Zide is, what it actually gates behind each plan, and which vendor claims to check before you install it.
- 36 ai
Ox Alpha: The Free 1M-Context Model With a Data-Terms Catch
A free 1M-context multimodal frontier model appeared on OpenRouter. How to wire it up this week — and the data-terms contradiction you must understand first.
- 37 ai
Who Made Ox Alpha? The Fingerprinting Evidence, the Alternatives, and What Would Disprove It
Nobody has claimed the stealth model. The fingerprinting evidence strongly favors one lab — here's the weighted case, the alternatives, and what would prove it wrong.
- 38 ai
The Capability-to-Context Shift
Frontier models got good enough that the binding constraint moved. The operators winning in 2026–2027 are the ones closing context gaps, not chasing capability.
- 39 ai
The AI Engineering Skills Map for Knowledge Workers
Andrew Ng mapped four AI skills for developers. Nathaniel Whittemore extended the map to five skills for everyone else — and put domain judgment underneath all of them.
- 40 ai
The AI Delivery Gap: Why Curing Cancer Won't Fix the Trust Problem
Anthropic's Dario Amodei conceded the AI industry hasn't delivered on its biggest promises. OpenAI's counter — the pharma parallel — is sharper than it first sounds.
- 41 ai
Should I Buy a DGX Spark?
A buyer's-decision guide: who the $4,699 NVIDIA DGX Spark fits, who should buy a Mac Studio or Strix Halo instead, and the 2.7 tok/s reality behind the 1 PFLOP marketing.
- 42 ai
Frontier-Model Stack-Fit: Why "Best Model" Is the Wrong Question in 2026
In late 2026 the right question for picking an AI model is where each fits in your stack, not which scores highest on the leaderboard.
- 43 ai
AI Deputization Audit
Five-criterion rubric (worth-it, teachability, checkability, stakes, integral-to-process; 0-10) for deciding which work AI should deputize, duet on, or defend.
- 44 ai
Stop Counting Tokens: The 4-Question Scorecard for AI
OpenAI's CFO proposed replacing 'seats sold' and 'tokens consumed' with a 4-question scorecard.
- 45 ai
Qwen3.8-27B on DGX Spark (GB10): vLLM 0.24 + NVFP4, the sm_121 kernel gotcha
Set up Qwen3.8-27B on DGX Spark (GB10). vLLM 0.24+ with ModelOpt NVFP4 + MTP n=3 is the stack. Qwen3.8 NVFP4 checkpoint not yet published; Qwen3.6 reference at 144 tok/s at concurrency 16.
- 46 ai
Qwen3.8-27B on a Mac: Ollama MLX, MTPLX, and the MTP path
Set up Qwen3.8-27B on Apple Silicon: the Ollama MLX engine, MTPLX with native Qwen MTP heads, and the NVFP4 MLX quant per chip. No Qwen3.8-27B Mac benchmark yet.
- 47 ai
Qwen3.8-27B on a single RTX 3090: MTP via llama.cpp-from-source or vLLM
Run Qwen3.8-27B on one RTX 3090 (24 GB): MTP via llama.cpp-from-source or vLLM, Q4_K_M quant. No Qwen3.8-27B benchmark yet; Qwen3.6-27B reference data cited in body.
- 48 ai
AI Bias: The Concern With the Narrowest Gap
64% of the public and 73% of experts worry about AI bias — a 9-point gap. But agreement on the concern hasn't produced agreement on the fix.
- 49 ai
AI and Kids: Where 89% Agreement Meets Real Incidents
89% of adults worry about kids' data privacy, AI companion incidents are documented, and the FTC's amended COPPA rule now names AI explicitly. The rules changed in October.
- 50 ai
AI's Energy Problem Is Now a Permitting Problem
61% of Americans worry about AI's electricity appetite. 70% oppose local data centers. And U.S. data centers may draw 6–12% of national electricity by 2028.
- 51 ai
AI Companionship and the Human Connection Worry
Two-thirds of Americans worry people will trade relationships for AI companions. It's the rare concern where the public and experts nearly agree — and usage is already real.
- 52 ai
Deepfakes and Election Integrity: AI's Most Concrete 2026 Worry
66% of the public and 70% of AI experts worry about AI misinformation — the rare full-alignment concern, and 2026 is its first big election year.
- 53 ai
AI Privacy and Data Misuse: The Concern Everyone Agrees On
84% of Europeans, 60% of AI experts, 82% globally — privacy is the rare AI concern where the public, experts, and regulators all align. The EU already legislated accordingly.
- 54 ai
The 50-Point Gap: AI Job Displacement and the Expert Credibility Problem
73% of AI experts expect positive job impact. 23% of the public agrees. That 50-point gap is the largest in AI surveying — and it explains every tense all-hands about AI.
- 55 ai
Who Regulates AI? Americans Don't Trust Their Own Answer
The U.S. has the lowest trust in its own AI regulator of any country surveyed — 31% vs. a 54% global average. Meanwhile the public worries regulation won't go far enough.
- 56 ai
The 8 Concerns Americans Have About AI: Ranked and Weighted
Job displacement, misinformation, privacy, and the rest — what the surveys actually stack up to when you put them on one page, and where the public-expert gaps widen or close.
- 57 ai
The Five Engineering Disciplines Behind Every AI Agent
The five engineering disciplines behind every AI agent, and when to reach for each.
- 58 ai
Context Engineering
Context engineering: curating exactly what an AI agent sees, the discipline named in mid-2025.
- 59 ai
Graph Engineering
Graph engineering: wiring many agents into one team, and the origin credit it actually deserves.
- 60 ai
Harness Engineering
Harness engineering: the invisible scaffolding built around a model that makes an agent work.
- 61 ai
Loop Engineering
Loop engineering: designing the observe-plan-act-check-repeat cycle an AI agent runs.
- 62 ai
Claude Opus 5
Claude Opus 5 (Anthropic, July 24, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 63 ai
GPT-5.6 Cyber
GPT-5.6 Cyber (OpenAI, August 10, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 64 ai
Grok 4.6
Grok 4.6 (xAI, August 12, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 65 ai
Kimi K3
Kimi K3 (Moonshot AI, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 66 ai
Ling 3.0 Flash
Ling 3.0 Flash (InclusionAI, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 67 ai
LFM2.5-VL-3B
LFM2.5-VL-3B (Liquid AI, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 68 ai
Muse Glimmer-30B
Muse Glimmer-30B (Meta / MSL, August 9, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 69 ai
Muse Spark 1.2
Muse Spark 1.2 (Meta / MSL, August 5, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 70 ai
Nemotron 3.5 Lightning
Nemotron 3.5 Lightning (NVIDIA, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 71 ai
Qwen Image 3.0
Qwen Image 3.0 (Alibaba / Qwen, July 21, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 72 ai
Qwen3.7 Flash
Qwen3.7 Flash (Alibaba / Qwen, July 27, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 73 ai
Qwen3.8-2.4T-A95B
Qwen3.8-2.4T-A95B (Alibaba / Qwen, August 11-12, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 74 ai
Solar Pro 4
Solar Pro 4 (Upstage, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 75 ai
Qwen3.8-Max Weights: Out
Qwen3.8-Max Weights: Out (Alibaba / Qwen, August 11-12, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.
- 76 ai
Prompt Engineering
What prompt engineering is in 2026, how to do it, and when it stops being enough.
- 77 ai
The Self-Driving Company
Replit's CEO coined the term for an organization where people set the destination and agents do the driving. Their field report has real numbers — and a hard prerequisite most enterprises skip.
- 78 ai
The AI Trust Deficit: Why Nobody Believes the Boosters
Zuckerberg's 6,500-word AI manifesto was meant to reassure. Instead it demonstrated exactly why the public doesn't trust tech executives — and the numbers behind that distrust are hardening.
- 79 computers
Cheapest Way to 128GB of Local AI Memory in 2026 (the Chart That Actually Matters)
The viral 128GB-local-AI chart gets prices right but conflates capacity with usefulness. Here is the chart that actually matters: bandwidth, interconnect, and real tok/s.
- 80 ai
MiniMax H3 Hardware Map: What Runs, What Doesn't, and Why the Answer Is Probably Your Card
MiniMax H3 is a 33B video model with native stereo audio. This operator's hardware map covers what runs it, what doesn't, the territory restriction, and the right rig for you.
- 81 ai
How I Ran H3 on a 48 GB Mac Mini (and How You Can on Whatever You Have)
An honest failure log: 5 broken local H3 generations, a Mac freeze, and 6 hours lost before pivoting to the Hailuo API. The mistakes, the receipts, and what I'd do differently.
- 82 computers
Your Agent Reviewing Its Own Work Is Not a Check
Why a same-model self-review is not an independent safety gate. Schema proves shape; intent and permission need a different layer. With a bounded retry pattern and the right list of real checks.
- 83 computers
Hermes vs OpenClaw: Pick a Path (Not a Fandom)
A fair, source-cited side-by-side of Hermes Agent and OpenClaw. Where each is strong, where each is wrong-fit, and how to choose honestly. Not a verdict.
- 84 computers
Start Here: Local AI and Agents Without Wasting a Weekend
A working first-night path for ABS operators: pick API or local, set up Hermes, cap cost, add one log line, then stop. Verifiable against our live guides and the OpenClaw comparison.
- 85 ai
Standard Compute review: unlimited flat-rate LLM API for Hermes Agent (setup + plans)
Flat-rate unlimited LLM API wired into Hermes Agent via a custom OpenAI-compatible endpoint. Plans, fair-use pacing, setup.
- 86 ai
Featherless for Hermes Agent: Plans, Compatibility, and Setup
Featherless Chat (32K) sits below Hermes Agent's 64K minimum. Use the Developer or token-based plan (256K) for self-hosted Hermes, or the managed Featherless Hermes — with exact setup for both.
- 87 ai
OpenCode Go: $5 First Month, $10 After, and Hermes Agent Setup
A current guide to OpenCode Go: real 5-hour / weekly / monthly dollar caps, 18 open models, top-up behavior, then exact Hermes Agent setup.
- 88 ai
QwenCloud Token Plan: Lite vs Standard vs Pro, and Hermes Agent Setup
QwenCloud Token Plan Individual compared ($6/$18/$68 limited-time): 5-hour and 7-day Credit windows, then exact Hermes Agent setup with the dedicated Token Plan API key and base URL.
- 89 ai
Nous Portal Subscription: Plans, Credits, and Hermes Agent Setup
Free, Plus, Super, and Ultra tiers on Nous Portal, plus the one-command Hermes Agent setup and how to verify every tool routes through your subscription.
- 90 ai
Hermes Agent Provider: MiniMax M3 (Direct)
Wire MiniMax M3 directly to Hermes Agent as the primary model: the exact provider id, model string, environment variable, setup steps, and smoke test.
- 91 ai
MiniMax Token Plan: Plans, Limits, and Hermes Agent Setup
A current guide to MiniMax Plus, Max, and Ultra: real quotas and caveats, then exact setup and testing with Hermes Agent.
- 92 policy
Laws as Experiments: A Political System Built on the Scientific Method
What if every law had to specify what it's trying to do, offered states a range of options to test, expired on a timer, and was evaluated by the outcomes — like a real experiment?
- 93 ai
Cost-aware model routing for agents
A three-tier routing decision tree, a 15-line router, an escalation pattern with retry budget, and the cost math showing 70/20/10 split saves ~88% vs all-large.
- 94 ai
From JSONL to dashboard with zero infra
Four stages with the exact tool for each: grep + awk, jq + cron, DuckDB, Metabase. Plus the graduation rule for when to upgrade and when to stop.
- 95 computers
GitHub Actions CI for Astro: lint, build, and deploy on every push
Add a three-job GitHub Actions workflow that checks an Astro site, builds it, and deploys to GitHub Pages whenever main changes.
- 96 ai
How to find, download, and evaluate local models on Hermes Agent or OpenClaw
A repeatable procedure for finding a model that fits your hardware, pulling it down with Ollama, and evaluating it against your own tasks on Hermes or OpenClaw.
- 97 ai
The 1-line observability hook that fits any agent
A 30-line decorator that wraps every model call, writes JSONL, and answers the four questions you actually ask. The smallest hook that is useful; smaller and you fly blind.
- 98 ai
The four tiers of agent memory (overview)
A shared vocabulary for memory and a decision rule for which tier to write to at any moment. Start here, then pick a tier to go deeper on.
- 99 ai
The three-phase context loop: grounding, messy middle, landing
A working frame for thinking about agent prompts as three deliveries, not one — and a checklist for what context belongs in each phase.
- 100 ai
Tier 1 working memory: prompt hygiene beats size
Five rules for keeping the model's context window lean — recency is not enough, compress don't copy, de-duplicate by reference, strip the system preamble, drop successful retries.
- 101 ai
Tier 2 session memory: the run log you actually use
A schema for one-line-per-step JSONL run logs, a directory layout with an index, three small tools that turn the log into something you will actually open, and a what-to-log / what-not-to-log list.
- 102 ai
Tier 3 project memory: the file-based second brain
Four files that an agent reads at session start — MEMORY, USER, DECISIONS, RUNBOOK — with sharp rules on what belongs in each, when to graduate to a vector DB, and what NOT to put in project memory.
- 103 ai
Tier 4 long-term user memory: don't be creepy, be useful
The rule for what to remember about an operator, the audit pattern, the trimming cadence, the four failure modes, and the legal floor under GDPR right-to-erasure.
- 104 ai
Why every agent needs a cost cap on day one
Three numbers, one enforcement point, and a hard ceiling before the bill lands. The cheapest pattern that separates an agent you run from one you babysit.
- 105 ai
Your first local-first agent on a Mac mini in 90 minutes
Pick an agent and a local model, then walk through 90 minutes yourself or ask the agent to do most of it. Hermes Desktop path + Claude Code + Ollama wiring.
- 106 ai
AI Coding Is a Nightmare. You're Not the Only One — and There's a Pattern.
The seven failure modes HN developers are reporting, what the Anthropic 2026 RCT found about AI-assisted coding, and seven interaction patterns that produce better code.
- 107 ai
Different Ways of Using LLMs for Coding: Eight Patterns Beyond the Prompt-Response Loop
A field guide to the alternative coding-agent patterns developers are actually using in 2026 — hermetic agents, tab-model editors, literate programming, workboxes, containerized sessions, and more.
- 108 ai
SEO for AI Answer Engines in 2026: The robots.txt, llms.txt, and Content Decision
How to be cited by ChatGPT, Perplexity, Claude, and Google's AI Overviews — the Search/Agent/Training bot taxonomy, llms.txt, and a per-goal decision matrix.
- 109 computers
Turn Off Ring Video Descriptions: The Fast Way (and What You're Actually Turning Off)
A 30-second path to disable Ring's AI-generated Video Descriptions across every camera, plus the per-device path, what the feature does, and who can't see the toggle.
- 110 ai
Anthropic's Doom Ad and the Messenger Paradox
Anthropic bet its brand on a 90-second film of burning houses and cemetery rows. The internet called it the best anti-AI ad ever made. What the campaign actually teaches.
- 111 computers
ABS Search Bar: The Cheap State Update, Not a Popup
A single client-side script, four pages, no backend. The ABS search bar filters a JSON index already on the page — sub-50ms on every keystroke, no popup.
- 112 ai
AI Model Matrix Quickstart: A Five-Axis, Three-Filter Way to Pick a Model
Use the ABS Model Matrix to pick an LLM in under two minutes: five axes, three filters, the cheapest-viable heuristic, and a worked Sonnet 4.5 / GPT-5 / M3 / Gemini 2.5 Pro example.
- 113 computers
Cron Job Prompts That Stay Self-Contained: The Discipline
A Hermes cron runs in a fresh session with zero memory. Six elements make the prompt self-contained; three anti-patterns break it. Skip any and the job runs but does nothing.
- 114 computers
Daily Memory Audit and the 4-Step Merge Plan: From Cron Flag to Clean Memory
Operator playbook for the nightly memory-audit cron: the 4-step merge plan (read / propose / dedupe / commit), the 3 categories of stale memory, and the 5 traps that look mergeable but are not.
- 115 ai
Gemini 2.5 Flash vs 2.5 Pro vs 1.5 Pro: Which One Should You Use?
Gemini 2.5 Flash is the cheap workhorse, 2.5 Pro is the reasoning model, and 1.5 Pro is now legacy. Here is the practical capability and cost ladder.
- 116 ai
GPT-5 vs GPT-5 Mini: When Each Makes Sense
A practical decision rubric for the OpenAI flagship tier vs the mini tier: per-token cost, the six capability axes, three tasks where mini is wrong, and three where it is right.
- 117 ai
Hermes Cron Job Authoring: The Shape and the Rules
A practical Hermes cron playbook: jobs.json fields, supported schedule shapes, a daily-traffic example, and four checks that catch bad jobs before they run.
- 118 computers
Hermes Curator: The Weekly Lint Gate for Skills and Memory
Operator playbook for hermes curator as the weekly lint gate: the 13-verb CLI surface, the 4 lint rules (orphan / stale / oversized / duplicate), and the pin-vs-prune decision.
- 119 computers
Hermes Memory Trim: A Proposal-Only Workflow
Operator playbook for the weekly memory trim: the proposal-only contract (never mutate), the 4 verdicts and 5 trim candidates, the diff-and-confirm UX, and the 3 verification checks.
- 120 ai
Hermes Memory: What to Save and What Not To
Eight-category rule for Hermes MEMORY.md and USER.md: four always-save kinds, four never-save kinds, the 2k-char ceiling, and the Johnny5 routing gate.
- 121 computers
How to Read an LLM Pricing Table Without Fooling Yourself
A practical guide to the LLM pricing table: compare input, output, and context window first, then treat cached, batch, and free-tier prices with care.
- 122 ai
Local LLM with Ollama: The Easy Mode (L7)
Ollama is the easiest path to a local LLM — three install steps, the 8B/14B/32B/70B shorthand, the q4_K_M / q8_0 / :instruct tag conventions, and the three boxes where it wins and loses.
- 123 ai
Set Up vLLM on a Single GPU: From Drivers to an OpenAI-Compatible Server
Install vLLM on one NVIDIA GPU, choose a model that fits, launch its OpenAI-compatible API, and avoid startup and production memory traps.
- 124 ai
MiniMax M3 vs Haiku 4.5 vs Gemini 2.5 Flash vs GPT-5-mini: The Budget-Tier Map
Four cheap LLMs compared on the only axes that matter on a budget: input, output, context, and the workloads where each one quietly wins or quietly costs you.
- 125 ai
Temperature, Top-P, and the Other Sampling Knobs: When to Touch Them and When to Leave Them Alone
When to touch temperature, top_p, top_k, frequency/presence penalties, max_tokens, and seed — and the default-everything rule that keeps most calls boring.
- 126 ai
Provider Rate Limits and the Fallback Rule: When Hermes Must Switch vs When It Must Hold
The three rate-limit shapes (RPM, TPM, RPD), the hermes fallback chain (anthropic → openrouter → local), when the chain fires, and the fallback must never pay twice rule.
- 127 computers
ABS Chat with Skills and Memory: A Practical Scoping Guide
Run ABS chat sessions with skills, memory, and fact_store in context, then scope the load when cost, latency, or focus matters.
- 128 computers
ABS Cron Job Self-Contained Prompts: The Rule
A Hermes cron prompt runs in a fresh session — no memory of past runs, no inline handoff. Six elements make it self-contained. Skip any one and the job silently returns nothing useful.
- 129 computers
ABS Frontmatter: Required vs Optional Fields, the Rule
ABS frontmatter is enforced by six Zod validators in src/content.config.ts — title ≤120, description ≤200, four enums, and a URL array. Build fails loudly on violation. Defaults fill the rest.
- 130 computers
ABS Content Frontmatter: The Canonical Shape
A field-by-field guide to the canonical ABS post header, including key order, image conventions, valid enums, and the Astro publishing pipeline.
- 131 computers
ABS Keyboard Shortcuts: n, r, b (No Prefix)
ABS ships three single-letter shortcuts, no prefix, no modifiers, no armed state. The bindings, why we left a g-prefix design behind, what we considered and cut, and the skip-rules.
- 132 computers
ABS pubDate: Timezone UTC-Only, the Decision
ABS coerces pubDate to a UTC Date via z.coerce.date() and sorts by epoch ms. Local-clock pubDates are an anti-pattern; DST swaps a post's rank twice a year. The schema, the comparator, the verify.
- 133 computers
ABS Mobile Nav: The Three Rules
Mobile nav on ABS runs through a drawer (not modal), with aria-expanded state on the toggle, and focus stays trapped inside until closed. The body-scroll-lock and the close behaviors, in three rules.
- 134 computers
ABS Regression Suite: Eight Failure Modes and Their Cheap-First Checks
A troubleshooting reference for the eight ABS regression scripts: what each failure looks like, the cheap-first check before anything else, and the canonical fix.
- 135 computers
Hermes `--resume` vs `--continue`: When Each Fits and the Stale-Session Trap
When to use `hermes chat --resume <id>` (deterministic) vs `--continue` (convenient): the trade-off, the decision rule, and three stale-session failure shapes.
- 136 computers
ABS RSS and Reader-Friendly Output: The Decisions
ABS publishes a single RSS feed at /rss.xml carrying news, reviews, and guides sorted newest first. The autodiscovery link tag, the no-JS reader-friendly HTML pages, and the verify commands.
- 137 computers
Hermes Terminal Output: The 50 KB Cap and the Pipe Fix
Hermes terminal output can stop before the line you need. Here is how to recognise the 50 KB cap, reduce noisy output, and recover the missing bottom safely.
- 138 computers
Hermes write_file: When to Write to Which Path, the Rule
Hermes write_file uses three buckets: /tmp/ for ephemeral scripts, /srv/abs-site/src/ for canonical in-tree files, /opt/data/ for operator scripts and backups. Includes the wrong-path failure shape.
- 139 computers
ABS Sort Pages: By Grade and Date, the Rules
ABS review listings default to newest-first; the secondary key is grade. URL params (?sort=grade|oldest|new) reach a /reviews/[sort].astro. The four tie-breakers and why we don't sort by score.
- 140 computers
Affiliate Link Auto-Tag: On Every Publish, the Rule
Every review publish runs an affiliate-link-rewrite pass: amazon.com URLs become amzn.to/agenticbotsit-20 tracked. Idempotent, run-at-deploy, audit-visible. Script, verify, failure shapes.
- 141 computers
Agents and the Critique Loop: When to Self-Review
When an agent should self-review before surfacing, when to defer to a steelman judge in a multi-agent swarm, and the four-question self-check that catches the most common regression.
- 142 computers
Agents and the Deploy Gate: When They Push
How an agent runs the ABS deploy playbook end-to-end: three-SHA check, regression gate, rsync to webroot, CF cache purge, commit, github push, ship-ready for operator review.
- 143 computers
Agents and Their Rules of Engagement: No Bulk Surprises
An agent's contract with the operator: no surprise bulk actions, surface decisions, ask before irreversible changes, quiet mode for crons. Shipping 10 guides unchecked is a broken rule.
- 144 ai
Anthropic Provider on ABS: Quick Setup and the Cheap Mistakes
Wiring Anthropic as the primary Hermes provider on ABS: API key, model pin, fallback chain position. Plus the 6 cheap mistakes operators make at setup that compound over months.
- 145 computers
Building and Deploying an Astro Static Site Behind a Cloudflare Tunnel
An operator's recipe for taking a static Astro site from `npm run build` to a live URL behind a Cloudflare Tunnel, with rsync, edge cache purges, and the three-SHA deploy gate.
- 146 computers
ABS Affiliate Link Convention: Voice Note + amzn.to, Built-In Tag
How the botsitter-review skill handles affiliate links: the voice note + amzn.to input shape, the auto-tag rewrite to agenticbotsit-20 at build, and the rules for writing tag-safe from URLs.
- 147 ai
Claude Sonnet vs Haiku vs Opus: When to Pick Which
A practical decision rubric for the Anthropic Claude tiers: per-tier cost, speed, capability, and a 7-row table mapping everyday tasks to the right model.
- 148 computers
Cloudflare Access Policies for Internal Tools
Put Cloudflare Access (Zero Trust) in front of an internal tool: app definition, Allow policy bound to an email domain, service-token back door for crons, and the audit log to subscribe to.
- 149 computers
Cloudflare Edge Cache: Purging Without Stale 404 Leaks
How to purge CF edge cache so a deploy doesn't serve 404s to users mid-flight: ordered purge, HTML-first then assets, deployment window, and verification probes.
- 150 computers
Cloudflare HSTS Preload: The Hard-to-Undo Directive
HSTS preload pins your domain into the browser-shipped list. Once on hstspreload.org, removing the entry takes 6-12 weeks to reach Chrome and Firefox. Avoid shipping it prematurely.
- 151 computers
Cloudflare Tunnel: Adding Multiple Hostnames Without Breaking the Others
How to put several public hostnames on a single Cloudflare Tunnel: cloudflared tunnel route dns per hostname, ingress rule order, incremental test pattern, and the catch-all 404 fallback.
- 152 computers
Cloudflare Tunnel: When the Daemon Dies, How to Recover (cfOrigin;dur=122)
Diagnose and recover a broken Cloudflare Tunnel: cloudflared is up but origin is silent, cfOrigin;dur=N ms traps, tunnel info diff, restart workflow, and what NOT to try before a restart.
- 153 computers
Cron Watchdogs vs Cron Monitors: When Each Fits and the Silent-Watchdog Trap
The semantic gap between watchdogs (fix-then-alert) and monitors (alert-only), the 3 shapes where a monitor is right, and the 3 shapes where a watchdog is the only correct choice.
- 154 computers
First-Hand Versus Research-Only Product Reviews: The ABS Editorial Bar
The ABS editorial bar is first-hand: did the operator actually use the product, in what setting, compared to what. Research-only reviews don't ship. The botsitter-review skill enforces the line.
- 155 computers
Git Commands an AI Agent Should Know
The ~15 git commands an agent uses in 95% of work, the failure modes for each, and the discipline that keeps an agent's git history clean.
- 156 computers
Git Commit and Push as the Deploy Gate: The ABS Convention
Why the live ABS site only deploys when local/origin/github are on the same commit. What this convention prevents, what it doesn't, and the post-receive hook that enforces it.
- 157 computers
Hermes Agent Cron Jobs: Authoring, Scheduling, and the Self-Contained Prompt Rule
Author Hermes cron jobs that survive a fresh session: prompt self-containment, schedule patterns, dry-run, pause/resume, run-once, and the difference between agent and no_agent modes.
- 158 computers
Hermes Agent Gateway: When Messages Don't Arrive
A troubleshooting reference for the Hermes messaging gateway: token/allow-list, adapter, runtime, routing — separating failures by cheap-first check order when the bot doesn't respond.
- 159 computers
Connecting Hermes Agent to Telegram: A First-Time Setup Guide
Step-by-step setup of the Hermes messaging gateway for Telegram: BotFather token, allow-list, home chat ID, and proof that the first message actually arrives back in the chat.
- 160 computers
Hermes Agent Memory vs Skills: Where Things Belong
A practical decision rule for the four Hermes storage surfaces — memory, skills, fact_store, session handoff — so durable preferences land in memory and procedures land in skills.
- 161 computers
Hermes Agent on a VPS: Hostinger, DigitalOcean, or Vultr First-Run Setup
Step-by-step setup of Hermes Agent on a Linux VPS: SSH key, non-root user, Python 3.11, Ubuntu 22.04 base, cloudflared install, and the first 'Reply with OK' chat.
- 162 computers
Hermes Agent on Your Local PC: Windows, Linux, macOS, and WSL2
How to install Hermes Agent on a workstation: native macOS, native Linux, Windows via WSL2, and what changes vs the VPS path.
- 163 computers
Hermes Agent Plugins: The Shape, the Register Entry Point, and How to Test Locally
How to author a Hermes plugin: the directory layout, the register(ctx) entry point, three real plugin examples on this build, and how to test a plugin locally before installing.
- 164 ai
Hermes Agent Provider: Anthropic API (Claude family — Slack-bot, Computer-Use)
Wiring Anthropic as a Hermes provider: API key from console.anthropic.com, the -latest model-pin rule, model-name quick-reference for the Claude family, and the Slack-bot and Computer-Use use cases.
- 165 ai
Hermes Agent Provider: Gemini (Google AI Studio, 2M-Token Context King)
Wiring Google Gemini on Hermes: API key from Google AI Studio, the model family (2.5 Pro, 3 Flash, 3.5 Pro), the 2M-token context window as the standout strength, and the verify pattern.
- 166 ai
Hermes Agent Provider: Local OpenAI-Compatible (vLLM, Ollama, LM Studio)
Wire any OpenAI-compatible local server as a Hermes provider: the --base-url pattern, three case studies, and the token-budget trade-off.
- 167 ai
Hermes Agent Provider: Nous Portal (Hermes-4-70B, Hermes-4-405B, First-Class)
Wiring the first-class nous provider on Hermes v0.18.2 — hermes chat --setup nous, the Portal OAuth/API-key exchange, Hermes-4-70B and Hermes-4-405B, and how it differs from MiniMax and OpenRouter.
- 168 ai
Hermes Agent Provider: OpenAI and Codex Auth (GPT + Image Gen)
Wiring OpenAI on Hermes: API key path (per-token), Codex OAuth path (subscription-based, drives image gen too), dashboard-vs-realtime usage trap, 14-day OAuth refresh.
- 169 ai
Hermes Agent Provider: OpenRouter for Cross-Vendor A/B Tests (and as a Cost Audit)
When and how to wire OpenRouter as a fallback into Hermes: model strings, cost audit use case, fallback chain position. OpenRouter is a cross-vendor search and budget audit, not a primary.
- 170 computers
Hermes Agent: Switching Providers Mid-Run When a Provider Fails
How to detect and recover from a mid-task provider failure (Anthropic 529, OpenAI 429, local Ollama OOM): the symptom per provider, the runtime fallback, and the override path.
- 171 computers
Hermes Agent Sessions: Listing, Searching, Resuming, and Continuing
How Hermes persists sessions: what `hermes sessions list --source` shows, when to use --resume vs --continue, and how to grep old sessions via the SQLite store.
- 172 ai
Picking a Model Provider for Hermes Agent: A First-Run Decision Guide
How to choose between Nous Portal, OpenRouter, Anthropic, OpenAI, Gemini, MiniMax, and local OpenAI-compatible endpoints for Hermes Agent. Compares auth paths, costs, and capability tradeoffs.
- 173 computers
Hermes Agent Skills: Authoring, Curating, and the Frontmatter Rules
Where skills live, the SKILL.md frontmatter fields, the cross_profile flag and pinning, when to update vs add a new skill, and how curator catches drift in the weekly sweep.
- 174 computers
Hermes Agent: Updating Without Breaking Running Crons or the Live Gateway
How to bring upstream Hermes commits in safely: the 10 active jobs to inventory, the pre-update disable list, the post-update smoke-test pattern, and the rollback rule.
- 175 computers
Hermes Agent Tools: Enabling Only What You Need
A practical guide to Hermes' toolsets: how to list them, enable what you actually use, scope per-call via --toolsets, and avoid loading the full platform surface by default.
- 176 computers
Hermes Agent CLI Tools: List and What Each One Does
The Hermes Agent CLI surface today: chat, config, sessions list --source, cron, skills, plugins, memory, fact_store, curator. The --toolsets flag for per-call scoped tool loading.
- 177 computers
Hermes Context Window: Budgeting for Long Sessions
How to keep a long Hermes session productive: per-tool, per-skill, per-conversation-history budgets. What to keep, what to compress, what to drop. Operator discipline for long sessions.
- 178 computers
Hermes Agent on a Mac mini M4 (or other Apple Silicon Macs)
How to install Hermes Agent on macOS: Homebrew for Python 3.11, Xcode CLI tools, launchd for cron, Ollama for local inference, and Apple Silicon-specific gotchas.
- 179 computers
Hermes Output: Streaming vs Final Blocks, When to Pick Which
Hermes output arrives two ways: streamed token-by-token, or as a single final block. Each fits different surfaces — CLI, Telegram, mobile push, automation logs. The matrix of when to pick which.
- 180 computers
Hermes Protocol: The System Prompt, the Context Window, and the Rule Boundary
What lives in the system prompt vs in the context window vs in user-side memory: how Hermes treats the instruction hierarchy, what travels where, and what the operator's contract actually is.
- 181 computers
Hermes Skills vs Memory vs Fact Store: The Storage Decision
Three storage surfaces, three jobs. Use this duration / reuse / scope rubric to put a fact in the right place, plus the bad-shape markers that tell you the rubric is failing.
- 182 computers
How an AI Agent Reads a Multi-Message Input Batch
When the operator sends multiple messages at once, the agent concatenates them into context with timestamps. A multi-message batch is one turn, not a multi-turn dialogue. The P43 rule.
- 183 computers
How an AI Agent Reads Your Memory on the Next Turn
Memory is loaded at session start, not mid-session; conversation history is in-memory and dies with the session. The full disk-to-context read path and the curation discipline that keeps memory clean.
- 184 gadgets
Image Style: The Retro-Robot Motif (Locked Style Block)
The locked prompt prefix behind every ABS featured illustration: crimson + electric blue on warm cream, comic-book outlines, friendly retro robot, 16:9, no text. The canonical style block.
- 185 computers
Image Gen Failover Chain: In Priority Order
When the openai plugin is degraded, ABS failover chain is: openai (real key) → cached PNG set → cache-bust query string → text-only social card. Each step's trigger and verify.
- 186 computers
Image Gen: OpenAI Backend (Real Key, Not Codex OAuth)
The working image-gen backend for ABS: the openai Python package driven by OPENAI_API_KEY, the model gpt-image-2-medium, ~$0.04 per 1024x1024 image, and why openai-codex OAuth is broken.
- 187 ai
OpenAI Provider on ABS: Quick Setup, API Key vs Codex OAuth
Wiring OpenAI on Hermes Agent: the API key path (paid per-token), the Codex OAuth path (subscription-based, also drives image gen), when to pick which, and the dashboard-vs-realtime usage trap.
- 188 computers
Pause and Resume Hermes Cron Jobs Without Deleting History
Why `hermes cron pause <id>` and `hermes cron resume <id>` beat `remove` for three operator cases. `paused_at` is added; prompt and history are preserved. Verified against 35 paused jobs on this VPS.
- 189 ai
Picking the Default LLM for the Botsitter-Review Skill
How to pick the default LLM for the botsitter-review skill: a four-axis decision (cost, latency, instruction following, context), per-skill alternates, and the live AI Model Matrix as the price list.
- 190 computers
Publishing a Product Review on Agentic Bot Sitter, End-to-End
The complete review-publishing loop on ABS: voice note + amzn.to → draft → image gen → WebP → build → rsync → CF purge → commit → push → three-SHA → live.
- 191 computers
Python Venv on This VPS: The Conventions
The canonical Python venv on this VPS is /opt/hermes/.venv/ with Python 3.11. Library surface today: openai 2.33.0, requests, pillow, ruff. Why we never call system python3 for agent scripts.
- 192 ai
Review Publishing Pipeline: The Voice Note to Post Chain
The seven-stage review-publishing pipeline: voice note + amzn.to URL in, post out. Capture, STT, first-pass draft, review pass, image gen, commit, ship. Each stage's failure shape and verify.
- 193 computers
Static Site Deploy: rsync, Cloudflare Purge, CDN Cache — The 3-Step Loop
The static-site deploy loop and what each step does: rsync dist/ to the webroot, ordered CF purge (HTML -> ASSETS -> BUNDLE), then verify. Plus failure shapes and the order that minimizes 404 leaks.
- 194 computers
Search the Codebase Before Fixing Anything
Before fixing any bug, search the codebase for prior attempts, related patterns, and existing utilities. The agent's most expensive mistake is fixing the wrong layer. A 4-step search workflow.
- 195 computers
site-watchdog-sweep: Every 5min, the Consolidation Story
Why one Python sweeper replaced five separate 5min watchdogs, what its six checks actually probe, and how it stays silent when the operator surface is healthy and only shouts when something is red.
- 196 computers
Terminal Commands Versus Python: When an Agent Should Write a Script
When to use one-liner terminal commands and when to write a 5-line Python script. The trade-off is readability and verifiability, not tool sophistication. Five signals that tip the choice.
- 197 computers
Terminal Commands Versus Python: When an Agent Should Write a Script
When to use one-liner terminal commands and when to write a 5-line Python script. The trade-off is readability and verifiability, not tool sophistication. Five signals that tip the choice.
- 198 computers
Three-SHA Check: Asserting local == origin == github == the Commit Going Live
The bash assertion that runs before every ABS deploy: Local and origin and github must all carry the same commit SHA as the change about to ship. Failure shapes and the auto-repair pattern for each.
- 199 computers
Tmp Folder vs Write_file: The Rule
When the agent writes to /tmp/ vs when to use write_file: long scripts (>300 lines) go to /tmp/, short content uses write_file. The bash heredoc trap and the agent-tool boundary.
- 200 computers
Voice Note to Published Review: The Input, the Pipeline, the Output
From a 60-second voice memo to a live ABS review: transcription, draft, image, build, deploy. What arrives in the voice note, what comes out the other side, what the operator does in between.
- 201 computers
WebP and PNG on the Same Deploy: The P42 Rule
P42: every ABS featured image deploys as both PNG and WebP in the same cycle. CF serves WebP via Accept-header; the PNG is canonical. The script, the 4h cache quirk, and the verify pattern.
- 202 computers
WebP Companion Images for Static Sites: A Sibling PNG+WebP Pattern
How to ship a PNG and a WebP companion for every static-site image with Pillow's quality=82, method=4, plus the <picture> helper that lets browsers pick the lighter file.
- 203 computers
Cloudflare Tunnel Setup for a Static Site Without Opening Any Ports
Setup guide for installing cloudflared, creating a named tunnel, routing a hostname to it, and verifying the public hostname reaches your local origin over the tunnel.
- 204 computers
How to troubleshoot a Cloudflare Tunnel that is not serving your site
Troubleshooting guide for separating local-origin, connector, DNS, and policy-layer failures in a named Cloudflare Tunnel. Cheap-first diagnostic order so you can tell which layer is broken.
- 205 computers
How to Set Up Hermes Agent Without Turning Everything On at Once
Setup guide for Hermes Agent: install the CLI, prove a plain chat works, then layer on tools, memory, skills, and messaging only after the base is healthy.
