Daily Brief - 2026-09-17

From 37 items, 14 important content pieces were selected

Nvidia Announces Native GPU Programming in Rust via CUDA ⭐️ 8.0/10
4B model generates query plans 81% faster than Postgres ⭐️ 7.0/10
Xiaomi Launches Live Post-Training Dashboard for MiMo 2.6 ⭐️ 7.0/10
New Method Pushes Ternary LLM Storage Below 1.58 Bits ⭐️ 7.0/10
Backups Aren’t Simple: A Deep Dive Into Backup Complexity ⭐️ 7.0/10
Anthropic Merges Claude Cowork and Chat Into One Unified Claude ⭐️ 7.0/10
ZGCM-1: A Fully Open 7B Model for Math and Agentic Search ⭐️ 7.0/10
CNSF: Causal Neural Set Filtering for Efficient Multi-Target Tracking ⭐️ 7.0/10
Few-Shot Degradation Is Task-Dependent, Not Representational Distortion ⭐️ 7.0/10
The Functionalizer: Lossless Functional Decomposition for Subword Tokenization ⭐️ 7.0/10
Blog Post on Small Programming Tricks Sparks Hacker News Debate ⭐️ 6.0/10
OpenSpec: A Lightweight Spec Framework for LLM Coding Agents ⭐️ 6.0/10
Datasette 1.0a40 adds background task API and security fix ⭐️ 6.0/10
Datasette 0.65.5 Patches Trailing Newline Permission Bypass ⭐️ 6.0/10

Nvidia Announces Native GPU Programming in Rust via CUDA ⭐️ 8.0/10

Nvidia has officially announced support for writing native GPU kernels in Rust through its CUDA platform, introducing two tracks for Rust-based GPU kernel development. The announcement was made on Nvidia’s developer blog and quickly sparked a lively Hacker News discussion with 293 points and 115 comments. This is a significant development for both the Rust and GPU computing ecosystems, as it brings Rust’s safety and zero-cost abstractions to Nvidia’s dominant CUDA platform. It could accelerate Rust adoption in high-performance computing and AI workloads, while also reigniting debates about CUDA’s proprietary vendor lock-in. The announcement outlines two tracks for writing GPU kernels in Rust via CUDA, though the article’s LLM-generated writing style drew criticism from commenters. Rust’s expressive type system and zero-cost abstractions are highlighted as key advantages for writing high-level, reusable GPU code without sacrificing performance.

hackernews · nonmaskable · Sep 16, 11:15 · Discussion

Background: CUDA is Nvidia’s proprietary GPU computing platform and programming model that exposes hardware-level parallel execution to software, but it only works on Nvidia GPUs, creating vendor lock-in. Historically, GPU programming has relied on specialized languages like HLSL, GLSL, MSL, and Triton, while projects like Rust GPU (from Embark Studios) have been working to make Rust a first-class language for GPU shaders and compute. Rust is a systems programming language known for memory safety and performance, making it an attractive candidate for GPU kernel development.

References:

Discussion: Commenters expressed strong opinions on CUDA’s proprietary nature, with some arguing it leads to vendor lock-in and #ifdef hell, and advocating for separate kernel files and manual launches as in Metal, OpenCL, and D3D12. Others welcomed the move, noting Hugging Face’s Candle crate for Rust inference and the excitement of working with technology that LLMs haven’t yet been trained on. Several commenters criticized the article’s LLM-generated style, saying it reads like Claude rather than typical Nvidia posts.

Tags: #Rust, #GPU Programming, #CUDA, #Nvidia, #Hacker News

4B model generates query plans 81% faster than Postgres ⭐️ 7.0/10

Rohan Bansal trained a 4-billion-parameter model, called QORL, that produces Postgres query plans claimed to be 81% faster, with a 44.7% latency reduction across 113 join-heavy queries. The model was initially unable to produce a plan for 99 of those queries before training. If learned query planners can reliably beat cost-based optimizers, it could reshape how databases handle complex joins and reduce reliance on manual hints or index tuning. However, the result is a research prototype, and the community debate highlights how far it is from production-grade reliability. The benchmark used an 8 GB in-memory dataset with shared_buffers constrained to a fraction of that, warmed queries, read-only SELECTs, and no indexes beyond primary keys. The author also designed a custom GRPO variant to score RL rollouts in a noisy measurement environment.

hackernews · polyphilz · Sep 16, 18:50 · Discussion

Background: Postgres uses a cost-based query planner that estimates the cheapest execution plan for a SQL query, but its heuristics can be suboptimal for complex joins or correlated columns. Recent research has explored learned query optimization, using machine learning to replace or augment these heuristics, though practical adoption remains limited.

References:

Discussion: HN commenters were skeptical: they noted the benchmark’s unrealistic setup (in-memory, no indexes, read-only) and worried about overfitting, hallucination risks, and whether LLMs are the right tool compared to algorithmic or neural heuristics. Some argued that missing statistics, not planner limitations, are the usual cause of bad plans, and that hints often paper over deeper problems.

Tags: #databases, #query-optimization, #LLM, #Postgres, #benchmarking

Xiaomi Launches Live Post-Training Dashboard for MiMo 2.6 ⭐️ 7.0/10

Xiaomi launched a live public dashboard at mimo.xiaomi.com/rl/ that shows the post-training and reinforcement learning progress of its MiMo 2.6 model in real time. The page drew 265 points and 66 comments on Hacker News, with users reporting hands-on experience using earlier MiMo versions. Publishing a live training dashboard is a rare transparency move in frontier AI development, where labs typically keep training runs and RL environment scores secret. If other developers follow suit, it could shift norms around how openly near-frontier models are built and evaluated. The dashboard focuses on post-training and RL progress rather than pretraining, and community members noted that MiMo-V2.5-Pro scored only 19% on the DeepSWE 1.1 benchmark, far behind Fable (70%), Kimi K3 (69%), and Astra (74%). Users also reported that MiMo-V2.5 is very cost-effective for software engineering work, though it occasionally falls into hallucination loops.

hackernews · krackers · Sep 16, 20:09 · Discussion

Background: Large language models are typically built in two stages: pretraining on massive text corpora, followed by post-training that refines the model using instruction data, preference comparisons, and reinforcement learning (often RLHF). Reinforcement learning from human feedback uses reward signals to align model outputs with human preferences, and it is a key part of turning a base model into a usable assistant. Xiaomi’s MiMo family includes open-sourced mixture-of-experts models such as MiMo-V2-Flash, which has 309 billion total parameters and 15 billion active parameters.

References:

Discussion: Commenters largely praised the transparency, with one asking why IBM or Google don’t do the same for Granite or Gemini. Others shared firsthand experience that MiMo-V2.5 is powerful and extremely cheap for coding work, while some framed open-source frontier AI as a threat to closed labs’ business models.

Tags: #AI/ML, #LLM, #open-source, #model-training, #transparency

New Method Pushes Ternary LLM Storage Below 1.58 Bits ⭐️ 7.0/10

A new paper presents a method that reduces ternary LLM weight storage from the theoretical 1.58 bits per weight to 1.48 bits by exploiting the fact that actual ternary weights are zero about 51% of the time. This is achieved through a presence-bitmap packing scheme that skips storing the zero values. This reduction could significantly shrink ternary LLMs for embedded systems and on-device inference, making them more portable and enabling custom silicon or ASIC designs with record power efficiency. It matters for the broader trend of efficient LLM deployment, where memory footprint and hardware acceleration are key constraints. The method exploits sparsity in ternary weights, which are zero 51% of the time, to pack weights more densely than the log2(3) ≈ 1.58-bit limit. However, the approach may face trade-offs in decoding complexity and hardware support, and some community members question whether ternary quantization is the best choice compared to vector quantization or trellis-based methods.

hackernews · matt_d · Sep 16, 20:59 · Discussion

Background: Ternary LLMs, also known as 1.58-bit models, use weights restricted to three values: -1, 0, and +1. This reduces memory footprint and allows multiplication to be replaced by addition, making inference more efficient. The name ‘1.58-bit’ comes from log2(3) ≈ 1.58 bits of information per weight. Microsoft’s BitNet b1.58 demonstrated that such models can match full-precision counterparts up to several billion parameters.

References:

Discussion: Community reactions are mixed: some find the sparsity exploitation ‘neat’ and see potential for shockingly efficient custom silicon and embedded deployment, while others are skeptical, arguing that ternary quantization is suboptimal compared to vector quantization or trellis-based methods. One commenter suggests arithmetic coding could squeeze out even more centi-bits, and another notes that quantization-aware training may require ~30% more weights for comparable quality.

Tags: #ternary-llms, #quantization, #model-compression, #efficient-inference, #hardware-acceleration

Backups Aren’t Simple: A Deep Dive Into Backup Complexity ⭐️ 7.0/10

A technical article titled “Backups Aren’t Simple” was published on filipovski.net, arguing that backup systems are deceptively complex despite appearing straightforward. The piece sparked a substantive Hacker News discussion with 90 upvotes and 36 comments, where engineers shared real-world data loss stories and concrete tooling recommendations. Backups are a foundational concern for anyone running infrastructure, yet the article and discussion highlight that many engineers underestimate the complexity until they actually need to restore data. The conversation reinforces that the real goal is restoration, not just making copies, which affects how teams design and validate their data protection strategies. Commenters highlighted specific tools and approaches: Restic combined with Backrest for a 3-2-1-style setup across three hosts, and a minimal tar | zstd | gpg pipeline using GNU tar’s --listed-incremental index (the .snar file) to avoid complex repository formats. Others noted vendor-specific solutions like ReaR RPM for Oracle Linux and the distinction that enterprise backup vendors consider themselves in the “restoration business.”

hackernews · afilipovski · Sep 16, 20:27 · Discussion

Background: The 3-2-1 backup rule is a widely recommended strategy: keep at least three copies of data, on two different media types, with one copy stored offsite. Tools like Restic and Borg use their own repository formats that support deduplication and incremental snapshots, while simpler approaches like tar archives remain portable but require manual handling of incremental state. The discussion reflects a broader industry tension between convenience, portability, and reliable restoration.

References:

Discussion: The Hacker News discussion was rich with practical experience: one commenter recounted four personal data loss incidents, including a lightning strike that fried a fax modem and a OneDrive terms change. A key insight came from a commenter whose Veritas-employed friend corrected him: “We are not in the backup business. We are in the restoration business.” Others shared concrete setups, including Restic + Backrest for a three-host homelab and a minimal tar/zstd/gpg pipeline, showing a preference for simple, portable, verifiable solutions.

Tags: #backups, #data-recovery, #infrastructure, #devops, #hacker-news

Anthropic Merges Claude Cowork and Chat Into One Unified Claude ⭐️ 7.0/10

Anthropic announced that Claude Cowork and Claude chat are merging into a single Claude product, rolling out first to Pro and Max plans across web, desktop, and mobile over the coming weeks. The unified Claude can handle both quick questions and long-running delegated tasks, continuing work even after the user closes their laptop. The consolidation positions Claude as a general-purpose agent rather than a chatbot plus separate agentic tool, clarifying a product lineup that had grown confusing with Cowork, Claude, and Claude Code. It mirrors OpenAI’s recent move of renaming its Codex desktop app to ChatGPT, signaling that major AI vendors are converging on unified agentic assistants. The rollout begins with Pro and Max subscribers and will reach existing and new users on those plans across the Claude app on web, desktop, and mobile in the coming weeks. Commentator Simon Willison notes that figuring out what the merge actually means in terms of features and surfaces will still take considerable work, and he had planned a follow-up to his piece on Understanding ChatGPT Work.

rss · Simon Willison · Sep 16, 18:09

Background: Claude is Anthropic’s family of large language models, first released as a chatbot in March 2023, with model tiers named Haiku, Sonnet, and Opus. Anthropic also sells agentic tools: Claude Code, a terminal-based coding agent, and Claude Cowork, a similar tool aimed at non-programmers that can access user folders on macOS to read, edit, and create files and perform office tasks asynchronously. A general-purpose AI agent is an assistant that can handle a wide range of tasks across domains, from writing and research to coding and data analysis, and can take autonomous actions.

References:

Discussion: The item was surfaced via Hacker News, and the author’s framing suggests a mix of relief at reduced product confusion and skepticism that the practical boundaries of the merged product will remain hard to pin down. No detailed comment sentiment was provided in the source content.

Tags: #Anthropic, #Claude, #AI agents, #product update, #Simon Willison

ZGCM-1: A Fully Open 7B Model for Math and Agentic Search ⭐️ 7.0/10

Researchers released ZGCM-1, a fully open 7B dense foundation model trained from scratch with a 256K context window, combining internal reasoning with external tool use. The team open-sourced weights from pre-training, mid-training, and post-training stages, intermediate checkpoints, training code, per-stage data recipes, and W&B logs. It shows a compact 7B model can stay competitive with frontier models orders of magnitude larger on math reasoning and agentic search, while its full openness gives the community a rare end-to-end reproducible training recipe. The claimed ~4.2x efficiency gain in 16K pre-training time-to-loss could lower the cost of building capable small models. The architecture interleaves gated sliding-window attention with full attention and uses a stable FP8 Muon optimizer, while training scales context progressively through 16K, 64K, and 256K stages and reformulates interaction traces as Markov Decision Processes. Evaluation results in the preprint are truncated, and no community discussion is available yet, so the claims still need independent validation.

rss · arXiv cs.AI · Sep 16, 04:00

Background: Sliding-window attention limits each token to a local window to cut compute, while full attention lets tokens attend globally; interleaving the two aims to balance efficiency and long-range recall. Muon is an optimizer that has shown faster convergence than AdamW but traditionally keeps FP32 state, so running it in FP8 is a notable systems challenge. Markov Decision Processes are the standard mathematical framework of states, actions, and rewards used to model sequential decision making, which the authors apply to agent interaction traces.

References:

Tags: #foundation-models, #efficient-training, #long-context, #agentic-search, #open-source-ai

CNSF: Causal Neural Set Filtering for Efficient Multi-Target Tracking ⭐️ 7.0/10

Researchers introduced Causal Neural Set Filtering (CNSF), a neural set filter for online multi-target tracking that encodes only current measurements while carrying past evidence in a structured recursive track state. On a held-out three-regime simulated test set, CNSF reduced mean GOSPA and T-GOSPA by 19.3% and 30.4% versus Track-MT3, with 55.9% fewer parameters and a 3.76x speedup in single-thread CPU inference. Transformer-based trackers like MT3 and Track-MT3 repeatedly re-encode measurement windows, causing redundant computation that limits real-time deployment. CNSF’s recursive design shows that combining classical filtering structure with neural components can deliver both better accuracy and substantially lower compute, which matters for online tracking in robotics, autonomous driving, and surveillance. CNSF combines exclusive Sinkhorn association, association-conditioned Kalman-shaped updates with moment matching, and recurrent Bernoulli lifecycle modeling with measurement-driven birth, imposing soft one-to-one constraints and propagating association-induced state uncertainty. The code is available at https://github.com/daihuangyu/CNSF, and results are reported on a simulated three-regime test set rather than real-world benchmarks.

rss · arXiv cs.LG · Sep 16, 04:00

Background: Multi-target tracking (MTT) aims to estimate the number and states of multiple moving objects over time from noisy sensor measurements, requiring both data association (deciding which measurement belongs to which target) and state estimation. Transformer-based trackers such as MT3 jointly learn these tasks but re-encode measurement windows at each step, which is computationally wasteful. Classical approaches like the Bernoulli filter, rooted in random finite set (RFS) statistics, recursively propagate a probabilistic track state and naturally handle missed detections and target birth/death. GOSPA and its trajectory variant T-GOSPA are standard metrics that jointly penalize localization error, missed targets, and false tracks.

References:

Tags: #multi-target tracking, #neural set filtering, #transformer efficiency, #state estimation, #data association

Few-Shot Degradation Is Task-Dependent, Not Representational Distortion ⭐️ 7.0/10

A new arXiv paper evaluates 12 open-weight models on two Ukrainian tasks—news classification and legal case outcome prediction—and finds that few-shot prompting gains +24 pp on news but only +3.4 pp on legal text, with two models actually degrading. The authors propose a length-matched random-text control to isolate the representational shift caused by demonstration content rather than prompt length, yielding a ‘content delta’ metric that correlates with few-shot benefit (rho = +0.65, p = 0.043), whereas raw shift does not (r = 0.20). This work challenges the intuitive ‘distortion’ explanation of few-shot degradation and offers a simple methodological fix that could reshape how researchers measure in-context learning effects. Practitioners relying on few-shot prompting should note that benefits are highly task-dependent, and that representation-shift metrics must control for prompt length to be meaningful. The proposed content delta subtracts the shift caused by length-matched random text from the raw zero-shot-to-few-shot shift, revealing that models which restructure representations more from demonstration content benefit more—the opposite of the distortion hypothesis. A causal check via masking demonstrations in Llama 3.3 70B recovers accuracy above the zero-shot baseline, confirming the finding.

rss · arXiv cs.CL · Sep 16, 04:00

Background: Few-shot prompting, popularized by GPT-3, lets language models adapt to new tasks by conditioning on a few input-output examples in the prompt without updating weights. Prior work measured how hidden states shift between zero-shot and few-shot modes, but few-shot prompts are much longer, and length alone moves representations, confounding earlier analyses. This paper introduces a length-matched random-text control to isolate the effect of demonstration content from prompt length.

References:

Tags: #few-shot learning, #in-context learning, #representation analysis, #language models, #prompting

The Functionalizer: Lossless Functional Decomposition for Subword Tokenization ⭐️ 7.0/10

A new arXiv preprint (2609.15991) introduces the Functionalizer, a lossless pre-tokenizer that decomposes orthographic variations such as casing, diacritics, and character repetition into reversible opcode/operand prefix streams encoded in the Unicode Private Use Area before standard subword tokenization. Across six natural language and code corpora, it achieves complete corpus coverage with up to 16% fewer vocabulary slots, and preliminary tests on 25M-parameter GPT-2 scale models show drastically improved code syntax validity while maintaining similar prose coherence. Subword tokenizers like BPE and WordPiece typically treat orthographic variants (e.g., hello, Hello, HELLO) as unrelated tokens, fragmenting the embedding space, or discard them via lossy normalization. The Functionalizer’s lossless, reversible decomposition offers a vocabulary-efficient and structurally aware alternative that could improve tokenization efficiency and embedding quality, particularly for code and multilingual text. The framework introduces fully reversible operators for casing (CAPITALIZE), diacritics (13 dedicated opcodes), and character repetition (REPEAT, MULTIREPEAT), encoded as Unicode Private Use Area characters. A key tradeoff is domain-dependent: it compresses indentation-heavy code sequences but inflates natural-language prose sequences, and the downstream evaluation is limited to 25M-parameter GPT-2 scale models, motivating further validation at production scale.

rss · arXiv cs.CL · Sep 16, 04:00

Background: Subword tokenization algorithms such as Byte Pair Encoding (BPE), Unigram, and WordPiece split text into units between words and characters, keeping the vocabulary compact while capturing meaningful pieces. Pre-tokenizers are the initial step that normalizes or splits raw text before the main tokenizer is applied. The Unicode Private Use Area consists of code points intentionally left undefined so third parties can assign their own characters without conflicting with standard Unicode assignments.

References:

Tags: #tokenization, #NLP, #subword, #lossless, #embeddings

Blog Post on Small Programming Tricks Sparks Hacker News Debate ⭐️ 6.0/10

Will Keleher published a blog post titled ‘Small programming tricks matter’ compiling small programming and command-line tricks, which reached the front page of Hacker News with 405 points and 181 comments. The post and its discussion highlight productivity shortcuts that many developers overlook. The enthusiastic discussion shows that even experienced developers value practical, incremental productivity tips, and it underscores a broader gap in how people learn to use their daily tools efficiently. The debate also touches on whether such tricks should be called ‘programming’ or ‘computing’ tricks, reflecting differing views on skill categorization. The blog post itself is a collection of small tricks, and the Hacker News thread adds many more, such as using Ctrl+r with fzf for shell history, leveraging AI assistants to discover unfamiliar commands like perf, and a gist for navigating back to exact directories. Commenters note that adopting these tricks requires deliberate habit formation, as people often default to inefficient methods like arrow-key scrolling.

hackernews · signa11 · Sep 16, 15:56 · Discussion

Background: Command-line tricks are shortcuts or lesser-known features of shells (like Bash or Zsh) and utilities that can speed up common tasks. Tools like fzf (a fuzzy finder) and Zoxide (a smarter cd command) are popular among developers for enhancing navigation and history search. Hacker News is a well-known forum where technology enthusiasts share and debate such tips, often leading to valuable community-driven knowledge.

Discussion: Commenters shared additional tricks and debated the distinction between programming and computing tricks, with some arguing that many are general computing tips. A key theme was that adopting these tricks requires conscious habit-building, and one commenter suggested watching AI assistants to learn new commands. Another lamented that most people use computers inefficiently, proposing better training over AI agents.

Tags: #programming, #productivity, #command-line, #tips, #hacker-news

OpenSpec: A Lightweight Spec Framework for LLM Coding Agents ⭐️ 6.0/10

OpenSpec has launched as a lightweight, configurable framework for writing AI specifications, designed to keep teams and coding agents aligned on requirements before any code is written. It adds a spec layer on top of AI coding assistants so that requirements no longer live only in chat history. As LLM coding agents become common, spec-driven development (SDD) is emerging as a way to make agent output more predictable and reviewable, and OpenSpec offers a minimal, no-API-key entry point into that workflow. It matters most to developers already using tools like Codex or Claude Code who want more structure without heavy process overhead. OpenSpec requires no API keys and focuses on capturing intent in a spec, refining requirements, and verifying that implementation matches, rather than generating code itself. It is one of several competing SDD approaches, alongside tools like Kiro and GitHub’s spec-kit, and some evaluations suggest spec-driven agents can consume significantly more tokens due to the extra specification step.

hackernews · etoxin · Sep 16, 23:06 · Discussion

Background: Spec-driven development (SDD) flips the usual relationship between specs and code: instead of specs merely guiding implementation, the specification becomes the source that generates and constrains the implementation. AI coding assistants are powerful but unpredictable when requirements live only in chat history, so frameworks like OpenSpec add a lightweight spec layer to agree on what to build first. This category has grown quickly, with tools such as Kiro and GitHub’s spec-kit promoting similar ideas.

References:

Discussion: Hacker News commenters were skeptical about whether OpenSpec is necessary, with some saying they already get good results by having the agent maintain a checklist file in /tmp or a todo folder. Others noted that recent LLMs have become decent at planning on their own, while one user said they feed OpenSpec-generated task lists into a “Ralph loop” bash script to save tokens, and another asked how people evaluate the various spec frameworks.

Tags: #AI, #spec-driven development, #LLM, #developer tools, #framework

Datasette 1.0a40 adds background task API and security fix ⭐️ 6.0/10

Datasette 1.0a40 is an alpha release that ships the same security fix as version 0.65.5, introduces a new datasette.add_background_task() method letting plugins launch and manage background tasks, and migrates Datasette to the httpx2 HTTP client for internal calls such as datasette.client.get(). The background task API gives plugin authors a supported way to run long-running work without blocking requests, which matters as Datasette approaches a stable 1.0 release, while the security fix should also be applied by users on the 0.65.x line. The release also includes a large batch of bug fixes, many produced by a recent issue-triage effort aimed at the 1.0 stable release, and the httpx2 migration affects internal features like datasette.client.get() rather than the public HTTP interface.

rss · Simon Willison · Sep 16, 23:51

Background: Datasette is an open-source tool by Simon Willison for exploring and publishing SQLite databases as browsable websites and JSON APIs. Plugins extend its functionality, and datasette.client.get() lets plugins call Datasette’s own internal API without the overhead of a real HTTP request. httpx2 is a next-generation Python HTTP client from the Pydantic project supporting HTTP/1.1 and HTTP/2 with sync and async APIs.

References:

Tags: #datasette, #release, #security, #python, #background-tasks

Datasette 0.65.5 Patches Trailing Newline Permission Bypass ⭐️ 6.0/10

Datasette 0.65.5 was released as a security fix for an issue where a trailing newline in a requested table name could bypass table permissions and expose private rows. The vulnerability was reported by dpfkdlemtp in advisory GHSA-h547-rmjf-5m2m. This patch matters because Datasette is widely used by journalists and researchers to publish SQLite databases, and a permission bypass could leak data that was meant to remain private. Users running affected versions should upgrade promptly to close the exposure. The advisory notes that in the 1.0 alpha series, users with table creation and alteration permissions can also rename protected tables, extending the impact beyond the 0.65.x line. The fix is a routine point release rather than a feature change.

rss · Simon Willison · Sep 16, 23:51

Background: Datasette is an open-source tool that turns any SQLite database into a queryable, shareable website with full SQL query support, often used to publish data publicly. It includes a permissions system that controls which tables and rows a given user or token can access. A trailing newline in a table name is a subtle input-handling edge case that can cause permission checks to be evaluated against a different string than the one actually used to fetch data.

References:

Tags: #datasette, #security, #release, #permissions, #open-source