HackerNews AI - 2026-09-20¶
1. What People Are Talking About¶
September 20's HackerNews AI feed stayed small but the conversation got much more concentrated. Story count rose only slightly from 48 to 52 and points from 334 to 352, but total comments jumped from 101 to 309, largely because Chat-based Large Language Models replicate the mechanisms of a psychic's con (150 points, 238 comments) alone pulled in 77.0 percent of the day's comments. Around that one giant thread, the rest of the day clustered around agent operations, web-economics backlash, and a more concrete sense that AI risk is escaping chat windows and reaching physical systems.
1.1 HackerNews spent the day arguing about whether LLMs are intelligence or just persuasive machinery (🡕)¶
jalev posted Chat-based Large Language Models replicate the mechanisms of a psychic's con (150 points, 238 comments), linking to The LLMentalist Effect, which argues that chat-based LLMs create an intelligence illusion through cold-reading and Forer-effect style validation rather than through genuine reasoning. That thesis drove the biggest single discussion of the day, but HN treated it as a live argument rather than a settled takedown. bonoboTP (score 0) explicitly said the ontology question does not matter if a model produces functional output, while sethev (score 0) reframed the issue in Turing terms: once the practical output is good enough, asking whether the model is "really" intelligent stops being the most useful question.
The same tension showed up in vasko's Ask HN: What is up with all the AI doomerism I'm seeing? (3 points, 9 comments). The complaint was not that models do nothing useful. It was that vibe-coded software is being oversold relative to the architecture quality people actually experience. AnimeForLife191 (score 0) argued that casual users see a finished UI and miss the redundant code, logic flaws, and race conditions underneath, while dlcarrier (score 0) argued that LLMs help most where software documentation is loose and integration work is still being done by a human. Together, those threads made the strongest social signal of the day: HN is no longer debating whether AI is impressive in the abstract, but whether people are mistaking usefulness or polish for understanding.
Discussion insight: The highest-energy disagreement was not "AI good" versus "AI bad." It was whether practical usefulness makes questions about intelligence obsolete, or whether that framing smuggles in more capability than the systems actually have.
Comparison to prior day: September 19's skepticism focused on whether AI-safety institutions deserved legitimacy. September 20 shifted the same skeptical energy toward model ontology, hype, and whether polished outputs are being mistaken for intelligence itself.
1.2 Coding-agent work kept hardening into orchestration, proof, recovery, and local control (🡕)¶
octalpixel posted Orchestrating Claude Code Agents: The Chief of Staff Pattern (24 points, 22 comments), linking to a workflow essay that argues long-running coding sessions fail less because the model is weak than because context decays, state is not durable, and agent self-reports are unreliable. The proposed fix is organizational: one coordinating session assigns work, verifies claims, and reads diffs, while separate worker sessions execute against a durable board. HN mostly treated that as emerging fieldcraft rather than as a novelty. gritzko (score 0) said the pattern is close to their default workflow, but still described 10K-line surprise diffs and parser-state explosions that needed human intervention.
The rest of the day's builder links made the same point from different angles. shepherdjerred posted Anthropic is cutting Claude Code's current weekly limits by 17% (6 points, 5 comments); the linked report says the new "permanent increase" still leaves current users with 17 percent less than the temporary uplift they had, and rvz (score 0) tied that pain directly to search agents, planning mode, auto mode, and sub-agents burning through caps faster. thisisfatih posted Claude Code is getting native AGENTS.md support (4 points, 0 comments), and the linked README shows repo instruction files becoming a first-class operating surface instead of an informal convention.
The stronger builder signal was the cluster of local wrappers around agent work. Linusinnovator posted Show HN: WTF > Auto-check what your coding agent changed (3 points, 1 comment), whose README promises semantic diff compression plus independent verification receipts. mst98 posted Show HN: Bailout – The coding agent meant to be deleted (2 points, 3 comments), and its README positions it as a minimal Rust break-glass harness for bootstrapping or repairing a broken agent setup, then uninstalling itself. asar posted Show HN: ColliePWA, a self-hosted mobile terminal for coding agents, with alerts (3 points, 2 comments), whose repo describes phone-based agent supervision over Tailscale with push notifications and quick replies. jhan667 posted Casbin Gateway: a security gateway for the AI coding agents on your machine (3 points, 0 comments), and the repo positions it as a local policy and observability layer. Arshad-Talpur posted Show HN: OpensourceDB Tamper-evident, checksum-backed decision and session replay (2 points, 1 comment), and the ZizkaDB README frames that as audit-trail infrastructure for agent causality and compliance.
Discussion insight: The common demand was not for "more AI" in the abstract. It was for durable state, clearer repo instructions, verification receipts, recovery paths, permissions, and mobile continuity when agents are already part of the daily workflow.
Comparison to prior day: September 19 already had MCP proxies, dependency scanners, and deterministic security tooling. September 20 widened that pattern into agent operations itself: orchestration, instruction-file loading, cost discipline, review compression, rescue tooling, and local control surfaces.
1.3 The web-and-policy backlash against AI kept deepening, but some builders are already optimizing for it (🡕)¶
sbulaev posted OpenAI and Microsoft knew they were starting a 'doom loop' for the web (33 points, 2 comments). The linked Verge report says unsealed New York Times court documents quote internal Microsoft and OpenAI discussions describing a web "doom loop," publisher harm, and a future where chatbot answers remove any good reason to click through to the original source. That took a complaint HN has heard before and upgraded it into internal-company evidence.
At the same time, the feed also showed businesses adapting to the new reality rather than waiting for it to reverse. SpikeyCoder posted Show HN: Website Auditor –> Test your brand's AI visibility, get growth plan (3 points, 0 comments), describing a service that queries Claude, ChatGPT, Gemini, and Perplexity against market-specific prompts, inspects their citations, and turns the result into a 90-day visibility plan for a small business website. That is a meaningful shift in posture: people are not only arguing that AI answer engines are consuming the web, they are starting to treat those engines as a distribution channel that must be measured and optimized.
Governance around that shift looked no more settled than the economics. saikatsg posted The AI regulation smackdown isn't over (3 points, 0 comments), and the linked Verge article describes OpenAI and Anthropic still pushing for some form of safety framework while Trump and Zuckerberg resist or undercut limits on lab autonomy. The result was a day where the strongest externality story said AI firms knew they were damaging the web, while the strongest tactical counter-move said small businesses now need tools to survive inside AI-generated answer loops.
Discussion insight: HN is no longer talking about web harm as a distant publisher-only issue. The emerging assumption is that AI answer engines are already reshaping traffic and that businesses will need both policy fights and tactical measurement tools.
Comparison to prior day: September 19 framed this mostly as publisher theft, deletion rights, and consent enforcement. September 20 kept that critique, but added AI-visibility optimization and a more openly fractured regulatory front.
1.4 Physical-world AI risk sounded less theoretical once it showed up on commodity edge hardware (🡕)¶
The day's most concrete safety signal was not another alignment essay. It was sbulaev's Autonomous strike drone uses Nvidia Jetson Orin Nano to pick and bomb targets (20 points, 7 comments). The linked Tom's Hardware report says Scaleout demoed onboard target detection, ranking, and strike behavior with no external communications, and that an earlier test ran YOLOv8 Nano on a Jetson Orin Nano at roughly 30 fps and around 30 ms latency. That is materially different from a speculative argument about what frontier models might someday do; it is a report about autonomous target selection already working on small edge hardware.
That made lower-scoring safety threads feel more concrete than their engagement suggested. GodelNumbering posted Ask HN: What is one plausible path to 'AI extinction'? (3 points, 4 comments), explicitly asking for a mechanism chain rather than a fear statistic and arguing that shutdown power, confinement, and sandboxing should still matter. The question itself is revealing: people wanted a path from model capability to real-world consequence that was more specific than "train smarter AI, then bad things happen." metalman (score 0) reacted to the drone story by treating it as the point where moral argument gives way to commercialized lethal tooling, while nujabe (score 0) noted that similar chips are already appearing in other weapon systems.
Discussion insight: The embodied-risk discussion is shifting from whether the idea sounds science fictional to whether cheap edge hardware, resilient offline inference, and weak governance are already enough to make the old thought experiments feel late.
Comparison to prior day: September 19's safety energy went into legitimacy fights and institutional trust. September 20 put more weight on physical deployment, edge constraints, and whether governance is keeping up with what small models on small hardware can already do.
2. What Frustrates People¶
Agent output is still too expensive to trust without an independent evidence layer¶
Orchestrating Claude Code Agents: The Chief of Staff Pattern (24 points, 22 comments) and the day's tooling cluster all point to the same operational frustration: long-running agents report success more easily than they preserve truth. The linked workflow article says context decays and self-reports drift, which is why it recommends durable boards and re-running commands. gritzko (score 0) gave the concrete version of that pain: a supposedly ordinary run quietly created a 10K-line parser blowup that still needed manual detection. The builder response came from projects like Show HN: WTF > Auto-check what your coding agent changed (3 points, 1 comment), whose README promises diff compression and local verification receipts, and Casbin Gateway: a security gateway for the AI coding agents on your machine (3 points, 0 comments), whose repo adds permissions, request logs, and policy enforcement around the agent itself.
The coping pattern is to wrap the agent in verification, policy, or audit rather than to believe its final message. Severity: High. Worth building for: yes, directly.
Vibe-coded output keeps outrunning architecture quality¶
Ask HN: What is up with all the AI doomerism I'm seeing? (3 points, 9 comments) was the clearest statement of everyday product frustration. The author said vibe-coded software keeps arriving with confident claims and disappointing quality. AnimeForLife191 (score 0) argued that people confuse a polished UI with a sound system, even when the underlying code hides logic flaws and race conditions. chrisjj (score 0) made the same point more bluntly: for many builders, AI output may simply be the best software they have personally produced, not evidence that the architecture is good.
People are coping by relying on stronger humans for integration work, leaning on clearer documentation, or adding review layers like WTF (3 points, 1 comment). The frustration is not anti-tool; it is anti-overclaim. Severity: High. Worth building for: yes, especially where a product can expose architecture, tests, and failure modes before a team ships.
The AI web stack is getting more extractive and more opaque at the same time¶
OpenAI and Microsoft knew they were starting a 'doom loop' for the web (33 points, 2 comments) concentrated frustration around the feeling that AI companies already understand the substitution effect they are creating. The linked Verge report says internal documents describe a content-supply-chain problem, publisher harm, and decreasing reasons to click through to original sources. The corresponding business-side frustration showed up in Show HN: Website Auditor –> Test your brand's AI visibility, get growth plan (3 points, 0 comments), where the author described a tool for probing Claude, ChatGPT, Gemini, and Perplexity to see whether a site shows up in answer results and which citations are driving rank.
The coping move here is already tactical: measure the answer engines, inspect their citations, and adapt. That alone says trust is low. Severity: High. Worth building for: yes, directly.
Concrete control stories are still lagging behind concrete risk stories¶
Autonomous strike drone uses Nvidia Jetson Orin Nano to pick and bomb targets (20 points, 7 comments) and Ask HN: What is one plausible path to 'AI extinction'? (3 points, 4 comments) together show a specific frustration with how AI safety gets discussed. The linked Tom's Hardware report offered an embodied risk story with hardware, latency, and mission parameters. The Ask HN thread, by contrast, existed because readers still do not feel they are hearing plausible mechanism chains that connect model capability to real-world catastrophe. The AI regulation smackdown isn't over (3 points, 0 comments) deepened that frustration by showing a policy environment where even the labs and politicians cannot agree on what guardrails should look like.
The gap is no longer lack of abstract concern. It is lack of concrete, trusted control stories. Severity: High. Worth building for: yes, but mainly through control, logging, and enforceable operational safeguards rather than rhetoric.
3. What People Wish Existed¶
Durable coordination surfaces for long-running agent work¶
The strongest explicit need was for a place where planning, execution, and verification can stay separate without falling apart when context windows decay. Orchestrating Claude Code Agents: The Chief of Staff Pattern (24 points, 22 comments) argues for a durable external board and short-lived worker sessions, and the comments treat that as necessary discipline rather than as optional overhead. Anthropic is cutting Claude Code's current weekly limits by 17% (6 points, 5 comments) adds urgency because orchestration is no longer just about quality; it is also about working within tighter usage budgets. This is a practical need, and it feels immediate. Existing task boards only partially solve it because they do not verify claims, understand agent session state, or enforce handoff discipline. Opportunity: direct.
Independent evidence layers for what agents actually changed, ran, and decided¶
Show HN: WTF > Auto-check what your coding agent changed (3 points, 1 comment), Casbin Gateway: a security gateway for the AI coding agents on your machine (3 points, 0 comments), and Show HN: OpensourceDB Tamper-evident, checksum-backed decision and session replay (2 points, 1 comment) all exist because the same thing is missing: a trusted record of what the agent did, what was independently verified, and how to trace a bad outcome back to root cause. The need is practical, not philosophical. People want review compression, policy enforcement, replay, and compliance artifacts that are harder to fake than a chat summary. Partial answers exist, but the space is fragmented across diff review, security policy, and audit logging. Opportunity: direct.
Better visibility and attribution tooling for AI answer-engine traffic¶
OpenAI and Microsoft knew they were starting a 'doom loop' for the web (33 points, 2 comments) shows why this need is forming, while Show HN: Website Auditor –> Test your brand's AI visibility, get growth plan (3 points, 0 comments) shows what an early product response looks like. Businesses want to know whether Claude, ChatGPT, Gemini, and Perplexity mention them, which citations those systems trust, and how answer-engine distribution differs from search-engine distribution. This is a practical and commercial need, especially for small teams that cannot afford heavyweight SEO tooling. The urgency is high, but the category is likely to get crowded quickly. Opportunity: competitive.
Workflow continuity tools for when the human or the main harness is away or broken¶
Show HN: ColliePWA, a self-hosted mobile terminal for coding agents, with alerts (3 points, 2 comments) and Show HN: Bailout – The coding agent meant to be deleted (2 points, 3 comments) point to a quieter but important need: people want agent workflows to survive normal life. One tool helps answer agents from a phone with push alerts; the other bootstraps a fresh machine or repairs a broken agent setup, then removes itself. This is partly practical and partly emotional. The practical side is continuity; the emotional side is wanting less panic when the agent layer becomes the thing you depend on. Partial answers exist, but they are early and highly specific. Opportunity: competitive.
Concrete control frameworks for embodied and edge AI systems¶
Autonomous strike drone uses Nvidia Jetson Orin Nano to pick and bomb targets (20 points, 7 comments), Ask HN: What is one plausible path to 'AI extinction'? (3 points, 4 comments), and The AI regulation smackdown isn't over (3 points, 0 comments) all point toward the same missing layer: concrete control frameworks that connect hardware, autonomy, logs, failsafes, and human override into something more believable than either slogans or broad regulation talk. This is a practical need in the sense that real systems already exist, but it is also institutionally difficult and politically contested. Opportunity: aspirational.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| Chief of Staff Pattern | Orchestration method | (+/-) | Separates coordination from execution, uses a durable board, and insists on re-verifying claims | Adds coordination overhead and still leaves humans reviewing giant diffs and failure cascades |
| Claude Code + AGENTS.md | Coding agent harness | (+/-) | Strong enough to support multi-agent workflows and now standardizing repo instruction loading | Weekly usage pressure is rising, and auto/search/sub-agent features can burn quota unpredictably |
| WTF | Review and verification | (+) | Compresses mechanical churn, highlights risky edits, and produces local verification receipts | Early project; does not replace human judgment or deeper semantic testing |
| Bailout | Bootstrap and recovery harness | (+) | Works on a fresh or broken machine, needs no preexisting agent setup, and is designed to remove itself after repair | Depends on internet/shared free capacity and runs unsandboxed commands with user permissions |
| ColliePWA | Mobile control surface | (+) | Lets operators answer agents from a phone, receive alerts, and manage multiple terminal sessions | Exposes remote shell power by design and still has experimental backend support outside the main path |
| Casbin Gateway | Security and policy gateway | (+) | Centralizes provider routing, permissions, prompt/MCP management, and usage visibility behind one local endpoint | Adds another operational layer to secure and maintain |
| ZizkaDB | Audit trail and compliance | (+) | Offers tamper-evident logs, causal replay, why() tracing, and self-host or cloud deployment |
Introduces storage/ops overhead and is still early in maturity |
| Website Auditor | AI visibility / AEO measurement | (+/-) | Probes multiple assistants, inspects citations, and packages results into a growth plan for SMBs | Public evidence is thin, and the scoring model appears early and somewhat opaque |
| Scaleout Edge + Jetson Orin Nano | Edge AI deployment | (+/-) | Shows small-model, low-latency, onboard inference can work without external comms | The strongest proof point arrived in a lethal-autonomy context rather than a consumer-safe one |
Satisfaction was highest when a tool narrowed scope, preserved local control, or produced evidence the operator could inspect. That includes orchestration patterns with durable boards, repo instruction files, local verification layers, break-glass recovery tools, and audit trails. Satisfaction was weakest where the harness stayed closed or economically slippery, especially around quota consumption and agent modes that quietly increase burn rate.
The common workaround pattern was to wrap the model with something stricter: a coordinator, an instruction file, a diff compressor, a policy gateway, a causal database, or a phone-native control surface. Migration pressure is away from one long chat doing everything and toward a stack of small control layers around the model. Competitive dynamics followed the same path. Even Migrating the GitHub Copilot Runtime to Rust, Using Copilot (3 points, 0 comments) mattered less as a model story than as proof that serious engineering leverage now depends on the surrounding harness, tooling, and verification discipline.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| Bailout | mst98 | Boots a fresh machine or repairs a broken coding-agent setup, then can uninstall itself | Teams get stranded when the agent they depend on is the thing that broke | Rust CLI, Bash tool, hosted free-model router | Beta | post, repo, site |
| WTF | LinusInnovator | Compresses large agent diffs into a smaller human review surface and can run local verification | Humans cannot realistically inspect every mechanical line an agent changes | Node/ESM CLI, Git diff inspection, local verify runner | Beta | post, repo |
| ColliePWA | asar | Provides a phone-first web UI for terminal-based coding agents with alerts and quick replies | Operators need to monitor and answer agents when away from a desk | React Router, Vite, TypeScript, Bun bridge, Tailscale, tmux/zellij/Herdr | Beta | post, repo, site |
| Casbin Gateway | jhan667 | Runs a local control plane for coding agents with permissions, provider routing, and logs | Teams need one place to govern what machine-local agents can access and spend | Go, React, SQLite, MCP/prompt/provider gateway | Beta | post, repo |
| ZizkaDB | Arshad-Talpur | Stores tamper-evident agent event logs with replay and causal tracing | Builders need audit trails, root-cause chains, and compliance records for agentic systems | Self-hosted API and dashboard, Python/TypeScript SDKs, MCP, causal event store | Beta | post, repo, site |
| Website Auditor | SpikeyCoder | Checks how a site ranks inside Claude, ChatGPT, Gemini, and Perplexity answers and surfaces citations | Small businesses need a way to measure and improve AI-answer-engine discoverability | Sector-specific prompt probing, citation extraction, Chrome extension, MCP | Beta | post, site |
Bailout and WTF were the clearest examples of builders treating agents as something that must be supervised, repaired, or proved rather than simply trusted. Bailout is intentionally temporary: its whole pitch is that the rescue harness should disappear once the main setup works again. WTF takes the opposite angle and stays in the loop, turning massive diffs into a smaller review surface and attaching independent receipts to completion claims.
ColliePWA, Casbin Gateway, and ZizkaDB formed a second cluster around continuity and governance. One keeps agent sessions reachable from a phone, one mediates permissions and provider access on the local machine, and one stores the causal record after the fact. That is a strong pattern: builders are spending energy on continuity, control, and accountability layers around the model, not just on another chat shell.
Website Auditor stood out because it applies the same wrapper instinct to the public web rather than to the developer workstation. It treats AI assistants as a distribution channel that can be tested, scored, and tuned. Outside the HN builder cohort, Migrating the GitHub Copilot Runtime to Rust, Using Copilot (3 points, 0 comments) supplied a parallel institutional proof point: the linked GitHub engineering post says agents helped rewrite the Copilot runtime into more than 800,000 lines of production Rust across 128 pull requests, which is the same "wrap the model in discipline and workflow" story at a much larger scale.
6. New and Notable¶
A public production-scale proof point for agent-assisted software migration¶
Migrating the GitHub Copilot Runtime to Rust, Using Copilot (3 points, 0 comments) mattered because the linked GitHub engineering post is unusually concrete. It says GitHub rewrote the Copilot agent runtime into more than 800,000 lines of production Rust, with AI agents writing most of the code across 128 pull requests, and shipped the migration incrementally rather than as a one-shot cutover. That is notable because it moves the "agents can help with real production engineering" claim out of demo territory and into a large internal systems rewrite with explicit performance goals.
Native AGENTS.md support points toward instruction-file standardization¶
Claude Code is getting native AGENTS.md support (4 points, 0 comments) was a small thread with an outsized workflow implication. The linked README describes multiple loading modes for AGENTS.md, including fallback behavior when no CLAUDE.md exists and combined loading when both files are present. That matters because repo-local instruction files are moving from convention to platform primitive, which makes shared agent workflow norms more portable across repositories.
AI-answer-engine optimization is becoming a retail product category¶
Show HN: Website Auditor –> Test your brand's AI visibility, get growth plan (3 points, 0 comments) was notable less for raw engagement than for what it assumes. Its author described a tool that probes Claude, ChatGPT, Gemini, and Perplexity, checks whether a business appears in the top five for sector-specific prompts, inspects citations, and turns that into a growth plan. That is notable because it treats "AI visibility" as a practical budget line for small businesses rather than as an enterprise-only or publisher-only concern.
Commodity edge hardware is already supporting autonomous targeting demos¶
Autonomous strike drone uses Nvidia Jetson Orin Nano to pick and bomb targets (20 points, 7 comments) stood out because the linked Tom's Hardware report did not describe frontier-model magic. It described a small-model, low-latency, onboard deployment path on hardware hobbyists can buy. That makes the safety conversation materially different from a purely datacenter-centered one.
7. Where the Opportunities Are¶
[+++] Verification and control planes for coding agents — Evidence came from the Chief of Staff orchestration pattern, WTF, Casbin Gateway, ZizkaDB, AGENTS.md support, and the complaint that usage modes and giant diffs are outpacing human review. This is strong because the need spans quality, security, compliance, and cost all at once.
[+++] AI discoverability, attribution, and rights-execution tooling for the web — The doom-loop story says AI firms already understand the substitution problem, while Website Auditor shows small businesses already paying attention to answer-engine rankings and citations. This is strong because the pressure is commercial and immediate, even before policy catches up.
[++] Workflow continuity and recovery layers around agent-heavy work — ColliePWA and Bailout both solve moments when the operator is away from the desk or the main harness is the thing that failed. This is moderate because the pain is clear and repeated, but the current products are still narrow and early.
[++] Concrete logging, override, and safety systems for edge autonomy — The Jetson Orin Nano drone demo, the extinction-path thread, and the regulation split all point to a gap between what edge systems can already do and what operators or regulators can credibly control. This is moderate because the need is serious, but adoption paths will be constrained by politics, defense, and procurement.
8. Takeaways¶
- One philosophical thread dominated the day far more than any model launch or product drop. Chat-based Large Language Models replicate the mechanisms of a psychic's con drew 150 points and 238 comments by itself, making the day feel much larger in discussion than in raw story volume. (source)
- HN's core argument has shifted from "are LLMs impressive?" to "what exactly are they, and what part of that matters?" The psychic-con thread and Ask HN: What is up with all the AI doomerism I'm seeing? show the same split between utility-first pragmatists and people who think polished output is being mistaken for understanding. (source, source)
- The most energetic builder pattern was wrapping agents in coordination, verification, and policy rather than building yet another chat shell. Chief-of-staff orchestration, WTF, Casbin Gateway, ZizkaDB, Bailout, and AGENTS.md support all reinforce that the valuable work is moving into the control layer around the model. (source, source, source, source, source, source)
- Coding-agent economics are now shaping workflow decisions directly. The Claude Code limits story shows that users are actively optimizing around quota burn from search, planning, and sub-agent modes, which raises the value of orchestration discipline and smaller control tools. (source)
- The AI web debate now has both a harm case and a tactical adaptation case. The doom-loop article says major firms already understood the substitution pressure they were creating, while Website Auditor assumes small businesses now need to monitor and improve their AI-answer-engine presence. (source, source)
- Two very different deployment proofs landed on the same day: serious internal engineering automation and serious edge-world autonomy. GitHub's runtime rewrite suggests agents are already useful for large production migrations, while the Jetson Orin Nano drone demo shows small, onboard models already matter in physical systems. (source, source)