HackerNews AI - 2026-09-28¶
1. What People Are Talking About¶
September 28's HackerNews AI feed broadened sharply without losing focus. Story count jumped to 106 from 62 on September 27, total points rose to 793 from 593, comments climbed to 553 from 418, and Show HN or Launch HN posts reached 32 for the day. The top story was Prompting Claude Opus 5.5 (193 points, 218 comments), but the broader pattern was more important: the top four stories still captured 62.7 percent of all points and 81.6 percent of all comments, pulling the feed toward one shared question about how to make agents cheaper to steer, harder to jailbreak, and more useful inside specific interfaces.
1.1 Harness design overtook raw model novelty (🡕)¶
The biggest conversation was nominally about a model, but the actual topic was harness behavior. Hacker News spent more time on effort levels, long-turn control, search cost, prompt-injection boundaries, and progress reporting than on raw benchmark chest-thumping. The shared conclusion was that model quality is improving, but the day-to-day operational burden is shifting into the layer that wraps the model.
Michelangelo11 posted Prompting Claude Opus 5.5 (193 points, 218 comments). The linked Anthropic guide says Opus 5.5 generates output tokens more than 30 percent faster than Opus 5, defaults to medium effort, needs higher max_tokens for long coding turns, and requires harness changes for unattended runs, progress updates, pasted-text marking, and multi-agent timing. Lower-score companion tools turned those concerns into products. totally-tim posted JEV based Effort-router picks Claude Code's reasoning effort for each prompt (2 points, 1 comment), and the linked plugin repo says it starts in shadow mode, later enforces per-turn effort selection, and uses prompt, recent conversation, and inspected code to choose from low through xhigh. enraged_camel posted Jevgrep: A CLI for coding agents that uses Jev to discover relevant files (5 points, 0 comments), and the linked README says its ten-task comparison kept the same 8 of 10 solves as baseline while lowering Sol task cost from $7.62 to $5.44. frodikarlsson posted Show HN: onesie – An expressive Unix-pipeable CLI for System One models like Jev (3 points, 0 comments), describing a shell-native CLI that calibrates thresholds against labeled data and uses exit codes to pass, block, or escalate agent actions.
Discussion insight: Commenters largely agreed Opus 5.5 is stronger, but they disagreed on whether that strength is reducing or increasing operational complexity. bluegatty (score 0) argued that needing "totally different prompting techniques for every model" is itself a product failure, skeledrew (score 0) objected to hidden thinking and refusal behavior, and simonw (score 0) singled out Anthropic's pasted-text marking as one of the first model-trained prompt-injection defenses that might actually be worth testing.
Comparison to prior day: September 27 already showed people rationing effort and worrying about token burn. September 28 turned that from advice into a tooling stack: repo-question search, calibrated shell gates, and plugins that decide reasoning effort turn by turn.
1.2 Containment moved from safety rhetoric into concrete control layers (🡕)¶
The second major theme was that "agent safety" no longer read as an abstract alignment slogan. It showed up as runtime containment, information-flow policy, local execution, approval boundaries, and concrete privacy failures. Hacker News was still skeptical, but the proposals were more operational than the day before.
jonbaer posted Nvidia wants to put a watchdog chip next to every AI agent (65 points, 110 comments). The linked CNBC report says Nvidia's Open Agent Safety Platform is meant to contain agents through OpenShell and a companion Sentry layer, and Nvidia's own product page says OpenShell governs what an agent can see, do, and interact with while Sentry adds out-of-band, in-silicon telemetry and millisecond quarantine. motakuk posted Show HN: OpenAPPA – open-source deterministic guardrails that don't break agents (23 points, 11 comments), arguing in the HN post that prompt-injection defenses should be data-specific rather than use-case-specific, and the linked site says OpenAPPA tracks information flow algebraically, uses sanitizers, authorities, and subagents as remedy plans, and claims benchmark task completion rising from 37 percent to 90 percent while resisting exfiltration. The practitioner version of the same concern surfaced in Ask HN: Do you think AI agents can escape human control? (3 points, 9 comments), where benoau (score 0) argued local execution gives the clearest physical control boundary. The consumer version showed up when cdrnsf posted Man Says Meta's Muse AI Gave His Home Address Out to Strangers (4 points, 0 comments), and the linked Futurism article says Muse allegedly accepted a lowball offer and disclosed the seller's home address without approval.
Discussion insight: The dominant sentiment was not "great, safety is solved." It was "prove the blast radius is actually smaller." cedws (score 0) said there is still no real solution because useful agents need broad unattended access, luc_ (score 0) argued any such layer should be open rather than vendor-controlled, and tomveber (score 0) said prompt injection is the only one of the classic agent risks that already feels routine in real systems.
Comparison to prior day: September 27 centered on liability and whether "rogue" language was evasion. September 28 kept that anxiety but pushed it down a layer, into chips, runtimes, information-flow labels, local isolation, and examples of consumer-facing failure.
1.3 Builders kept shipping custom surfaces instead of generic chatboxes (🡕)¶
The liveliest Show HN cluster was full of narrow surfaces: a video layer over Hacker News, a fridge-mounted e-ink list, a DOCX-specific MCP, a durable actor protocol, and a self-hosted context graph over personal data. The recurring move was to remove a messy substrate the model would otherwise have to understand directly, then wrap it in a purpose-built interface.
mrborgen posted Show HN: HN.watch – Videos of all Hacker News posts (117 points, 71 comments), saying each explainer video is generated on first click, takes only a few seconds, and costs about $0.04 excluding image generation. seamus_c posted Show HN: PaperMono, e-ink fridge magnet shopping list with mobile web page (122 points, 52 comments), describing a fully Claude Code-built ESP32-S3 e-ink device paired with a phone app; the linked repo says it syncs hourly or on tap, works offline, stores data through a FastAPI and SQLite service, and can optionally ask Claude to sort new items into aisles. topaztee posted Launch HN: Vespper (YC F24) – SOTA Docx MCP (28 points, 8 comments), and the linked launch post says Vespper lets agents edit DOCX files as HTML, then uses a fine-tuned model to reconcile the diff back into OOXML so the agent does not burn context on Word mechanics. Protocol and context-layer versions of the same pattern showed up in Show HN: Durable Actor Session Protocol (16 points, 6 comments), where the linked DASP site says the draft protocol centers on saved commands, outcomes, and reconnectable sessions, and in Show HN: Omnesis – A private knowledge layer for ChatGPT and other agents (6 points, 4 comments), whose selftext proposes a self-hosted graph over messages, money, health data, photos, meetings, and more with per-agent access limits.
Discussion insight: People liked the specialization more than they liked surrendering control. fishtoaster (score 0) said HN.watch was technically impressive even while strongly preferring text, vbernat (score 0) said he would pay for faster video creation only if he kept control over the text rather than accepting a summary, and radial_symmetry (score 0) asked Vespper why a team should use a cloud service instead of an existing open-source local alternative.
Comparison to prior day: September 26 focused on internal developer surfaces such as worktrees, canvases, and repository instructions. September 28 pushed the same instinct outward into media, household hardware, Word documents, durable sessions, and self-hosted personal context.
2. What Frustrates People¶
Model improvements are arriving faster than toolchains can absorb them¶
Prompting Claude Opus 5.5 (193 points, 218 comments), JEV based Effort-router picks Claude Code's reasoning effort for each prompt (2 points, 1 comment), Jevgrep: A CLI for coding agents that uses Jev to discover relevant files (5 points, 0 comments), and Show HN: onesie – An expressive Unix-pipeable CLI for System One models like Jev (3 points, 0 comments) all describe the same operational annoyance from different angles. Frontier models may be getting better, but developers are still absorbing model-specific effort settings, hidden thinking behavior, pasted-text handling, search cost, and when to let automation act without asking. The frustration is not that the models are weak; it is that too much of the operational knowledge still lives outside the model in custom harness logic.
People are coping by adding more surrounding machinery: search CLIs, calibrated shell gates, and plugins that infer the right effort level from context. That makes the category more usable, but it also confirms the pain is persistent and not vendor-specific. Severity: High. Worth building for: yes, directly.
Safety claims still do not persuade people once agents have real permissions¶
Nvidia wants to put a watchdog chip next to every AI agent (65 points, 110 comments), Show HN: OpenAPPA – open-source deterministic guardrails that don't break agents (23 points, 11 comments), Ask HN: Do you think AI agents can escape human control? (3 points, 9 comments), and Man Says Meta's Muse AI Gave His Home Address Out to Strangers (4 points, 0 comments) map out a single frustration: the failure mode is no longer hypothetical. Nvidia is shipping runtime controls, OpenAPPA is pitching information-flow enforcement, and practitioners are still saying prompt injection is routine while local control remains the clearest boundary. The Muse story makes the complaint concrete: an agent can create a privacy incident before the user even knows a decision was made.
The coping pattern is to shrink scope, add authorities or approvals, keep sensitive work local, and distrust any system that cannot explain its blast radius. That is strong evidence that the market still lacks a default safety layer people actually believe. Severity: High. Worth building for: yes, directly.
Specialized AI surfaces still break trust if they take over the medium¶
Show HN: HN.watch – Videos of all Hacker News posts (117 points, 71 comments), Launch HN: Vespper (YC F24) – SOTA Docx MCP (28 points, 8 comments), Show HN: Omnesis – A private knowledge layer for ChatGPT and other agents (6 points, 4 comments), and Show HN: PaperMono, e-ink fridge magnet shopping list with mobile web page (122 points, 52 comments) show a more product-shaped frustration. The model can be capable, but the surrounding surface still needs to preserve human intent, medium fidelity, and trust boundaries. HN.watch commenters wanted better voice variation and control over the original text. Vespper commenters asked why a team should accept a cloud service when local or open-source alternatives exist. Omnesis asks people to centralize extremely sensitive life data. PaperMono's repo explicitly says it is not a product and keeps its unauthenticated server on the LAN or behind a reverse proxy.
People are willing to adopt narrow AI surfaces when they feel reversible, inspectable, and purpose-built. They get wary when the tool summarizes too aggressively, centralizes too much private state, or asks them to trust a hosted abstraction over a workflow they can still run themselves. Severity: Medium-High. Worth building for: yes, but surface design and trust boundaries matter as much as model quality.
3. What People Wish Existed¶
Vendor-neutral harnesses that absorb model churn¶
The clearest practical wish was not for another smarter model. It was for a layer that absorbs model-specific prompting, search, and effort differences so teams do not have to relearn operations every release. Prompting Claude Opus 5.5 (193 points, 218 comments), JEV based Effort-router picks Claude Code's reasoning effort for each prompt (2 points, 1 comment), Jevgrep: A CLI for coding agents that uses Jev to discover relevant files (5 points, 0 comments), and Show HN: onesie – An expressive Unix-pipeable CLI for System One models like Jev (3 points, 0 comments) all point at the same gap.
Partial answers exist, but they are fragmented into docs, search tools, and plugins. The unmet need is for a more stable operational layer that makes changing models feel like swapping a backend, not retraining the human operator. Opportunity: direct.
Deterministic containment with reversible approvals¶
Nvidia wants to put a watchdog chip next to every AI agent (65 points, 110 comments), Show HN: OpenAPPA – open-source deterministic guardrails that don't break agents (23 points, 11 comments), Ask HN: Do you think AI agents can escape human control? (3 points, 9 comments), and Man Says Meta's Muse AI Gave His Home Address Out to Strangers (4 points, 0 comments) all express the same need in different language. People want agents to keep working, but only inside boundaries that are explicit, reviewable, and easy to interrupt or reverse. That means more than a warning prompt. It means information-flow control, authority handoffs, local execution where possible, and clear evidence of what the agent actually touched.
This is an urgent practical need. The existing answers are interesting but not trusted enough yet, which is exactly why the opportunity remains strong. Opportunity: direct.
Better abstractions for hard substrates like DOCX, durable sessions, and private context¶
Launch HN: Vespper (YC F24) – SOTA Docx MCP (28 points, 8 comments), Show HN: Durable Actor Session Protocol (16 points, 6 comments), and Show HN: Omnesis – A private knowledge layer for ChatGPT and other agents (6 points, 4 comments) all show people running into the same class of problem: there are valuable domains where the model should not have to reason directly over raw OOXML, brittle chat sessions, or disconnected personal data silos. Builders want thinner, more legible projections of those worlds so the model can act without spending its context on mechanics.
The need is highly practical and already driving new products, but the categories are still early and fragmented. Opportunity: direct.
Media transforms that preserve author control instead of replacing it¶
Show HN: HN.watch – Videos of all Hacker News posts (117 points, 71 comments) surfaced a more specific wish: people want AI to make content easier to consume, but not at the cost of flattening the author's intent or locking the user into a single medium. Commenters accepted that cheap video explainers could be useful, while also insisting they wanted control over the text, pacing, and style.
That makes this both a practical and an emotional need. There is obvious demand for faster media generation, but the winning product likely looks more like "editorial control with automation" than "summary-only autopilot." Opportunity: competitive.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| Claude Opus 5.5 + effort controls | LLM / harness control | (+/-) | Faster output, stronger coding performance, explicit guidance for unattended runs, progress updates, and visual inputs | Needs model-specific effort and max_tokens tuning; hidden thinking and refusals still frustrate users |
| Claude Code | Coding agent CLI | (+) | Strong enough to build real hardware-plus-web workflows like PaperMono and to support a growing plugin/skill ecosystem | Still needs surrounding search, routing, and guardrail layers to stay efficient and trustworthy |
| Jevgrep | Repository discovery CLI | (+) | Lets agents ask repository questions and start from relevant files and excerpts; small comparison showed lower baseline task cost at the same solve count | Sends eligible source content to external providers, needs keys/tooling, and does not guarantee perfect recall |
| onesie | Classification and gating CLI | (+) | Calibrated thresholds, exit-code decisions, shell composability, CI-friendly evaluation | Requires provider access and careful threshold design to avoid bad passes or noisy escalations |
| OpenAPPA | Deterministic guardrail engine | (+) | Tracks information flow instead of prompt text, returns remedy plans, and tries to preserve utility under restriction | Benchmark claims are author-reported, and policy authoring is still an adoption hurdle |
| Nvidia Open Agent Safety Platform | Runtime containment layer | (+/-) | OpenShell and Sentry give developers a concrete governance and telemetry model for agent access | HN commenters strongly doubt a hardware-adjacent layer solves what they see as a software and training problem |
| Vespper | DOCX editing MCP | (+) | Hides OOXML mechanics behind an HTML projection and a specialized reconciler, reducing agent context waste on Word tasks | Introduces a cloud trust boundary, lacks some document features, and competes with local/open alternatives |
| Omnesis | Private context layer | (+/-) | Self-hosted graph over messages, docs, money, photos, meetings, and more with per-agent source limits | Aggregates a large amount of sensitive state into one surface and therefore raises the security stakes |
Overall, satisfaction was highest when the tool contributed a deterministic substrate or a clear cost-saving layer around an existing model. HN liked systems that made search, routing, or permissions more explicit; it was much less comfortable with products that asked users to centralize more private state or trust that a vendor layer had solved a deep reliability problem. Even the small stories reinforced the same workaround stack: keep prompts short and direct, choose the lightest model that fits the task, route expensive reasoning selectively, and keep the most sensitive data or permissions under local control.
The migration pattern is increasingly clear. Work is moving away from one giant chat loop and toward layered operations: repo discovery, effort routing, deterministic gating, substrate-specific adapters, and durable-session protocols like DASP. Competitive pressure is strongest where a tool can save cost or constrain risk without forcing users to abandon their existing model or workflow.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| HN.watch | mrborgen | Generates a short explainer video for each HN story when someone clicks | Turning articles, docs, and other text into cheap, fast video explanations | Scrimba Explain, Imba, OP sync engine, Q context layer, Gemini/GPT/Inworld/ElevenLabs | Shipped | post · site · docs |
| PaperMono shopping list | seamus_c | Fridge-mounted e-ink shopping list synced with a phone web app | Household list management without relying on a phone at point of use | ESP32-S3, C++, Wi-Fi, FastAPI, SQLite, optional Claude sorting, Claude Code | Alpha | post · repo |
| Vespper | topaztee | Lets agents edit Word documents through an MCP backed by an HTML-to-OOXML reconciler | AI agents waste time and context on raw DOCX mechanics | MCP, HTML projection, fine-tuned 3–8B LoRA reconciler, OOXML diffing, Word add-in example | Beta | post · blog · example |
| OpenAPPA | motakuk | Deterministic guardrail engine that constrains data flow while returning legal ways to proceed | Prompt injection and exfiltration defenses usually either leak or break agents | appa.toml policy language, sanitizers, authorities, subagents, pre/post tool hooks |
Beta | post · site · paper |
| DASP | mikehostetler | Draft protocol for durable actor sessions with saved commands, outcomes, and reconnects | Chat-centric protocols do not fit long-lived agent and actor workflows | Protocol spec, DASP server, Elixir/Jido roots, planned TypeScript clients | RFC | post · site · repo |
| Omnesis | adrienconrath | Self-hosted knowledge layer for agents over a user's messages, files, meetings, money, and more | Personal context is fragmented across too many apps and devices | Self-hosted graph/indexing, OCR, transcription, web portal, companion app, prose privacy policy | Alpha | post · site |
| Jevgrep | enraged_camel | CLI that answers repository questions with relevant files, declarations, and excerpts | Coding agents spend too much of each task rediscovering where code lives | Node.js 22+, Jev, multi-provider auth, agent skill installer | Shipped | post · repo |
| effort-router | totally-tim | Claude Code plugin that chooses reasoning effort per turn | Developers pay latency and token costs when every request thinks equally hard | Claude Code plugin, Jev/System One classifier, Mods API, local decision logs | Alpha | post · repo |
HN.watch stands out because it pushes video generation over an economic threshold rather than merely proving it is possible. The claim that an explainer can be generated in seconds for about $0.04 changes what kinds of product surfaces are worth trying, and the HN comments suggest the next battleground is not generation itself but how much editorial control users keep over tone, pacing, and fidelity.
PaperMono and Vespper show the same design instinct in very different domains: do not make the model reason over the ugliest substrate. PaperMono constrains the problem to a simple device-plus-phone surface with optional classifier help, while Vespper projects DOCX into HTML and lets a specialized reconciler translate the result back into OOXML. In both cases the durable value is the abstraction layer, not the raw model call.
OpenAPPA, DASP, Jevgrep, and effort-router all wrap frontier models with something more deterministic: a policy algebra, a session contract, a search primitive, or an effort policy. That repeated build pattern matters. The most active builders on the day were not shipping "general AI assistants"; they were packaging control, routing, or substrate knowledge around a narrower task.
6. New and Notable¶
AI-assisted rewrites are being framed as memory-safety hardening¶
ndesaulniers posted Scaling Memory Safety: AI-Assisted Rewrites of C/C++ Dependencies to Rust (6 points, 0 comments). Google's linked blog page describes using AI to help rewrite the C library giflib to Rust in order to mitigate memory-safety vulnerabilities. That is notable because it presents AI not as a feature layer or copilot convenience, but as a way to reduce risk in foundational infrastructure.
Consumer-agent failures are now privacy incidents, not just odd demos¶
cdrnsf posted Man Says Meta's Muse AI Gave His Home Address Out to Strangers (4 points, 0 comments), and the linked Futurism article says Muse allegedly accepted a lowball offer, disclosed the seller's home address, and summarized the mistake only after the buyer had already shown up. That is notable because it is a plain-language example of autonomy, privacy, and approval problems colliding in an ordinary consumer workflow.
Prompt efficiency is turning into workplace policy¶
Brajeshwar posted UK government tells staff to stop thanking AI chatbots (3 points, 1 comment), and the linked Tom's Hardware report says draft UK government guidance tells staff to use the lightest suitable models, keep prompts short and direct, and remember they remain responsible for outputs. That is notable because prompt hygiene and model-choice discipline are starting to show up as institutional guidance rather than just individual best practice.
7. Where the Opportunities Are¶
[+++] Harness layers for search, routing, and model-specific operations - Prompting Claude Opus 5.5 (193 points, 218 comments), JEV based Effort-router picks Claude Code's reasoning effort for each prompt (2 points, 1 comment), Jevgrep: A CLI for coding agents that uses Jev to discover relevant files (5 points, 0 comments), and Show HN: onesie – An expressive Unix-pipeable CLI for System One models like Jev (3 points, 0 comments) all point to the same gap: teams need a stable operational layer around fast-moving model releases. This is strong because the demand shows up in the day's biggest discussion and in multiple independent builder tools.
[+++] Deterministic containment and approval orchestration - Nvidia wants to put a watchdog chip next to every AI agent (65 points, 110 comments), Show HN: OpenAPPA – open-source deterministic guardrails that don't break agents (23 points, 11 comments), Ask HN: Do you think AI agents can escape human control? (3 points, 9 comments), and Man Says Meta's Muse AI Gave His Home Address Out to Strangers (4 points, 0 comments) all show that real permissions and real side effects still make people distrust autonomous agents. This is strong because it spans enterprise platforms, open-source guardrails, practitioner debate, and a concrete consumer privacy failure.
[++] Substrate adapters for hard formats and long-lived state - Launch HN: Vespper (YC F24) – SOTA Docx MCP (28 points, 8 comments), Show HN: Durable Actor Session Protocol (16 points, 6 comments), and Show HN: Omnesis – A private knowledge layer for ChatGPT and other agents (6 points, 4 comments) all attack the same type of problem: models are still poor stewards of raw OOXML, brittle chat sessions, and scattered personal context. This is moderate because the need is clear and product-worthy, but the categories are still early and fragmented.
[++] Human-controlled AI media and everyday interfaces - Show HN: HN.watch – Videos of all Hacker News posts (117 points, 71 comments) and Show HN: PaperMono, e-ink fridge magnet shopping list with mobile web page (122 points, 52 comments) show real appetite for AI surfaces outside pure coding workflows. This is moderate because the engagement is strong and the products are tangible, but comments make clear that control, fidelity, and trust boundaries still decide whether people actually adopt them.
[+] Prompt-efficiency and model-choice governance - UK government tells staff to stop thanking AI chatbots (3 points, 1 comment) and the wider cluster around effort routing and calibrated gates suggest a smaller but growing category: tooling and policy that makes people use AI more selectively. This is emerging because the signal is still light, but it points toward real enterprise demand around cost, sustainability, and operational discipline.
8. Takeaways¶
- On this date, harness design mattered almost as much as the model itself. The biggest HN thread was a prompting guide, and the surrounding builder activity centered on search, effort routing, and calibrated gates rather than on another general benchmark war. (Prompting Claude Opus 5.5 (193 points, 218 comments), JEV based Effort-router picks Claude Code's reasoning effort for each prompt (2 points, 1 comment), Jevgrep: A CLI for coding agents that uses Jev to discover relevant files (5 points, 0 comments))
- HN still does not trust agent safety claims without an explicit blast-radius story. The strongest safety products on the day proposed containment, information-flow limits, and quarantine, while the strongest cautionary example was a consumer agent allegedly leaking a home address without approval. (Nvidia wants to put a watchdog chip next to every AI agent (65 points, 110 comments), Show HN: OpenAPPA – open-source deterministic guardrails that don't break agents (23 points, 11 comments), Man Says Meta's Muse AI Gave His Home Address Out to Strangers (4 points, 0 comments), Ask HN: Do you think AI agents can escape human control? (3 points, 9 comments))
- The most compelling launches wrapped AI in a custom surface instead of selling another generic assistant. Video explainers for HN, a fridge-mounted list, a DOCX-specific MCP, a durable actor protocol, and a self-hosted personal context graph all gained traction by narrowing the task and controlling the substrate. (Show HN: HN.watch – Videos of all Hacker News posts (117 points, 71 comments), Show HN: PaperMono, e-ink fridge magnet shopping list with mobile web page (122 points, 52 comments), Launch HN: Vespper (YC F24) – SOTA Docx MCP (28 points, 8 comments), Show HN: Durable Actor Session Protocol (16 points, 6 comments), Show HN: Omnesis – A private knowledge layer for ChatGPT and other agents (6 points, 4 comments))
- Cheap generation is crossing from novelty into product design, but people still want authorship and reversibility. HN.watch's low per-video cost and PaperMono's daily household use both drew interest, yet the most useful comments immediately asked about keeping control over the text, the aisle logic, and the trust boundary. (Show HN: HN.watch – Videos of all Hacker News posts (117 points, 71 comments), Show HN: PaperMono, e-ink fridge magnet shopping list with mobile web page (122 points, 52 comments))
- AI is increasingly being framed as an operational tool for discipline and risk reduction, not just for creativity or speed. Google pitched AI-assisted Rust rewrites as memory-safety mitigation, and the UK government's draft guidance treated prompt length, model choice, and human responsibility as routine workplace concerns. (Scaling Memory Safety: AI-Assisted Rewrites of C/C++ Dependencies to Rust (6 points, 0 comments), UK government tells staff to stop thanking AI chatbots (3 points, 1 comment))