← Sep 17Friday, September 18, 2026 · all daysSep 19 → · today »
1.Astra for Law(openai.com)
578 points by vertigoruntime 26 days ago | 678 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↓Lawyers will still be needed as AI outputs require significant expert correction and validation
  • ~Legal field is heterogeneous; LLM impact varies drastically by practice area and shouldn't be generalized
  • ↓Skepticism about OpenAI's motives, data privacy claims, and marketing framing of the product
  • ↓Astra's benchmark performance is actually behind competing frontier models like Claude and Muse
  • ↓Law firm partnerships with AI labs will erode firms' competitive advantage and commoditize expertise
2.Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint(prismml.com)
576 points by JonSchneider 26 days ago | 197 comments | permalink
tl;dr: PrismML released Ternary Bonsai 2 27B, a compressed version of Qwen3.8 27B using ternary {−1, 0, +1} weights with FP16 scaling (1.76 bits/weight), shrinking the model to 5.9GB while retaining 98.2% of full-precision benchmark performance. It hits 143 tokens/sec on an RTX 5090 and 46.8 tokens/sec on M5 Max, supports 262K context and multimodal input, and runs via CUDA or MLX under Apache 2.0.
HN Discussion:
  • •Practical setup instructions for running the model locally with required forks
  • ↓The 'near-lossless' claim is overstated; model struggles on real tasks
  • ↓Linguistic pedantry about '9x smaller' phrasing being mathematically incorrect
  • ↓Skepticism about lack of comparison to existing quantization methods like Q2 or Unsloth
  • ↑Sharing benchmark results on personal hardware confirming impressive performance
3.Bend – A language that blocks AI mistakes via proof, on CPU and GPU(bend-lang.com)
591 points by nicolas-siplis 26 days ago | 302 comments | permalink
tl;dr: Bend is a programming language with Python-like syntax that compiles to fast native code, runs in parallel across CPU cores and GPUs, and includes a Lean/Rocq-style proof checker fast enough to run on every AI code change. Its key pitch is a `LAWS.bend` file where developers declare invariants that AI-generated code must mathematically prove it doesn't violate before committing, aimed at making "vibe-coded" apps trustworthy without human code review. It's early-stage, back-end focused, and works on Linux/macOS.
HN Discussion:
  • ↓AIs won't reliably follow the LAWS or will modify them to bypass constraints
  • ~The standard library of proofs is too sparse, requiring users to prove basic facts themselves
  • ↓Suspicious metrics (stars vs forks/issues) suggest the project's popularity is inflated or untrustworthy
  • ↑Technically interesting language design combining QTT and staging, worth exploring further
  • ↓Humans still need to write and validate the laws, so vibe-coding trust problem isn't solved
4.Qwen 3.8 Omni Flash(qwen.ai)
327 points by jjcm 26 days ago | 127 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • •Requests a tool to help pick appropriate models given the overwhelming variety
  • ↑Qwen offers dramatic cost reduction compared to Gemini with comparable performance
  • ↑Impressive if the audio benchmarks matching/exceeding Gemini Flash hold up
  • ↑Qwen models are reliable and well-designed, making them a preferred choice
  • ↓The blog post presentation is poor with technical issues and shallow content
5.Hister: A private search engine for the pages you visit and the files you keep(github.com)
696 points by bookofjoe 26 days ago | 191 comments | permalink
tl;dr: Hister is a self-hosted, private search engine that full-text indexes web pages you visit (via a Firefox/Chrome extension) and local files, letting you search them from a web UI, terminal, or AI assistant via MCP. It runs locally with no telemetry or mandatory cloud, supports field filters, wildcards, and optional semantic search through a configurable embeddings endpoint. Written in Go and licensed AGPLv3, it's installable via binary, Homebrew, Docker, or Nix.
HN Discussion:
  • ↑Author introduces project and shares background/motivation for building it
  • ~Concern about resource usage (CPU/battery) from indexing activities
  • •Sharing similar personal projects or alternative tools that solve the same problem
  • ↑Nostalgia for Chrome's old full-text history search feature validates the need
  • ↓Hesitation to install unvetted software due to supply-chain/security risks
6.Wax motor(en.wikipedia.org)
504 points by mhb 27 days ago | 106 comments | permalink
tl;dr: Wax motors are linear actuators that exploit the 5-20% volume expansion of wax during melting to convert thermal energy into mechanical force, capable of producing up to 4000 N. Compared to solenoids, they offer smooth actuation, resistive (non-inductive) electrical behavior, and can operate passively from ambient heat sources. Common applications include front-load washing machine door locks, dishwasher detergent dispensers, thermostatic mixing valves, hydronic zone valves, greenhouse vent controls, and aerospace fluid systems.
HN Discussion:
  • ~Points out a factual error or omission in the article/Wikipedia page about wax motors
  • ↑Shares supplementary demonstrations or videos of wax motors in action
  • ↑Adds additional real-world applications not mentioned in the article, like automotive thermostats and faucet valves
  • ↑Enthusiastic about alternative actuator technologies as promising future replacements for electromagnetic motors
  • •Shares personal experience with wax's expansion properties or draws analogies to similar simple mechanisms
7.Fujitsu launches made-in-Japan next-generation CPU FUJITSU-MONAKA(global.fujitsu)
626 points by my123 28 days ago | 246 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ~Points out the CPU is ARM-based, which the announcement downplays
  • ↓Questions the 'made in Japan' sovereignty claim since fabrication likely relies on TSMC
  • •Provides additional technical context and links about the Monaka chip
  • ↑Frames technology sovereignty push as a reasonable response to US political instability
  • ↓Skeptical of Fujitsu given past scandals and corporate behavior
8.A heap overflow and SSO misconfiguration to compromise OpenAI internal repos(hacktron.ai)
484 points by Handy-Man 26 days ago | 206 comments | permalink
tl;dr: Researchers at Hacktron chained a heap overflow in libheif (exploitable via image uploads to Discourse, which OpenAI's community forum runs on) with an OpenAI SSO misconfiguration to take over employee ChatGPT/Codex accounts and open a PR in OpenAI's internal monorepo—all within 72 hours, earning a $6,500 bounty. They leaned heavily on Claude Opus 5, which produced a reliable ASLR-enabled exploit within hours of release after Opus 4.8 had failed. The authors argue AI is collapsing the cost of weaponizing known memory-corruption bugs, eroding the "security through complexity" that previously protected typical targets.
HN Discussion:
  • ↑Attack surface of image parsers like libheif/ImageMagick is too large and needs safer alternatives or sandboxing
  • ~Bounty amount is insultingly low compared to black market value of such an exploit chain
  • ↑Bloated code and unsafe languages are the root problem; we need less software and safer languages like Rust
  • ↑Curiosity about how AI models achieved this capability so quickly, possibly due to machine-verifiable training signals
  • •Tangential wonder at why AI model weights themselves have never leaked despite lax security elsewhere
9.Why I didn’t sign the Fields medallists’ letter(gowers.wordpress.com)
274 points by simianwords 27 days ago | 400 comments | permalink
tl;dr: Timothy Gowers explains why he declined to sign a letter from 25 Fields medallists warning that AI's rapid problem-solving threatens mathematical understanding. He argues the letter overweights conceptual understanding relative to problem-solving, and that a flood of AI-generated results, while imperfectly digested, would likely increase both the volume of properly digested mathematics and the choice of results to study. His real concern is social: that young people will lose motivation to pursue math PhDs, funding will dry up, and the human mathematical tradition may not be transmitted to future generations.
HN Discussion:
  • ↑The letter fails to justify funding mathematicians merely for understanding without producing new proofs
  • ↑This reflects broader AI-displacement concerns about eroding social structures and training pipelines
  • ↓AI companies exploit curated mathematical problems as raw material without caring about the field's advancement
  • ~Process/journey of doing math has intrinsic value distinct from just producing results
  • ↑AI raises both the floor and ceiling of what mathematicians can tackle, expanding rather than replacing work
10.The American Religion of Self-Storage Facilities(newyorker.com)
257 points by pseudolus 26 days ago | 442 comments | permalink
tl;dr: The US dominates global self-storage with ~90% of capacity, and there are more facilities domestically than Starbucks, McDonald's, Walmart, Home Depot, Domino's, Dunkin', and Costco combined, generating over $40 billion annually. The industry, which took off in the 1970s and evolved into an institutional asset class, is driven by "the four D's": death, displacement, divorce, and downsizing. Texas leads all states in facility count, while Wichita was once dubbed the Self-Storage Capital of the US.
HN Discussion:
  • ↑Self-storage thrives as a business due to cheap buildout, low operating costs, and steady cashflow
  • ~Storage units serve legitimate practical purposes for people in small homes with hobbies or life transitions
  • ↑Storage units enable unhealthy consumerism and hoarding; better to declutter and avoid them entirely
  • ↑Frustration that self-storage facilities are built instead of much-needed housing
  • ↑Storage industry has consumer-hostile practices like opaque pricing and legal immunity from theft losses
11.Show HN: An e-ink frame that hears birds and draws them as 1800s illustrations(github.com)
2389 points by arnemunthekaas 28 days ago | 267 comments | permalink
tl;dr: A Raspberry Pi project that uses BirdNET-Go to detect birds by audio in real time and displays the matched species on an Inky Impression e-ink panel (or web kiosk) as hand-curated 1800s natural-history illustrations. The library includes 800+ public-domain cutouts covering 400+ species, arranged on a textured page with larger birds centered and sized by body mass. Currently biased toward Scandinavian/British/Central European species, with a live demo running from the author's kitchen in Bergen, Norway.
HN Discussion:
  • ↑General admiration for the project as magical and inspiring builder work
  • ↑Enthusiasm for e-ink displays as a versatile, joy-bringing medium for small projects
  • •Technical clarification that BirdNET is a traditional neural network, not an LLM
  • •Recognition of BirdNET-Go as the catalyst enabling a wave of bird-detection projects
  • ↑Sharing related BirdTech projects inspired by or parallel to this work
12.Show HN: Share your AI Setup, Learn from others(mysetup.ai)
231 points by steveybrown 26 days ago | 135 comments | permalink
tl;dr: mysetup.ai is a new community site for sharing AI tooling setups—agents, workflows, skills, and systems—created by someone who felt overwhelmed trying to piece together how others are actually working with AI from scattered Twitter posts. The creator admits it may end up being a community of one, but hopes others will contribute their own configurations.
HN Discussion:
  • ↓Refuses to connect MCP/GitHub credentials to third-party site; wants markdown submissions instead
  • ↓Sharing AI workflows is bad for job security in the AI era
  • ↑Likes the idea and actively contributed, suggesting feature improvements like RSS and surveys
  • ~Wants more useful metadata like monthly cost and local/low-resource setups
  • ~The underlying prompt is more valuable than the tool itself
13.Nvidia announces native GPU programming in Rust(developer.nvidia.com)
957 points by nonmaskable 27 days ago | 402 comments | permalink
tl;dr: NVIDIA is officially supporting native Rust for CUDA GPU kernel programming via two projects: cuda-oxide (a SIMT-style rustc codegen backend targeting PTX, requiring nightly) and cutile-rs (a higher-level tile-based model on stable Rust that JIT-compiles through CUDA Tile IR). Both leverage Rust's ownership model to catch aliasing bugs and data races at compile time, though cuda-oxide is early alpha while cutile-rs is already used in HuggingFace's Grout and mistral.rs. NVIDIA plans to mature these toolchains through 2027 alongside CUDA C++ and Python.
HN Discussion:
  • ↓CUDA is proprietary lock-in; GPU kernels should be written separately in vendor-neutral ways
  • ~GPU vendors should expose low-level ISA/documentation so developers can target hardware directly
  • ↑Rust's growing momentum across kernels, GPUs, and verified software is exciting and validating
  • ↑NVIDIA's Huggingface connection and existing Rust ecosystem make this a natural, promising step
  • ~Concerns about how single-file CPU/GPU code handles dependency configuration and compilation targets
14.How GLM built its own inference infrastructure(z.ai)
401 points by whiteros_e 27 days ago | 279 comments | permalink
tl;dr: Summary not available
HN Discussion:
  • ↑US chip export restrictions are accelerating China's domestic AI chip development
  • ↑Impressive technical achievement in building production inference on domestic accelerators
  • ~Skeptical whether the 100k accelerators are truly end-to-end domestic Chinese hardware
  • ↓Real-world GLM service performance is slow with strict usage limits, contradicting claims
  • ↓Article reads as positioning for government policy influence rather than genuine technical disclosure
15.CCC invites all model citizens to 40C3(events.ccc.de)
370 points by antonly 27 days ago | 223 comments | permalink
tl;dr: The Chaos Computer Club has announced 40C3, taking place December 27-30, 2026 at the Hamburg Exhibition halls under the motto "Model Citizens" — a call to develop new societal models amid rising authoritarianism. Calls for Participation are now open for talks, art, music, DJs, and punk performances at what remains Europe's largest non-commercial hacker conference, organized by volunteers and drawing over 16,000 attendees.
HN Discussion:
  • ~Personal experience at CCC was mixed with many small negative social interactions
  • ↑Recommending related smaller CCC events and enthusiasm for the community
  • ↓The scheduling over the holidays makes attendance difficult for many
  • •Concerns about ticket availability and accessibility for outsiders
  • ~Nostalgia for the older, edgier hacker culture versus today's more sanitized version
16.Cloudflare/Security-Audit-Skill(github.com)
208 points by donk8r 27 days ago | 38 comments | permalink
tl;dr: Cloudflare open-sourced a "skill" for coding agents that runs a structured six-phase security audit: reconnaissance, coverage-led hunting, adversarial validation by separate agents, schema-validated findings output, independent verification, and target-neutral reporting. It enforces principles like separating hunters from validators, requiring source-grounded confirmation, and treating defense-in-depth gaps as hardening notes rather than vulnerabilities. This is the single-repo precursor to Cloudflare's internal vulnerability discovery harness, installable via the Skills CLI and requiring Node.js plus a sandboxed environment for executing target code.
HN Discussion:
  • ↓Skill is token-inefficient, wasting context and money on codebases
  • ↓Design is bloated with too many schemas and fragmented skills polluting context
  • ~Alternative in-house approaches or reframing can work better than this skill
  • ↓Skepticism that a markdown-based 'skill' warrants this much attention
  • •Genuine curiosity about implementation details like sandboxing and skill vs prompt distinction
17.One year of sponsored Servo development(servo.org)
354 points by AshleysBrain 27 days ago | 144 comments | permalink
tl;dr: Servo maintainer Josh Bowman-Matthews (@jdm) recaps his first year in a part-time role funded entirely by OpenCollective/GitHub donations, during which he reviewed 1,150 PRs, filed 114 newcomer-friendly issues (92% fixed), nominated 8 new maintainers, and wrote onboarding documentation. Notable work included supporting a large rewrite of Servo's JS engine integration to fix GC-related panics and stabilizing flaky tests to smooth the contribution process.
HN Discussion:
  • ↑Additional funding sources like NLnet also support Servo development
  • ~Servo needs a corporate patron to adopt it in shipping products
  • ↓Questioning cost efficiency and whether SF-level salaries are justified
  • ↓Criticism that Servo still relies on C++ SpiderMonkey, undermining Rust's safety goals
  • ↓Skepticism about Servo's practical usefulness or purpose today
18.Training a 4B model to produce 81% faster query plans than Postgres(rohanbansal.com)
689 points by polyphilz 27 days ago | 143 comments | permalink
tl;dr: A developer post-trained a 4B parameter model (Qwen3.8-4B-Distill) using supervised fine-tuning on GPT-6 Astra trajectories plus custom GRPO reinforcement learning to generate Postgres query hints via the pg_hint_plan extension. The resulting agent achieved a 44.7% latency reduction across 113 join-heavy JOB benchmark queries, up from a baseline where the untrained model couldn't even produce valid plans for 99 of them. Key engineering work included building a low-noise Postgres measurement rig (tuning shared_buffers was critical) and designing reward signals robust to execution-time variance.
HN Discussion:
  • ↓Benchmark conditions are unrealistic and results likely don't generalize to real workloads
  • ↓LLMs are the wrong tool; query planning needs math/algorithms or specialized neural heuristics
  • ↓The speedup is likely explained by fixing misconfigured Postgres defaults rather than genuine LLM optimization
  • ↓Using hints papers over the real problem of missing indices and bad statistics
  • ↑Appreciation for the clear writing and accessible explanation of advanced ML concepts
19.Reversing Factorio's RNG(gegell.github.io)
251 points by jheitmann 31 days ago | 48 comments | permalink
tl;dr: Factorio uses the taus88 PRNG (three XORed LFSRs), which is linear over GF(2) and therefore trivially breakable—observing enough output bits lets you reconstruct the internal state via matrix inversion and predict all future "random" rolls. The author built an in-game implementation using repair pack recyclers as entropy sources (each craft leaks ~2 bits via 50/50 output splits), needing 44 recycling operations to recover the full 88-bit state and predict quality upgrade outcomes. Note: this works in Factorio 2.0 but was broken by RNG changes in 2.1.
HN Discussion:
  • ↑Admiration for the in-game combinator implementation beyond just the RNG cracking
  • ↑Criticism of Factorio's choice of an outdated PRNG and boost's verbose API
  • ↑Games should use CSPRNGs to prevent this class of reverse-engineering exploits
  • •Sharing related examples of RNG cracking in other games like Stardew Valley and Minecraft
  • ↑General appreciation for the article's depth and niche subject matter
20.My temporary PHP fix from 2014 has nearly 20M installs. Today I'm deprecating it(jakeasmith.com)
337 points by jakeasmith 28 days ago | 104 comments | permalink
tl;dr: A developer's 174-line PHP polyfill for `http_build_url()`, written in 2014 as a temporary shim during an AOL CMS upgrade, has been installed nearly 20 million times via Packagist and bundled into widely-used projects like WPML and Debian packages. After rediscovering the package (which contains an amusing bug that strips the letter "a" from paths with trailing slashes), the author is deprecating it rather than handing it off, citing supply-chain risks à la the xz backdoor and pointing users to PHP League's URI library or PHP 8.5's native URI API.
HN Discussion:
  • ↑Temporary fixes inevitably become permanent load-bearing infrastructure
  • ↑The 'a' bug is remarkably amusing and highlights how trivial polyfills shouldn't be external dependencies
  • ~Author should provide a final release with deprecation notice pointing to migration options
  • •PHP's ecosystem design (extensions/PECL) is what forced these one-off polyfill packages to exist
  • ↑Personal appreciation or acknowledgment of the author